[llvm] [AMDGCN][SDAG] Expand (INSERT|EXTRACT)_VECTOR_ELT for dynamic indices even with -O0 (PR #170323)
Juan Manuel Martinez Caamaño via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 2 08:32:23 PST 2025
https://github.com/jmmartinez created https://github.com/llvm/llvm-project/pull/170323
Before this patch, `insertelement/extractelement` with dynamic indices would
fail to select with `-O0` for vector types that do not map to a
`SI_INDIRECT_SRC/DST` pattern.
To solve this issue, we allow this expansion with `-O0` by hoisting the
code that does the expansion outside of the second `switch` statement in
`SITargetLowering::PerformDAGCombine` and into the first `switch`.
>From 5dad4b1b9fd03ee2daa7beea78c68d5829e804d8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juan=20Manuel=20Martinez=20Caama=C3=B1o?=
<jmartinezcaamao at gmail.com>
Date: Tue, 2 Dec 2025 16:18:42 +0100
Subject: [PATCH 1/3] Pre commit test: Add <6 x 32-bit> test cases for
insert/extract element with a dynamic index
---
.../CodeGen/AMDGPU/extract_vector_dynelt.ll | 104 +++++++++++++++
.../CodeGen/AMDGPU/insert_vector_dynelt.ll | 126 ++++++++++++++++++
2 files changed, 230 insertions(+)
diff --git a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
index c69b0cce3d208..42a6e4adc4e89 100644
--- a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
@@ -1167,3 +1167,107 @@ entry:
%ext = extractelement <16 x double> <double 1.1, double 2.1, double 3.1, double 4.1, double 5.1, double 6.1, double 7.1, double 8.1, double 9.1, double 10.1, double 11.1, double 12.1, double 13.1, double 14.1, double 15.1, double 16.1>, i32 %sel
ret double %ext
}
+
+define i32 @extract_dyn_i32_6(<6 x i32> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_i32_6:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: v_mov_b32_e32 v2, s17
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s18
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 2, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s19
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 3, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s20
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 4, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s21
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 5, v0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %x = extractelement <6 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define i32 @extract_dyn_i32_6_inreg(<6 x i32> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_i32_6_inreg:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s22, 1
+; GCN-NEXT: s_cselect_b32 s4, s17, s16
+; GCN-NEXT: s_cmp_eq_u32 s22, 2
+; GCN-NEXT: s_cselect_b32 s4, s18, s4
+; GCN-NEXT: s_cmp_eq_u32 s22, 3
+; GCN-NEXT: s_cselect_b32 s4, s19, s4
+; GCN-NEXT: s_cmp_eq_u32 s22, 4
+; GCN-NEXT: s_cselect_b32 s4, s20, s4
+; GCN-NEXT: s_cmp_eq_u32 s22, 5
+; GCN-NEXT: s_cselect_b32 s4, s21, s4
+; GCN-NEXT: v_mov_b32_e32 v0, s4
+; GCN-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %x = extractelement <6 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define float @extract_dyn_float_6(<6 x float> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_float_6:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: v_mov_b32_e32 v2, s17
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s18
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 2, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s19
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 3, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s20
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 4, v0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s21
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 5, v0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %x = extractelement <6 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define float @extract_dyn_float_6_inreg(<6 x float> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_float_6_inreg:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s22, 1
+; GCN-NEXT: v_mov_b32_e32 v0, s16
+; GCN-NEXT: v_mov_b32_e32 v1, s17
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 2
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s18
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 3
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s19
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 4
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s20
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 5
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s21
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %x = extractelement <6 x float> %arg, i32 %idx
+ ret float %x
+}
diff --git a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
index 11826aa0b360d..44938a2b5fedf 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
@@ -1949,3 +1949,129 @@ entry:
%v = insertelement <8 x double> %vec, double 1.000000e+00, i32 %sel
ret <8 x double> %v
}
+
+define <6 x i32> @insert_dyn_i32_6(<6 x i32> inreg %arg, i32 %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_i32_6:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v2, s16
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GCN-NEXT: v_cndmask_b32_e32 v6, v2, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s17
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
+; GCN-NEXT: v_cndmask_b32_e32 v7, v2, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s18
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 2, v0
+; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v3, s19
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 3, v0
+; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v4, s20
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 4, v0
+; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v5, s21
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 5, v0
+; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v6
+; GCN-NEXT: v_mov_b32_e32 v1, v7
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <6 x i32> %arg, i32 %val, i32 %idx
+ ret <6 x i32> %x
+}
+
+define <6 x i32> @insert_dyn_i32_6_inreg(<6 x i32> inreg %arg, i32 inreg %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_i32_6_inreg:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s22, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 1
+; GCN-NEXT: v_cndmask_b32_e32 v6, v1, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s17
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 2
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s18
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 3
+; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v3, s19
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 4
+; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v4, s20
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 5
+; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v5, s21
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v6
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <6 x i32> %arg, i32 %val, i32 %idx
+ ret <6 x i32> %x
+}
+
+define <6 x float> @insert_dyn_float_6(<6 x float> inreg %arg, i32 %idx, float %val) {
+; GCN-LABEL: insert_dyn_float_6:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v2, s16
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GCN-NEXT: v_cndmask_b32_e32 v6, v2, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s17
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
+; GCN-NEXT: v_cndmask_b32_e32 v7, v2, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s18
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 2, v0
+; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v3, s19
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 3, v0
+; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v4, s20
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 4, v0
+; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v5, s21
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 5, v0
+; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v6
+; GCN-NEXT: v_mov_b32_e32 v1, v7
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <6 x float> %arg, float %val, i32 %idx
+ ret <6 x float> %x
+}
+
+define <6 x float> @insert_dyn_float_6_inreg(<6 x float> inreg %arg, i32 inreg %idx, float %val) {
+; GCN-LABEL: insert_dyn_float_6_inreg:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s22, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 1
+; GCN-NEXT: v_cndmask_b32_e32 v6, v1, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s17
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 2
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s18
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 3
+; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v3, s19
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 4
+; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v4, s20
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s22, 5
+; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v5, s21
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v6
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <6 x float> %arg, float %val, i32 %idx
+ ret <6 x float> %x
+}
>From 4d04e2e012d61f596cd4dc510aca9401bf4ee8a8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juan=20Manuel=20Martinez=20Caama=C3=B1o?=
<jmartinezcaamao at gmail.com>
Date: Tue, 2 Dec 2025 16:29:58 +0100
Subject: [PATCH 2/3] Add llc -O0 line that triggers instruction selection
failure
---
llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll | 2 ++
llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll | 2 ++
2 files changed, 4 insertions(+)
diff --git a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
index 42a6e4adc4e89..10ea80ba24a3d 100644
--- a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
@@ -1,5 +1,7 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -check-prefixes=GCN %s
+; RUN: not --crash llc -O0 -mtriple=amdgcn -mcpu=fiji < %s -o /dev/null 2>&1 | FileCheck --check-prefixes=GCN-O0 %s
+; GCN-O0: LLVM ERROR: Cannot select: t{{[0-9]+}}: i32 = extract_vector_elt
define amdgpu_kernel void @float4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-LABEL: float4_extelt:
diff --git a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
index 44938a2b5fedf..6ad62a675944b 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
@@ -1,5 +1,7 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -enable-var-scope -check-prefixes=GCN %s
+; RUN: not --crash llc -O0 -mtriple=amdgcn -mcpu=fiji < %s -o /dev/null 2>&1 | FileCheck --check-prefixes=GCN-O0 %s
+; GCN-O0: LLVM ERROR: Cannot select: t{{[0-9]+}}: {{.*}} = insert_vector_elt
define amdgpu_kernel void @float4_inselt(ptr addrspace(1) %out, <4 x float> %vec, i32 %sel) {
; GCN-LABEL: float4_inselt:
>From 93d6942afefbe4a6443460c10d666dab0af0b18e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juan=20Manuel=20Martinez=20Caama=C3=B1o?=
<jmartinezcaamao at gmail.com>
Date: Tue, 2 Dec 2025 13:38:55 +0100
Subject: [PATCH 3/3] [AMDGCN][SDAG] Expand (INSERT|EXTRACT)_VECTOR_ELT for
dynamic indices even with -O0
Before this patch, insertelement/extractelement with dynamic indices would
fail to select with -O0 for vector types that do not map to a
SI_INDIRECT_SRC/DST pattern.
To solve this issue, we allow this expansion with -O0 by hoisting the
code that does the expansion outside of the second `switch` statement in
`SITargetLowering::PerformDAGCombine` and into the first `switch`.
---
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 69 +-
llvm/lib/Target/AMDGPU/SIISelLowering.h | 5 +-
.../CodeGen/AMDGPU/extract_vector_dynelt.ll | 2046 +++++++-
.../CodeGen/AMDGPU/indirect-addressing-si.ll | 1881 +++-----
.../AMDGPU/indirect-addressing-term.ll | 142 +-
.../CodeGen/AMDGPU/insert_vector_dynelt.ll | 4162 ++++++++++++++++-
6 files changed, 6898 insertions(+), 1407 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index a4f376aceaa4b..80ea9fc0789d5 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15336,22 +15336,6 @@ SITargetLowering::performExtractVectorEltCombine(SDNode *N,
}
}
- // EXTRACT_VECTOR_ELT (<n x e>, var-idx) => n x select (e, const-idx)
- if (shouldExpandVectorDynExt(N)) {
- SDLoc SL(N);
- SDValue Idx = N->getOperand(1);
- SDValue V;
- for (unsigned I = 0, E = VecVT.getVectorNumElements(); I < E; ++I) {
- SDValue IC = DAG.getVectorIdxConstant(I, SL);
- SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, ResVT, Vec, IC);
- if (I == 0)
- V = Elt;
- else
- V = DAG.getSelectCC(SL, Idx, IC, Elt, V, ISD::SETEQ);
- }
- return V;
- }
-
if (!DCI.isBeforeLegalize())
return SDValue();
@@ -15393,19 +15377,45 @@ SITargetLowering::performExtractVectorEltCombine(SDNode *N,
return SDValue();
}
-SDValue
-SITargetLowering::performInsertVectorEltCombine(SDNode *N,
- DAGCombinerInfo &DCI) const {
+// EXTRACT_VECTOR_ELT (<n x e>, var-idx) => n x select (e, const-idx)
+SDValue SITargetLowering::performExtractVectorDynEltCombine(
+ SDNode *N, DAGCombinerInfo &DCI) const {
+ if (!shouldExpandVectorDynExt(N))
+ return SDValue();
+
SDValue Vec = N->getOperand(0);
- SDValue Idx = N->getOperand(2);
+ SelectionDAG &DAG = DCI.DAG;
+
EVT VecVT = Vec.getValueType();
- EVT EltVT = VecVT.getVectorElementType();
+ EVT ResVT = N->getValueType(0);
+
+ SDLoc SL(N);
+ SDValue Idx = N->getOperand(1);
+ SDValue V;
+ for (unsigned I = 0, E = VecVT.getVectorNumElements(); I < E; ++I) {
+ SDValue IC = DAG.getVectorIdxConstant(I, SL);
+ SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, ResVT, Vec, IC);
+ if (I == 0)
+ V = Elt;
+ else
+ V = DAG.getSelectCC(SL, Idx, IC, Elt, V, ISD::SETEQ);
+ }
+ return V;
+}
- // INSERT_VECTOR_ELT (<n x e>, var-idx)
- // => BUILD_VECTOR n x select (e, const-idx)
+// INSERT_VECTOR_ELT (<n x e>, var-idx)
+// => BUILD_VECTOR n x select (e, const-idx)
+SDValue
+SITargetLowering::performInsertVectorDynEltCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
if (!shouldExpandVectorDynExt(N))
return SDValue();
+ SDValue Vec = N->getOperand(0);
+ SDValue Idx = N->getOperand(2);
+ EVT VecVT = Vec.getValueType();
+ EVT EltVT = VecVT.getVectorElementType();
+
SelectionDAG &DAG = DCI.DAG;
SDLoc SL(N);
SDValue Ins = N->getOperand(1);
@@ -16943,12 +16953,21 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
if (auto Res = promoteUniformOpToI32(SDValue(N, 0), DCI))
return Res;
break;
+ case ISD::EXTRACT_VECTOR_ELT:
+ if (SDValue V = performExtractVectorDynEltCombine(N, DCI))
+ return V;
+ break;
+ case ISD::INSERT_VECTOR_ELT:
+ if (SDValue V = performInsertVectorDynEltCombine(N, DCI))
+ return V;
+ break;
default:
break;
}
- if (getTargetMachine().getOptLevel() == CodeGenOptLevel::None)
+ if (getTargetMachine().getOptLevel() == CodeGenOptLevel::None) {
return SDValue();
+ }
switch (N->getOpcode()) {
case ISD::ADD:
@@ -17063,8 +17082,6 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
}
case ISD::EXTRACT_VECTOR_ELT:
return performExtractVectorEltCombine(N, DCI);
- case ISD::INSERT_VECTOR_ELT:
- return performInsertVectorEltCombine(N, DCI);
case ISD::FP_ROUND:
return performFPRoundCombine(N, DCI);
case ISD::LOAD: {
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 40c03ca024c6c..55e883e4c78c6 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -223,7 +223,10 @@ class SITargetLowering final : public AMDGPUTargetLowering {
SDValue performFMed3Combine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue performCvtPkRTZCombine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue performExtractVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
- SDValue performInsertVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+ SDValue performExtractVectorDynEltCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const;
+ SDValue performInsertVectorDynEltCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const;
SDValue performFPRoundCombine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue performSelectCombine(SDNode *N, DAGCombinerInfo &DCI) const;
diff --git a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
index 10ea80ba24a3d..a4037c817c359 100644
--- a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
@@ -1,7 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -check-prefixes=GCN %s
-; RUN: not --crash llc -O0 -mtriple=amdgcn -mcpu=fiji < %s -o /dev/null 2>&1 | FileCheck --check-prefixes=GCN-O0 %s
-; GCN-O0: LLVM ERROR: Cannot select: t{{[0-9]+}}: i32 = extract_vector_elt
+; RUN: llc -O0 -mtriple=amdgcn -mcpu=fiji < %s | FileCheck --check-prefixes=GCN-O0 %s
define amdgpu_kernel void @float4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-LABEL: float4_extelt:
@@ -22,6 +21,36 @@ define amdgpu_kernel void @float4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float4_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 1.0
+; GCN-O0-NEXT: s_mov_b32 s6, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 2.0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[2:3], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s4, 4.0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v0, v1, s[2:3]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <4 x float> <float 0.0, float 1.0, float 2.0, float 4.0>, i32 %sel
store float %ext, ptr addrspace(1) %out
@@ -45,6 +74,28 @@ define amdgpu_kernel void @int4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: int4_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s1, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_cselect_b32 s1, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s0, 4
+; GCN-O0-NEXT: s_cselect_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 4>, i32 %sel
store i32 %ext, ptr addrspace(1) %out
@@ -74,6 +125,56 @@ define amdgpu_kernel void @double4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double4_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s5, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s2, 0x3f847ae1
+; GCN-O0-NEXT: s_mov_b32 s8, 0x47ae147b
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s9, s2
+; GCN-O0-NEXT: s_mov_b32 s3, s9
+; GCN-O0-NEXT: s_mov_b32 s2, 0x3ff028f5
+; GCN-O0-NEXT: s_mov_b32 s6, 0xc28f5c29
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s4
+; GCN-O0-NEXT: s_cselect_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s4, s8
+; GCN-O0-NEXT: s_mov_b32 s2, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s2, s4
+; GCN-O0-NEXT: s_mov_b32 s2, 0x4000147a
+; GCN-O0-NEXT: s_mov_b32 s6, 0xe147ae14
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s6
+; GCN-O0-NEXT: s_mov_b32 s8, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s8
+; GCN-O0-NEXT: s_cselect_b32 s4, s2, s4
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_cselect_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0x40100a3d
+; GCN-O0-NEXT: s_mov_b32 s6, 0x70a3d70a
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_mov_b32 s8, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s8
+; GCN-O0-NEXT: s_cselect_b32 s2, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: s_cselect_b32 s3, s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <4 x double> <double 0.01, double 1.01, double 2.01, double 4.01>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -111,6 +212,65 @@ define amdgpu_kernel void @double5_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double5_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s5, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s4, 0x3f847ae1
+; GCN-O0-NEXT: s_mov_b32 s2, 0x47ae147b
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s2
+; GCN-O0-NEXT: s_mov_b32 s4, 0x3ff028f5
+; GCN-O0-NEXT: s_mov_b32 s6, 0xc28f5c29
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s6
+; GCN-O0-NEXT: s_mov_b32 s9, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s9
+; GCN-O0-NEXT: s_cselect_b32 s4, s4, s8
+; GCN-O0-NEXT: ; kill: def $sgpr3 killed $sgpr3 killed $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_cselect_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0x4000147a
+; GCN-O0-NEXT: s_mov_b32 s6, 0xe147ae14
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_mov_b32 s8, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s8
+; GCN-O0-NEXT: s_cselect_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s2, s4
+; GCN-O0-NEXT: s_mov_b32 s2, 0x40100a3d
+; GCN-O0-NEXT: s_mov_b32 s6, 0x70a3d70a
+; GCN-O0-NEXT: s_mov_b32 s8, s6
+; GCN-O0-NEXT: s_mov_b32 s9, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s8
+; GCN-O0-NEXT: s_mov_b32 s7, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s7
+; GCN-O0-NEXT: s_cselect_b32 s4, s2, s4
+; GCN-O0-NEXT: s_mov_b32 s2, s9
+; GCN-O0-NEXT: s_cselect_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0x40140a3d
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_mov_b32 s8, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s8
+; GCN-O0-NEXT: s_cselect_b32 s2, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: s_cselect_b32 s3, s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <5 x double> <double 0.01, double 1.01, double 2.01, double 4.01, double 5.01>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -132,6 +292,25 @@ define amdgpu_kernel void @half4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: half4_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s4, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s5, 0x44004200
+; GCN-O0-NEXT: s_mov_b32 s0, 0x40003c00
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s5, 4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s4, s4, s5
+; GCN-O0-NEXT: s_lshr_b64 s[0:1], s[0:1], s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_short v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <4 x half> <half 1.0, half 2.0, half 3.0, half 4.0>, i32 %sel
store half %ext, ptr addrspace(1) %out
@@ -151,6 +330,24 @@ define amdgpu_kernel void @float2_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float2_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[2:3], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s4, 1.0
+; GCN-O0-NEXT: s_mov_b32 s5, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v0, v1, s[2:3]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <2 x float> <float 0.0, float 1.0>, i32 %sel
store float %ext, ptr addrspace(1) %out
@@ -174,6 +371,36 @@ define amdgpu_kernel void @double2_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double2_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s8, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s2, 0x3f847ae1
+; GCN-O0-NEXT: s_mov_b32 s4, 0x47ae147b
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s2
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: s_mov_b32 s2, 0x3ff028f5
+; GCN-O0-NEXT: s_mov_b32 s6, 0xc28f5c29
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_mov_b32 s9, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s8, s9
+; GCN-O0-NEXT: s_cselect_b32 s2, s2, s3
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: s_cselect_b32 s3, s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <2 x double> <double 0.01, double 1.01>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -219,6 +446,60 @@ define amdgpu_kernel void @half8_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: half8_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4000
+; GCN-O0-NEXT: s_mov_b32 s6, 0x3c00
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4200
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4400
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4500
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4600
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 6
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4700
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s3, 7
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[2:3], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s4, 0x4800
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v0, v1, s[2:3]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_short v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <8 x half> <half 1.0, half 2.0, half 3.0, half 4.0, half 5.0, half 6.0, half 7.0, half 8.0>, i32 %sel
store half %ext, ptr addrspace(1) %out
@@ -250,6 +531,39 @@ define amdgpu_kernel void @short8_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_short v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: short8_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 6
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 7
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s0, 8
+; GCN-O0-NEXT: s_cselect_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_short v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <8 x i16> <i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8>, i32 %sel
store i16 %ext, ptr addrspace(1) %out
@@ -276,6 +590,42 @@ define amdgpu_kernel void @float8_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float8_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 0x41000000
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40e00000
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40c00000
+; GCN-O0-NEXT: s_mov_b32 s6, 0x40a00000
+; GCN-O0-NEXT: s_mov_b32 s7, 4.0
+; GCN-O0-NEXT: s_mov_b32 s8, 0x40400000
+; GCN-O0-NEXT: s_mov_b32 s9, 2.0
+; GCN-O0-NEXT: s_mov_b32 s10, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s3
+; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v8
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <8 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0>, i32 %sel
store float %ext, ptr addrspace(1) %out
@@ -327,6 +677,101 @@ define amdgpu_kernel void @double8_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s17
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[15:16]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double8_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40200000
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: s_mov_b32 s20, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x401c0000
+; GCN-O0-NEXT: s_mov_b32 s21, s5
+; GCN-O0-NEXT: s_mov_b32 s22, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40180000
+; GCN-O0-NEXT: s_mov_b32 s23, s5
+; GCN-O0-NEXT: s_mov_b32 s24, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40140000
+; GCN-O0-NEXT: s_mov_b32 s25, s5
+; GCN-O0-NEXT: s_mov_b32 s26, s4
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 4.0
+; GCN-O0-NEXT: s_mov_b32 s27, s5
+; GCN-O0-NEXT: s_mov_b32 s28, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40080000
+; GCN-O0-NEXT: s_mov_b32 s29, s5
+; GCN-O0-NEXT: s_mov_b32 s30, s4
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 2.0
+; GCN-O0-NEXT: s_mov_b32 s31, s5
+; GCN-O0-NEXT: s_mov_b32 s33, s4
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: s_mov_b32 s34, s5
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+; GCN-O0-NEXT: s_mov_b32 s5, s34
+; GCN-O0-NEXT: s_mov_b32 s6, s33
+; GCN-O0-NEXT: s_mov_b32 s7, s31
+; GCN-O0-NEXT: s_mov_b32 s8, s30
+; GCN-O0-NEXT: s_mov_b32 s9, s29
+; GCN-O0-NEXT: s_mov_b32 s10, s28
+; GCN-O0-NEXT: s_mov_b32 s11, s27
+; GCN-O0-NEXT: s_mov_b32 s12, s26
+; GCN-O0-NEXT: s_mov_b32 s13, s25
+; GCN-O0-NEXT: s_mov_b32 s14, s24
+; GCN-O0-NEXT: s_mov_b32 s15, s23
+; GCN-O0-NEXT: s_mov_b32 s16, s22
+; GCN-O0-NEXT: s_mov_b32 s17, s21
+; GCN-O0-NEXT: s_mov_b32 s18, s20
+; GCN-O0-NEXT: s_mov_b32 s19, s3
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s19
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s19
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v1
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <8 x double> <double 1.0, double 2.0, double 3.0, double 4.0, double 5.0, double 6.0, double 7.0, double 8.0>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -376,6 +821,101 @@ define amdgpu_kernel void @double7_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s15
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[15:16]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double7_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x401c0000
+; GCN-O0-NEXT: s_mov_b32 s21, s5
+; GCN-O0-NEXT: s_mov_b32 s22, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40180000
+; GCN-O0-NEXT: s_mov_b32 s23, s5
+; GCN-O0-NEXT: s_mov_b32 s24, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40140000
+; GCN-O0-NEXT: s_mov_b32 s25, s5
+; GCN-O0-NEXT: s_mov_b32 s26, s4
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 4.0
+; GCN-O0-NEXT: s_mov_b32 s27, s5
+; GCN-O0-NEXT: s_mov_b32 s28, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40080000
+; GCN-O0-NEXT: s_mov_b32 s29, s5
+; GCN-O0-NEXT: s_mov_b32 s30, s4
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 2.0
+; GCN-O0-NEXT: s_mov_b32 s31, s5
+; GCN-O0-NEXT: s_mov_b32 s33, s4
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: s_mov_b32 s34, s5
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
+; GCN-O0-NEXT: ; implicit-def: $sgpr20
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr5
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+; GCN-O0-NEXT: s_mov_b32 s5, s34
+; GCN-O0-NEXT: s_mov_b32 s6, s33
+; GCN-O0-NEXT: s_mov_b32 s7, s31
+; GCN-O0-NEXT: s_mov_b32 s8, s30
+; GCN-O0-NEXT: s_mov_b32 s9, s29
+; GCN-O0-NEXT: s_mov_b32 s10, s28
+; GCN-O0-NEXT: s_mov_b32 s11, s27
+; GCN-O0-NEXT: s_mov_b32 s12, s26
+; GCN-O0-NEXT: s_mov_b32 s13, s25
+; GCN-O0-NEXT: s_mov_b32 s14, s24
+; GCN-O0-NEXT: s_mov_b32 s15, s23
+; GCN-O0-NEXT: s_mov_b32 s16, s22
+; GCN-O0-NEXT: s_mov_b32 s17, s21
+; GCN-O0-NEXT: s_mov_b32 s18, s20
+; GCN-O0-NEXT: s_mov_b32 s19, s3
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s19
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s19
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v1
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <7 x double> <double 1.0, double 2.0, double 3.0, double 4.0, double 5.0, double 6.0, double 7.0>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -410,6 +950,66 @@ define amdgpu_kernel void @float16_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float16_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 0x41800000
+; GCN-O0-NEXT: s_mov_b32 s4, 0x41700000
+; GCN-O0-NEXT: s_mov_b32 s5, 0x41600000
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41500000
+; GCN-O0-NEXT: s_mov_b32 s7, 0x41400000
+; GCN-O0-NEXT: s_mov_b32 s8, 0x41300000
+; GCN-O0-NEXT: s_mov_b32 s9, 0x41200000
+; GCN-O0-NEXT: s_mov_b32 s10, 0x41100000
+; GCN-O0-NEXT: s_mov_b32 s11, 0x41000000
+; GCN-O0-NEXT: s_mov_b32 s12, 0x40e00000
+; GCN-O0-NEXT: s_mov_b32 s13, 0x40c00000
+; GCN-O0-NEXT: s_mov_b32 s14, 0x40a00000
+; GCN-O0-NEXT: s_mov_b32 s15, 4.0
+; GCN-O0-NEXT: s_mov_b32 s16, 0x40400000
+; GCN-O0-NEXT: s_mov_b32 s17, 2.0
+; GCN-O0-NEXT: s_mov_b32 s18, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v30, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v29, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v28, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v27, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v26, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v23, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s3
+; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v19
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v16
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <16 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0, float 9.0, float 10.0, float 11.0, float 12.0, float 13.0, float 14.0, float 15.0, float 16.0>, i32 %sel
store float %ext, ptr addrspace(1) %out
@@ -491,6 +1091,181 @@ define amdgpu_kernel void @double15_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[31:32]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double15_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s6, 0
+; GCN-O0-NEXT: s_mov_b32 s7, 0x402e0000
+; GCN-O0-NEXT: s_mov_b32 s5, s7
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 killed $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s8, 0
+; GCN-O0-NEXT: s_mov_b32 s9, 0x402c0000
+; GCN-O0-NEXT: s_mov_b32 s7, s9
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s10, 0
+; GCN-O0-NEXT: s_mov_b32 s11, 0x402a0000
+; GCN-O0-NEXT: s_mov_b32 s9, s11
+; GCN-O0-NEXT: ; kill: def $sgpr10 killed $sgpr10 killed $sgpr10_sgpr11
+; GCN-O0-NEXT: s_mov_b32 s12, 0
+; GCN-O0-NEXT: s_mov_b32 s13, 0x40280000
+; GCN-O0-NEXT: s_mov_b32 s11, s13
+; GCN-O0-NEXT: ; kill: def $sgpr12 killed $sgpr12 killed $sgpr12_sgpr13
+; GCN-O0-NEXT: s_mov_b32 s14, 0
+; GCN-O0-NEXT: s_mov_b32 s15, 0x40260000
+; GCN-O0-NEXT: s_mov_b32 s13, s15
+; GCN-O0-NEXT: ; kill: def $sgpr14 killed $sgpr14 killed $sgpr14_sgpr15
+; GCN-O0-NEXT: s_mov_b32 s16, 0
+; GCN-O0-NEXT: s_mov_b32 s17, 0x40240000
+; GCN-O0-NEXT: s_mov_b32 s15, s17
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 killed $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s18, 0
+; GCN-O0-NEXT: s_mov_b32 s19, 0x40220000
+; GCN-O0-NEXT: s_mov_b32 s17, s19
+; GCN-O0-NEXT: ; kill: def $sgpr18 killed $sgpr18 killed $sgpr18_sgpr19
+; GCN-O0-NEXT: s_mov_b32 s20, 0
+; GCN-O0-NEXT: s_mov_b32 s21, 0x40200000
+; GCN-O0-NEXT: s_mov_b32 s19, s21
+; GCN-O0-NEXT: ; kill: def $sgpr20 killed $sgpr20 killed $sgpr20_sgpr21
+; GCN-O0-NEXT: s_mov_b32 s22, 0
+; GCN-O0-NEXT: s_mov_b32 s23, 0x401c0000
+; GCN-O0-NEXT: s_mov_b32 s21, s23
+; GCN-O0-NEXT: ; kill: def $sgpr22 killed $sgpr22 killed $sgpr22_sgpr23
+; GCN-O0-NEXT: s_mov_b32 s24, 0
+; GCN-O0-NEXT: s_mov_b32 s25, 0x40180000
+; GCN-O0-NEXT: s_mov_b32 s23, s25
+; GCN-O0-NEXT: ; kill: def $sgpr24 killed $sgpr24 killed $sgpr24_sgpr25
+; GCN-O0-NEXT: s_mov_b32 s26, 0
+; GCN-O0-NEXT: s_mov_b32 s27, 0x40140000
+; GCN-O0-NEXT: s_mov_b32 s25, s27
+; GCN-O0-NEXT: ; kill: def $sgpr26 killed $sgpr26 killed $sgpr26_sgpr27
+; GCN-O0-NEXT: s_mov_b64 s[28:29], 4.0
+; GCN-O0-NEXT: s_mov_b32 s27, s29
+; GCN-O0-NEXT: ; kill: def $sgpr28 killed $sgpr28 killed $sgpr28_sgpr29
+; GCN-O0-NEXT: s_mov_b32 s30, 0
+; GCN-O0-NEXT: s_mov_b32 s31, 0x40080000
+; GCN-O0-NEXT: s_mov_b32 s29, s31
+; GCN-O0-NEXT: ; kill: def $sgpr30 killed $sgpr30 killed $sgpr30_sgpr31
+; GCN-O0-NEXT: s_mov_b64 s[34:35], 2.0
+; GCN-O0-NEXT: s_mov_b32 s31, s35
+; GCN-O0-NEXT: s_mov_b32 s33, s34
+; GCN-O0-NEXT: s_mov_b64 s[36:37], 1.0
+; GCN-O0-NEXT: s_mov_b32 s34, s37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37
+; GCN-O0-NEXT: ; implicit-def: $sgpr4
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr35
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
+; GCN-O0-NEXT: s_mov_b32 s37, s34
+; GCN-O0-NEXT: s_mov_b32 s38, s33
+; GCN-O0-NEXT: s_mov_b32 s39, s31
+; GCN-O0-NEXT: s_mov_b32 s40, s30
+; GCN-O0-NEXT: s_mov_b32 s41, s29
+; GCN-O0-NEXT: s_mov_b32 s42, s28
+; GCN-O0-NEXT: s_mov_b32 s43, s27
+; GCN-O0-NEXT: s_mov_b32 s44, s26
+; GCN-O0-NEXT: s_mov_b32 s45, s25
+; GCN-O0-NEXT: s_mov_b32 s46, s24
+; GCN-O0-NEXT: s_mov_b32 s47, s23
+; GCN-O0-NEXT: s_mov_b32 s48, s22
+; GCN-O0-NEXT: s_mov_b32 s49, s21
+; GCN-O0-NEXT: s_mov_b32 s50, s20
+; GCN-O0-NEXT: s_mov_b32 s51, s19
+; GCN-O0-NEXT: s_mov_b32 s52, s18
+; GCN-O0-NEXT: s_mov_b32 s53, s17
+; GCN-O0-NEXT: s_mov_b32 s54, s16
+; GCN-O0-NEXT: s_mov_b32 s55, s15
+; GCN-O0-NEXT: s_mov_b32 s56, s14
+; GCN-O0-NEXT: s_mov_b32 s57, s13
+; GCN-O0-NEXT: s_mov_b32 s58, s12
+; GCN-O0-NEXT: s_mov_b32 s59, s11
+; GCN-O0-NEXT: s_mov_b32 s60, s10
+; GCN-O0-NEXT: s_mov_b32 s61, s9
+; GCN-O0-NEXT: s_mov_b32 s62, s8
+; GCN-O0-NEXT: s_mov_b32 s63, s7
+; GCN-O0-NEXT: s_mov_b32 s64, s6
+; GCN-O0-NEXT: s_mov_b32 s65, s5
+; GCN-O0-NEXT: s_mov_b32 s66, s4
+; GCN-O0-NEXT: s_mov_b32 s67, s3
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s37
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s38
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s39
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s40
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s41
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s42
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s43
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s46
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s47
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s48
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s49
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s50
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s51
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s52
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s53
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s54
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s55
