[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-private.mir
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-private.mir?rev=371533&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-private.mir (added)
+++ llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load-private.mir Tue Sep 10 09:20:14 2019
@@ -0,0 +1,10223 @@
+# 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
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=CI %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-mesa-mesa3d -mcpu=gfx900 -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefix=GFX9 %s
+
+---
+name: test_load_private_s1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s1_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; CI: $vgpr0 = COPY [[AND]](s32)
+    ; VI-LABEL: name: test_load_private_s1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s1_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; GFX9: $vgpr0 = COPY [[AND]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s1) = G_LOAD %0 :: (load 1, align 1, addrspace 5)
+    %2:_(s32) = G_ZEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s2_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s2_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s2_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; CI: $vgpr0 = COPY [[AND]](s32)
+    ; VI-LABEL: name: test_load_private_s2_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s2_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
+    ; GFX9: $vgpr0 = COPY [[AND]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s2) = G_LOAD %0 :: (load 1, align 1, addrspace 5)
+    %2:_(s32) = G_ZEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-LABEL: name: test_load_private_s8_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_private_s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-LABEL: name: test_load_private_s8_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s8) = G_LOAD %0 :: (load 1, align 4, addrspace 5)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-LABEL: name: test_load_private_s8_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_private_s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-LABEL: name: test_load_private_s8_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s8) = G_LOAD %0 :: (load 1, align 1, addrspace 5)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-LABEL: name: test_load_private_s16_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_private_s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-LABEL: name: test_load_private_s16_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s16) = G_LOAD %0 :: (load 2, align 4, addrspace 5)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: $vgpr0 = COPY [[COPY1]](s32)
+    ; CI-LABEL: name: test_load_private_s16_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: $vgpr0 = COPY [[COPY1]](s32)
+    ; VI-LABEL: name: test_load_private_s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-LABEL: name: test_load_private_s16_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s16) = G_LOAD %0 :: (load 2, align 2, addrspace 5)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s16_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; CI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; VI-LABEL: name: test_load_private_s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s16_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s16) = G_LOAD %0 :: (load 2, align 1, addrspace 5)
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...
+
+---
+name: test_load_private_s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-LABEL: name: test_load_private_s32_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_load_private_s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-LABEL: name: test_load_private_s32_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s32) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_s32_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s32_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s32_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: $vgpr0 = COPY [[MV]](s32)
+    ; VI-LABEL: name: test_load_private_s32_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s32_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: $vgpr0 = COPY [[MV]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s32) = G_LOAD %0 :: (load 4, align 2, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_s32_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s32_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s32_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: $vgpr0 = COPY [[MV]](s32)
+    ; VI-LABEL: name: test_load_private_s32_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_s32_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: $vgpr0 = COPY [[MV]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s32) = G_LOAD %0 :: (load 4, align 1, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_s48_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s48_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, align 4, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[DEF:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
+    ; SI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[DEF]](s64)
+    ; SI: [[INSERT:%[0-9]+]]:_(s64) = G_INSERT [[COPY1]], [[LOAD]](s32), 0
+    ; SI: [[COPY2:%[0-9]+]]:_(s64) = COPY [[INSERT]](s64)
+    ; SI: [[INSERT1:%[0-9]+]]:_(s64) = G_INSERT [[COPY2]], [[TRUNC]](s16), 32
+    ; SI: [[COPY3:%[0-9]+]]:_(s64) = COPY [[INSERT1]](s64)
+    ; SI: $vgpr0_vgpr1 = COPY [[COPY3]](s64)
+    ; CI-LABEL: name: test_load_private_s48_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, align 4, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[DEF:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
+    ; CI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[DEF]](s64)
+    ; CI: [[INSERT:%[0-9]+]]:_(s64) = G_INSERT [[COPY1]], [[LOAD]](s32), 0
+    ; CI: [[COPY2:%[0-9]+]]:_(s64) = COPY [[INSERT]](s64)
+    ; CI: [[INSERT1:%[0-9]+]]:_(s64) = G_INSERT [[COPY2]], [[TRUNC]](s16), 32
+    ; CI: [[COPY3:%[0-9]+]]:_(s64) = COPY [[INSERT1]](s64)
+    ; CI: $vgpr0_vgpr1 = COPY [[COPY3]](s64)
+    ; VI-LABEL: name: test_load_private_s48_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, align 4, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[DEF:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
+    ; VI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[DEF]](s64)
+    ; VI: [[INSERT:%[0-9]+]]:_(s64) = G_INSERT [[COPY1]], [[LOAD]](s32), 0
+    ; VI: [[COPY2:%[0-9]+]]:_(s64) = COPY [[INSERT]](s64)
+    ; VI: [[INSERT1:%[0-9]+]]:_(s64) = G_INSERT [[COPY2]], [[TRUNC]](s16), 32
+    ; VI: [[COPY3:%[0-9]+]]:_(s64) = COPY [[INSERT1]](s64)
+    ; VI: $vgpr0_vgpr1 = COPY [[COPY3]](s64)
+    ; GFX9-LABEL: name: test_load_private_s48_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, align 4, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[DEF:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s64) = COPY [[DEF]](s64)
+    ; GFX9: [[INSERT:%[0-9]+]]:_(s64) = G_INSERT [[COPY1]], [[LOAD]](s32), 0
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s64) = COPY [[INSERT]](s64)
+    ; GFX9: [[INSERT1:%[0-9]+]]:_(s64) = G_INSERT [[COPY2]], [[TRUNC]](s16), 32
+    ; GFX9: [[COPY3:%[0-9]+]]:_(s64) = COPY [[INSERT1]](s64)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[COPY3]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s48) = G_LOAD %0 :: (load 6, align 8, addrspace 5)
+    %2:_(s64) = G_ANYEXT %1
+    $vgpr0_vgpr1 = COPY %2
+...
