[llvm] r371533 - AMDGPU/GlobalISel: First pass at attempting to legalize load/stores

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Tue Sep 10 09:20:14 PDT 2019


Added: llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir?rev=371533&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir (added)
+++ llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir Tue Sep 10 09:20:14 2019
@@ -0,0 +1,11361 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=SI %s
+
+#  Unaligned access is assumed on for HSA, but not mesa
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=bonaire -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=CI-HSA %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=CI-MESA %s
+
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefix=VI %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefixes=GFX9-HSA %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefixes=GFX9-MESA %s
+
+---
+name: test_load_global_s1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; SI: $vgpr0 = COPY [[AND]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s1_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; CI-HSA: $vgpr0 = COPY [[AND]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s1_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; CI-MESA: $vgpr0 = COPY [[AND]](s32)
+    ; VI-LABEL: name: test_load_global_s1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; VI: $vgpr0 = COPY [[AND]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s1_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; GFX9-HSA: $vgpr0 = COPY [[AND]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s1_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; GFX9-MESA: $vgpr0 = COPY [[AND]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s1) = G_LOAD %0 :: (load 1, align 1, addrspace 1)
+    %2:_(s32) = G_ZEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s2_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s2_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; SI: $vgpr0 = COPY [[AND]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s2_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; CI-HSA: $vgpr0 = COPY [[AND]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s2_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; CI-MESA: $vgpr0 = COPY [[AND]](s32)
+    ; VI-LABEL: name: test_load_global_s2_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; VI: $vgpr0 = COPY [[AND]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s2_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; GFX9-HSA: $vgpr0 = COPY [[AND]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s2_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; GFX9-MESA: $vgpr0 = COPY [[AND]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s2) = G_LOAD %0 :: (load 1, align 1, addrspace 1)
+    %2:_(s32) = G_ZEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_s8_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: $vgpr0 = COPY [[COPY1]](s32)
+    ; SI-LABEL: name: test_load_global_s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s8_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s8_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_global_s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s8_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s8_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s8) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s8_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s8_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_global_s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s8_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s8_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s8) = G_LOAD %0 :: (load 1, align 1, addrspace 1)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s16_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s16_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_global_s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s16_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s16_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s16) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s16_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s16_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_global_s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s16_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s16_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s16) = G_LOAD %0 :: (load 2, align 2, addrspace 1)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_s16_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[C1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[TRUNC]]
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[AND1]](s32)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC2]]
+    ; CI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; CI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; SI-LABEL: name: test_load_global_s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s16_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s16_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; VI-LABEL: name: test_load_global_s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s16_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s16_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s16) = G_LOAD %0 :: (load 2, align 1, addrspace 1)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_global_s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s32_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s32_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_load_global_s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s32_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s32_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_s32_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s32_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; SI: $vgpr0 = COPY [[MV]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s32_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s32_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[MV]](s32)
+    ; VI-LABEL: name: test_load_global_s32_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; VI: $vgpr0 = COPY [[MV]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s32_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s32_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[MV]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_s32_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s32_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: $vgpr0 = COPY [[MV]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_s32_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_s32_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[MV]](s32)
+    ; VI-LABEL: name: test_load_global_s32_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; VI: $vgpr0 = COPY [[MV]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_s32_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_s32_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[MV]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_s48_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_s48_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; CI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; CI: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; CI: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    ; SI-LABEL: name: test_load_global_s48_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; SI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; SI: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; SI: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    ; CI-HSA-LABEL: name: test_load_global_s48_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; CI-HSA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    ; CI-MESA-LABEL: name: test_load_global_s48_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    ; VI-LABEL: name: test_load_global_s48_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; VI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; VI: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; VI: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    ; GFX9-HSA-LABEL: name: test_load_global_s48_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    ; GFX9-MESA-LABEL: name: test_load_global_s48_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[AND]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s48) = G_LOAD %0 :: (load 6, align 8, addrspace 1)
+    %2:_(s64) = G_ZEXT %1
+    $vgpr0_vgpr1 = COPY %2
+...
+
+---
+name: test_load_global_s64_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s64_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_load_global_s64_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_load_global_s64_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_load_global_s64_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_load_global_s64_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_load_global_s64_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_s64_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s64_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_load_global_s64_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_load_global_s64_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_load_global_s64_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_load_global_s64_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_load_global_s64_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_s64_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s64_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; CI-HSA-LABEL: name: test_load_global_s64_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_load_global_s64_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; VI-LABEL: name: test_load_global_s64_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; GFX9-HSA-LABEL: name: test_load_global_s64_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_load_global_s64_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_s64_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s64_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; CI-HSA-LABEL: name: test_load_global_s64_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_load_global_s64_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; VI-LABEL: name: test_load_global_s64_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; GFX9-HSA-LABEL: name: test_load_global_s64_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_load_global_s64_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_s96_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s96_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-HSA-LABEL: name: test_load_global_s96_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-MESA-LABEL: name: test_load_global_s96_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; VI-LABEL: name: test_load_global_s96_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-HSA-LABEL: name: test_load_global_s96_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-MESA-LABEL: name: test_load_global_s96_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_global_s96_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s96_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-HSA-LABEL: name: test_load_global_s96_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-MESA-LABEL: name: test_load_global_s96_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; VI-LABEL: name: test_load_global_s96_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-HSA-LABEL: name: test_load_global_s96_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-MESA-LABEL: name: test_load_global_s96_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_global_s96_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s96_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-HSA-LABEL: name: test_load_global_s96_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-MESA-LABEL: name: test_load_global_s96_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; VI-LABEL: name: test_load_global_s96_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-HSA-LABEL: name: test_load_global_s96_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-MESA-LABEL: name: test_load_global_s96_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_global_s96_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s96_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; CI-HSA-LABEL: name: test_load_global_s96_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-MESA-LABEL: name: test_load_global_s96_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; VI-LABEL: name: test_load_global_s96_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; GFX9-HSA-LABEL: name: test_load_global_s96_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-MESA-LABEL: name: test_load_global_s96_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 2, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_global_s96_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s96_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; SI: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C13]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C13]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C13]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C13]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C13]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C13]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C12]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; CI-HSA-LABEL: name: test_load_global_s96_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; CI-MESA-LABEL: name: test_load_global_s96_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; CI-MESA: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C13]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C13]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C13]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C13]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C13]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C13]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C12]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; VI-LABEL: name: test_load_global_s96_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C11]]
+    ; VI: [[C12:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C12]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C11]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C12]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C11]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C12]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C11]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C12]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C11]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C12]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C11]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C12]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; GFX9-HSA-LABEL: name: test_load_global_s96_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s96) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](s96)
+    ; GFX9-MESA-LABEL: name: test_load_global_s96_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C11]]
+    ; GFX9-MESA: [[C12:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C12]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C11]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C12]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C11]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C12]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C11]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C12]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C11]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C12]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C11]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C12]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 1, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_global_s160_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_s160_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; CI: S_NOP 0, implicit [[TRUNC]](s160)
+    ; SI-LABEL: name: test_load_global_s160_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; SI: S_NOP 0, implicit [[TRUNC]](s160)
+    ; CI-HSA-LABEL: name: test_load_global_s160_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; CI-HSA: S_NOP 0, implicit [[TRUNC]](s160)
+    ; CI-MESA-LABEL: name: test_load_global_s160_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; CI-MESA: S_NOP 0, implicit [[TRUNC]](s160)
+    ; VI-LABEL: name: test_load_global_s160_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; VI: S_NOP 0, implicit [[TRUNC]](s160)
+    ; GFX9-HSA-LABEL: name: test_load_global_s160_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; GFX9-HSA: S_NOP 0, implicit [[TRUNC]](s160)
+    ; GFX9-MESA-LABEL: name: test_load_global_s160_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
+    ; GFX9-MESA: S_NOP 0, implicit [[TRUNC]](s160)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s160) = G_LOAD %0 :: (load 20, align 4, addrspace 1)
+    S_NOP 0, implicit %1
+...
