Skip to content

[VPlan] Add initial loop-invariant code motion transform. #107894

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 12 commits into from
Sep 20, 2024
Merged

[VPlan] Add initial loop-invariant code motion transform. #107894

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
7e9a5f8
[VPlan] Add initial loop-invariant code motion transform.
fhahn Aug 27, 2024
3caeea3
!fixup add TODO
fhahn Sep 10, 2024
f93524c
Merge remote-tracking branch 'origin/main' into vplan-licm
fhahn Sep 17, 2024
578ec3e
!fixup remove unrelated change, thanks!
fhahn Sep 17, 2024
a3ac937
Merge remote-tracking branch 'origin/main' into vplan-licm
fhahn Sep 19, 2024
62a24af
!fixup update tests
fhahn Sep 19, 2024
eb9dc17
Merge remote-tracking branch 'origin/main' into vplan-licm
fhahn Sep 19, 2024
b996541
Merge remote-tracking branch 'origin/main' into vplan-licm
fhahn Sep 19, 2024
9807dfe
Merge remote-tracking branch 'origin/main' into vplan-licm
fhahn Sep 19, 2024
7817971
Merge remote-tracking branch 'origin/main' into vplan-licm
fhahn Sep 20, 2024
fa1d94f
!fixup address latest comments, thanks!
fhahn Sep 20, 2024
46636d8
!fixup add tes t for
fhahn Sep 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2378,7 +2378,8 @@ void InnerLoopVectorizer::scalarizeInstruction(const Instruction *Instr,
AC->registerAssumption(II);

// End if-block.
bool IfPredicateInstr = RepRecipe->getParent()->getParent()->isReplicator();
VPRegionBlock *Parent = RepRecipe->getParent()->getParent();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(strictly speaking, qualifies as GrandParent ;-))

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps we should introduce getParentRegion or getRegion, similar to Instruction::getFunction?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe getEnclosingRegion(), to complement and serve getEnclosingLoopRegion()?

bool IfPredicateInstr = Parent ? Parent->isReplicator() : false;
if (IfPredicateInstr)
PredicatedInstructions.push_back(Cloned);
}
Expand Down
36 changes: 36 additions & 0 deletions llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -971,6 +971,41 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
return R.getVPSingleValue()->replaceAllUsesWith(A);
}

/// Move loop-invariant recipes out of the vector loop region in \p Plan.
static void licm(VPlan &Plan) {
VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
VPBasicBlock *Preheader =
cast<VPBasicBlock>(LoopRegion->getSinglePredecessor());

// Return true if we do not know how to (mechanically) hoist a given recipe
// out of a loop region. Does not address legality concerns such as aliasing
// or speculation safety.
auto CannotHoistRecipe = [](VPRecipeBase &R) {
// Allocas cannot be hoisted.
auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
return RepR && RepR->getOpcode() == Instruction::Alloca;
Comment on lines +985 to +986
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: match?

};

// Hoist any loop invariant recipes from the vector loop region to the
// preheader. Preform a shallow traversal of the vector loop region, to
// exclude recipes in replicate regions.
for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
vp_depth_first_shallow(LoopRegion->getEntry()))) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shallow to exclude going into replicate regions (we also check for side effects), visit only blocks that post-dom the header?
Should be fine also when folding the tail (right?)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, I'll add a comment to clarify re shallow traversal.

for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
if (CannotHoistRecipe(R))
continue;
// TODO: Relax checks in the future, e.g. we could also hoist reads, if
// their memory location is not modified in the vector loop.
if (R.mayHaveSideEffects() || R.mayReadFromMemory() || R.isPhi() ||
any_of(R.operands(), [](VPValue *Op) {
return !Op->isDefinedOutsideLoopRegions();
}))
continue;
R.moveBefore(*Preheader, Preheader->end());
}
}
}

/// Try to simplify the recipes in \p Plan.
static void simplifyRecipes(VPlan &Plan) {
ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
Expand Down Expand Up @@ -1123,6 +1158,7 @@ void VPlanTransforms::optimize(VPlan &Plan) {
removeRedundantInductionCasts(Plan);

simplifyRecipes(Plan);
licm(Plan);
legalizeAndOptimizeInductions(Plan);
removeDeadRecipes(Plan);

Expand Down
295 changes: 75 additions & 220 deletions llvm/test/Transforms/LoopVectorize/AArch64/conditional-branches-cost.ll
View file
Open in desktop

Large diffs are not rendered by default.