+
+---
+name: test_load_private_s64_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s64_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; CI-LABEL: name: test_load_private_s64_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; VI-LABEL: name: test_load_private_s64_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; GFX9-LABEL: name: test_load_private_s64_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 8, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_s64_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s64_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; CI-LABEL: name: test_load_private_s64_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; VI-LABEL: name: test_load_private_s64_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    ; GFX9-LABEL: name: test_load_private_s64_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_s64_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s64_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    ; CI-LABEL: name: test_load_private_s64_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    ; VI-LABEL: name: test_load_private_s64_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    ; GFX9-LABEL: name: test_load_private_s64_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 2, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_s64_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s64_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    ; CI-LABEL: name: test_load_private_s64_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    ; VI-LABEL: name: test_load_private_s64_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    ; GFX9-LABEL: name: test_load_private_s64_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV2]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 8, align 1, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_s96_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s96_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; CI-LABEL: name: test_load_private_s96_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; VI-LABEL: name: test_load_private_s96_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; GFX9-LABEL: name: test_load_private_s96_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 1, addrspace 56)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_s96_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s96_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; CI-LABEL: name: test_load_private_s96_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; VI-LABEL: name: test_load_private_s96_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; GFX9-LABEL: name: test_load_private_s96_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 8, addrspace 5)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_s96_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s96_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; CI-LABEL: name: test_load_private_s96_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; VI-LABEL: name: test_load_private_s96_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    ; GFX9-LABEL: name: test_load_private_s96_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[MV]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_s96_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s96_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; SI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; CI-LABEL: name: test_load_private_s96_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; CI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; VI-LABEL: name: test_load_private_s96_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; VI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; GFX9-LABEL: name: test_load_private_s96_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 2, addrspace 5)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_s96_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s96_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; CI-LABEL: name: test_load_private_s96_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; VI-LABEL: name: test_load_private_s96_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    ; GFX9-LABEL: name: test_load_private_s96_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[MV3]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s96) = G_LOAD %0 :: (load 12, align 1, addrspace 5)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_s128_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s128_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; CI-LABEL: name: test_load_private_s128_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; CI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; CI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; CI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; CI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; VI-LABEL: name: test_load_private_s128_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; GFX9-LABEL: name: test_load_private_s128_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; GFX9: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; GFX9: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](s16)
+    ; GFX9: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; GFX9: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; GFX9: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; GFX9: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; GFX9: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 1, addrspace 56)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_s128_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s128_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; CI-LABEL: name: test_load_private_s128_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; VI-LABEL: name: test_load_private_s128_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; GFX9-LABEL: name: test_load_private_s128_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 8, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_s128_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s128_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; CI-LABEL: name: test_load_private_s128_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; VI-LABEL: name: test_load_private_s128_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    ; GFX9-LABEL: name: test_load_private_s128_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_s128_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s128_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; SI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; CI-LABEL: name: test_load_private_s128_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; CI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; VI-LABEL: name: test_load_private_s128_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; VI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; GFX9-LABEL: name: test_load_private_s128_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; GFX9: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 2, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_s128_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_s128_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; CI-LABEL: name: test_load_private_s128_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; CI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; CI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; CI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; CI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; VI-LABEL: name: test_load_private_s128_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    ; GFX9-LABEL: name: test_load_private_s128_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; GFX9: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; GFX9: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](s16)
+    ; GFX9: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; GFX9: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; GFX9: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; GFX9: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; GFX9: [[MV4:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV4]](s128)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s128) = G_LOAD %0 :: (load 16, align 1, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_p1_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p1_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; CI-LABEL: name: test_load_private_p1_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; VI-LABEL: name: test_load_private_p1_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; GFX9-LABEL: name: test_load_private_p1_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 8, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_p1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; CI-LABEL: name: test_load_private_p1_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; VI-LABEL: name: test_load_private_p1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    ; GFX9-LABEL: name: test_load_private_p1_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV]](p1)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_p1_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p1_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    ; CI-LABEL: name: test_load_private_p1_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    ; VI-LABEL: name: test_load_private_p1_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    ; GFX9-LABEL: name: test_load_private_p1_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 2, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_p1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    ; CI-LABEL: name: test_load_private_p1_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    ; VI-LABEL: name: test_load_private_p1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    ; GFX9-LABEL: name: test_load_private_p1_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[MV2]](p1)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p1) = G_LOAD %0 :: (load 8, align 1, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_p3_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p3_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](p3)
+    ; CI-LABEL: name: test_load_private_p3_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](p3)
+    ; VI-LABEL: name: test_load_private_p3_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](p3)
+    ; GFX9-LABEL: name: test_load_private_p3_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](p3)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p3) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_p3_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p3_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; SI: $vgpr0 = COPY [[MV]](p3)
+    ; CI-LABEL: name: test_load_private_p3_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: $vgpr0 = COPY [[MV]](p3)
+    ; VI-LABEL: name: test_load_private_p3_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; VI: $vgpr0 = COPY [[MV]](p3)
+    ; GFX9-LABEL: name: test_load_private_p3_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: $vgpr0 = COPY [[MV]](p3)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p3) = G_LOAD %0 :: (load 4, align 2, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_p3_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p3_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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: $vgpr0 = COPY [[MV]](p3)
+    ; CI-LABEL: name: test_load_private_p3_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: $vgpr0 = COPY [[MV]](p3)
+    ; VI-LABEL: name: test_load_private_p3_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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: $vgpr0 = COPY [[MV]](p3)
+    ; GFX9-LABEL: name: test_load_private_p3_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(p3) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: $vgpr0 = COPY [[MV]](p3)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p3) = G_LOAD %0 :: (load 4, align 1, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_p5_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p5_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](p5)
+    ; CI-LABEL: name: test_load_private_p5_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](p5)
+    ; VI-LABEL: name: test_load_private_p5_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](p5)
+    ; GFX9-LABEL: name: test_load_private_p5_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](p5)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p5) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_p5_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p5_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_p5_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: $vgpr0 = COPY [[MV]](p5)
+    ; VI-LABEL: name: test_load_private_p5_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_p5_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: $vgpr0 = COPY [[MV]](p5)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p5) = G_LOAD %0 :: (load 4, align 2, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_p5_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_p5_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_p5_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: $vgpr0 = COPY [[MV]](p5)
+    ; VI-LABEL: name: test_load_private_p5_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_p5_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(p5) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: $vgpr0 = COPY [[MV]](p5)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(p5) = G_LOAD %0 :: (load 4, align 1, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_v2s8_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s8_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 2, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, align 2, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s8_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 2, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, align 2, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; CI: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; CI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; CI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; VI-LABEL: name: test_load_private_v2s8_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 2, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, align 2, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s8_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 2, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, align 2, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9: [[EXTRACT:%[0-9]+]]:_(<2 x s8>) = G_EXTRACT [[TRUNC]](<4 x s8>), 0
+    ; GFX9: [[BITCAST:%[0-9]+]]:_(s16) = G_BITCAST [[EXTRACT]](<2 x s8>)
+    ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s16)
+    ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 2, addrspace 5)
+    %2:_(s16) = G_BITCAST %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_private_v2s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[COPY2]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY3]](s32), [[COPY4]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-LABEL: name: test_load_private_v2s8_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[COPY2]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY3]](s32), [[COPY4]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_load_private_v2s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[COPY2]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY3]](s32), [[COPY4]](s32)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v2s8_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[COPY2]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 1, addrspace 5)
+    %2:_(<2 x s32>) = G_ANYEXT %1
+    $vgpr0_vgpr1 = COPY %2
+...
