AMDGPU: Reduce readfirstlane for single demanded vector element #128647
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This was referenced Feb 25, 2025
|
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-backend-amdgpu Author: Matt Arsenault (arsenm) ChangesIf we are only extracting a single element, rewrite the intrinsic call Full diff: https://github.com/llvm/llvm-project/pull/128647.diff 3 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
index 617974713d6f0..99016fdd0ff91 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
@@ -1538,6 +1538,49 @@ static Value *simplifyAMDGCNMemoryIntrinsicDemanded(InstCombiner &IC,
return NewCall;
}
+Value *GCNTTIImpl::simplifyAMDGCNLaneIntrinsicDemanded(
+ InstCombiner &IC, IntrinsicInst &II, const APInt &DemandedElts,
+ APInt &UndefElts) const {
+ auto *VT = dyn_cast<FixedVectorType>(II.getType());
+ if (!VT)
+ return nullptr;
+
+ const unsigned FirstElt = DemandedElts.countr_zero();
+ const unsigned LastElt = DemandedElts.getActiveBits() - 1;
+ const unsigned MaskLen = LastElt - FirstElt + 1;
+
+ // TODO: Handle general subvector extract.
+ if (MaskLen != 1)
+ return nullptr;
+
+ Type *EltTy = VT->getElementType();
+ if (!isTypeLegal(EltTy))
+ return nullptr;
+
+ Value *Src = II.getArgOperand(0);
+
+ assert(FirstElt == LastElt);
+ Value *Extract = IC.Builder.CreateExtractElement(Src, FirstElt);
+
+ // Make sure convergence tokens are preserved.
+ // TODO: CreateIntrinsic should allow directly copying bundles
+ SmallVector<OperandBundleDef, 2> OpBundles;
+ II.getOperandBundlesAsDefs(OpBundles);
+
+ Module *M = IC.Builder.GetInsertBlock()->getModule();
+ Function *Remangled = Intrinsic::getOrInsertDeclaration(
+ M, II.getIntrinsicID(), {Extract->getType()});
+
+ // TODO: Preserve callsite attributes?
+ CallInst *NewCall = IC.Builder.CreateCall(Remangled, {Extract}, OpBundles);
+
+ Value *Result = IC.Builder.CreateInsertElement(PoisonValue::get(II.getType()),
+ NewCall, FirstElt);
+ IC.replaceInstUsesWith(II, Result);
+ IC.eraseInstFromFunction(II);
+ return Result;
+}
+
std::optional<Value *> GCNTTIImpl::simplifyDemandedVectorEltsIntrinsic(
InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts,
APInt &UndefElts2, APInt &UndefElts3,
@@ -1545,9 +1588,8 @@ std::optional<Value *> GCNTTIImpl::simplifyDemandedVectorEltsIntrinsic(
SimplifyAndSetOp) const {
switch (II.getIntrinsicID()) {
case Intrinsic::amdgcn_readfirstlane:
- // TODO: For a vector extract, should reduce the intrinsic call type.
SimplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
- return std::nullopt;
+ return simplifyAMDGCNLaneIntrinsicDemanded(IC, II, DemandedElts, UndefElts);
case Intrinsic::amdgcn_raw_buffer_load:
case Intrinsic::amdgcn_raw_ptr_buffer_load:
case Intrinsic::amdgcn_raw_buffer_load_format:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
index a0d62008d9ddc..f5062070ac6f4 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
@@ -226,6 +226,12 @@ class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> {
std::optional<Instruction *> instCombineIntrinsic(InstCombiner &IC,
IntrinsicInst &II) const;
+
+ Value *simplifyAMDGCNLaneIntrinsicDemanded(InstCombiner &IC,
+ IntrinsicInst &II,
+ const APInt &DemandedElts,
+ APInt &UndefElts) const;
+
std::optional<Value *> simplifyDemandedVectorEltsIntrinsic(
InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts,
APInt &UndefElts2, APInt &UndefElts3,
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/simplify-demanded-vector-elts-lane-intrinsics.ll b/llvm/test/Transforms/InstCombine/AMDGPU/simplify-demanded-vector-elts-lane-intrinsics.ll
index 836c739048411..e9d3b5e963b35 100644
--- a/llvm/test/Transforms/InstCombine/AMDGPU/simplify-demanded-vector-elts-lane-intrinsics.ll
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/simplify-demanded-vector-elts-lane-intrinsics.ll
@@ -4,8 +4,8 @@