4 changes: 2 additions & 2 deletions llvm/test/Transforms/LoopVectorize/AArch64/divs-with-scalable-vfs.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,8 @@ define void @sdiv_feeding_gep(ptr %dst, i32 %x, i64 %M, i64 %conv6, i64 %N) {
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
; CHECK-NEXT: [[TMP10:%.*]] = call i64 @llvm.vscale.i64()
; CHECK-NEXT: [[TMP11:%.*]] = mul i64 [[TMP10]], 4
; CHECK-NEXT: [[TMP18:%.*]] = sdiv i64 [[M]], [[CONV6]]
; CHECK-NEXT: [[TMP19:%.*]] = sdiv i64 [[M]], [[CONV6]]
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
Expand All @@ -35,8 +37,6 @@ define void @sdiv_feeding_gep(ptr %dst, i32 %x, i64 %M, i64 %conv6, i64 %N) {
; CHECK-NEXT: [[TMP15:%.*]] = add i64 [[TMP14]], 0
; CHECK-NEXT: [[TMP16:%.*]] = mul i64 [[TMP15]], 1
; CHECK-NEXT: [[TMP17:%.*]] = add i64 [[INDEX]], [[TMP16]]
; CHECK-NEXT: [[TMP18:%.*]] = sdiv i64 [[M]], [[CONV6]]
; CHECK-NEXT: [[TMP19:%.*]] = sdiv i64 [[M]], [[CONV6]]
; CHECK-NEXT: [[TMP20:%.*]] = trunc i64 [[TMP18]] to i32
; CHECK-NEXT: [[TMP21:%.*]] = trunc i64 [[TMP19]] to i32
; CHECK-NEXT: [[TMP22:%.*]] = mul i64 [[TMP18]], [[CONV61]]
Expand Down
54 changes: 28 additions & 26 deletions llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,22 +17,23 @@

; Check that the extractvalue operands are actually free in vector code.

; FORCED-LABEL: vector.body: ; preds = %vector.body, %vector.ph
; FORCED-NEXT: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
; FORCED-NEXT: %0 = add i32 %index, 0
; FORCED-NEXT: %1 = extractvalue { i64, i64 } %sv, 0
; FORCED-NEXT: %broadcast.splatinsert = insertelement <2 x i64> poison, i64 %1, i64 0
; FORCED: [[E1:%.+]] = extractvalue { i64, i64 } %sv, 0
; FORCED-NEXT: %broadcast.splatinsert = insertelement <2 x i64> poison, i64 [[E1]], i64 0
; FORCED-NEXT: %broadcast.splat = shufflevector <2 x i64> %broadcast.splatinsert, <2 x i64> poison, <2 x i32> zeroinitializer
; FORCED-NEXT: %2 = extractvalue { i64, i64 } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert1 = insertelement <2 x i64> poison, i64 %2, i64 0
; FORCED-NEXT: [[E2:%.+]] = extractvalue { i64, i64 } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert1 = insertelement <2 x i64> poison, i64 [[E2]], i64 0
; FORCED-NEXT: %broadcast.splat2 = shufflevector <2 x i64> %broadcast.splatinsert1, <2 x i64> poison, <2 x i32> zeroinitializer
; FORCED-NEXT: %3 = getelementptr i64, ptr %dst, i32 %0
; FORCED-NEXT: %4 = add <2 x i64> %broadcast.splat, %broadcast.splat2
; FORCED-NEXT: %5 = getelementptr i64, ptr %3, i32 0
; FORCED-NEXT: store <2 x i64> %4, ptr %5, align 4