+
+---
+name: test_load_private_v3s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (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-LABEL: name: test_load_private_v3s8_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (load 3, align 4, addrspace 1)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; CI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; CI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; VI-LABEL: name: test_load_private_v3s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (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-LABEL: name: test_load_private_v3s8_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (load 3, align 4, addrspace 1)
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; GFX9: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; GFX9: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, addrspace 1, align 4)
+    %2:_(<4 x s8>) = G_IMPLICIT_DEF
+    %3:_(<4 x s8>) = G_INSERT %2, %1, 0
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_private_v3s8_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s8_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (load 3, align 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s8_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (load 3, align 1, addrspace 5)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; CI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; CI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    ; VI-LABEL: name: test_load_private_v3s8_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (load 3, align 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s8_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p5) :: (load 3, align 1, addrspace 5)
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
+    ; GFX9: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF]](<4 x s8>)
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[LOAD]](<3 x s8>), 0
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; GFX9: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, align 1, addrspace 5)
+    %2:_(<4 x s8>) = G_IMPLICIT_DEF
+    %3:_(<4 x s8>) = G_INSERT %2, %1, 0
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_load_private_v4s8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; CI-LABEL: name: test_load_private_v4s8_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; VI-LABEL: name: test_load_private_v4s8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    ; GFX9-LABEL: name: test_load_private_v4s8_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](<4 x s8>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_v8s8_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v8s8_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[LOAD]](<4 x s8>), [[LOAD1]](<4 x s8>)
+    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
+    ; CI-LABEL: name: test_load_private_v8s8_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[LOAD]](<4 x s8>), [[LOAD1]](<4 x s8>)
+    ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
+    ; VI-LABEL: name: test_load_private_v8s8_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[LOAD]](<4 x s8>), [[LOAD1]](<4 x s8>)
+    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
+    ; GFX9-LABEL: name: test_load_private_v8s8_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[LOAD]](<4 x s8>), [[LOAD1]](<4 x s8>)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v16s8_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v16s8_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
+    ; CI-LABEL: name: test_load_private_v16s8_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
+    ; VI-LABEL: name: test_load_private_v16s8_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
+    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
+    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
+    ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
+    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
+    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
+    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
+    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
+    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
+    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
+    ; GFX9-LABEL: name: test_load_private_v16s8_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
+    ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
+    ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
+    ; GFX9: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
+    ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
+    ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; GFX9: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
+    ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
+    ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; GFX9: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 1, addrspace 56)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v2s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; CI-LABEL: name: test_load_private_v2s16_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; VI-LABEL: name: test_load_private_v2s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    ; GFX9-LABEL: name: test_load_private_v2s16_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](<2 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_v2s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s16_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; VI-LABEL: name: test_load_private_v2s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s16_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 2, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_v2s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s16_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; CI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    ; VI-LABEL: name: test_load_private_v2s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s16_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; GFX9: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 1, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_load_private_v3s16_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s16_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, align 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s16_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, align 4, addrspace 5)
+    ; 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: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; CI: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; CI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v3s16_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, align 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s16_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, align 4, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; GFX9: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 8, addrspace 5)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_private_v3s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s16_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; 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: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; CI: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; CI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v3s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s16_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC]](<3 x s16>), 0
+    ; GFX9: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 2, addrspace 5)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_private_v3s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; 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: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; SI: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
+    ; SI: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32)
+    ; SI: [[TRUNC6:%[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]], [[TRUNC6]](<3 x s16>), 0
+    ; SI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; CI-LABEL: name: test_load_private_v3s16_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; CI: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
+    ; CI: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32)
+    ; CI: [[TRUNC6:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC6]](<3 x s16>), 0
+    ; CI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v3s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; 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: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; VI: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
+    ; VI: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32)
+    ; VI: [[TRUNC6:%[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]], [[TRUNC6]](<3 x s16>), 0
+    ; VI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    ; GFX9-LABEL: name: test_load_private_v3s16_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
+    ; GFX9: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[ANYEXT]](s32), [[ANYEXT1]](s32), [[ANYEXT2]](s32)
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(<3 x s16>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[TRUNC6]](<3 x s16>), 0
+    ; GFX9: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 1, addrspace 5)
+    %2:_(<4 x s16>) = G_IMPLICIT_DEF
+    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
+    $vgpr0_vgpr1 = COPY %3
+...
+
+---
+name: test_load_private_v4s16_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+    ; SI-LABEL: name: test_load_private_v4s16_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; CI-LABEL: name: test_load_private_v4s16_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v4s16_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; GFX9-LABEL: name: test_load_private_v4s16_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 8, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v4s16_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s16_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; CI-LABEL: name: test_load_private_v4s16_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v4s16_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; GFX9-LABEL: name: test_load_private_v4s16_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[LOAD]](<2 x s16>), [[LOAD1]](<2 x s16>)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v4s16_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+    ; SI-LABEL: name: test_load_private_v4s16_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC2]](s16), [[TRUNC3]](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-LABEL: name: test_load_private_v4s16_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v4s16_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC2]](s16), [[TRUNC3]](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-LABEL: name: test_load_private_v4s16_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 2, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v4s16_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s16_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(p5) = G_GEP [[GEP3]], [[C4]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v4s16_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C4]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    ; VI-LABEL: name: test_load_private_v4s16_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(p5) = G_GEP [[GEP3]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v4s16_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s16>), [[BUILD_VECTOR1]](<2 x s16>)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 1, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v2s32_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s32_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s32_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_load_private_v2s32_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s32_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 8, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v2s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s32_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_load_private_v2s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s32_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v2s32_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s32_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s32_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_load_private_v2s32_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v2s32_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 1, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_load_private_v3s32_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s32_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; CI-LABEL: name: test_load_private_v3s32_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; VI-LABEL: name: test_load_private_v3s32_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v3s32_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, align 1, addrspace 56)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_v3s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s32_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; VI-LABEL: name: test_load_private_v3s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v3s32_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_load_private_v4s32_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s32_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-LABEL: name: test_load_private_v4s32_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; CI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; CI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; CI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; CI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; VI-LABEL: name: test_load_private_v4s32_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v4s32_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 56)
+    ; GFX9: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; GFX9: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; GFX9: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](s16)
+    ; GFX9: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; GFX9: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; GFX9: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; GFX9: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 1, addrspace 56)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v4s32_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s32_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v4s32_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; VI-LABEL: name: test_load_private_v4s32_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v4s32_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 8, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v4s32_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s32_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v4s32_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; VI-LABEL: name: test_load_private_v4s32_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_load_private_v4s32_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v4s32_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s32_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-LABEL: name: test_load_private_v4s32_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; VI-LABEL: name: test_load_private_v4s32_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v4s32_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 2, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v4s32_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s32_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; CI-LABEL: name: test_load_private_v4s32_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; CI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; CI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; CI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; VI-LABEL: name: test_load_private_v4s32_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v4s32_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; GFX9: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; GFX9: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](s16)
+    ; GFX9: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; GFX9: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; GFX9: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; GFX9: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 1, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v8s32_align32
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v8s32_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD6]](s32), [[LOAD7]](s32)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>), [[BUILD_VECTOR2]](<2 x s32>), [[BUILD_VECTOR3]](<2 x s32>)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<8 x s32>)
+    ; CI-LABEL: name: test_load_private_v8s32_align32
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD6]](s32), [[LOAD7]](s32)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>), [[BUILD_VECTOR2]](<2 x s32>), [[BUILD_VECTOR3]](<2 x s32>)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<8 x s32>)
+    ; VI-LABEL: name: test_load_private_v8s32_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD6]](s32), [[LOAD7]](s32)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>), [[BUILD_VECTOR2]](<2 x s32>), [[BUILD_VECTOR3]](<2 x s32>)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<8 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v8s32_align32
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR2:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR3:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD6]](s32), [[LOAD7]](s32)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>), [[BUILD_VECTOR2]](<2 x s32>), [[BUILD_VECTOR3]](<2 x s32>)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<8 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<8 x s32>) = G_LOAD %0 :: (load 16, align 32, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_private_v16s32_align32
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v16s32_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 32
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 4, align 32, addrspace 5)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 48
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 4, align 16, addrspace 5)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<4 x s32>), [[BUILD_VECTOR1]](<4 x s32>), [[BUILD_VECTOR2]](<4 x s32>), [[BUILD_VECTOR3]](<4 x s32>)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[CONCAT_VECTORS]](<16 x s32>)
+    ; CI-LABEL: name: test_load_private_v16s32_align32
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 32
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 4, align 32, addrspace 5)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 48
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 4, align 16, addrspace 5)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<4 x s32>), [[BUILD_VECTOR1]](<4 x s32>), [[BUILD_VECTOR2]](<4 x s32>), [[BUILD_VECTOR3]](<4 x s32>)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[CONCAT_VECTORS]](<16 x s32>)
+    ; VI-LABEL: name: test_load_private_v16s32_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 32
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 4, align 32, addrspace 5)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32)
+    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 48
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 4, align 16, addrspace 5)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<4 x s32>), [[BUILD_VECTOR1]](<4 x s32>), [[BUILD_VECTOR2]](<4 x s32>), [[BUILD_VECTOR3]](<4 x s32>)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[CONCAT_VECTORS]](<16 x s32>)
+    ; GFX9-LABEL: name: test_load_private_v16s32_align32
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32)
+    ; GFX9: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 32
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 4, align 32, addrspace 5)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 48
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 4, align 16, addrspace 5)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<4 x s32>), [[BUILD_VECTOR1]](<4 x s32>), [[BUILD_VECTOR2]](<4 x s32>), [[BUILD_VECTOR3]](<4 x s32>)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[CONCAT_VECTORS]](<16 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<16 x s32>) = G_LOAD %0 :: (load 16, align 32, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %1
+...