define i16 @extract_elt0_v2i16_readfirstlane(<2 x i16> %src) {
; CHECK-LABEL: define i16 @extract_elt0_v2i16_readfirstlane(
; CHECK-SAME: <2 x i16> [[SRC:%.*]]) #[[ATTR0:[0-9]+]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i16> @llvm.amdgcn.readfirstlane.v2i16(<2 x i16> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i16> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i16> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i16 @llvm.amdgcn.readfirstlane.i16(i16 [[TMP1]])
; CHECK-NEXT: ret i16 [[ELT]]
;
%vec = call <2 x i16> @llvm.amdgcn.readfirstlane.v2i16(<2 x i16> %src)
@@ -16,8 +16,8 @@ define i16 @extract_elt0_v2i16_readfirstlane(<2 x i16> %src) {
define i16 @extract_elt0_v1i16_readfirstlane(<1 x i16> %src) {
; CHECK-LABEL: define i16 @extract_elt0_v1i16_readfirstlane(
; CHECK-SAME: <1 x i16> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <1 x i16> @llvm.amdgcn.readfirstlane.v1i16(<1 x i16> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <1 x i16> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <1 x i16> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i16 @llvm.amdgcn.readfirstlane.i16(i16 [[TMP1]])
; CHECK-NEXT: ret i16 [[ELT]]
;
%vec = call <1 x i16> @llvm.amdgcn.readfirstlane.v1i16(<1 x i16> %src)
@@ -28,8 +28,8 @@ define i16 @extract_elt0_v1i16_readfirstlane(<1 x i16> %src) {
define i16 @extract_elt1_v2i16_readfirstlane(<2 x i16> %src) {
; CHECK-LABEL: define i16 @extract_elt1_v2i16_readfirstlane(
; CHECK-SAME: <2 x i16> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i16> @llvm.amdgcn.readfirstlane.v2i16(<2 x i16> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i16> [[VEC]], i64 1
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i16> [[SRC]], i64 1
+; CHECK-NEXT: [[ELT:%.*]] = call i16 @llvm.amdgcn.readfirstlane.i16(i16 [[TMP1]])
; CHECK-NEXT: ret i16 [[ELT]]
;
%vec = call <2 x i16> @llvm.amdgcn.readfirstlane.v2i16(<2 x i16> %src)
@@ -40,8 +40,8 @@ define i16 @extract_elt1_v2i16_readfirstlane(<2 x i16> %src) {
define i16 @extract_elt0_v4i16_readfirstlane(<4 x i16> %src) {
; CHECK-LABEL: define i16 @extract_elt0_v4i16_readfirstlane(
; CHECK-SAME: <4 x i16> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <4 x i16> @llvm.amdgcn.readfirstlane.v4i16(<4 x i16> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <4 x i16> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x i16> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i16 @llvm.amdgcn.readfirstlane.i16(i16 [[TMP1]])
; CHECK-NEXT: ret i16 [[ELT]]
;
%vec = call <4 x i16> @llvm.amdgcn.readfirstlane.v4i16(<4 x i16> %src)
@@ -52,8 +52,8 @@ define i16 @extract_elt0_v4i16_readfirstlane(<4 x i16> %src) {
define i16 @extract_elt2_v4i16_readfirstlane(<4 x i16> %src) {
; CHECK-LABEL: define i16 @extract_elt2_v4i16_readfirstlane(
; CHECK-SAME: <4 x i16> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <4 x i16> @llvm.amdgcn.readfirstlane.v4i16(<4 x i16> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <4 x i16> [[VEC]], i64 2
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x i16> [[SRC]], i64 2
+; CHECK-NEXT: [[ELT:%.*]] = call i16 @llvm.amdgcn.readfirstlane.i16(i16 [[TMP1]])
; CHECK-NEXT: ret i16 [[ELT]]
;
%vec = call <4 x i16> @llvm.amdgcn.readfirstlane.v4i16(<4 x i16> %src)
@@ -136,8 +136,8 @@ define <2 x i16> @extract_elt30_v4i16_readfirstlane(<4 x i16> %src) {
define half @extract_elt0_v2f16_readfirstlane(<2 x half> %src) {
; CHECK-LABEL: define half @extract_elt0_v2f16_readfirstlane(
; CHECK-SAME: <2 x half> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x half> @llvm.amdgcn.readfirstlane.v2f16(<2 x half> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x half> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x half> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call half @llvm.amdgcn.readfirstlane.f16(half [[TMP1]])
; CHECK-NEXT: ret half [[ELT]]
;
%vec = call <2 x half> @llvm.amdgcn.readfirstlane.v2i16(<2 x half> %src)
