[llvm] [AMDGPU] Allow sinking of free vector ops (PR #162580)
Shilei Tian via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 14 21:04:06 PDT 2025
================
@@ -0,0 +1,152 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 < %s | FileCheck -check-prefix=OPT %s
+
+; testing insert case
+define amdgpu_kernel void @runningSum(ptr addrspace(1) %out0, ptr addrspace(1) %out1, i32 %inputElement0, i32 %inputElement1, i32 %inputIter) {
+; OPT-LABEL: define amdgpu_kernel void @runningSum(
+; OPT-SAME: ptr addrspace(1) [[OUT0:%.*]], ptr addrspace(1) [[OUT1:%.*]], i32 [[INPUTELEMENT0:%.*]], i32 [[INPUTELEMENT1:%.*]], i32 [[INPUTITER:%.*]]) #[[ATTR0:[0-9]+]] {
+; OPT-NEXT: [[PREHEADER:.*]]:
+; OPT-NEXT: [[VECELEMENT0:%.*]] = insertelement <2 x i32> poison, i32 [[INPUTELEMENT0]], i64 0
+; OPT-NEXT: [[BROADCAST0:%.*]] = shufflevector <2 x i32> [[VECELEMENT0]], <2 x i32> poison, <2 x i32> zeroinitializer
+; OPT-NEXT: [[VECELEMENT1:%.*]] = insertelement <2 x i32> poison, i32 [[INPUTELEMENT1]], i64 0
+; OPT-NEXT: [[BROADCAST1:%.*]] = shufflevector <2 x i32> [[VECELEMENT1]], <2 x i32> poison, <2 x i32> zeroinitializer
+; OPT-NEXT: br label %[[LOOPBODY:.*]]
+; OPT: [[LOOPBODY]]:
+; OPT-NEXT: [[PREVIOUSSUM:%.*]] = phi <2 x i32> [ [[BROADCAST1]], %[[PREHEADER]] ], [ [[RUNNINGSUM:%.*]], %[[LOOPBODY]] ]
+; OPT-NEXT: [[ITERCOUNT:%.*]] = phi i32 [ [[INPUTITER]], %[[PREHEADER]] ], [ [[ITERSLEFT:%.*]], %[[LOOPBODY]] ]
+; OPT-NEXT: [[TMP0:%.*]] = shufflevector <2 x i32> [[VECELEMENT1]], <2 x i32> poison, <2 x i32> zeroinitializer
+; OPT-NEXT: [[RUNNINGSUM]] = add <2 x i32> [[TMP0]], [[PREVIOUSSUM]]
+; OPT-NEXT: [[ITERSLEFT]] = sub i32 [[ITERCOUNT]], 1
+; OPT-NEXT: [[COND:%.*]] = icmp eq i32 [[ITERSLEFT]], 0
+; OPT-NEXT: br i1 [[COND]], label %[[LOOPEXIT:.*]], label %[[LOOPBODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; OPT: [[LOOPEXIT]]:
+; OPT-NEXT: [[SUMELEMENT0:%.*]] = extractelement <2 x i32> [[RUNNINGSUM]], i64 0
+; OPT-NEXT: [[SUMELEMENT1:%.*]] = extractelement <2 x i32> [[RUNNINGSUM]], i64 1
+; OPT-NEXT: store i32 [[SUMELEMENT0]], ptr addrspace(1) [[OUT0]], align 4
+; OPT-NEXT: store i32 [[SUMELEMENT1]], ptr addrspace(1) [[OUT1]], align 4
+; OPT-NEXT: ret void
+;
+preheader:
+ %vecElement0 = insertelement <2 x i32> poison, i32 %inputElement0, i64 0
+ %broadcast0 = shufflevector <2 x i32> %vecElement0, <2 x i32> poison, <2 x i32> zeroinitializer
+ %vecElement1 = insertelement <2 x i32> poison, i32 %inputElement1, i64 0
+ %broadcast1 = shufflevector <2 x i32> %vecElement1, <2 x i32> poison, <2 x i32> zeroinitializer
+ br label %loopBody
+
+loopBody:
+ %previousSum = phi <2 x i32> [ %broadcast1, %preheader ], [ %runningSum, %loopBody ]
+ %iterCount = phi i32 [ %inputIter, %preheader ], [ %itersLeft, %loopBody ]
+ %runningSum = add <2 x i32> %broadcast1, %previousSum
+ %itersLeft = sub i32 %iterCount, 1