+
+---
+name: test_load_private_v2s64_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s64_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; 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-LABEL: name: test_load_private_v2s64_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; VI-LABEL: name: test_load_private_v2s64_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; 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-LABEL: name: test_load_private_v2s64_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v2s64_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2s64_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C6]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[MV5:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV3]](s32), [[MV4]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV2]](s64), [[MV5]](s64)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; CI-LABEL: name: test_load_private_v2s64_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C6]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; CI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; CI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; CI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; CI: [[MV5:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV3]](s32), [[MV4]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV2]](s64), [[MV5]](s64)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; VI-LABEL: name: test_load_private_v2s64_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C5]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[MV5:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV3]](s32), [[MV4]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV2]](s64), [[MV5]](s64)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; GFX9-LABEL: name: test_load_private_v2s64_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C5]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; GFX9: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; GFX9: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](s16)
+    ; GFX9: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; GFX9: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; GFX9: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; GFX9: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; GFX9: [[MV5:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[MV3]](s32), [[MV4]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV2]](s64), [[MV5]](s64)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 1, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v3s64_align32
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v3s64_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; 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-LABEL: name: test_load_private_v3s64_align32
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64)
+    ; CI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<3 x s64>), 0
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    ; VI-LABEL: name: test_load_private_v3s64_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; 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-LABEL: name: test_load_private_v3s64_align32
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64)
+    ; GFX9: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
+    ; GFX9: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<3 x s64>), 0
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 32, addrspace 5)
+    %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_private_v4s64_align32
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4s64_align32
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; 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-LABEL: name: test_load_private_v4s64_align32
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
+    ; VI-LABEL: name: test_load_private_v4s64_align32
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; 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-LABEL: name: test_load_private_v4s64_align32
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 32, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 16, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 32, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_private_v2p1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2p1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; 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-LABEL: name: test_load_private_v2p1_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x p1>)
+    ; VI-LABEL: name: test_load_private_v2p1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; 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-LABEL: name: test_load_private_v2p1_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x p1>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_load_private_v4p1_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v4p1_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV3:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1), [[MV2]](p1), [[MV3]](p1)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x p1>)
+    ; CI-LABEL: name: test_load_private_v4p1_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV3:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1), [[MV2]](p1), [[MV3]](p1)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x p1>)
+    ; VI-LABEL: name: test_load_private_v4p1_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV3:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1), [[MV2]](p1), [[MV3]](p1)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x p1>)
+    ; GFX9-LABEL: name: test_load_private_v4p1_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV2:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD4]](s32), [[LOAD5]](s32)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV3:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[LOAD6]](s32), [[LOAD7]](s32)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x p1>) = G_BUILD_VECTOR [[MV]](p1), [[MV1]](p1), [[MV2]](p1), [[MV3]](p1)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x p1>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x p1>) = G_LOAD %0 :: (load 32, align 8, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
+...
+
+---
+name: test_load_private_v2p3_align8
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_load_private_v2p3_align8
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(p3) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[LOAD]](p3), [[LOAD1]](p3)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    ; CI-LABEL: name: test_load_private_v2p3_align8
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(p3) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[LOAD]](p3), [[LOAD1]](p3)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    ; VI-LABEL: name: test_load_private_v2p3_align8
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(p3) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[LOAD]](p3), [[LOAD1]](p3)
+    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    ; GFX9-LABEL: name: test_load_private_v2p3_align8
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(p3) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[LOAD]](p3), [[LOAD1]](p3)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 8, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_private_s32_from_1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s32_from_1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-LABEL: name: test_ext_load_private_s32_from_1_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_ext_load_private_s32_from_1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-LABEL: name: test_ext_load_private_s32_from_1_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s32) = G_LOAD %0 :: (load 1, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_ext_load_private_s32_from_2_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s32_from_2_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-LABEL: name: test_ext_load_private_s32_from_2_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; CI: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_ext_load_private_s32_from_2_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-LABEL: name: test_ext_load_private_s32_from_2_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; GFX9: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s32) = G_LOAD %0 :: (load 2, align 4, addrspace 5)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_ext_load_private_s64_from_1_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+
+    ; SI-LABEL: name: test_ext_load_private_s64_from_1_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-LABEL: name: test_ext_load_private_s64_from_1_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_private_s64_from_1_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-LABEL: name: test_ext_load_private_s64_from_1_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 1, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_private_s64_from_2_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s64_from_2_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-LABEL: name: test_ext_load_private_s64_from_2_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_private_s64_from_2_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-LABEL: name: test_ext_load_private_s64_from_2_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 2, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_private_s64_from_4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s64_from_4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-LABEL: name: test_ext_load_private_s64_from_4_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_private_s64_from_4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-LABEL: name: test_ext_load_private_s64_from_4_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_private_s128_from_4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s128_from_4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; CI-LABEL: name: test_ext_load_private_s128_from_4_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; VI-LABEL: name: test_ext_load_private_s128_from_4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    ; GFX9-LABEL: name: test_ext_load_private_s128_from_4_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s128) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_ext_load_private_s64_from_2_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s64_from_2_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-LABEL: name: test_ext_load_private_s64_from_2_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_private_s64_from_2_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-LABEL: name: test_ext_load_private_s64_from_2_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 2, align 4, addrspace 5)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 2, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_ext_load_private_s64_from_1_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_ext_load_private_s64_from_1_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; CI-LABEL: name: test_ext_load_private_s64_from_1_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; VI-LABEL: name: test_ext_load_private_s64_from_1_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    ; GFX9-LABEL: name: test_ext_load_private_s64_from_1_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p5) :: (load 1, align 4, addrspace 5)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(s64) = G_LOAD %0 :: (load 1, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_private_v2s32_from_4_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s32_from_4_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v2s32_from_4_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_extload_private_v2s32_from_4_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v2s32_from_4_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 1, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_private_v2s32_from_4_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s32_from_4_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v2s32_from_4_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_extload_private_v2s32_from_4_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v2s32_from_4_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 2, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_private_v2s32_from_4_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s32_from_4_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v2s32_from_4_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; VI-LABEL: name: test_extload_private_v2s32_from_4_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v2s32_from_4_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 4, addrspace 5)
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_extload_private_v3s32_from_6_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v3s32_from_6_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v3s32_from_6_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    ; VI-LABEL: name: test_extload_private_v3s32_from_6_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; 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-LABEL: name: test_extload_private_v3s32_from_6_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[BUILD_VECTOR]](<3 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<3 x s32>) = G_LOAD %0 :: (load 6, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2 = COPY %1
+...
+
+---
+name: test_extload_private_v4s32_from_8_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v4s32_from_8_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<4 x s32>)
+    ; CI-LABEL: name: test_extload_private_v4s32_from_8_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<4 x s32>)
+    ; VI-LABEL: name: test_extload_private_v4s32_from_8_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>)
+    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<4 x s32>)
+    ; GFX9-LABEL: name: test_extload_private_v4s32_from_8_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD2]](s32), [[LOAD3]](s32)
+    ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>)
+    ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<4 x s32>)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<4 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 5)
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...