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s56
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s57
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s58
+; GCN-O0-NEXT: v_mov_b32_e32 v23, s59
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s60
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s61
+; GCN-O0-NEXT: v_mov_b32_e32 v26, s62
+; GCN-O0-NEXT: v_mov_b32_e32 v27, s63
+; GCN-O0-NEXT: v_mov_b32_e32 v28, s64
+; GCN-O0-NEXT: v_mov_b32_e32 v29, s65
+; GCN-O0-NEXT: v_mov_b32_e32 v30, s66
+; GCN-O0-NEXT: v_mov_b32_e32 v31, s67
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s37
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s38
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s39
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s40
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s41
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s42
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s43
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s46
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s47
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s48
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s49
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s50
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s51
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s52
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s53
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s54
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s55
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s56
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s57
+; GCN-O0-NEXT: v_mov_b32_e32 v23, s58
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s59
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s60
+; GCN-O0-NEXT: v_mov_b32_e32 v26, s61
+; GCN-O0-NEXT: v_mov_b32_e32 v27, s62
+; GCN-O0-NEXT: v_mov_b32_e32 v28, s63
+; GCN-O0-NEXT: v_mov_b32_e32 v29, s64
+; GCN-O0-NEXT: v_mov_b32_e32 v30, s65
+; GCN-O0-NEXT: v_mov_b32_e32 v31, s66
+; GCN-O0-NEXT: v_mov_b32_e32 v32, s67
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v1
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <15 x double> <double 1.0, double 2.0, double 3.0, double 4.0, double 5.0, double 6.0, double 7.0, double 8.0, double 9.0, double 10.0, double 11.0, double 12.0, double 13.0, double 14.0, double 15.0>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -574,6 +1349,181 @@ define amdgpu_kernel void @double16_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[31:32]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double16_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_mov_b32 s5, 0x40300000
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s6, 0
+; GCN-O0-NEXT: s_mov_b32 s7, 0x402e0000
+; GCN-O0-NEXT: s_mov_b32 s5, s7
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 killed $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s8, 0
+; GCN-O0-NEXT: s_mov_b32 s9, 0x402c0000
+; GCN-O0-NEXT: s_mov_b32 s7, s9
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s10, 0
+; GCN-O0-NEXT: s_mov_b32 s11, 0x402a0000
+; GCN-O0-NEXT: s_mov_b32 s9, s11
+; GCN-O0-NEXT: ; kill: def $sgpr10 killed $sgpr10 killed $sgpr10_sgpr11
+; GCN-O0-NEXT: s_mov_b32 s12, 0
+; GCN-O0-NEXT: s_mov_b32 s13, 0x40280000
+; GCN-O0-NEXT: s_mov_b32 s11, s13
+; GCN-O0-NEXT: ; kill: def $sgpr12 killed $sgpr12 killed $sgpr12_sgpr13
+; GCN-O0-NEXT: s_mov_b32 s14, 0
+; GCN-O0-NEXT: s_mov_b32 s15, 0x40260000
+; GCN-O0-NEXT: s_mov_b32 s13, s15
+; GCN-O0-NEXT: ; kill: def $sgpr14 killed $sgpr14 killed $sgpr14_sgpr15
+; GCN-O0-NEXT: s_mov_b32 s16, 0
+; GCN-O0-NEXT: s_mov_b32 s17, 0x40240000
+; GCN-O0-NEXT: s_mov_b32 s15, s17
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 killed $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s18, 0
+; GCN-O0-NEXT: s_mov_b32 s19, 0x40220000
+; GCN-O0-NEXT: s_mov_b32 s17, s19
+; GCN-O0-NEXT: ; kill: def $sgpr18 killed $sgpr18 killed $sgpr18_sgpr19
+; GCN-O0-NEXT: s_mov_b32 s20, 0
+; GCN-O0-NEXT: s_mov_b32 s21, 0x40200000
+; GCN-O0-NEXT: s_mov_b32 s19, s21
+; GCN-O0-NEXT: ; kill: def $sgpr20 killed $sgpr20 killed $sgpr20_sgpr21
+; GCN-O0-NEXT: s_mov_b32 s22, 0
+; GCN-O0-NEXT: s_mov_b32 s23, 0x401c0000
+; GCN-O0-NEXT: s_mov_b32 s21, s23
+; GCN-O0-NEXT: ; kill: def $sgpr22 killed $sgpr22 killed $sgpr22_sgpr23
+; GCN-O0-NEXT: s_mov_b32 s24, 0
+; GCN-O0-NEXT: s_mov_b32 s25, 0x40180000
+; GCN-O0-NEXT: s_mov_b32 s23, s25
+; GCN-O0-NEXT: ; kill: def $sgpr24 killed $sgpr24 killed $sgpr24_sgpr25
+; GCN-O0-NEXT: s_mov_b32 s26, 0
+; GCN-O0-NEXT: s_mov_b32 s27, 0x40140000
+; GCN-O0-NEXT: s_mov_b32 s25, s27
+; GCN-O0-NEXT: ; kill: def $sgpr26 killed $sgpr26 killed $sgpr26_sgpr27
+; GCN-O0-NEXT: s_mov_b64 s[28:29], 4.0
+; GCN-O0-NEXT: s_mov_b32 s27, s29
+; GCN-O0-NEXT: ; kill: def $sgpr28 killed $sgpr28 killed $sgpr28_sgpr29
+; GCN-O0-NEXT: s_mov_b32 s30, 0
+; GCN-O0-NEXT: s_mov_b32 s31, 0x40080000
+; GCN-O0-NEXT: s_mov_b32 s29, s31
+; GCN-O0-NEXT: ; kill: def $sgpr30 killed $sgpr30 killed $sgpr30_sgpr31
+; GCN-O0-NEXT: s_mov_b64 s[34:35], 2.0
+; GCN-O0-NEXT: s_mov_b32 s31, s35
+; GCN-O0-NEXT: s_mov_b32 s33, s34
+; GCN-O0-NEXT: s_mov_b64 s[36:37], 1.0
+; GCN-O0-NEXT: s_mov_b32 s34, s37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
+; GCN-O0-NEXT: s_mov_b32 s37, s34
+; GCN-O0-NEXT: s_mov_b32 s38, s33
+; GCN-O0-NEXT: s_mov_b32 s39, s31
+; GCN-O0-NEXT: s_mov_b32 s40, s30
+; GCN-O0-NEXT: s_mov_b32 s41, s29
+; GCN-O0-NEXT: s_mov_b32 s42, s28
+; GCN-O0-NEXT: s_mov_b32 s43, s27
+; GCN-O0-NEXT: s_mov_b32 s44, s26
+; GCN-O0-NEXT: s_mov_b32 s45, s25
+; GCN-O0-NEXT: s_mov_b32 s46, s24
+; GCN-O0-NEXT: s_mov_b32 s47, s23
+; GCN-O0-NEXT: s_mov_b32 s48, s22
+; GCN-O0-NEXT: s_mov_b32 s49, s21
+; GCN-O0-NEXT: s_mov_b32 s50, s20
+; GCN-O0-NEXT: s_mov_b32 s51, s19
+; GCN-O0-NEXT: s_mov_b32 s52, s18
+; GCN-O0-NEXT: s_mov_b32 s53, s17
+; GCN-O0-NEXT: s_mov_b32 s54, s16
+; GCN-O0-NEXT: s_mov_b32 s55, s15
+; GCN-O0-NEXT: s_mov_b32 s56, s14
+; GCN-O0-NEXT: s_mov_b32 s57, s13
+; GCN-O0-NEXT: s_mov_b32 s58, s12
+; GCN-O0-NEXT: s_mov_b32 s59, s11
+; GCN-O0-NEXT: s_mov_b32 s60, s10
+; GCN-O0-NEXT: s_mov_b32 s61, s9
+; GCN-O0-NEXT: s_mov_b32 s62, s8
+; GCN-O0-NEXT: s_mov_b32 s63, s7
+; GCN-O0-NEXT: s_mov_b32 s64, s6
+; GCN-O0-NEXT: s_mov_b32 s65, s5
+; GCN-O0-NEXT: s_mov_b32 s66, s4
+; GCN-O0-NEXT: s_mov_b32 s67, s3
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s37
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s38
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s39
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s40
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s41
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s42
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s43
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s46
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s47
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s48
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s49
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s50
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s51
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s52
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s53
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s54
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s55
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s56
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s57
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s58
+; GCN-O0-NEXT: v_mov_b32_e32 v23, s59
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s60
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s61
+; GCN-O0-NEXT: v_mov_b32_e32 v26, s62
+; GCN-O0-NEXT: v_mov_b32_e32 v27, s63
+; GCN-O0-NEXT: v_mov_b32_e32 v28, s64
+; GCN-O0-NEXT: v_mov_b32_e32 v29, s65
+; GCN-O0-NEXT: v_mov_b32_e32 v30, s66
+; GCN-O0-NEXT: v_mov_b32_e32 v31, s67
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s37
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s38
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s39
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s40
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s41
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s42
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s43
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s46
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s47
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s48
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s49
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s50
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s51
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s52
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s53
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s54
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s55
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s56
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s57
+; GCN-O0-NEXT: v_mov_b32_e32 v23, s58
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s59
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s60
+; GCN-O0-NEXT: v_mov_b32_e32 v26, s61
+; GCN-O0-NEXT: v_mov_b32_e32 v27, s62
+; GCN-O0-NEXT: v_mov_b32_e32 v28, s63
+; GCN-O0-NEXT: v_mov_b32_e32 v29, s64
+; GCN-O0-NEXT: v_mov_b32_e32 v30, s65
+; GCN-O0-NEXT: v_mov_b32_e32 v31, s66
+; GCN-O0-NEXT: v_mov_b32_e32 v32, s67
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v1
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <16 x double> <double 1.0, double 2.0, double 3.0, double 4.0, double 5.0, double 6.0, double 7.0, double 8.0, double 9.0, double 10.0, double 11.0, double 12.0, double 13.0, double 14.0, double 15.0, double 16.0>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -624,6 +1574,114 @@ define amdgpu_kernel void @float32_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float32_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 0x42000000
+; GCN-O0-NEXT: s_mov_b32 s4, 0x41f80000
+; GCN-O0-NEXT: s_mov_b32 s5, 0x41f00000
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41e80000
+; GCN-O0-NEXT: s_mov_b32 s7, 0x41e00000
+; GCN-O0-NEXT: s_mov_b32 s8, 0x41d80000
+; GCN-O0-NEXT: s_mov_b32 s9, 0x41d00000
+; GCN-O0-NEXT: s_mov_b32 s10, 0x41c80000
+; GCN-O0-NEXT: s_mov_b32 s11, 0x41c00000
+; GCN-O0-NEXT: s_mov_b32 s12, 0x41b80000
+; GCN-O0-NEXT: s_mov_b32 s13, 0x41b00000
+; GCN-O0-NEXT: s_mov_b32 s14, 0x41a80000
+; GCN-O0-NEXT: s_mov_b32 s15, 0x41a00000
+; GCN-O0-NEXT: s_mov_b32 s16, 0x41980000
+; GCN-O0-NEXT: s_mov_b32 s17, 0x41900000
+; GCN-O0-NEXT: s_mov_b32 s18, 0x41880000
+; GCN-O0-NEXT: s_mov_b32 s19, 0x41800000
+; GCN-O0-NEXT: s_mov_b32 s20, 0x41700000
+; GCN-O0-NEXT: s_mov_b32 s21, 0x41600000
+; GCN-O0-NEXT: s_mov_b32 s22, 0x41500000
+; GCN-O0-NEXT: s_mov_b32 s23, 0x41400000
+; GCN-O0-NEXT: s_mov_b32 s24, 0x41300000
+; GCN-O0-NEXT: s_mov_b32 s25, 0x41200000
+; GCN-O0-NEXT: s_mov_b32 s26, 0x41100000
+; GCN-O0-NEXT: s_mov_b32 s27, 0x41000000
+; GCN-O0-NEXT: s_mov_b32 s28, 0x40e00000
+; GCN-O0-NEXT: s_mov_b32 s29, 0x40c00000
+; GCN-O0-NEXT: s_mov_b32 s30, 0x40a00000
+; GCN-O0-NEXT: s_mov_b32 s31, 4.0
+; GCN-O0-NEXT: s_mov_b32 s33, 0x40400000
+; GCN-O0-NEXT: s_mov_b32 s34, 2.0
+; GCN-O0-NEXT: s_mov_b32 s35, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s35
+; GCN-O0-NEXT: v_mov_b32_e32 v62, s34
+; GCN-O0-NEXT: v_mov_b32_e32 v61, s33
+; GCN-O0-NEXT: v_mov_b32_e32 v60, s31
+; GCN-O0-NEXT: v_mov_b32_e32 v59, s30
+; GCN-O0-NEXT: v_mov_b32_e32 v58, s29
+; GCN-O0-NEXT: v_mov_b32_e32 v57, s28
+; GCN-O0-NEXT: v_mov_b32_e32 v56, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v55, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v54, s25
+; GCN-O0-NEXT: v_mov_b32_e32 v53, s24
+; GCN-O0-NEXT: v_mov_b32_e32 v52, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v51, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v50, s21
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v48, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v47, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v46, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v45, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v44, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v43, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v42, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v40, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v39, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v38, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v37, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v36, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v35, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v34, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v32, s3
+; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v62
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v61
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v60
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v59
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v58
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v57
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v56
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v55
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v54
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v53
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v52
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v51
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v50
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v49
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v48
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v47
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v42
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v40
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v39
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v38
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v37
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v36
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v35
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v34
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v33
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v32
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movrels_b32_e32 v2, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <32 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0, float 9.0, float 10.0, float 11.0, float 12.0, float 13.0, float 14.0, float 15.0, float 16.0, float 17.0, float 18.0, float 19.0, float 20.0, float 21.0, float 22.0, float 23.0, float 24.0, float 25.0, float 26.0, float 27.0, float 28.0, float 29.0, float 30.0, float 31.0, float 32.0>, i32 %sel
store float %ext, ptr addrspace(1) %out
@@ -645,6 +1703,25 @@ define amdgpu_kernel void @byte8_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: byte8_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s1, 3
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s4, s0, s1
+; GCN-O0-NEXT: s_mov_b32 s5, 0x8070605
+; GCN-O0-NEXT: s_mov_b32 s0, 0x4030201
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_lshr_b64 s[0:1], s[0:1], s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <8 x i8> <i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8>, i32 %sel
store i8 %ext, ptr addrspace(1) %out
@@ -692,6 +1769,63 @@ define amdgpu_kernel void @byte16_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: byte16_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 6
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 7
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 8
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 9
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 10
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 11
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 12
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 13
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 14
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 15
+; GCN-O0-NEXT: s_cselect_b32 s1, s4, s1
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s0, 16
+; GCN-O0-NEXT: s_cselect_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <16 x i8> <i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 16>, i32 %sel
store i8 %ext, ptr addrspace(1) %out
@@ -712,6 +1846,23 @@ define amdgpu_kernel void @bit4_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: bit4_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s1, 3
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s1, s0, s1
+; GCN-O0-NEXT: s_mov_b32 s0, 0x1000100
+; GCN-O0-NEXT: s_lshr_b32 s0, s0, s1
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_and_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <4 x i1> <i1 0, i1 1, i1 0, i1 1>, i32 %sel
%zext = zext i1 %ext to i32
@@ -987,6 +2138,402 @@ define amdgpu_kernel void @bit128_extelt(ptr addrspace(1) %out, i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: bit128_extelt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s5, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_lg_u32 s5, s1
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[6:7]
+; GCN-O0-NEXT: s_mov_b32 s0, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s0
+; GCN-O0-NEXT: v_readfirstlane_b32 s0, v0
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s0
+; GCN-O0-NEXT: s_mov_b32 s0, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s0
+; GCN-O0-NEXT: s_mov_b32 s0, 0
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 6
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 7
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 8
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 9
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 10
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 11
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 12
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 13
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 14
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 15
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 16
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 17
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 18
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 19
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 20
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 21
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 22
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 23
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 24
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 25
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 26
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 27
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 28
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 29
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 30
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 31
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 32
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 33
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 34
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 35
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 36
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 37
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 38
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 39
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 40
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 41
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 42
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 43
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 44
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 45
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 46
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 47
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 48
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 49
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 50
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 51
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 52
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 53
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 54
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 55
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 56
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 57
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 58
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 59
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 60
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 61
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 62
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 63
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 64
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x42
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x43
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x44
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x45
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x46
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x47
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x48
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x49
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4a
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4b
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4c
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4d
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4e
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4f
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x50
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x51
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x52
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x53
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x54
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x55
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x56
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x57
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x58
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x59
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x5a
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x5b
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x5c
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x5d
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x5e
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x5f
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x60
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x61
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x62
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x63
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x64
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x65
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x66
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x67
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x68
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x69
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x6a
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x6b
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x6c
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x6d
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x6e
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x6f
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x70
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x71
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x72
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x73
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x74
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x75
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x76
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x77
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x78
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x79
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x7a
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x7b
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x7c
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x7d
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s0, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x7e
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s4, s1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x7f
+; GCN-O0-NEXT: s_cmp_eq_u32 s5, s6
+; GCN-O0-NEXT: s_cselect_b32 s0, s0, s4
+; GCN-O0-NEXT: s_and_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%ext = extractelement <128 x i1> <i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0>, i32 %sel
%zext = zext i1 %ext to i32
@@ -1090,6 +2637,169 @@ define float @float32_extelt_vec(i32 %sel) {
; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 31, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: float32_extelt_vec:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 2.0
+; GCN-O0-NEXT: s_mov_b32 s7, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x40400000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 4.0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x40a00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x40c00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x40e00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41000000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 8
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41100000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 9
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41200000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 10
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41300000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 11
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41400000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 12
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41500000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 13
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41600000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 14
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41700000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 15
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41800000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 16
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41880000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 17
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41900000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 18
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41980000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 19
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41a00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 20
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41a80000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 21
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41b00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 22
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41b80000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 23
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41c00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 24
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41c80000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 25
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41d00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 26
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41d80000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 27
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41e00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 28
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41e80000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 29
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41f00000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 30
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41f80000
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 31
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x42000000
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%ext = extractelement <32 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0, float 9.0, float 10.0, float 11.0, float 12.0, float 13.0, float 14.0, float 15.0, float 16.0, float 17.0, float 18.0, float 19.0, float 20.0, float 21.0, float 22.0, float 23.0, float 24.0, float 25.0, float 26.0, float 27.0, float 28.0, float 29.0, float 30.0, float 31.0, float 32.0>, i32 %sel
ret float %ext
@@ -1165,6 +2875,208 @@ define double @double16_extelt_vec(i32 %sel) {
; GCN-NEXT: v_mov_b32_e32 v1, 0x40301999
; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: double16_extelt_vec:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v4, s30, 0
+; GCN-O0-NEXT: v_writelane_b32 v4, s31, 1
+; GCN-O0-NEXT: v_writelane_b32 v4, s34, 2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[94:95], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x4000cccc
+; GCN-O0-NEXT: s_mov_b32 s92, 0xcccccccd
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, s92
+; GCN-O0-NEXT: s_mov_b32 vcc_hi, s4
+; GCN-O0-NEXT: s_mov_b32 s4, vcc_lo
+; GCN-O0-NEXT: s_mov_b32 s5, 0x3ff19999
+; GCN-O0-NEXT: s_mov_b32 s6, 0x9999999a
+; GCN-O0-NEXT: s_mov_b32 s30, s6
+; GCN-O0-NEXT: s_mov_b32 s31, s5
+; GCN-O0-NEXT: s_mov_b32 s5, s30
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[94:95]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[90:91], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x4008cccc
+; GCN-O0-NEXT: ; kill: def $sgpr92 killed $sgpr92 def $sgpr92_sgpr93
+; GCN-O0-NEXT: s_mov_b32 s93, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s92
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[90:91]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[78:79], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40106666
+; GCN-O0-NEXT: s_mov_b32 s60, 0x66666666
+; GCN-O0-NEXT: s_mov_b32 s88, s60
+; GCN-O0-NEXT: s_mov_b32 s89, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s88
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[78:79]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[74:75], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40146666
+; GCN-O0-NEXT: s_mov_b32 s76, s60
+; GCN-O0-NEXT: s_mov_b32 s77, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s76
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[74:75]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[62:63], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40186666
+; GCN-O0-NEXT: s_mov_b32 s72, s60
+; GCN-O0-NEXT: s_mov_b32 s73, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s72
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[62:63]
+; GCN-O0-NEXT: s_mov_b32 s4, 6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[58:59], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x401c6666
+; GCN-O0-NEXT: ; kill: def $sgpr60 killed $sgpr60 def $sgpr60_sgpr61
+; GCN-O0-NEXT: s_mov_b32 s61, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s60
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[58:59]
+; GCN-O0-NEXT: s_mov_b32 s4, 7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[46:47], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40203333
+; GCN-O0-NEXT: s_mov_b32 s10, 0x33333333
+; GCN-O0-NEXT: s_mov_b32 s56, s10
+; GCN-O0-NEXT: s_mov_b32 s57, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s56
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[46:47]
+; GCN-O0-NEXT: s_mov_b32 s4, 8
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[42:43], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40223333
+; GCN-O0-NEXT: s_mov_b32 s44, s10
+; GCN-O0-NEXT: s_mov_b32 s45, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[42:43]
+; GCN-O0-NEXT: s_mov_b32 s4, 9
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[28:29], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40243333
+; GCN-O0-NEXT: s_mov_b32 s40, s10
+; GCN-O0-NEXT: s_mov_b32 s41, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s40
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[28:29]
+; GCN-O0-NEXT: s_mov_b32 s4, 10
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[24:25], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40263333
+; GCN-O0-NEXT: s_mov_b32 s26, s10
+; GCN-O0-NEXT: s_mov_b32 s27, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[24:25]
+; GCN-O0-NEXT: s_mov_b32 s4, 11
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[20:21], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40283333
+; GCN-O0-NEXT: s_mov_b32 s22, s10
+; GCN-O0-NEXT: s_mov_b32 s23, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[20:21]
+; GCN-O0-NEXT: s_mov_b32 s4, 12
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[16:17], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x402a3333
+; GCN-O0-NEXT: s_mov_b32 s18, s10
+; GCN-O0-NEXT: s_mov_b32 s19, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[16:17]
+; GCN-O0-NEXT: s_mov_b32 s4, 13
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[12:13], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x402c3333
+; GCN-O0-NEXT: s_mov_b32 s14, s10
+; GCN-O0-NEXT: s_mov_b32 s15, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s4, 14
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[8:9], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s4, 0x402e3333
+; GCN-O0-NEXT: ; kill: def $sgpr10 killed $sgpr10 def $sgpr10_sgpr11
+; GCN-O0-NEXT: s_mov_b32 s11, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[8:9]
+; GCN-O0-NEXT: s_mov_b32 s4, 15
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: s_mov_b32 s34, 0x40301999
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s34
+; GCN-O0-NEXT: s_mov_b32 s34, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s34
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, vcc_hi
+; GCN-O0-NEXT: s_mov_b32 vcc_hi, s31
+; GCN-O0-NEXT: v_mov_b32_e32 v1, vcc_hi
+; GCN-O0-NEXT: v_mov_b32_e32 v2, vcc_lo
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[94:95]
+; GCN-O0-NEXT: s_mov_b32 s92, s93
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s92
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[90:91]
+; GCN-O0-NEXT: s_mov_b32 s88, s89
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s88
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[78:79]
+; GCN-O0-NEXT: s_mov_b32 s76, s77
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s76
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[74:75]
+; GCN-O0-NEXT: s_mov_b32 s72, s73
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s72
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[62:63]
+; GCN-O0-NEXT: s_mov_b32 s60, s61
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s60
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[58:59]
+; GCN-O0-NEXT: s_mov_b32 s56, s57
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s56
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[46:47]
+; GCN-O0-NEXT: s_mov_b32 s44, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s44
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[42:43]
+; GCN-O0-NEXT: s_mov_b32 s40, s41
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s40
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[28:29]
+; GCN-O0-NEXT: s_mov_b32 s26, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s26
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[24:25]
+; GCN-O0-NEXT: s_mov_b32 s22, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s22
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[20:21]
+; GCN-O0-NEXT: s_mov_b32 s18, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[16:17]
+; GCN-O0-NEXT: s_mov_b32 s14, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s14
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s10, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s10
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[8:9]
+; GCN-O0-NEXT: s_mov_b32 s6, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v1, v2, s[4:5]
+; GCN-O0-NEXT: ; implicit-def: $sgpr4
+; GCN-O0-NEXT: ; implicit-def: $sgpr5
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v3
