[llvm] [AMDGPU] Use v_mov_b64 in codegen on gfx1250 (PR #148272)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jul 11 11:30:12 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Stanislav Mekhanoshin (rampitec)
<details>
<summary>Changes</summary>
---
Patch is 31.35 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/148272.diff
5 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/GCNSubtarget.h (+1-1)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/copy_phys_vgpr64.mir (+103)
- (modified) llvm/test/CodeGen/AMDGPU/global-load-xcnt.ll (+45-52)
- (modified) llvm/test/CodeGen/AMDGPU/v_mov_b64_expansion.mir (+18-9)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 8ea60871b6613..7d3b821825e82 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -1154,7 +1154,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
bool hasMadF16() const;
- bool hasMovB64() const { return GFX940Insts; }
+ bool hasMovB64() const { return GFX940Insts || GFX1250Insts; }
bool hasLshlAddU64Inst() const { return HasLshlAddU64Inst; }
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index ec7ef66f2c1aa..1bf152bb34851 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2214,7 +2214,7 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
if (ST.hasMovB64()) {
MI.setDesc(get(AMDGPU::V_MOV_B64_e32));
if (SrcOp.isReg() || isInlineConstant(MI, 1) ||
- isUInt<32>(SrcOp.getImm()))
+ isUInt<32>(SrcOp.getImm()) || ST.has64BitLiterals())
break;
}
if (SrcOp.isImm()) {
diff --git a/llvm/test/CodeGen/AMDGPU/copy_phys_vgpr64.mir b/llvm/test/CodeGen/AMDGPU/copy_phys_vgpr64.mir
index d5dfb5dd0848d..cc976fe13c47c 100644
--- a/llvm/test/CodeGen/AMDGPU/copy_phys_vgpr64.mir
+++ b/llvm/test/CodeGen/AMDGPU/copy_phys_vgpr64.mir
@@ -4,6 +4,7 @@
# RUN: llc -mtriple=amdgcn -mcpu=gfx942 -run-pass postrapseudos -verify-machineinstrs %s -o - | FileCheck -check-prefix=GFX942 %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass postrapseudos -verify-machineinstrs %s -o - | FileCheck -check-prefix=GFX10 %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -run-pass postrapseudos -verify-machineinstrs %s -o - | FileCheck -check-prefix=GFX10 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -run-pass postrapseudos %s -o - | FileCheck -check-prefix=GFX1250 %s
---
name: copy_v64_to_v64
@@ -32,6 +33,10 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr2_vgpr3
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit killed $vgpr2_vgpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_v64_to_v64
+ ; GFX1250: liveins: $vgpr2_vgpr3
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 killed $vgpr2_vgpr3, implicit $exec, implicit $exec
$vgpr0_vgpr1 = COPY killed $vgpr2_vgpr3, implicit $exec
...
@@ -62,6 +67,10 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $sgpr2, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $sgpr2_sgpr3
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr3, implicit $exec, implicit killed $sgpr2_sgpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_s64_to_v64
+ ; GFX1250: liveins: $sgpr2_sgpr3
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 killed $sgpr2_sgpr3, implicit $exec, implicit $exec
$vgpr0_vgpr1 = COPY killed $sgpr2_sgpr3, implicit $exec
...
@@ -94,6 +103,11 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $agpr2_agpr3
; GFX10-NEXT: $vgpr1 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec, implicit killed $agpr2_agpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_a64_to_v64
+ ; GFX1250: liveins: $agpr2_agpr3
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $agpr2_agpr3
+ ; GFX1250-NEXT: $vgpr1 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec, implicit killed $agpr2_agpr3, implicit $exec
$vgpr0_vgpr1 = COPY killed $agpr2_agpr3, implicit $exec
...
@@ -130,6 +144,11 @@ body: |
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit $vgpr2_vgpr3_vgpr4_vgpr5
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr4, implicit $exec, implicit $vgpr2_vgpr3_vgpr4_vgpr5
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr5, implicit $exec, implicit $vgpr2_vgpr3_vgpr4_vgpr5, implicit $exec
+ ; GFX1250-LABEL: name: copy_v128_to_v128_fwd
+ ; GFX1250: liveins: $vgpr2_vgpr3_vgpr4_vgpr5
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $vgpr2_vgpr3, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr2_vgpr3_vgpr4_vgpr5
+ ; GFX1250-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 $vgpr4_vgpr5, implicit $exec, implicit $vgpr2_vgpr3_vgpr4_vgpr5, implicit $exec
$vgpr0_vgpr1_vgpr2_vgpr3 = COPY killed $vgpr2_vgpr3_vgpr4_vgpr5, implicit $exec
...