+ %cond = icmp eq i32 %itersLeft, 0
+ br i1 %cond, label %loopExit, label %loopBody, !llvm.loop !0
+
+loopExit:
+ %sumElement0 = extractelement <2 x i32> %runningSum, i64 0
+ %sumElement1 = extractelement <2 x i32> %runningSum, i64 1
+ store i32 %sumElement0, ptr addrspace(1) %out0
+ store i32 %sumElement1, ptr addrspace(1) %out1
+ ret void
+}
+
+; testing extract case with single use
+define amdgpu_kernel void @test_sink_extract_single_use_operands(ptr addrspace(1) %out0, <2 x i32> %inputVec) {
+; OPT-LABEL: define amdgpu_kernel void @test_sink_extract_single_use_operands(
+; OPT-SAME: ptr addrspace(1) [[OUT0:%.*]], <2 x i32> [[INPUTVEC:%.*]]) #[[ATTR0]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[RUNNINGSUM:%.*]] = add <2 x i32> [[INPUTVEC]], splat (i32 1)
+; OPT-NEXT: [[SUMELEMENT0:%.*]] = extractelement <2 x i32> [[RUNNINGSUM]], i64 0
+; OPT-NEXT: [[RESULT:%.*]] = add i32 [[SUMELEMENT0]], 100
+; OPT-NEXT: store i32 [[RESULT]], ptr addrspace(1) [[OUT0]], align 4
+; OPT-NEXT: ret void
+;
+entry:
+ %runningSum = add <2 x i32> %inputVec, <i32 1, i32 1>
+ %sumElement0 = extractelement <2 x i32> %runningSum, i64 0
+ %result = add i32 %sumElement0, 100
+ store i32 %result, ptr addrspace(1) %out0
+ ret void
+}
+
+; testing extract case with multiple uses
+define amdgpu_kernel void @test_sink_extract_operands(ptr addrspace(1) %ptr, <4 x i32> %input_vec) {
+; OPT-LABEL: define amdgpu_kernel void @test_sink_extract_operands(
+; OPT-SAME: ptr addrspace(1) [[PTR:%.*]], <4 x i32> [[INPUT_VEC:%.*]]) #[[ATTR0]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[VEC_FULL:%.*]] = add <4 x i32> [[INPUT_VEC]], <i32 42, i32 43, i32 44, i32 45>
+; OPT-NEXT: [[EXTRACT0:%.*]] = extractelement <4 x i32> [[VEC_FULL]], i64 0
+; OPT-NEXT: [[VEC_SHIFTED:%.*]] = shl <4 x i32> [[VEC_FULL]], splat (i32 1)
+; OPT-NEXT: [[RESULT0:%.*]] = add i32 [[EXTRACT0]], 100
+; OPT-NEXT: store i32 [[RESULT0]], ptr addrspace(1) [[PTR]], align 4
+; OPT-NEXT: store <4 x i32> [[VEC_SHIFTED]], ptr addrspace(1) [[PTR]], align 16
+; OPT-NEXT: ret void
+;
+entry:
+ %vec_full = add <4 x i32> %input_vec, <i32 42, i32 43, i32 44, i32 45>
+ %extract0 = extractelement <4 x i32> %vec_full, i64 0
+ %vec_shifted = shl <4 x i32> %vec_full, <i32 1, i32 1, i32 1, i32 1>
+ %result0 = add i32 %extract0, 100
+ store i32 %result0, ptr addrspace(1) %ptr
+ store <4 x i32> %vec_shifted, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_kernel void @test_shuffle_insert_subvector(ptr addrspace(1) %ptr, <4 x i16> %vec1, <4 x i16> %vec2) {
+; OPT-LABEL: define amdgpu_kernel void @test_shuffle_insert_subvector(
+; OPT-SAME: ptr addrspace(1) [[PTR:%.*]], <4 x i16> [[VEC1:%.*]], <4 x i16> [[VEC2:%.*]]) #[[ATTR0]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x i16> [[VEC1]], <4 x i16> [[VEC2]], <4 x i32> <i32 0, i32 1, i32 4, i32 5>