+; GCN-O0-NEXT: s_mov_b32 s4, 32
+; GCN-O0-NEXT: v_lshrrev_b64 v[1:2], s4, v[1:2]
+; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec
+; GCN-O0-NEXT: v_readlane_b32 s34, v4, 2
+; GCN-O0-NEXT: v_readlane_b32 s31, v4, 1
+; GCN-O0-NEXT: v_readlane_b32 s30, v4, 0
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%ext = extractelement <16 x double> <double 1.1, double 2.1, double 3.1, double 4.1, double 5.1, double 6.1, double 7.1, double 8.1, double 9.1, double 10.1, double 11.1, double 12.1, double 13.1, double 14.1, double 15.1, double 16.1>, i32 %sel
ret double %ext
@@ -1191,6 +3103,39 @@ define i32 @extract_dyn_i32_6(<6 x i32> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 5, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_i32_6:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%x = extractelement <6 x i32> %arg, i32 %idx
ret i32 %x
@@ -1212,6 +3157,33 @@ define i32 @extract_dyn_i32_6_inreg(<6 x i32> inreg %arg, i32 inreg %idx) {
; GCN-NEXT: s_cselect_b32 s4, s21, s4
; GCN-NEXT: v_mov_b32_e32 v0, s4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_i32_6_inreg:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b32 s4, s17, s16
+; GCN-O0-NEXT: s_mov_b32 s5, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s5
+; GCN-O0-NEXT: s_cselect_b32 s4, s18, s4
+; GCN-O0-NEXT: s_mov_b32 s5, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s5
+; GCN-O0-NEXT: s_cselect_b32 s4, s19, s4
+; GCN-O0-NEXT: s_mov_b32 s5, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s5
+; GCN-O0-NEXT: s_cselect_b32 s4, s20, s4
+; GCN-O0-NEXT: s_mov_b32 s5, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s5
+; GCN-O0-NEXT: s_cselect_b32 s4, s21, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%x = extractelement <6 x i32> %arg, i32 %idx
ret i32 %x
@@ -1238,6 +3210,39 @@ define float @extract_dyn_float_6(<6 x float> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 5, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_float_6:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%x = extractelement <6 x float> %arg, i32 %idx
ret float %x
@@ -1269,6 +3274,43 @@ define float @extract_dyn_float_6_inreg(<6 x float> inreg %arg, i32 inreg %idx)
; GCN-NEXT: s_cselect_b64 vcc, -1, 0
; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_float_6_inreg:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%x = extractelement <6 x float> %arg, i32 %idx
ret float %x
diff --git a/llvm/test/CodeGen/AMDGPU/indirect-addressing-si.ll b/llvm/test/CodeGen/AMDGPU/indirect-addressing-si.ll
index 8fcf1ad3fbc95..771621e9a61a8 100644
--- a/llvm/test/CodeGen/AMDGPU/indirect-addressing-si.ll
+++ b/llvm/test/CodeGen/AMDGPU/indirect-addressing-si.ll
@@ -1500,13 +1500,6 @@ define amdgpu_kernel void @extract_neg_offset_vgpr(ptr addrspace(1) %out) {
;
; NOOPT-LABEL: extract_neg_offset_vgpr:
; NOOPT: ; %bb.0: ; %entry
-; NOOPT-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
-; NOOPT-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
-; NOOPT-NEXT: s_mov_b32 s22, -1
-; NOOPT-NEXT: s_mov_b32 s23, 0xe8f000
-; NOOPT-NEXT: s_add_u32 s20, s20, s11
-; NOOPT-NEXT: s_addc_u32 s21, s21, 0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:68 ; 4-byte Folded Spill
; NOOPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; NOOPT-NEXT: s_waitcnt lgkmcnt(0)
; NOOPT-NEXT: s_mov_b32 s6, s1
@@ -1517,146 +1510,71 @@ define amdgpu_kernel void @extract_neg_offset_vgpr(ptr addrspace(1) %out) {
; NOOPT-NEXT: s_mov_b32 s1, s6
; NOOPT-NEXT: s_mov_b32 s2, s5
; NOOPT-NEXT: s_mov_b32 s3, s4
-; NOOPT-NEXT: ; implicit-def: $vgpr31 : SGPR spill to VGPR lane
-; NOOPT-NEXT: v_writelane_b32 v31, s0, 0
-; NOOPT-NEXT: v_writelane_b32 v31, s1, 1
-; NOOPT-NEXT: v_writelane_b32 v31, s2, 2
-; NOOPT-NEXT: v_writelane_b32 v31, s3, 3
-; NOOPT-NEXT: s_mov_b32 s0, 16
-; NOOPT-NEXT: s_mov_b32 s1, 15
-; NOOPT-NEXT: s_mov_b32 s2, 14
-; NOOPT-NEXT: s_mov_b32 s3, 13
+; NOOPT-NEXT: s_mov_b32 s4, 0xfffffe00
+; NOOPT-NEXT: v_add_i32_e64 v1, s[4:5], v0, s4
+; NOOPT-NEXT: s_mov_b32 s6, 1
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s6
+; NOOPT-NEXT: s_mov_b32 s7, 0
+; NOOPT-NEXT: v_mov_b32_e32 v0, s7
+; NOOPT-NEXT: v_mov_b32_e32 v2, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; NOOPT-NEXT: s_mov_b32 s6, 2
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s6
+; NOOPT-NEXT: v_mov_b32_e32 v2, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; NOOPT-NEXT: s_mov_b32 s6, 3
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s6
+; NOOPT-NEXT: v_mov_b32_e32 v2, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[4:5]
+; NOOPT-NEXT: s_mov_b32 s4, 4
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 5
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 6
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 7
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 8
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 9
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 10
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 11
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
; NOOPT-NEXT: s_mov_b32 s4, 12
-; NOOPT-NEXT: s_mov_b32 s5, 11
-; NOOPT-NEXT: s_mov_b32 s6, 10
-; NOOPT-NEXT: s_mov_b32 s7, 9
-; NOOPT-NEXT: s_mov_b32 s8, 8
-; NOOPT-NEXT: s_mov_b32 s9, 7
-; NOOPT-NEXT: s_mov_b32 s10, 6
-; NOOPT-NEXT: s_mov_b32 s11, 5
-; NOOPT-NEXT: s_mov_b32 s12, 3
-; NOOPT-NEXT: s_mov_b32 s13, 2
-; NOOPT-NEXT: s_mov_b32 s14, 1
-; NOOPT-NEXT: s_mov_b32 s15, 0
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v0, s15
-; NOOPT-NEXT: v_mov_b32_e32 v30, s14
-; NOOPT-NEXT: v_mov_b32_e32 v29, s13
-; NOOPT-NEXT: v_mov_b32_e32 v28, s12
-; NOOPT-NEXT: v_mov_b32_e32 v27, s11
-; NOOPT-NEXT: v_mov_b32_e32 v26, s10
-; NOOPT-NEXT: v_mov_b32_e32 v25, s9
-; NOOPT-NEXT: v_mov_b32_e32 v24, s8
-; NOOPT-NEXT: v_mov_b32_e32 v23, s7
-; NOOPT-NEXT: v_mov_b32_e32 v22, s6
-; NOOPT-NEXT: v_mov_b32_e32 v21, s5
-; NOOPT-NEXT: v_mov_b32_e32 v20, s4
-; NOOPT-NEXT: v_mov_b32_e32 v19, s3
-; NOOPT-NEXT: v_mov_b32_e32 v18, s2
-; NOOPT-NEXT: v_mov_b32_e32 v17, s1
-; NOOPT-NEXT: v_mov_b32_e32 v16, s0
-; NOOPT-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 killed $exec
-; NOOPT-NEXT: v_mov_b32_e32 v1, v30
-; NOOPT-NEXT: v_mov_b32_e32 v2, v29
-; NOOPT-NEXT: v_mov_b32_e32 v3, v28
-; NOOPT-NEXT: v_mov_b32_e32 v4, v27
-; NOOPT-NEXT: v_mov_b32_e32 v5, v26
-; NOOPT-NEXT: v_mov_b32_e32 v6, v25
-; NOOPT-NEXT: v_mov_b32_e32 v7, v24
-; NOOPT-NEXT: v_mov_b32_e32 v8, v23
-; NOOPT-NEXT: v_mov_b32_e32 v9, v22
-; NOOPT-NEXT: v_mov_b32_e32 v10, v21
-; NOOPT-NEXT: v_mov_b32_e32 v11, v20
-; NOOPT-NEXT: v_mov_b32_e32 v12, v19
-; NOOPT-NEXT: v_mov_b32_e32 v13, v18
-; NOOPT-NEXT: v_mov_b32_e32 v14, v17
-; NOOPT-NEXT: v_mov_b32_e32 v15, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v31, s0, 4
-; NOOPT-NEXT: v_writelane_b32 v31, s1, 5
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_store_dword v31, off, s[20:23], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: ; implicit-def: $vgpr0
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 6
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 7
-; NOOPT-NEXT: buffer_load_dword v17, off, s[20:23], 0 offset:72 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[20:23], 0 offset:68 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v16
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v16
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_add_i32 m0, s2, 0xfffffe00
-; NOOPT-NEXT: v_movrels_b32_e32 v0, v0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:76 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:72 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v31, s2, 6
-; NOOPT-NEXT: v_writelane_b32 v31, s3, 7
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_store_dword v31, off, s[20:23], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB5_1
-; NOOPT-NEXT: ; %bb.2:
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 4
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 5
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.3:
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 0
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 1
-; NOOPT-NEXT: v_readlane_b32 s2, v31, 2
-; NOOPT-NEXT: v_readlane_b32 s3, v31, 3
-; NOOPT-NEXT: buffer_load_dword v0, off, s[20:23], 0 offset:76 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 13
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 14
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; NOOPT-NEXT: s_mov_b32 s4, 15
+; NOOPT-NEXT: v_mov_b32_e32 v2, s4
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[6:7]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v1, s4
+; NOOPT-NEXT: s_mov_b32 s6, 16
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; NOOPT-NEXT: buffer_store_dword v0, off, s[0:3], 0
; NOOPT-NEXT: s_endpgm
;
@@ -3999,13 +3917,6 @@ define amdgpu_kernel void @insert_neg_offset_vgpr(ptr addrspace(1) %in, ptr addr
;
; NOOPT-LABEL: insert_neg_offset_vgpr:
; NOOPT: ; %bb.0: ; %entry
-; NOOPT-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
-; NOOPT-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
-; NOOPT-NEXT: s_mov_b32 s22, -1
-; NOOPT-NEXT: s_mov_b32 s23, 0xe8f000
-; NOOPT-NEXT: s_add_u32 s20, s20, s11
-; NOOPT-NEXT: s_addc_u32 s21, s21, 0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:136 ; 4-byte Folded Spill
; NOOPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0xb
; NOOPT-NEXT: s_waitcnt lgkmcnt(0)
; NOOPT-NEXT: s_mov_b32 s6, s1
@@ -4016,234 +3927,90 @@ define amdgpu_kernel void @insert_neg_offset_vgpr(ptr addrspace(1) %in, ptr addr
; NOOPT-NEXT: s_mov_b32 s1, s6
; NOOPT-NEXT: s_mov_b32 s2, s5
; NOOPT-NEXT: s_mov_b32 s3, s4
-; NOOPT-NEXT: ; implicit-def: $vgpr31 : SGPR spill to VGPR lane
-; NOOPT-NEXT: v_writelane_b32 v31, s0, 0
-; NOOPT-NEXT: v_writelane_b32 v31, s1, 1
-; NOOPT-NEXT: v_writelane_b32 v31, s2, 2
-; NOOPT-NEXT: v_writelane_b32 v31, s3, 3
-; NOOPT-NEXT: s_mov_b32 s0, 16
-; NOOPT-NEXT: s_mov_b32 s1, 15
-; NOOPT-NEXT: s_mov_b32 s2, 14
-; NOOPT-NEXT: s_mov_b32 s3, 13
-; NOOPT-NEXT: s_mov_b32 s4, 12
+; NOOPT-NEXT: s_mov_b32 s4, 0xfffffe00
+; NOOPT-NEXT: v_add_i32_e64 v15, s[4:5], v0, s4
+; NOOPT-NEXT: s_mov_b32 s5, 3
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: s_mov_b32 s6, 33
+; NOOPT-NEXT: s_mov_b32 s4, 4
+; NOOPT-NEXT: v_mov_b32_e32 v0, s4
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v4, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 2
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v0, s5
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v5, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 1
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v0, s7
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v6, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 0
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v0, s5
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s10, 7
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s10
+; NOOPT-NEXT: s_mov_b32 s7, 8
+; NOOPT-NEXT: v_mov_b32_e32 v1, s7
+; NOOPT-NEXT: v_mov_b32_e32 v2, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 6
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v2, s10
+; NOOPT-NEXT: v_mov_b32_e32 v3, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s8, 5
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v15, s8
+; NOOPT-NEXT: v_mov_b32_e32 v3, s5
+; NOOPT-NEXT: v_mov_b32_e32 v7, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v3, v3, v7, s[10:11]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v15, s4
+; NOOPT-NEXT: v_mov_b32_e32 v7, s8
+; NOOPT-NEXT: v_mov_b32_e32 v8, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[4:5]
; NOOPT-NEXT: s_mov_b32 s5, 11
-; NOOPT-NEXT: s_mov_b32 s6, 10
-; NOOPT-NEXT: s_mov_b32 s7, 9
-; NOOPT-NEXT: s_mov_b32 s8, 8
-; NOOPT-NEXT: s_mov_b32 s9, 7
-; NOOPT-NEXT: s_mov_b32 s10, 6
-; NOOPT-NEXT: s_mov_b32 s11, 5
-; NOOPT-NEXT: s_mov_b32 s12, 4
-; NOOPT-NEXT: s_mov_b32 s13, 3
-; NOOPT-NEXT: s_mov_b32 s14, 2
-; NOOPT-NEXT: s_mov_b32 s15, 1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v0, s15
-; NOOPT-NEXT: v_mov_b32_e32 v30, s14
-; NOOPT-NEXT: v_mov_b32_e32 v29, s13
-; NOOPT-NEXT: v_mov_b32_e32 v28, s12
-; NOOPT-NEXT: v_mov_b32_e32 v27, s11
-; NOOPT-NEXT: v_mov_b32_e32 v26, s10
-; NOOPT-NEXT: v_mov_b32_e32 v25, s9
-; NOOPT-NEXT: v_mov_b32_e32 v24, s8
-; NOOPT-NEXT: v_mov_b32_e32 v23, s7
-; NOOPT-NEXT: v_mov_b32_e32 v22, s6
-; NOOPT-NEXT: v_mov_b32_e32 v21, s5
-; NOOPT-NEXT: v_mov_b32_e32 v20, s4
-; NOOPT-NEXT: v_mov_b32_e32 v19, s3
-; NOOPT-NEXT: v_mov_b32_e32 v18, s2
-; NOOPT-NEXT: v_mov_b32_e32 v17, s1
-; NOOPT-NEXT: v_mov_b32_e32 v16, s0
-; NOOPT-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 killed $exec
-; NOOPT-NEXT: v_mov_b32_e32 v1, v30
-; NOOPT-NEXT: v_mov_b32_e32 v2, v29
-; NOOPT-NEXT: v_mov_b32_e32 v3, v28
-; NOOPT-NEXT: v_mov_b32_e32 v4, v27
-; NOOPT-NEXT: v_mov_b32_e32 v5, v26
-; NOOPT-NEXT: v_mov_b32_e32 v6, v25
-; NOOPT-NEXT: v_mov_b32_e32 v7, v24
-; NOOPT-NEXT: v_mov_b32_e32 v8, v23
-; NOOPT-NEXT: v_mov_b32_e32 v9, v22
-; NOOPT-NEXT: v_mov_b32_e32 v10, v21
-; NOOPT-NEXT: v_mov_b32_e32 v11, v20
-; NOOPT-NEXT: v_mov_b32_e32 v12, v19
-; NOOPT-NEXT: v_mov_b32_e32 v13, v18
-; NOOPT-NEXT: v_mov_b32_e32 v14, v17
-; NOOPT-NEXT: v_mov_b32_e32 v15, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:72 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:76 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:80 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:84 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:88 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:92 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:96 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:100 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:104 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:108 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:112 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:116 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:120 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:124 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:128 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:132 ; 4-byte Folded Spill
-; NOOPT-NEXT: v_mov_b32_e32 v16, 33
-; NOOPT-NEXT: buffer_store_dword v16, off, s[20:23], 0 offset:68 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v31, s0, 4
-; NOOPT-NEXT: v_writelane_b32 v31, s1, 5
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_store_dword v31, off, s[20:23], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 6
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 7
-; NOOPT-NEXT: buffer_load_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(6)
-; NOOPT-NEXT: buffer_load_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(5)
-; NOOPT-NEXT: buffer_load_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(4)
-; NOOPT-NEXT: buffer_load_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(3)
-; NOOPT-NEXT: buffer_load_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(2)
-; NOOPT-NEXT: buffer_load_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(1)
-; NOOPT-NEXT: buffer_load_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[20:23], 0 offset:68 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[20:23], 0 offset:136 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v17
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v17
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_add_i32 m0, s2, 0xfffffe00
-; NOOPT-NEXT: v_movreld_b32_e32 v0, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:140 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:144 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:148 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:152 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:156 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:160 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:164 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:168 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:172 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:176 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:180 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:184 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:188 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:192 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:196 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:200 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v31, s2, 6
-; NOOPT-NEXT: v_writelane_b32 v31, s3, 7
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_store_dword v31, off, s[20:23], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB14_1
-; NOOPT-NEXT: ; %bb.2:
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 4
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 5
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.3:
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 0
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 1
-; NOOPT-NEXT: v_readlane_b32 s2, v31, 2
-; NOOPT-NEXT: v_readlane_b32 s3, v31, 3
-; NOOPT-NEXT: buffer_load_dword v15, off, s[20:23], 0 offset:140 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[20:23], 0 offset:144 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[20:23], 0 offset:148 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v18, off, s[20:23], 0 offset:152 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v19, off, s[20:23], 0 offset:156 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v20, off, s[20:23], 0 offset:160 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v21, off, s[20:23], 0 offset:164 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v22, off, s[20:23], 0 offset:168 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v23, off, s[20:23], 0 offset:172 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v24, off, s[20:23], 0 offset:176 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v25, off, s[20:23], 0 offset:180 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v26, off, s[20:23], 0 offset:184 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v27, off, s[20:23], 0 offset:188 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v28, off, s[20:23], 0 offset:192 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v29, off, s[20:23], 0 offset:196 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v30, off, s[20:23], 0 offset:200 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(12)
-; NOOPT-NEXT: v_mov_b32_e32 v4, v18
-; NOOPT-NEXT: v_mov_b32_e32 v5, v17
-; NOOPT-NEXT: v_mov_b32_e32 v6, v16
-; NOOPT-NEXT: v_mov_b32_e32 v0, v15
-; NOOPT-NEXT: s_waitcnt vmcnt(8)
-; NOOPT-NEXT: v_mov_b32_e32 v1, v22
-; NOOPT-NEXT: v_mov_b32_e32 v2, v21
-; NOOPT-NEXT: v_mov_b32_e32 v3, v20
-; NOOPT-NEXT: v_mov_b32_e32 v7, v19
-; NOOPT-NEXT: s_waitcnt vmcnt(4)
-; NOOPT-NEXT: v_mov_b32_e32 v12, v26
-; NOOPT-NEXT: v_mov_b32_e32 v13, v25
-; NOOPT-NEXT: v_mov_b32_e32 v14, v24
-; NOOPT-NEXT: v_mov_b32_e32 v8, v23
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v9, v30
-; NOOPT-NEXT: v_mov_b32_e32 v10, v29
-; NOOPT-NEXT: v_mov_b32_e32 v11, v28
-; NOOPT-NEXT: v_mov_b32_e32 v15, v27
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: s_mov_b32 s4, 12
+; NOOPT-NEXT: v_mov_b32_e32 v8, s4
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v12, v8, v9, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s10, 10
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s10
+; NOOPT-NEXT: v_mov_b32_e32 v8, s5
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v13, v8, v9, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 9
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v8, s10
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v14, v8, v9, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v8, s5
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v8, v8, v9, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 15
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: s_mov_b32 s5, 16
+; NOOPT-NEXT: v_mov_b32_e32 v9, s5
+; NOOPT-NEXT: v_mov_b32_e32 v10, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v9, v9, v10, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 14
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v10, s7
+; NOOPT-NEXT: v_mov_b32_e32 v11, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v10, v10, v11, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 13
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v11, s5
+; NOOPT-NEXT: v_mov_b32_e32 v16, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v11, v11, v16, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v15, s4
+; NOOPT-NEXT: v_mov_b32_e32 v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v16, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v15, v15, v16, s[4:5]
; NOOPT-NEXT: ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16_vgpr17_vgpr18 killed $exec
; NOOPT-NEXT: v_mov_b32_e32 v16, v11
; NOOPT-NEXT: v_mov_b32_e32 v17, v10
@@ -4472,13 +4239,6 @@ define amdgpu_kernel void @insert_neg_inline_offset_vgpr(ptr addrspace(1) %in, p
;
; NOOPT-LABEL: insert_neg_inline_offset_vgpr:
; NOOPT: ; %bb.0: ; %entry
-; NOOPT-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
-; NOOPT-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
-; NOOPT-NEXT: s_mov_b32 s22, -1
-; NOOPT-NEXT: s_mov_b32 s23, 0xe8f000
-; NOOPT-NEXT: s_add_u32 s20, s20, s11
-; NOOPT-NEXT: s_addc_u32 s21, s21, 0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:136 ; 4-byte Folded Spill
; NOOPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0xb
; NOOPT-NEXT: s_waitcnt lgkmcnt(0)
; NOOPT-NEXT: s_mov_b32 s6, s1
@@ -4489,234 +4249,90 @@ define amdgpu_kernel void @insert_neg_inline_offset_vgpr(ptr addrspace(1) %in, p
; NOOPT-NEXT: s_mov_b32 s1, s6
; NOOPT-NEXT: s_mov_b32 s2, s5
; NOOPT-NEXT: s_mov_b32 s3, s4
-; NOOPT-NEXT: ; implicit-def: $vgpr31 : SGPR spill to VGPR lane
-; NOOPT-NEXT: v_writelane_b32 v31, s0, 0
-; NOOPT-NEXT: v_writelane_b32 v31, s1, 1
-; NOOPT-NEXT: v_writelane_b32 v31, s2, 2
-; NOOPT-NEXT: v_writelane_b32 v31, s3, 3
-; NOOPT-NEXT: s_mov_b32 s0, 16
-; NOOPT-NEXT: s_mov_b32 s1, 15
-; NOOPT-NEXT: s_mov_b32 s2, 14
-; NOOPT-NEXT: s_mov_b32 s3, 13
-; NOOPT-NEXT: s_mov_b32 s4, 12
+; NOOPT-NEXT: s_mov_b32 s4, -16
+; NOOPT-NEXT: v_add_i32_e64 v15, s[4:5], v0, s4
+; NOOPT-NEXT: s_mov_b32 s5, 3
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: s_mov_b32 s6, 0x1f4
+; NOOPT-NEXT: s_mov_b32 s4, 4
+; NOOPT-NEXT: v_mov_b32_e32 v0, s4
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v4, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 2
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v0, s5
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v5, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 1
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v0, s7
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v6, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 0
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v0, s5
+; NOOPT-NEXT: v_mov_b32_e32 v1, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s10, 7
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s10
+; NOOPT-NEXT: s_mov_b32 s7, 8
+; NOOPT-NEXT: v_mov_b32_e32 v1, s7
+; NOOPT-NEXT: v_mov_b32_e32 v2, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 6
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v2, s10
+; NOOPT-NEXT: v_mov_b32_e32 v3, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s8, 5
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v15, s8
+; NOOPT-NEXT: v_mov_b32_e32 v3, s5
+; NOOPT-NEXT: v_mov_b32_e32 v7, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v3, v3, v7, s[10:11]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v15, s4
+; NOOPT-NEXT: v_mov_b32_e32 v7, s8
+; NOOPT-NEXT: v_mov_b32_e32 v8, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v7, v7, v8, s[4:5]
; NOOPT-NEXT: s_mov_b32 s5, 11
-; NOOPT-NEXT: s_mov_b32 s6, 10
-; NOOPT-NEXT: s_mov_b32 s7, 9
-; NOOPT-NEXT: s_mov_b32 s8, 8
-; NOOPT-NEXT: s_mov_b32 s9, 7
-; NOOPT-NEXT: s_mov_b32 s10, 6
-; NOOPT-NEXT: s_mov_b32 s11, 5
-; NOOPT-NEXT: s_mov_b32 s12, 4
-; NOOPT-NEXT: s_mov_b32 s13, 3
-; NOOPT-NEXT: s_mov_b32 s14, 2
-; NOOPT-NEXT: s_mov_b32 s15, 1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v0, s15
-; NOOPT-NEXT: v_mov_b32_e32 v30, s14
-; NOOPT-NEXT: v_mov_b32_e32 v29, s13
-; NOOPT-NEXT: v_mov_b32_e32 v28, s12
-; NOOPT-NEXT: v_mov_b32_e32 v27, s11
-; NOOPT-NEXT: v_mov_b32_e32 v26, s10
-; NOOPT-NEXT: v_mov_b32_e32 v25, s9
-; NOOPT-NEXT: v_mov_b32_e32 v24, s8
-; NOOPT-NEXT: v_mov_b32_e32 v23, s7
-; NOOPT-NEXT: v_mov_b32_e32 v22, s6
-; NOOPT-NEXT: v_mov_b32_e32 v21, s5
-; NOOPT-NEXT: v_mov_b32_e32 v20, s4
-; NOOPT-NEXT: v_mov_b32_e32 v19, s3
-; NOOPT-NEXT: v_mov_b32_e32 v18, s2
-; NOOPT-NEXT: v_mov_b32_e32 v17, s1
-; NOOPT-NEXT: v_mov_b32_e32 v16, s0
-; NOOPT-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 killed $exec
-; NOOPT-NEXT: v_mov_b32_e32 v1, v30
-; NOOPT-NEXT: v_mov_b32_e32 v2, v29
-; NOOPT-NEXT: v_mov_b32_e32 v3, v28
-; NOOPT-NEXT: v_mov_b32_e32 v4, v27
-; NOOPT-NEXT: v_mov_b32_e32 v5, v26
-; NOOPT-NEXT: v_mov_b32_e32 v6, v25
-; NOOPT-NEXT: v_mov_b32_e32 v7, v24
-; NOOPT-NEXT: v_mov_b32_e32 v8, v23
-; NOOPT-NEXT: v_mov_b32_e32 v9, v22
-; NOOPT-NEXT: v_mov_b32_e32 v10, v21
-; NOOPT-NEXT: v_mov_b32_e32 v11, v20
-; NOOPT-NEXT: v_mov_b32_e32 v12, v19
-; NOOPT-NEXT: v_mov_b32_e32 v13, v18
-; NOOPT-NEXT: v_mov_b32_e32 v14, v17
-; NOOPT-NEXT: v_mov_b32_e32 v15, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:72 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:76 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:80 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:84 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:88 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:92 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:96 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:100 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:104 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:108 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:112 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:116 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:120 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:124 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:128 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:132 ; 4-byte Folded Spill
-; NOOPT-NEXT: v_mov_b32_e32 v16, 0x1f4
-; NOOPT-NEXT: buffer_store_dword v16, off, s[20:23], 0 offset:68 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v31, s0, 4
-; NOOPT-NEXT: v_writelane_b32 v31, s1, 5
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_store_dword v31, off, s[20:23], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB15_1: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 6
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 7
-; NOOPT-NEXT: buffer_load_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(6)
-; NOOPT-NEXT: buffer_load_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(5)
-; NOOPT-NEXT: buffer_load_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(4)
-; NOOPT-NEXT: buffer_load_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(3)
-; NOOPT-NEXT: buffer_load_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(2)
-; NOOPT-NEXT: buffer_load_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(1)
-; NOOPT-NEXT: buffer_load_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[20:23], 0 offset:68 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[20:23], 0 offset:136 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v17
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v17
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_add_i32 m0, s2, -16
-; NOOPT-NEXT: v_movreld_b32_e32 v0, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:140 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:144 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:148 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:152 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:156 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:160 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:164 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:168 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:172 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:176 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:180 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:184 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:188 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:192 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:196 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:200 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[20:23], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[20:23], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[20:23], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[20:23], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[20:23], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[20:23], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[20:23], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[20:23], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[20:23], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[20:23], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[20:23], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[20:23], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[20:23], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[20:23], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[20:23], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[20:23], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v31, s2, 6
-; NOOPT-NEXT: v_writelane_b32 v31, s3, 7
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_store_dword v31, off, s[20:23], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB15_1
-; NOOPT-NEXT: ; %bb.2:
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 4
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 5
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.3:
-; NOOPT-NEXT: s_or_saveexec_b64 s[16:17], -1
-; NOOPT-NEXT: buffer_load_dword v31, off, s[20:23], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[16:17]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v31, 0
-; NOOPT-NEXT: v_readlane_b32 s1, v31, 1
-; NOOPT-NEXT: v_readlane_b32 s2, v31, 2
-; NOOPT-NEXT: v_readlane_b32 s3, v31, 3
-; NOOPT-NEXT: buffer_load_dword v15, off, s[20:23], 0 offset:140 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[20:23], 0 offset:144 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[20:23], 0 offset:148 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v18, off, s[20:23], 0 offset:152 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v19, off, s[20:23], 0 offset:156 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v20, off, s[20:23], 0 offset:160 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v21, off, s[20:23], 0 offset:164 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v22, off, s[20:23], 0 offset:168 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v23, off, s[20:23], 0 offset:172 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v24, off, s[20:23], 0 offset:176 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v25, off, s[20:23], 0 offset:180 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v26, off, s[20:23], 0 offset:184 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v27, off, s[20:23], 0 offset:188 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v28, off, s[20:23], 0 offset:192 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v29, off, s[20:23], 0 offset:196 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v30, off, s[20:23], 0 offset:200 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(12)
-; NOOPT-NEXT: v_mov_b32_e32 v4, v18
-; NOOPT-NEXT: v_mov_b32_e32 v5, v17
-; NOOPT-NEXT: v_mov_b32_e32 v6, v16
-; NOOPT-NEXT: v_mov_b32_e32 v0, v15
-; NOOPT-NEXT: s_waitcnt vmcnt(8)
-; NOOPT-NEXT: v_mov_b32_e32 v1, v22
-; NOOPT-NEXT: v_mov_b32_e32 v2, v21
-; NOOPT-NEXT: v_mov_b32_e32 v3, v20
-; NOOPT-NEXT: v_mov_b32_e32 v7, v19
-; NOOPT-NEXT: s_waitcnt vmcnt(4)
-; NOOPT-NEXT: v_mov_b32_e32 v12, v26
-; NOOPT-NEXT: v_mov_b32_e32 v13, v25
-; NOOPT-NEXT: v_mov_b32_e32 v14, v24
-; NOOPT-NEXT: v_mov_b32_e32 v8, v23
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v9, v30
-; NOOPT-NEXT: v_mov_b32_e32 v10, v29
-; NOOPT-NEXT: v_mov_b32_e32 v11, v28
-; NOOPT-NEXT: v_mov_b32_e32 v15, v27
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: s_mov_b32 s4, 12
+; NOOPT-NEXT: v_mov_b32_e32 v8, s4
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v12, v8, v9, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s10, 10
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s10
+; NOOPT-NEXT: v_mov_b32_e32 v8, s5
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v13, v8, v9, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 9
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v8, s10
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v14, v8, v9, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v8, s5
+; NOOPT-NEXT: v_mov_b32_e32 v9, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v8, v8, v9, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 15
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: s_mov_b32 s5, 16
+; NOOPT-NEXT: v_mov_b32_e32 v9, s5
+; NOOPT-NEXT: v_mov_b32_e32 v10, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v9, v9, v10, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s5, 14
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s5
+; NOOPT-NEXT: v_mov_b32_e32 v10, s7
+; NOOPT-NEXT: v_mov_b32_e32 v11, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v10, v10, v11, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s7, 13
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v11, s5
+; NOOPT-NEXT: v_mov_b32_e32 v16, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v11, v11, v16, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v15, s4
+; NOOPT-NEXT: v_mov_b32_e32 v15, s7
+; NOOPT-NEXT: v_mov_b32_e32 v16, s6
+; NOOPT-NEXT: v_cndmask_b32_e64 v15, v15, v16, s[4:5]
; NOOPT-NEXT: ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16_vgpr17_vgpr18 killed $exec
; NOOPT-NEXT: v_mov_b32_e32 v16, v11
; NOOPT-NEXT: v_mov_b32_e32 v17, v10
@@ -4996,336 +4612,182 @@ define amdgpu_kernel void @extract_vgpr_offset_multiple_in_block(ptr addrspace(1
;
; NOOPT-LABEL: extract_vgpr_offset_multiple_in_block:
; NOOPT: ; %bb.0: ; %entry
-; NOOPT-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
-; NOOPT-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
-; NOOPT-NEXT: s_mov_b32 s38, -1
-; NOOPT-NEXT: s_mov_b32 s39, 0xe8f000
-; NOOPT-NEXT: s_add_u32 s36, s36, s11
-; NOOPT-NEXT: s_addc_u32 s37, s37, 0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:76 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x9
-; NOOPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0xd
+; NOOPT-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
+; NOOPT-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
+; NOOPT-NEXT: s_mov_b32 s26, -1
+; NOOPT-NEXT: s_mov_b32 s27, 0xe8f000
+; NOOPT-NEXT: s_add_u32 s24, s24, s11
+; NOOPT-NEXT: s_addc_u32 s25, s25, 0
+; NOOPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; NOOPT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0xd
; NOOPT-NEXT: s_waitcnt lgkmcnt(0)
-; NOOPT-NEXT: s_mov_b32 s8, s3
-; NOOPT-NEXT: s_mov_b32 s4, s2
-; NOOPT-NEXT: s_mov_b32 s2, 0xf000
-; NOOPT-NEXT: s_mov_b32 s3, -1
-; NOOPT-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
-; NOOPT-NEXT: s_mov_b32 s5, s8
-; NOOPT-NEXT: s_mov_b32 s6, s3
-; NOOPT-NEXT: s_mov_b32 s7, s2
-; NOOPT-NEXT: ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
-; NOOPT-NEXT: v_writelane_b32 v18, s4, 0
-; NOOPT-NEXT: v_writelane_b32 v18, s5, 1
-; NOOPT-NEXT: v_writelane_b32 v18, s6, 2
-; NOOPT-NEXT: v_writelane_b32 v18, s7, 3
-; NOOPT-NEXT: s_mov_b32 s4, 0
-; NOOPT-NEXT: v_writelane_b32 v18, s4, 4
-; NOOPT-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
-; NOOPT-NEXT: s_mov_b32 s5, s2
-; NOOPT-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[4:5]
-; NOOPT-NEXT: s_mov_b32 s4, 2
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_lshlrev_b32_e64 v0, s4, v0
