[llvm] [AMDGPU][SDAG] Add missing cases for SI_INDIRECT_SRC/DST (PR #170323)
Juan Manuel Martinez Caamaño via llvm-commits
llvm-commits at lists.llvm.org
Wed Dec 10 08:27:50 PST 2025
https://github.com/jmmartinez updated https://github.com/llvm/llvm-project/pull/170323
>From 7b8d549945f391a33866f21116c16d035fa1faec 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 1/5] 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 e1b4cad370f96..786ef4a12f1fb 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 65ccbc0f3b9f87a55e8f7903361d793b328828a3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juan=20Manuel=20Martinez=20Caama=C3=B1o?=
<jmartinezcaamao at gmail.com>
Date: Wed, 10 Dec 2025 16:51:17 +0100
Subject: [PATCH 2/5] Revert "Pre commit test: Add <6 x 32-bit> test cases for
insert/extract element with a dynamic index"
This reverts commit a57ca1f52a77e7d3bff59ffaeee93d175cbb9d55.
---
.../CodeGen/AMDGPU/extract_vector_dynelt.ll | 104 ---------------
.../CodeGen/AMDGPU/insert_vector_dynelt.ll | 126 ------------------
2 files changed, 230 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
index 42a6e4adc4e89..c69b0cce3d208 100644
--- a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
@@ -1167,107 +1167,3 @@ 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 786ef4a12f1fb..e1b4cad370f96 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
@@ -1949,129 +1949,3 @@ 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 7dc8fd498bc5703ea64bf01e9d236f5092337999 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juan=20Manuel=20Martinez=20Caama=C3=B1o?=
<jmartinezcaamao at gmail.com>
Date: Wed, 10 Dec 2025 17:04:14 +0100
Subject: [PATCH 3/5] Pre commit test: Add <(3, 5, 6, 7) x 32-bit> test cases
for insert/extract element with a dynamic index
---
.../CodeGen/AMDGPU/extract_vector_dynelt.ll | 364 ++++++++++++++
.../CodeGen/AMDGPU/insert_vector_dynelt.ll | 462 ++++++++++++++++++
2 files changed, 826 insertions(+)
diff --git a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
index c69b0cce3d208..a08de66d71819 100644
--- a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
@@ -1167,3 +1167,367 @@ 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_3(<3 x i32> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_i32_3:
+; GCN: ; %bb.0:
+; 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 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <3 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define i32 @extract_dyn_inreg_i32_3(<3 x i32> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_i32_3:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s19, 1
+; GCN-NEXT: s_cselect_b32 s4, s17, s16
+; GCN-NEXT: s_cmp_eq_u32 s19, 2
+; GCN-NEXT: s_cselect_b32 s4, s18, s4
+; GCN-NEXT: v_mov_b32_e32 v0, s4
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <3 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define float @extract_dyn_float_3(<3 x float> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_float_3:
+; GCN: ; %bb.0:
+; 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 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <3 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define float @extract_dyn_inreg_float_3(<3 x float> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_float_3:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s19, 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 s19, 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: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <3 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define i32 @extract_dyn_i32_5(<5 x i32> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_i32_5:
+; GCN: ; %bb.0:
+; 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 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <5 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define i32 @extract_dyn_inreg_i32_5(<5 x i32> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_i32_5:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s21, 1
+; GCN-NEXT: s_cselect_b32 s4, s17, s16
+; GCN-NEXT: s_cmp_eq_u32 s21, 2
+; GCN-NEXT: s_cselect_b32 s4, s18, s4
+; GCN-NEXT: s_cmp_eq_u32 s21, 3
+; GCN-NEXT: s_cselect_b32 s4, s19, s4
+; GCN-NEXT: s_cmp_eq_u32 s21, 4
+; GCN-NEXT: s_cselect_b32 s4, s20, s4
+; GCN-NEXT: v_mov_b32_e32 v0, s4
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <5 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define float @extract_dyn_float_5(<5 x float> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_float_5:
+; GCN: ; %bb.0:
+; 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 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <5 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define float @extract_dyn_inreg_float_5(<5 x float> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_float_5:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s21, 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 s21, 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 s21, 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 s21, 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: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <5 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define i32 @extract_dyn_i32_6(<6 x i32> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_i32_6:
+; GCN: ; %bb.0:
+; 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]
+ %x = extractelement <6 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define i32 @extract_dyn_inreg_i32_6(<6 x i32> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_i32_6:
+; GCN: ; %bb.0:
+; 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]
+ %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:
+; 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]
+ %x = extractelement <6 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define float @extract_dyn_inreg_float_6(<6 x float> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_float_6:
+; GCN: ; %bb.0:
+; 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]
+ %x = extractelement <6 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define i32 @extract_dyn_i32_7(<7 x i32> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_i32_7:
+; GCN: ; %bb.0:
+; 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 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s22
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 6, v0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <7 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define i32 @extract_dyn_inreg_i32_7(<7 x i32> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_i32_7:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s23, 1
+; GCN-NEXT: s_cselect_b32 s4, s17, s16
+; GCN-NEXT: s_cmp_eq_u32 s23, 2
+; GCN-NEXT: s_cselect_b32 s4, s18, s4
+; GCN-NEXT: s_cmp_eq_u32 s23, 3
+; GCN-NEXT: s_cselect_b32 s4, s19, s4
+; GCN-NEXT: s_cmp_eq_u32 s23, 4
+; GCN-NEXT: s_cselect_b32 s4, s20, s4
+; GCN-NEXT: s_cmp_eq_u32 s23, 5
+; GCN-NEXT: s_cselect_b32 s4, s21, s4
+; GCN-NEXT: s_cmp_eq_u32 s23, 6
+; GCN-NEXT: s_cselect_b32 s4, s22, s4
+; GCN-NEXT: v_mov_b32_e32 v0, s4
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <7 x i32> %arg, i32 %idx
+ ret i32 %x
+}
+
+define float @extract_dyn_float_7(<7 x float> inreg %arg, i32 %idx) {
+; GCN-LABEL: extract_dyn_float_7:
+; GCN: ; %bb.0:
+; 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 v1, v1, v2, vcc
+; GCN-NEXT: v_mov_b32_e32 v2, s22
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 6, v0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = extractelement <7 x float> %arg, i32 %idx
+ ret float %x
+}
+
+define float @extract_dyn_inreg_float_7(<7 x float> inreg %arg, i32 inreg %idx) {
+; GCN-LABEL: extract_dyn_inreg_float_7:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s23, 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 s23, 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 s23, 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 s23, 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 s23, 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: s_cmp_eq_u32 s23, 6
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v1, s22
+; 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]
+ %x = extractelement <7 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 e1b4cad370f96..95685c4bda916 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll
@@ -1949,3 +1949,465 @@ entry:
%v = insertelement <8 x double> %vec, double 1.000000e+00, i32 %sel
ret <8 x double> %v
}
+
+define <3 x i32> @insert_dyn_i32_3(<3 x i32> inreg %arg, i32 %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_i32_3:
+; 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 v4, 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 v3, 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 v0, v4
+; GCN-NEXT: v_mov_b32_e32 v1, v3
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <3 x i32> %arg, i32 %val, i32 %idx
+ ret <3 x i32> %x
+}
+
+define <3 x i32> @insert_dyn_inreg_i32_3(<3 x i32> inreg %arg, i32 inreg %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_inreg_i32_3:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s19, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s19, 1
+; GCN-NEXT: v_cndmask_b32_e32 v3, 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 s19, 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: v_cndmask_b32_e32 v2, v2, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v3
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <3 x i32> %arg, i32 %val, i32 %idx
+ ret <3 x i32> %x
+}
+
+define <3 x float> @insert_dyn_float_3(<3 x float> inreg %arg, i32 %idx, float %val) {
+; GCN-LABEL: insert_dyn_float_3:
+; 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 v4, 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 v3, 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 v0, v4
+; GCN-NEXT: v_mov_b32_e32 v1, v3
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <3 x float> %arg, float %val, i32 %idx
+ ret <3 x float> %x
+}
+
+define <3 x float> @insert_dyn_inreg_float_3(<3 x float> inreg %arg, i32 inreg %idx, float %val) {
+; GCN-LABEL: insert_dyn_inreg_float_3:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s19, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s19, 1
+; GCN-NEXT: v_cndmask_b32_e32 v3, 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 s19, 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: v_cndmask_b32_e32 v2, v2, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v3
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <3 x float> %arg, float %val, i32 %idx
+ ret <3 x float> %x
+}
+
+define <5 x i32> @insert_dyn_i32_5(<5 x i32> inreg %arg, i32 %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_i32_5:
+; 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 v5, 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 v0, v6
+; GCN-NEXT: v_mov_b32_e32 v1, v5
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <5 x i32> %arg, i32 %val, i32 %idx
+ ret <5 x i32> %x
+}
+
+define <5 x i32> @insert_dyn_inreg_i32_5(<5 x i32> inreg %arg, i32 inreg %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_inreg_i32_5:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s21, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s21, 1
+; GCN-NEXT: v_cndmask_b32_e32 v5, 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 s21, 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 s21, 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 s21, 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: v_cndmask_b32_e32 v4, v4, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v5
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <5 x i32> %arg, i32 %val, i32 %idx
+ ret <5 x i32> %x
+}
+
+define <5 x float> @insert_dyn_float_5(<5 x float> inreg %arg, i32 %idx, float %val) {
+; GCN-LABEL: insert_dyn_float_5:
+; 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 v5, 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 v0, v6
+; GCN-NEXT: v_mov_b32_e32 v1, v5
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <5 x float> %arg, float %val, i32 %idx
+ ret <5 x float> %x
+}
+
+define <5 x float> @insert_dyn_inreg_float_5(<5 x float> inreg %arg, i32 inreg %idx, float %val) {
+; GCN-LABEL: insert_dyn_inreg_float_5:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s21, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s21, 1
+; GCN-NEXT: v_cndmask_b32_e32 v5, 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 s21, 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 s21, 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 s21, 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: v_cndmask_b32_e32 v4, v4, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v5
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <5 x float> %arg, float %val, i32 %idx
+ ret <5 x float> %x
+}
+
+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_inreg_i32_6(<6 x i32> inreg %arg, i32 inreg %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_inreg_i32_6:
+; 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_inreg_float_6(<6 x float> inreg %arg, i32 inreg %idx, float %val) {
+; GCN-LABEL: insert_dyn_inreg_float_6:
+; 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
+}
+
+define <7 x i32> @insert_dyn_i32_7(<7 x i32> inreg %arg, i32 %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_i32_7:
+; 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 v8, 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 v6, s22
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 6, v0
+; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v8
+; GCN-NEXT: v_mov_b32_e32 v1, v7
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <7 x i32> %arg, i32 %val, i32 %idx
+ ret <7 x i32> %x
+}
+
+define <7 x i32> @insert_dyn_inreg_i32_7(<7 x i32> inreg %arg, i32 inreg %idx, i32 %val) {
+; GCN-LABEL: insert_dyn_inreg_i32_7:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s23, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s23, 1
+; GCN-NEXT: v_cndmask_b32_e32 v7, 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 s23, 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 s23, 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 s23, 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 s23, 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: s_cmp_eq_u32 s23, 6
+; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v6, s22
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v7
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <7 x i32> %arg, i32 %val, i32 %idx
+ ret <7 x i32> %x
+}
+
+define <7 x float> @insert_dyn_float_7(<7 x float> inreg %arg, i32 %idx, float %val) {
+; GCN-LABEL: insert_dyn_float_7:
+; 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 v8, 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 v6, s22
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 6, v0
+; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v1, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v8
+; GCN-NEXT: v_mov_b32_e32 v1, v7
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <7 x float> %arg, float %val, i32 %idx
+ ret <7 x float> %x
+}
+
+define <7 x float> @insert_dyn_inreg_float_7(<7 x float> inreg %arg, i32 inreg %idx, float %val) {
+; GCN-LABEL: insert_dyn_inreg_float_7:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_cmp_eq_u32 s23, 0
+; GCN-NEXT: v_mov_b32_e32 v1, s16
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_cmp_eq_u32 s23, 1
+; GCN-NEXT: v_cndmask_b32_e32 v7, 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 s23, 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 s23, 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 s23, 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 s23, 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: s_cmp_eq_u32 s23, 6
+; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v6, s22
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v0, vcc
+; GCN-NEXT: v_mov_b32_e32 v0, v7
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %x = insertelement <7 x float> %arg, float %val, i32 %idx
+ ret <7 x float> %x
+}
>From 27242d902a33600ac5d7129461e48428ee94ea8e 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 4/5] 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 a08de66d71819..0379ee6c7af65 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 95685c4bda916..bb514db5c2e51 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 417134d879bf48c85fcd2843aada5918b36ce289 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juan=20Manuel=20Martinez=20Caama=C3=B1o?=
<jmartinezcaamao at gmail.com>
Date: Wed, 3 Dec 2025 11:08:56 +0100
Subject: [PATCH 5/5] [AMDGPU][SDAG] Add missing cases for SI_INDIRECT_SRC/DST
Before, instruction selection would fail to select extract/insert
elements for i32/float vectors of sizes 3, 5, 6 and 7 when -O0 was used.
This patch adds the missing SI_INDIRECT_SRC/DST cases for those sizes.