@@ -166,6 +185,11 @@ body: |
; GFX10-NEXT: $vgpr4 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr1, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_v128_to_v128_back
+ ; GFX1250: liveins: $vgpr0_vgpr1_vgpr2_vgpr3
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $vgpr2_vgpr3, implicit $exec, implicit-def $vgpr2_vgpr3_vgpr4_vgpr5, implicit $vgpr0_vgpr1_vgpr2_vgpr3
+ ; GFX1250-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 $vgpr0_vgpr1, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3, implicit $exec
$vgpr2_vgpr3_vgpr4_vgpr5 = COPY killed $vgpr0_vgpr1_vgpr2_vgpr3, implicit $exec
...
@@ -202,6 +226,12 @@ body: |
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr4, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr4_vgpr5_vgpr6
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr5, implicit $exec, implicit $vgpr4_vgpr5_vgpr6
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr6, implicit $exec, implicit killed $vgpr4_vgpr5_vgpr6, implicit $exec
+ ; GFX1250-LABEL: name: copy_v96_to_v96
+ ; GFX1250: liveins: $vgpr4_vgpr5_vgpr6
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr4, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr4_vgpr5_vgpr6
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr5, implicit $exec, implicit $vgpr4_vgpr5_vgpr6
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr6, implicit $exec, implicit killed $vgpr4_vgpr5_vgpr6, implicit $exec
$vgpr0_vgpr1_vgpr2 = COPY killed $vgpr4_vgpr5_vgpr6, implicit $exec
...
@@ -232,6 +262,10 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr2_vgpr3
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit killed $vgpr2_vgpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_v64_to_v64_undef_sub0
+ ; GFX1250: liveins: $vgpr3
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 killed $vgpr2_vgpr3, implicit $exec, implicit $exec
$vgpr0_vgpr1 = COPY killed $vgpr2_vgpr3, implicit $exec
...
@@ -262,6 +296,10 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr2_vgpr3
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit killed $vgpr2_vgpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_v64_to_v64_undef_sub1
+ ; GFX1250: liveins: $vgpr2
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 killed $vgpr2_vgpr3, implicit $exec, implicit $exec
$vgpr0_vgpr1 = COPY killed $vgpr2_vgpr3, implicit $exec
...
@@ -298,6 +336,11 @@ body: |
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr5, implicit $exec, implicit $sgpr4_sgpr5_sgpr6_sgpr7
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr6, implicit $exec, implicit $sgpr4_sgpr5_sgpr6_sgpr7
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $sgpr7, implicit $exec, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7
+ ; GFX1250-LABEL: name: copy_s128_to_v128_killed
+ ; GFX1250: liveins: $sgpr4_sgpr5_sgpr6_sgpr7
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr4_sgpr5, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr4_sgpr5_sgpr6_sgpr7
+ ; GFX1250-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 $sgpr6_sgpr7, implicit $exec, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7
$vgpr0_vgpr1_vgpr2_vgpr3 = COPY killed $sgpr4_sgpr5_sgpr6_sgpr7
...
@@ -330,6 +373,11 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit-def $vgpr1_vgpr2, implicit $vgpr2_vgpr3
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit $vgpr2_vgpr3, implicit $exec
+ ; GFX1250-LABEL: name: copy_v64_to_v64_unaligned
+ ; GFX1250: liveins: $vgpr2_vgpr3
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr2, implicit $exec, implicit-def $vgpr1_vgpr2, implicit $vgpr2_vgpr3
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit $vgpr2_vgpr3, implicit $exec
$vgpr1_vgpr2 = COPY killed $vgpr2_vgpr3, implicit $exec
...
@@ -362,6 +410,11 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr3_vgpr4
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr4, implicit $exec, implicit killed $vgpr3_vgpr4, implicit $exec
+ ; GFX1250-LABEL: name: copy_v64_unaligned_to_v64
+ ; GFX1250: liveins: $vgpr3_vgpr4
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr3, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr3_vgpr4
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr4, implicit $exec, implicit killed $vgpr3_vgpr4, implicit $exec
$vgpr0_vgpr1 = COPY killed $vgpr3_vgpr4, implicit $exec
...