+; OPT-NEXT: [[OTHER_SHUFFLE:%.*]] = shufflevector <4 x i16> [[SHUFFLE]], <4 x i16> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+; OPT-NEXT: [[RESULT_VEC:%.*]] = add <4 x i16> [[SHUFFLE]], <i16 100, i16 200, i16 300, i16 400>
+; OPT-NEXT: [[OTHER_RESULT:%.*]] = mul <4 x i16> [[OTHER_SHUFFLE]], splat (i16 2)
+; OPT-NEXT: store <4 x i16> [[RESULT_VEC]], ptr addrspace(1) [[PTR]], align 8
+; OPT-NEXT: store <4 x i16> [[OTHER_RESULT]], ptr addrspace(1) [[PTR]], align 8
+; OPT-NEXT: ret void
+;
+entry:
+ %shuffle = shufflevector <4 x i16> %vec1, <4 x i16> %vec2, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
+ %other_shuffle = shufflevector <4 x i16> %shuffle, <4 x i16> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+ %result_vec = add <4 x i16> %shuffle, <i16 100, i16 200, i16 300, i16 400>
+ %other_result = mul <4 x i16> %other_shuffle, <i16 2, i16 2, i16 2, i16 2>
+ store <4 x i16> %result_vec, ptr addrspace(1) %ptr
+ store <4 x i16> %other_result, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_kernel void @test_shuffle_extract_subvector(ptr addrspace(1) %ptr, <4 x i16> %input_vec) {
+; OPT-LABEL: define amdgpu_kernel void @test_shuffle_extract_subvector(
+; OPT-SAME: ptr addrspace(1) [[PTR:%.*]], <4 x i16> [[INPUT_VEC:%.*]]) #[[ATTR0]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x i16> [[INPUT_VEC]], <4 x i16> poison, <2 x i32> <i32 2, i32 3>
+; OPT-NEXT: [[RESULT_VEC:%.*]] = add <2 x i16> [[SHUFFLE]], <i16 100, i16 200>
+; OPT-NEXT: store <2 x i16> [[RESULT_VEC]], ptr addrspace(1) [[PTR]], align 4
+; OPT-NEXT: ret void
+;
+entry:
+ %shuffle = shufflevector <4 x i16> %input_vec, <4 x i16> poison, <2 x i32> <i32 2, i32 3>
+ %result_vec = add <2 x i16> %shuffle, <i16 100, i16 200>
+ store <2 x i16> %result_vec, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_kernel void @test_shuffle_sink_operands(ptr addrspace(1) %ptr, <2 x i16> %input_vec) {
+; OPT-LABEL: define amdgpu_kernel void @test_shuffle_sink_operands(
+; OPT-SAME: ptr addrspace(1) [[PTR:%.*]], <2 x i16> [[INPUT_VEC:%.*]]) #[[ATTR0]] {
+; OPT-NEXT: [[ENTRY:.*:]]
+; OPT-NEXT: [[SHUFFLE:%.*]] = shufflevector <2 x i16> [[INPUT_VEC]], <2 x i16> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>
+; OPT-NEXT: [[RESULT_VEC:%.*]] = add <4 x i16> [[SHUFFLE]], <i16 100, i16 200, i16 300, i16 400>
+; OPT-NEXT: store <4 x i16> [[RESULT_VEC]], ptr addrspace(1) [[PTR]], align 8
+; OPT-NEXT: ret void
+;
+entry:
+ %shuffle = shufflevector <2 x i16> %input_vec, <2 x i16> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>
+ %result_vec = add <4 x i16> %shuffle, <i16 100, i16 200, i16 300, i16 400>
+ store <4 x i16> %result_vec, ptr addrspace(1) %ptr
+ ret void
+}
+
+!0 = !{!"llvm.loop.mustprogress"}
----------------
shiltian wrote:
Is this necessary?
https://github.com/llvm/llvm-project/pull/162580
More information about the llvm-commits
mailing list