+
+---
+name: test_load_global_s224_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s224_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 28, align 4, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s224) = G_TRUNC [[LOAD]](s256)
+    ; SI: [[DEF:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(s256) = G_INSERT [[DEF]], [[TRUNC]](s224), 0
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](s256)
+    ; CI-HSA-LABEL: name: test_load_global_s224_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 28, align 4, addrspace 1)
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s224) = G_TRUNC [[LOAD]](s256)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(s256) = G_INSERT [[DEF]], [[TRUNC]](s224), 0
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](s256)
+    ; CI-MESA-LABEL: name: test_load_global_s224_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 28, align 4, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s224) = G_TRUNC [[LOAD]](s256)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(s256) = G_INSERT [[DEF]], [[TRUNC]](s224), 0
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](s256)
+    ; VI-LABEL: name: test_load_global_s224_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 28, align 4, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s224) = G_TRUNC [[LOAD]](s256)
+    ; VI: [[DEF:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(s256) = G_INSERT [[DEF]], [[TRUNC]](s224), 0
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](s256)
+    ; GFX9-HSA-LABEL: name: test_load_global_s224_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 28, align 4, addrspace 1)
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s224) = G_TRUNC [[LOAD]](s256)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(s256) = G_INSERT [[DEF]], [[TRUNC]](s224), 0
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](s256)
+    ; GFX9-MESA-LABEL: name: test_load_global_s224_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 28, align 4, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s224) = G_TRUNC [[LOAD]](s256)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(s256) = G_INSERT [[DEF]], [[TRUNC]](s224), 0
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](s256)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s224) = G_LOAD %0 :: (load 28, align 4, addrspace 1)
+     %2:_(s256) = G_IMPLICIT_DEF
+     %3:_(s256) = G_INSERT %2, %1, 0
+     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
+
+...
+
+---
+name: test_load_global_s128_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s128_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-HSA-LABEL: name: test_load_global_s128_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-MESA-LABEL: name: test_load_global_s128_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; VI-LABEL: name: test_load_global_s128_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-HSA-LABEL: name: test_load_global_s128_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-MESA-LABEL: name: test_load_global_s128_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_s128_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_s128_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; SI-LABEL: name: test_load_global_s128_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-HSA-LABEL: name: test_load_global_s128_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-MESA-LABEL: name: test_load_global_s128_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; VI-LABEL: name: test_load_global_s128_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-HSA-LABEL: name: test_load_global_s128_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-MESA-LABEL: name: test_load_global_s128_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_s128_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s128_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 13
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 15
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C15:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C15]]
+    ; SI: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C17]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C15]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C17]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C15]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C17]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C15]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C17]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C15]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C17]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C15]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C17]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C15]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C17]]
+    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C15]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C17]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C16]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; CI-HSA-LABEL: name: test_load_global_s128_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-MESA-LABEL: name: test_load_global_s128_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 13
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 15
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C15:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C15]]
+    ; CI-MESA: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C17]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C15]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C17]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C15]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C17]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C15]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C17]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C15]]
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C17]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C15]]
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C17]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C15]]
+    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C16]](s32)
+    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C17]]
+    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C15]]
+    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C17]]
+    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C16]](s32)
+    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; VI-LABEL: name: test_load_global_s128_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 13
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 15
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C15:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C15]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C15]]
+    ; VI: [[C16:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C16]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C15]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C15]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C16]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C15]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C15]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C16]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C15]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C15]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C16]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C15]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C15]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C16]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C15]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C15]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C16]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C15]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C15]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C16]](s16)
+    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C15]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C15]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C16]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; GFX9-HSA-LABEL: name: test_load_global_s128_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-MESA-LABEL: name: test_load_global_s128_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 13
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 15
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C15:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C15]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C15]]
+    ; GFX9-MESA: [[C16:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C16]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C15]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C15]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C16]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C15]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C15]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C16]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C15]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C15]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C16]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C15]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C15]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C16]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C15]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C15]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C16]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C15]]
+    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C15]]
+    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C16]](s16)
+    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C15]]
+    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C15]]
+    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C16]](s16)
+    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 1, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_s256_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_s256_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](s256)
+    ; CI-HSA-LABEL: name: test_load_global_s256_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](s256)
+    ; CI-MESA-LABEL: name: test_load_global_s256_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](s256)
+    ; VI-LABEL: name: test_load_global_s256_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](s256)
+    ; GFX9-HSA-LABEL: name: test_load_global_s256_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](s256)
+    ; GFX9-MESA-LABEL: name: test_load_global_s256_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](s256)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s256) = G_LOAD %0 :: (load 32, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_p1_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p1_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; CI-HSA-LABEL: name: test_load_global_p1_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; CI-MESA-LABEL: name: test_load_global_p1_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; VI-LABEL: name: test_load_global_p1_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; GFX9-HSA-LABEL: name: test_load_global_p1_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; GFX9-MESA-LABEL: name: test_load_global_p1_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; CI-HSA-LABEL: name: test_load_global_p1_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; CI-MESA-LABEL: name: test_load_global_p1_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; VI-LABEL: name: test_load_global_p1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; GFX9-HSA-LABEL: name: test_load_global_p1_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; GFX9-MESA-LABEL: name: test_load_global_p1_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; CI-HSA-LABEL: name: test_load_global_p1_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; CI-MESA-LABEL: name: test_load_global_p1_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; VI-LABEL: name: test_load_global_p1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; GFX9-HSA-LABEL: name: test_load_global_p1_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
+    ; GFX9-MESA-LABEL: name: test_load_global_p1_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p3_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p3_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](p3)
+    ; CI-HSA-LABEL: name: test_load_global_p3_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p3)
+    ; CI-MESA-LABEL: name: test_load_global_p3_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](p3)
+    ; VI-LABEL: name: test_load_global_p3_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](p3)
+    ; GFX9-HSA-LABEL: name: test_load_global_p3_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p3)
+    ; GFX9-MESA-LABEL: name: test_load_global_p3_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](p3)
+     %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p3) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_p4_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_p4_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; SI-LABEL: name: test_load_global_p4_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; CI-HSA-LABEL: name: test_load_global_p4_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; CI-MESA-LABEL: name: test_load_global_p4_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; VI-LABEL: name: test_load_global_p4_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; GFX9-HSA-LABEL: name: test_load_global_p4_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; GFX9-MESA-LABEL: name: test_load_global_p4_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p4) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; CI-HSA-LABEL: name: test_load_global_p4_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; CI-MESA-LABEL: name: test_load_global_p4_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; VI-LABEL: name: test_load_global_p4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; GFX9-HSA-LABEL: name: test_load_global_p4_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; GFX9-MESA-LABEL: name: test_load_global_p4_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p4) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p4_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p4_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    ; CI-HSA-LABEL: name: test_load_global_p4_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; CI-MESA-LABEL: name: test_load_global_p4_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    ; VI-LABEL: name: test_load_global_p4_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    ; GFX9-HSA-LABEL: name: test_load_global_p4_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; GFX9-MESA-LABEL: name: test_load_global_p4_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p4) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p4_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p4_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    ; CI-HSA-LABEL: name: test_load_global_p4_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; CI-MESA-LABEL: name: test_load_global_p4_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    ; VI-LABEL: name: test_load_global_p4_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    ; GFX9-HSA-LABEL: name: test_load_global_p4_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
+    ; GFX9-MESA-LABEL: name: test_load_global_p4_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p4) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_p5_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p5_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](p5)
+    ; CI-HSA-LABEL: name: test_load_global_p5_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p5)
+    ; CI-MESA-LABEL: name: test_load_global_p5_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](p5)
+    ; VI-LABEL: name: test_load_global_p5_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](p5)