@@ -402,6 +455,13 @@ body: |
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit $vgpr8_vgpr9_vgpr10_vgpr11
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr10, implicit $exec, implicit $vgpr8_vgpr9_vgpr10_vgpr11
; GFX10-NEXT: $vgpr4 = V_MOV_B32_e32 $vgpr11, implicit $exec, implicit killed $vgpr8_vgpr9_vgpr10_vgpr11, implicit $exec
+ ; GFX1250-LABEL: name: copy_v128_to_v128_unaligned
+ ; GFX1250: liveins: $vgpr8_vgpr9_vgpr10_vgpr11
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit-def $vgpr1_vgpr2_vgpr3_vgpr4, implicit $vgpr8_vgpr9_vgpr10_vgpr11
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit $vgpr8_vgpr9_vgpr10_vgpr11
+ ; GFX1250-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr10, implicit $exec, implicit $vgpr8_vgpr9_vgpr10_vgpr11
+ ; GFX1250-NEXT: $vgpr4 = V_MOV_B32_e32 $vgpr11, implicit $exec, implicit killed $vgpr8_vgpr9_vgpr10_vgpr11, implicit $exec
$vgpr1_vgpr2_vgpr3_vgpr4 = COPY killed $vgpr8_vgpr9_vgpr10_vgpr11, implicit $exec
...
@@ -442,6 +502,13 @@ body: |
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit $vgpr7_vgpr8_vgpr9_vgpr10
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit $vgpr7_vgpr8_vgpr9_vgpr10
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr10, implicit $exec, implicit killed $vgpr7_vgpr8_vgpr9_vgpr10, implicit $exec
+ ; GFX1250-LABEL: name: copy_v128_unaligned_to_v128
+ ; GFX1250: liveins: $vgpr7_vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr7, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr7_vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit $vgpr7_vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit $vgpr7_vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr10, implicit $exec, implicit killed $vgpr7_vgpr8_vgpr9_vgpr10, implicit $exec
$vgpr0_vgpr1_vgpr2_vgpr3 = COPY killed $vgpr7_vgpr8_vgpr9_vgpr10, implicit $exec
...
@@ -474,6 +541,11 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr8, implicit $exec, implicit-def $vgpr1_vgpr2, implicit $sgpr8_sgpr9
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr9, implicit $exec, implicit killed $sgpr8_sgpr9, implicit $exec
+ ; GFX1250-LABEL: name: copy_s64_to_v64_unaligned
+ ; GFX1250: liveins: $sgpr8_sgpr9
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr8, implicit $exec, implicit-def $vgpr1_vgpr2, implicit $sgpr8_sgpr9
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr9, implicit $exec, implicit killed $sgpr8_sgpr9, implicit $exec
$vgpr1_vgpr2 = COPY killed $sgpr8_sgpr9, implicit $exec
...
@@ -514,6 +586,13 @@ body: |
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr9, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $sgpr10, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11
; GFX10-NEXT: $vgpr4 = V_MOV_B32_e32 $sgpr11, implicit $exec, implicit killed $sgpr8_sgpr9_sgpr10_sgpr11, implicit $exec
+ ; GFX1250-LABEL: name: copy_s128_to_v128_unaligned
+ ; GFX1250: liveins: $sgpr8_sgpr9_sgpr10_sgpr11
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr8, implicit $exec, implicit-def $vgpr1_vgpr2_vgpr3_vgpr4, implicit $sgpr8_sgpr9_sgpr10_sgpr11
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr9, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11
+ ; GFX1250-NEXT: $vgpr3 = V_MOV_B32_e32 $sgpr10, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11
+ ; GFX1250-NEXT: $vgpr4 = V_MOV_B32_e32 $sgpr11, implicit $exec, implicit killed $sgpr8_sgpr9_sgpr10_sgpr11, implicit $exec
$vgpr1_vgpr2_vgpr3_vgpr4 = COPY killed $sgpr8_sgpr9_sgpr10_sgpr11, implicit $exec
...