---
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 8 +
llvm/lib/Target/AMDGPU/SIInstructions.td | 16 +
.../CodeGen/AMDGPU/extract_vector_dynelt.ll | 2956 ++++++++-
.../CodeGen/AMDGPU/insert_vector_dynelt.ll | 5503 ++++++++++++++++-
4 files changed, 8475 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 1d74de9cf687b..345e4701c0ed8 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6304,7 +6304,11 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
}
case AMDGPU::SI_INDIRECT_SRC_V1:
case AMDGPU::SI_INDIRECT_SRC_V2:
+ case AMDGPU::SI_INDIRECT_SRC_V3:
case AMDGPU::SI_INDIRECT_SRC_V4:
+ case AMDGPU::SI_INDIRECT_SRC_V5:
+ case AMDGPU::SI_INDIRECT_SRC_V6:
+ case AMDGPU::SI_INDIRECT_SRC_V7:
case AMDGPU::SI_INDIRECT_SRC_V8:
case AMDGPU::SI_INDIRECT_SRC_V9:
case AMDGPU::SI_INDIRECT_SRC_V10:
@@ -6315,7 +6319,11 @@ SITargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
return emitIndirectSrc(MI, *BB, *getSubtarget());
case AMDGPU::SI_INDIRECT_DST_V1:
case AMDGPU::SI_INDIRECT_DST_V2:
+ case AMDGPU::SI_INDIRECT_DST_V3:
case AMDGPU::SI_INDIRECT_DST_V4:
+ case AMDGPU::SI_INDIRECT_DST_V5:
+ case AMDGPU::SI_INDIRECT_DST_V6:
+ case AMDGPU::SI_INDIRECT_DST_V7:
case AMDGPU::SI_INDIRECT_DST_V8:
case AMDGPU::SI_INDIRECT_DST_V9:
case AMDGPU::SI_INDIRECT_DST_V10:
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 984d1a4db4cd6..643b2463344e5 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -969,7 +969,11 @@ class SI_INDIRECT_DST<RegisterClass rc> : VPseudoInstSI <
def SI_INDIRECT_SRC_V1 : SI_INDIRECT_SRC<VGPR_32>;
def SI_INDIRECT_SRC_V2 : SI_INDIRECT_SRC<VReg_64>;
+def SI_INDIRECT_SRC_V3 : SI_INDIRECT_SRC<VReg_96>;
def SI_INDIRECT_SRC_V4 : SI_INDIRECT_SRC<VReg_128>;
+def SI_INDIRECT_SRC_V5 : SI_INDIRECT_SRC<VReg_160>;
+def SI_INDIRECT_SRC_V6 : SI_INDIRECT_SRC<VReg_192>;
+def SI_INDIRECT_SRC_V7 : SI_INDIRECT_SRC<VReg_224>;
def SI_INDIRECT_SRC_V8 : SI_INDIRECT_SRC<VReg_256>;
def SI_INDIRECT_SRC_V9 : SI_INDIRECT_SRC<VReg_288>;
def SI_INDIRECT_SRC_V10 : SI_INDIRECT_SRC<VReg_320>;
@@ -980,7 +984,11 @@ def SI_INDIRECT_SRC_V32 : SI_INDIRECT_SRC<VReg_1024>;
def SI_INDIRECT_DST_V1 : SI_INDIRECT_DST<VGPR_32>;
def SI_INDIRECT_DST_V2 : SI_INDIRECT_DST<VReg_64>;
+def SI_INDIRECT_DST_V3 : SI_INDIRECT_DST<VReg_96>;
def SI_INDIRECT_DST_V4 : SI_INDIRECT_DST<VReg_128>;
+def SI_INDIRECT_DST_V5 : SI_INDIRECT_DST<VReg_160>;
+def SI_INDIRECT_DST_V6 : SI_INDIRECT_DST<VReg_192>;
+def SI_INDIRECT_DST_V7 : SI_INDIRECT_DST<VReg_224>;
def SI_INDIRECT_DST_V8 : SI_INDIRECT_DST<VReg_256>;
def SI_INDIRECT_DST_V9 : SI_INDIRECT_DST<VReg_288>;
def SI_INDIRECT_DST_V10 : SI_INDIRECT_DST<VReg_320>;
@@ -2779,7 +2787,11 @@ multiclass SI_INDIRECT_Pattern <ValueType vt, ValueType eltvt, string VecSize> {
}
defm : SI_INDIRECT_Pattern <v2f32, f32, "V2">;
+defm : SI_INDIRECT_Pattern<v3f32, f32, "V3">;
defm : SI_INDIRECT_Pattern <v4f32, f32, "V4">;
+defm : SI_INDIRECT_Pattern<v5f32, f32, "V5">;
+defm : SI_INDIRECT_Pattern<v6f32, f32, "V6">;
+defm : SI_INDIRECT_Pattern<v7f32, f32, "V7">;
defm : SI_INDIRECT_Pattern <v8f32, f32, "V8">;
defm : SI_INDIRECT_Pattern <v9f32, f32, "V9">;
defm : SI_INDIRECT_Pattern <v10f32, f32, "V10">;
@@ -2789,7 +2801,11 @@ defm : SI_INDIRECT_Pattern <v16f32, f32, "V16">;
defm : SI_INDIRECT_Pattern <v32f32, f32, "V32">;
defm : SI_INDIRECT_Pattern <v2i32, i32, "V2">;
+defm : SI_INDIRECT_Pattern<v3i32, i32, "V3">;
defm : SI_INDIRECT_Pattern <v4i32, i32, "V4">;
+defm : SI_INDIRECT_Pattern<v5i32, i32, "V5">;
+defm : SI_INDIRECT_Pattern<v6i32, i32, "V6">;
+defm : SI_INDIRECT_Pattern<v7i32, i32, "V7">;
defm : SI_INDIRECT_Pattern <v8i32, i32, "V8">;
defm : SI_INDIRECT_Pattern <v9i32, i32, "V9">;
defm : SI_INDIRECT_Pattern <v10i32, i32, "V10">;
diff --git a/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll b/llvm/test/CodeGen/AMDGPU/extract_vector_dynelt.ll
index 0379ee6c7af65..4b340f308d5f6 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 +21,30 @@ 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, 4.0
+; GCN-O0-NEXT: s_mov_b32 s4, 2.0
+; GCN-O0-NEXT: s_mov_b32 s5, 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 v6, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s3
+; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
+; 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 <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 +68,30 @@ 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[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 4
+; GCN-O0-NEXT: s_mov_b32 s4, 2
+; GCN-O0-NEXT: s_mov_b32 s5, 1
+; GCN-O0-NEXT: s_mov_b32 s6, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s3
+; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
+; 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 <4 x i32> <i32 0, i32 1, i32 2, i32 4>, i32 %sel
store i32 %ext, ptr addrspace(1) %out
@@ -74,6 +121,72 @@ 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 s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 0x40100a3d
+; GCN-O0-NEXT: s_mov_b32 s4, 0x70a3d70a
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: s_mov_b32 s12, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x4000147a
+; GCN-O0-NEXT: s_mov_b32 s4, 0xe147ae14
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s13, s5
+; GCN-O0-NEXT: s_mov_b32 s14, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 0x3ff028f5
+; GCN-O0-NEXT: s_mov_b32 s4, 0xc28f5c29
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s15, s5
+; GCN-O0-NEXT: s_mov_b32 s16, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 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, s6
+; GCN-O0-NEXT: s_mov_b32 s17, 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
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s16
+; GCN-O0-NEXT: s_mov_b32 s7, s15
+; GCN-O0-NEXT: s_mov_b32 s8, s14
+; GCN-O0-NEXT: s_mov_b32 s9, s13
+; GCN-O0-NEXT: s_mov_b32 s10, s12
+; GCN-O0-NEXT: s_mov_b32 s11, 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: 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: 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 <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 +224,113 @@ 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 s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 0x40140a3d
+; GCN-O0-NEXT: s_mov_b32 s4, 0x70a3d70a
+; GCN-O0-NEXT: s_mov_b32 s6, s4
+; GCN-O0-NEXT: s_mov_b32 s7, s3
+; GCN-O0-NEXT: s_mov_b32 s25, s7
+; GCN-O0-NEXT: s_mov_b32 s26, s6
+; GCN-O0-NEXT: s_mov_b32 s3, 0x40100a3d
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s3
+; GCN-O0-NEXT: s_mov_b32 s27, s5
+; GCN-O0-NEXT: s_mov_b32 s28, s4
+; GCN-O0-NEXT: s_mov_b32 s3, 0x4000147a
+; GCN-O0-NEXT: s_mov_b32 s4, 0xe147ae14
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s3
+; GCN-O0-NEXT: s_mov_b32 s29, s5
+; GCN-O0-NEXT: s_mov_b32 s30, s4
+; GCN-O0-NEXT: s_mov_b32 s3, 0x3ff028f5
+; GCN-O0-NEXT: s_mov_b32 s4, 0xc28f5c29
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s3
+; GCN-O0-NEXT: s_mov_b32 s31, s5
+; GCN-O0-NEXT: s_mov_b32 s33, s4
+; GCN-O0-NEXT: s_mov_b32 s3, 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, s3
+; GCN-O0-NEXT: s_mov_b32 s34, s5
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
+; GCN-O0-NEXT: ; implicit-def: $sgpr24
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr23
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr22
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; GCN-O0-NEXT: ; implicit-def: $sgpr21
+; GCN-O0-NEXT: ; implicit-def: $sgpr3
+; 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 <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 +352,25 @@ 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 +390,24 @@ 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, 1.0
+; GCN-O0-NEXT: s_mov_b32 s4, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s3
+; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v2
+; 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 <2 x float> <float 0.0, float 1.0>, i32 %sel
store float %ext, ptr addrspace(1) %out
@@ -174,6 +431,48 @@ 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 s2, s[4:5], 0x2c
+; GCN-O0-NEXT: s_mov_b32 s3, 0x3ff028f5
+; GCN-O0-NEXT: s_mov_b32 s4, 0xc28f5c29
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
+; GCN-O0-NEXT: s_mov_b32 s5, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: s_mov_b32 s8, s4
+; GCN-O0-NEXT: s_mov_b32 s6, 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, s6
+; GCN-O0-NEXT: s_mov_b32 s9, 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
+; GCN-O0-NEXT: s_mov_b32 s5, s9
+; GCN-O0-NEXT: s_mov_b32 s6, s8
+; GCN-O0-NEXT: s_mov_b32 s7, 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: 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: 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 <2 x double> <double 0.01, double 1.01>, i32 %sel
store double %ext, ptr addrspace(1) %out
@@ -219,6 +518,60 @@ 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 +603,39 @@ 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 +662,42 @@ 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 +749,101 @@ 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 +893,101 @@ 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 +1022,66 @@ 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 +1163,181 @@ 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 +1421,181 @@ 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 +1646,114 @@ 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 +1775,25 @@ 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 +1841,61 @@ 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_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_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, 15
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_and_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0
+; GCN-O0-NEXT: s_or_b32 s2, s2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 16
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:15
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 15
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:14
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 14
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:13
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 13
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:12
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 12
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:11
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 11
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:10
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 10
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:9
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 9
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:8
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 8
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:7
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 7
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:6
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 6
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:5
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 5
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 4
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 3
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 2
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: buffer_load_ubyte v2, v0, s[12:15], 0 offen
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; 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 +1916,23 @@ 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 +2208,161 @@ 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_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_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, 0x7f
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_and_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0
+; GCN-O0-NEXT: s_add_i32 s2, s2, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v1, 0
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:127
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:126
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:125
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:124
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:123
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:122
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:121
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:120
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:119
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:118
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:117
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:116
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:115
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:114
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:113
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:112
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:111
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:110
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:109
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:108
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:107
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:106
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:105
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:104
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:103
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:102
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:101
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:100
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:99
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:98
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:97
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:96
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:95
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:94
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:93
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:92
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:91
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:90
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:89
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:88
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:87
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:86
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:85
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:84
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:83
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:82
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:81
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:80
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:79
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:78
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:77
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:76
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:75
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:74
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:73
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:72
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:71
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:70
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:69
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:68
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:67
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:66
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:65
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:64
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:63
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:62
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:61
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:60
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:59
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:58
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:57
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:56
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:55
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:54
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:53
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:52
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:51
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:50
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:49
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:48
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:47
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:46
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:45
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:44
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:43
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:42
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:41
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:40
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:39
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:38
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:37
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:36
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:35
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:34
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:33
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:32
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:31
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:30
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:29
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:28
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:27
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:26
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:25
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:24
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:23
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:22
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:21
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:20
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:19
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:18
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:17
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:16
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:15
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:14
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:13
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:12
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:11
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:10
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:9
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:8
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:7
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:6
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:5
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:4
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:3
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0 offset:2
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:1
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s2
+; GCN-O0-NEXT: buffer_load_ubyte v0, v0, s[12:15], 0 offen
+; GCN-O0-NEXT: s_mov_b32 s2, 1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; 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_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 +2466,253 @@ 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: s_or_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, 0x42000000
+; GCN-O0-NEXT: s_mov_b32 s5, 0x41f80000
+; GCN-O0-NEXT: s_mov_b32 s6, 0x41f00000
+; GCN-O0-NEXT: s_mov_b32 s7, 0x41e80000
+; GCN-O0-NEXT: s_mov_b32 s8, 0x41e00000
+; GCN-O0-NEXT: s_mov_b32 s9, 0x41d80000
+; GCN-O0-NEXT: s_mov_b32 s10, 0x41d00000
+; GCN-O0-NEXT: s_mov_b32 s11, 0x41c80000
+; GCN-O0-NEXT: s_mov_b32 s12, 0x41c00000
+; GCN-O0-NEXT: s_mov_b32 s13, 0x41b80000
+; GCN-O0-NEXT: s_mov_b32 s14, 0x41b00000
+; GCN-O0-NEXT: s_mov_b32 s15, 0x41a80000
+; GCN-O0-NEXT: s_mov_b32 s16, 0x41a00000
+; GCN-O0-NEXT: s_mov_b32 s17, 0x41980000
+; GCN-O0-NEXT: s_mov_b32 s18, 0x41900000
+; GCN-O0-NEXT: s_mov_b32 s19, 0x41880000
+; GCN-O0-NEXT: s_mov_b32 s20, 0x41800000
+; GCN-O0-NEXT: s_mov_b32 s21, 0x41700000
+; GCN-O0-NEXT: s_mov_b32 s22, 0x41600000
+; GCN-O0-NEXT: s_mov_b32 s23, 0x41500000
+; GCN-O0-NEXT: s_mov_b32 s24, 0x41400000
+; GCN-O0-NEXT: s_mov_b32 s25, 0x41300000
+; GCN-O0-NEXT: s_mov_b32 s26, 0x41200000
+; GCN-O0-NEXT: s_mov_b32 s27, 0x41100000
+; GCN-O0-NEXT: s_mov_b32 s28, 0x41000000
+; GCN-O0-NEXT: s_mov_b32 s29, 0x40e00000
+; GCN-O0-NEXT: s_mov_b32 s40, 0x40c00000
+; GCN-O0-NEXT: s_mov_b32 s41, 0x40a00000
+; GCN-O0-NEXT: s_mov_b32 s42, 4.0
+; GCN-O0-NEXT: s_mov_b32 s43, 0x40400000
+; GCN-O0-NEXT: s_mov_b32 s44, 2.0
+; GCN-O0-NEXT: s_mov_b32 s45, 1.0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s45
+; GCN-O0-NEXT: v_mov_b32_e32 v62, s44
+; GCN-O0-NEXT: v_mov_b32_e32 v61, s43
+; GCN-O0-NEXT: v_mov_b32_e32 v60, s42
+; GCN-O0-NEXT: v_mov_b32_e32 v59, s41
+; GCN-O0-NEXT: v_mov_b32_e32 v58, s40
+; 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 v47, s27
+; GCN-O0-NEXT: v_mov_b32_e32 v46, s26
+; GCN-O0-NEXT: v_mov_b32_e32 v45, s25
+; GCN-O0-NEXT: v_mov_b32_e32 v44, s24
+; GCN-O0-NEXT: v_mov_b32_e32 v43, s23
+; GCN-O0-NEXT: v_mov_b32_e32 v42, s22
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s21
+; GCN-O0-NEXT: v_mov_b32_e32 v40, s20
+; GCN-O0-NEXT: v_mov_b32_e32 v55, s19
+; GCN-O0-NEXT: v_mov_b32_e32 v54, s18
+; GCN-O0-NEXT: v_mov_b32_e32 v53, s17
+; GCN-O0-NEXT: v_mov_b32_e32 v52, s16
+; GCN-O0-NEXT: v_mov_b32_e32 v51, s15