+    ; GFX9-HSA-LABEL: name: test_load_global_p5_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p5)
+    ; GFX9-MESA-LABEL: name: test_load_global_p5_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](p5)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p5) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_p5_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p5_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; SI: $vgpr0 = COPY [[MV]](p5)
+    ; CI-HSA-LABEL: name: test_load_global_p5_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p5)
+    ; CI-MESA-LABEL: name: test_load_global_p5_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[MV]](p5)
+    ; VI-LABEL: name: test_load_global_p5_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; VI: $vgpr0 = COPY [[MV]](p5)
+    ; GFX9-HSA-LABEL: name: test_load_global_p5_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p5)
+    ; GFX9-MESA-LABEL: name: test_load_global_p5_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[MV]](p5)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p5) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_p5_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_p5_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: $vgpr0 = COPY [[MV]](p5)
+    ; CI-HSA-LABEL: name: test_load_global_p5_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p5)
+    ; CI-MESA-LABEL: name: test_load_global_p5_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[MV]](p5)
+    ; VI-LABEL: name: test_load_global_p5_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; VI: $vgpr0 = COPY [[MV]](p5)
+    ; GFX9-HSA-LABEL: name: test_load_global_p5_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p5)
+    ; GFX9-MESA-LABEL: name: test_load_global_p5_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[MV]](p5)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(p5) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v2s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; SI: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; SI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_v2s8_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_v2s8_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-MESA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; VI-LABEL: name: test_load_global_v2s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; VI: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; VI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s8_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s8_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-MESA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
+    %2:_(s16) = G_BITCAST %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_global_v2s8_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s8_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; SI: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; SI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_v2s8_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_v2s8_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-MESA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; VI-LABEL: name: test_load_global_v2s8_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; VI: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; VI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s8_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s8_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-MESA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 2, addrspace 1)
+    %2:_(s16) = G_BITCAST %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_global_v2s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32)
+    ; SI: [[TRUNC:%[0-9]+]]:_(<2 x s8>) = G_TRUNC [[BUILD_VECTOR]](<2 x s32>)
+    ; SI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[TRUNC]](<2 x s8>)
+    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-HSA-LABEL: name: test_load_global_v2s8_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-MESA-LABEL: name: test_load_global_v2s8_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<2 x s8>) = G_TRUNC [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[TRUNC]](<2 x s8>)
+    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; VI-LABEL: name: test_load_global_v2s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32)
+    ; VI: [[TRUNC:%[0-9]+]]:_(<2 x s8>) = G_TRUNC [[BUILD_VECTOR]](<2 x s32>)
+    ; VI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[TRUNC]](<2 x s8>)
+    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s8_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s8_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<2 x s8>) = G_TRUNC [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[TRUNC]](<2 x s8>)
+    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 1, addrspace 1)
+    %2:_(s16) = G_BITCAST %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_global_v3s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; SI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s8_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s8_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; VI-LABEL: name: test_load_global_v3s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; VI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s8_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s8_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, align 4, addrspace 1)
+    %2:_(<4 x s8>) = G_IMPLICIT_DEF
+    %3:_(<4 x s8>) = G_INSERT %2, %1, 0
+    $vgpr0 = COPY %3
+
+...
+
+---
+name: test_load_global_v3s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; SI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s8_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s8_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; VI-LABEL: name: test_load_global_v3s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; VI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s8_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s8_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, align 1, addrspace 1)
+    %2:_(<4 x s8>) = G_IMPLICIT_DEF
+    %3:_(<4 x s8>) = G_INSERT %2, %1, 0
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_global_v4s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s8_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s8_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; VI-LABEL: name: test_load_global_v4s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s8_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s8_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v4s8_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_v4s8_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; SI-LABEL: name: test_load_global_v4s8_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s8_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s8_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; VI-LABEL: name: test_load_global_v4s8_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s8_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s8_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v4s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s8_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s8_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; VI-LABEL: name: test_load_global_v4s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s8_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s8_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v8s8_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v8s8_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<8 x s8>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<8 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v8s8_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s8>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<8 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v8s8_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s8>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<8 x s8>)
+    ; VI-LABEL: name: test_load_global_v8s8_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<8 x s8>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<8 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v8s8_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s8>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<8 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v8s8_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s8>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<8 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v16s8_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v16s8_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<16 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v16s8_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<16 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v16s8_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<16 x s8>)
+    ; VI-LABEL: name: test_load_global_v16s8_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<16 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v16s8_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<16 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v16s8_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<16 x s8>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<16 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v32s8_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v32s8_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<32 x s8>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<32 x s8>)
+    ; CI-HSA-LABEL: name: test_load_global_v32s8_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<32 x s8>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<32 x s8>)
+    ; CI-MESA-LABEL: name: test_load_global_v32s8_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<32 x s8>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<32 x s8>)
+    ; VI-LABEL: name: test_load_global_v32s8_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<32 x s8>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<32 x s8>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v32s8_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<32 x s8>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<32 x s8>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v32s8_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<32 x s8>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<32 x s8>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<32 x s8>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+
+name: test_load_global_v2s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s16_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s16_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; VI-LABEL: name: test_load_global_v2s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s16_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s16_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v2s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; SI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s16_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s16_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; VI-LABEL: name: test_load_global_v2s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; VI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s16_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s16_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v2s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_v2s16_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[C1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[TRUNC]]
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[AND1]](s32)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC2]]
+    ; CI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[TRUNC]]
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C6]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C4]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[AND4]](s32)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND3]], [[TRUNC4]]
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; CI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; SI-LABEL: name: test_load_global_v2s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; SI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s16_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s16_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; VI-LABEL: name: test_load_global_v2s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; VI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s16_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s16_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_global_v3s16_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s16_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; SI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s16_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s16_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v3s16_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; VI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 8, addrspace 1)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_global_v3s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_v3s16_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; SI-LABEL: name: test_load_global_v3s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; SI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s16_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s16_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v3s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; VI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 4, addrspace 1)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_global_v3s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; SI: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; SI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s16_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 2, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s16_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v3s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; VI: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; VI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 2, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 2, addrspace 1)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_global_v3s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[DEF]](<4 x s16>), 0
+    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[EXTRACT]](<3 x s16>), 0
+    ; SI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR]](<2 x s16>), 0
+    ; SI: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT1]](<4 x s16>), 0
+    ; SI: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; SI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR2]](s16), 32
+    ; SI: [[EXTRACT2:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT3]](<4 x s16>), 0
+    ; SI: [[DEF3:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF3]], [[EXTRACT2]](<3 x s16>), 0
+    ; SI: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s16_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 1, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s16_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[DEF]](<4 x s16>), 0
+    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[EXTRACT]](<3 x s16>), 0
+    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR]](<2 x s16>), 0
+    ; CI-MESA: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT1]](<4 x s16>), 0
+    ; CI-MESA: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR2]](s16), 32
+    ; CI-MESA: [[EXTRACT2:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT3]](<4 x s16>), 0
+    ; CI-MESA: [[DEF3:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF3]], [[EXTRACT2]](<3 x s16>), 0
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v3s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[DEF]](<4 x s16>), 0
+    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[EXTRACT]](<3 x s16>), 0
+    ; VI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR]](<2 x s16>), 0
+    ; VI: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT1]](<4 x s16>), 0