-; NOOPT-NEXT: v_mov_b32_e32 v2, 0
-; NOOPT-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
-; NOOPT-NEXT: v_mov_b32_e32 v1, v2
-; NOOPT-NEXT: buffer_load_dword v0, v[0:1], s[0:3], 0 addr64 glc
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:72 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b32 s0, 1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_add_i32_e64 v0, s[0:1], v0, s0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:68 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b32 s16, 16
-; NOOPT-NEXT: s_mov_b32 s17, 15
-; NOOPT-NEXT: s_mov_b32 s18, 14
-; NOOPT-NEXT: s_mov_b32 s20, 12
-; NOOPT-NEXT: s_mov_b32 s22, 10
-; NOOPT-NEXT: s_mov_b32 s24, 8
-; NOOPT-NEXT: s_mov_b32 s26, 6
-; NOOPT-NEXT: s_mov_b32 s27, 5
-; NOOPT-NEXT: s_mov_b32 s19, 13
-; NOOPT-NEXT: s_mov_b32 s21, 11
-; NOOPT-NEXT: s_mov_b32 s23, 9
-; NOOPT-NEXT: s_mov_b32 s25, 7
-; NOOPT-NEXT: s_mov_b32 s0, s25
-; NOOPT-NEXT: s_mov_b32 s1, s23
-; NOOPT-NEXT: s_mov_b32 s2, s21
-; NOOPT-NEXT: s_mov_b32 s3, s19
-; NOOPT-NEXT: s_mov_b32 s4, s27
-; NOOPT-NEXT: s_mov_b32 s5, s26
-; NOOPT-NEXT: s_mov_b32 s6, s25
-; NOOPT-NEXT: s_mov_b32 s7, s24
-; NOOPT-NEXT: s_mov_b32 s8, s23
-; NOOPT-NEXT: s_mov_b32 s9, s22
-; NOOPT-NEXT: s_mov_b32 s10, s21
-; NOOPT-NEXT: s_mov_b32 s11, s20
-; NOOPT-NEXT: s_mov_b32 s12, s19
-; NOOPT-NEXT: s_mov_b32 s13, s18
-; NOOPT-NEXT: s_mov_b32 s14, s17
-; NOOPT-NEXT: s_mov_b32 s15, s16
-; NOOPT-NEXT: v_writelane_b32 v18, s0, 5
-; NOOPT-NEXT: v_writelane_b32 v18, s1, 6
-; NOOPT-NEXT: v_writelane_b32 v18, s2, 7
-; NOOPT-NEXT: v_writelane_b32 v18, s3, 8
-; NOOPT-NEXT: v_writelane_b32 v18, s4, 9
-; NOOPT-NEXT: v_writelane_b32 v18, s5, 10
-; NOOPT-NEXT: v_writelane_b32 v18, s6, 11
-; NOOPT-NEXT: v_writelane_b32 v18, s7, 12
-; NOOPT-NEXT: v_writelane_b32 v18, s8, 13
-; NOOPT-NEXT: v_writelane_b32 v18, s9, 14
-; NOOPT-NEXT: v_writelane_b32 v18, s10, 15
-; NOOPT-NEXT: v_writelane_b32 v18, s11, 16
-; NOOPT-NEXT: v_writelane_b32 v18, s12, 17
-; NOOPT-NEXT: v_writelane_b32 v18, s13, 18
-; NOOPT-NEXT: v_writelane_b32 v18, s14, 19
-; NOOPT-NEXT: v_writelane_b32 v18, s15, 20
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v0, s0
-; NOOPT-NEXT: v_mov_b32_e32 v1, s1
-; NOOPT-NEXT: v_mov_b32_e32 v2, s2
-; NOOPT-NEXT: v_mov_b32_e32 v3, s3
-; NOOPT-NEXT: v_mov_b32_e32 v4, s4
-; NOOPT-NEXT: v_mov_b32_e32 v5, s5
-; NOOPT-NEXT: v_mov_b32_e32 v6, s6
-; NOOPT-NEXT: v_mov_b32_e32 v7, s7
-; NOOPT-NEXT: v_mov_b32_e32 v8, s8
-; NOOPT-NEXT: v_mov_b32_e32 v9, s9
-; NOOPT-NEXT: v_mov_b32_e32 v10, s10
-; NOOPT-NEXT: v_mov_b32_e32 v11, s11
-; NOOPT-NEXT: v_mov_b32_e32 v12, s12
-; NOOPT-NEXT: v_mov_b32_e32 v13, s13
-; NOOPT-NEXT: v_mov_b32_e32 v14, s14
-; NOOPT-NEXT: v_mov_b32_e32 v15, s15
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[36:39], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[36:39], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[36:39], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[36:39], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[36:39], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[36:39], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[36:39], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[36:39], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[36:39], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[36:39], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[36:39], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[36:39], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[36:39], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[36:39], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[36:39], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v18, s0, 21
-; NOOPT-NEXT: v_writelane_b32 v18, s1, 22
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_store_dword v18, off, s[36:39], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: ; implicit-def: $vgpr0
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB16_1: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 23
-; NOOPT-NEXT: v_readlane_b32 s1, v18, 24
-; NOOPT-NEXT: buffer_load_dword v17, off, s[36:39], 0 offset:80 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v0, off, s[36:39], 0 offset:4 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[36:39], 0 offset:8 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[36:39], 0 offset:12 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[36:39], 0 offset:16 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[36:39], 0 offset:20 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[36:39], 0 offset:24 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[36:39], 0 offset:28 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[36:39], 0 offset:32 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[36:39], 0 offset:36 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v9, off, s[36:39], 0 offset:40 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v10, off, s[36:39], 0 offset:44 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v11, off, s[36:39], 0 offset:48 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v12, off, s[36:39], 0 offset:52 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v13, off, s[36:39], 0 offset:56 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v14, off, s[36:39], 0 offset:60 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v15, off, s[36:39], 0 offset:64 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[36:39], 0 offset:72 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v16
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v16
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_mov_b32 m0, s2
-; NOOPT-NEXT: v_movrels_b32_e32 v0, v0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:84 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:80 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v18, s2, 23
-; NOOPT-NEXT: v_writelane_b32 v18, s3, 24
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_store_dword v18, off, s[36:39], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB16_1
-; NOOPT-NEXT: ; %bb.2:
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
+; NOOPT-NEXT: s_mov_b32 s2, s1
+; NOOPT-NEXT: s_mov_b32 s8, s0
+; NOOPT-NEXT: s_mov_b32 s1, 0xf000
+; NOOPT-NEXT: s_mov_b32 s0, -1
+; NOOPT-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
+; NOOPT-NEXT: s_mov_b32 s9, s2
+; NOOPT-NEXT: s_mov_b32 s10, s0
+; NOOPT-NEXT: s_mov_b32 s11, s1
+; NOOPT-NEXT: ; implicit-def: $vgpr5 : SGPR spill to VGPR lane
+; NOOPT-NEXT: v_writelane_b32 v5, s8, 0
+; NOOPT-NEXT: v_writelane_b32 v5, s9, 1
+; NOOPT-NEXT: v_writelane_b32 v5, s10, 2
+; NOOPT-NEXT: v_writelane_b32 v5, s11, 3
+; NOOPT-NEXT: s_mov_b32 s0, 0
+; NOOPT-NEXT: s_mov_b32 s2, s0
+; NOOPT-NEXT: s_mov_b32 s3, s1
+; NOOPT-NEXT: ; kill: def $sgpr4_sgpr5 killed $sgpr4_sgpr5 def $sgpr4_sgpr5_sgpr6_sgpr7
+; NOOPT-NEXT: s_mov_b64 s[6:7], s[2:3]
+; NOOPT-NEXT: s_mov_b32 s19, 2
+; NOOPT-NEXT: v_lshlrev_b32_e64 v1, s19, v0
+; NOOPT-NEXT: v_mov_b32_e32 v3, 0
+; NOOPT-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
+; NOOPT-NEXT: v_mov_b32_e32 v2, v3
+; NOOPT-NEXT: buffer_load_dword v2, v[1:2], s[4:7], 0 addr64 glc
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 21
-; NOOPT-NEXT: v_readlane_b32 s1, v18, 22
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.3:
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
+; NOOPT-NEXT: buffer_store_dword v2, off, s[24:27], 0 offset:8 ; 4-byte Folded Spill
+; NOOPT-NEXT: s_mov_b32 s20, 1
+; NOOPT-NEXT: v_add_i32_e64 v3, s[2:3], v2, s20
+; NOOPT-NEXT: buffer_store_dword v3, off, s[24:27], 0 offset:4 ; 4-byte Folded Spill
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s20
+; NOOPT-NEXT: s_mov_b32 s12, 9
+; NOOPT-NEXT: s_mov_b32 s14, 7
+; NOOPT-NEXT: v_mov_b32_e32 v1, s14
+; NOOPT-NEXT: v_mov_b32_e32 v4, s12
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s19
+; NOOPT-NEXT: s_mov_b32 s10, 11
+; NOOPT-NEXT: v_mov_b32_e32 v4, s10
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: s_mov_b32 s18, 3
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s18
+; NOOPT-NEXT: s_mov_b32 s8, 13
+; NOOPT-NEXT: v_mov_b32_e32 v4, s8
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: s_mov_b32 s17, 4
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s17
+; NOOPT-NEXT: s_mov_b32 s16, 5
+; NOOPT-NEXT: v_mov_b32_e32 v4, s16
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s16
+; NOOPT-NEXT: s_mov_b32 s15, 6
+; NOOPT-NEXT: v_mov_b32_e32 v4, s15
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s15
+; NOOPT-NEXT: v_mov_b32_e32 v4, s14
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s14
+; NOOPT-NEXT: s_mov_b32 s13, 8
+; NOOPT-NEXT: v_mov_b32_e32 v4, s13
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s13
+; NOOPT-NEXT: v_mov_b32_e32 v4, s12
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s12
+; NOOPT-NEXT: s_mov_b32 s11, 10
+; NOOPT-NEXT: v_mov_b32_e32 v4, s11
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s11
+; NOOPT-NEXT: v_mov_b32_e32 v4, s10
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s10
+; NOOPT-NEXT: s_mov_b32 s9, 12
+; NOOPT-NEXT: v_mov_b32_e32 v4, s9
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v2, s9
+; NOOPT-NEXT: v_mov_b32_e32 v4, s8
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[2:3]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v2, s8
+; NOOPT-NEXT: s_mov_b32 s3, 14
+; NOOPT-NEXT: v_mov_b32_e32 v4, s3
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[4:5]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v2, s3
+; NOOPT-NEXT: s_mov_b32 s2, 15
+; NOOPT-NEXT: v_mov_b32_e32 v4, s2
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[4:5]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[4:5], v2, s2
+; NOOPT-NEXT: s_mov_b32 s1, 16
+; NOOPT-NEXT: s_waitcnt expcnt(1)
+; NOOPT-NEXT: v_mov_b32_e32 v2, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[4:5]
; NOOPT-NEXT: ;;#ASMSTART
; NOOPT-NEXT: s_mov_b32 s4, 17
; NOOPT-NEXT: ;;#ASMEND
-; NOOPT-NEXT: s_mov_b32 s16, s4
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 5
-; NOOPT-NEXT: v_readlane_b32 s1, v18, 6
-; NOOPT-NEXT: v_readlane_b32 s2, v18, 7
-; NOOPT-NEXT: v_readlane_b32 s3, v18, 8
-; NOOPT-NEXT: v_readlane_b32 s4, v18, 9
-; NOOPT-NEXT: v_readlane_b32 s5, v18, 10
-; NOOPT-NEXT: v_readlane_b32 s6, v18, 11
-; NOOPT-NEXT: v_readlane_b32 s7, v18, 12
-; NOOPT-NEXT: v_readlane_b32 s8, v18, 13
-; NOOPT-NEXT: v_readlane_b32 s9, v18, 14
-; NOOPT-NEXT: v_readlane_b32 s10, v18, 15
-; NOOPT-NEXT: v_readlane_b32 s11, v18, 16
-; NOOPT-NEXT: v_readlane_b32 s12, v18, 17
-; NOOPT-NEXT: v_readlane_b32 s13, v18, 18
-; NOOPT-NEXT: v_readlane_b32 s14, v18, 19
-; NOOPT-NEXT: v_readlane_b32 s15, v18, 20
-; NOOPT-NEXT: v_writelane_b32 v18, s16, 25
-; NOOPT-NEXT: v_mov_b32_e32 v0, s0
-; NOOPT-NEXT: v_mov_b32_e32 v1, s1
-; NOOPT-NEXT: v_mov_b32_e32 v2, s2
-; NOOPT-NEXT: v_mov_b32_e32 v3, s3
-; NOOPT-NEXT: v_mov_b32_e32 v4, s4
-; NOOPT-NEXT: v_mov_b32_e32 v5, s5
-; NOOPT-NEXT: v_mov_b32_e32 v6, s6
-; NOOPT-NEXT: v_mov_b32_e32 v7, s7
-; NOOPT-NEXT: v_mov_b32_e32 v8, s8
-; NOOPT-NEXT: v_mov_b32_e32 v9, s9
-; NOOPT-NEXT: v_mov_b32_e32 v10, s10
-; NOOPT-NEXT: v_mov_b32_e32 v11, s11
-; NOOPT-NEXT: v_mov_b32_e32 v12, s12
-; NOOPT-NEXT: v_mov_b32_e32 v13, s13
-; NOOPT-NEXT: v_mov_b32_e32 v14, s14
-; NOOPT-NEXT: v_mov_b32_e32 v15, s15
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:88 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[36:39], 0 offset:92 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[36:39], 0 offset:96 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[36:39], 0 offset:100 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[36:39], 0 offset:104 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[36:39], 0 offset:108 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[36:39], 0 offset:112 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[36:39], 0 offset:116 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[36:39], 0 offset:120 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[36:39], 0 offset:124 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[36:39], 0 offset:128 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[36:39], 0 offset:132 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[36:39], 0 offset:136 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[36:39], 0 offset:140 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[36:39], 0 offset:144 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[36:39], 0 offset:148 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v18, s0, 26
-; NOOPT-NEXT: v_writelane_b32 v18, s1, 27
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_store_dword v18, off, s[36:39], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: ; implicit-def: $vgpr0
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB16_4: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 28
-; NOOPT-NEXT: v_readlane_b32 s1, v18, 29
-; NOOPT-NEXT: buffer_load_dword v17, off, s[36:39], 0 offset:152 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v0, off, s[36:39], 0 offset:88 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[36:39], 0 offset:92 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[36:39], 0 offset:96 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[36:39], 0 offset:100 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[36:39], 0 offset:104 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[36:39], 0 offset:108 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[36:39], 0 offset:112 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[36:39], 0 offset:116 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[36:39], 0 offset:120 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v9, off, s[36:39], 0 offset:124 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v10, off, s[36:39], 0 offset:128 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v11, off, s[36:39], 0 offset:132 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v12, off, s[36:39], 0 offset:136 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v13, off, s[36:39], 0 offset:140 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v14, off, s[36:39], 0 offset:144 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v15, off, s[36:39], 0 offset:148 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[36:39], 0 offset:68 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v16
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v16
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_mov_b32 m0, s2
-; NOOPT-NEXT: v_movrels_b32_e32 v0, v0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:156 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[36:39], 0 offset:152 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v18, s2, 28
-; NOOPT-NEXT: v_writelane_b32 v18, s3, 29
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_store_dword v18, off, s[36:39], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB16_4
-; NOOPT-NEXT: ; %bb.5:
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
+; NOOPT-NEXT: s_mov_b32 s21, s4
+; NOOPT-NEXT: v_readlane_b32 s4, v5, 0
+; NOOPT-NEXT: v_readlane_b32 s5, v5, 1
+; NOOPT-NEXT: v_readlane_b32 s6, v5, 2
+; NOOPT-NEXT: v_readlane_b32 s7, v5, 3
+; NOOPT-NEXT: v_writelane_b32 v5, s21, 4
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v3, s20
+; NOOPT-NEXT: v_mov_b32_e32 v1, s14
+; NOOPT-NEXT: v_mov_b32_e32 v4, s12
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v3, s19
+; NOOPT-NEXT: v_mov_b32_e32 v4, s10
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[18:19], v3, s18
+; NOOPT-NEXT: v_mov_b32_e32 v4, s8
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[18:19]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[18:19], v3, s17
+; NOOPT-NEXT: v_mov_b32_e32 v4, s16
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[18:19]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[16:17], v3, s16
+; NOOPT-NEXT: v_mov_b32_e32 v4, s15
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[16:17]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[16:17], v3, s15
+; NOOPT-NEXT: v_mov_b32_e32 v4, s14
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[16:17]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[14:15], v3, s14
+; NOOPT-NEXT: v_mov_b32_e32 v4, s13
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[14:15]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[14:15], v3, s13
+; NOOPT-NEXT: v_mov_b32_e32 v4, s12
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[14:15]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[12:13], v3, s12
+; NOOPT-NEXT: v_mov_b32_e32 v4, s11
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[12:13]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[12:13], v3, s11
+; NOOPT-NEXT: v_mov_b32_e32 v4, s10
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[12:13]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v3, s10
+; NOOPT-NEXT: v_mov_b32_e32 v4, s9
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[10:11]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v3, s9
+; NOOPT-NEXT: v_mov_b32_e32 v4, s8
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[10:11]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v3, s8
+; NOOPT-NEXT: v_mov_b32_e32 v4, s3
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v3, s3
+; NOOPT-NEXT: v_mov_b32_e32 v4, s2
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v3, s2
; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 26
-; NOOPT-NEXT: v_readlane_b32 s1, v18, 27
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.6:
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 4
-; NOOPT-NEXT: v_readlane_b32 s4, v18, 0
-; NOOPT-NEXT: v_readlane_b32 s5, v18, 1
-; NOOPT-NEXT: v_readlane_b32 s6, v18, 2
-; NOOPT-NEXT: v_readlane_b32 s7, v18, 3
-; NOOPT-NEXT: buffer_load_dword v0, off, s[36:39], 0 offset:76 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[36:39], 0 offset:156 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[36:39], 0 offset:84 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
+; NOOPT-NEXT: v_mov_b32_e32 v3, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v3, s[2:3]
; NOOPT-NEXT: buffer_store_dword v2, off, s[4:7], 0
; NOOPT-NEXT: s_waitcnt vmcnt(0)
; NOOPT-NEXT: buffer_store_dword v1, off, s[4:7], 0
; NOOPT-NEXT: s_waitcnt vmcnt(0)
; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v0, s0
; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v18, s0, 30
-; NOOPT-NEXT: v_writelane_b32 v18, s1, 31
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
-; NOOPT-NEXT: buffer_store_dword v18, off, s[36:39], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
+; NOOPT-NEXT: v_writelane_b32 v5, s0, 5
+; NOOPT-NEXT: v_writelane_b32 v5, s1, 6
+; NOOPT-NEXT: s_or_saveexec_b64 s[22:23], -1
+; NOOPT-NEXT: buffer_store_dword v5, off, s[24:27], 0 ; 4-byte Folded Spill
+; NOOPT-NEXT: s_mov_b64 exec, s[22:23]
; NOOPT-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execz .LBB16_8
-; NOOPT-NEXT: ; %bb.7: ; %bb1
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
+; NOOPT-NEXT: s_cbranch_execz .LBB16_2
+; NOOPT-NEXT: ; %bb.1: ; %bb1
+; NOOPT-NEXT: s_or_saveexec_b64 s[22:23], -1
; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
+; NOOPT-NEXT: buffer_load_dword v5, off, s[24:27], 0 ; 4-byte Folded Reload
+; NOOPT-NEXT: s_mov_b64 exec, s[22:23]
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s4, v18, 25
+; NOOPT-NEXT: v_readlane_b32 s4, v5, 4
; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
; NOOPT-NEXT: s_mov_b32 s7, s1
; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
@@ -5339,14 +4801,14 @@ define amdgpu_kernel void @extract_vgpr_offset_multiple_in_block(ptr addrspace(1
; NOOPT-NEXT: v_mov_b32_e32 v0, s4
; NOOPT-NEXT: buffer_store_dword v0, off, s[0:3], 0
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: .LBB16_8: ; %bb2
-; NOOPT-NEXT: s_or_saveexec_b64 s[28:29], -1
+; NOOPT-NEXT: .LBB16_2: ; %bb2
+; NOOPT-NEXT: s_or_saveexec_b64 s[22:23], -1
; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v18, off, s[36:39], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[28:29]
+; NOOPT-NEXT: buffer_load_dword v5, off, s[24:27], 0 ; 4-byte Folded Reload
+; NOOPT-NEXT: s_mov_b64 exec, s[22:23]
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v18, 30
-; NOOPT-NEXT: v_readlane_b32 s1, v18, 31
+; NOOPT-NEXT: v_readlane_b32 s0, v5, 5
+; NOOPT-NEXT: v_readlane_b32 s1, v5, 6
; NOOPT-NEXT: s_or_b64 exec, exec, s[0:1]
; NOOPT-NEXT: s_endpgm
;
@@ -5767,362 +5229,185 @@ define amdgpu_kernel void @insert_vgpr_offset_multiple_in_block(ptr addrspace(1)
; NOOPT-NEXT: s_mov_b32 s31, 0xe8f000
; NOOPT-NEXT: s_add_u32 s28, s28, s11
; NOOPT-NEXT: s_addc_u32 s29, s29, 0
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:84 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_load_dwordx2 s[18:19], s[4:5], 0x9
-; NOOPT-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0xd
-; NOOPT-NEXT: s_load_dwordx16 s[0:15], s[4:5], 0x19
+; NOOPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; NOOPT-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd
+; NOOPT-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x19
+; NOOPT-NEXT: ; implicit-def: $vgpr20 : SGPR spill to VGPR lane
; NOOPT-NEXT: s_waitcnt lgkmcnt(0)
-; NOOPT-NEXT: s_mov_b32 s24, s19
-; NOOPT-NEXT: s_mov_b32 s20, s18
-; NOOPT-NEXT: s_mov_b32 s18, 0xf000
-; NOOPT-NEXT: s_mov_b32 s19, -1
-; NOOPT-NEXT: ; kill: def $sgpr20 killed $sgpr20 def $sgpr20_sgpr21_sgpr22_sgpr23
-; NOOPT-NEXT: s_mov_b32 s21, s24
-; NOOPT-NEXT: s_mov_b32 s22, s19
-; NOOPT-NEXT: s_mov_b32 s23, s18
-; NOOPT-NEXT: ; implicit-def: $vgpr32 : SGPR spill to VGPR lane
-; NOOPT-NEXT: v_writelane_b32 v32, s20, 0
-; NOOPT-NEXT: v_writelane_b32 v32, s21, 1
-; NOOPT-NEXT: v_writelane_b32 v32, s22, 2
-; NOOPT-NEXT: v_writelane_b32 v32, s23, 3
-; NOOPT-NEXT: s_mov_b32 s20, 0
-; NOOPT-NEXT: v_writelane_b32 v32, s20, 4
-; NOOPT-NEXT: ; kill: def $sgpr20 killed $sgpr20 def $sgpr20_sgpr21
-; NOOPT-NEXT: s_mov_b32 s21, s18
-; NOOPT-NEXT: ; kill: def $sgpr16_sgpr17 killed $sgpr16_sgpr17 def $sgpr16_sgpr17_sgpr18_sgpr19
-; NOOPT-NEXT: s_mov_b64 s[18:19], s[20:21]
-; NOOPT-NEXT: s_mov_b32 s20, 2
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_lshlrev_b32_e64 v0, s20, v0
-; NOOPT-NEXT: v_mov_b32_e32 v2, 0
-; NOOPT-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
-; NOOPT-NEXT: v_mov_b32_e32 v1, v2
-; NOOPT-NEXT: buffer_load_dword v0, v[0:1], s[16:19], 0 addr64 glc
+; NOOPT-NEXT: v_writelane_b32 v20, s36, 0
+; NOOPT-NEXT: v_writelane_b32 v20, s37, 1
+; NOOPT-NEXT: v_writelane_b32 v20, s38, 2
+; NOOPT-NEXT: v_writelane_b32 v20, s39, 3
+; NOOPT-NEXT: v_writelane_b32 v20, s40, 4
+; NOOPT-NEXT: v_writelane_b32 v20, s41, 5
+; NOOPT-NEXT: v_writelane_b32 v20, s42, 6
+; NOOPT-NEXT: v_writelane_b32 v20, s43, 7
+; NOOPT-NEXT: v_writelane_b32 v20, s44, 8
+; NOOPT-NEXT: v_writelane_b32 v20, s45, 9
+; NOOPT-NEXT: v_writelane_b32 v20, s46, 10
+; NOOPT-NEXT: v_writelane_b32 v20, s47, 11
+; NOOPT-NEXT: v_writelane_b32 v20, s48, 12
+; NOOPT-NEXT: v_writelane_b32 v20, s49, 13
+; NOOPT-NEXT: v_writelane_b32 v20, s50, 14
+; NOOPT-NEXT: v_writelane_b32 v20, s51, 15
+; NOOPT-NEXT: s_mov_b32 s2, s1
+; NOOPT-NEXT: s_mov_b32 s4, s0
+; NOOPT-NEXT: s_mov_b32 s1, 0xf000
+; NOOPT-NEXT: s_mov_b32 s0, -1
+; NOOPT-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
+; NOOPT-NEXT: s_mov_b32 s5, s2
+; NOOPT-NEXT: s_mov_b32 s6, s0
+; NOOPT-NEXT: s_mov_b32 s7, s1
+; NOOPT-NEXT: s_mov_b32 s0, 0
+; NOOPT-NEXT: s_mov_b32 s2, s0
+; NOOPT-NEXT: s_mov_b32 s3, s1
+; NOOPT-NEXT: ; kill: def $sgpr8_sgpr9 killed $sgpr8_sgpr9 def $sgpr8_sgpr9_sgpr10_sgpr11
+; NOOPT-NEXT: s_mov_b64 s[10:11], s[2:3]
+; NOOPT-NEXT: s_mov_b32 s19, 2
+; NOOPT-NEXT: v_lshlrev_b32_e64 v1, s19, v0
+; NOOPT-NEXT: v_mov_b32_e32 v3, 0
+; NOOPT-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
+; NOOPT-NEXT: v_mov_b32_e32 v2, v3
+; NOOPT-NEXT: buffer_load_dword v5, v[1:2], s[8:11], 0 addr64 glc
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:80 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b32 s16, 1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_add_i32_e64 v0, s[16:17], v0, s16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:76 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_waitcnt expcnt(0)
+; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:16 ; 4-byte Folded Spill
+; NOOPT-NEXT: s_mov_b32 s18, 1
+; NOOPT-NEXT: v_add_i32_e64 v17, s[2:3], v5, s18
+; NOOPT-NEXT: buffer_store_dword v17, off, s[28:31], 0 offset:12 ; 4-byte Folded Spill
; NOOPT-NEXT: ;;#ASMSTART
-; NOOPT-NEXT: v_mov_b32 v0, 62
+; NOOPT-NEXT: v_mov_b32 v18, 62
; NOOPT-NEXT: ;;#ASMEND
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:72 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:68 ; 4-byte Folded Spill
+; NOOPT-NEXT: buffer_store_dword v18, off, s[28:31], 0 offset:8 ; 4-byte Folded Spill
+; NOOPT-NEXT: v_mov_b32_e32 v1, v18
+; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:4 ; 4-byte Folded Spill
+; NOOPT-NEXT: s_mov_b32 s2, 12
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v5, s2
+; NOOPT-NEXT: s_mov_b32 s1, s48
; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v0, s0
; NOOPT-NEXT: v_mov_b32_e32 v1, s1
-; NOOPT-NEXT: v_mov_b32_e32 v2, s2
-; NOOPT-NEXT: v_mov_b32_e32 v3, s3
-; NOOPT-NEXT: v_mov_b32_e32 v4, s4
-; NOOPT-NEXT: v_mov_b32_e32 v5, s5
-; NOOPT-NEXT: v_mov_b32_e32 v6, s6
-; NOOPT-NEXT: v_mov_b32_e32 v7, s7
-; NOOPT-NEXT: v_mov_b32_e32 v8, s8
-; NOOPT-NEXT: v_mov_b32_e32 v9, s9
-; NOOPT-NEXT: v_mov_b32_e32 v10, s10
-; NOOPT-NEXT: v_mov_b32_e32 v11, s11
-; NOOPT-NEXT: v_mov_b32_e32 v12, s12
-; NOOPT-NEXT: v_mov_b32_e32 v13, s13
-; NOOPT-NEXT: v_mov_b32_e32 v14, s14
-; NOOPT-NEXT: v_mov_b32_e32 v15, s15
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v32, s0, 5
-; NOOPT-NEXT: v_writelane_b32 v32, s1, 6
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_store_dword v32, off, s[28:31], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[28:31], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[28:31], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[28:31], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[28:31], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[28:31], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[28:31], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[28:31], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[28:31], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[28:31], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[28:31], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[28:31], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[28:31], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[28:31], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v32, 7
-; NOOPT-NEXT: v_readlane_b32 s1, v32, 8
-; NOOPT-NEXT: buffer_load_dword v0, off, s[28:31], 0 offset:4 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[28:31], 0 offset:8 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[28:31], 0 offset:12 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[28:31], 0 offset:16 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[28:31], 0 offset:20 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[28:31], 0 offset:24 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[28:31], 0 offset:28 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[28:31], 0 offset:32 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[28:31], 0 offset:36 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(6)
-; NOOPT-NEXT: buffer_load_dword v9, off, s[28:31], 0 offset:40 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(5)
-; NOOPT-NEXT: buffer_load_dword v10, off, s[28:31], 0 offset:44 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(4)
-; NOOPT-NEXT: buffer_load_dword v11, off, s[28:31], 0 offset:48 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(3)
-; NOOPT-NEXT: buffer_load_dword v12, off, s[28:31], 0 offset:52 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(2)
-; NOOPT-NEXT: buffer_load_dword v13, off, s[28:31], 0 offset:56 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(1)
-; NOOPT-NEXT: buffer_load_dword v14, off, s[28:31], 0 offset:60 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v15, off, s[28:31], 0 offset:64 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[28:31], 0 offset:72 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[28:31], 0 offset:80 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v17
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v17
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_mov_b32 m0, s2
-; NOOPT-NEXT: v_movreld_b32_e32 v0, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:88 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:92 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[28:31], 0 offset:96 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[28:31], 0 offset:100 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[28:31], 0 offset:104 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:108 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[28:31], 0 offset:112 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[28:31], 0 offset:116 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[28:31], 0 offset:120 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[28:31], 0 offset:124 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[28:31], 0 offset:128 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[28:31], 0 offset:132 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[28:31], 0 offset:136 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[28:31], 0 offset:140 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[28:31], 0 offset:144 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[28:31], 0 offset:148 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:4 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:8 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[28:31], 0 offset:12 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[28:31], 0 offset:16 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[28:31], 0 offset:20 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:24 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[28:31], 0 offset:28 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[28:31], 0 offset:32 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[28:31], 0 offset:36 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[28:31], 0 offset:40 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[28:31], 0 offset:44 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[28:31], 0 offset:48 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[28:31], 0 offset:52 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[28:31], 0 offset:56 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[28:31], 0 offset:60 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[28:31], 0 offset:64 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v32, s2, 7
-; NOOPT-NEXT: v_writelane_b32 v32, s3, 8
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_store_dword v32, off, s[28:31], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB17_1
-; NOOPT-NEXT: ; %bb.2:
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v32, 5
-; NOOPT-NEXT: v_readlane_b32 s1, v32, 6
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.3:
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: buffer_load_dword v0, off, s[28:31], 0 offset:88 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[28:31], 0 offset:92 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[28:31], 0 offset:96 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[28:31], 0 offset:100 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[28:31], 0 offset:104 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[28:31], 0 offset:108 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[28:31], 0 offset:112 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[28:31], 0 offset:116 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[28:31], 0 offset:120 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v9, off, s[28:31], 0 offset:124 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v10, off, s[28:31], 0 offset:128 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v11, off, s[28:31], 0 offset:132 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v12, off, s[28:31], 0 offset:136 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v13, off, s[28:31], 0 offset:140 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v14, off, s[28:31], 0 offset:144 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v15, off, s[28:31], 0 offset:148 ; 4-byte Folded Reload
-; NOOPT-NEXT: v_mov_b32_e32 v16, 63
-; NOOPT-NEXT: buffer_store_dword v16, off, s[28:31], 0 offset:216 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: s_waitcnt vmcnt(14)
-; NOOPT-NEXT: v_writelane_b32 v32, s0, 9
-; NOOPT-NEXT: v_writelane_b32 v32, s1, 10
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_store_dword v32, off, s[28:31], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:152 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:156 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[28:31], 0 offset:160 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_waitcnt vmcnt(14)
-; NOOPT-NEXT: buffer_store_dword v3, off, s[28:31], 0 offset:164 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[28:31], 0 offset:168 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:172 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[28:31], 0 offset:176 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_waitcnt vmcnt(14)
-; NOOPT-NEXT: buffer_store_dword v7, off, s[28:31], 0 offset:180 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[28:31], 0 offset:184 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[28:31], 0 offset:188 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[28:31], 0 offset:192 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_waitcnt vmcnt(14)
-; NOOPT-NEXT: buffer_store_dword v11, off, s[28:31], 0 offset:196 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[28:31], 0 offset:200 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[28:31], 0 offset:204 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[28:31], 0 offset:208 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_waitcnt vmcnt(14)
-; NOOPT-NEXT: buffer_store_dword v15, off, s[28:31], 0 offset:212 ; 4-byte Folded Spill
-; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
-; NOOPT-NEXT: .LBB17_4: ; =>This Inner Loop Header: Depth=1
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v32, 11
-; NOOPT-NEXT: v_readlane_b32 s1, v32, 12
-; NOOPT-NEXT: buffer_load_dword v0, off, s[28:31], 0 offset:152 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v1, off, s[28:31], 0 offset:156 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v2, off, s[28:31], 0 offset:160 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v3, off, s[28:31], 0 offset:164 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v4, off, s[28:31], 0 offset:168 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v5, off, s[28:31], 0 offset:172 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v6, off, s[28:31], 0 offset:176 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v7, off, s[28:31], 0 offset:180 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v8, off, s[28:31], 0 offset:184 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(6)
-; NOOPT-NEXT: buffer_load_dword v9, off, s[28:31], 0 offset:188 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(5)
-; NOOPT-NEXT: buffer_load_dword v10, off, s[28:31], 0 offset:192 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(4)
-; NOOPT-NEXT: buffer_load_dword v11, off, s[28:31], 0 offset:196 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(3)
-; NOOPT-NEXT: buffer_load_dword v12, off, s[28:31], 0 offset:200 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(2)
-; NOOPT-NEXT: buffer_load_dword v13, off, s[28:31], 0 offset:204 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(1)