@@ -550,6 +629,12 @@ body: |
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit-def $vgpr1_vgpr2_vgpr3, implicit $vgpr8_vgpr9_vgpr10
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit $vgpr8_vgpr9_vgpr10
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr10, implicit $exec, implicit killed $vgpr8_vgpr9_vgpr10, implicit $exec
+ ; GFX1250-LABEL: name: copy_v96_to_v96_unaligned
+ ; GFX1250: liveins: $vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit-def $vgpr1_vgpr2_vgpr3, implicit $vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit $vgpr8_vgpr9_vgpr10
+ ; GFX1250-NEXT: $vgpr3 = V_MOV_B32_e32 $vgpr10, implicit $exec, implicit killed $vgpr8_vgpr9_vgpr10, implicit $exec
$vgpr1_vgpr2_vgpr3 = COPY killed $vgpr8_vgpr9_vgpr10, implicit $exec
...
@@ -586,6 +671,12 @@ body: |
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr7, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr7_vgpr8_vgpr9
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit $vgpr7_vgpr8_vgpr9
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit killed $vgpr7_vgpr8_vgpr9, implicit $exec
+ ; GFX1250-LABEL: name: copy_v96_unaligned_to_v96
+ ; GFX1250: liveins: $vgpr7_vgpr8_vgpr9
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0 = V_MOV_B32_e32 $vgpr7, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr7_vgpr8_vgpr9
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $vgpr8, implicit $exec, implicit $vgpr7_vgpr8_vgpr9
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr9, implicit $exec, implicit killed $vgpr7_vgpr8_vgpr9, implicit $exec
$vgpr0_vgpr1_vgpr2 = COPY killed $vgpr7_vgpr8_vgpr9, implicit $exec
...
@@ -622,6 +713,12 @@ body: |
; GFX10-NEXT: $vgpr0 = V_MOV_B32_e32 $sgpr0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $sgpr0_sgpr1_sgpr2
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr2, implicit $exec, implicit killed $sgpr0_sgpr1_sgpr2, implicit $exec
+ ; GFX1250-LABEL: name: copy_s96_to_v96
+ ; GFX1250: liveins: $sgpr0_sgpr1_sgpr2
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr0 = V_MOV_B32_e32 $sgpr0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $sgpr0_sgpr1_sgpr2
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr2, implicit $exec, implicit killed $sgpr0_sgpr1_sgpr2, implicit $exec
$vgpr0_vgpr1_vgpr2 = COPY killed $sgpr0_sgpr1_sgpr2, implicit $exec
...
@@ -658,5 +755,11 @@ body: |
; GFX10-NEXT: $vgpr3 = V_MOV_B32_e32 $sgpr2, implicit $exec, implicit-def $vgpr1_vgpr2_vgpr3, implicit $sgpr0_sgpr1_sgpr2
; GFX10-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2
; GFX10-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr0, implicit $exec, implicit killed $sgpr0_sgpr1_sgpr2, implicit $exec
+ ; GFX1250-LABEL: name: copy_s96_to_v96_unaligned
+ ; GFX1250: liveins: $sgpr0_sgpr1_sgpr2
+ ; GFX1250-NEXT: {{ $}}
+ ; GFX1250-NEXT: $vgpr3 = V_MOV_B32_e32 $sgpr2, implicit $exec, implicit-def $vgpr1_vgpr2_vgpr3, implicit $sgpr0_sgpr1_sgpr2
+ ; GFX1250-NEXT: $vgpr2 = V_MOV_B32_e32 $sgpr1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2
+ ; GFX1250-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr0, implicit $exec, implicit killed $sgpr0_sgpr1_sgpr2, implicit $exec
$vgpr1_vgpr2_vgpr3 = COPY killed $sgpr0_sgpr1_sgpr2, implicit $exec
...