+; GCN-O0-NEXT: v_mov_b32_e32 v50, s14
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s13
+; GCN-O0-NEXT: v_mov_b32_e32 v48, s12
+; GCN-O0-NEXT: v_mov_b32_e32 v39, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v38, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v37, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v36, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v35, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v34, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v32, s4
+; 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, v47
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v42
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v40
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v55
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v54
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v53
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v52
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v51
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v50
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v49
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v48
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr63 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v63, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v63, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
+; GCN-O0-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[46:47]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB20_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
+; GCN-O0-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[46:47]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v63, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v63, 3
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v32
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v32
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v63, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v63, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
+; GCN-O0-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[46:47]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB20_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
+; GCN-O0-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[46:47]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v63, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v63, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v63, off, s[0:3], s32 offset:204 ; 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 <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 +2788,527 @@ 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 v34, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v34, s36, 0
+; GCN-O0-NEXT: v_writelane_b32 v34, s37, 1
+; GCN-O0-NEXT: v_writelane_b32 v34, s38, 2
+; GCN-O0-NEXT: v_writelane_b32 v34, s39, 3
+; GCN-O0-NEXT: v_writelane_b32 v34, s48, 4
+; GCN-O0-NEXT: v_writelane_b32 v34, s49, 5
+; GCN-O0-NEXT: v_writelane_b32 v34, s50, 6
+; GCN-O0-NEXT: v_writelane_b32 v34, s51, 7
+; GCN-O0-NEXT: v_writelane_b32 v34, s52, 8
+; GCN-O0-NEXT: v_writelane_b32 v34, s53, 9
+; GCN-O0-NEXT: v_writelane_b32 v34, s54, 10
+; GCN-O0-NEXT: v_writelane_b32 v34, s55, 11
+; GCN-O0-NEXT: v_writelane_b32 v34, s64, 12
+; GCN-O0-NEXT: v_writelane_b32 v34, s65, 13
+; GCN-O0-NEXT: v_writelane_b32 v34, s66, 14
+; GCN-O0-NEXT: v_writelane_b32 v34, s67, 15
+; GCN-O0-NEXT: s_mov_b32 s4, 0x40301999
+; GCN-O0-NEXT: s_mov_b32 s40, 0x9999999a
+; GCN-O0-NEXT: s_mov_b32 s6, s40
+; GCN-O0-NEXT: s_mov_b32 s7, s4
+; GCN-O0-NEXT: s_mov_b32 s4, s7
+; GCN-O0-NEXT: s_mov_b32 s5, s6
+; GCN-O0-NEXT: s_mov_b32 s6, 0x402e3333
+; GCN-O0-NEXT: s_mov_b32 s22, 0x33333333
+; GCN-O0-NEXT: ; implicit-def: $vgpr35 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v35, s22, 0
+; GCN-O0-NEXT: s_mov_b32 s8, s22
+; GCN-O0-NEXT: s_mov_b32 s9, s6
+; GCN-O0-NEXT: s_mov_b32 s6, s9
+; GCN-O0-NEXT: s_mov_b32 s7, s8
+; GCN-O0-NEXT: s_mov_b32 s8, 0x402c3333
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: s_mov_b32 s11, s8
+; GCN-O0-NEXT: s_mov_b32 s8, s11
+; GCN-O0-NEXT: s_mov_b32 s9, s10
+; GCN-O0-NEXT: s_mov_b32 s10, 0x402a3333
+; GCN-O0-NEXT: s_mov_b32 s12, s22
+; GCN-O0-NEXT: s_mov_b32 s13, s10
+; GCN-O0-NEXT: s_mov_b32 s10, s13
+; GCN-O0-NEXT: s_mov_b32 s11, s12
+; GCN-O0-NEXT: s_mov_b32 s12, 0x40283333
+; GCN-O0-NEXT: s_mov_b32 s14, s22
+; GCN-O0-NEXT: s_mov_b32 s15, s12
+; GCN-O0-NEXT: s_mov_b32 s12, s15
+; GCN-O0-NEXT: s_mov_b32 s13, s14
+; GCN-O0-NEXT: s_mov_b32 s14, 0x40263333
+; GCN-O0-NEXT: s_mov_b32 s16, s22
+; GCN-O0-NEXT: s_mov_b32 s17, s14
+; GCN-O0-NEXT: s_mov_b32 s14, s17
+; GCN-O0-NEXT: s_mov_b32 s15, s16
+; GCN-O0-NEXT: s_mov_b32 s16, 0x40243333
+; GCN-O0-NEXT: s_mov_b32 s18, s22
+; GCN-O0-NEXT: s_mov_b32 s19, s16
+; GCN-O0-NEXT: s_mov_b32 s16, s19
+; GCN-O0-NEXT: s_mov_b32 s17, s18
+; GCN-O0-NEXT: s_mov_b32 s18, 0x40223333
+; GCN-O0-NEXT: s_mov_b32 s20, s22
+; GCN-O0-NEXT: s_mov_b32 s21, s18
+; GCN-O0-NEXT: s_mov_b32 s18, s21
+; GCN-O0-NEXT: s_mov_b32 s19, s20
+; GCN-O0-NEXT: s_mov_b32 s20, 0x40203333
+; GCN-O0-NEXT: ; kill: def $sgpr22 killed $sgpr22 def $sgpr22_sgpr23
+; GCN-O0-NEXT: s_mov_b32 s23, s20
+; GCN-O0-NEXT: s_mov_b32 s20, s23
+; GCN-O0-NEXT: s_mov_b32 s21, s22
+; GCN-O0-NEXT: s_mov_b32 s22, 0x401c6666
+; GCN-O0-NEXT: s_mov_b32 s42, 0x66666666
+; GCN-O0-NEXT: s_mov_b32 s24, s42
+; GCN-O0-NEXT: s_mov_b32 s25, s22
+; GCN-O0-NEXT: s_mov_b32 s22, s25
+; GCN-O0-NEXT: s_mov_b32 s23, s24
+; GCN-O0-NEXT: s_mov_b32 s24, 0x40186666
+; GCN-O0-NEXT: s_mov_b32 s26, s42
+; GCN-O0-NEXT: s_mov_b32 s27, s24
+; GCN-O0-NEXT: s_mov_b32 s24, s27
+; GCN-O0-NEXT: s_mov_b32 s25, s26
+; GCN-O0-NEXT: s_mov_b32 s26, 0x40146666
+; GCN-O0-NEXT: s_mov_b32 s28, s42
+; GCN-O0-NEXT: s_mov_b32 s29, s26
+; GCN-O0-NEXT: s_mov_b32 s26, s29
+; GCN-O0-NEXT: s_mov_b32 s27, s28
+; GCN-O0-NEXT: s_mov_b32 s28, 0x40106666
+; GCN-O0-NEXT: ; kill: def $sgpr42 killed $sgpr42 def $sgpr42_sgpr43
+; GCN-O0-NEXT: s_mov_b32 s43, s28
+; GCN-O0-NEXT: s_mov_b32 s28, s43
+; GCN-O0-NEXT: s_mov_b32 s29, s42
+; GCN-O0-NEXT: s_mov_b32 s41, 0x4008cccc
+; GCN-O0-NEXT: s_mov_b32 s42, 0xcccccccd
+; GCN-O0-NEXT: s_mov_b32 s44, s42
+; GCN-O0-NEXT: s_mov_b32 s45, s41
+; GCN-O0-NEXT: s_mov_b32 s72, s45
+; GCN-O0-NEXT: s_mov_b32 s73, s44
+; GCN-O0-NEXT: s_mov_b32 s41, 0x4000cccc
+; GCN-O0-NEXT: ; kill: def $sgpr42 killed $sgpr42 def $sgpr42_sgpr43
+; GCN-O0-NEXT: s_mov_b32 s43, s41
+; GCN-O0-NEXT: s_mov_b32 s74, s43
+; GCN-O0-NEXT: s_mov_b32 s75, s42
+; GCN-O0-NEXT: s_mov_b32 s42, 0x3ff19999
+; GCN-O0-NEXT: ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41
+; GCN-O0-NEXT: s_mov_b32 s41, s42
+; GCN-O0-NEXT: s_mov_b32 s76, s41
+; GCN-O0-NEXT: s_mov_b32 s36, s40
+; 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, s76
+; GCN-O0-NEXT: s_mov_b32 s38, s75
+; GCN-O0-NEXT: s_mov_b32 s39, s74
+; GCN-O0-NEXT: s_mov_b32 s40, s73
+; GCN-O0-NEXT: s_mov_b32 s41, s72
+; GCN-O0-NEXT: s_mov_b32 s42, s29
+; GCN-O0-NEXT: s_mov_b32 s43, s28
+; GCN-O0-NEXT: s_mov_b32 s44, s27
+; GCN-O0-NEXT: s_mov_b32 s45, s26
+; GCN-O0-NEXT: s_mov_b32 s46, s25
+; GCN-O0-NEXT: s_mov_b32 s47, s24
+; GCN-O0-NEXT: s_mov_b32 s48, s23
+; GCN-O0-NEXT: s_mov_b32 s49, s22
+; GCN-O0-NEXT: s_mov_b32 s50, s21
+; GCN-O0-NEXT: s_mov_b32 s51, s20
+; GCN-O0-NEXT: s_mov_b32 s52, s19
+; GCN-O0-NEXT: s_mov_b32 s53, s18
+; GCN-O0-NEXT: s_mov_b32 s54, s17
+; GCN-O0-NEXT: s_mov_b32 s55, s16
+; GCN-O0-NEXT: s_mov_b32 s56, s15
+; GCN-O0-NEXT: s_mov_b32 s57, s14
+; GCN-O0-NEXT: s_mov_b32 s58, s13
+; GCN-O0-NEXT: s_mov_b32 s59, s12
+; GCN-O0-NEXT: s_mov_b32 s60, s11
+; GCN-O0-NEXT: s_mov_b32 s61, s10
+; GCN-O0-NEXT: s_mov_b32 s62, s9
+; GCN-O0-NEXT: s_mov_b32 s63, s8
+; GCN-O0-NEXT: s_mov_b32 s64, s7
+; GCN-O0-NEXT: s_mov_b32 s65, s6
+; GCN-O0-NEXT: s_mov_b32 s66, s5
+; GCN-O0-NEXT: s_mov_b32 s67, s4
+; GCN-O0-NEXT: v_writelane_b32 v35, s36, 1
+; GCN-O0-NEXT: v_writelane_b32 v35, s37, 2
+; GCN-O0-NEXT: v_writelane_b32 v35, s38, 3
+; GCN-O0-NEXT: v_writelane_b32 v35, s39, 4
+; GCN-O0-NEXT: v_writelane_b32 v35, s40, 5
+; GCN-O0-NEXT: v_writelane_b32 v35, s41, 6
+; GCN-O0-NEXT: v_writelane_b32 v35, s42, 7
+; GCN-O0-NEXT: v_writelane_b32 v35, s43, 8
+; GCN-O0-NEXT: v_writelane_b32 v35, s44, 9
+; GCN-O0-NEXT: v_writelane_b32 v35, s45, 10
+; GCN-O0-NEXT: v_writelane_b32 v35, s46, 11
+; GCN-O0-NEXT: v_writelane_b32 v35, s47, 12
+; GCN-O0-NEXT: v_writelane_b32 v35, s48, 13
+; GCN-O0-NEXT: v_writelane_b32 v35, s49, 14
+; GCN-O0-NEXT: v_writelane_b32 v35, s50, 15
+; GCN-O0-NEXT: v_writelane_b32 v35, s51, 16
+; GCN-O0-NEXT: v_writelane_b32 v35, s52, 17
+; GCN-O0-NEXT: v_writelane_b32 v35, s53, 18
+; GCN-O0-NEXT: v_writelane_b32 v35, s54, 19
+; GCN-O0-NEXT: v_writelane_b32 v35, s55, 20
+; GCN-O0-NEXT: v_writelane_b32 v35, s56, 21
+; GCN-O0-NEXT: v_writelane_b32 v35, s57, 22
+; GCN-O0-NEXT: v_writelane_b32 v35, s58, 23
+; GCN-O0-NEXT: v_writelane_b32 v35, s59, 24
+; GCN-O0-NEXT: v_writelane_b32 v35, s60, 25
+; GCN-O0-NEXT: v_writelane_b32 v35, s61, 26
+; GCN-O0-NEXT: v_writelane_b32 v35, s62, 27
+; GCN-O0-NEXT: v_writelane_b32 v35, s63, 28
+; GCN-O0-NEXT: v_writelane_b32 v35, s64, 29
+; GCN-O0-NEXT: v_writelane_b32 v35, s65, 30
+; GCN-O0-NEXT: v_writelane_b32 v35, s66, 31
+; GCN-O0-NEXT: v_writelane_b32 v35, s67, 32
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_lshlrev_b32_e64 v0, s4, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: v_writelane_b32 v35, s4, 33
+; GCN-O0-NEXT: v_writelane_b32 v35, s5, 34
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB21_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v35, 35
+; GCN-O0-NEXT: v_readlane_b32 s5, v35, 36
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v32
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v32
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v35, s6, 35
+; GCN-O0-NEXT: v_writelane_b32 v35, s7, 36
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB21_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v35, 33
+; GCN-O0-NEXT: v_readlane_b32 s5, v35, 34
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s36, v35, 1
+; GCN-O0-NEXT: v_readlane_b32 s37, v35, 2
+; GCN-O0-NEXT: v_readlane_b32 s38, v35, 3
+; GCN-O0-NEXT: v_readlane_b32 s39, v35, 4
+; GCN-O0-NEXT: v_readlane_b32 s40, v35, 5
+; GCN-O0-NEXT: v_readlane_b32 s41, v35, 6
+; GCN-O0-NEXT: v_readlane_b32 s42, v35, 7
+; GCN-O0-NEXT: v_readlane_b32 s43, v35, 8
+; GCN-O0-NEXT: v_readlane_b32 s44, v35, 9
+; GCN-O0-NEXT: v_readlane_b32 s45, v35, 10
+; GCN-O0-NEXT: v_readlane_b32 s46, v35, 11
+; GCN-O0-NEXT: v_readlane_b32 s47, v35, 12
+; GCN-O0-NEXT: v_readlane_b32 s48, v35, 13
+; GCN-O0-NEXT: v_readlane_b32 s49, v35, 14
+; GCN-O0-NEXT: v_readlane_b32 s50, v35, 15
+; GCN-O0-NEXT: v_readlane_b32 s51, v35, 16
+; GCN-O0-NEXT: v_readlane_b32 s52, v35, 17
+; GCN-O0-NEXT: v_readlane_b32 s53, v35, 18
+; GCN-O0-NEXT: v_readlane_b32 s54, v35, 19
+; GCN-O0-NEXT: v_readlane_b32 s55, v35, 20
+; GCN-O0-NEXT: v_readlane_b32 s56, v35, 21
+; GCN-O0-NEXT: v_readlane_b32 s57, v35, 22
+; GCN-O0-NEXT: v_readlane_b32 s58, v35, 23
+; GCN-O0-NEXT: v_readlane_b32 s59, v35, 24
+; GCN-O0-NEXT: v_readlane_b32 s60, v35, 25
+; GCN-O0-NEXT: v_readlane_b32 s61, v35, 26
+; GCN-O0-NEXT: v_readlane_b32 s62, v35, 27
+; GCN-O0-NEXT: v_readlane_b32 s63, v35, 28
+; GCN-O0-NEXT: v_readlane_b32 s64, v35, 29
+; GCN-O0-NEXT: v_readlane_b32 s65, v35, 30
+; GCN-O0-NEXT: v_readlane_b32 s66, v35, 31
+; GCN-O0-NEXT: v_readlane_b32 s67, v35, 32
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: v_writelane_b32 v35, s4, 37
+; GCN-O0-NEXT: v_writelane_b32 v35, s5, 38
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB21_4: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v35, 39
+; GCN-O0-NEXT: v_readlane_b32 s5, v35, 40
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v32
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v32
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v35, s6, 39
+; GCN-O0-NEXT: v_writelane_b32 v35, s7, 40
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB21_4
+; GCN-O0-NEXT: ; %bb.5:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[78:79], -1
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[78:79]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v35, 37
+; GCN-O0-NEXT: v_readlane_b32 s5, v35, 38
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.6:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; GCN-O0-NEXT: ; implicit-def: $sgpr4
+; GCN-O0-NEXT: ; implicit-def: $sgpr5
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; 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 s67, v34, 15
+; GCN-O0-NEXT: v_readlane_b32 s66, v34, 14
+; GCN-O0-NEXT: v_readlane_b32 s65, v34, 13
+; GCN-O0-NEXT: v_readlane_b32 s64, v34, 12
+; GCN-O0-NEXT: v_readlane_b32 s55, v34, 11
+; GCN-O0-NEXT: v_readlane_b32 s54, v34, 10
+; GCN-O0-NEXT: v_readlane_b32 s53, v34, 9
+; GCN-O0-NEXT: v_readlane_b32 s52, v34, 8
+; GCN-O0-NEXT: v_readlane_b32 s51, v34, 7
+; GCN-O0-NEXT: v_readlane_b32 s50, v34, 6
+; GCN-O0-NEXT: v_readlane_b32 s49, v34, 5
+; GCN-O0-NEXT: v_readlane_b32 s48, v34, 4
+; GCN-O0-NEXT: v_readlane_b32 s39, v34, 3
+; GCN-O0-NEXT: v_readlane_b32 s38, v34, 2
+; GCN-O0-NEXT: v_readlane_b32 s37, v34, 1
+; GCN-O0-NEXT: v_readlane_b32 s36, v34, 0
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:284 ; 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
@@ -1182,6 +3326,77 @@ define i32 @extract_dyn_i32_3(<3 x i32> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 2, 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_3:
+; GCN-O0: ; %bb.0:
+; 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 v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr5 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v5, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v5, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB22_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 3
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v3
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v5, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v5, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB22_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28 ; 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]
%x = extractelement <3 x i32> %arg, i32 %idx
ret i32 %x
}
@@ -1196,6 +3411,21 @@ define i32 @extract_dyn_inreg_i32_3(<3 x i32> inreg %arg, i32 inreg %idx) {
; GCN-NEXT: s_cselect_b32 s4, s18, s4
; GCN-NEXT: v_mov_b32_e32 v0, s4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_inreg_i32_3:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; 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: s_mov_b32 m0, s19
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = extractelement <3 x i32> %arg, i32 %idx
ret i32 %x
}
@@ -1212,6 +3442,77 @@ define float @extract_dyn_float_3(<3 x float> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 2, 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_3:
+; GCN-O0: ; %bb.0:
+; 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 v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr5 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v5, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v5, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB24_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 3
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v3
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v3
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v5, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v5, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB24_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28 ; 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]
%x = extractelement <3 x float> %arg, i32 %idx
ret float %x
}
@@ -1230,6 +3531,21 @@ define float @extract_dyn_inreg_float_3(<3 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_inreg_float_3:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; 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: s_mov_b32 m0, s19
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = extractelement <3 x float> %arg, i32 %idx
ret float %x
}
@@ -1252,6 +3568,85 @@ define i32 @extract_dyn_i32_5(<5 x i32> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 4, 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_5:
+; GCN-O0: ; %bb.0:
+; 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 v7, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr7 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v7, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v7, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB26_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v7, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v7, 3
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v5
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v7, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v7, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB26_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v7, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v7, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:36 ; 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]
%x = extractelement <5 x i32> %arg, i32 %idx
ret i32 %x
}
@@ -1270,6 +3665,25 @@ define i32 @extract_dyn_inreg_i32_5(<5 x i32> inreg %arg, i32 inreg %idx) {
; GCN-NEXT: s_cselect_b32 s4, s20, s4
; GCN-NEXT: v_mov_b32_e32 v0, s4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_inreg_i32_5:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; 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: s_mov_b32 m0, s21
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = extractelement <5 x i32> %arg, i32 %idx
ret i32 %x
}
@@ -1292,6 +3706,85 @@ define float @extract_dyn_float_5(<5 x float> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 4, 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_5:
+; GCN-O0: ; %bb.0:
+; 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 v7, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr7 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v7, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v7, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB28_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v7, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v7, 3
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v5
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v5
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v7, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v7, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB28_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v7, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v7, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:36 ; 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]
%x = extractelement <5 x float> %arg, i32 %idx
ret float %x
}
@@ -1318,13 +3811,32 @@ define float @extract_dyn_inreg_float_5(<5 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_inreg_float_5:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; 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: s_mov_b32 m0, s21
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = extractelement <5 x float> %arg, i32 %idx
ret float %x
}
define i32 @extract_dyn_i32_6(<6 x i32> inreg %arg, i32 %idx) {
; GCN-LABEL: extract_dyn_i32_6:
-; GCN: ; %bb.0:
+; 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
@@ -1343,13 +3855,97 @@ 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: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr8 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v8, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v8, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB30_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v8, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v8, 3
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v6
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v8, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v8, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB30_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v8, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v8, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:40 ; 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:
%x = extractelement <6 x i32> %arg, i32 %idx
ret i32 %x
}
define i32 @extract_dyn_inreg_i32_6(<6 x i32> inreg %arg, i32 inreg %idx) {
; GCN-LABEL: extract_dyn_inreg_i32_6:
-; GCN: ; %bb.0:
+; 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
@@ -1363,13 +3959,35 @@ define i32 @extract_dyn_inreg_i32_6(<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_inreg_i32_6:
+; 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 $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; 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: s_mov_b32 m0, s22
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-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:
+; 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
@@ -1388,13 +4006,97 @@ 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: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr8 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v8, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v8, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB32_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v8, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v8, 3
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v6
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v8, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v8, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB32_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v8, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v8, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:40 ; 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:
%x = extractelement <6 x float> %arg, i32 %idx
ret float %x
}
define float @extract_dyn_inreg_float_6(<6 x float> inreg %arg, i32 inreg %idx) {
; GCN-LABEL: extract_dyn_inreg_float_6:
-; GCN: ; %bb.0:
+; 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
@@ -1418,6 +4120,28 @@ define float @extract_dyn_inreg_float_6(<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_inreg_float_6:
+; 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 $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; 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: s_mov_b32 m0, s22
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
+entry:
%x = extractelement <6 x float> %arg, i32 %idx
ret float %x
}
@@ -1446,6 +4170,93 @@ define i32 @extract_dyn_i32_7(<7 x i32> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 6, 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_7:
+; GCN-O0: ; %bb.0:
+; 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 v9, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr9 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v9, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v9, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB34_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 3
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v7
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v9, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v9, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB34_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:44 ; 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]
%x = extractelement <7 x i32> %arg, i32 %idx
ret i32 %x
}
@@ -1468,6 +4279,29 @@ define i32 @extract_dyn_inreg_i32_7(<7 x i32> inreg %arg, i32 inreg %idx) {
; GCN-NEXT: s_cselect_b32 s4, s22, s4
; GCN-NEXT: v_mov_b32_e32 v0, s4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: extract_dyn_inreg_i32_7:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; 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: s_mov_b32 m0, s23
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = extractelement <7 x i32> %arg, i32 %idx
ret i32 %x
}
@@ -1496,6 +4330,93 @@ define float @extract_dyn_float_7(<7 x float> inreg %arg, i32 %idx) {
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 6, 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_7:
+; GCN-O0: ; %bb.0:
+; 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 v9, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; 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: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr9 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v9, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v9, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: ; implicit-def: $vgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB36_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 3
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v7
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v9, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v9, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB36_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:44 ; 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]
%x = extractelement <7 x float> %arg, i32 %idx
ret float %x
}
@@ -1530,6 +4451,29 @@ define float @extract_dyn_inreg_float_7(<7 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_inreg_float_7:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; 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: s_mov_b32 m0, s23
+; GCN-O0-NEXT: v_movrels_b32_e32 v0, v0
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = extractelement <7 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 bb514db5c2e51..beeeaa32cacfd 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 +29,25 @@ 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_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_dwordx4 s[4:7], s[2:3], 0x34
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0x44
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1.0
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s7
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_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_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 +67,24 @@ 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_load_dword s2, s[4:5], 0x2c
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1.0
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5_sgpr6_sgpr7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s7
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_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_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 +114,25 @@ 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_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_dwordx4 s[4:7], s[2:3], 0x34
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0x44
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s7
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_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_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 +157,23 @@ 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_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_dwordx2 s[4:5], s[2:3], 0x2c
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0x34
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1.0
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_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_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 +207,57 @@ 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 +311,105 @@ 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 +504,267 @@ 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 +790,30 @@ define amdgpu_kernel void @half4_inselt(ptr addrspace(1) %out, <4 x half> %vec,
; GCN-NEXT: v_mov_b32_e32 v3, s3
; 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 +835,26 @@ 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 +916,56 @@ 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_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_dwordx4 s[4:7], s[2:3], 0x34
+; GCN-O0-NEXT: s_load_dword s2, s[2:3], 0x44
+; GCN-O0-NEXT: s_mov_b32 s3, 7
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_and_b32 s2, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s3, 1
+; GCN-O0-NEXT: s_lshl_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0
+; GCN-O0-NEXT: s_or_b32 s2, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s3, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:12
+; GCN-O0-NEXT: s_mov_b32 s3, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:8
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4
+; GCN-O0-NEXT: s_mov_b32 s3, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 0x3c00
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s2
+; GCN-O0-NEXT: buffer_store_short v0, v1, s[12:15], 0 offen
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[12:15], 0 offset:4
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:12
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; 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 +987,26 @@ 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 +1032,30 @@ define amdgpu_kernel void @short4_inselt(ptr addrspace(1) %out, <4 x i16> %vec,
; GCN-NEXT: v_mov_b32_e32 v3, s3
; 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 +1080,140 @@ define amdgpu_kernel void @byte8_inselt(ptr addrspace(1) %out, <8 x i8> %vec, i3
; GCN-NEXT: v_mov_b32_e32 v3, s3
; 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 +1315,435 @@ 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_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[6:7], 52
+; 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_b64 s[6:7], 53
+; 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 v2, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s0
+; GCN-O0-NEXT: flat_load_ubyte v1, v[1:2]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 54
+; 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 v3, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s0
+; GCN-O0-NEXT: flat_load_ubyte v2, v[2:3]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 55
+; 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 v4, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s0
+; GCN-O0-NEXT: flat_load_ubyte v3, v[3:4]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 56
+; 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 v5, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s0
+; GCN-O0-NEXT: flat_load_ubyte v4, v[4:5]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 57
+; 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 v6, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s0
+; GCN-O0-NEXT: flat_load_ubyte v5, v[5:6]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 58
+; 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 v7, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s0
+; GCN-O0-NEXT: flat_load_ubyte v6, v[6:7]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 59
+; 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 v8, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s0
+; GCN-O0-NEXT: flat_load_ubyte v7, v[7:8]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 60
+; 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 v9, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s0
+; GCN-O0-NEXT: flat_load_ubyte v8, v[8:9]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 61
+; 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 v10, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s0
+; GCN-O0-NEXT: flat_load_ubyte v9, v[9:10]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 62
+; 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 v11, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s0
+; GCN-O0-NEXT: flat_load_ubyte v10, v[10:11]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 63
+; 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 v12, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s0
+; GCN-O0-NEXT: flat_load_ubyte v11, v[11:12]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 64
+; 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 v13, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s0
+; GCN-O0-NEXT: flat_load_ubyte v12, v[12:13]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x41
+; 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 v14, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v13, s0
+; GCN-O0-NEXT: flat_load_ubyte v13, v[13:14]
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 0x42
+; 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 v15, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v14, s0
+; GCN-O0-NEXT: flat_load_ubyte v14, v[14:15]
+; 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 v16, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v15, s0
+; GCN-O0-NEXT: flat_load_ubyte v15, v[15:16]
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x44
+; GCN-O0-NEXT: s_mov_b32 s3, 15
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_and_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0
+; GCN-O0-NEXT: s_or_b32 s2, s2, s3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v15, off, s[12:15], 0 offset:15
+; GCN-O0-NEXT: buffer_store_byte v14, off, s[12:15], 0 offset:14
+; GCN-O0-NEXT: buffer_store_byte v13, off, s[12:15], 0 offset:13
+; GCN-O0-NEXT: buffer_store_byte v12, off, s[12:15], 0 offset:12
+; GCN-O0-NEXT: buffer_store_byte v11, off, s[12:15], 0 offset:11
+; GCN-O0-NEXT: buffer_store_byte v10, off, s[12:15], 0 offset:10
+; GCN-O0-NEXT: buffer_store_byte v9, off, s[12:15], 0 offset:9
+; GCN-O0-NEXT: buffer_store_byte v8, off, s[12:15], 0 offset:8
+; GCN-O0-NEXT: buffer_store_byte v7, off, s[12:15], 0 offset:7
+; GCN-O0-NEXT: buffer_store_byte v6, off, s[12:15], 0 offset:6
+; GCN-O0-NEXT: buffer_store_byte v5, off, s[12:15], 0 offset:5
+; GCN-O0-NEXT: buffer_store_byte v4, off, s[12:15], 0 offset:4
+; GCN-O0-NEXT: buffer_store_byte v3, off, s[12:15], 0 offset:3
+; GCN-O0-NEXT: buffer_store_byte v2, off, s[12:15], 0 offset:2
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[12:15], 0 offset:1
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[12:15], 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v1, s2
+; GCN-O0-NEXT: buffer_store_byte v0, v1, s[12:15], 0 offen
+; GCN-O0-NEXT: buffer_load_ubyte v2, off, s[12:15], 0
+; GCN-O0-NEXT: buffer_load_ubyte v3, off, s[12:15], 0 offset:1
+; GCN-O0-NEXT: buffer_load_ubyte v4, off, s[12:15], 0 offset:2
+; GCN-O0-NEXT: buffer_load_ubyte v5, off, s[12:15], 0 offset:3
+; GCN-O0-NEXT: buffer_load_ubyte v6, off, s[12:15], 0 offset:4
+; GCN-O0-NEXT: buffer_load_ubyte v7, off, s[12:15], 0 offset:5
+; GCN-O0-NEXT: buffer_load_ubyte v8, off, s[12:15], 0 offset:6
+; GCN-O0-NEXT: buffer_load_ubyte v9, off, s[12:15], 0 offset:7
+; GCN-O0-NEXT: buffer_load_ubyte v10, off, s[12:15], 0 offset:8
+; GCN-O0-NEXT: buffer_load_ubyte v11, off, s[12:15], 0 offset:9
+; GCN-O0-NEXT: buffer_load_ubyte v12, off, s[12:15], 0 offset:10
+; GCN-O0-NEXT: buffer_load_ubyte v13, off, s[12:15], 0 offset:11
+; GCN-O0-NEXT: buffer_load_ubyte v14, off, s[12:15], 0 offset:12
+; GCN-O0-NEXT: buffer_load_ubyte v15, off, s[12:15], 0 offset:13
+; GCN-O0-NEXT: buffer_load_ubyte v16, off, s[12:15], 0 offset:14
+; GCN-O0-NEXT: buffer_load_ubyte v17, off, s[12:15], 0 offset:15
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 15
+; 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: s_waitcnt vmcnt(0)
+; 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 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_byte v[0:1], v16
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 13
+; 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_byte v[0:1], v15
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 12
+; 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_byte v[0:1], v14
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 11
+; 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_byte v[0:1], v13
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 10
+; 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_byte v[0:1], v12
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 9
+; 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_byte v[0:1], v11
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 8
+; 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_byte v[0:1], v10
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 7
+; 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_byte v[0:1], v9
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 6
+; 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_byte v[0:1], v8
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 5
+; 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_byte v[0:1], v7
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 4
+; 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_byte v[0:1], v6
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 3
+; 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_byte v[0:1], v5
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 2
+; 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_byte v[0:1], v4
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 1
+; 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_byte v[0:1], v3
+; 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 +1771,32 @@ 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 s2, s[4:5], 0x44
+; 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 v2, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s11
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s3
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v2, 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 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_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 +1851,129 @@ 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_load_dwordx2 s[2:3], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x84
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s10, s1
+; GCN-O0-NEXT: s_mov_b32 s11, s0
+; 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: $sgpr9
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr8
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr7
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; GCN-O0-NEXT: ; implicit-def: $sgpr6
+; GCN-O0-NEXT: ; implicit-def: $sgpr0
+; 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 v1, 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 v22, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v21, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v19, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s0
+; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v19
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v16, 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 v1, 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 v2, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v5
+; GCN-O0-NEXT: ; kill: def $vgpr22 killed $vgpr22 def $vgpr22_vgpr23_vgpr24_vgpr25 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v20
+; 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 v21, s1
+; GCN-O0-NEXT: v_mov_b32_e32 v20, s0
+; GCN-O0-NEXT: flat_store_dwordx4 v[20:21], v[22:25]
+; GCN-O0-NEXT: ; kill: def $vgpr19 killed $vgpr19 def $vgpr19_vgpr20_vgpr21_vgpr22 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v18, s3
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s2
+; GCN-O0-NEXT: flat_store_dwordx4 v[17:18], v[19:22]
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v9
+; 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_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 +2029,112 @@ 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 +2188,147 @@ 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 +2426,275 @@ 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 +2790,277 @@ 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 +3116,63 @@ 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 +4015,1599 @@ 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_b32 s16, SCRATCH_RSRC_DWORD0
+; GCN-O0-NEXT: s_mov_b32 s17, SCRATCH_RSRC_DWORD1
+; GCN-O0-NEXT: s_mov_b32 s18, -1
+; GCN-O0-NEXT: s_mov_b32 s19, 0xe80000
+; GCN-O0-NEXT: s_add_u32 s16, s16, s11
+; GCN-O0-NEXT: s_addc_u32 s17, s17, 0
+; GCN-O0-NEXT: s_mov_b64 s[6:7], 52
+; 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 s1, 1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:388 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:648 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v2, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v3, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v4, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v5, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v6, v0, 6, 1
+; GCN-O0-NEXT: s_mov_b32 s0, 7
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v7, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 53
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v8, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v9, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v10, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v11, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v12, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v13, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v14, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v15, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 54
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v16, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v17, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v18, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v19, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v20, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v21, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v22, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v23, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 55
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v24, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v25, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v26, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v27, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v28, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v29, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v30, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v31, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 56
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v32, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v33, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v34, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v35, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v36, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v37, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v38, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v39, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 57
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v40, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v41, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v42, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v43, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v44, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v45, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v46, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v47, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 58