-; NOOPT-NEXT: buffer_load_dword v14, off, s[28:31], 0 offset:208 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v15, off, s[28:31], 0 offset:212 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[28:31], 0 offset:216 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[28:31], 0 offset:76 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readfirstlane_b32 s2, v17
-; NOOPT-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v17
-; NOOPT-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
-; NOOPT-NEXT: s_mov_b32 m0, s2
-; NOOPT-NEXT: v_movreld_b32_e32 v0, v16
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:220 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:224 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[28:31], 0 offset:228 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[28:31], 0 offset:232 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[28:31], 0 offset:236 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:240 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[28:31], 0 offset:244 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[28:31], 0 offset:248 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[28:31], 0 offset:252 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[28:31], 0 offset:256 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[28:31], 0 offset:260 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[28:31], 0 offset:264 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[28:31], 0 offset:268 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[28:31], 0 offset:272 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[28:31], 0 offset:276 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[28:31], 0 offset:280 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v0, off, s[28:31], 0 offset:152 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v1, off, s[28:31], 0 offset:156 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v2, off, s[28:31], 0 offset:160 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v3, off, s[28:31], 0 offset:164 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v4, off, s[28:31], 0 offset:168 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v5, off, s[28:31], 0 offset:172 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v6, off, s[28:31], 0 offset:176 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v7, off, s[28:31], 0 offset:180 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v8, off, s[28:31], 0 offset:184 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v9, off, s[28:31], 0 offset:188 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v10, off, s[28:31], 0 offset:192 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v11, off, s[28:31], 0 offset:196 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v12, off, s[28:31], 0 offset:200 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v13, off, s[28:31], 0 offset:204 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v14, off, s[28:31], 0 offset:208 ; 4-byte Folded Spill
-; NOOPT-NEXT: buffer_store_dword v15, off, s[28:31], 0 offset:212 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 s[2:3], s[0:1]
-; NOOPT-NEXT: v_writelane_b32 v32, s2, 11
-; NOOPT-NEXT: v_writelane_b32 v32, s3, 12
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_store_dword v32, off, s[28:31], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_xor_b64 exec, exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execnz .LBB17_4
-; NOOPT-NEXT: ; %bb.5:
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v32, 9
-; NOOPT-NEXT: v_readlane_b32 s1, v32, 10
-; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: ; %bb.6:
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v32, 4
-; NOOPT-NEXT: v_readlane_b32 s4, v32, 0
-; NOOPT-NEXT: v_readlane_b32 s5, v32, 1
-; NOOPT-NEXT: v_readlane_b32 s6, v32, 2
-; NOOPT-NEXT: v_readlane_b32 s7, v32, 3
-; NOOPT-NEXT: buffer_load_dword v0, off, s[28:31], 0 offset:84 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v16, off, s[28:31], 0 offset:220 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v17, off, s[28:31], 0 offset:224 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v18, off, s[28:31], 0 offset:228 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v19, off, s[28:31], 0 offset:232 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v20, off, s[28:31], 0 offset:236 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v21, off, s[28:31], 0 offset:240 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v22, off, s[28:31], 0 offset:244 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v23, off, s[28:31], 0 offset:248 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v24, off, s[28:31], 0 offset:252 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v25, off, s[28:31], 0 offset:256 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v26, off, s[28:31], 0 offset:260 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v27, off, s[28:31], 0 offset:264 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v28, off, s[28:31], 0 offset:268 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v29, off, s[28:31], 0 offset:272 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v30, off, s[28:31], 0 offset:276 ; 4-byte Folded Reload
-; NOOPT-NEXT: buffer_load_dword v31, off, s[28:31], 0 offset:280 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_waitcnt vmcnt(12)
-; NOOPT-NEXT: v_mov_b32_e32 v5, v19
-; NOOPT-NEXT: v_mov_b32_e32 v6, v18
-; NOOPT-NEXT: v_mov_b32_e32 v7, v17
-; NOOPT-NEXT: v_mov_b32_e32 v1, v16
-; NOOPT-NEXT: s_waitcnt vmcnt(8)
-; NOOPT-NEXT: v_mov_b32_e32 v2, v23
-; NOOPT-NEXT: v_mov_b32_e32 v3, v22
-; NOOPT-NEXT: v_mov_b32_e32 v4, v21
-; NOOPT-NEXT: v_mov_b32_e32 v8, v20
-; NOOPT-NEXT: s_waitcnt vmcnt(4)
-; NOOPT-NEXT: v_mov_b32_e32 v13, v27
-; NOOPT-NEXT: v_mov_b32_e32 v14, v26
-; NOOPT-NEXT: v_mov_b32_e32 v15, v25
-; NOOPT-NEXT: v_mov_b32_e32 v9, v24
-; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_mov_b32_e32 v10, v31
-; NOOPT-NEXT: v_mov_b32_e32 v11, v30
-; NOOPT-NEXT: v_mov_b32_e32 v12, v29
-; NOOPT-NEXT: v_mov_b32_e32 v16, v28
+; NOOPT-NEXT: v_cndmask_b32_e64 v16, v1, v18, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s3, 13
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v5, s3
+; NOOPT-NEXT: s_mov_b32 s1, s49
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v12, v1, v18, s[8:9]
+; NOOPT-NEXT: s_mov_b32 s8, 14
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v5, s8
+; NOOPT-NEXT: s_mov_b32 s1, s50
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v11, v1, v18, s[10:11]
+; NOOPT-NEXT: s_mov_b32 s9, 15
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v5, s9
+; NOOPT-NEXT: s_mov_b32 s1, s51
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v10, v1, v18, s[10:11]
+; NOOPT-NEXT: s_mov_b32 s10, 8
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[12:13], v5, s10
+; NOOPT-NEXT: s_mov_b32 s1, s44
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v9, v1, v18, s[12:13]
+; NOOPT-NEXT: s_mov_b32 s11, 9
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[12:13], v5, s11
+; NOOPT-NEXT: s_mov_b32 s1, s45
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v15, v1, v18, s[12:13]
+; NOOPT-NEXT: s_mov_b32 s12, 10
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[14:15], v5, s12
+; NOOPT-NEXT: s_mov_b32 s1, s46
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v14, v1, v18, s[14:15]
+; NOOPT-NEXT: s_mov_b32 s13, 11
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[14:15], v5, s13
+; NOOPT-NEXT: s_mov_b32 s1, s47
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v13, v1, v18, s[14:15]
+; NOOPT-NEXT: s_mov_b32 s14, 4
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[16:17], v5, s14
+; NOOPT-NEXT: s_mov_b32 s1, s40
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v8, v1, v18, s[16:17]
+; NOOPT-NEXT: s_mov_b32 s15, 5
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[16:17], v5, s15
+; NOOPT-NEXT: s_mov_b32 s1, s41
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v4, v1, v18, s[16:17]
+; NOOPT-NEXT: s_mov_b32 s16, 6
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v5, s16
+; NOOPT-NEXT: s_mov_b32 s1, s42
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v3, v1, v18, s[20:21]
+; NOOPT-NEXT: s_mov_b32 s17, 7
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v5, s17
+; NOOPT-NEXT: s_mov_b32 s1, s43
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v2, v1, v18, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v5, s0
+; NOOPT-NEXT: s_mov_b32 s1, s36
+; NOOPT-NEXT: v_mov_b32_e32 v1, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v18, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v5, s18
+; NOOPT-NEXT: s_mov_b32 s1, s37
+; NOOPT-NEXT: v_mov_b32_e32 v6, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v7, v6, v18, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v5, s19
+; NOOPT-NEXT: s_mov_b32 s1, s38
+; NOOPT-NEXT: v_mov_b32_e32 v6, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v6, v6, v18, s[20:21]
+; NOOPT-NEXT: s_mov_b32 s1, 3
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v5, s1
+; NOOPT-NEXT: s_mov_b32 s22, s39
+; NOOPT-NEXT: v_mov_b32_e32 v5, s22
+; NOOPT-NEXT: v_cndmask_b32_e64 v5, v5, v18, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v17, s1
+; NOOPT-NEXT: s_mov_b32 s1, 63
+; NOOPT-NEXT: v_writelane_b32 v20, s1, 16
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v5, v5, v18, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[20:21], v17, s19
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v6, v6, v18, s[20:21]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[18:19], v17, s18
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v7, v7, v18, s[18:19]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[18:19], v17, s0
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v1, v1, v18, s[18:19]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[18:19], v17, s17
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v2, v2, v18, s[18:19]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[16:17], v17, s16
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v3, v3, v18, s[16:17]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[16:17], v17, s15
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v4, v4, v18, s[16:17]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[14:15], v17, s14
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v8, v8, v18, s[14:15]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[14:15], v17, s13
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v13, v13, v18, s[14:15]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[12:13], v17, s12
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v14, v14, v18, s[12:13]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[12:13], v17, s11
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v15, v15, v18, s[12:13]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v17, s10
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v9, v9, v18, s[10:11]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[10:11], v17, s9
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v10, v10, v18, s[10:11]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v17, s8
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v11, v11, v18, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[8:9], v17, s3
+; NOOPT-NEXT: v_mov_b32_e32 v18, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v12, v12, v18, s[8:9]
+; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v17, s2
+; NOOPT-NEXT: v_mov_b32_e32 v17, s1
+; NOOPT-NEXT: v_cndmask_b32_e64 v16, v16, v17, s[2:3]
; NOOPT-NEXT: ; kill: def $vgpr16 killed $vgpr16 def $vgpr16_vgpr17_vgpr18_vgpr19 killed $exec
; NOOPT-NEXT: v_mov_b32_e32 v17, v12
; NOOPT-NEXT: v_mov_b32_e32 v18, v11
@@ -6149,16 +5434,16 @@ define amdgpu_kernel void @insert_vgpr_offset_multiple_in_block(ptr addrspace(1)
; NOOPT-NEXT: s_waitcnt vmcnt(0)
; NOOPT-NEXT: v_cmp_eq_u32_e64 s[2:3], v0, s0
; NOOPT-NEXT: s_mov_b64 s[0:1], exec
-; NOOPT-NEXT: v_writelane_b32 v32, s0, 13
-; NOOPT-NEXT: v_writelane_b32 v32, s1, 14
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
-; NOOPT-NEXT: buffer_store_dword v32, off, s[28:31], 0 ; 4-byte Folded Spill
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
+; NOOPT-NEXT: v_writelane_b32 v20, s0, 17
+; NOOPT-NEXT: v_writelane_b32 v20, s1, 18
+; NOOPT-NEXT: s_or_saveexec_b64 s[24:25], -1
+; NOOPT-NEXT: buffer_store_dword v20, off, s[28:31], 0 ; 4-byte Folded Spill
+; NOOPT-NEXT: s_mov_b64 exec, s[24:25]
; NOOPT-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; NOOPT-NEXT: s_mov_b64 exec, s[0:1]
-; NOOPT-NEXT: s_cbranch_execz .LBB17_8
-; NOOPT-NEXT: ; %bb.7: ; %bb1
-; NOOPT-NEXT: buffer_load_dword v0, off, s[28:31], 0 offset:68 ; 4-byte Folded Reload
+; NOOPT-NEXT: s_cbranch_execz .LBB17_2
+; NOOPT-NEXT: ; %bb.1: ; %bb1
+; NOOPT-NEXT: buffer_load_dword v0, off, s[28:31], 0 offset:4 ; 4-byte Folded Reload
; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
; NOOPT-NEXT: s_mov_b32 s6, s1
; NOOPT-NEXT: ; implicit-def: $sgpr0_sgpr1
@@ -6172,14 +5457,14 @@ define amdgpu_kernel void @insert_vgpr_offset_multiple_in_block(ptr addrspace(1)
; NOOPT-NEXT: s_waitcnt vmcnt(0)
; NOOPT-NEXT: buffer_store_dword v0, off, s[0:3], 0
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: .LBB17_8: ; %bb2
-; NOOPT-NEXT: s_or_saveexec_b64 s[26:27], -1
+; NOOPT-NEXT: .LBB17_2: ; %bb2
+; NOOPT-NEXT: s_or_saveexec_b64 s[24:25], -1
; NOOPT-NEXT: s_waitcnt expcnt(0)
-; NOOPT-NEXT: buffer_load_dword v32, off, s[28:31], 0 ; 4-byte Folded Reload
-; NOOPT-NEXT: s_mov_b64 exec, s[26:27]
+; NOOPT-NEXT: buffer_load_dword v20, off, s[28:31], 0 ; 4-byte Folded Reload
+; NOOPT-NEXT: s_mov_b64 exec, s[24:25]
; NOOPT-NEXT: s_waitcnt vmcnt(0)
-; NOOPT-NEXT: v_readlane_b32 s0, v32, 13
-; NOOPT-NEXT: v_readlane_b32 s1, v32, 14
+; NOOPT-NEXT: v_readlane_b32 s0, v20, 17
+; NOOPT-NEXT: v_readlane_b32 s1, v20, 18
; NOOPT-NEXT: s_or_b64 exec, exec, s[0:1]
; NOOPT-NEXT: s_endpgm
;
diff --git a/llvm/test/CodeGen/AMDGPU/indirect-addressing-term.ll b/llvm/test/CodeGen/AMDGPU/indirect-addressing-term.ll
index 98658ded03903..1da83f8b231ac 100644
--- a/llvm/test/CodeGen/AMDGPU/indirect-addressing-term.ll
+++ b/llvm/test/CodeGen/AMDGPU/indirect-addressing-term.ll
@@ -10,7 +10,6 @@ declare i32 @llvm.amdgcn.workitem.id.x() #1
define amdgpu_kernel void @extract_w_offset_vgpr(ptr addrspace(1) %out) {
; GCN-LABEL: name: extract_w_offset_vgpr
; GCN: bb.0.entry:
- ; GCN-NEXT: successors: %bb.1(0x80000000)
; GCN-NEXT: liveins: $vgpr0, $sgpr4_sgpr5
; GCN-NEXT: {{ $}}
; GCN-NEXT: [[COPY:%[0-9]+]]:vgpr_32(s32) = COPY killed $vgpr0
@@ -23,84 +22,71 @@ define amdgpu_kernel void @extract_w_offset_vgpr(ptr addrspace(1) %out) {
; GCN-NEXT: renamable $sgpr1 = COPY killed renamable $sgpr6
; GCN-NEXT: renamable $sgpr2 = COPY killed renamable $sgpr5
; GCN-NEXT: renamable $sgpr3 = COPY killed renamable $sgpr4
- ; GCN-NEXT: SI_SPILL_S128_SAVE killed $sgpr0_sgpr1_sgpr2_sgpr3, %stack.1, implicit $exec, implicit $sgpr32 :: (store (s128) into %stack.1, align 4, addrspace 5)
- ; GCN-NEXT: renamable $sgpr0 = S_MOV_B32 16
- ; GCN-NEXT: renamable $sgpr1 = S_MOV_B32 15
- ; GCN-NEXT: renamable $sgpr2 = S_MOV_B32 14
- ; GCN-NEXT: renamable $sgpr3 = S_MOV_B32 13
+ ; GCN-NEXT: renamable $sgpr6 = S_MOV_B32 1
+ ; GCN-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]](s32), $sgpr6, 0, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4_sgpr5 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], $sgpr6, implicit $exec
+ ; GCN-NEXT: renamable $sgpr7 = S_MOV_B32 0
+ ; GCN-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr7
+ ; GCN-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr6
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[COPY1]], 0, [[COPY2]], killed $sgpr4_sgpr5, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6 = S_MOV_B32 2
+ ; GCN-NEXT: renamable $sgpr4_sgpr5 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], $sgpr6, implicit $exec
+ ; GCN-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr6
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_1:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_]], 0, [[COPY3]], killed $sgpr4_sgpr5, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6 = S_MOV_B32 3
+ ; GCN-NEXT: renamable $sgpr4_sgpr5 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], $sgpr6, implicit $exec
+ ; GCN-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr6
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_2:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_1]], 0, [[COPY4]], killed $sgpr4_sgpr5, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 4
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 5
+ ; GCN-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_3:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_2]], 0, [[COPY5]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 6
+ ; GCN-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_4:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_3]], 0, [[COPY6]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 7
+ ; GCN-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_5:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_4]], 0, [[COPY7]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 8
+ ; GCN-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_6:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_5]], 0, [[COPY8]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 9
+ ; GCN-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_7:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_6]], 0, [[COPY9]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 10
+ ; GCN-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_8:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_7]], 0, [[COPY10]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 11
+ ; GCN-NEXT: [[COPY11:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_9:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_8]], 0, [[COPY11]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 12
- ; GCN-NEXT: renamable $sgpr5 = S_MOV_B32 11
- ; GCN-NEXT: renamable $sgpr6 = S_MOV_B32 10
- ; GCN-NEXT: renamable $sgpr7 = S_MOV_B32 9
- ; GCN-NEXT: renamable $sgpr8 = S_MOV_B32 8
- ; GCN-NEXT: renamable $sgpr9 = S_MOV_B32 7
- ; GCN-NEXT: renamable $sgpr10 = S_MOV_B32 6
- ; GCN-NEXT: renamable $sgpr11 = S_MOV_B32 5
- ; GCN-NEXT: renamable $sgpr12 = S_MOV_B32 3
- ; GCN-NEXT: renamable $sgpr13 = S_MOV_B32 2
- ; GCN-NEXT: renamable $sgpr14 = S_MOV_B32 1
- ; GCN-NEXT: renamable $sgpr15 = S_MOV_B32 0
- ; GCN-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr15
- ; GCN-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr14
- ; GCN-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr13
- ; GCN-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr12
- ; GCN-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr11
- ; GCN-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr10
- ; GCN-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr9
- ; GCN-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr8
- ; GCN-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr7
- ; GCN-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr6
- ; GCN-NEXT: [[COPY11:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr5
- ; GCN-NEXT: [[COPY12:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr4
- ; GCN-NEXT: [[COPY13:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr3
- ; GCN-NEXT: [[COPY14:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr2
- ; GCN-NEXT: [[COPY15:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr1
- ; GCN-NEXT: [[COPY16:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr0
- ; GCN-NEXT: undef [[COPY17:%[0-9]+]].sub0:vreg_512 = COPY [[COPY1]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub1:vreg_512 = COPY [[COPY2]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub2:vreg_512 = COPY [[COPY3]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub3:vreg_512 = COPY [[COPY4]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub4:vreg_512 = COPY [[COPY5]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub5:vreg_512 = COPY [[COPY6]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub6:vreg_512 = COPY [[COPY7]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub7:vreg_512 = COPY [[COPY8]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub8:vreg_512 = COPY [[COPY9]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub9:vreg_512 = COPY [[COPY10]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub10:vreg_512 = COPY [[COPY11]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub11:vreg_512 = COPY [[COPY12]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub12:vreg_512 = COPY [[COPY13]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub13:vreg_512 = COPY [[COPY14]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub14:vreg_512 = COPY [[COPY15]]
- ; GCN-NEXT: [[COPY17:%[0-9]+]].sub15:vreg_512 = COPY [[COPY16]]
- ; GCN-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 $exec
- ; GCN-NEXT: SI_SPILL_S64_SAVE killed $sgpr0_sgpr1, %stack.0, implicit $exec, implicit $sgpr32 :: (store (s64) into %stack.0, align 4, addrspace 5)
- ; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; GCN-NEXT: renamable $sgpr0_sgpr1 = IMPLICIT_DEF
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.1:
- ; GCN-NEXT: successors: %bb.1(0x40000000), %bb.3(0x40000000)
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr0_sgpr1 = SI_SPILL_S64_RESTORE %stack.2, implicit $exec, implicit $sgpr32 :: (load (s64) from %stack.2, align 4, addrspace 5)
- ; GCN-NEXT: dead [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[DEF]]
- ; GCN-NEXT: renamable $sgpr2 = V_READFIRSTLANE_B32 [[COPY]](s32), implicit $exec
- ; GCN-NEXT: renamable $sgpr0_sgpr1 = V_CMP_EQ_U32_e64 $sgpr2, [[COPY]](s32), implicit $exec
- ; GCN-NEXT: renamable $sgpr0_sgpr1 = S_AND_SAVEEXEC_B64 killed renamable $sgpr0_sgpr1, implicit-def $exec, implicit-def dead $scc, implicit $exec
- ; GCN-NEXT: [[V_INDIRECT_REG_READ_GPR_IDX_B32_V16_:%[0-9]+]]:vgpr_32 = V_INDIRECT_REG_READ_GPR_IDX_B32_V16 [[COPY17]], killed $sgpr2, 11, implicit-def $m0, implicit $m0, implicit $exec
- ; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = COPY [[V_INDIRECT_REG_READ_GPR_IDX_B32_V16_]]
- ; GCN-NEXT: renamable $sgpr2_sgpr3 = COPY renamable $sgpr0_sgpr1
- ; GCN-NEXT: SI_SPILL_S64_SAVE killed $sgpr2_sgpr3, %stack.2, implicit $exec, implicit $sgpr32 :: (store (s64) into %stack.2, align 4, addrspace 5)
- ; GCN-NEXT: $exec = S_XOR_B64_term $exec, killed renamable $sgpr0_sgpr1, implicit-def dead $scc
- ; GCN-NEXT: S_CBRANCH_EXECNZ %bb.1, implicit $exec
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.3:
- ; GCN-NEXT: successors: %bb.2(0x80000000)
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr0_sgpr1 = SI_SPILL_S64_RESTORE %stack.0, implicit $exec, implicit $sgpr32 :: (load (s64) from %stack.0, align 4, addrspace 5)
- ; GCN-NEXT: $exec = S_MOV_B64 renamable $sgpr0_sgpr1
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.2:
- ; GCN-NEXT: $sgpr0_sgpr1_sgpr2_sgpr3 = SI_SPILL_S128_RESTORE %stack.1, implicit $exec, implicit $sgpr32 :: (load (s128) from %stack.1, align 4, addrspace 5)
- ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET [[V_INDIRECT_REG_READ_GPR_IDX_B32_V16_]], killed renamable $sgpr0_sgpr1_sgpr2_sgpr3, 0, 0, 0, 0, implicit $exec :: (store (s32) into %ir.out.load, addrspace 1)
+ ; GCN-NEXT: [[COPY12:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_10:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_9]], 0, [[COPY12]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 13
+ ; GCN-NEXT: [[COPY13:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_11:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_10]], 0, [[COPY13]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 14
+ ; GCN-NEXT: [[COPY14:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_12:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_11]], 0, [[COPY14]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6_sgpr7 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4 = S_MOV_B32 15
+ ; GCN-NEXT: [[COPY15:%[0-9]+]]:vgpr_32 = COPY renamable $sgpr4
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_13:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_12]], 0, [[COPY15]], killed $sgpr6_sgpr7, implicit $exec
+ ; GCN-NEXT: renamable $sgpr4_sgpr5 = V_CMP_EQ_U32_e64 [[V_ADD_U32_e64_]], killed $sgpr4, implicit $exec
+ ; GCN-NEXT: renamable $sgpr6 = S_MOV_B32 16
+ ; GCN-NEXT: [[COPY16:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr6
+ ; GCN-NEXT: [[V_CNDMASK_B32_e64_14:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, [[V_CNDMASK_B32_e64_13]], 0, [[COPY16]], killed $sgpr4_sgpr5, implicit $exec
+ ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET [[V_CNDMASK_B32_e64_14]], killed renamable $sgpr0_sgpr1_sgpr2_sgpr3, 0, 0, 0, 0, implicit $exec :: (store (s32) into %ir.out.load, addrspace 1)
; GCN-NEXT: S_ENDPGM 0
entry:
%id = call i32 @llvm.amdgcn.workitem.id.x() #1
diff --git a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
index 6ad62a675944b..92c97fa5d8b1e 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
@@ -1,7 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -enable-var-scope -check-prefixes=GCN %s
-; RUN: not --crash llc -O0 -mtriple=amdgcn -mcpu=fiji < %s -o /dev/null 2>&1 | FileCheck --check-prefixes=GCN-O0 %s
-; GCN-O0: LLVM ERROR: Cannot select: t{{[0-9]+}}: {{.*}} = insert_vector_elt
+; RUN: llc -O0 -mtriple=amdgcn -mcpu=fiji < %s | FileCheck --check-prefixes=GCN-O0 %s
define amdgpu_kernel void @float4_inselt(ptr addrspace(1) %out, <4 x float> %vec, i32 %sel) {
; GCN-LABEL: float4_inselt:
@@ -30,6 +29,50 @@ define amdgpu_kernel void @float4_inselt(ptr addrspace(1) %out, <4 x float> %vec
; GCN-NEXT: v_mov_b32_e32 v5, s5
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float4_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x34
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x44
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s3, s11
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s4
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s4, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[6:7]
+; GCN-O0-NEXT: s_mov_b32 s3, s10
+; GCN-O0-NEXT: s_mov_b32 s5, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s5
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[6:7]
+; GCN-O0-NEXT: s_mov_b32 s3, s9
+; GCN-O0-NEXT: s_mov_b32 s5, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s5
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v6, v2, v3, s[6:7]
+; GCN-O0-NEXT: s_mov_b32 s5, s8
+; GCN-O0-NEXT: s_mov_b32 s3, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[2:3], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[2:3]
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <4 x float> %vec, float 1.000000e+00, i32 %sel
store <4 x float> %v, ptr addrspace(1) %out
@@ -49,6 +92,24 @@ define amdgpu_kernel void @float4_inselt_undef(ptr addrspace(1) %out, i32 %sel)
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float4_inselt_undef:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_mov_b32 s2, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <4 x float> poison, float 1.000000e+00, i32 %sel
store <4 x float> %v, ptr addrspace(1) %out
@@ -78,6 +139,41 @@ define amdgpu_kernel void @int4_inselt(ptr addrspace(1) %out, <4 x i32> %vec, i3
; GCN-NEXT: v_mov_b32_e32 v5, s5
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: int4_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x34
+; GCN-O0-NEXT: s_load_dword s7, s[4:5], 0x44
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s2, s11
+; GCN-O0-NEXT: s_mov_b32 s3, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s7, s3
+; GCN-O0-NEXT: s_mov_b32 s5, 1
+; GCN-O0-NEXT: s_cselect_b32 s2, s5, s2
+; GCN-O0-NEXT: s_mov_b32 s3, s10
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s7, s4
+; GCN-O0-NEXT: s_cselect_b32 s3, s5, s3
+; GCN-O0-NEXT: s_mov_b32 s4, s9
+; GCN-O0-NEXT: s_cmp_eq_u32 s7, s5
+; GCN-O0-NEXT: s_cselect_b32 s4, s5, s4
+; GCN-O0-NEXT: s_mov_b32 s6, s8
+; GCN-O0-NEXT: s_mov_b32 s8, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s7, s8
+; GCN-O0-NEXT: s_cselect_b32 s5, s5, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <4 x i32> %vec, i32 1, i32 %sel
store <4 x i32> %v, ptr addrspace(1) %out
@@ -102,6 +198,34 @@ define amdgpu_kernel void @float2_inselt(ptr addrspace(1) %out, <2 x float> %vec
; GCN-NEXT: v_mov_b32_e32 v3, s5
; GCN-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float2_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x2c
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x34
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s3, s7
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s4
+; GCN-O0-NEXT: s_cselect_b64 s[8:9], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s4, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[8:9]
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s3, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s2, s3
+; GCN-O0-NEXT: s_cselect_b64 s[2:3], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[2:3]
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <2 x float> %vec, float 1.000000e+00, i32 %sel
store <2 x float> %v, ptr addrspace(1) %out
@@ -135,6 +259,57 @@ define amdgpu_kernel void @float8_inselt(ptr addrspace(1) %out, <8 x float> %vec
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float8_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x24
+; GCN-O0-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x44
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0x64
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1.0
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s4
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v11
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <8 x float> %vec, float 1.000000e+00, i32 %sel
store <8 x float> %v, ptr addrspace(1) %out
@@ -188,6 +363,105 @@ define amdgpu_kernel void @float16_inselt(ptr addrspace(1) %out, <16 x float> %v
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float16_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x24
+; GCN-O0-NEXT: s_load_dwordx16 s[4:19], s[2:3], 0x64
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0xa4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1.0
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s4
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v19
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 48
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v15
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 32
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v11
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <16 x float> %vec, float 1.000000e+00, i32 %sel
store <16 x float> %v, ptr addrspace(1) %out
@@ -282,6 +556,267 @@ define amdgpu_kernel void @float32_inselt(ptr addrspace(1) %out, <32 x float> %v
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: float32_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0xe4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s2, s51
+; GCN-O0-NEXT: s_mov_b32 s3, s50
+; GCN-O0-NEXT: s_mov_b32 s6, s49
+; GCN-O0-NEXT: s_mov_b32 s7, s48
+; GCN-O0-NEXT: s_mov_b32 s8, s47
+; GCN-O0-NEXT: s_mov_b32 s9, s46
+; GCN-O0-NEXT: s_mov_b32 s10, s45
+; GCN-O0-NEXT: s_mov_b32 s11, s44
+; GCN-O0-NEXT: s_mov_b32 s12, s43
+; GCN-O0-NEXT: s_mov_b32 s13, s42
+; GCN-O0-NEXT: s_mov_b32 s14, s41
+; GCN-O0-NEXT: s_mov_b32 s15, s40
+; GCN-O0-NEXT: s_mov_b32 s16, s39
+; GCN-O0-NEXT: s_mov_b32 s17, s38
+; GCN-O0-NEXT: s_mov_b32 s18, s37
+; GCN-O0-NEXT: s_mov_b32 s19, s36
+; GCN-O0-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0xa4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s20, s51
+; GCN-O0-NEXT: s_mov_b32 s21, s50
+; GCN-O0-NEXT: s_mov_b32 s22, s49
+; GCN-O0-NEXT: s_mov_b32 s23, s48
+; GCN-O0-NEXT: s_mov_b32 s24, s47
+; GCN-O0-NEXT: s_mov_b32 s25, s46
+; GCN-O0-NEXT: s_mov_b32 s26, s45
+; GCN-O0-NEXT: s_mov_b32 s27, s44
+; GCN-O0-NEXT: s_mov_b32 s28, s43
+; GCN-O0-NEXT: s_mov_b32 s29, s42
+; GCN-O0-NEXT: s_mov_b32 s30, s41
+; GCN-O0-NEXT: s_mov_b32 s31, s40
+; GCN-O0-NEXT: s_mov_b32 s33, s39
+; GCN-O0-NEXT: s_mov_b32 s34, s38
+; GCN-O0-NEXT: s_mov_b32 s35, s37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v62, s35
+; GCN-O0-NEXT: v_mov_b32_e32 v61, s34
+; GCN-O0-NEXT: v_mov_b32_e32 v60, s33
+; GCN-O0-NEXT: v_mov_b32_e32 v59, s31
+; GCN-O0-NEXT: v_mov_b32_e32 v58, s30
+; GCN-O0-NEXT: v_mov_b32_e32 v57, s29
+; GCN-O0-NEXT: v_mov_b32_e32 v56, s28
+; GCN-O0-NEXT: v_mov_b32_e32 v55, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v54, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v53, s25
+; GCN-O0-NEXT: v_mov_b32_e32 v52, s24
+; GCN-O0-NEXT: v_mov_b32_e32 v51, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v50, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s21
+; GCN-O0-NEXT: v_mov_b32_e32 v48, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v47, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v46, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v45, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v44, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v43, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v42, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v40, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v39, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v62
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v61
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v60
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v59
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v58
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v57
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v56
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v55
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v54
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v53
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v52
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v51
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v50
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v49
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v48
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v47
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v42
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v40
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v39
+; GCN-O0-NEXT: v_mov_b32_e32 v32, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v33, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v34, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v35, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v36, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v37, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v38, v0
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x124
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1.0
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v38
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v37
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v36
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v35
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x70
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v34
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v33
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v32
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v31
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x60
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v27
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x50
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v23
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 64
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v19
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 48
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v15
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 32
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v11
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <32 x float> %vec, float 1.000000e+00, i32 %sel
store <32 x float> %v, ptr addrspace(1) %out
@@ -307,6 +842,30 @@ define amdgpu_kernel void @half4_inselt(ptr addrspace(1) %out, <4 x half> %vec,