+    ; VI: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; VI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR2]](s16), 32
+    ; VI: [[EXTRACT2:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT3]](<4 x s16>), 0
+    ; VI: [[DEF3:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF3]], [[EXTRACT2]](<3 x s16>), 0
+    ; VI: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 1, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[EXTRACT:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[DEF]](<4 x s16>), 0
+    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[EXTRACT]](<3 x s16>), 0
+    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR]](<2 x s16>), 0
+    ; GFX9-MESA: [[EXTRACT1:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT1]](<4 x s16>), 0
+    ; GFX9-MESA: [[DEF2:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF2]], [[EXTRACT1]](<3 x s16>), 0
+    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR2]](s16), 32
+    ; GFX9-MESA: [[EXTRACT2:%[0-9]+]]:_(<3 x s16>) = G_EXTRACT [[INSERT3]](<4 x s16>), 0
+    ; GFX9-MESA: [[DEF3:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF3]], [[EXTRACT2]](<3 x s16>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 1, addrspace 1)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_global_v4s16_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s16_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s16_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s16_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v4s16_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v4s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s16_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s16_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v4s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v4s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s16_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s16_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v4s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v4s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C4]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s16_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s16_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C4]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; VI-LABEL: name: test_load_global_v4s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C3]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v8s16_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v8s16_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 8, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, align 4, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, align 8, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, align 4, addrspace 1)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    ; CI-HSA-LABEL: name: test_load_global_v8s16_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 8, addrspace 1)
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-HSA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-HSA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-HSA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-HSA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-HSA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, align 8, addrspace 1)
+    ; CI-HSA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-HSA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-HSA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-HSA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; CI-HSA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI-HSA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-HSA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-HSA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-HSA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-HSA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; CI-HSA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-HSA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; CI-HSA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    ; CI-MESA-LABEL: name: test_load_global_v8s16_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 8, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, align 8, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    ; VI-LABEL: name: test_load_global_v8s16_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 8, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, align 4, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, align 8, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, align 4, addrspace 1)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v8s16_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 8, addrspace 1)
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-HSA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-HSA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-HSA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-HSA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, align 8, addrspace 1)
+    ; GFX9-HSA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-HSA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-HSA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; GFX9-HSA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-HSA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-HSA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-HSA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-HSA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-HSA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; GFX9-HSA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-HSA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; GFX9-HSA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v8s16_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 8, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, align 8, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 14
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<8 x s16>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2s32_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s32_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s32_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s32_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; VI-LABEL: name: test_load_global_v2s32_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v2s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s32_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s32_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; VI-LABEL: name: test_load_global_v2s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+
+---
+name: test_load_global_v2s32_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s32_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s32_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s32_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; VI-LABEL: name: test_load_global_v2s32_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v3s32_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s32_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s32_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s32_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; VI-LABEL: name: test_load_global_v3s32_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s32_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s32_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+
+
+---
+name: test_load_global_v3s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s32_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s32_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; VI-LABEL: name: test_load_global_v3s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s32_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s32_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_global_v4s32_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s32_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s32_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s32_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; VI-LABEL: name: test_load_global_v4s32_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s32_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s32_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v4s32_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s32_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s32_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s32_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; VI-LABEL: name: test_load_global_v4s32_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s32_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s32_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v4s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s32_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s32_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; VI-LABEL: name: test_load_global_v4s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s32_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s32_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v8s32_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v8s32_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v8s32_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI-HSA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-HSA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI-HSA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; CI-HSA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-HSA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; CI-HSA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-HSA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-HSA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; CI-HSA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-HSA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v8s32_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>)
+    ; VI-LABEL: name: test_load_global_v8s32_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v8s32_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; GFX9-HSA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-HSA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; GFX9-HSA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-HSA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; GFX9-HSA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-HSA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-HSA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; GFX9-HSA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-HSA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v8s32_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<8 x s32>) = G_LOAD %0 :: (load 16, align 32, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_v16s32_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_load_global_v16s32_align32
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; CI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; CI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; CI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; CI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; CI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; CI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; CI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; CI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; CI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; CI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; CI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; CI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    ; SI-LABEL: name: test_load_global_v16s32_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; SI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; SI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    ; CI-HSA-LABEL: name: test_load_global_v16s32_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI-HSA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-HSA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI-HSA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; CI-HSA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-HSA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; CI-HSA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-HSA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-HSA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; CI-HSA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-HSA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; CI-HSA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; CI-HSA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI-HSA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; CI-HSA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; CI-HSA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; CI-HSA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; CI-HSA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-HSA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; CI-HSA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-HSA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; CI-HSA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; CI-HSA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI-HSA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; CI-HSA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; CI-HSA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; CI-HSA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; CI-HSA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-HSA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; CI-HSA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; CI-HSA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    ; CI-MESA-LABEL: name: test_load_global_v16s32_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; CI-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; CI-MESA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    ; VI-LABEL: name: test_load_global_v16s32_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; VI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; VI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v16s32_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; GFX9-HSA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-HSA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; GFX9-HSA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-HSA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; GFX9-HSA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-HSA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-HSA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; GFX9-HSA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-HSA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; GFX9-HSA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; GFX9-HSA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; GFX9-HSA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; GFX9-HSA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; GFX9-HSA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; GFX9-HSA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-HSA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-HSA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; GFX9-HSA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-HSA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; GFX9-HSA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; GFX9-HSA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; GFX9-HSA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; GFX9-HSA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; GFX9-HSA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; GFX9-HSA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; GFX9-HSA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-HSA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; GFX9-HSA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; GFX9-HSA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v16s32_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
+    ; GFX9-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
+    ; GFX9-MESA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<16 x s32>) = G_LOAD %0 :: (load 16, align 32, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %1
+...