@@ -148,8 +148,8 @@ define half @extract_elt0_v2f16_readfirstlane(<2 x half> %src) {
define half @extract_elt1_v2f16_readfirstlane(<2 x half> %src) {
; CHECK-LABEL: define half @extract_elt1_v2f16_readfirstlane(
; CHECK-SAME: <2 x half> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x half> @llvm.amdgcn.readfirstlane.v2f16(<2 x half> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x half> [[VEC]], i64 1
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x half> [[SRC]], i64 1
+; CHECK-NEXT: [[ELT:%.*]] = call half @llvm.amdgcn.readfirstlane.f16(half [[TMP1]])
; CHECK-NEXT: ret half [[ELT]]
;
%vec = call <2 x half> @llvm.amdgcn.readfirstlane.v2i16(<2 x half> %src)
@@ -186,8 +186,8 @@ define i32 @extract_elt0_nxv4i32_readfirstlane(<vscale x 2 x i32> %src) {
define i32 @extract_elt0_v2i32_readfirstlane(<2 x i32> %src) {
; CHECK-LABEL: define i32 @extract_elt0_v2i32_readfirstlane(
; CHECK-SAME: <2 x i32> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i32> @llvm.amdgcn.readfirstlane.v2i32(<2 x i32> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i32> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i32> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i32 @llvm.amdgcn.readfirstlane.i32(i32 [[TMP1]])
; CHECK-NEXT: ret i32 [[ELT]]
;
%vec = call <2 x i32> @llvm.amdgcn.readfirstlane.v2i32(<2 x i32> %src)
@@ -198,8 +198,8 @@ define i32 @extract_elt0_v2i32_readfirstlane(<2 x i32> %src) {
define ptr addrspace(3) @extract_elt0_v2p3_readfirstlane(<2 x ptr addrspace(3)> %src) {
; CHECK-LABEL: define ptr addrspace(3) @extract_elt0_v2p3_readfirstlane(
; CHECK-SAME: <2 x ptr addrspace(3)> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x ptr addrspace(3)> @llvm.amdgcn.readfirstlane.v2p3(<2 x ptr addrspace(3)> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x ptr addrspace(3)> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x ptr addrspace(3)> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call ptr addrspace(3) @llvm.amdgcn.readfirstlane.p3(ptr addrspace(3) [[TMP1]])
; CHECK-NEXT: ret ptr addrspace(3) [[ELT]]
;
%vec = call <2 x ptr addrspace(3)> @llvm.amdgcn.readfirstlane.v2p3(<2 x ptr addrspace(3)> %src)
@@ -210,8 +210,8 @@ define ptr addrspace(3) @extract_elt0_v2p3_readfirstlane(<2 x ptr addrspace(3)>
define i64 @extract_elt0_v2i64_readfirstlane(<2 x i64> %src) {
; CHECK-LABEL: define i64 @extract_elt0_v2i64_readfirstlane(
; CHECK-SAME: <2 x i64> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i64> @llvm.amdgcn.readfirstlane.v2i64(<2 x i64> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i64> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i64> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i64 @llvm.amdgcn.readfirstlane.i64(i64 [[TMP1]])
; CHECK-NEXT: ret i64 [[ELT]]
;
%vec = call <2 x i64> @llvm.amdgcn.readfirstlane.v2i64(<2 x i64> %src)
@@ -222,8 +222,8 @@ define i64 @extract_elt0_v2i64_readfirstlane(<2 x i64> %src) {
define i64 @extract_elt1_v2i64_readfirstlane(<2 x i64> %src) {
; CHECK-LABEL: define i64 @extract_elt1_v2i64_readfirstlane(
; CHECK-SAME: <2 x i64> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i64> @llvm.amdgcn.readfirstlane.v2i64(<2 x i64> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i64> [[VEC]], i64 1
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i64> [[SRC]], i64 1
+; CHECK-NEXT: [[ELT:%.*]] = call i64 @llvm.amdgcn.readfirstlane.i64(i64 [[TMP1]])
; CHECK-NEXT: ret i64 [[ELT]]
;
%vec = call <2 x i64> @llvm.amdgcn.readfirstlane.v2i64(<2 x i64> %src)
@@ -306,9 +306,8 @@ define <2 x i16> @extract_elt13_v4i16readfirstlane(<4 x i16> %src) {
define <2 x i32> @extract_elt13_v4i32_readfirstlane_source_simplify0(i32 %src0, i32 %src2) {
; CHECK-LABEL: define <2 x i32> @extract_elt13_v4i32_readfirstlane_source_simplify0(