diff --git a/llvm/test/CodeGen/AMDGPU/global-load-xcnt.ll b/llvm/test/CodeGen/AMDGPU/global-load-xcnt.ll
index d8079651787ad..5d35adc8cbe0a 100644
--- a/llvm/test/CodeGen/AMDGPU/global-load-xcnt.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-load-xcnt.ll
@@ -65,15 +65,15 @@ define i16 @test_v7i16_load_store(ptr addrspace(1) %ptr1, ptr addrspace(1) %ptr2
; GCN-SDAG-NEXT: s_wait_kmcnt 0x0
; GCN-SDAG-NEXT: global_load_b128 v[4:7], v[0:1], off
; GCN-SDAG-NEXT: global_load_b128 v[0:3], v[2:3], off
+; GCN-SDAG-NEXT: v_mov_b64_e32 v[8:9], 0
; GCN-SDAG-NEXT: s_wait_loadcnt 0x0
; GCN-SDAG-NEXT: v_pk_add_u16 v10, v6, v2
; GCN-SDAG-NEXT: v_pk_add_u16 v11, v7, v3
; GCN-SDAG-NEXT: s_wait_xcnt 0x0
-; GCN-SDAG-NEXT: v_dual_mov_b32 v2, 12 :: v_dual_mov_b32 v6, 8
-; GCN-SDAG-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v7, 0
-; GCN-SDAG-NEXT: v_mov_b32_e32 v8, 0
+; GCN-SDAG-NEXT: v_mov_b64_e32 v[2:3], 12
+; GCN-SDAG-NEXT: v_mov_b64_e32 v[6:7], 8
; GCN-SDAG-NEXT: v_pk_add_u16 v4, v4, v0
-; GCN-SDAG-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_lshrrev_b32 v0, 16, v10
+; GCN-SDAG-NEXT: v_lshrrev_b32_e32 v0, 16, v10
; GCN-SDAG-NEXT: v_pk_add_u16 v5, v5, v1
; GCN-SDAG-NEXT: s_clause 0x2
; GCN-SDAG-NEXT: global_store_b16 v[2:3], v11, off
@@ -87,19 +87,17 @@ define i16 @test_v7i16_load_store(ptr addrspace(1) %ptr1, ptr addrspace(1) %ptr2
; GCN-GISEL-NEXT: s_wait_kmcnt 0x0
; GCN-GISEL-NEXT: global_load_b128 v[4:7], v[0:1], off
; GCN-GISEL-NEXT: global_load_b128 v[0:3], v[2:3], off
-; GCN-GISEL-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v10, 2
-; GCN-GISEL-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v11, 0
-; GCN-GISEL-NEXT: v_dual_mov_b32 v12, 4 :: v_dual_mov_b32 v14, 6
-; GCN-GISEL-NEXT: v_dual_mov_b32 v13, 0 :: v_dual_mov_b32 v15, 0
-; GCN-GISEL-NEXT: v_dual_mov_b32 v16, 8 :: v_dual_mov_b32 v18, 10
-; GCN-GISEL-NEXT: v_dual_mov_b32 v17, 0 :: v_dual_mov_b32 v19, 0
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[8:9], 0
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[10:11], 2
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[12:13], 4
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[14:15], 6
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[16:17], 8
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[18:19], 10
+; GCN-GISEL-NEXT: v_mov_b64_e32 v[20:21], 12
; GCN-GISEL-NEXT: s_wait_loadcnt 0x0
; GCN-GISEL-NEXT: v_pk_add_u16 v2, v6, v2
; GCN-GISEL-NEXT: v_pk_add_u16 v4, v4, v0
-; GCN-GISEL-NEXT: v_mov_b32_e32 v20, 12
; GCN-GISEL-NEXT: v_pk_add_u16 v1, v5, v1
-; GCN-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GCN-GISEL-NEXT: v_dual_mov_b32 v21, 0 :: v_dual_lshrrev_b32 v0, 16, v2
; GCN-GISEL-NEXT: v_pk_add_u16 v3, v7, v3
; GCN-GISEL-NEXT: s_clause 0x6
; GCN-GISEL-NEXT: global_store_b16 v[8:9], v4, off
@@ -109,6 +107,7 @@ define i16 @test_v7i16_load_store(ptr addrspace(1) %ptr1, ptr addrspace(1) %ptr2
; GCN-GISEL-NEXT: global_store_b16 v[16:17], v2, off
; GCN-GISEL-NEXT: global_store_d16_hi_b16 v[18:19], v2, off
; GCN-GISEL-NEXT: global_store_b16 v[20:21], v3, off
+; GCN-GISEL-NEXT: v_lshrrev_b32_e32 v0, 16, v2
; GCN-GISEL-NEXT: s_set_pc_i64 s[30:31]
%vec1 = load <7 x i16>, ptr addrspace...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/148272
More information about the llvm-commits
mailing list