+
+---
+name: test_load_global_v2s64_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s64_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s64_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s64_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; VI-LABEL: name: test_load_global_v2s64_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2s64_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s64_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s64_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s64_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; VI-LABEL: name: test_load_global_v2s64_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2s64_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s64_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s64_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s64_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; VI-LABEL: name: test_load_global_v2s64_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2s64_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s64_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s64_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 2, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s64_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; VI-LABEL: name: test_load_global_v2s64_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 2, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16), [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 2, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2s64_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s64_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s64_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s64_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; VI-LABEL: name: test_load_global_v2s64_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 1, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v3s64_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s64_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 32, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s64_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 32, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s64_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 32, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; VI-LABEL: name: test_load_global_v3s64_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 32, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s64_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 32, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s64_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 32, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 32, addrspace 1)
+    %2:_(<4 x s64>) = G_IMPLICIT_DEF
+    %3:_(<4 x s64>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
+...
+
+---
+name: test_load_global_v3s64_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s64_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 8, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s64_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 8, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s64_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 8, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; VI-LABEL: name: test_load_global_v3s64_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 8, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s64_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 8, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s64_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 8, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 8, addrspace 1)
+    %2:_(<4 x s64>) = G_IMPLICIT_DEF
+    %3:_(<4 x s64>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
+...
+
+---
+name: test_load_global_v3s64_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v3s64_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; SI: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; SI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
+    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
+    ; SI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; SI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; SI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; SI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
+    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
+    ; SI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; SI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; SI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
+    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
+    ; SI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; SI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; SI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; SI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
+    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
+    ; SI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; SI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64)
+    ; SI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<3 x s64>), 0
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v3s64_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 1, addrspace 1)
+    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v3s64_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI-MESA: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; CI-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; CI-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; CI-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; CI-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; CI-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; CI-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; CI-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; CI-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; CI-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
+    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
+    ; CI-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; CI-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; CI-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; CI-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
+    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
+    ; CI-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; CI-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; CI-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
+    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
+    ; CI-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; CI-MESA: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; CI-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; CI-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; CI-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
+    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
+    ; CI-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; CI-MESA: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; CI-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<3 x s64>), 0
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; VI-LABEL: name: test_load_global_v3s64_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; VI: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; VI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; VI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; VI: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
+    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
+    ; VI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; VI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; VI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; VI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; VI: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
+    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
+    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; VI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; VI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; VI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; VI: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
+    ; VI: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
+    ; VI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; VI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; VI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; VI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; VI: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
+    ; VI: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
+    ; VI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64)
+    ; VI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<3 x s64>), 0
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v3s64_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s64>) = G_LOAD [[COPY]](p1) :: (load 24, align 1, addrspace 1)
+    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[LOAD]](<3 x s64>), 0
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v3s64_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; GFX9-MESA: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; GFX9-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; GFX9-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; GFX9-MESA: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
+    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
+    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; GFX9-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; GFX9-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; GFX9-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; GFX9-MESA: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
+    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
+    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; GFX9-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; GFX9-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; GFX9-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; GFX9-MESA: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
+    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
+    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; GFX9-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; GFX9-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; GFX9-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; GFX9-MESA: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
+    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
+    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; GFX9-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<3 x s64>), 0
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 1, addrspace 1)
+    %2:_(<4 x s64>) = G_IMPLICIT_DEF
+    %3:_(<4 x s64>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
+...
+
+---
+name: test_load_global_v4s64_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s64_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s64_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s64_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; VI-LABEL: name: test_load_global_v4s64_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s64_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s64_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_v4s64_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s64_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s64_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s64_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; VI-LABEL: name: test_load_global_v4s64_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s64_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s64_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_v4s64_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4s64_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; SI: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; SI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
+    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
+    ; SI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; SI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; SI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; SI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
+    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
+    ; SI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; SI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; SI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
+    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
+    ; SI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; SI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; SI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; SI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
+    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
+    ; SI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; SI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; SI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; SI: [[GEP23:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; SI: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[GEP23]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP24:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C]](s64)
+    ; SI: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[GEP24]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP25:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C1]](s64)
+    ; SI: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[GEP25]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP26:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C2]](s64)
+    ; SI: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[GEP26]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP27:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C3]](s64)
+    ; SI: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[GEP27]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP28:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C4]](s64)
+    ; SI: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[GEP28]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP29:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C5]](s64)
+    ; SI: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[GEP29]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP30:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C6]](s64)
+    ; SI: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[GEP30]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
+    ; SI: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
+    ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD25]](s32)
+    ; SI: [[AND25:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C9]]
+    ; SI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND25]], [[COPY24]](s32)
+    ; SI: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[SHL12]](s32)
+    ; SI: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[TRUNC25]]
+    ; SI: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
+    ; SI: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
+    ; SI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LOAD27]](s32)
+    ; SI: [[AND27:%[0-9]+]]:_(s32) = G_AND [[COPY27]], [[C9]]
+    ; SI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND27]], [[COPY26]](s32)
+    ; SI: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[SHL13]](s32)
+    ; SI: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[TRUNC27]]
+    ; SI: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
+    ; SI: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
+    ; SI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LOAD29]](s32)
+    ; SI: [[AND29:%[0-9]+]]:_(s32) = G_AND [[COPY29]], [[C9]]
+    ; SI: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND29]], [[COPY28]](s32)
+    ; SI: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[SHL14]](s32)
+    ; SI: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[TRUNC29]]
+    ; SI: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
+    ; SI: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
+    ; SI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LOAD31]](s32)
+    ; SI: [[AND31:%[0-9]+]]:_(s32) = G_AND [[COPY31]], [[C9]]
+    ; SI: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND31]], [[COPY30]](s32)
+    ; SI: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[SHL15]](s32)
+    ; SI: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[TRUNC31]]
+    ; SI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR12]](s16), [[OR13]](s16), [[OR14]](s16), [[OR15]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