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: half4_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[0:1], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2c
+; GCN-O0-NEXT: s_load_dword s6, s[0:1], 0x34
+; GCN-O0-NEXT: s_mov_b32 s7, 0x3c003c00
+; GCN-O0-NEXT: s_mov_b32 s0, s7
+; GCN-O0-NEXT: s_mov_b32 s1, s7
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_xor_b64 s[0:1], s[4:5], s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s7, 4
+; GCN-O0-NEXT: s_lshl_b32 s8, s6, s7
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0xffff
+; GCN-O0-NEXT: s_lshl_b64 s[6:7], s[6:7], s8
+; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[6:7]
+; GCN-O0-NEXT: s_xor_b64 s[0:1], s[0:1], s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <4 x half> %vec, half 1.000000e+00, i32 %sel
store <4 x half> %v, ptr addrspace(1) %out
@@ -328,6 +887,26 @@ define amdgpu_kernel void @half2_inselt(ptr addrspace(1) %out, <2 x half> %vec,
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: half2_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s1, s[4:5], 0x2c
+; GCN-O0-NEXT: s_load_dword s4, s[4:5], 0x30
+; GCN-O0-NEXT: s_mov_b32 s0, 0x3c003c00
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_xor_b32 s0, s1, s0
+; GCN-O0-NEXT: s_mov_b32 s5, 4
+; GCN-O0-NEXT: s_lshl_b32 s5, s4, s5
+; GCN-O0-NEXT: s_mov_b32 s4, 0xffff
+; GCN-O0-NEXT: s_lshl_b32 s4, s4, s5
+; GCN-O0-NEXT: s_and_b32 s0, s0, s4
+; GCN-O0-NEXT: s_xor_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <2 x half> %vec, half 1.000000e+00, i32 %sel
store <2 x half> %v, ptr addrspace(1) %out
@@ -389,6 +968,106 @@ define amdgpu_kernel void @half8_inselt(ptr addrspace(1) %out, <8 x half> %vec,
; GCN-NEXT: v_mov_b32_e32 v5, s5
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: half8_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x34
+; GCN-O0-NEXT: s_load_dword s4, s[4:5], 0x44
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s2, s15
+; GCN-O0-NEXT: s_mov_b32 s5, s14
+; GCN-O0-NEXT: s_mov_b32 s10, s5
+; GCN-O0-NEXT: s_mov_b32 s11, s2
+; GCN-O0-NEXT: s_mov_b32 s2, s10
+; GCN-O0-NEXT: s_mov_b32 s3, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s3
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: s_mov_b32 s3, 0x3c00
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[6:7]
+; GCN-O0-NEXT: s_mov_b32 s8, 0xffff
+; GCN-O0-NEXT: v_and_b32_e64 v0, s8, v0
+; GCN-O0-NEXT: s_mov_b32 s2, 16
+; GCN-O0-NEXT: s_lshr_b32 s5, s5, s2
+; GCN-O0-NEXT: s_mov_b32 s6, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s6
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[6:7]
+; GCN-O0-NEXT: v_lshlrev_b32_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v1, v0, v1
+; GCN-O0-NEXT: s_mov_b32 s9, s13
+; GCN-O0-NEXT: s_mov_b32 s5, s12
+; GCN-O0-NEXT: s_mov_b32 s6, s5
+; GCN-O0-NEXT: s_mov_b32 s7, s9
+; GCN-O0-NEXT: s_mov_b32 s9, s6
+; GCN-O0-NEXT: s_mov_b32 s12, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s12
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v2, s[12:13]
+; GCN-O0-NEXT: v_and_b32_e64 v0, s8, v0
+; GCN-O0-NEXT: s_lshr_b32 s5, s5, s2
+; GCN-O0-NEXT: s_mov_b32 s9, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s9
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[12:13]
+; GCN-O0-NEXT: v_lshlrev_b32_e64 v2, s2, v2
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v2
+; GCN-O0-NEXT: s_mov_b32 s9, 32
+; GCN-O0-NEXT: s_lshr_b64 s[12:13], s[10:11], s9
+; GCN-O0-NEXT: s_mov_b32 s5, s12
+; GCN-O0-NEXT: s_mov_b32 s12, 6
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s12
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v3, s[12:13]
+; GCN-O0-NEXT: v_and_b32_e64 v0, s8, v0
+; GCN-O0-NEXT: s_mov_b32 s5, 48
+; GCN-O0-NEXT: s_lshr_b64 s[10:11], s[10:11], s5
+; GCN-O0-NEXT: s_mov_b32 s12, s10
+; GCN-O0-NEXT: s_mov_b32 s10, 7
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s10
+; GCN-O0-NEXT: s_cselect_b64 s[10:11], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v4, s[10:11]
+; GCN-O0-NEXT: v_lshlrev_b32_e64 v3, s2, v3
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: s_lshr_b64 s[10:11], s[6:7], s9
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s10, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s10
+; GCN-O0-NEXT: s_cselect_b64 s[10:11], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v4, s[10:11]
+; GCN-O0-NEXT: v_and_b32_e64 v3, s8, v3
+; GCN-O0-NEXT: s_lshr_b64 s[6:7], s[6:7], s5
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 killed $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s5, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s4, s5
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s3
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v5, s[4:5]
+; GCN-O0-NEXT: v_lshlrev_b32_e64 v4, s2, v4
+; GCN-O0-NEXT: v_or_b32_e64 v6, v3, v4
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <8 x half> %vec, half 1.000000e+00, i32 %sel
store <8 x half> %v, ptr addrspace(1) %out
@@ -410,6 +1089,26 @@ define amdgpu_kernel void @short2_inselt(ptr addrspace(1) %out, <2 x i16> %vec,
; GCN-NEXT: v_mov_b32_e32 v2, s2
; GCN-NEXT: flat_store_dword v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: short2_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s1, s[4:5], 0x2c
+; GCN-O0-NEXT: s_load_dword s4, s[4:5], 0x30
+; GCN-O0-NEXT: s_mov_b32 s0, 0x10001
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_xor_b32 s0, s1, s0
+; GCN-O0-NEXT: s_mov_b32 s5, 4
+; GCN-O0-NEXT: s_lshl_b32 s5, s4, s5
+; GCN-O0-NEXT: s_mov_b32 s4, 0xffff
+; GCN-O0-NEXT: s_lshl_b32 s4, s4, s5
+; GCN-O0-NEXT: s_and_b32 s0, s0, s4
+; GCN-O0-NEXT: s_xor_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dword v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <2 x i16> %vec, i16 1, i32 %sel
store <2 x i16> %v, ptr addrspace(1) %out
@@ -435,6 +1134,30 @@ define amdgpu_kernel void @short4_inselt(ptr addrspace(1) %out, <4 x i16> %vec,
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: short4_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[0:1], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2c
+; GCN-O0-NEXT: s_load_dword s6, s[0:1], 0x34
+; GCN-O0-NEXT: s_mov_b32 s7, 0x10001
+; GCN-O0-NEXT: s_mov_b32 s0, s7
+; GCN-O0-NEXT: s_mov_b32 s1, s7
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_xor_b64 s[0:1], s[4:5], s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s7, 4
+; GCN-O0-NEXT: s_lshl_b32 s8, s6, s7
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0xffff
+; GCN-O0-NEXT: s_lshl_b64 s[6:7], s[6:7], s8
+; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[6:7]
+; GCN-O0-NEXT: s_xor_b64 s[0:1], s[0:1], s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <4 x i16> %vec, i16 1, i32 %sel
store <4 x i16> %v, ptr addrspace(1) %out
@@ -459,6 +1182,140 @@ define amdgpu_kernel void @byte8_inselt(ptr addrspace(1) %out, <8 x i8> %vec, i3
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: byte8_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[0:1], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
+; GCN-O0-NEXT: s_load_dword s6, s[0:1], 0x34
+; GCN-O0-NEXT: s_mov_b32 s7, 0x1010101
+; GCN-O0-NEXT: s_mov_b32 s0, s7
+; GCN-O0-NEXT: s_mov_b32 s1, s7
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_xor_b64 s[0:1], s[2:3], s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s7, 3
+; GCN-O0-NEXT: s_lshl_b32 s8, s6, s7
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0xff
+; GCN-O0-NEXT: s_lshl_b64 s[6:7], s[6:7], s8
+; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[6:7]
+; GCN-O0-NEXT: s_xor_b64 s[10:11], s[0:1], s[2:3]
+; GCN-O0-NEXT: s_mov_b32 s3, s10
+; GCN-O0-NEXT: s_mov_b32 s0, 8
+; GCN-O0-NEXT: s_lshr_b32 s0, s3, s0
+; GCN-O0-NEXT: s_mov_b32 s1, s10
+; GCN-O0-NEXT: s_mov_b32 s2, 16
+; GCN-O0-NEXT: s_lshr_b32 s2, s3, s2
+; GCN-O0-NEXT: s_mov_b32 s6, 24
+; GCN-O0-NEXT: s_lshr_b32 s3, s3, s6
+; GCN-O0-NEXT: s_mov_b32 s6, 32
+; GCN-O0-NEXT: s_lshr_b64 s[6:7], s[10:11], s6
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 killed $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, 40
+; GCN-O0-NEXT: s_lshr_b64 s[8:9], s[10:11], s7
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: s_mov_b32 s8, 48
+; GCN-O0-NEXT: s_lshr_b64 s[8:9], s[10:11], s8
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s9, 56
+; GCN-O0-NEXT: s_lshr_b64 s[10:11], s[10:11], s9
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b64 s[14:15], 7
+; GCN-O0-NEXT: s_mov_b32 s10, s4
+; GCN-O0-NEXT: s_mov_b32 s11, s5
+; GCN-O0-NEXT: s_mov_b32 s13, s14
+; GCN-O0-NEXT: s_mov_b32 s12, s15
+; GCN-O0-NEXT: s_add_u32 s10, s10, s13
+; GCN-O0-NEXT: s_addc_u32 s12, s11, s12
+; GCN-O0-NEXT: ; kill: def $sgpr10 killed $sgpr10 def $sgpr10_sgpr11
+; GCN-O0-NEXT: s_mov_b32 s11, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s9
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_mov_b64 s[14:15], 6
+; GCN-O0-NEXT: s_mov_b32 s10, s4
+; GCN-O0-NEXT: s_mov_b32 s9, s5
+; GCN-O0-NEXT: s_mov_b32 s12, s14
+; GCN-O0-NEXT: s_mov_b32 s11, s15
+; GCN-O0-NEXT: s_add_u32 s10, s10, s12
+; GCN-O0-NEXT: s_addc_u32 s9, s9, s11
+; GCN-O0-NEXT: ; kill: def $sgpr10 killed $sgpr10 def $sgpr10_sgpr11
+; GCN-O0-NEXT: s_mov_b32 s11, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s8
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 5
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s9, s5
+; GCN-O0-NEXT: s_mov_b32 s11, s12
+; GCN-O0-NEXT: s_mov_b32 s10, s13
+; GCN-O0-NEXT: s_add_u32 s8, s8, s11
+; GCN-O0-NEXT: s_addc_u32 s10, s9, s10
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s7
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 4
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s7, s5
+; GCN-O0-NEXT: s_mov_b32 s10, s12
+; GCN-O0-NEXT: s_mov_b32 s9, s13
+; GCN-O0-NEXT: s_add_u32 s8, s8, s10
+; GCN-O0-NEXT: s_addc_u32 s7, s7, s9
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s9, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_mov_b64 s[10:11], 3
+; GCN-O0-NEXT: s_mov_b32 s6, s4
+; GCN-O0-NEXT: s_mov_b32 s7, s5
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s8, s11
+; GCN-O0-NEXT: s_add_u32 s6, s6, s9
+; GCN-O0-NEXT: s_addc_u32 s8, s7, s8
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_mov_b64 s[10:11], 2
+; GCN-O0-NEXT: s_mov_b32 s6, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: s_mov_b32 s8, s10
+; GCN-O0-NEXT: s_mov_b32 s7, s11
+; GCN-O0-NEXT: s_add_u32 s6, s6, s8
+; GCN-O0-NEXT: s_addc_u32 s3, s3, s7
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 1
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s4, s6
+; GCN-O0-NEXT: s_mov_b32 s3, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s4
+; GCN-O0-NEXT: s_addc_u32 s1, s1, s3
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <8 x i8> %vec, i8 1, i32 %sel
store <8 x i8> %v, ptr addrspace(1) %out
@@ -560,6 +1417,418 @@ define amdgpu_kernel void @byte16_inselt(ptr addrspace(1) %out, <16 x i8> %vec,
; GCN-NEXT: v_mov_b32_e32 v5, s5
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: byte16_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[0:1], s[4:5]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 52
+; GCN-O0-NEXT: s_mov_b32 s4, s0
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s3, s7
+; GCN-O0-NEXT: s_add_u32 s4, s4, s5
+; GCN-O0-NEXT: s_addc_u32 s2, s2, s3
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s2
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 0x43
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: s_mov_b32 s6, s9
+; GCN-O0-NEXT: s_add_u32 s2, s2, s7
+; GCN-O0-NEXT: s_addc_u32 s6, s3, s6
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_load_ubyte v0, v[0:1]
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 0x42
+; GCN-O0-NEXT: s_mov_b32 s6, s0
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: s_mov_b32 s3, s9
+; GCN-O0-NEXT: s_add_u32 s6, s6, s7
+; GCN-O0-NEXT: s_addc_u32 s2, s2, s3
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s2
+; GCN-O0-NEXT: s_mov_b64 s[10:11], 64
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s8, s11
+; GCN-O0-NEXT: s_add_u32 s2, s2, s9
+; GCN-O0-NEXT: s_addc_u32 s8, s3, s8
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: flat_load_ushort v1, v[1:2]
+; GCN-O0-NEXT: s_mov_b64 s[10:11], 63
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s8, s11
+; GCN-O0-NEXT: s_add_u32 s2, s2, s9
+; GCN-O0-NEXT: s_addc_u32 s8, s3, s8
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: flat_load_ubyte v13, v[2:3]
+; GCN-O0-NEXT: s_mov_b64 s[10:11], 62
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s3, s11
+; GCN-O0-NEXT: s_add_u32 s8, s8, s9
+; GCN-O0-NEXT: s_addc_u32 s2, s2, s3
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s9, s2
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 60
+; GCN-O0-NEXT: s_mov_b32 s10, s0
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: s_mov_b32 s11, s12
+; GCN-O0-NEXT: s_mov_b32 s3, s13
+; GCN-O0-NEXT: s_add_u32 s10, s10, s11
+; GCN-O0-NEXT: s_addc_u32 s2, s2, s3
+; GCN-O0-NEXT: ; kill: def $sgpr10 killed $sgpr10 def $sgpr10_sgpr11
+; GCN-O0-NEXT: s_mov_b32 s11, s2
+; GCN-O0-NEXT: s_mov_b64 s[14:15], 59
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s13, s14
+; GCN-O0-NEXT: s_mov_b32 s12, s15
+; GCN-O0-NEXT: s_add_u32 s2, s2, s13
+; GCN-O0-NEXT: s_addc_u32 s12, s3, s12
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: flat_load_ubyte v9, v[2:3]
+; GCN-O0-NEXT: s_mov_b64 s[14:15], 58
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s13, s14
+; GCN-O0-NEXT: s_mov_b32 s12, s15
+; GCN-O0-NEXT: s_add_u32 s2, s2, s13
+; GCN-O0-NEXT: s_addc_u32 s12, s3, s12
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: flat_load_ubyte v8, v[2:3]
+; GCN-O0-NEXT: s_mov_b64 s[14:15], 56
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s13, s14
+; GCN-O0-NEXT: s_mov_b32 s12, s15
+; GCN-O0-NEXT: s_add_u32 s2, s2, s13
+; GCN-O0-NEXT: s_addc_u32 s12, s3, s12
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: flat_load_ushort v7, v[2:3]
+; GCN-O0-NEXT: s_mov_b64 s[14:15], 54
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s13, s14
+; GCN-O0-NEXT: s_mov_b32 s12, s15
+; GCN-O0-NEXT: s_add_u32 s2, s2, s13
+; GCN-O0-NEXT: s_addc_u32 s12, s3, s12
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s3
+; GCN-O0-NEXT: flat_load_ushort v5, v[2:3]
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN-O0-NEXT: s_load_dword s0, s[0:1], 0x44
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: flat_load_ushort v3, v[2:3]
+; GCN-O0-NEXT: s_mov_b32 s1, 8
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v2, s1, v3
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v4, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s5, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v4, s[12:13]
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v4, s1, v5
+; GCN-O0-NEXT: s_mov_b32 s5, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v6, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s5, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v6, s[12:13]
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v6, s1, v7
+; GCN-O0-NEXT: s_mov_b32 s5, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v6, v6, v10, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s5, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v7, v7, v10, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s5, 6
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v8, v8, v10, s[12:13]
+; GCN-O0-NEXT: s_mov_b32 s5, 7
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[12:13], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v9, v9, v10, s[12:13]
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s11
+; GCN-O0-NEXT: flat_load_ushort v11, v[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v10, s1, v11
+; GCN-O0-NEXT: s_mov_b32 s5, 9
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[10:11], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v10, v10, v12, s[10:11]
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_cselect_b64 s[10:11], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v11, v11, v12, s[10:11]
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s8
+; GCN-O0-NEXT: flat_load_ubyte v12, v[14:15]
+; GCN-O0-NEXT: s_mov_b32 s5, 10
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[8:9], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s4
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_cndmask_b32_e64 v12, v12, v14, s[8:9]
+; GCN-O0-NEXT: s_mov_b32 s5, 11
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s5
+; GCN-O0-NEXT: s_cselect_b64 s[8:9], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v13, v13, v14, s[8:9]
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v14, s1, v1
+; GCN-O0-NEXT: s_mov_b32 s1, 13
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_cselect_b64 s[8:9], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v14, v14, v15, s[8:9]
+; GCN-O0-NEXT: s_mov_b32 s1, 12
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_cselect_b64 s[8:9], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v15, v1, v15, s[8:9]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s6
+; GCN-O0-NEXT: flat_load_ubyte v1, v[16:17]
+; GCN-O0-NEXT: s_mov_b32 s1, 14
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_cselect_b64 s[6:7], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s4
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_cndmask_b32_e64 v16, v1, v16, s[6:7]
+; GCN-O0-NEXT: s_mov_b32 s1, 15
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: s_cselect_b64 s[0:1], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: v_cndmask_b32_e64 v17, v0, v1, s[0:1]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 15
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v17
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 14
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v16
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 12
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v15
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 13
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v14
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 11
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v13
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 10
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v12
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 8
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v11
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 9
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v10
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 7
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v9
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 6
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v8
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 4
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v7
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 5
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v6
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 2
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v5
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 3
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_byte v[0:1], v3
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: s_mov_b32 s2, s5
+; GCN-O0-NEXT: s_add_u32 s0, s0, s3
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <16 x i8> %vec, i8 1, i32 %sel
store <16 x i8> %v, ptr addrspace(1) %out
@@ -587,6 +1856,40 @@ define amdgpu_kernel void @double2_inselt(ptr addrspace(1) %out, <2 x double> %v
; GCN-NEXT: v_mov_b32_e32 v5, s5
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double2_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x34
+; GCN-O0-NEXT: s_load_dword s7, s[4:5], 0x44
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 1.0
+; GCN-O0-NEXT: s_mov_b32 s4, s13
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s2, s11
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s7, s3
+; GCN-O0-NEXT: s_cselect_b32 s2, s4, s2
+; GCN-O0-NEXT: s_mov_b32 s5, s12
+; GCN-O0-NEXT: s_mov_b32 s3, s10
+; GCN-O0-NEXT: s_cselect_b32 s3, s5, s3
+; GCN-O0-NEXT: s_mov_b32 s6, s9
+; GCN-O0-NEXT: s_mov_b32 s12, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s7, s12
+; GCN-O0-NEXT: s_cselect_b32 s4, s4, s6
+; GCN-O0-NEXT: s_mov_b32 s6, s8
+; GCN-O0-NEXT: s_cselect_b32 s5, s5, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <2 x double> %vec, double 1.000000e+00, i32 %sel
store <2 x double> %v, ptr addrspace(1) %out
@@ -641,6 +1944,97 @@ define amdgpu_kernel void @double5_inselt(ptr addrspace(1) %out, <5 x double> %v
; GCN-NEXT: v_mov_b32_e32 v2, s0
; GCN-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double5_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[0:1], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GCN-O0-NEXT: s_load_dwordx8 s[12:19], s[0:1], 0x64
+; GCN-O0-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x84
+; GCN-O0-NEXT: s_load_dword s11, s[0:1], 0xa4
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 1.0
+; GCN-O0-NEXT: s_mov_b32 s8, s7
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s0, s5
+; GCN-O0-NEXT: s_mov_b32 s1, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s11, s1
+; GCN-O0-NEXT: s_cselect_b32 s0, s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s6
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: s_cselect_b32 s1, s9, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
+; GCN-O0-NEXT: s_mov_b32 s0, s15
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s11, s1
+; GCN-O0-NEXT: s_cselect_b32 s0, s8, s0
+; GCN-O0-NEXT: s_mov_b32 s1, s14
+; GCN-O0-NEXT: s_cselect_b32 s1, s9, s1
+; GCN-O0-NEXT: s_mov_b32 s4, s13
+; GCN-O0-NEXT: s_mov_b32 s5, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s11, s5
+; GCN-O0-NEXT: s_cselect_b32 s4, s8, s4
+; GCN-O0-NEXT: s_mov_b32 s5, s12
+; GCN-O0-NEXT: s_cselect_b32 s5, s9, s5
+; GCN-O0-NEXT: s_mov_b32 s6, s19
+; GCN-O0-NEXT: s_mov_b32 s7, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s11, s7
+; GCN-O0-NEXT: s_cselect_b32 s6, s8, s6
+; GCN-O0-NEXT: s_mov_b32 s7, s18
+; GCN-O0-NEXT: s_cselect_b32 s7, s9, s7
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_mov_b32 s20, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s11, s20
+; GCN-O0-NEXT: s_cselect_b32 s8, s8, s10
+; GCN-O0-NEXT: s_mov_b32 s10, s16
+; GCN-O0-NEXT: s_cselect_b32 s9, s9, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5_vgpr6_vgpr7 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v0
+; GCN-O0-NEXT: s_mov_b64 s[10:11], 16
+; GCN-O0-NEXT: s_mov_b32 s6, s2
+; GCN-O0-NEXT: s_mov_b32 s7, s3
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s8, s11
+; GCN-O0-NEXT: s_add_u32 s6, s6, s9
+; GCN-O0-NEXT: s_addc_u32 s8, s7, s8
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s7
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
+; GCN-O0-NEXT: s_nop 0
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5_vgpr6_vgpr7 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 32
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: s_mov_b32 s2, s5
+; GCN-O0-NEXT: s_add_u32 s0, s0, s3
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <5 x double> %vec, double 1.000000e+00, i32 %sel
store <5 x double> %v, ptr addrspace(1) %out
@@ -696,6 +2090,112 @@ define amdgpu_kernel void @double8_inselt(ptr addrspace(1) %out, <8 x double> %v
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double8_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x64
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0xa4
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s3
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s21
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v22, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v8, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v19
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 48
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v15
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 32
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v11
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <8 x double> %vec, double 1.000000e+00, i32 %sel
store <8 x double> %v, ptr addrspace(1) %out
@@ -749,6 +2249,147 @@ define amdgpu_kernel void @double7_inselt(ptr addrspace(1) %out, <7 x double> %v
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dwordx4 v[0:1], v[8:11]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double7_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x94
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s6, s1
+; GCN-O0-NEXT: s_mov_b32 s7, s0
+; GCN-O0-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x84
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s8, s15
+; GCN-O0-NEXT: s_mov_b32 s9, s14
+; GCN-O0-NEXT: s_mov_b32 s10, s13
+; GCN-O0-NEXT: s_mov_b32 s11, s12
+; GCN-O0-NEXT: s_load_dwordx8 s[20:27], s[4:5], 0x64
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s12, s27
+; GCN-O0-NEXT: s_mov_b32 s13, s26
+; GCN-O0-NEXT: s_mov_b32 s14, s25
+; GCN-O0-NEXT: s_mov_b32 s15, s24
+; GCN-O0-NEXT: s_mov_b32 s16, s23
+; GCN-O0-NEXT: s_mov_b32 s17, s22
+; GCN-O0-NEXT: s_mov_b32 s18, s21
+; GCN-O0-NEXT: s_mov_b32 s19, s20
+; GCN-O0-NEXT: ; implicit-def: $sgpr1
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr20
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v30, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v29, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v28, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v27, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v26, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v23, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v0
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0xa4
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s0, s0, s1
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s1
+; GCN-O0-NEXT: s_mov_b32 m0, s0
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s1
+; GCN-O0-NEXT: s_mov_b32 m0, s0
+; GCN-O0-NEXT: v_movreld_b32_e32 v8, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v15
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v11
+; GCN-O0-NEXT: ; kill: def $vgpr26 killed $vgpr26 def $vgpr26_vgpr27_vgpr28_vgpr29 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v24
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v24, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[24:25], v[26:29]
+; GCN-O0-NEXT: ; kill: def $vgpr23 killed $vgpr23 def $vgpr23_vgpr24_vgpr25_vgpr26 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s2
+; GCN-O0-NEXT: flat_store_dwordx4 v[3:4], v[23:26]
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v19
+; GCN-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v3
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 48
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s0
+; GCN-O0-NEXT: flat_store_dwordx2 v[3:4], v[7:8]
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 32
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: s_mov_b32 s2, s5
+; GCN-O0-NEXT: s_add_u32 s0, s0, s3
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <7 x double> %vec, double 1.000000e+00, i32 %sel
store <7 x double> %v, ptr addrspace(1) %out
@@ -846,6 +2487,275 @@ define amdgpu_kernel void @double16_inselt(ptr addrspace(1) %out, <16 x double>
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double16_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0xe4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s2, s51
+; GCN-O0-NEXT: s_mov_b32 s3, s50
+; GCN-O0-NEXT: s_mov_b32 s6, s49
+; GCN-O0-NEXT: s_mov_b32 s7, s48
+; GCN-O0-NEXT: s_mov_b32 s8, s47
+; GCN-O0-NEXT: s_mov_b32 s9, s46
+; GCN-O0-NEXT: s_mov_b32 s10, s45
+; GCN-O0-NEXT: s_mov_b32 s11, s44
+; GCN-O0-NEXT: s_mov_b32 s12, s43
+; GCN-O0-NEXT: s_mov_b32 s13, s42
+; GCN-O0-NEXT: s_mov_b32 s14, s41
+; GCN-O0-NEXT: s_mov_b32 s15, s40
+; GCN-O0-NEXT: s_mov_b32 s16, s39
+; GCN-O0-NEXT: s_mov_b32 s17, s38
+; GCN-O0-NEXT: s_mov_b32 s18, s37
+; GCN-O0-NEXT: s_mov_b32 s19, s36
+; GCN-O0-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0xa4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s20, s51
+; GCN-O0-NEXT: s_mov_b32 s21, s50
+; GCN-O0-NEXT: s_mov_b32 s22, s49
+; GCN-O0-NEXT: s_mov_b32 s23, s48
+; GCN-O0-NEXT: s_mov_b32 s24, s47
+; GCN-O0-NEXT: s_mov_b32 s25, s46
+; GCN-O0-NEXT: s_mov_b32 s26, s45
+; GCN-O0-NEXT: s_mov_b32 s27, s44
+; GCN-O0-NEXT: s_mov_b32 s28, s43
+; GCN-O0-NEXT: s_mov_b32 s29, s42
+; GCN-O0-NEXT: s_mov_b32 s30, s41
+; GCN-O0-NEXT: s_mov_b32 s31, s40
+; GCN-O0-NEXT: s_mov_b32 s33, s39
+; GCN-O0-NEXT: s_mov_b32 s34, s38
+; GCN-O0-NEXT: s_mov_b32 s35, s37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v62, s35
+; GCN-O0-NEXT: v_mov_b32_e32 v61, s34
+; GCN-O0-NEXT: v_mov_b32_e32 v60, s33
+; GCN-O0-NEXT: v_mov_b32_e32 v59, s31
+; GCN-O0-NEXT: v_mov_b32_e32 v58, s30
+; GCN-O0-NEXT: v_mov_b32_e32 v57, s29
+; GCN-O0-NEXT: v_mov_b32_e32 v56, s28
+; GCN-O0-NEXT: v_mov_b32_e32 v55, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v54, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v53, s25
+; GCN-O0-NEXT: v_mov_b32_e32 v52, s24
+; GCN-O0-NEXT: v_mov_b32_e32 v51, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v50, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s21
+; GCN-O0-NEXT: v_mov_b32_e32 v48, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v47, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v46, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v45, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v44, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v43, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v42, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v40, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v39, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v62
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v61
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v60
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v59
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v58
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v57
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v56
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v55
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v54
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v53
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v52
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v51
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v50
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v49
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v48
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v47
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v42
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v40
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v39
+; GCN-O0-NEXT: v_mov_b32_e32 v32, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v33, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v34, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v35, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v36, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v37, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v38, v0
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x124
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s3
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v8, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v38
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v37
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v36
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v35
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x70
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v34
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v33
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v32
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v31
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x60
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v27
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x50
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v23
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 64
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v19
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 48
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v15
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 32
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v11
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: s_mov_b32 s3, s1
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s2, s2, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s3, s4
+; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <16 x double> %vec, double 1.000000e+00, i32 %sel
store <16 x double> %v, ptr addrspace(1) %out
@@ -941,6 +2851,277 @@ define amdgpu_kernel void @double15_inselt(ptr addrspace(1) %out, <15 x double>
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_dwordx4 v[0:1], v[24:27]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: double15_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x114
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s6, s1
+; GCN-O0-NEXT: s_mov_b32 s7, s0
+; GCN-O0-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x104
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s8, s15
+; GCN-O0-NEXT: s_mov_b32 s9, s14
+; GCN-O0-NEXT: s_mov_b32 s10, s13
+; GCN-O0-NEXT: s_mov_b32 s11, s12
+; GCN-O0-NEXT: s_load_dwordx8 s[20:27], s[4:5], 0xe4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s12, s27
+; GCN-O0-NEXT: s_mov_b32 s13, s26
+; GCN-O0-NEXT: s_mov_b32 s14, s25
+; GCN-O0-NEXT: s_mov_b32 s15, s24
+; GCN-O0-NEXT: s_mov_b32 s16, s23
+; GCN-O0-NEXT: s_mov_b32 s17, s22
+; GCN-O0-NEXT: s_mov_b32 s18, s21
+; GCN-O0-NEXT: s_mov_b32 s19, s20
+; GCN-O0-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0xa4
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s20, s51
+; GCN-O0-NEXT: s_mov_b32 s21, s50
+; GCN-O0-NEXT: s_mov_b32 s22, s49
+; GCN-O0-NEXT: s_mov_b32 s23, s48
+; GCN-O0-NEXT: s_mov_b32 s24, s47
+; GCN-O0-NEXT: s_mov_b32 s25, s46
+; GCN-O0-NEXT: s_mov_b32 s26, s45
+; GCN-O0-NEXT: s_mov_b32 s27, s44
+; GCN-O0-NEXT: s_mov_b32 s28, s43
+; GCN-O0-NEXT: s_mov_b32 s29, s42
+; GCN-O0-NEXT: s_mov_b32 s30, s41
+; GCN-O0-NEXT: s_mov_b32 s31, s40
+; GCN-O0-NEXT: s_mov_b32 s33, s39
+; GCN-O0-NEXT: s_mov_b32 s34, s38
+; GCN-O0-NEXT: s_mov_b32 s35, s37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51
+; GCN-O0-NEXT: ; implicit-def: $sgpr1
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr37
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v62, s35
+; GCN-O0-NEXT: v_mov_b32_e32 v61, s34
+; GCN-O0-NEXT: v_mov_b32_e32 v60, s33
+; GCN-O0-NEXT: v_mov_b32_e32 v59, s31
+; GCN-O0-NEXT: v_mov_b32_e32 v58, s30
+; GCN-O0-NEXT: v_mov_b32_e32 v57, s29
+; GCN-O0-NEXT: v_mov_b32_e32 v56, s28
+; GCN-O0-NEXT: v_mov_b32_e32 v55, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v54, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v53, s25
+; GCN-O0-NEXT: v_mov_b32_e32 v52, s24
+; GCN-O0-NEXT: v_mov_b32_e32 v51, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v50, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s21
+; GCN-O0-NEXT: v_mov_b32_e32 v48, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v47, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v46, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v45, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v44, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v43, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v42, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v40, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v39, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v62
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v61
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v60
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v59
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v58
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v57
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v56
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v55
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v54
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v53
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v52