+
+---
+name: test_extload_private_v2s96_from_24_align1
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s96_from_24_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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 [[C4]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; 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: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; 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: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; 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: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; SI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; SI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; 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]], [[C3]]
+    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; SI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; SI: [[GEP15:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C6]](s32)
+    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP16:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C]](s32)
+    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP17:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C1]](s32)
+    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP18:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C2]](s32)
+    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; SI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C3]]
+    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C5]]
+    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY17]](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]], [[C3]]
+    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C5]]
+    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY19]](s32)
+    ; SI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; SI: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16)
+    ; SI: [[GEP19:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C4]](s32)
+    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP20:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C]](s32)
+    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP21:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C1]](s32)
+    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p5) :: (load 1, addrspace 5)
+    ; SI: [[GEP22:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C2]](s32)
+    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p5) :: (load 1, addrspace 5)
+    ; SI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; SI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C3]]
+    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY22]], [[C5]]
+    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY21]](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]], [[C3]]
+    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY24]], [[C5]]
+    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY23]](s32)
+    ; SI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; SI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; SI: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR10]](s16), [[OR11]](s16)
+    ; SI: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; SI: [[COPY25:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; SI: [[COPY26:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY25]](s96)
+    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY26]](s96)
+    ; CI-LABEL: name: test_extload_private_v2s96_from_24_align1
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; CI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
+    ; CI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
+    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C5]]
+    ; CI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
+    ; CI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
+    ; CI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C5]]
+    ; CI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY7]](s32)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
+    ; CI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C4]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
+    ; CI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C5]]
+    ; CI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY9]](s32)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
+    ; CI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
+    ; CI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C5]]
+    ; CI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY11]](s32)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL5]](s32)
+    ; CI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; CI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; CI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; CI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
+    ; CI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C5]]
+    ; CI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY13]](s32)
+    ; CI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
+    ; CI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
+    ; CI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; CI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
+    ; CI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C5]]
+    ; CI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY15]](s32)
+    ; CI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
+    ; CI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
+    ; CI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; CI: [[GEP15:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C6]](s32)
+    ; CI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP16:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C]](s32)
+    ; CI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP17:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C1]](s32)
+    ; CI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP18:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C2]](s32)
+    ; CI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; CI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C3]]
+    ; CI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
+    ; CI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C5]]
+    ; CI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY17]](s32)
+    ; CI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
+    ; CI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
+    ; CI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; CI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C3]]
+    ; CI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
+    ; CI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C5]]
+    ; CI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY19]](s32)
+    ; CI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
+    ; CI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
+    ; CI: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16)
+    ; CI: [[GEP19:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C4]](s32)
+    ; CI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP20:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C]](s32)
+    ; CI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP21:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C1]](s32)
+    ; CI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p5) :: (load 1, addrspace 5)
+    ; CI: [[GEP22:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C2]](s32)
+    ; CI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p5) :: (load 1, addrspace 5)
+    ; CI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; CI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C3]]
+    ; CI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
+    ; CI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY22]], [[C5]]
+    ; CI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY21]](s32)
+    ; CI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL10]](s32)
+    ; CI: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
+    ; CI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; CI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C3]]
+    ; CI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; CI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
+    ; CI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY24]], [[C5]]
+    ; CI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY23]](s32)
+    ; CI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL11]](s32)
+    ; CI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
+    ; CI: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR10]](s16), [[OR11]](s16)
+    ; CI: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; CI: [[COPY25:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; CI: [[COPY26:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY25]](s96)
+    ; CI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY26]](s96)
+    ; VI-LABEL: name: test_extload_private_v2s96_from_24_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; 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: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; VI: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; VI: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; VI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; VI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; VI: [[GEP15:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C5]](s32)
+    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP16:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C]](s32)
+    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP17:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C1]](s32)
+    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP18:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C2]](s32)
+    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; VI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C3]]
+    ; VI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; VI: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C3]]
+    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; VI: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C3]]
+    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C4]](s16)
+    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; VI: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16)
+    ; VI: [[GEP19:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C6]](s32)
+    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP20:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C]](s32)
+    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP21:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C1]](s32)
+    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p5) :: (load 1, addrspace 5)
+    ; VI: [[GEP22:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C2]](s32)
+    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p5) :: (load 1, addrspace 5)
+    ; VI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; VI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C3]]
+    ; VI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; VI: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C3]]
+    ; VI: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C4]](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]], [[C3]]
+    ; VI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; VI: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C3]]
+    ; VI: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C4]](s16)
+    ; VI: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; VI: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR10]](s16), [[OR11]](s16)
+    ; VI: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; VI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; VI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; GFX9-LABEL: name: test_extload_private_v2s96_from_24_align1
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
+    ; GFX9: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
+    ; GFX9: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C3]]
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C3]]
+    ; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C4]](s16)
+    ; GFX9: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR]](s16), [[OR1]](s16)
+    ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C5]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C1]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C2]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C3]]
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C3]]
+    ; GFX9: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C4]](s16)
+    ; GFX9: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C3]]
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C3]]
+    ; GFX9: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C4]](s16)
+    ; GFX9: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR2]](s16), [[OR3]](s16)
+    ; GFX9: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C6]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C1]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C2]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C3]]
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C3]]
+    ; GFX9: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C4]](s16)
+    ; GFX9: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C3]]
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C3]]
+    ; GFX9: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C4]](s16)
+    ; GFX9: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR4]](s16), [[OR5]](s16)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; GFX9: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP11:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C7]](s32)
+    ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP12:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C]](s32)
+    ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP13:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C1]](s32)
+    ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP14:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C2]](s32)
+    ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
+    ; GFX9: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C3]]
+    ; GFX9: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
+    ; GFX9: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C3]]
+    ; GFX9: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C4]](s16)
+    ; GFX9: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
+    ; GFX9: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
+    ; GFX9: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C3]]
+    ; GFX9: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
+    ; GFX9: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C3]]
+    ; GFX9: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C4]](s16)
+    ; GFX9: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL7]]
+    ; GFX9: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR6]](s16), [[OR7]](s16)
+    ; GFX9: [[GEP15:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C5]](s32)
+    ; GFX9: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[GEP15]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP16:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C]](s32)
+    ; GFX9: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[GEP16]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP17:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C1]](s32)
+    ; GFX9: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[GEP17]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP18:%[0-9]+]]:_(p5) = G_GEP [[GEP15]], [[C2]](s32)