+    ; CI-HSA-LABEL: name: test_load_global_v4s64_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; CI-MESA-LABEL: name: test_load_global_v4s64_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CI-MESA: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; CI-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; CI-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; CI-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; CI-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; CI-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; CI-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; CI-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; CI-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; CI-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
+    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
+    ; CI-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; CI-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; CI-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; CI-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
+    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
+    ; CI-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; CI-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; CI-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
+    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
+    ; CI-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; CI-MESA: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; CI-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; CI-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; CI-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
+    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
+    ; CI-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; CI-MESA: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; CI-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; CI-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; CI-MESA: [[GEP23:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C12]](s64)
+    ; CI-MESA: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[GEP23]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP24:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C]](s64)
+    ; CI-MESA: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[GEP24]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP25:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C1]](s64)
+    ; CI-MESA: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[GEP25]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP26:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C2]](s64)
+    ; CI-MESA: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[GEP26]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP27:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C3]](s64)
+    ; CI-MESA: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[GEP27]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP28:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C4]](s64)
+    ; CI-MESA: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[GEP28]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP29:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C5]](s64)
+    ; CI-MESA: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[GEP29]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP30:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C6]](s64)
+    ; CI-MESA: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[GEP30]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
+    ; CI-MESA: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
+    ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD25]](s32)
+    ; CI-MESA: [[AND25:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C9]]
+    ; CI-MESA: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND25]], [[COPY24]](s32)
+    ; CI-MESA: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[SHL12]](s32)
+    ; CI-MESA: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[TRUNC25]]
+    ; CI-MESA: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
+    ; CI-MESA: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
+    ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LOAD27]](s32)
+    ; CI-MESA: [[AND27:%[0-9]+]]:_(s32) = G_AND [[COPY27]], [[C9]]
+    ; CI-MESA: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND27]], [[COPY26]](s32)
+    ; CI-MESA: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[SHL13]](s32)
+    ; CI-MESA: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[TRUNC27]]
+    ; CI-MESA: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
+    ; CI-MESA: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
+    ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LOAD29]](s32)
+    ; CI-MESA: [[AND29:%[0-9]+]]:_(s32) = G_AND [[COPY29]], [[C9]]
+    ; CI-MESA: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND29]], [[COPY28]](s32)
+    ; CI-MESA: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[SHL14]](s32)
+    ; CI-MESA: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[TRUNC29]]
+    ; CI-MESA: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
+    ; CI-MESA: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
+    ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LOAD31]](s32)
+    ; CI-MESA: [[AND31:%[0-9]+]]:_(s32) = G_AND [[COPY31]], [[C9]]
+    ; CI-MESA: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND31]], [[COPY30]](s32)
+    ; CI-MESA: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[SHL15]](s32)
+    ; CI-MESA: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[TRUNC31]]
+    ; CI-MESA: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR12]](s16), [[OR13]](s16), [[OR14]](s16), [[OR15]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
+    ; VI-LABEL: name: test_load_global_v4s64_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; VI: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; VI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; VI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; VI: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
+    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
+    ; VI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; VI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; VI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; VI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; VI: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
+    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
+    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; VI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; VI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; VI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; VI: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
+    ; VI: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
+    ; VI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; VI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; VI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; VI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; VI: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
+    ; VI: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
+    ; VI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; VI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; VI: [[GEP23:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; VI: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[GEP23]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP24:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C]](s64)
+    ; VI: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[GEP24]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP25:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C1]](s64)
+    ; VI: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[GEP25]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP26:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C2]](s64)
+    ; VI: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[GEP26]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP27:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C3]](s64)
+    ; VI: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[GEP27]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP28:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C4]](s64)
+    ; VI: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[GEP28]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP29:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C5]](s64)
+    ; VI: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[GEP29]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP30:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C6]](s64)
+    ; VI: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[GEP30]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
+    ; VI: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
+    ; VI: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD25]](s32)
+    ; VI: [[AND25:%[0-9]+]]:_(s16) = G_AND [[TRUNC25]], [[C7]]
+    ; VI: [[SHL12:%[0-9]+]]:_(s16) = G_SHL [[AND25]], [[C8]](s16)
+    ; VI: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[SHL12]]
+    ; VI: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
+    ; VI: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
+    ; VI: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD27]](s32)
+    ; VI: [[AND27:%[0-9]+]]:_(s16) = G_AND [[TRUNC27]], [[C7]]
+    ; VI: [[SHL13:%[0-9]+]]:_(s16) = G_SHL [[AND27]], [[C8]](s16)
+    ; VI: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[SHL13]]
+    ; VI: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
+    ; VI: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
+    ; VI: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD29]](s32)
+    ; VI: [[AND29:%[0-9]+]]:_(s16) = G_AND [[TRUNC29]], [[C7]]
+    ; VI: [[SHL14:%[0-9]+]]:_(s16) = G_SHL [[AND29]], [[C8]](s16)
+    ; VI: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[SHL14]]
+    ; VI: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
+    ; VI: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
+    ; VI: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD31]](s32)
+    ; VI: [[AND31:%[0-9]+]]:_(s16) = G_AND [[TRUNC31]], [[C7]]
+    ; VI: [[SHL15:%[0-9]+]]:_(s16) = G_SHL [[AND31]], [[C8]](s16)
+    ; VI: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[SHL15]]
+    ; VI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR12]](s16), [[OR13]](s16), [[OR14]](s16), [[OR15]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4s64_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4s64_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; GFX9-MESA: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP15]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; GFX9-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
+    ; GFX9-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; GFX9-MESA: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
+    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
+    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; GFX9-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; GFX9-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
+    ; GFX9-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; GFX9-MESA: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
+    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
+    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; GFX9-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; GFX9-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
+    ; GFX9-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; GFX9-MESA: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
+    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
+    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; GFX9-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; GFX9-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
+    ; GFX9-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; GFX9-MESA: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
+    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
+    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; GFX9-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
+    ; GFX9-MESA: [[GEP23:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C11]](s64)
+    ; GFX9-MESA: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[GEP23]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP24:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[GEP24]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP25:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[GEP25]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP26:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[GEP26]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP27:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[GEP27]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP28:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[GEP28]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP29:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[GEP29]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP30:%[0-9]+]]:_(p1) = G_GEP [[GEP23]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[GEP30]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
+    ; GFX9-MESA: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
+    ; GFX9-MESA: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD25]](s32)
+    ; GFX9-MESA: [[AND25:%[0-9]+]]:_(s16) = G_AND [[TRUNC25]], [[C7]]
+    ; GFX9-MESA: [[SHL12:%[0-9]+]]:_(s16) = G_SHL [[AND25]], [[C8]](s16)
+    ; GFX9-MESA: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[SHL12]]
+    ; GFX9-MESA: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
+    ; GFX9-MESA: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
+    ; GFX9-MESA: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD27]](s32)
+    ; GFX9-MESA: [[AND27:%[0-9]+]]:_(s16) = G_AND [[TRUNC27]], [[C7]]
+    ; GFX9-MESA: [[SHL13:%[0-9]+]]:_(s16) = G_SHL [[AND27]], [[C8]](s16)
+    ; GFX9-MESA: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[SHL13]]
+    ; GFX9-MESA: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
+    ; GFX9-MESA: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
+    ; GFX9-MESA: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD29]](s32)
+    ; GFX9-MESA: [[AND29:%[0-9]+]]:_(s16) = G_AND [[TRUNC29]], [[C7]]
+    ; GFX9-MESA: [[SHL14:%[0-9]+]]:_(s16) = G_SHL [[AND29]], [[C8]](s16)
+    ; GFX9-MESA: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[SHL14]]
+    ; GFX9-MESA: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
+    ; GFX9-MESA: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
+    ; GFX9-MESA: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD31]](s32)
+    ; GFX9-MESA: [[AND31:%[0-9]+]]:_(s16) = G_AND [[TRUNC31]], [[C7]]
+    ; GFX9-MESA: [[SHL15:%[0-9]+]]:_(s16) = G_SHL [[AND31]], [[C8]](s16)
+    ; GFX9-MESA: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[SHL15]]
+    ; GFX9-MESA: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR12]](s16), [[OR13]](s16), [[OR14]](s16), [[OR15]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 1, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_v2s128_align32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2s128_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s128>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<2 x s128>)
+    ; CI-HSA-LABEL: name: test_load_global_v2s128_align32
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s128>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<2 x s128>)
+    ; CI-MESA-LABEL: name: test_load_global_v2s128_align32
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s128>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<2 x s128>)
+    ; VI-LABEL: name: test_load_global_v2s128_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s128>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<2 x s128>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2s128_align32
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s128>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<2 x s128>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2s128_align32
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s128>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<2 x s128>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s128>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_v2p1_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p1_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p1_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p1_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; VI-LABEL: name: test_load_global_v2p1_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2p1_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p1_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p1_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p1_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; VI-LABEL: name: test_load_global_v2p1_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2p1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p1_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p1_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; VI-LABEL: name: test_load_global_v2p1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v2p1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x p1>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p1_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p1_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