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v51
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v50
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v49
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v48
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v47
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v42
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v40
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v39
+; GCN-O0-NEXT: v_mov_b32_e32 v32, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v33, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v34, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v35, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v36, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v37, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v38, v0
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x124
+; GCN-O0-NEXT: s_mov_b32 s1, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_lshl_b32 s0, s0, s1
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s1
+; GCN-O0-NEXT: s_mov_b32 m0, s0
+; GCN-O0-NEXT: v_movreld_b32_e32 v7, v0
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s1
+; GCN-O0-NEXT: s_mov_b32 m0, s0
+; GCN-O0-NEXT: v_movreld_b32_e32 v8, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v34
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v33
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v32
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v31
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v39, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v40, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v41, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v46, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v42, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v43, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v44, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v45, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v47, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v48, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v49, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v54, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v50, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v51, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v52, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v53, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v55, v15
+; GCN-O0-NEXT: v_mov_b32_e32 v56, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v57, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v62, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v58, v19
+; GCN-O0-NEXT: ; kill: def $vgpr58 killed $vgpr58 def $vgpr58_vgpr59_vgpr60_vgpr61 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v59, v62
+; GCN-O0-NEXT: v_mov_b32_e32 v60, v57
+; GCN-O0-NEXT: v_mov_b32_e32 v61, v56
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 48
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v57, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v56, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[56:57], v[58:61]
+; GCN-O0-NEXT: ; kill: def $vgpr55 killed $vgpr55 def $vgpr55_vgpr56_vgpr57_vgpr58 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v56, v53
+; GCN-O0-NEXT: v_mov_b32_e32 v57, v52
+; GCN-O0-NEXT: v_mov_b32_e32 v58, v51
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 32
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v52, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v51, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[51:52], v[55:58]
+; GCN-O0-NEXT: ; kill: def $vgpr50 killed $vgpr50 def $vgpr50_vgpr51_vgpr52_vgpr53 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v51, v54
+; GCN-O0-NEXT: v_mov_b32_e32 v52, v49
+; GCN-O0-NEXT: v_mov_b32_e32 v53, v48
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 16
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v48, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[48:49], v[50:53]
+; GCN-O0-NEXT: ; kill: def $vgpr47 killed $vgpr47 def $vgpr47_vgpr48_vgpr49_vgpr50 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v48, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v49, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v50, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v44, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v43, s2
+; GCN-O0-NEXT: flat_store_dwordx4 v[43:44], v[47:50]
+; GCN-O0-NEXT: ; kill: def $vgpr42 killed $vgpr42 def $vgpr42_vgpr43_vgpr44_vgpr45 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v43, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v44, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v45, v40
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x50
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v40, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[40:41], v[42:45]
+; GCN-O0-NEXT: ; kill: def $vgpr39 killed $vgpr39 def $vgpr39_vgpr40_vgpr41_vgpr42 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v40, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v41, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v42, v3
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 64
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[3:4], v[39:42]
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v36
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v35
+; GCN-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v3
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x70
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s5
+; GCN-O0-NEXT: s_addc_u32 s4, s1, s4
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s0
+; GCN-O0-NEXT: flat_store_dwordx2 v[3:4], v[7:8]
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 0x60
+; GCN-O0-NEXT: s_mov_b32 s0, s2
+; GCN-O0-NEXT: s_mov_b32 s1, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: s_mov_b32 s2, s5
+; GCN-O0-NEXT: s_add_u32 s0, s0, s3
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <15 x double> %vec, double 1.000000e+00, i32 %sel
store <15 x double> %v, ptr addrspace(1) %out
@@ -996,6 +3177,63 @@ define amdgpu_kernel void @bit4_inselt(ptr addrspace(1) %out, <4 x i1> %vec, i32
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: flat_store_byte v[0:1], v2
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: bit4_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN-O0-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN-O0-NEXT: s_mov_b32 s14, -1
+; GCN-O0-NEXT: s_mov_b32 s15, 0xe80000
+; GCN-O0-NEXT: s_add_u32 s12, s12, s11
+; GCN-O0-NEXT: s_addc_u32 s13, s13, 0
+; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x24
+; GCN-O0-NEXT: s_load_dword s4, s[2:3], 0x2c
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0x30
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_bfe_u32 s3, s4, 0x10001
+; GCN-O0-NEXT: s_bfe_u32 s5, s4, 0x20002
+; GCN-O0-NEXT: s_bfe_u32 s6, s4, 0x10003
+; GCN-O0-NEXT: s_mov_b32 s7, 3
+; GCN-O0-NEXT: s_and_b32 s7, s2, s7
+; GCN-O0-NEXT: s_mov_b32 s2, 0
+; GCN-O0-NEXT: s_or_b32 s2, s2, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s5
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:1
+; GCN-O0-NEXT: v_mov_b32_e32 v3, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: buffer_store_byte v3, v0, s[12:15], 0 offen
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[12:15], 0
+; GCN-O0-NEXT: buffer_load_ubyte v4, off, s[12:15], 0 offset:1
+; GCN-O0-NEXT: buffer_load_ubyte v2, off, s[12:15], 0 offset:2
+; GCN-O0-NEXT: buffer_load_ubyte v1, off, s[12:15], 0 offset:3
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_and_b32_e64 v4, v4, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v4, v3, v4
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v4
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: s_mov_b32 s2, 2
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s2, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_mov_b32 s2, 3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: s_mov_b32 s2, 15
+; GCN-O0-NEXT: v_and_b32_e64 v2, v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <4 x i1> %vec, i1 1, i32 %sel
store <4 x i1> %v, ptr addrspace(1) %out
@@ -1838,6 +4076,1531 @@ define amdgpu_kernel void @bit128_inselt(ptr addrspace(1) %out, <128 x i1> %vec,
; GCN-NEXT: v_mov_b32_e32 v4, s0
; GCN-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NEXT: s_endpgm
+;
+; GCN-O0-LABEL: bit128_inselt:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 52
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s0, s5
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: s_mov_b32 s1, s7
+; GCN-O0-NEXT: s_add_u32 vcc_lo, s2, s3
+; GCN-O0-NEXT: s_addc_u32 s0, s0, s1
+; GCN-O0-NEXT: ; kill: def $vcc_lo killed $vcc_lo def $vcc
+; GCN-O0-NEXT: s_mov_b32 vcc_hi, s0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x43
+; GCN-O0-NEXT: s_mov_b32 s0, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: s_mov_b32 s2, s7
+; GCN-O0-NEXT: s_add_u32 s0, s0, s3
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: flat_load_ubyte v0, v[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 0xffff
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v0
+; GCN-O0-NEXT: s_and_b32 s1, s0, s1
+; GCN-O0-NEXT: s_mov_b32 s7, 1
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s7
+; GCN-O0-NEXT: ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v18, s2, 0
+; GCN-O0-NEXT: s_mov_b32 s15, 2
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s15
+; GCN-O0-NEXT: v_writelane_b32 v18, s2, 1
+; GCN-O0-NEXT: s_mov_b32 s13, 3
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s13
+; GCN-O0-NEXT: v_writelane_b32 v18, s2, 2
+; GCN-O0-NEXT: s_mov_b32 s11, 4
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s11
+; GCN-O0-NEXT: v_writelane_b32 v18, s2, 3
+; GCN-O0-NEXT: s_mov_b32 s9, 5
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s9
+; GCN-O0-NEXT: v_writelane_b32 v18, s2, 4
+; GCN-O0-NEXT: s_mov_b32 s6, 6
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s6
+; GCN-O0-NEXT: v_writelane_b32 v18, s2, 5
+; GCN-O0-NEXT: s_mov_b32 s24, 7
+; GCN-O0-NEXT: v_writelane_b32 v18, s24, 6
+; GCN-O0-NEXT: s_lshr_b32 s2, s1, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 0x42
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s10, s16
+; GCN-O0-NEXT: s_mov_b32 s3, s17
+; GCN-O0-NEXT: s_add_u32 s16, s8, s10
+; GCN-O0-NEXT: s_addc_u32 s1, s1, s3
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s17
+; GCN-O0-NEXT: flat_load_ubyte v1, v[1:2]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v1
+; GCN-O0-NEXT: s_and_b32 s1, s0, s1
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s7
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 7
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s15
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 8
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s13
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 9
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s11
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 10
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s9
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 11
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s6
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 12
+; GCN-O0-NEXT: s_lshr_b32 s1, s1, s24
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 13
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 0x41
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s10, s16
+; GCN-O0-NEXT: s_mov_b32 s3, s17
+; GCN-O0-NEXT: s_add_u32 s16, s8, s10
+; GCN-O0-NEXT: s_addc_u32 s1, s1, s3
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s17
+; GCN-O0-NEXT: flat_load_ubyte v2, v[2:3]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v2
+; GCN-O0-NEXT: s_and_b32 s1, s0, s1
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s7
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 14
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s15
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 15
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s13
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 16
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s11
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 17
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s9
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 18
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s6
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 19
+; GCN-O0-NEXT: s_lshr_b32 s1, s1, s24
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 20
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 64
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s10, s16
+; GCN-O0-NEXT: s_mov_b32 s3, s17
+; GCN-O0-NEXT: s_add_u32 s16, s8, s10
+; GCN-O0-NEXT: s_addc_u32 s1, s1, s3
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s17
+; GCN-O0-NEXT: flat_load_ubyte v3, v[3:4]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v3
+; GCN-O0-NEXT: s_and_b32 s1, s0, s1
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s7
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 21
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s15
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 22
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s13
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 23
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s11
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 24
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s9
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 25
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s6
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 26
+; GCN-O0-NEXT: s_lshr_b32 s1, s1, s24
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 27
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 63
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s10, s16
+; GCN-O0-NEXT: s_mov_b32 s3, s17
+; GCN-O0-NEXT: s_add_u32 s16, s8, s10
+; GCN-O0-NEXT: s_addc_u32 s1, s1, s3
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s17
+; GCN-O0-NEXT: flat_load_ubyte v4, v[4:5]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v4
+; GCN-O0-NEXT: s_and_b32 s1, s0, s1
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s7
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 28
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s15
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 29
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s13
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 30
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s11
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 31
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s9
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 32
+; GCN-O0-NEXT: s_lshr_b32 s3, s1, s6
+; GCN-O0-NEXT: s_lshr_b32 s1, s1, s24
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 33
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 62
+; GCN-O0-NEXT: s_mov_b32 s10, s4
+; GCN-O0-NEXT: s_mov_b32 s1, s5
+; GCN-O0-NEXT: s_mov_b32 s12, s16
+; GCN-O0-NEXT: s_mov_b32 s8, s17
+; GCN-O0-NEXT: s_add_u32 s16, s10, s12
+; GCN-O0-NEXT: s_addc_u32 s1, s1, s8
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s17
+; GCN-O0-NEXT: flat_load_ubyte v5, v[5:6]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v5
+; GCN-O0-NEXT: s_and_b32 s8, s0, s1
+; GCN-O0-NEXT: s_lshr_b32 s1, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s90, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s95, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s94, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s93, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s92, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s91, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 61
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s17
+; GCN-O0-NEXT: flat_load_ubyte v6, v[6:7]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v6
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s82, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s88, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s87, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s86, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s85, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s84, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s83, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 60
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s17
+; GCN-O0-NEXT: flat_load_ubyte v7, v[7:8]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v7
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s74, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s80, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s79, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s78, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s77, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s76, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s75, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 59
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s17
+; GCN-O0-NEXT: flat_load_ubyte v8, v[8:9]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v8
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s66, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s72, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s71, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s70, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s69, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s68, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s67, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 58
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s17
+; GCN-O0-NEXT: flat_load_ubyte v9, v[9:10]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v9
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s58, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s64, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s63, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s62, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s61, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s60, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s59, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 57
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s17
+; GCN-O0-NEXT: flat_load_ubyte v10, v[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v10
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s50, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s56, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s55, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s54, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s53, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s52, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s51, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 56
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s17
+; GCN-O0-NEXT: flat_load_ubyte v11, v[11:12]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v11
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s42, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s48, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s47, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s46, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s45, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s44, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s43, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 55
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s17
+; GCN-O0-NEXT: flat_load_ubyte v12, v[12:13]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v12
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s34, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s40, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s39, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s38, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s37, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s36, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s35, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 54
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s17
+; GCN-O0-NEXT: flat_load_ubyte v13, v[13:14]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v13
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s25, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s31, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s30, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s29, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s28, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s27, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s26, s8, s24
+; GCN-O0-NEXT: s_mov_b64 s[16:17], 53
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s8, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s16
+; GCN-O0-NEXT: s_mov_b32 s10, s17
+; GCN-O0-NEXT: s_add_u32 s16, s12, s14
+; GCN-O0-NEXT: s_addc_u32 s8, s8, s10
+; GCN-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17
+; GCN-O0-NEXT: s_mov_b32 s17, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s17
+; GCN-O0-NEXT: flat_load_ubyte v14, v[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v14
+; GCN-O0-NEXT: s_and_b32 s8, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s17, s8, s7
+; GCN-O0-NEXT: s_lshr_b32 s23, s8, s15
+; GCN-O0-NEXT: s_lshr_b32 s22, s8, s13
+; GCN-O0-NEXT: s_lshr_b32 s21, s8, s11
+; GCN-O0-NEXT: s_lshr_b32 s20, s8, s9
+; GCN-O0-NEXT: s_lshr_b32 s19, s8, s6
+; GCN-O0-NEXT: s_lshr_b32 s18, s8, s24
+; GCN-O0-NEXT: v_mov_b32_e32 v15, vcc_lo
+; GCN-O0-NEXT: v_mov_b32_e32 v16, vcc_hi
+; GCN-O0-NEXT: flat_load_ubyte v15, v[15:16]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s8, v15
+; GCN-O0-NEXT: s_and_b32 s0, s0, s8
+; GCN-O0-NEXT: s_lshr_b32 s16, s0, s7
+; GCN-O0-NEXT: s_lshr_b32 s14, s0, s15
+; GCN-O0-NEXT: s_lshr_b32 s12, s0, s13
+; GCN-O0-NEXT: s_lshr_b32 s10, s0, s11
+; GCN-O0-NEXT: s_lshr_b32 s8, s0, s9
+; GCN-O0-NEXT: s_lshr_b32 s33, s0, s6
+; GCN-O0-NEXT: v_writelane_b32 v18, s33, 34
+; GCN-O0-NEXT: s_lshr_b32 s24, s0, s24
+; GCN-O0-NEXT: s_load_dwordx2 vcc, s[4:5], 0x24
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_writelane_b32 v18, vcc_lo, 35
+; GCN-O0-NEXT: v_writelane_b32 v18, vcc_hi, 36
+; GCN-O0-NEXT: s_load_dword s0, s[4:5], 0x44
+; GCN-O0-NEXT: v_readlane_b32 s5, v18, 34
+; GCN-O0-NEXT: v_readlane_b32 s4, v18, 6
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s4
+; GCN-O0-NEXT: s_cselect_b32 s24, s7, s24
+; GCN-O0-NEXT: v_writelane_b32 v18, s24, 37
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s6
+; GCN-O0-NEXT: s_cselect_b32 s5, s7, s5
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s9
+; GCN-O0-NEXT: s_cselect_b32 s8, s7, s8
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s11
+; GCN-O0-NEXT: s_cselect_b32 s10, s7, s10
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s13
+; GCN-O0-NEXT: s_cselect_b32 s12, s7, s12
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s15
+; GCN-O0-NEXT: s_cselect_b32 s14, s7, s14
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s7
+; GCN-O0-NEXT: s_cselect_b32 s16, s7, s16
+; GCN-O0-NEXT: v_readfirstlane_b32 s24, v15
+; GCN-O0-NEXT: s_mov_b32 s33, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s24, s7, s24
+; GCN-O0-NEXT: v_writelane_b32 v18, s24, 38
+; GCN-O0-NEXT: s_mov_b32 s24, 15
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s18, s7, s18
+; GCN-O0-NEXT: s_mov_b32 s24, 14
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s19, s7, s19
+; GCN-O0-NEXT: s_mov_b32 s24, 13
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s20, s7, s20
+; GCN-O0-NEXT: s_mov_b32 s24, 12
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s21, s7, s21
+; GCN-O0-NEXT: s_mov_b32 s24, 11
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s22, s7, s22
+; GCN-O0-NEXT: s_mov_b32 s24, 10
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s23, s7, s23
+; GCN-O0-NEXT: s_mov_b32 s24, 9
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s24
+; GCN-O0-NEXT: s_cselect_b32 s24, s7, s17
+; GCN-O0-NEXT: v_readfirstlane_b32 s17, v14
+; GCN-O0-NEXT: s_mov_b32 s33, 8
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s17, s7, s17
+; GCN-O0-NEXT: s_mov_b32 s33, 23
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s26, s7, s26
+; GCN-O0-NEXT: s_mov_b32 s33, 22
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s27, s7, s27
+; GCN-O0-NEXT: s_mov_b32 s33, 21
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s28, s7, s28
+; GCN-O0-NEXT: s_mov_b32 s33, 20
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s29, s7, s29
+; GCN-O0-NEXT: s_mov_b32 s33, 19
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s30, s7, s30
+; GCN-O0-NEXT: s_mov_b32 s33, 18
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s31, s7, s31
+; GCN-O0-NEXT: s_mov_b32 s33, 17
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s33
+; GCN-O0-NEXT: s_cselect_b32 s33, s7, s25
+; GCN-O0-NEXT: v_readfirstlane_b32 s25, v13
+; GCN-O0-NEXT: s_mov_b32 s41, 16
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s25, s7, s25
+; GCN-O0-NEXT: s_mov_b32 s41, 31
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s35, s7, s35
+; GCN-O0-NEXT: s_mov_b32 s41, 30
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s36, s7, s36
+; GCN-O0-NEXT: s_mov_b32 s41, 29
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s37, s7, s37
+; GCN-O0-NEXT: s_mov_b32 s41, 28
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s38, s7, s38
+; GCN-O0-NEXT: s_mov_b32 s41, 27
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s39, s7, s39
+; GCN-O0-NEXT: s_mov_b32 s41, 26
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s40, s7, s40
+; GCN-O0-NEXT: s_mov_b32 s41, 25
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s41
+; GCN-O0-NEXT: s_cselect_b32 s41, s7, s34
+; GCN-O0-NEXT: v_readfirstlane_b32 s34, v12
+; GCN-O0-NEXT: s_mov_b32 s49, 24
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s34, s7, s34
+; GCN-O0-NEXT: s_mov_b32 s49, 39
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s43, s7, s43
+; GCN-O0-NEXT: s_mov_b32 s49, 38
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s44, s7, s44
+; GCN-O0-NEXT: s_mov_b32 s49, 37
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s45, s7, s45
+; GCN-O0-NEXT: s_mov_b32 s49, 36
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s46, s7, s46
+; GCN-O0-NEXT: s_mov_b32 s49, 35
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s47, s7, s47
+; GCN-O0-NEXT: s_mov_b32 s49, 34
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s48, s7, s48
+; GCN-O0-NEXT: s_mov_b32 s49, 33
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s49
+; GCN-O0-NEXT: s_cselect_b32 s49, s7, s42
+; GCN-O0-NEXT: v_readfirstlane_b32 s42, v11
+; GCN-O0-NEXT: s_mov_b32 s57, 32
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s42, s7, s42
+; GCN-O0-NEXT: s_mov_b32 s57, 47
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s51, s7, s51
+; GCN-O0-NEXT: s_mov_b32 s57, 46
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s52, s7, s52
+; GCN-O0-NEXT: s_mov_b32 s57, 45
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s53, s7, s53
+; GCN-O0-NEXT: s_mov_b32 s57, 44
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s54, s7, s54
+; GCN-O0-NEXT: s_mov_b32 s57, 43
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s55, s7, s55
+; GCN-O0-NEXT: s_mov_b32 s57, 42
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s56, s7, s56
+; GCN-O0-NEXT: s_mov_b32 s57, 41
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s57
+; GCN-O0-NEXT: s_cselect_b32 s57, s7, s50
+; GCN-O0-NEXT: v_readfirstlane_b32 s50, v10
+; GCN-O0-NEXT: s_mov_b32 s65, 40
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s50, s7, s50
+; GCN-O0-NEXT: s_mov_b32 s65, 55
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s59, s7, s59
+; GCN-O0-NEXT: s_mov_b32 s65, 54
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s60, s7, s60
+; GCN-O0-NEXT: s_mov_b32 s65, 53
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s61, s7, s61
+; GCN-O0-NEXT: s_mov_b32 s65, 52
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s62, s7, s62
+; GCN-O0-NEXT: s_mov_b32 s65, 51
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s63, s7, s63
+; GCN-O0-NEXT: s_mov_b32 s65, 50
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s64, s7, s64
+; GCN-O0-NEXT: s_mov_b32 s65, 49
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s65
+; GCN-O0-NEXT: s_cselect_b32 s65, s7, s58
+; GCN-O0-NEXT: v_readfirstlane_b32 s58, v9
+; GCN-O0-NEXT: s_mov_b32 s73, 48
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s58, s7, s58
+; GCN-O0-NEXT: s_mov_b32 s73, 63
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s67, s7, s67
+; GCN-O0-NEXT: s_mov_b32 s73, 62
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s68, s7, s68
+; GCN-O0-NEXT: s_mov_b32 s73, 61
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s69, s7, s69
+; GCN-O0-NEXT: s_mov_b32 s73, 60
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s70, s7, s70
+; GCN-O0-NEXT: s_mov_b32 s73, 59
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s71, s7, s71
+; GCN-O0-NEXT: s_mov_b32 s73, 58
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s72, s7, s72
+; GCN-O0-NEXT: s_mov_b32 s73, 57
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s73
+; GCN-O0-NEXT: s_cselect_b32 s73, s7, s66
+; GCN-O0-NEXT: v_readfirstlane_b32 s66, v8
+; GCN-O0-NEXT: s_mov_b32 s81, 56
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s66, s7, s66
+; GCN-O0-NEXT: s_mov_b32 s81, 0x47
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s75, s7, s75
+; GCN-O0-NEXT: s_mov_b32 s81, 0x46
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s76, s7, s76
+; GCN-O0-NEXT: s_mov_b32 s81, 0x45
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s77, s7, s77
+; GCN-O0-NEXT: s_mov_b32 s81, 0x44
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s78, s7, s78
+; GCN-O0-NEXT: s_mov_b32 s81, 0x43
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s79, s7, s79
+; GCN-O0-NEXT: s_mov_b32 s81, 0x42
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s80, s7, s80
+; GCN-O0-NEXT: s_mov_b32 s81, 0x41
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s81
+; GCN-O0-NEXT: s_cselect_b32 s81, s7, s74
+; GCN-O0-NEXT: v_readfirstlane_b32 s74, v7
+; GCN-O0-NEXT: s_mov_b32 s89, 64
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s74, s7, s74
+; GCN-O0-NEXT: s_mov_b32 s89, 0x4f
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s83, s7, s83
+; GCN-O0-NEXT: s_mov_b32 s89, 0x4e
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s84, s7, s84
+; GCN-O0-NEXT: s_mov_b32 s89, 0x4d
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s85, s7, s85
+; GCN-O0-NEXT: s_mov_b32 s89, 0x4c
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s86, s7, s86
+; GCN-O0-NEXT: s_mov_b32 s89, 0x4b
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s87, s7, s87
+; GCN-O0-NEXT: s_mov_b32 s89, 0x4a
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s88, s7, s88
+; GCN-O0-NEXT: s_mov_b32 s89, 0x49
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s89
+; GCN-O0-NEXT: s_cselect_b32 s89, s7, s82
+; GCN-O0-NEXT: v_readfirstlane_b32 s82, v6
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x48
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 s82, s7, s82
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x57
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 s91, s7, s91
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x56
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 s92, s7, s92
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x55
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 s93, s7, s93
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x54
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 s94, s7, s94
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x53
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 s95, s7, s95
+; GCN-O0-NEXT: s_mov_b32 vcc_lo, 0x52
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, vcc_lo
+; GCN-O0-NEXT: s_cselect_b32 vcc_lo, s7, s90
+; GCN-O0-NEXT: s_mov_b32 s90, 0x51
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s90
+; GCN-O0-NEXT: s_cselect_b32 vcc_hi, s7, s1
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v5
+; GCN-O0-NEXT: s_mov_b32 s90, 0x50
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s90
+; GCN-O0-NEXT: s_cselect_b32 s90, s7, s1
+; GCN-O0-NEXT: s_mov_b32 s1, 0x5f
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 33
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 39
+; GCN-O0-NEXT: s_mov_b32 s1, 0x5e
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 32
+; GCN-O0-NEXT: s_cselect_b32 s3, s7, s3
+; GCN-O0-NEXT: v_writelane_b32 v18, s3, 40
+; GCN-O0-NEXT: s_mov_b32 s3, 0x5d
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s3
+; GCN-O0-NEXT: s_cselect_b32 s3, s7, s1
+; GCN-O0-NEXT: s_mov_b32 s1, 0x5c
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 31
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 41
+; GCN-O0-NEXT: s_mov_b32 s1, 0x5b
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 30
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 42
+; GCN-O0-NEXT: s_mov_b32 s1, 0x5a
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 29
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 43
+; GCN-O0-NEXT: s_mov_b32 s1, 0x59
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 28
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 44
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v4
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 45
+; GCN-O0-NEXT: s_mov_b32 s1, 0x58
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 45
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 46
+; GCN-O0-NEXT: s_mov_b32 s1, 0x67
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 27
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 47
+; GCN-O0-NEXT: s_mov_b32 s1, 0x66
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 26
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 48
+; GCN-O0-NEXT: s_mov_b32 s1, 0x65
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 25
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 49
+; GCN-O0-NEXT: s_mov_b32 s1, 0x64
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 24
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 50
+; GCN-O0-NEXT: s_mov_b32 s1, 0x63
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 23
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 51
+; GCN-O0-NEXT: s_mov_b32 s1, 0x62
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 22
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 52
+; GCN-O0-NEXT: s_mov_b32 s1, 0x61
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 21
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 53
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v3
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 54
+; GCN-O0-NEXT: s_mov_b32 s1, 0x60
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 54
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 55
+; GCN-O0-NEXT: s_mov_b32 s1, 0x6f
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 20
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 56
+; GCN-O0-NEXT: s_mov_b32 s1, 0x6e
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 19
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 57
+; GCN-O0-NEXT: s_mov_b32 s1, 0x6d
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 18
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 58
+; GCN-O0-NEXT: s_mov_b32 s1, 0x6c
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 17
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 59
+; GCN-O0-NEXT: s_mov_b32 s1, 0x6b
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 16
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 60
+; GCN-O0-NEXT: s_mov_b32 s1, 0x6a
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 15
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 61
+; GCN-O0-NEXT: s_mov_b32 s1, 0x69
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 14
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 62
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v2
+; GCN-O0-NEXT: v_writelane_b32 v18, s1, 63
+; GCN-O0-NEXT: s_mov_b32 s1, 0x68
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 63
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: ; implicit-def: $vgpr17 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 0