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v48, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v49, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v50, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v51, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v52, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v53, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v54, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v55, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 59
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v56, v0, s1
+; GCN-O0-NEXT: v_bfe_u32 v57, v0, 1, 1
+; GCN-O0-NEXT: v_bfe_u32 v58, v0, 2, 1
+; GCN-O0-NEXT: v_bfe_u32 v59, v0, 3, 1
+; GCN-O0-NEXT: v_bfe_u32 v60, v0, 4, 1
+; GCN-O0-NEXT: v_bfe_u32 v61, v0, 5, 1
+; GCN-O0-NEXT: v_bfe_u32 v62, v0, 6, 1
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v63, s0, v0
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 60
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:392 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:396 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:400 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:404 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:408 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:412 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:416 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:420 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 61
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:424 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:428 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:432 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:436 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:440 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:444 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:448 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:452 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 62
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:456 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:460 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:464 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:468 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:472 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:476 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:480 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:484 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 63
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:488 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:492 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:496 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:500 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:504 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:508 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:512 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:516 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 64
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:520 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:524 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:528 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:532 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:536 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:540 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:544 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:548 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 0x41
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:552 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:556 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:560 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:564 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:568 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:572 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:576 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:580 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 0x42
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:584 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:588 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:592 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:596 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:600 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:604 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_bfe_u32 v1, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[16:19], 0 offset:608 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:612 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[8:9], 0x43
+; GCN-O0-NEXT: s_mov_b32 s2, s4
+; GCN-O0-NEXT: s_mov_b32 s3, s5
+; 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: buffer_load_dword v1, off, s[16:19], 0 offset:648 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, s1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:616 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_bfe_u32 v0, v0, 1, 1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:620 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_bfe_u32 v0, v0, 2, 1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:624 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_bfe_u32 v0, v0, 3, 1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:628 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_bfe_u32 v0, v0, 4, 1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:632 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_bfe_u32 v0, v0, 5, 1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:636 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_bfe_u32 v0, v0, 6, 1
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:640 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:644 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshrrev_b32_e64 v0, s0, v0
+; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GCN-O0-NEXT: s_load_dword s2, s[4:5], 0x44
+; GCN-O0-NEXT: s_mov_b32 s3, 0x7f
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: s_and_b32 s3, s2, s3
+; GCN-O0-NEXT: s_mov_b32 s2, 0
+; GCN-O0-NEXT: s_add_i32 s2, s2, s3
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:127
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:640 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:126
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:636 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:125
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:632 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:124
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:628 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:123
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:624 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:122
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:620 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:121
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:616 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:120
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:612 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:119
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:608 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:118
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:604 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:117
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:600 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:116
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:596 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:115
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:592 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:114
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:588 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:113
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:584 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:112
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:580 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:111
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:576 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:110
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:572 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:109
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:568 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:108
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:564 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:107
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:560 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:106
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:556 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:105
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:552 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:104
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:548 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:103
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:544 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:102
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:540 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:101
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:536 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:100
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:532 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:99
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:528 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:98
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:524 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:97
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:520 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:96
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:516 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:95
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:512 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:94
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:508 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:93
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:504 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:92
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:500 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:91
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:496 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:90
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:492 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:89
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:488 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:88
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:484 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:87
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:480 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:86
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:476 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:85
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:472 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:84
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:468 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:83
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:464 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:82
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:460 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:81
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:456 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:80
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:452 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:79
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:448 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:78
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:444 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:77
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:440 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:76
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:436 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:75
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:432 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:74
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:428 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:73
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:424 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:72
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:420 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:71
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:416 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:70
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:412 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:69
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:408 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:68
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:404 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:67
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:400 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:66
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:396 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:65
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:392 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0 offset:64
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:388 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_store_byte v63, off, s[16:19], 0 offset:63
+; GCN-O0-NEXT: buffer_store_byte v62, off, s[16:19], 0 offset:62
+; GCN-O0-NEXT: buffer_store_byte v61, off, s[16:19], 0 offset:61
+; GCN-O0-NEXT: buffer_store_byte v60, off, s[16:19], 0 offset:60
+; GCN-O0-NEXT: buffer_store_byte v59, off, s[16:19], 0 offset:59
+; GCN-O0-NEXT: buffer_store_byte v58, off, s[16:19], 0 offset:58
+; GCN-O0-NEXT: buffer_store_byte v57, off, s[16:19], 0 offset:57
+; GCN-O0-NEXT: buffer_store_byte v56, off, s[16:19], 0 offset:56
+; GCN-O0-NEXT: buffer_store_byte v55, off, s[16:19], 0 offset:55
+; GCN-O0-NEXT: buffer_store_byte v54, off, s[16:19], 0 offset:54
+; GCN-O0-NEXT: buffer_store_byte v53, off, s[16:19], 0 offset:53
+; GCN-O0-NEXT: buffer_store_byte v52, off, s[16:19], 0 offset:52
+; GCN-O0-NEXT: buffer_store_byte v51, off, s[16:19], 0 offset:51
+; GCN-O0-NEXT: buffer_store_byte v50, off, s[16:19], 0 offset:50
+; GCN-O0-NEXT: buffer_store_byte v49, off, s[16:19], 0 offset:49
+; GCN-O0-NEXT: buffer_store_byte v48, off, s[16:19], 0 offset:48
+; GCN-O0-NEXT: buffer_store_byte v47, off, s[16:19], 0 offset:47
+; GCN-O0-NEXT: buffer_store_byte v46, off, s[16:19], 0 offset:46
+; GCN-O0-NEXT: buffer_store_byte v45, off, s[16:19], 0 offset:45
+; GCN-O0-NEXT: buffer_store_byte v44, off, s[16:19], 0 offset:44
+; GCN-O0-NEXT: buffer_store_byte v43, off, s[16:19], 0 offset:43
+; GCN-O0-NEXT: buffer_store_byte v42, off, s[16:19], 0 offset:42
+; GCN-O0-NEXT: buffer_store_byte v41, off, s[16:19], 0 offset:41
+; GCN-O0-NEXT: buffer_store_byte v40, off, s[16:19], 0 offset:40
+; GCN-O0-NEXT: buffer_store_byte v39, off, s[16:19], 0 offset:39
+; GCN-O0-NEXT: buffer_store_byte v38, off, s[16:19], 0 offset:38
+; GCN-O0-NEXT: buffer_store_byte v37, off, s[16:19], 0 offset:37
+; GCN-O0-NEXT: buffer_store_byte v36, off, s[16:19], 0 offset:36
+; GCN-O0-NEXT: buffer_store_byte v35, off, s[16:19], 0 offset:35
+; GCN-O0-NEXT: buffer_store_byte v34, off, s[16:19], 0 offset:34
+; GCN-O0-NEXT: buffer_store_byte v33, off, s[16:19], 0 offset:33
+; GCN-O0-NEXT: buffer_store_byte v32, off, s[16:19], 0 offset:32
+; GCN-O0-NEXT: buffer_store_byte v31, off, s[16:19], 0 offset:31
+; GCN-O0-NEXT: buffer_store_byte v30, off, s[16:19], 0 offset:30
+; GCN-O0-NEXT: buffer_store_byte v29, off, s[16:19], 0 offset:29
+; GCN-O0-NEXT: buffer_store_byte v28, off, s[16:19], 0 offset:28
+; GCN-O0-NEXT: buffer_store_byte v27, off, s[16:19], 0 offset:27
+; GCN-O0-NEXT: buffer_store_byte v26, off, s[16:19], 0 offset:26
+; GCN-O0-NEXT: buffer_store_byte v25, off, s[16:19], 0 offset:25
+; GCN-O0-NEXT: buffer_store_byte v24, off, s[16:19], 0 offset:24
+; GCN-O0-NEXT: buffer_store_byte v23, off, s[16:19], 0 offset:23
+; GCN-O0-NEXT: buffer_store_byte v22, off, s[16:19], 0 offset:22
+; GCN-O0-NEXT: buffer_store_byte v21, off, s[16:19], 0 offset:21
+; GCN-O0-NEXT: buffer_store_byte v20, off, s[16:19], 0 offset:20
+; GCN-O0-NEXT: buffer_store_byte v19, off, s[16:19], 0 offset:19
+; GCN-O0-NEXT: buffer_store_byte v18, off, s[16:19], 0 offset:18
+; GCN-O0-NEXT: buffer_store_byte v17, off, s[16:19], 0 offset:17
+; GCN-O0-NEXT: buffer_store_byte v16, off, s[16:19], 0 offset:16
+; GCN-O0-NEXT: buffer_store_byte v15, off, s[16:19], 0 offset:15
+; GCN-O0-NEXT: buffer_store_byte v14, off, s[16:19], 0 offset:14
+; GCN-O0-NEXT: buffer_store_byte v13, off, s[16:19], 0 offset:13
+; GCN-O0-NEXT: buffer_store_byte v12, off, s[16:19], 0 offset:12
+; GCN-O0-NEXT: buffer_store_byte v11, off, s[16:19], 0 offset:11
+; GCN-O0-NEXT: buffer_store_byte v10, off, s[16:19], 0 offset:10
+; GCN-O0-NEXT: buffer_store_byte v9, off, s[16:19], 0 offset:9
+; GCN-O0-NEXT: buffer_store_byte v8, off, s[16:19], 0 offset:8
+; GCN-O0-NEXT: buffer_store_byte v7, off, s[16:19], 0 offset:7
+; GCN-O0-NEXT: buffer_store_byte v6, off, s[16:19], 0 offset:6
+; GCN-O0-NEXT: buffer_store_byte v5, off, s[16:19], 0 offset:5
+; GCN-O0-NEXT: buffer_store_byte v4, off, s[16:19], 0 offset:4
+; GCN-O0-NEXT: buffer_store_byte v3, off, s[16:19], 0 offset:3
+; GCN-O0-NEXT: buffer_store_byte v2, off, s[16:19], 0 offset:2
+; GCN-O0-NEXT: buffer_store_byte v1, off, s[16:19], 0 offset:1
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: buffer_store_byte v0, off, s[16:19], 0
+; 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[16:19], 0 offen
+; GCN-O0-NEXT: buffer_load_ubyte v18, off, s[16:19], 0 offset:23
+; GCN-O0-NEXT: buffer_load_ubyte v19, off, s[16:19], 0 offset:22
+; GCN-O0-NEXT: buffer_load_ubyte v20, off, s[16:19], 0 offset:21
+; GCN-O0-NEXT: buffer_load_ubyte v21, off, s[16:19], 0 offset:20
+; GCN-O0-NEXT: buffer_load_ubyte v22, off, s[16:19], 0 offset:19
+; GCN-O0-NEXT: buffer_load_ubyte v23, off, s[16:19], 0 offset:18
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:128 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v8, off, s[16:19], 0 offset:1
+; GCN-O0-NEXT: buffer_load_ubyte v7, off, s[16:19], 0 offset:2
+; GCN-O0-NEXT: buffer_load_ubyte v6, off, s[16:19], 0 offset:3
+; GCN-O0-NEXT: buffer_load_ubyte v5, off, s[16:19], 0 offset:4
+; GCN-O0-NEXT: buffer_load_ubyte v4, off, s[16:19], 0 offset:5
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:6
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:140 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:7
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:136 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v9, off, s[16:19], 0 offset:8
+; GCN-O0-NEXT: buffer_load_ubyte v16, off, s[16:19], 0 offset:9
+; GCN-O0-NEXT: buffer_load_ubyte v15, off, s[16:19], 0 offset:10
+; GCN-O0-NEXT: buffer_load_ubyte v14, off, s[16:19], 0 offset:11
+; GCN-O0-NEXT: buffer_load_ubyte v13, off, s[16:19], 0 offset:12
+; GCN-O0-NEXT: buffer_load_ubyte v12, off, s[16:19], 0 offset:13
+; GCN-O0-NEXT: buffer_load_ubyte v11, off, s[16:19], 0 offset:14
+; GCN-O0-NEXT: buffer_load_ubyte v10, off, s[16:19], 0 offset:15
+; GCN-O0-NEXT: buffer_load_ubyte v17, off, s[16:19], 0 offset:16
+; GCN-O0-NEXT: buffer_load_ubyte v24, off, s[16:19], 0 offset:17
+; GCN-O0-NEXT: buffer_load_ubyte v26, off, s[16:19], 0 offset:31
+; GCN-O0-NEXT: buffer_load_ubyte v27, off, s[16:19], 0 offset:30
+; GCN-O0-NEXT: buffer_load_ubyte v28, off, s[16:19], 0 offset:29
+; GCN-O0-NEXT: buffer_load_ubyte v29, off, s[16:19], 0 offset:28
+; GCN-O0-NEXT: buffer_load_ubyte v30, off, s[16:19], 0 offset:27
+; GCN-O0-NEXT: buffer_load_ubyte v31, off, s[16:19], 0 offset:26
+; GCN-O0-NEXT: buffer_load_ubyte v32, off, s[16:19], 0 offset:25
+; GCN-O0-NEXT: buffer_load_ubyte v25, off, s[16:19], 0 offset:24
+; GCN-O0-NEXT: buffer_load_ubyte v34, off, s[16:19], 0 offset:39
+; GCN-O0-NEXT: buffer_load_ubyte v35, off, s[16:19], 0 offset:38
+; GCN-O0-NEXT: buffer_load_ubyte v36, off, s[16:19], 0 offset:37
+; GCN-O0-NEXT: buffer_load_ubyte v37, off, s[16:19], 0 offset:36
+; GCN-O0-NEXT: buffer_load_ubyte v38, off, s[16:19], 0 offset:35
+; GCN-O0-NEXT: buffer_load_ubyte v39, off, s[16:19], 0 offset:34
+; GCN-O0-NEXT: buffer_load_ubyte v40, off, s[16:19], 0 offset:33
+; GCN-O0-NEXT: buffer_load_ubyte v33, off, s[16:19], 0 offset:32
+; GCN-O0-NEXT: buffer_load_ubyte v42, off, s[16:19], 0 offset:47
+; GCN-O0-NEXT: buffer_load_ubyte v43, off, s[16:19], 0 offset:46
+; GCN-O0-NEXT: buffer_load_ubyte v44, off, s[16:19], 0 offset:45
+; GCN-O0-NEXT: buffer_load_ubyte v45, off, s[16:19], 0 offset:44
+; GCN-O0-NEXT: buffer_load_ubyte v46, off, s[16:19], 0 offset:43
+; GCN-O0-NEXT: buffer_load_ubyte v47, off, s[16:19], 0 offset:42
+; GCN-O0-NEXT: buffer_load_ubyte v48, off, s[16:19], 0 offset:41
+; GCN-O0-NEXT: buffer_load_ubyte v41, off, s[16:19], 0 offset:40
+; GCN-O0-NEXT: buffer_load_ubyte v50, off, s[16:19], 0 offset:55
+; GCN-O0-NEXT: buffer_load_ubyte v51, off, s[16:19], 0 offset:54
+; GCN-O0-NEXT: buffer_load_ubyte v52, off, s[16:19], 0 offset:53
+; GCN-O0-NEXT: buffer_load_ubyte v53, off, s[16:19], 0 offset:52
+; GCN-O0-NEXT: buffer_load_ubyte v54, off, s[16:19], 0 offset:51
+; GCN-O0-NEXT: buffer_load_ubyte v55, off, s[16:19], 0 offset:50
+; GCN-O0-NEXT: buffer_load_ubyte v56, off, s[16:19], 0 offset:49
+; GCN-O0-NEXT: buffer_load_ubyte v49, off, s[16:19], 0 offset:48
+; GCN-O0-NEXT: buffer_load_ubyte v58, off, s[16:19], 0 offset:63