+    ; GFX9: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[GEP18]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
+    ; GFX9: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C3]]
+    ; GFX9: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
+    ; GFX9: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C3]]
+    ; GFX9: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C4]](s16)
+    ; GFX9: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL8]]
+    ; GFX9: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
+    ; GFX9: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C3]]
+    ; GFX9: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
+    ; GFX9: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C3]]
+    ; GFX9: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C4]](s16)
+    ; GFX9: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL9]]
+    ; GFX9: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR8]](s16), [[OR9]](s16)
+    ; GFX9: [[GEP19:%[0-9]+]]:_(p5) = G_GEP [[GEP11]], [[C6]](s32)
+    ; GFX9: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[GEP19]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP20:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C]](s32)
+    ; GFX9: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[GEP20]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP21:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C1]](s32)
+    ; GFX9: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[GEP21]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[GEP22:%[0-9]+]]:_(p5) = G_GEP [[GEP19]], [[C2]](s32)
+    ; GFX9: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[GEP22]](p5) :: (load 1, addrspace 5)
+    ; GFX9: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
+    ; GFX9: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C3]]
+    ; GFX9: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
+    ; GFX9: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C3]]
+    ; GFX9: [[SHL10:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C4]](s16)
+    ; GFX9: [[OR10:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL10]]
+    ; GFX9: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
+    ; GFX9: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C3]]
+    ; GFX9: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
+    ; GFX9: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C3]]
+    ; GFX9: [[SHL11:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C4]](s16)
+    ; GFX9: [[OR11:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL11]]
+    ; GFX9: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[OR10]](s16), [[OR11]](s16)
+    ; GFX9: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; GFX9: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 1, addrspace 5)
+    %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_private_v2s96_from_24_align2
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s96_from_24_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; SI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; SI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; SI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; SI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; SI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C1]](s32)
+    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; SI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; SI: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s16), [[TRUNC9]](s16)
+    ; SI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C2]](s32)
+    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; SI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP9]], [[C]](s32)
+    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 2, addrspace 5)
+    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; SI: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; SI: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; SI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; SI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; CI-LABEL: name: test_extload_private_v2s96_from_24_align2
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; CI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; CI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; CI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; CI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C1]](s32)
+    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; CI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; CI: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s16), [[TRUNC9]](s16)
+    ; CI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C2]](s32)
+    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; CI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP9]], [[C]](s32)
+    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 2, addrspace 5)
+    ; CI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; CI: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; CI: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; CI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; CI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; CI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; VI-LABEL: name: test_extload_private_v2s96_from_24_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; 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]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; VI: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; VI: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; VI: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; VI: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C1]](s32)
+    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; VI: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; VI: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s16), [[TRUNC9]](s16)
+    ; VI: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C2]](s32)
+    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; VI: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP9]], [[C]](s32)
+    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 2, addrspace 5)
+    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; VI: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; VI: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; VI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; VI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; GFX9-LABEL: name: test_extload_private_v2s96_from_24_align2
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s16), [[TRUNC1]](s16)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[GEP1]], [[C]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC2]](s16), [[TRUNC3]](s16)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP3]], [[C]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
+    ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s16), [[TRUNC5]](s16)
+    ; GFX9: [[MV3:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV]](s32), [[MV1]](s32), [[MV2]](s32)
+    ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP5:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C3]](s32)
+    ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
+    ; GFX9: [[GEP6:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C]](s32)
+    ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
+    ; GFX9: [[MV4:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC6]](s16), [[TRUNC7]](s16)
+    ; GFX9: [[GEP7:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C1]](s32)
+    ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
+    ; GFX9: [[GEP8:%[0-9]+]]:_(p5) = G_GEP [[GEP7]], [[C]](s32)
+    ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
+    ; GFX9: [[MV5:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s16), [[TRUNC9]](s16)
+    ; GFX9: [[GEP9:%[0-9]+]]:_(p5) = G_GEP [[GEP5]], [[C2]](s32)
+    ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
+    ; GFX9: [[GEP10:%[0-9]+]]:_(p5) = G_GEP [[GEP9]], [[C]](s32)
+    ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p5) :: (load 2, addrspace 5)
+    ; GFX9: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
+    ; GFX9: [[MV6:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC10]](s16), [[TRUNC11]](s16)
+    ; GFX9: [[MV7:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[MV4]](s32), [[MV5]](s32), [[MV6]](s32)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV3]](s96)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV7]](s96)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; GFX9: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 2, addrspace 5)
+    %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_private_v2s96_from_24_align4
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s96_from_24_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; 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-LABEL: name: test_extload_private_v2s96_from_24_align4
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; CI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; CI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; CI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; VI-LABEL: name: test_extload_private_v2s96_from_24_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; 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-LABEL: name: test_extload_private_v2s96_from_24_align4
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; GFX9: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 4, addrspace 5)
+    %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_private_v2s96_from_24_align16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_extload_private_v2s96_from_24_align16
+    ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 16, addrspace 5)
+    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; SI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; SI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; SI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; SI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; SI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; 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-LABEL: name: test_extload_private_v2s96_from_24_align16
+    ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 16, addrspace 5)
+    ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; CI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; CI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; CI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; CI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; CI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; CI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; CI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    ; VI-LABEL: name: test_extload_private_v2s96_from_24_align16
+    ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 16, addrspace 5)
+    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; VI: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; VI: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; VI: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; VI: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; VI: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; 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-LABEL: name: test_extload_private_v2s96_from_24_align16
+    ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
+    ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 16, addrspace 5)
+    ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; GFX9: [[GEP:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C]](s32)
+    ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9: [[GEP1:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C1]](s32)
+    ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p5) :: (load 4, align 8, addrspace 5)
+    ; GFX9: [[MV:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32)
+    ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; GFX9: [[GEP2:%[0-9]+]]:_(p5) = G_GEP [[COPY]], [[C2]](s32)
+    ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP3:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C]](s32)
+    ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[GEP4:%[0-9]+]]:_(p5) = G_GEP [[GEP2]], [[C1]](s32)
+    ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p5) :: (load 4, addrspace 5)
+    ; GFX9: [[MV1:%[0-9]+]]:_(s96) = G_MERGE_VALUES [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32)
+    ; GFX9: [[COPY1:%[0-9]+]]:_(s96) = COPY [[MV]](s96)
+    ; GFX9: [[COPY2:%[0-9]+]]:_(s96) = COPY [[MV1]](s96)
+    ; GFX9: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
+    ; GFX9: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
+    %0:_(p5) = COPY $vgpr0
+    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 16, addrspace 5)
+    %2:_(s96) = G_EXTRACT %1, 0
+    %3:_(s96) = G_EXTRACT %1, 96
+    $vgpr0_vgpr1_vgpr2 = COPY %2
+    $vgpr3_vgpr4_vgpr5 = COPY %3
+...

Removed: llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load.mir
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load.mir?rev=371532&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load.mir (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-load.mir (removed)
@@ -1,443 +0,0 @@
-# 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
-# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=VI %s
-
----
-name: test_load_global_i32
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_i32
-    ; 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)
-    ; VI-LABEL: name: test_load_global_i32
-    ; 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)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s32) = G_LOAD %0 :: (load 4, addrspace 1)
-    $vgpr0 = COPY %1
-...
-
----
-name: test_load_global_i64
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_i64
-    ; 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)
-    ; VI-LABEL: name: test_load_global_i64
-    ; 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)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s32) = G_LOAD %0 :: (load 4, addrspace 1)
-    $vgpr0 = COPY %1
-...
-
----
-name: test_load_global_p1
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_p1
-    ; 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)
-    ; VI-LABEL: name: test_load_global_p1
-    ; 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)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(p1) = G_LOAD %0 :: (load 8, addrspace 1)
-    $vgpr0_vgpr1 = COPY %1
-...
-
----
-name: test_load_global_p4
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_p4
-    ; 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)
-    ; VI-LABEL: name: test_load_global_p4
-    ; 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)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(p4) = G_LOAD %0 :: (load 8, addrspace 1)
-    $vgpr0_vgpr1 = COPY %1
-...
-
----
-name: test_load_global_p3
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_p3
-    ; 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)
-    ; VI-LABEL: name: test_load_global_p3
-    ; 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)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(p3) = G_LOAD %0 :: (load 4, addrspace 1)
-    $vgpr0 = COPY %1
-...
-
----
-name: test_load_global_v2s32
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_v2s32
-    ; 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>)
-    ; VI-LABEL: name: test_load_global_v2s32
-    ; 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>)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, addrspace 1)
-    $vgpr0_vgpr1 = COPY %1
-...