+; GCN-O0-NEXT: s_mov_b32 s1, 0x77
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 13
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 1
+; GCN-O0-NEXT: s_mov_b32 s1, 0x76
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 12
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 2
+; GCN-O0-NEXT: s_mov_b32 s1, 0x75
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 11
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 3
+; GCN-O0-NEXT: s_mov_b32 s1, 0x74
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 10
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 4
+; GCN-O0-NEXT: s_mov_b32 s1, 0x73
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 9
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 5
+; GCN-O0-NEXT: s_mov_b32 s1, 0x72
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 8
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 6
+; GCN-O0-NEXT: s_mov_b32 s1, 0x71
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 7
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 7
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 8
+; GCN-O0-NEXT: s_mov_b32 s1, 0x70
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 8
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 9
+; GCN-O0-NEXT: s_mov_b32 s1, 0x7f
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 5
+; GCN-O0-NEXT: s_cselect_b32 s2, s7, s2
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 10
+; GCN-O0-NEXT: s_mov_b32 s2, 0x7e
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s2
+; GCN-O0-NEXT: s_cselect_b32 s2, s7, s1
+; GCN-O0-NEXT: s_mov_b32 s1, 0x7d
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 4
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 11
+; GCN-O0-NEXT: s_mov_b32 s1, 0x7c
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 3
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 12
+; GCN-O0-NEXT: s_mov_b32 s1, 0x7b
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 2
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 13
+; GCN-O0-NEXT: s_mov_b32 s1, 0x7a
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 1
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 14
+; GCN-O0-NEXT: s_mov_b32 s1, 0x79
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 0
+; GCN-O0-NEXT: s_cselect_b32 s1, s7, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 15
+; GCN-O0-NEXT: v_readfirstlane_b32 s1, v0
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 16
+; GCN-O0-NEXT: s_mov_b32 s1, 0x78
+; GCN-O0-NEXT: s_cmp_eq_u32 s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 15
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 16
+; GCN-O0-NEXT: s_cselect_b32 s0, s7, s0
+; GCN-O0-NEXT: s_and_b32 s0, s0, s7
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s7
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 14
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s15
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 13
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s13
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 12
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s11
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 11
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s9
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 10
+; GCN-O0-NEXT: s_and_b32 s2, s2, s7
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s6
+; GCN-O0-NEXT: s_or_b32 s0, s0, s2
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s4
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 17
+; GCN-O0-NEXT: s_mov_b64 s[0:1], 15
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 18
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 19
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 36
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 20
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 18
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 19
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 21
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 22
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 22
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 20
+; GCN-O0-NEXT: s_add_u32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 21
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_readlane_b32 s2, v17, 2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 7
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 17
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 9
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s0, s0, s7
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s7
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 6
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s15
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 5
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s13
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 4
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s11
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 3
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s9
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 1
+; GCN-O0-NEXT: s_and_b32 s2, s2, s7
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s6
+; GCN-O0-NEXT: s_or_b32 s0, s0, s2
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s4
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 23
+; GCN-O0-NEXT: s_mov_b64 s[0:1], 14
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 24
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 25
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 36
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 26
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 24
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 25
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 27
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 28
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 28
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 26
+; GCN-O0-NEXT: s_add_u32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 27
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_readlane_b32 s2, v18, 57
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 62
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 23
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 0
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s0, s0, s7
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s7
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 61
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s15
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 60
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s13
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 59
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s11
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 58
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s9
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 56
+; GCN-O0-NEXT: s_and_b32 s2, s2, s7
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s6
+; GCN-O0-NEXT: s_or_b32 s0, s0, s2
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s4
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 29
+; GCN-O0-NEXT: s_mov_b64 s[0:1], 13
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 30
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 31
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 36
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 32
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 30
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 31
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 33
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 34
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 34
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 32
+; GCN-O0-NEXT: s_add_u32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 33
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_readlane_b32 s2, v18, 48
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 53
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 29
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 55
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s0, s0, s7
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s7
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 52
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s15
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 51
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s13
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 50
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s11
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 49
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s9
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 47
+; GCN-O0-NEXT: s_and_b32 s2, s2, s7
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s6
+; GCN-O0-NEXT: s_or_b32 s0, s0, s2
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s4
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 35
+; GCN-O0-NEXT: s_mov_b64 s[0:1], 12
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 36
+; GCN-O0-NEXT: v_writelane_b32 v17, s1, 37
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 36
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 38
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 36
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 37
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 39
+; GCN-O0-NEXT: s_mov_b32 s2, s0
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 40
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 40
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 38
+; GCN-O0-NEXT: s_add_u32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 39
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_readlane_b32 s2, v18, 40
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 44
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 35
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 46
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s0, s0, s7
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s7
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 43
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s15
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 42
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s13
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 41
+; GCN-O0-NEXT: s_and_b32 s1, s1, s7
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s11
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 39
+; GCN-O0-NEXT: s_and_b32 s3, s3, s7
+; GCN-O0-NEXT: s_lshl_b32 s3, s3, s9
+; GCN-O0-NEXT: s_or_b32 s0, s0, s3
+; GCN-O0-NEXT: s_and_b32 s2, s2, s7
+; GCN-O0-NEXT: s_lshl_b32 s2, s2, s6
+; GCN-O0-NEXT: s_or_b32 s0, s0, s2
+; GCN-O0-NEXT: v_readlane_b32 s2, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s3, v18, 36
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s4
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_writelane_b32 v17, s0, 41
+; GCN-O0-NEXT: s_mov_b64 s[0:1], 11
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 42
+; GCN-O0-NEXT: v_readlane_b32 s2, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s3, v18, 36
+; GCN-O0-NEXT: s_mov_b32 s2, s3
+; GCN-O0-NEXT: v_writelane_b32 v17, s2, 43
+; GCN-O0-NEXT: s_mov_b32 s3, s0
+; GCN-O0-NEXT: s_mov_b32 s2, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v17, 43
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 42
+; GCN-O0-NEXT: s_add_u32 s0, s0, s3
+; GCN-O0-NEXT: s_addc_u32 s2, s1, s2
+; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
+; GCN-O0-NEXT: s_mov_b32 s1, s2
+; GCN-O0-NEXT: v_readlane_b32 s2, v18, 35
+; GCN-O0-NEXT: v_readlane_b32 s3, v18, 36
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: v_readlane_b32 s1, v18, 37
+; GCN-O0-NEXT: v_readlane_b32 s0, v17, 41
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: v_readlane_b32 s0, v18, 38
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s90, s90, s7
+; GCN-O0-NEXT: s_and_b32 vcc_hi, vcc_hi, s7
+; GCN-O0-NEXT: s_lshl_b32 vcc_hi, vcc_hi, s7
+; GCN-O0-NEXT: s_or_b32 s90, s90, vcc_hi
+; GCN-O0-NEXT: s_and_b32 vcc_lo, vcc_lo, s7
+; GCN-O0-NEXT: s_lshl_b32 vcc_lo, vcc_lo, s15
+; GCN-O0-NEXT: s_or_b32 s90, s90, vcc_lo
+; GCN-O0-NEXT: s_and_b32 s95, s95, s7
+; GCN-O0-NEXT: s_lshl_b32 s95, s95, s13
+; GCN-O0-NEXT: s_or_b32 s90, s90, s95
+; GCN-O0-NEXT: s_and_b32 s94, s94, s7
+; GCN-O0-NEXT: s_lshl_b32 s94, s94, s11
+; GCN-O0-NEXT: s_or_b32 s90, s90, s94
+; GCN-O0-NEXT: s_and_b32 s93, s93, s7
+; GCN-O0-NEXT: s_lshl_b32 s93, s93, s9
+; GCN-O0-NEXT: s_or_b32 s90, s90, s93
+; GCN-O0-NEXT: s_and_b32 s92, s92, s7
+; GCN-O0-NEXT: s_lshl_b32 s92, s92, s6
+; GCN-O0-NEXT: s_or_b32 s90, s90, s92
+; GCN-O0-NEXT: s_lshl_b32 s91, s91, s4
+; GCN-O0-NEXT: s_or_b32 s90, s90, s91
+; GCN-O0-NEXT: s_mov_b64 vcc, 10
+; GCN-O0-NEXT: s_mov_b32 s92, s2
+; GCN-O0-NEXT: s_mov_b32 s91, s3
+; GCN-O0-NEXT: s_mov_b32 s94, vcc_lo
+; GCN-O0-NEXT: s_mov_b32 s93, vcc_hi
+; GCN-O0-NEXT: s_add_u32 s92, s92, s94
+; GCN-O0-NEXT: s_addc_u32 s91, s91, s93
+; GCN-O0-NEXT: ; kill: def $sgpr92 killed $sgpr92 def $sgpr92_sgpr93
+; GCN-O0-NEXT: s_mov_b32 s93, s91
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s92
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s93
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s90
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s82, s82, s7
+; GCN-O0-NEXT: s_and_b32 s89, s89, s7
+; GCN-O0-NEXT: s_lshl_b32 s89, s89, s7
+; GCN-O0-NEXT: s_or_b32 s82, s82, s89
+; GCN-O0-NEXT: s_and_b32 s88, s88, s7
+; GCN-O0-NEXT: s_lshl_b32 s88, s88, s15
+; GCN-O0-NEXT: s_or_b32 s82, s82, s88
+; GCN-O0-NEXT: s_and_b32 s87, s87, s7
+; GCN-O0-NEXT: s_lshl_b32 s87, s87, s13
+; GCN-O0-NEXT: s_or_b32 s82, s82, s87
+; GCN-O0-NEXT: s_and_b32 s86, s86, s7
+; GCN-O0-NEXT: s_lshl_b32 s86, s86, s11
+; GCN-O0-NEXT: s_or_b32 s82, s82, s86
+; GCN-O0-NEXT: s_and_b32 s85, s85, s7
+; GCN-O0-NEXT: s_lshl_b32 s85, s85, s9
+; GCN-O0-NEXT: s_or_b32 s82, s82, s85
+; GCN-O0-NEXT: s_and_b32 s84, s84, s7
+; GCN-O0-NEXT: s_lshl_b32 s84, s84, s6
+; GCN-O0-NEXT: s_or_b32 s82, s82, s84
+; GCN-O0-NEXT: s_lshl_b32 s83, s83, s4
+; GCN-O0-NEXT: s_or_b32 s82, s82, s83
+; GCN-O0-NEXT: s_mov_b64 s[88:89], 9
+; GCN-O0-NEXT: s_mov_b32 s84, s2
+; GCN-O0-NEXT: s_mov_b32 s83, s3
+; GCN-O0-NEXT: s_mov_b32 s86, s88
+; GCN-O0-NEXT: s_mov_b32 s85, s89
+; GCN-O0-NEXT: s_add_u32 s84, s84, s86
+; GCN-O0-NEXT: s_addc_u32 s83, s83, s85
+; GCN-O0-NEXT: ; kill: def $sgpr84 killed $sgpr84 def $sgpr84_sgpr85
+; GCN-O0-NEXT: s_mov_b32 s85, s83
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s84
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s85
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s82
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s74, s74, s7
+; GCN-O0-NEXT: s_and_b32 s81, s81, s7
+; GCN-O0-NEXT: s_lshl_b32 s81, s81, s7
+; GCN-O0-NEXT: s_or_b32 s74, s74, s81
+; GCN-O0-NEXT: s_and_b32 s80, s80, s7
+; GCN-O0-NEXT: s_lshl_b32 s80, s80, s15
+; GCN-O0-NEXT: s_or_b32 s74, s74, s80
+; GCN-O0-NEXT: s_and_b32 s79, s79, s7
+; GCN-O0-NEXT: s_lshl_b32 s79, s79, s13
+; GCN-O0-NEXT: s_or_b32 s74, s74, s79
+; GCN-O0-NEXT: s_and_b32 s78, s78, s7
+; GCN-O0-NEXT: s_lshl_b32 s78, s78, s11
+; GCN-O0-NEXT: s_or_b32 s74, s74, s78
+; GCN-O0-NEXT: s_and_b32 s77, s77, s7
+; GCN-O0-NEXT: s_lshl_b32 s77, s77, s9
+; GCN-O0-NEXT: s_or_b32 s74, s74, s77
+; GCN-O0-NEXT: s_and_b32 s76, s76, s7
+; GCN-O0-NEXT: s_lshl_b32 s76, s76, s6
+; GCN-O0-NEXT: s_or_b32 s74, s74, s76
+; GCN-O0-NEXT: s_lshl_b32 s75, s75, s4
+; GCN-O0-NEXT: s_or_b32 s74, s74, s75
+; GCN-O0-NEXT: s_mov_b64 s[80:81], 8
+; GCN-O0-NEXT: s_mov_b32 s76, s2
+; GCN-O0-NEXT: s_mov_b32 s75, s3
+; GCN-O0-NEXT: s_mov_b32 s78, s80
+; GCN-O0-NEXT: s_mov_b32 s77, s81
+; GCN-O0-NEXT: s_add_u32 s76, s76, s78
+; GCN-O0-NEXT: s_addc_u32 s75, s75, s77
+; GCN-O0-NEXT: ; kill: def $sgpr76 killed $sgpr76 def $sgpr76_sgpr77
+; GCN-O0-NEXT: s_mov_b32 s77, s75
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s76
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s77
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s74
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s66, s66, s7
+; GCN-O0-NEXT: s_and_b32 s73, s73, s7
+; GCN-O0-NEXT: s_lshl_b32 s73, s73, s7
+; GCN-O0-NEXT: s_or_b32 s66, s66, s73
+; GCN-O0-NEXT: s_and_b32 s72, s72, s7
+; GCN-O0-NEXT: s_lshl_b32 s72, s72, s15
+; GCN-O0-NEXT: s_or_b32 s66, s66, s72
+; GCN-O0-NEXT: s_and_b32 s71, s71, s7
+; GCN-O0-NEXT: s_lshl_b32 s71, s71, s13
+; GCN-O0-NEXT: s_or_b32 s66, s66, s71
+; GCN-O0-NEXT: s_and_b32 s70, s70, s7
+; GCN-O0-NEXT: s_lshl_b32 s70, s70, s11
+; GCN-O0-NEXT: s_or_b32 s66, s66, s70
+; GCN-O0-NEXT: s_and_b32 s69, s69, s7
+; GCN-O0-NEXT: s_lshl_b32 s69, s69, s9
+; GCN-O0-NEXT: s_or_b32 s66, s66, s69
+; GCN-O0-NEXT: s_and_b32 s68, s68, s7
+; GCN-O0-NEXT: s_lshl_b32 s68, s68, s6
+; GCN-O0-NEXT: s_or_b32 s66, s66, s68
+; GCN-O0-NEXT: s_lshl_b32 s67, s67, s4
+; GCN-O0-NEXT: s_or_b32 s66, s66, s67
+; GCN-O0-NEXT: s_mov_b64 s[72:73], 7
+; GCN-O0-NEXT: s_mov_b32 s68, s2
+; GCN-O0-NEXT: s_mov_b32 s67, s3
+; GCN-O0-NEXT: s_mov_b32 s70, s72
+; GCN-O0-NEXT: s_mov_b32 s69, s73
+; GCN-O0-NEXT: s_add_u32 s68, s68, s70
+; GCN-O0-NEXT: s_addc_u32 s67, s67, s69
+; GCN-O0-NEXT: ; kill: def $sgpr68 killed $sgpr68 def $sgpr68_sgpr69
+; GCN-O0-NEXT: s_mov_b32 s69, s67
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s68
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s69
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s66
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s58, s58, s7
+; GCN-O0-NEXT: s_and_b32 s65, s65, s7
+; GCN-O0-NEXT: s_lshl_b32 s65, s65, s7
+; GCN-O0-NEXT: s_or_b32 s58, s58, s65
+; GCN-O0-NEXT: s_and_b32 s64, s64, s7
+; GCN-O0-NEXT: s_lshl_b32 s64, s64, s15
+; GCN-O0-NEXT: s_or_b32 s58, s58, s64
+; GCN-O0-NEXT: s_and_b32 s63, s63, s7
+; GCN-O0-NEXT: s_lshl_b32 s63, s63, s13
+; GCN-O0-NEXT: s_or_b32 s58, s58, s63
+; GCN-O0-NEXT: s_and_b32 s62, s62, s7
+; GCN-O0-NEXT: s_lshl_b32 s62, s62, s11
+; GCN-O0-NEXT: s_or_b32 s58, s58, s62
+; GCN-O0-NEXT: s_and_b32 s61, s61, s7
+; GCN-O0-NEXT: s_lshl_b32 s61, s61, s9
+; GCN-O0-NEXT: s_or_b32 s58, s58, s61
+; GCN-O0-NEXT: s_and_b32 s60, s60, s7
+; GCN-O0-NEXT: s_lshl_b32 s60, s60, s6
+; GCN-O0-NEXT: s_or_b32 s58, s58, s60
+; GCN-O0-NEXT: s_lshl_b32 s59, s59, s4
+; GCN-O0-NEXT: s_or_b32 s58, s58, s59
+; GCN-O0-NEXT: s_mov_b64 s[64:65], 6
+; GCN-O0-NEXT: s_mov_b32 s60, s2
+; GCN-O0-NEXT: s_mov_b32 s59, s3
+; GCN-O0-NEXT: s_mov_b32 s62, s64
+; GCN-O0-NEXT: s_mov_b32 s61, s65
+; GCN-O0-NEXT: s_add_u32 s60, s60, s62
+; GCN-O0-NEXT: s_addc_u32 s59, s59, s61
+; GCN-O0-NEXT: ; kill: def $sgpr60 killed $sgpr60 def $sgpr60_sgpr61
+; GCN-O0-NEXT: s_mov_b32 s61, s59
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s60
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s61
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s58
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s50, s50, s7
+; GCN-O0-NEXT: s_and_b32 s57, s57, s7
+; GCN-O0-NEXT: s_lshl_b32 s57, s57, s7
+; GCN-O0-NEXT: s_or_b32 s50, s50, s57
+; GCN-O0-NEXT: s_and_b32 s56, s56, s7
+; GCN-O0-NEXT: s_lshl_b32 s56, s56, s15
+; GCN-O0-NEXT: s_or_b32 s50, s50, s56
+; GCN-O0-NEXT: s_and_b32 s55, s55, s7
+; GCN-O0-NEXT: s_lshl_b32 s55, s55, s13
+; GCN-O0-NEXT: s_or_b32 s50, s50, s55
+; GCN-O0-NEXT: s_and_b32 s54, s54, s7
+; GCN-O0-NEXT: s_lshl_b32 s54, s54, s11
+; GCN-O0-NEXT: s_or_b32 s50, s50, s54
+; GCN-O0-NEXT: s_and_b32 s53, s53, s7
+; GCN-O0-NEXT: s_lshl_b32 s53, s53, s9
+; GCN-O0-NEXT: s_or_b32 s50, s50, s53
+; GCN-O0-NEXT: s_and_b32 s52, s52, s7
+; GCN-O0-NEXT: s_lshl_b32 s52, s52, s6
+; GCN-O0-NEXT: s_or_b32 s50, s50, s52
+; GCN-O0-NEXT: s_lshl_b32 s51, s51, s4
+; GCN-O0-NEXT: s_or_b32 s50, s50, s51
+; GCN-O0-NEXT: s_mov_b64 s[56:57], 5
+; GCN-O0-NEXT: s_mov_b32 s52, s2
+; GCN-O0-NEXT: s_mov_b32 s51, s3
+; GCN-O0-NEXT: s_mov_b32 s54, s56
+; GCN-O0-NEXT: s_mov_b32 s53, s57
+; GCN-O0-NEXT: s_add_u32 s52, s52, s54
+; GCN-O0-NEXT: s_addc_u32 s51, s51, s53
+; GCN-O0-NEXT: ; kill: def $sgpr52 killed $sgpr52 def $sgpr52_sgpr53
+; GCN-O0-NEXT: s_mov_b32 s53, s51
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s52
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s53
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s50
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s42, s42, s7
+; GCN-O0-NEXT: s_and_b32 s49, s49, s7
+; GCN-O0-NEXT: s_lshl_b32 s49, s49, s7
+; GCN-O0-NEXT: s_or_b32 s42, s42, s49
+; GCN-O0-NEXT: s_and_b32 s48, s48, s7
+; GCN-O0-NEXT: s_lshl_b32 s48, s48, s15
+; GCN-O0-NEXT: s_or_b32 s42, s42, s48
+; GCN-O0-NEXT: s_and_b32 s47, s47, s7
+; GCN-O0-NEXT: s_lshl_b32 s47, s47, s13
+; GCN-O0-NEXT: s_or_b32 s42, s42, s47
+; GCN-O0-NEXT: s_and_b32 s46, s46, s7
+; GCN-O0-NEXT: s_lshl_b32 s46, s46, s11
+; GCN-O0-NEXT: s_or_b32 s42, s42, s46
+; GCN-O0-NEXT: s_and_b32 s45, s45, s7
+; GCN-O0-NEXT: s_lshl_b32 s45, s45, s9
+; GCN-O0-NEXT: s_or_b32 s42, s42, s45
+; GCN-O0-NEXT: s_and_b32 s44, s44, s7
+; GCN-O0-NEXT: s_lshl_b32 s44, s44, s6
+; GCN-O0-NEXT: s_or_b32 s42, s42, s44
+; GCN-O0-NEXT: s_lshl_b32 s43, s43, s4
+; GCN-O0-NEXT: s_or_b32 s42, s42, s43
+; GCN-O0-NEXT: s_mov_b64 s[48:49], 4
+; GCN-O0-NEXT: s_mov_b32 s44, s2
+; GCN-O0-NEXT: s_mov_b32 s43, s3
+; GCN-O0-NEXT: s_mov_b32 s46, s48
+; GCN-O0-NEXT: s_mov_b32 s45, s49
+; GCN-O0-NEXT: s_add_u32 s44, s44, s46
+; GCN-O0-NEXT: s_addc_u32 s43, s43, s45
+; GCN-O0-NEXT: ; kill: def $sgpr44 killed $sgpr44 def $sgpr44_sgpr45
+; GCN-O0-NEXT: s_mov_b32 s45, s43
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s42
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s34, s34, s7
+; GCN-O0-NEXT: s_and_b32 s41, s41, s7
+; GCN-O0-NEXT: s_lshl_b32 s41, s41, s7
+; GCN-O0-NEXT: s_or_b32 s34, s34, s41
+; GCN-O0-NEXT: s_and_b32 s40, s40, s7
+; GCN-O0-NEXT: s_lshl_b32 s40, s40, s15
+; GCN-O0-NEXT: s_or_b32 s34, s34, s40
+; GCN-O0-NEXT: s_and_b32 s39, s39, s7
+; GCN-O0-NEXT: s_lshl_b32 s39, s39, s13
+; GCN-O0-NEXT: s_or_b32 s34, s34, s39
+; GCN-O0-NEXT: s_and_b32 s38, s38, s7
+; GCN-O0-NEXT: s_lshl_b32 s38, s38, s11
+; GCN-O0-NEXT: s_or_b32 s34, s34, s38
+; GCN-O0-NEXT: s_and_b32 s37, s37, s7
+; GCN-O0-NEXT: s_lshl_b32 s37, s37, s9
+; GCN-O0-NEXT: s_or_b32 s34, s34, s37
+; GCN-O0-NEXT: s_and_b32 s36, s36, s7
+; GCN-O0-NEXT: s_lshl_b32 s36, s36, s6
+; GCN-O0-NEXT: s_or_b32 s34, s34, s36
+; GCN-O0-NEXT: s_lshl_b32 s35, s35, s4
+; GCN-O0-NEXT: s_or_b32 s34, s34, s35
+; GCN-O0-NEXT: s_mov_b64 s[40:41], 3
+; GCN-O0-NEXT: s_mov_b32 s36, s2
+; GCN-O0-NEXT: s_mov_b32 s35, s3
+; GCN-O0-NEXT: s_mov_b32 s38, s40
+; GCN-O0-NEXT: s_mov_b32 s37, s41
+; GCN-O0-NEXT: s_add_u32 s36, s36, s38
+; GCN-O0-NEXT: s_addc_u32 s35, s35, s37
+; GCN-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37
+; GCN-O0-NEXT: s_mov_b32 s37, s35
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s36
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s37
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s34
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s25, s25, s7
+; GCN-O0-NEXT: s_and_b32 s33, s33, s7
+; GCN-O0-NEXT: s_lshl_b32 s33, s33, s7
+; GCN-O0-NEXT: s_or_b32 s25, s25, s33
+; GCN-O0-NEXT: s_and_b32 s31, s31, s7
+; GCN-O0-NEXT: s_lshl_b32 s31, s31, s15
+; GCN-O0-NEXT: s_or_b32 s25, s25, s31
+; GCN-O0-NEXT: s_and_b32 s30, s30, s7
+; GCN-O0-NEXT: s_lshl_b32 s30, s30, s13
+; GCN-O0-NEXT: s_or_b32 s25, s25, s30
+; GCN-O0-NEXT: s_and_b32 s29, s29, s7
+; GCN-O0-NEXT: s_lshl_b32 s29, s29, s11
+; GCN-O0-NEXT: s_or_b32 s25, s25, s29
+; GCN-O0-NEXT: s_and_b32 s28, s28, s7
+; GCN-O0-NEXT: s_lshl_b32 s28, s28, s9
+; GCN-O0-NEXT: s_or_b32 s25, s25, s28
+; GCN-O0-NEXT: s_and_b32 s27, s27, s7
+; GCN-O0-NEXT: s_lshl_b32 s27, s27, s6
+; GCN-O0-NEXT: s_or_b32 s25, s25, s27
+; GCN-O0-NEXT: s_lshl_b32 s26, s26, s4
+; GCN-O0-NEXT: s_or_b32 s25, s25, s26
+; GCN-O0-NEXT: s_mov_b64 s[30:31], 2
+; GCN-O0-NEXT: s_mov_b32 s26, s2
+; GCN-O0-NEXT: s_mov_b32 s27, s3
+; GCN-O0-NEXT: s_mov_b32 s29, s30
+; GCN-O0-NEXT: s_mov_b32 s28, s31
+; GCN-O0-NEXT: s_add_u32 s26, s26, s29
+; GCN-O0-NEXT: s_addc_u32 s28, s27, s28
+; GCN-O0-NEXT: ; kill: def $sgpr26 killed $sgpr26 def $sgpr26_sgpr27
+; GCN-O0-NEXT: s_mov_b32 s27, s28
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s25
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s17, s17, s7
+; GCN-O0-NEXT: s_and_b32 s24, s24, s7
+; GCN-O0-NEXT: s_lshl_b32 s24, s24, s7
+; GCN-O0-NEXT: s_or_b32 s17, s17, s24
+; GCN-O0-NEXT: s_and_b32 s23, s23, s7
+; GCN-O0-NEXT: s_lshl_b32 s23, s23, s15
+; GCN-O0-NEXT: s_or_b32 s17, s17, s23
+; GCN-O0-NEXT: s_and_b32 s22, s22, s7
+; GCN-O0-NEXT: s_lshl_b32 s22, s22, s13
+; GCN-O0-NEXT: s_or_b32 s17, s17, s22
+; GCN-O0-NEXT: s_and_b32 s21, s21, s7
+; GCN-O0-NEXT: s_lshl_b32 s21, s21, s11
+; GCN-O0-NEXT: s_or_b32 s17, s17, s21
+; GCN-O0-NEXT: s_and_b32 s20, s20, s7
+; GCN-O0-NEXT: s_lshl_b32 s20, s20, s9
+; GCN-O0-NEXT: s_or_b32 s17, s17, s20
+; GCN-O0-NEXT: s_and_b32 s19, s19, s7
+; GCN-O0-NEXT: s_lshl_b32 s19, s19, s6
+; GCN-O0-NEXT: s_or_b32 s17, s17, s19
+; GCN-O0-NEXT: s_lshl_b32 s18, s18, s4
+; GCN-O0-NEXT: s_or_b32 s17, s17, s18
+; GCN-O0-NEXT: s_mov_b64 s[22:23], 1
+; GCN-O0-NEXT: s_mov_b32 s18, s2
+; GCN-O0-NEXT: s_mov_b32 s19, s3
+; GCN-O0-NEXT: s_mov_b32 s21, s22
+; GCN-O0-NEXT: s_mov_b32 s20, s23
+; GCN-O0-NEXT: s_add_u32 s18, s18, s21
+; GCN-O0-NEXT: s_addc_u32 s20, s19, s20
+; GCN-O0-NEXT: ; kill: def $sgpr18 killed $sgpr18 def $sgpr18_sgpr19
+; GCN-O0-NEXT: s_mov_b32 s19, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s17
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_and_b32 s0, s0, s7
+; GCN-O0-NEXT: s_and_b32 s16, s16, s7
+; GCN-O0-NEXT: s_lshl_b32 s16, s16, s7
+; GCN-O0-NEXT: s_or_b32 s0, s0, s16
+; GCN-O0-NEXT: s_and_b32 s14, s14, s7
+; GCN-O0-NEXT: s_lshl_b32 s14, s14, s15
+; GCN-O0-NEXT: s_or_b32 s0, s0, s14
+; GCN-O0-NEXT: s_and_b32 s12, s12, s7
+; GCN-O0-NEXT: s_lshl_b32 s12, s12, s13
+; GCN-O0-NEXT: s_or_b32 s0, s0, s12
+; GCN-O0-NEXT: s_and_b32 s10, s10, s7
+; GCN-O0-NEXT: s_lshl_b32 s10, s10, s11
+; GCN-O0-NEXT: s_or_b32 s0, s0, s10
+; GCN-O0-NEXT: s_and_b32 s8, s8, s7
+; GCN-O0-NEXT: s_lshl_b32 s8, s8, s9
+; GCN-O0-NEXT: s_or_b32 s0, s0, s8
+; GCN-O0-NEXT: s_and_b32 s5, s5, s7
+; GCN-O0-NEXT: s_lshl_b32 s5, s5, s6
+; GCN-O0-NEXT: s_or_b32 s0, s0, s5
+; GCN-O0-NEXT: s_lshl_b32 s1, s1, s4
+; GCN-O0-NEXT: s_or_b32 s0, s0, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: s_endpgm
entry:
%v = insertelement <128 x i1> %vec, i1 1, i32 %sel
store <128 x i1> %v, ptr addrspace(1) %out
@@ -1912,6 +5675,171 @@ define amdgpu_ps <32 x float> @float32_inselt_vec(<32 x float> %vec, i32 %sel) {
; GCN-NEXT: v_cndmask_b32_e64 v30, 1.0, v30, s[58:59]
; GCN-NEXT: v_cndmask_b32_e64 v31, 1.0, v31, s[60:61]
; GCN-NEXT: ; return to shader part epilog
+;
+; GCN-O0-LABEL: float32_inselt_vec:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: v_mov_b32_e32 v33, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v34, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v35, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v36, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v37, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v38, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v39, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v40, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v41, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v42, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v43, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v44, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v45, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v46, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v47, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v48, v15
+; GCN-O0-NEXT: v_mov_b32_e32 v49, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v50, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v51, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v52, v19
+; GCN-O0-NEXT: v_mov_b32_e32 v53, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v54, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v55, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v56, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v57, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v58, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v59, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v60, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v61, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v62, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v63, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v64, v31
+; GCN-O0-NEXT: s_mov_b32 s0, 0
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: s_mov_b32 s2, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v6, v6, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v7, v7, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 8
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v8, v8, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 9
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v9, v9, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 10
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v10, v10, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 11
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v11, v11, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 12
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v12, v12, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 13
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v13, v13, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 14
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v14, v14, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 15
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v15, v15, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 16
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v16, v16, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 17
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v17, v17, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 18
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v18, v18, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 19
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v19, v19, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 20
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v20, v20, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 21
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v21, v21, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 22
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v22, v22, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 23
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v23, v23, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 24
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v24, v24, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 25
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v25, v25, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 26
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v26, v26, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 27
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v27, v27, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 28
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v28, v28, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 29
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v29, v29, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 30
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v30, v30, v33, s[0:1]
+; GCN-O0-NEXT: s_mov_b32 s0, 31
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], v32, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v32, s2
+; GCN-O0-NEXT: v_cndmask_b32_e64 v31, v31, v32, s[0:1]
+; GCN-O0-NEXT: ; return to shader part epilog
entry:
%v = insertelement <32 x float> %vec, float 1.000000e+00, i32 %sel
ret <32 x float> %v
@@ -1947,6 +5875,78 @@ define <8 x double> @double8_inselt_vec(<8 x double> %vec, i32 %sel) {
; GCN-NEXT: v_cndmask_b32_e64 v14, v14, 0, vcc
; GCN-NEXT: v_cndmask_b32_e32 v15, v15, v17, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: double8_inselt_vec:
+; GCN-O0: ; %bb.0: ; %entry
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v32, v15
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 1.0
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s6, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v17, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v17, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v6, v6, v17, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v7, v7, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v8, v8, v17, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v9, v9, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v10, v10, v17, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v11, v11, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v12, v12, v17, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v13, v13, v17, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v16, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s7
+; GCN-O0-NEXT: v_cndmask_b32_e64 v14, v14, v16, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s6
+; GCN-O0-NEXT: v_cndmask_b32_e64 v15, v15, v16, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
entry:
%v = insertelement <8 x double> %vec, double 1.000000e+00, i32 %sel
ret <8 x double> %v
@@ -1977,6 +5977,43 @@ define <6 x i32> @insert_dyn_i32_6(<6 x i32> inreg %arg, i32 %idx, i32 %val) {
; GCN-NEXT: v_mov_b32_e32 v0, v6
; GCN-NEXT: v_mov_b32_e32 v1, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_i32_6:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v6, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <6 x i32> %arg, i32 %val, i32 %idx
ret <6 x i32> %x
}
@@ -2011,6 +6048,48 @@ define <6 x i32> @insert_dyn_i32_6_inreg(<6 x i32> inreg %arg, i32 inreg %idx, i
; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v6
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_i32_6_inreg:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v0
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v6, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <6 x i32> %arg, i32 %val, i32 %idx
ret <6 x i32> %x
}
@@ -2040,6 +6119,43 @@ define <6 x float> @insert_dyn_float_6(<6 x float> inreg %arg, i32 %idx, float %
; GCN-NEXT: v_mov_b32_e32 v0, v6
; GCN-NEXT: v_mov_b32_e32 v1, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_float_6:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v0
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v6, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <6 x float> %arg, float %val, i32 %idx
ret <6 x float> %x
}
@@ -2074,6 +6190,48 @@ define <6 x float> @insert_dyn_float_6_inreg(<6 x float> inreg %arg, i32 inreg %
; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v6
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_float_6_inreg:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v0
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr5 killed $sgpr17
+; GCN-O0-NEXT: ; kill: def $sgpr6 killed $sgpr18
+; GCN-O0-NEXT: ; kill: def $sgpr7 killed $sgpr19
+; GCN-O0-NEXT: ; kill: def $sgpr8 killed $sgpr20
+; GCN-O0-NEXT: ; kill: def $sgpr9 killed $sgpr21
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s16
+; GCN-O0-NEXT: v_cndmask_b32_e64 v0, v0, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s17
+; GCN-O0-NEXT: v_cndmask_b32_e64 v1, v1, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s18
+; GCN-O0-NEXT: v_cndmask_b32_e64 v2, v2, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 3
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s19
+; GCN-O0-NEXT: v_cndmask_b32_e64 v3, v3, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 4
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s20
+; GCN-O0-NEXT: v_cndmask_b32_e64 v4, v4, v6, s[4:5]
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: s_cmp_eq_u32 s22, s4
+; GCN-O0-NEXT: s_cselect_b64 s[4:5], -1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s21
+; GCN-O0-NEXT: v_cndmask_b32_e64 v5, v5, v6, s[4:5]
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <6 x float> %arg, float %val, i32 %idx
ret <6 x float> %x
}
More information about the llvm-commits
mailing list