+; GCN-O0-NEXT: buffer_load_ubyte v59, off, s[16:19], 0 offset:62
+; GCN-O0-NEXT: buffer_load_ubyte v60, off, s[16:19], 0 offset:61
+; GCN-O0-NEXT: buffer_load_ubyte v61, off, s[16:19], 0 offset:60
+; GCN-O0-NEXT: buffer_load_ubyte v62, off, s[16:19], 0 offset:59
+; GCN-O0-NEXT: buffer_load_ubyte v63, off, s[16:19], 0 offset:58
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:57
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:132 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v57, off, s[16:19], 0 offset:56
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:71
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:144 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:70
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:172 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:69
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:148 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:68
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:152 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:67
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:156 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:66
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:160 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:65
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:168 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:64
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:164 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:79
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:176 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:78
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:204 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:77
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:180 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:76
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:184 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:75
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:188 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:74
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:192 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:73
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:200 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:72
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:196 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:87
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:208 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:86
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:236 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:85
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:212 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:84
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:216 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:83
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:220 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:82
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:224 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:81
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:232 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:80
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:228 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:95
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:240 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:94
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:268 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:93
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:244 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:92
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:248 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:91
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:252 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:90
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:256 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:89
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:264 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:88
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:260 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:103
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:272 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:102
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:300 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:101
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:276 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:100
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:280 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:99
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:284 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:98
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:288 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:97
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:296 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:96
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:292 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:111
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:304 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:110
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:332 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:109
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:308 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:108
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:312 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:107
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:316 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:106
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:320 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:105
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:328 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:104
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:324 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:119
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:336 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:118
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:364 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:117
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:340 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:116
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:344 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:115
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:348 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:114
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:352 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:113
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:360 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:112
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:356 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:127
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:368 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v2, off, s[16:19], 0 offset:126
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:125
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:372 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:124
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:376 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:123
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:380 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:122
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:384 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_load_ubyte v1, off, s[16:19], 0 offset:121
+; GCN-O0-NEXT: buffer_load_ubyte v0, off, s[16:19], 0 offset:120
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:384 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: s_mov_b32 s7, 2
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:380 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: s_mov_b32 s6, 3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:376 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: s_mov_b32 s5, 4
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:372 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: s_mov_b32 s4, 5
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:368 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: s_mov_b32 s3, 6
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_mov_b32 s2, 7
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 15
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:364 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:360 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:356 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:352 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:348 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:344 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:340 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:336 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 14
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:332 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:328 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:324 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:320 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:316 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:312 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:308 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:304 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 13
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:300 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:296 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:292 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:288 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:284 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:280 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:276 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:272 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 12
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:268 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:264 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:260 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:256 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:252 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:248 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:244 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:240 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 11
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:236 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:232 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:228 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:224 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:220 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:216 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:212 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:208 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 10
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:204 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:200 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:196 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:192 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:188 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:184 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:180 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:176 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 9
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:172 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:168 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:164 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, v3, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:160 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s7, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:156 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s6, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:152 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s5, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:148 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v1, v1, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s4, v1
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:144 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 8
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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: flat_store_byte v[0:1], v2
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[16:19], 0 offset:140 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[16:19], 0 offset:136 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:132 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v57, v57, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v0, v3, v0
+; GCN-O0-NEXT: v_or_b32_e64 v57, v57, v0
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 offset:128 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_and_b32_e64 v63, v63, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v63, s7, v63
+; GCN-O0-NEXT: v_or_b32_e64 v57, v57, v63
+; GCN-O0-NEXT: v_and_b32_e64 v62, v62, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v62, s6, v62
+; GCN-O0-NEXT: v_or_b32_e64 v57, v57, v62
+; GCN-O0-NEXT: v_and_b32_e64 v61, v61, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v61, s5, v61
+; GCN-O0-NEXT: v_or_b32_e64 v57, v57, v61
+; GCN-O0-NEXT: v_and_b32_e64 v60, v60, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v60, s4, v60
+; GCN-O0-NEXT: v_or_b32_e64 v57, v57, v60
+; GCN-O0-NEXT: v_and_b32_e64 v59, v59, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v59, s3, v59
+; GCN-O0-NEXT: v_or_b32_e64 v57, v57, v59
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v58, s2, v58
+; GCN-O0-NEXT: v_or_b32_e64 v59, v57, v58
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 7
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v58, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v57, s8
+; GCN-O0-NEXT: flat_store_byte v[57:58], v59
+; GCN-O0-NEXT: v_and_b32_e64 v49, v49, v3
+; GCN-O0-NEXT: v_and_b32_e64 v56, v56, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v56, v3, v56
+; GCN-O0-NEXT: v_or_b32_e64 v49, v49, v56
+; GCN-O0-NEXT: v_and_b32_e64 v55, v55, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v55, s7, v55
+; GCN-O0-NEXT: v_or_b32_e64 v49, v49, v55
+; GCN-O0-NEXT: v_and_b32_e64 v54, v54, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v54, s6, v54
+; GCN-O0-NEXT: v_or_b32_e64 v49, v49, v54
+; GCN-O0-NEXT: v_and_b32_e64 v53, v53, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v53, s5, v53
+; GCN-O0-NEXT: v_or_b32_e64 v49, v49, v53
+; GCN-O0-NEXT: v_and_b32_e64 v52, v52, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v52, s4, v52
+; GCN-O0-NEXT: v_or_b32_e64 v49, v49, v52
+; GCN-O0-NEXT: v_and_b32_e64 v51, v51, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v51, s3, v51
+; GCN-O0-NEXT: v_or_b32_e64 v49, v49, v51
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v50, s2, v50
+; GCN-O0-NEXT: v_or_b32_e64 v51, v49, v50
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 6
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v50, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v49, s8
+; GCN-O0-NEXT: flat_store_byte v[49:50], v51
+; GCN-O0-NEXT: v_and_b32_e64 v41, v41, v3
+; GCN-O0-NEXT: v_and_b32_e64 v48, v48, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v48, v3, v48
+; GCN-O0-NEXT: v_or_b32_e64 v41, v41, v48
+; GCN-O0-NEXT: v_and_b32_e64 v47, v47, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v47, s7, v47
+; GCN-O0-NEXT: v_or_b32_e64 v41, v41, v47
+; GCN-O0-NEXT: v_and_b32_e64 v46, v46, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v46, s6, v46
+; GCN-O0-NEXT: v_or_b32_e64 v41, v41, v46
+; GCN-O0-NEXT: v_and_b32_e64 v45, v45, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v45, s5, v45
+; GCN-O0-NEXT: v_or_b32_e64 v41, v41, v45
+; GCN-O0-NEXT: v_and_b32_e64 v44, v44, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v44, s4, v44
+; GCN-O0-NEXT: v_or_b32_e64 v41, v41, v44
+; GCN-O0-NEXT: v_and_b32_e64 v43, v43, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v43, s3, v43
+; GCN-O0-NEXT: v_or_b32_e64 v41, v41, v43
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v42, s2, v42
+; GCN-O0-NEXT: v_or_b32_e64 v43, v41, v42
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 5
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v42, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v41, s8
+; GCN-O0-NEXT: flat_store_byte v[41:42], v43
+; GCN-O0-NEXT: v_and_b32_e64 v33, v33, v3
+; GCN-O0-NEXT: v_and_b32_e64 v40, v40, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v40, v3, v40
+; GCN-O0-NEXT: v_or_b32_e64 v33, v33, v40
+; GCN-O0-NEXT: v_and_b32_e64 v39, v39, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v39, s7, v39
+; GCN-O0-NEXT: v_or_b32_e64 v33, v33, v39
+; GCN-O0-NEXT: v_and_b32_e64 v38, v38, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v38, s6, v38
+; GCN-O0-NEXT: v_or_b32_e64 v33, v33, v38
+; GCN-O0-NEXT: v_and_b32_e64 v37, v37, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v37, s5, v37
+; GCN-O0-NEXT: v_or_b32_e64 v33, v33, v37
+; GCN-O0-NEXT: v_and_b32_e64 v36, v36, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v36, s4, v36
+; GCN-O0-NEXT: v_or_b32_e64 v33, v33, v36
+; GCN-O0-NEXT: v_and_b32_e64 v35, v35, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v35, s3, v35
+; GCN-O0-NEXT: v_or_b32_e64 v33, v33, v35
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v34, s2, v34
+; GCN-O0-NEXT: v_or_b32_e64 v35, v33, v34
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 4
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v34, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v33, s8
+; GCN-O0-NEXT: flat_store_byte v[33:34], v35
+; GCN-O0-NEXT: v_and_b32_e64 v25, v25, v3
+; GCN-O0-NEXT: v_and_b32_e64 v32, v32, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v32, v3, v32
+; GCN-O0-NEXT: v_or_b32_e64 v25, v25, v32
+; GCN-O0-NEXT: v_and_b32_e64 v31, v31, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v31, s7, v31
+; GCN-O0-NEXT: v_or_b32_e64 v25, v25, v31
+; GCN-O0-NEXT: v_and_b32_e64 v30, v30, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v30, s6, v30
+; GCN-O0-NEXT: v_or_b32_e64 v25, v25, v30
+; GCN-O0-NEXT: v_and_b32_e64 v29, v29, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v29, s5, v29
+; GCN-O0-NEXT: v_or_b32_e64 v25, v25, v29
+; GCN-O0-NEXT: v_and_b32_e64 v28, v28, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v28, s4, v28
+; GCN-O0-NEXT: v_or_b32_e64 v25, v25, v28
+; GCN-O0-NEXT: v_and_b32_e64 v27, v27, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v27, s3, v27
+; GCN-O0-NEXT: v_or_b32_e64 v25, v25, v27
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v26, s2, v26
+; GCN-O0-NEXT: v_or_b32_e64 v27, v25, v26
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 3
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v26, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v25, s8
+; GCN-O0-NEXT: flat_store_byte v[25:26], v27
+; GCN-O0-NEXT: v_and_b32_e64 v17, v17, v3
+; GCN-O0-NEXT: v_and_b32_e64 v24, v24, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v24, v3, v24
+; GCN-O0-NEXT: v_or_b32_e64 v17, v17, v24
+; GCN-O0-NEXT: v_and_b32_e64 v23, v23, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v23, s7, v23
+; GCN-O0-NEXT: v_or_b32_e64 v17, v17, v23
+; GCN-O0-NEXT: v_and_b32_e64 v22, v22, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v22, s6, v22
+; GCN-O0-NEXT: v_or_b32_e64 v17, v17, v22
+; GCN-O0-NEXT: v_and_b32_e64 v21, v21, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v21, s5, v21
+; GCN-O0-NEXT: v_or_b32_e64 v17, v17, v21
+; GCN-O0-NEXT: v_and_b32_e64 v20, v20, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v20, s4, v20
+; GCN-O0-NEXT: v_or_b32_e64 v17, v17, v20
+; GCN-O0-NEXT: v_and_b32_e64 v19, v19, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v19, s3, v19
+; GCN-O0-NEXT: v_or_b32_e64 v17, v17, v19
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v18, s2, v18
+; GCN-O0-NEXT: v_or_b32_e64 v19, v17, v18
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 2
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v18, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v17, s8
+; GCN-O0-NEXT: flat_store_byte v[17:18], v19
+; GCN-O0-NEXT: v_and_b32_e64 v9, v9, v3
+; GCN-O0-NEXT: v_and_b32_e64 v16, v16, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v16, v3, v16
+; GCN-O0-NEXT: v_or_b32_e64 v9, v9, v16
+; GCN-O0-NEXT: v_and_b32_e64 v15, v15, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v15, s7, v15
+; GCN-O0-NEXT: v_or_b32_e64 v9, v9, v15
+; GCN-O0-NEXT: v_and_b32_e64 v14, v14, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v14, s6, v14
+; GCN-O0-NEXT: v_or_b32_e64 v9, v9, v14
+; GCN-O0-NEXT: v_and_b32_e64 v13, v13, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v13, s5, v13
+; GCN-O0-NEXT: v_or_b32_e64 v9, v9, v13
+; GCN-O0-NEXT: v_and_b32_e64 v12, v12, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v12, s4, v12
+; GCN-O0-NEXT: v_or_b32_e64 v9, v9, v12
+; GCN-O0-NEXT: v_and_b32_e64 v11, v11, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v11, s3, v11
+; GCN-O0-NEXT: v_or_b32_e64 v9, v9, v11
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v10, s2, v10
+; GCN-O0-NEXT: v_or_b32_e64 v11, v9, v10
+; GCN-O0-NEXT: s_mov_b64 s[12:13], 1
+; GCN-O0-NEXT: s_mov_b32 s8, s0
+; GCN-O0-NEXT: s_mov_b32 s9, s1
+; 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 v10, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s8