-
----
-
-name: test_load_global_v2s16
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_v2s16
-    ; 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>)
-    ; VI-LABEL: name: test_load_global_v2s16
-    ; 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>)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, addrspace 1)
-    $vgpr0 = COPY %1
-...
-
----
-name: test_load_global_v3i32
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_v3i32
-    ; 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: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
-    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
-    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
-    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
-    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[INSERT1]](<3 x s32>)
-    ; VI-LABEL: name: test_load_global_v3i32
-    ; 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>)
-    %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_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]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
-    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
-    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
-    ; VI-LABEL: name: test_ext_load_global_s64_from_1_align1
-    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
-    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
-    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s64) = G_LOAD %0 :: (load 1, addrspace 1, align 4)
-    $vgpr0_vgpr1 = 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]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
-    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
-    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
-    ; VI-LABEL: name: test_ext_load_global_s64_from_2_align2
-    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
-    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
-    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s64) = G_LOAD %0 :: (load 2, addrspace 1, align 4)
-    $vgpr0_vgpr1 = COPY %1
-...
-
----
-name: test_ext_load_global_s64_from_4_align4
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_ext_load_global_s64_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: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
-    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
-    ; VI-LABEL: name: test_ext_load_global_s64_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: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
-    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s64) = G_LOAD %0 :: (load 4, addrspace 1, align 4)
-    $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]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
-    ; SI: [[ANYEXT:%[0-9]+]]:_(s128) = G_ANYEXT [[LOAD]](s32)
-    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[ANYEXT]](s128)
-    ; VI-LABEL: name: test_ext_load_global_s128_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: [[ANYEXT:%[0-9]+]]:_(s128) = G_ANYEXT [[LOAD]](s32)
-    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[ANYEXT]](s128)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s128) = G_LOAD %0 :: (load 4, addrspace 1, align 4)
-    $vgpr0_vgpr1_vgpr2_vgpr3 = 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]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
-    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
-    ; SI: [[DEF:%[0-9]+]]:_(s96) = G_IMPLICIT_DEF
-    ; SI: [[INSERT:%[0-9]+]]:_(s96) = G_INSERT [[DEF]], [[LOAD]](s64), 0
-    ; SI: [[INSERT1:%[0-9]+]]:_(s96) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
-    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[INSERT1]](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)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s96) = G_LOAD %0 :: (load 12, addrspace 1, align 4)
-    $vgpr0_vgpr1_vgpr2 = COPY %1
-
-...
-
----
-name: test_load_global_s160_align4
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_s160_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: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; SI: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[GEP]](p1) :: (load 8, align 4, addrspace 1)
-    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
-    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
-    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
-    ; SI: [[DEF:%[0-9]+]]:_(s160) = G_IMPLICIT_DEF
-    ; SI: [[INSERT:%[0-9]+]]:_(s160) = G_INSERT [[DEF]], [[LOAD]](s64), 0
-    ; SI: [[INSERT1:%[0-9]+]]:_(s160) = G_INSERT [[INSERT]], [[LOAD1]](s64), 64
-    ; SI: [[INSERT2:%[0-9]+]]:_(s160) = G_INSERT [[INSERT1]], [[LOAD2]](s32), 128
-    ; SI: S_NOP 0, implicit [[INSERT2]](s160)
-    ; VI-LABEL: name: test_load_global_s160_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: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; VI: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[GEP]](p1) :: (load 8, align 4, addrspace 1)
-    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
-    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
-    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, addrspace 1)
-    ; VI: [[DEF:%[0-9]+]]:_(s160) = G_IMPLICIT_DEF
-    ; VI: [[INSERT:%[0-9]+]]:_(s160) = G_INSERT [[DEF]], [[LOAD]](s64), 0
-    ; VI: [[INSERT1:%[0-9]+]]:_(s160) = G_INSERT [[INSERT]], [[LOAD1]](s64), 64
-    ; VI: [[INSERT2:%[0-9]+]]:_(s160) = G_INSERT [[INSERT1]], [[LOAD2]](s32), 128
-    ; VI: S_NOP 0, implicit [[INSERT2]](s160)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s160) = G_LOAD %0 :: (load 20, addrspace 1, align 4)
-    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]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
-    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; SI: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[GEP]](p1) :: (load 8, align 4, addrspace 1)
-    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
-    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
-    ; SI: [[LOAD2:%[0-9]+]]:_(s64) = G_LOAD [[GEP1]](p1) :: (load 8, align 4, addrspace 1)
-    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
-    ; SI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
-    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
-    ; SI: [[DEF:%[0-9]+]]:_(s224) = G_IMPLICIT_DEF
-    ; SI: [[INSERT:%[0-9]+]]:_(s224) = G_INSERT [[DEF]], [[LOAD]](s64), 0
-    ; SI: [[INSERT1:%[0-9]+]]:_(s224) = G_INSERT [[INSERT]], [[LOAD1]](s64), 64
-    ; SI: [[INSERT2:%[0-9]+]]:_(s224) = G_INSERT [[INSERT1]], [[LOAD2]](s64), 128
-    ; SI: [[INSERT3:%[0-9]+]]:_(s224) = G_INSERT [[INSERT2]], [[LOAD3]](s32), 192
-    ; SI: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
-    ; SI: [[INSERT4:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[INSERT3]](s224), 0
-    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT4]](s256)
-    ; VI-LABEL: name: test_load_global_s224_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: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; VI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; VI: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[GEP]](p1) :: (load 8, align 4, addrspace 1)
-    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
-    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
-    ; VI: [[LOAD2:%[0-9]+]]:_(s64) = G_LOAD [[GEP1]](p1) :: (load 8, align 4, addrspace 1)
-    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
-    ; VI: [[GEP2:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C2]](s64)
-    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
-    ; VI: [[DEF:%[0-9]+]]:_(s224) = G_IMPLICIT_DEF
-    ; VI: [[INSERT:%[0-9]+]]:_(s224) = G_INSERT [[DEF]], [[LOAD]](s64), 0
-    ; VI: [[INSERT1:%[0-9]+]]:_(s224) = G_INSERT [[INSERT]], [[LOAD1]](s64), 64
-    ; VI: [[INSERT2:%[0-9]+]]:_(s224) = G_INSERT [[INSERT1]], [[LOAD2]](s64), 128
-    ; VI: [[INSERT3:%[0-9]+]]:_(s224) = G_INSERT [[INSERT2]], [[LOAD3]](s32), 192
-    ; VI: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
-    ; VI: [[INSERT4:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[INSERT3]](s224), 0
-    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT4]](s256)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(s224) = G_LOAD %0 :: (load 28, addrspace 1, align 4)
-
-     %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_v3s32
-body: |
-  bb.0:
-    liveins: $vgpr0_vgpr1
-
-    ; SI-LABEL: name: test_load_global_v3s32
-    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 16, addrspace 1)
-    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, align 8, addrspace 1)
-    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
-    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
-    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
-    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[INSERT1]](<3 x s32>)
-    ; VI-LABEL: name: test_load_global_v3s32
-    ; 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>)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, addrspace 1, align 16)
-    $vgpr0_vgpr1_vgpr2 = COPY %1
-...
-
----
-
-name: test_load_constant_v8i32
-body: |
-  bb.0:
-    liveins: $sgpr0_sgpr1
-
-    ; CHECK-LABEL: name: test_load_global_v8i32
-    ; CHECK: [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
-    ; CHECK: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p4) :: (load 32, addrspace 4)
-    ; CHECK: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7 = COPY [[LOAD]](<8 x s32>)
-    %0:_(p4) = COPY $sgpr0_sgpr1
-    %1:_(<8 x s32>) = G_LOAD %0 :: (load 32, addrspace 4)
-    $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7 = COPY %1
-...