+    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
+    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
+    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x p1>)
+    ; VI-LABEL: name: test_load_global_v2p1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x p1>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p1>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x p1>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP7]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
+    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
+    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
+    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
+    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
+    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
+    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16), [[OR6]](s16), [[OR7]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x p1>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 1, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_global_v4p1_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v4p1_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x p1>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x p1>)
+    ; CI-HSA-LABEL: name: test_load_global_v4p1_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x p1>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x p1>)
+    ; CI-MESA-LABEL: name: test_load_global_v4p1_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x p1>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x p1>)
+    ; VI-LABEL: name: test_load_global_v4p1_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x p1>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x p1>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v4p1_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x p1>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x p1>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v4p1_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x p1>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x p1>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x p1>) = G_LOAD %0 :: (load 32, align 8, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_global_v2p3_align8
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p3_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p3_align8
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p3_align8
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; VI-LABEL: name: test_load_global_v2p3_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p3_align8
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p3_align8
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v2p3_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p3_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p3_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p3_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; VI-LABEL: name: test_load_global_v2p3_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p3_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p3_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_global_v2p3_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_load_global_v2p3_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV1:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[MV]](p3), [[MV1]](p3)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    ; CI-HSA-LABEL: name: test_load_global_v2p3_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; CI-MESA-LABEL: name: test_load_global_v2p3_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[MV]](p3), [[MV1]](p3)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    ; VI-LABEL: name: test_load_global_v2p3_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV1:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[MV]](p3), [[MV1]](p3)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    ; GFX9-HSA-LABEL: name: test_load_global_v2p3_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
+    ; GFX9-MESA-LABEL: name: test_load_global_v2p3_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[MV]](p3), [[MV1]](p3)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_global_s32_from_1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s32_from_1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_1_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_1_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_ext_load_global_s32_from_1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_1_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_1_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_ext_load_global_s32_from_2_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s32_from_2_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_2_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_2_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_ext_load_global_s32_from_2_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_2_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_2_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_ext_load_global_s64_from_1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+
+    ; SI-LABEL: name: test_ext_load_global_s64_from_1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_1_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_1_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_global_s64_from_1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_1_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_1_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_global_s64_from_2_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s64_from_2_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_2_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_2_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_global_s64_from_2_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_2_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_2_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_global_s64_from_4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CI-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; SI-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_4_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_global_s128_from_4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s128_from_4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s128_from_4_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s128_from_4_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; VI-LABEL: name: test_ext_load_global_s128_from_4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s128_from_4_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s128_from_4_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s128) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_ext_load_global_s64_from_2_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s64_from_2_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_2_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_2_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_global_s64_from_2_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_2_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_2_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_global_s64_from_1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s64_from_1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_1_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_1_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_global_s64_from_1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_1_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_1_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_global_v2s32_from_4_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s32_from_4_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s32_from_4_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, align 1, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s32_from_4_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_extload_global_v2s32_from_4_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s32_from_4_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, align 1, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s32_from_4_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP3]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_global_v2s32_from_4_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s32_from_4_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s32_from_4_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, align 2, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s32_from_4_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_extload_global_v2s32_from_4_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s32_from_4_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, align 2, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s32_from_4_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[GEP1]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_global_v2s32_from_4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s32_from_4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s32_from_4_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI-HSA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s32_from_4_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_extload_global_v2s32_from_4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s32_from_4_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s32_from_4_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_global_v3s32_from_6_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v3s32_from_6_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; CI-HSA-LABEL: name: test_extload_global_v3s32_from_6_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; CI-MESA-LABEL: name: test_extload_global_v3s32_from_6_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; VI-LABEL: name: test_extload_global_v3s32_from_6_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v3s32_from_6_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v3s32_from_6_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<3 x s32>) = G_LOAD %0 :: (load 6, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_extload_global_v4s32_from_8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v4s32_from_8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-HSA-LABEL: name: test_extload_global_v4s32_from_8_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-MESA-LABEL: name: test_extload_global_v4s32_from_8_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; VI-LABEL: name: test_extload_global_v4s32_from_8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v4s32_from_8_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-HSA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-HSA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-HSA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-HSA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-HSA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; GFX9-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v4s32_from_8_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
+    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_extload_global_v2s96_from_24_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; SI: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; SI: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C13]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C13]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C13]]
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C13]]
+    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C13]]
+    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C13]]
+    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C12]](s32)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; SI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C]](s64)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C1]](s64)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C2]](s64)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C3]](s64)
+    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C4]](s64)
+    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C5]](s64)
+    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C6]](s64)
+    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C7]](s64)
+    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C8]](s64)
+    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C9]](s64)
+    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; SI: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C10]](s64)
+    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C11]]
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C13]]
+    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C11]]
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C13]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; SI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C11]]
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C13]]
+    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
+    ; SI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; SI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; SI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; SI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C11]]
+    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C13]]
+    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
+    ; SI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; SI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; SI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C11]]
+    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C13]]
+    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
+    ; SI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; SI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; SI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; SI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C11]]
+    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C13]]
+    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
+    ; SI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; SI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; SI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16), [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; SI: [[COPY24:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; SI: [[COPY25:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY24]](s96)
+    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY25]](s96)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 1, addrspace 1)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; CI-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; CI-MESA: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C13]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C13]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C13]]
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C13]]
+    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C13]]
+    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C13]]
+    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C12]](s32)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; CI-MESA: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C14]](s64)
+    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C]](s64)
+    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C1]](s64)
+    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C2]](s64)
+    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C3]](s64)
+    ; CI-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C4]](s64)
+    ; CI-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C5]](s64)
+    ; CI-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C6]](s64)
+    ; CI-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C7]](s64)
+    ; CI-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C8]](s64)
+    ; CI-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C9]](s64)
+    ; CI-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C10]](s64)
+    ; CI-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C11]]
+    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C13]]
+    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
+    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C11]]
+    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C13]]
+    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
+    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; CI-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C11]]
+    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C13]]
+    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