+; GCN-O0-NEXT: flat_store_byte v[9:10], v11
+; GCN-O0-NEXT: s_waitcnt vmcnt(7)
+; GCN-O0-NEXT: v_and_b32_e64 v0, v0, v3
+; GCN-O0-NEXT: v_and_b32_e64 v8, v8, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v8, v3, v8
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v8
+; GCN-O0-NEXT: v_and_b32_e64 v7, v7, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v7, s7, v7
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v7
+; GCN-O0-NEXT: v_and_b32_e64 v6, v6, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v6, s6, v6
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v6
+; GCN-O0-NEXT: v_and_b32_e64 v5, v5, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v5, s5, v5
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v5
+; GCN-O0-NEXT: v_and_b32_e64 v4, v4, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v4, s4, v4
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v4
+; GCN-O0-NEXT: v_and_b32_e64 v2, v2, v3
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v2, s3, v2
+; GCN-O0-NEXT: v_or_b32_e64 v0, v0, v2
+; GCN-O0-NEXT: v_lshlrev_b16_e64 v1, s2, v1
+; GCN-O0-NEXT: v_or_b32_e64 v2, v0, v1
+; 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 <128 x i1> %vec, i1 1, i32 %sel
store <128 x i1> %v, ptr addrspace(1) %out
@@ -1912,6 +5682,361 @@ 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: s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GCN-O0-NEXT: s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GCN-O0-NEXT: s_mov_b32 s10, -1
+; GCN-O0-NEXT: s_mov_b32 s11, 0xe80000
+; GCN-O0-NEXT: s_add_u32 s8, s8, s0
+; GCN-O0-NEXT: s_addc_u32 s9, s9, 0
+; GCN-O0-NEXT: buffer_store_dword v32, off, s[8:11], 0 offset:264 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_mov_b32_e32 v32, v31
+; GCN-O0-NEXT: v_mov_b32_e32 v33, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v34, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v35, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v36, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v37, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v38, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v39, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v40, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v41, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v42, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v43, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v44, v19
+; GCN-O0-NEXT: v_mov_b32_e32 v45, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v46, v17
+; GCN-O0-NEXT: v_mov_b32_e32 v47, v16
+; GCN-O0-NEXT: v_mov_b32_e32 v48, v15
+; GCN-O0-NEXT: v_mov_b32_e32 v49, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v50, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v51, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v52, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v53, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v54, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v55, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v56, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v57, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v58, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v59, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v60, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v61, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v62, v1
+; 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: buffer_store_dword v0, off, s[8:11], 0 offset:136 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[8:11], 0 offset:140 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[8:11], 0 offset:144 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[8:11], 0 offset:148 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[8:11], 0 offset:152 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[8:11], 0 offset:156 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[8:11], 0 offset:160 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[8:11], 0 offset:164 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[8:11], 0 offset:168 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[8:11], 0 offset:172 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[8:11], 0 offset:176 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[8:11], 0 offset:180 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[8:11], 0 offset:184 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[8:11], 0 offset:188 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[8:11], 0 offset:192 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[8:11], 0 offset:196 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[8:11], 0 offset:200 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[8:11], 0 offset:204 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[8:11], 0 offset:208 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[8:11], 0 offset:212 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[8:11], 0 offset:216 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[8:11], 0 offset:220 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[8:11], 0 offset:224 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[8:11], 0 offset:228 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[8:11], 0 offset:232 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[8:11], 0 offset:236 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[8:11], 0 offset:240 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[8:11], 0 offset:244 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[8:11], 0 offset:248 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[8:11], 0 offset:252 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[8:11], 0 offset:256 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[8:11], 0 offset:260 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_mov_b32_e32 v32, 1.0
+; GCN-O0-NEXT: buffer_store_dword v32, off, s[8:11], 0 offset:132 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr64 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v64, s0, 0
+; GCN-O0-NEXT: v_writelane_b32 v64, s1, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v64, off, s[8:11], 0 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[8:11], 0 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[8:11], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[8:11], 0 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[8:11], 0 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[8:11], 0 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[8:11], 0 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[8:11], 0 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[8:11], 0 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[8:11], 0 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[8:11], 0 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[8:11], 0 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[8:11], 0 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[8:11], 0 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[8:11], 0 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[8:11], 0 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[8:11], 0 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[8:11], 0 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[8:11], 0 offset:72 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[8:11], 0 offset:76 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[8:11], 0 offset:80 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[8:11], 0 offset:84 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[8:11], 0 offset:88 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[8:11], 0 offset:92 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[8:11], 0 offset:96 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[8:11], 0 offset:100 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[8:11], 0 offset:104 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[8:11], 0 offset:108 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[8:11], 0 offset:112 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[8:11], 0 offset:116 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[8:11], 0 offset:120 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[8:11], 0 offset:124 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[8:11], 0 offset:128 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr0_sgpr1
+; GCN-O0-NEXT: .LBB22_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v64, off, s[8:11], 0 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s0, v64, 2
+; GCN-O0-NEXT: v_readlane_b32 s1, v64, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[8:11], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[8:11], 0 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[8:11], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[8:11], 0 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[8:11], 0 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[8:11], 0 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[8:11], 0 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[8:11], 0 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[8:11], 0 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[8:11], 0 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[8:11], 0 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[8:11], 0 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[8:11], 0 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[8:11], 0 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[8:11], 0 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[8:11], 0 offset:64 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[8:11], 0 offset:68 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[8:11], 0 offset:72 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v18, off, s[8:11], 0 offset:76 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v19, off, s[8:11], 0 offset:80 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v20, off, s[8:11], 0 offset:84 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v21, off, s[8:11], 0 offset:88 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v22, off, s[8:11], 0 offset:92 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v23, off, s[8:11], 0 offset:96 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v24, off, s[8:11], 0 offset:100 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v25, off, s[8:11], 0 offset:104 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v26, off, s[8:11], 0 offset:108 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v27, off, s[8:11], 0 offset:112 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v28, off, s[8:11], 0 offset:116 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v29, off, s[8:11], 0 offset:120 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v30, off, s[8:11], 0 offset:124 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v31, off, s[8:11], 0 offset:128 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v32, off, s[8:11], 0 offset:132 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[8:11], 0 offset:264 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s2, v33
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[0:1], s2, v33
+; GCN-O0-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GCN-O0-NEXT: s_mov_b32 m0, s2
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v32
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[8:11], 0 offset:268 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[8:11], 0 offset:272 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[8:11], 0 offset:276 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[8:11], 0 offset:280 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[8:11], 0 offset:284 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[8:11], 0 offset:288 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[8:11], 0 offset:292 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[8:11], 0 offset:296 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[8:11], 0 offset:300 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[8:11], 0 offset:304 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[8:11], 0 offset:308 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[8:11], 0 offset:312 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[8:11], 0 offset:316 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[8:11], 0 offset:320 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[8:11], 0 offset:324 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[8:11], 0 offset:328 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[8:11], 0 offset:332 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[8:11], 0 offset:336 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[8:11], 0 offset:340 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[8:11], 0 offset:344 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[8:11], 0 offset:348 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[8:11], 0 offset:352 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[8:11], 0 offset:356 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[8:11], 0 offset:360 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[8:11], 0 offset:364 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[8:11], 0 offset:368 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[8:11], 0 offset:372 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[8:11], 0 offset:376 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[8:11], 0 offset:380 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[8:11], 0 offset:384 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[8:11], 0 offset:388 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[8:11], 0 offset:392 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[8:11], 0 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[8:11], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[8:11], 0 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[8:11], 0 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[8:11], 0 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[8:11], 0 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[8:11], 0 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[8:11], 0 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[8:11], 0 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[8:11], 0 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[8:11], 0 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[8:11], 0 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[8:11], 0 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[8:11], 0 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[8:11], 0 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[8:11], 0 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[8:11], 0 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v17, off, s[8:11], 0 offset:72 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v18, off, s[8:11], 0 offset:76 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v19, off, s[8:11], 0 offset:80 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v20, off, s[8:11], 0 offset:84 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v21, off, s[8:11], 0 offset:88 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v22, off, s[8:11], 0 offset:92 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v23, off, s[8:11], 0 offset:96 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v24, off, s[8:11], 0 offset:100 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v25, off, s[8:11], 0 offset:104 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v26, off, s[8:11], 0 offset:108 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v27, off, s[8:11], 0 offset:112 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v28, off, s[8:11], 0 offset:116 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v29, off, s[8:11], 0 offset:120 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v30, off, s[8:11], 0 offset:124 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v31, off, s[8:11], 0 offset:128 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[2:3], s[0:1]
+; GCN-O0-NEXT: v_writelane_b32 v64, s2, 2
+; GCN-O0-NEXT: v_writelane_b32 v64, s3, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v64, off, s[8:11], 0 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB22_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v64, off, s[8:11], 0 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s0, v64, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v64, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v31, off, s[8:11], 0 offset:268 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v32, off, s[8:11], 0 offset:272 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[8:11], 0 offset:276 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v34, off, s[8:11], 0 offset:280 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v35, off, s[8:11], 0 offset:284 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v36, off, s[8:11], 0 offset:288 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v37, off, s[8:11], 0 offset:292 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v38, off, s[8:11], 0 offset:296 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v39, off, s[8:11], 0 offset:300 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v40, off, s[8:11], 0 offset:304 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v41, off, s[8:11], 0 offset:308 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v42, off, s[8:11], 0 offset:312 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v43, off, s[8:11], 0 offset:316 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v44, off, s[8:11], 0 offset:320 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v45, off, s[8:11], 0 offset:324 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v46, off, s[8:11], 0 offset:328 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v47, off, s[8:11], 0 offset:332 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v48, off, s[8:11], 0 offset:336 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v49, off, s[8:11], 0 offset:340 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v50, off, s[8:11], 0 offset:344 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v51, off, s[8:11], 0 offset:348 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v52, off, s[8:11], 0 offset:352 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v53, off, s[8:11], 0 offset:356 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v54, off, s[8:11], 0 offset:360 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v55, off, s[8:11], 0 offset:364 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v56, off, s[8:11], 0 offset:368 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v57, off, s[8:11], 0 offset:372 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v58, off, s[8:11], 0 offset:376 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v59, off, s[8:11], 0 offset:380 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v60, off, s[8:11], 0 offset:384 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v61, off, s[8:11], 0 offset:388 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v62, off, s[8:11], 0 offset:392 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v31
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v32
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v33
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v34
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v35
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v36
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v37
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v38
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v39
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v40
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v41
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v42
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v43
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v44
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v45
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v46
+; GCN-O0-NEXT: v_mov_b32_e32 v16, v47
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v48
+; GCN-O0-NEXT: s_waitcnt vmcnt(13)
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v49
+; GCN-O0-NEXT: s_waitcnt vmcnt(12)
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v50
+; GCN-O0-NEXT: s_waitcnt vmcnt(11)
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v51
+; GCN-O0-NEXT: s_waitcnt vmcnt(10)
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v52
+; GCN-O0-NEXT: s_waitcnt vmcnt(9)
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v53
+; GCN-O0-NEXT: s_waitcnt vmcnt(8)
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v54
+; GCN-O0-NEXT: s_waitcnt vmcnt(7)
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v55
+; GCN-O0-NEXT: s_waitcnt vmcnt(6)
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v56
+; GCN-O0-NEXT: s_waitcnt vmcnt(5)
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v57
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v58
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v59
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v60
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v61
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v62
+; 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 +6072,356 @@ 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: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: v_mov_b32_e32 v17, v15
+; GCN-O0-NEXT: v_mov_b32_e32 v18, v14
+; GCN-O0-NEXT: v_mov_b32_e32 v19, v13