; CHECK-SAME: i32 [[SRC0:%.*]], i32 [[SRC2:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[INS_1:%.*]] = insertelement <4 x i32> poison, i32 [[SRC0]], i64 1
-; CHECK-NEXT: [[VEC:%.*]] = call <4 x i32> @llvm.amdgcn.readfirstlane.v4i32(<4 x i32> [[INS_1]])
-; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x i32> [[VEC]], <4 x i32> poison, <2 x i32> <i32 1, i32 poison>
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.amdgcn.readfirstlane.i32(i32 [[SRC0]])
+; CHECK-NEXT: [[SHUFFLE:%.*]] = insertelement <2 x i32> poison, i32 [[TMP1]], i64 0
; CHECK-NEXT: ret <2 x i32> [[SHUFFLE]]
;
%ins.0 = insertelement <4 x i32> poison, i32 %src0, i32 0
@@ -350,8 +349,8 @@ define i32 @extract_elt0_v2i32_readfirstlane_convergencetoken(<2 x i32> %src) co
; CHECK-LABEL: define i32 @extract_elt0_v2i32_readfirstlane_convergencetoken(
; CHECK-SAME: <2 x i32> [[SRC:%.*]]) #[[ATTR1:[0-9]+]] {
; CHECK-NEXT: [[T:%.*]] = call token @llvm.experimental.convergence.entry()
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i32> @llvm.amdgcn.readfirstlane.v2i32(<2 x i32> [[SRC]]) [ "convergencectrl"(token [[T]]) ]
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i32> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i32> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i32 @llvm.amdgcn.readfirstlane.i32(i32 [[TMP1]]) [ "convergencectrl"(token [[T]]) ]
; CHECK-NEXT: ret i32 [[ELT]]
;
%t = call token @llvm.experimental.convergence.entry()
@@ -381,8 +380,8 @@ define < 2 x i32> @extract_elt13_v4i32_readfirstlane_source_simplify1_convergenc
define i1 @extract_elt0_v2i1_readfirstlane(<2 x i1> %src) {
; CHECK-LABEL: define i1 @extract_elt0_v2i1_readfirstlane(
; CHECK-SAME: <2 x i1> [[SRC:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT: [[VEC:%.*]] = call <2 x i1> @llvm.amdgcn.readfirstlane.v2i1(<2 x i1> [[SRC]])
-; CHECK-NEXT: [[ELT:%.*]] = extractelement <2 x i1> [[VEC]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i1> [[SRC]], i64 0
+; CHECK-NEXT: [[ELT:%.*]] = call i1 @llvm.amdgcn.readfirstlane.i1(i1 [[TMP1]])
; CHECK-NEXT: ret i1 [[ELT]]
;
%vec = call <2 x i1> @llvm.amdgcn.readfirstlane.v2i1(<2 x i1> %src)
|
jayfoad
reviewed
Feb 25, 2025
Yes, this should be expanded to cover all of those cases (we should do the same for demanded bits). I was going to leave this as the first sample one, and open an issue to expand to the other cases |
arsenm
force-pushed
the
users/arsenm/amdgpu/simplify-demanded-vector-elts-readfirstlane-src
branch
from
ca0b3a4 to
e32caff
Compare
Base automatically changed from
users/arsenm/amdgpu/simplify-demanded-vector-elts-readfirstlane-src
to
main
February 28, 2025 06:01
If we are only extracting a single element, rewrite the intrinsic call to use the element type. We should extend this to arbitrary extract shuffles.
arsenm
force-pushed
the
users/arsenm/amdgpu/simplify-demanded-vector-elts-readfirstlane
branch
from
bfe67bc to
c80695c
Compare
rampitec
reviewed
Mar 4, 2025
Actually I think it does not work either way, <2 x i8> or just i8. So LGTM, but I think it would be better to also apply #114887. It will make i8 work, although <2 x i8> still does not. |
rampitec
approved these changes
Mar 4, 2025
arsenm
deleted the
users/arsenm/amdgpu/simplify-demanded-vector-elts-readfirstlane
branch
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/55/builds/7930 Here is the relevant piece of the build log for the reference |
Fixed by5c375c3283fcd2bf4f98fe8627658e056e25dc44 |
Should be able to do it in generic code for any |
jph-13
pushed a commit
to jph-13/llvm-project
that referenced
this pull request
Mar 21, 2025
…#128647) If we are only extracting a single element, rewrite the intrinsic call to use the element type. We should extend this to arbitrary extract shuffles.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
If we are only extracting a single element, rewrite the intrinsic call
to use the element type. We should extend this to arbitrary extract
shuffles.