+    ; CI-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; CI-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; CI-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; CI-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C11]]
+    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C13]]
+    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
+    ; CI-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; CI-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; CI-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C11]]
+    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C13]]
+    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
+    ; CI-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; CI-MESA: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; CI-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; CI-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C11]]
+    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C12]](s32)
+    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; CI-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C13]]
+    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
+    ; CI-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; CI-MESA: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16), [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; CI-MESA: [[COPY24:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; CI-MESA: [[COPY25:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY24]](s96)
+    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY25]](s96)
+    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; VI: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C11]]
+    ; VI: [[C12:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C12]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C11]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C12]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C11]]
+    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C12]](s16)
+    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C11]]
+    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C12]](s16)
+    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C11]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C12]](s16)
+    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C11]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C12]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C]](s64)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C1]](s64)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C2]](s64)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C3]](s64)
+    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C4]](s64)
+    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C5]](s64)
+    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C6]](s64)
+    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C7]](s64)
+    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C8]](s64)
+    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C9]](s64)
+    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; VI: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C10]](s64)
+    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C11]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C11]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C12]](s16)
+    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C11]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C11]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C12]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; VI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C11]]
+    ; VI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; VI: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C11]]
+    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C12]](s16)
+    ; VI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; VI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; VI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C11]]
+    ; VI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; VI: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C11]]
+    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C12]](s16)
+    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; VI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; VI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C11]]
+    ; VI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; VI: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C11]]
+    ; VI: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C12]](s16)
+    ; VI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; VI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; VI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C11]]
+    ; VI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; VI: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C11]]
+    ; VI: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C12]](s16)
+    ; VI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; VI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16), [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; VI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; VI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 1, addrspace 1)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; GFX9-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C7]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 9
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C8]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 11
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C11]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C11]]
+    ; GFX9-MESA: [[C12:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C12]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C11]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C11]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C12]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C11]]
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C11]]
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C12]](s16)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C11]]
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C11]]
+    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C12]](s16)
+    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C11]]
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C11]]
+    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C12]](s16)
+    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C11]]
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C11]]
+    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C12]](s16)
+    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16), [[OR2]](s16), [[OR3]](s16), [[OR4]](s16), [[OR5]](s16)
+    ; GFX9-MESA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP11:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C13]](s64)
+    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP12:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP13:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP14:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP15:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP16:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP17:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP18:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C6]](s64)
+    ; GFX9-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP19:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C7]](s64)
+    ; GFX9-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP20:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C8]](s64)
+    ; GFX9-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP21:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C9]](s64)
+    ; GFX9-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[GEP22:%[0-9]+]]:_(p1) = G_GEP [[GEP11]], [[C10]](s64)
+    ; GFX9-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p1) :: (load 1, addrspace 1)
+    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C11]]
+    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C11]]
+    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C12]](s16)
+    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C11]]
+    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C11]]
+    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C12]](s16)
+    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; GFX9-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C11]]
+    ; GFX9-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; GFX9-MESA: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C11]]
+    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C12]](s16)
+    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; GFX9-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; GFX9-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C11]]
+    ; GFX9-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; GFX9-MESA: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C11]]
+    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C12]](s16)
+    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; GFX9-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; GFX9-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C11]]
+    ; GFX9-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; GFX9-MESA: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C11]]
+    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C12]](s16)
+    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; GFX9-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; GFX9-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C11]]
+    ; GFX9-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; GFX9-MESA: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C11]]
+    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C12]](s16)
+    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16), [[OR8]](s16), [[OR9]](s16), [[OR10]](s16), [[OR11]](s16)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 1, addrspace 1)
+    %2:_(s96) = G_EXTRACT %1, 0
+    %3:_(s96) = G_EXTRACT %1, 96
+    $vgpr0_vgpr1_vgpr2 = COPY %2
+    $vgpr3_vgpr4_vgpr5 = COPY %3
+...
+
+---
+name: test_extload_global_v2s96_from_24_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; SI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; SI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; SI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C1]](s64)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C2]](s64)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; SI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C3]](s64)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C4]](s64)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 2, addrspace 1)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16), [[TRUNC8]](s16), [[TRUNC9]](s16), [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; SI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; SI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 2, addrspace 1)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; CI-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; CI-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CI-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CI-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CI-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; CI-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C1]](s64)
+    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C2]](s64)
+    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; CI-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C3]](s64)
+    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C4]](s64)
+    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 2, addrspace 1)
+    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; CI-MESA: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16), [[TRUNC8]](s16), [[TRUNC9]](s16), [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; VI: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; VI: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C1]](s64)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C2]](s64)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C3]](s64)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C4]](s64)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 2, addrspace 1)
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16), [[TRUNC8]](s16), [[TRUNC9]](s16), [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; VI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; VI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 2, addrspace 1)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; GFX9-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+    ; GFX9-MESA: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; GFX9-MESA: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; GFX9-MESA: [[GEP3:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; GFX9-MESA: [[GEP4:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16), [[TRUNC2]](s16), [[TRUNC3]](s16), [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+    ; GFX9-MESA: [[GEP5:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C5]](s64)
+    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9-MESA: [[GEP6:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9-MESA: [[GEP7:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9-MESA: [[GEP8:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C2]](s64)
+    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9-MESA: [[GEP9:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C3]](s64)
+    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9-MESA: [[GEP10:%[0-9]+]]:_(p1) = G_GEP [[GEP5]], [[C4]](s64)
+    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 2, addrspace 1)
+    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16), [[TRUNC8]](s16), [[TRUNC9]](s16), [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 2, addrspace 1)
+    %2:_(s96) = G_EXTRACT %1, 0
+    %3:_(s96) = G_EXTRACT %1, 96
+    $vgpr0_vgpr1_vgpr2 = COPY %2
+    $vgpr3_vgpr4_vgpr5 = COPY %3
+...
+
+---
+name: test_extload_global_v2s96_from_24_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 4, addrspace 1)
+    ; SI: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; SI: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 4, addrspace 1)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; CI-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 4, addrspace 1)
+    ; CI-MESA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; CI-MESA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 4, addrspace 1)
+    ; VI: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; VI: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 4, addrspace 1)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; GFX9-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 4, addrspace 1)
+    ; GFX9-MESA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; GFX9-MESA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 4, addrspace 1)
+    %2:_(s96) = G_EXTRACT %1, 0
+    %3:_(s96) = G_EXTRACT %1, 96
+    $vgpr0_vgpr1_vgpr2 = COPY %2
+    $vgpr3_vgpr4_vgpr5 = COPY %3
+...
+
+---
+name: test_extload_global_v2s96_from_24_align16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 16, addrspace 1)
+    ; SI: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; SI: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align16
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 16, addrspace 1)
+    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; CI-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align16
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 16, addrspace 1)
+    ; CI-MESA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; CI-MESA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 16, addrspace 1)
+    ; VI: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; VI: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align16
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 16, addrspace 1)
+    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; GFX9-HSA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align16
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s96>) = G_LOAD [[COPY]](p1) :: (load 24, align 16, addrspace 1)
+    ; GFX9-MESA: [[EXTRACT:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 0
+    ; GFX9-MESA: [[EXTRACT1:%[0-9]+]]:_(s96) = G_EXTRACT [[LOAD]](<2 x s96>), 96
+    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](s96)
+    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[EXTRACT1]](s96)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 16, addrspace 1)
+    %2:_(s96) = G_EXTRACT %1, 0
+    %3:_(s96) = G_EXTRACT %1, 96
+    $vgpr0_vgpr1_vgpr2 = COPY %2
+    $vgpr3_vgpr4_vgpr5 = COPY %3
+...




More information about the llvm-commits mailing list