; FORCED-LABEL: vector.body: ; preds = %vector.body, %vector.ph
; FORCED-NEXT: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
; FORCED-NEXT: [[IV_0:%.]] = add i32 %index, 0
; FORCED-NEXT: [[GEP:%.+]] = getelementptr i64, ptr %dst, i32 [[IV_0]]
; FORCED-NEXT: [[ADD:%.+]] = add <2 x i64> %broadcast.splat, %broadcast.splat2
; FORCED-NEXT: [[GEP2:%.+]] = getelementptr i64, ptr [[GEP]], i32 0
; FORCED-NEXT: store <2 x i64> [[ADD]], ptr [[GEP2]], align 4
; FORCED-NEXT: %index.next = add nuw i32 %index, 2
; FORCED-NEXT: %6 = icmp eq i32 %index.next, 1000
; FORCED-NEXT: br i1 %6, label %middle.block, label %vector.body
; FORCED-NEXT: [[C:%.+]] = icmp eq i32 %index.next, 1000
; FORCED-NEXT: br i1 [[C]], label %middle.block, label %vector.body

define void @test1(ptr %dst, {i64, i64} %sv) {
entry:
Expand Down Expand Up @@ -66,22 +67,23 @@ declare float @powf(float, float) readnone nounwind

; FORCED-LABEL: define void @test_getVectorCallCost

; FORCED-LABEL: vector.body: ; preds = %vector.body, %vector.ph
; FORCED-NEXT: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
; FORCED-NEXT: %0 = add i32 %index, 0
; FORCED-NEXT: %1 = extractvalue { float, float } %sv, 0
; FORCED-NEXT: %broadcast.splatinsert = insertelement <2 x float> poison, float %1, i64 0
; FORCED: [[E1:%.+]] = extractvalue { float, float } %sv, 0
; FORCED-NEXT: %broadcast.splatinsert = insertelement <2 x float> poison, float [[E1]], i64 0
; FORCED-NEXT: %broadcast.splat = shufflevector <2 x float> %broadcast.splatinsert, <2 x float> poison, <2 x i32> zeroinitializer
; FORCED-NEXT: %2 = extractvalue { float, float } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert1 = insertelement <2 x float> poison, float %2, i64 0
; FORCED-NEXT: [[E2:%.+]] = extractvalue { float, float } %sv, 1
; FORCED-NEXT: %broadcast.splatinsert1 = insertelement <2 x float> poison, float [[E2]], i64 0
; FORCED-NEXT: %broadcast.splat2 = shufflevector <2 x float> %broadcast.splatinsert1, <2 x float> poison, <2 x i32> zeroinitializer
; FORCED-NEXT: %3 = getelementptr float, ptr %dst, i32 %0
; FORCED-NEXT: %4 = call <2 x float> @llvm.pow.v2f32(<2 x float> %broadcast.splat, <2 x float> %broadcast.splat2)
; FORCED-NEXT: %5 = getelementptr float, ptr %3, i32 0
; FORCED-NEXT: store <2 x float> %4, ptr %5, align 4

; FORCED-LABEL: vector.body: ; preds = %vector.body, %vector.ph
; FORCED-NEXT: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
; FORCED-NEXT: [[IV0:%.+]] = add i32 %index, 0
; FORCED-NEXT: [[GEP1:%.+]] = getelementptr float, ptr %dst, i32 [[IV0]]
; FORCED-NEXT: [[POW:%.+]] = call <2 x float> @llvm.pow.v2f32(<2 x float> %broadcast.splat, <2 x float> %broadcast.splat2)
; FORCED-NEXT: [[GEP2:%.+]] = getelementptr float, ptr [[GEP1]], i32 0
; FORCED-NEXT: store <2 x float> [[POW]], ptr [[GEP2]], align 4
; FORCED-NEXT: %index.next = add nuw i32 %index, 2
; FORCED-NEXT: %6 = icmp eq i32 %index.next, 1000
; FORCED-NEXT: br i1 %6, label %middle.block, label %vector.body
; FORCED-NEXT: [[C:%.+]] = icmp eq i32 %index.next, 1000
; FORCED-NEXT: br i1 [[C]], label %middle.block, label %vector.body

define void @test_getVectorCallCost(ptr %dst, {float, float} %sv) {
entry:
Expand Down
Loading