-
----
-name: test_load_constant_v16i32
-body: |
-  bb.0:
-    liveins: $sgpr0_sgpr1
-
-    ; CHECK-LABEL: name: test_load_global_v16i32
-    ; CHECK: [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
-    ; CHECK: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p4) :: (load 64, addrspace 4)
-    ; CHECK: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = COPY [[LOAD]](<16 x s32>)
-    %0:_(p4) = COPY $sgpr0_sgpr1
-    %1:_(<16 x s32>) = G_LOAD %0 :: (load 64, addrspace 4)
-    $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = COPY %1
-...
-
-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]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
-    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s8>) = G_EXTRACT [[LOAD]](<4 x s8>), 0
-    ; SI: S_NOP 0, implicit [[EXTRACT]](<3 x s8>)
-    ; VI-LABEL: name: test_load_global_v3s8_align4
-    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
-    ; VI: [[EXTRACT:%[0-9]+]]:_(<3 x s8>) = G_EXTRACT [[LOAD]](<4 x s8>), 0
-    ; VI: S_NOP 0, implicit [[EXTRACT]](<3 x s8>)
-    %0:_(p1) = COPY $vgpr0_vgpr1
-    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, align 4, addrspace 1)
-    S_NOP 0, implicit %1
-...

Modified: llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-store.mir
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-store.mir?rev=371533&r1=371532&r2=371533&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-store.mir (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/GlobalISel/legalize-store.mir Tue Sep 10 09:20:14 2019
@@ -143,12 +143,7 @@ body: |
     ; SI-LABEL: name: test_store_global_v3s32
     ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
     ; SI: [[COPY1:%[0-9]+]]:_(<3 x s32>) = COPY $vgpr2_vgpr3_vgpr4
-    ; SI: [[EXTRACT:%[0-9]+]]:_(<2 x s32>) = G_EXTRACT [[COPY1]](<3 x s32>), 0
-    ; SI: [[EXTRACT1:%[0-9]+]]:_(s32) = G_EXTRACT [[COPY1]](<3 x s32>), 64
-    ; SI: G_STORE [[EXTRACT]](<2 x s32>), [[COPY]](p1) :: (store 8, align 4, addrspace 1)
-    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C]](s64)
-    ; SI: G_STORE [[EXTRACT1]](s32), [[GEP]](p1) :: (store 4, addrspace 1)
+    ; SI: G_STORE [[COPY1]](<3 x s32>), [[COPY]](p1) :: (store 12, align 4, addrspace 1)
     ; VI-LABEL: name: test_store_global_v3s32
     ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
     ; VI: [[COPY1:%[0-9]+]]:_(<3 x s32>) = COPY $vgpr2_vgpr3_vgpr4
@@ -327,12 +322,7 @@ body: |
     ; SI-LABEL: name: test_store_global_96
     ; SI: [[COPY:%[0-9]+]]:_(s96) = COPY $vgpr0_vgpr1_vgpr2
     ; SI: [[COPY1:%[0-9]+]]:_(p1) = COPY $vgpr3_vgpr4
-    ; SI: [[EXTRACT:%[0-9]+]]:_(s64) = G_EXTRACT [[COPY]](s96), 0
-    ; SI: [[EXTRACT1:%[0-9]+]]:_(s32) = G_EXTRACT [[COPY]](s96), 64
-    ; SI: G_STORE [[EXTRACT]](s64), [[COPY1]](p1) :: (store 8, align 16, addrspace 1)
-    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
-    ; SI: [[GEP:%[0-9]+]]:_(p1) = G_GEP [[COPY1]], [[C]](s64)
-    ; SI: G_STORE [[EXTRACT1]](s32), [[GEP]](p1) :: (store 4, align 8, addrspace 1)
+    ; SI: G_STORE [[COPY]](s96), [[COPY1]](p1) :: (store 12, align 16, addrspace 1)
     ; VI-LABEL: name: test_store_global_96
     ; VI: [[COPY:%[0-9]+]]:_(s96) = COPY $vgpr0_vgpr1_vgpr2
     ; VI: [[COPY1:%[0-9]+]]:_(p1) = COPY $vgpr3_vgpr4
@@ -392,14 +382,46 @@ body: |
     ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
     ; SI: [[DEF:%[0-9]+]]:_(<3 x s8>) = G_IMPLICIT_DEF
     ; SI: [[DEF1:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
-    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF1]], [[DEF]](<3 x s8>), 0
-    ; SI: G_STORE [[INSERT]](<4 x s8>), [[COPY]](p1) :: (store 3, align 4, addrspace 1)
+    ; SI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF1]](<4 x s8>)
+    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[DEF]](<3 x s8>), 0
+    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; SI: [[UV:%[0-9]+]]:_(s8), [[UV1:%[0-9]+]]:_(s8), [[UV2:%[0-9]+]]:_(s8), [[UV3:%[0-9]+]]:_(s8) = G_UNMERGE_VALUES [[TRUNC]](<4 x s8>)
+    ; SI: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s8)
+    ; SI: G_STORE [[ANYEXT1]](s32), [[COPY]](p1) :: (store 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: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s8)
+    ; SI: G_STORE [[ANYEXT2]](s32), [[GEP]](p1) :: (store 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; SI: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s8)
+    ; SI: G_STORE [[ANYEXT3]](s32), [[GEP1]](p1) :: (store 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: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s8)
+    ; SI: G_STORE [[ANYEXT4]](s32), [[GEP2]](p1) :: (store 1, addrspace 1)
     ; VI-LABEL: name: test_store_global_v3s8_align4
     ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
     ; VI: [[DEF:%[0-9]+]]:_(<3 x s8>) = G_IMPLICIT_DEF
     ; VI: [[DEF1:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
-    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF1]], [[DEF]](<3 x s8>), 0
-    ; VI: G_STORE [[INSERT]](<4 x s8>), [[COPY]](p1) :: (store 3, align 4, addrspace 1)
+    ; VI: [[ANYEXT:%[0-9]+]]:_(<4 x s16>) = G_ANYEXT [[DEF1]](<4 x s8>)
+    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[ANYEXT]], [[DEF]](<3 x s8>), 0
+    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[INSERT]](<4 x s16>)
+    ; VI: [[UV:%[0-9]+]]:_(s8), [[UV1:%[0-9]+]]:_(s8), [[UV2:%[0-9]+]]:_(s8), [[UV3:%[0-9]+]]:_(s8) = G_UNMERGE_VALUES [[TRUNC]](<4 x s8>)
+    ; VI: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[UV]](s8)
+    ; VI: G_STORE [[ANYEXT1]](s32), [[COPY]](p1) :: (store 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: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[UV1]](s8)
+    ; VI: G_STORE [[ANYEXT2]](s32), [[GEP]](p1) :: (store 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[GEP1:%[0-9]+]]:_(p1) = G_GEP [[COPY]], [[C1]](s64)
+    ; VI: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[UV2]](s8)
+    ; VI: G_STORE [[ANYEXT3]](s32), [[GEP1]](p1) :: (store 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: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[UV3]](s8)
+    ; VI: G_STORE [[ANYEXT4]](s32), [[GEP2]](p1) :: (store 1, addrspace 1)
     %0:_(p1) = COPY $vgpr0_vgpr1
     %1:_(<3 x s8>) = G_IMPLICIT_DEF
     G_STORE %1, %0 :: (store 3, addrspace 1, align 4)




More information about the llvm-commits mailing list