+; GCN-O0-NEXT: v_mov_b32_e32 v20, v12
+; GCN-O0-NEXT: v_mov_b32_e32 v21, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v22, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v23, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v24, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v25, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v26, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v27, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v28, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v29, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v30, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v31, v1
+; 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, v31
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v30
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v29
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v28
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v27
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v26
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v25
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v24
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v23
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v22
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v21
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v20
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v19
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v18
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v17
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, 1
+; GCN-O0-NEXT: v_lshlrev_b32_e64 v16, s4, v16
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], 1.0
+; GCN-O0-NEXT: ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v33, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v33, s5, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s4
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: v_writelane_b32 v33, s4, 2
+; GCN-O0-NEXT: v_writelane_b32 v33, s5, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB23_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v33, 4
+; GCN-O0-NEXT: v_readlane_b32 s5, v33, 5
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v17
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v17
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v16
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v33, s6, 4
+; GCN-O0-NEXT: v_writelane_b32 v33, s7, 5
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB23_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v33, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v33, 3
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v33, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v33, 1
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b32 s4, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v16, s4
+; GCN-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: v_writelane_b32 v33, s4, 6
+; GCN-O0-NEXT: v_writelane_b32 v33, s5, 7
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB23_4: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v33, 8
+; GCN-O0-NEXT: v_readlane_b32 s5, v33, 9
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v17
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v17
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v1, v16
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v33, s6, 8
+; GCN-O0-NEXT: v_writelane_b32 v33, s7, 9
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB23_4
+; GCN-O0-NEXT: ; %bb.5:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[10:11]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v33, 6
+; GCN-O0-NEXT: v_readlane_b32 s5, v33, 7
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.6:
+; GCN-O0-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(14)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v15
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v16
+; GCN-O0-NEXT: s_waitcnt vmcnt(13)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v17
+; GCN-O0-NEXT: s_waitcnt vmcnt(12)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v18
+; GCN-O0-NEXT: s_waitcnt vmcnt(11)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v19
+; GCN-O0-NEXT: s_waitcnt vmcnt(10)
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v20
+; GCN-O0-NEXT: s_waitcnt vmcnt(9)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v21
+; GCN-O0-NEXT: s_waitcnt vmcnt(8)
+; GCN-O0-NEXT: v_mov_b32_e32 v7, v22
+; GCN-O0-NEXT: s_waitcnt vmcnt(7)
+; GCN-O0-NEXT: v_mov_b32_e32 v8, v23
+; GCN-O0-NEXT: s_waitcnt vmcnt(6)
+; GCN-O0-NEXT: v_mov_b32_e32 v9, v24
+; GCN-O0-NEXT: s_waitcnt vmcnt(5)
+; GCN-O0-NEXT: v_mov_b32_e32 v10, v25
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v11, v26
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v12, v27
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v13, v28
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v14, v29
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v15, v30
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:336 ; 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:
%v = insertelement <8 x double> %vec, double 1.000000e+00, i32 %sel
ret <8 x double> %v
@@ -1968,6 +6443,89 @@ define <3 x i32> @insert_dyn_i32_3(<3 x i32> inreg %arg, i32 %idx, i32 %val) {
; GCN-NEXT: v_mov_b32_e32 v0, v4
; GCN-NEXT: v_mov_b32_e32 v1, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_i32_3:
+; GCN-O0: ; %bb.0:
+; 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 v5, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr5 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v5, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v5, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB24_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v4
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v3
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v5, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v5, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB24_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v4
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:36 ; 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]
%x = insertelement <3 x i32> %arg, i32 %val, i32 %idx
ret <3 x i32> %x
}
@@ -1990,6 +6548,24 @@ define <3 x i32> @insert_dyn_inreg_i32_3(<3 x i32> inreg %arg, i32 inreg %idx, i
; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_inreg_i32_3:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: s_mov_b32 m0, s19
+; GCN-O0-NEXT: v_movreld_b32_e32 v2, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v4
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <3 x i32> %arg, i32 %val, i32 %idx
ret <3 x i32> %x
}
@@ -2010,6 +6586,89 @@ define <3 x float> @insert_dyn_float_3(<3 x float> inreg %arg, i32 %idx, float %
; GCN-NEXT: v_mov_b32_e32 v0, v4
; GCN-NEXT: v_mov_b32_e32 v1, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_float_3:
+; GCN-O0: ; %bb.0:
+; 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 v5, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr5 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v5, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v5, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB26_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v4
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v4
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v3
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v5, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v5, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB26_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v5, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v5, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v4
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:36 ; 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]
%x = insertelement <3 x float> %arg, float %val, i32 %idx
ret <3 x float> %x
}
@@ -2032,6 +6691,24 @@ define <3 x float> @insert_dyn_inreg_float_3(<3 x float> inreg %arg, i32 inreg %
; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_inreg_float_3:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: ; kill: def $sgpr8_sgpr9_sgpr10 killed $sgpr4_sgpr5_sgpr6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: s_mov_b32 m0, s19
+; GCN-O0-NEXT: v_movreld_b32_e32 v2, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v4
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <3 x float> %arg, float %val, i32 %idx
ret <3 x float> %x
}
@@ -2058,6 +6735,107 @@ define <5 x i32> @insert_dyn_i32_5(<5 x i32> inreg %arg, i32 %idx, i32 %val) {
; GCN-NEXT: v_mov_b32_e32 v0, v6
; GCN-NEXT: v_mov_b32_e32 v1, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_i32_5:
+; GCN-O0: ; %bb.0:
+; 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 v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr9 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v9, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v9, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB28_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v6
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v5
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v9, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v9, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB28_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v4
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v5
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v7
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v8
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:52 ; 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]
%x = insertelement <5 x i32> %arg, i32 %val, i32 %idx
ret <5 x i32> %x
}
@@ -2088,6 +6866,30 @@ define <5 x i32> @insert_dyn_inreg_i32_5(<5 x i32> inreg %arg, i32 inreg %idx, i
; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_inreg_i32_5:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s8
+; GCN-O0-NEXT: s_mov_b32 m0, s21
+; GCN-O0-NEXT: v_movreld_b32_e32 v4, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v8
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <5 x i32> %arg, i32 %val, i32 %idx
ret <5 x i32> %x
}
@@ -2114,6 +6916,107 @@ define <5 x float> @insert_dyn_float_5(<5 x float> inreg %arg, i32 %idx, float %
; GCN-NEXT: v_mov_b32_e32 v0, v6
; GCN-NEXT: v_mov_b32_e32 v1, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_float_5:
+; GCN-O0: ; %bb.0:
+; 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 v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr9 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v9, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v9, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB30_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v6
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v5
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v9, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v9, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB30_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v9, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v9, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v4
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v5
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v7
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v8
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:52 ; 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]
%x = insertelement <5 x float> %arg, float %val, i32 %idx
ret <5 x float> %x
}
@@ -2144,6 +7047,30 @@ define <5 x float> @insert_dyn_inreg_float_5(<5 x float> inreg %arg, i32 inreg %
; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_inreg_float_5:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s8
+; GCN-O0-NEXT: s_mov_b32 m0, s21
+; GCN-O0-NEXT: v_movreld_b32_e32 v4, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v8
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <5 x float> %arg, float %val, i32 %idx
ret <5 x float> %x
}
@@ -2173,6 +7100,116 @@ 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: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr11 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v11, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v11, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB32_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v11, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v11, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v7
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v6
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v11, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v11, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB32_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v11, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v11, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(5)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v5
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v8
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v9
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v10
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:60 ; 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]
%x = insertelement <6 x i32> %arg, i32 %val, i32 %idx
ret <6 x i32> %x
}
@@ -2207,6 +7244,33 @@ define <6 x i32> @insert_dyn_inreg_i32_6(<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_inreg_i32_6:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s4
+; GCN-O0-NEXT: s_mov_b32 m0, s22
+; GCN-O0-NEXT: v_movreld_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v10
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <6 x i32> %arg, i32 %val, i32 %idx
ret <6 x i32> %x
}
@@ -2236,6 +7300,116 @@ 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: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr11 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v11, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v11, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB34_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v11, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v11, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v7
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v7
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v6
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v11, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v11, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB34_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[22:23], -1
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[22:23]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v11, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v11, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(5)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v5
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v8
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v9
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v10
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:60 ; 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]
%x = insertelement <6 x float> %arg, float %val, i32 %idx
ret <6 x float> %x
}
@@ -2270,6 +7444,33 @@ define <6 x float> @insert_dyn_inreg_float_6(<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_inreg_float_6:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s4
+; GCN-O0-NEXT: s_mov_b32 m0, s22
+; GCN-O0-NEXT: v_movreld_b32_e32 v5, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v5
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v10
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <6 x float> %arg, float %val, i32 %idx
ret <6 x float> %x
}
@@ -2302,6 +7503,125 @@ define <7 x i32> @insert_dyn_i32_7(<7 x i32> inreg %arg, i32 %idx, i32 %val) {
; GCN-NEXT: v_mov_b32_e32 v0, v8
; GCN-NEXT: v_mov_b32_e32 v1, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_i32_7:
+; GCN-O0: ; %bb.0:
+; 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 v13, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr13 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v13, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v13, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB36_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v13, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v13, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v8
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v8
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v7
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v13, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v13, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB36_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v13, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v13, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(6)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(5)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v7
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v8
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v9
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v10
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v11
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:68 ; 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]
%x = insertelement <7 x i32> %arg, i32 %val, i32 %idx
ret <7 x i32> %x
}
@@ -2340,6 +7660,36 @@ define <7 x i32> @insert_dyn_inreg_i32_7(<7 x i32> inreg %arg, i32 inreg %idx, i
; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_inreg_i32_7:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
+; GCN-O0-NEXT: s_mov_b32 m0, s23
+; GCN-O0-NEXT: v_movreld_b32_e32 v6, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <7 x i32> %arg, i32 %val, i32 %idx
ret <7 x i32> %x
}
@@ -2372,6 +7722,125 @@ define <7 x float> @insert_dyn_float_7(<7 x float> inreg %arg, i32 %idx, float %
; GCN-NEXT: v_mov_b32_e32 v0, v8
; GCN-NEXT: v_mov_b32_e32 v1, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_float_7:
+; GCN-O0: ; %bb.0:
+; 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 v13, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; 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: s_mov_b64 s[4:5], exec
+; GCN-O0-NEXT: ; implicit-def: $vgpr13 : SGPR spill to VGPR lane
+; GCN-O0-NEXT: v_writelane_b32 v13, s4, 0
+; GCN-O0-NEXT: v_writelane_b32 v13, s5, 1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
+; GCN-O0-NEXT: .LBB38_1: ; =>This Inner Loop Header: Depth=1
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v13, 2
+; GCN-O0-NEXT: v_readlane_b32 s5, v13, 3
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readfirstlane_b32 s6, v8
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], s6, v8
+; GCN-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GCN-O0-NEXT: s_mov_b32 m0, s6
+; GCN-O0-NEXT: v_movreld_b32_e32 v0, v7
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
+; GCN-O0-NEXT: v_writelane_b32 v13, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v13, s7, 3
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GCN-O0-NEXT: s_cbranch_execnz .LBB38_1
+; GCN-O0-NEXT: ; %bb.2:
+; GCN-O0-NEXT: s_or_saveexec_b64 s[26:27], -1
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_mov_b64 exec, s[26:27]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_readlane_b32 s4, v13, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v13, 1
+; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-O0-NEXT: ; %bb.3:
+; GCN-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt vmcnt(6)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v6
+; GCN-O0-NEXT: s_waitcnt vmcnt(5)
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v7
+; GCN-O0-NEXT: s_waitcnt vmcnt(4)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v8
+; GCN-O0-NEXT: s_waitcnt vmcnt(3)
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v9
+; GCN-O0-NEXT: s_waitcnt vmcnt(2)
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v10
+; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v11
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:68 ; 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]
%x = insertelement <7 x float> %arg, float %val, i32 %idx
ret <7 x float> %x
}
@@ -2410,6 +7879,36 @@ define <7 x float> @insert_dyn_inreg_float_7(<7 x float> inreg %arg, i32 inreg %
; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v0, vcc
; GCN-NEXT: v_mov_b32_e32 v0, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GCN-O0-LABEL: insert_dyn_inreg_float_7:
+; GCN-O0: ; %bb.0:
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-O0-NEXT: s_mov_b32 s4, s16
+; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: s_mov_b32 s5, s17
+; GCN-O0-NEXT: s_mov_b32 s6, s18
+; GCN-O0-NEXT: s_mov_b32 s7, s19
+; GCN-O0-NEXT: s_mov_b32 s8, s20
+; GCN-O0-NEXT: s_mov_b32 s9, s21
+; GCN-O0-NEXT: s_mov_b32 s10, s22
+; GCN-O0-NEXT: ; kill: def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18 killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+; GCN-O0-NEXT: v_mov_b32_e32 v12, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v11, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v10, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v9, s7
+; GCN-O0-NEXT: v_mov_b32_e32 v8, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v7, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
+; GCN-O0-NEXT: s_mov_b32 m0, s23
+; GCN-O0-NEXT: v_movreld_b32_e32 v6, v0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v6
+; GCN-O0-NEXT: v_mov_b32_e32 v1, v7
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, v10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, v11
+; GCN-O0-NEXT: v_mov_b32_e32 v6, v12
+; GCN-O0-NEXT: s_setpc_b64 s[30:31]
%x = insertelement <7 x float> %arg, float %val, i32 %idx
ret <7 x float> %x
}
More information about the llvm-commits
mailing list