[llvm] [PassBuilder] Move LoopFullUnrollPass to after vectorisation in the FullLTO pipeline (PR #197176)
via llvm-commits
llvm-commits at lists.llvm.org
Tue May 12 05:52:11 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-transforms
Author: Momchil Velikov (momchil-velikov)
<details>
<summary>Changes</summary>
This patch is in preparation for disabling the pre-LTO vectorisation in the FullLTO pipeline.
It moves `LoopFullUnrollPass` after `LoopVectorizePass` in the full LTO pipeline, replacing the `LoopUnrollPass` that was previously there.
Running the `LoopFullUnrollPass` before `LoopVectorizePass` causes functions to be unrolled before the vectoriser has had a chance to vectorise them (in the modified pipeline where vectorisation does not happen in the pre-link stage).
---
Full diff: https://github.com/llvm/llvm-project/pull/197176.diff
3 Files Affected:
- (modified) llvm/lib/Passes/PassBuilderPipelines.cpp (+12-9)
- (modified) llvm/test/Other/new-pm-lto-defaults.ll (+3-2)
- (modified) llvm/test/Transforms/PhaseOrdering/loop-vectorize-bfi.ll (+21-17)
``````````diff
diff --git a/llvm/lib/Passes/PassBuilderPipelines.cpp b/llvm/lib/Passes/PassBuilderPipelines.cpp
index 2d2867b9b84d1..ae329a420d1e0 100644
--- a/llvm/lib/Passes/PassBuilderPipelines.cpp
+++ b/llvm/lib/Passes/PassBuilderPipelines.cpp
@@ -1331,6 +1331,7 @@ void PassBuilder::addVectorPasses(OptimizationLevel Level,
FPM.addPass(InferAlignmentPass());
if (IsFullLTO) {
+ LoopPassManager LPM;
// The vectorizer may have significantly shortened a loop body; unroll
// again. Unroll small loops to hide loop backedge latency and saturate any
// parallel execution resources of an out-of-order processor. We also then
@@ -1340,11 +1341,15 @@ void PassBuilder::addVectorPasses(OptimizationLevel Level,
// across the loop nests.
// We do UnrollAndJam in a separate LPM to ensure it happens before unroll
if (EnableUnrollAndJam && PTO.LoopUnrolling)
- FPM.addPass(createFunctionToLoopPassAdaptor(
- LoopUnrollAndJamPass(Level.getSpeedupLevel())));
- FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
- Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
- PTO.ForgetAllSCEVInLoopUnroll)));
+ LPM.addPass(LoopUnrollAndJamPass(Level.getSpeedupLevel()));
+
+ // Unroll small loops and perform peeling.
+ LPM.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
+ /* OnlyWhenForced= */ !PTO.LoopUnrolling,
+ PTO.ForgetAllSCEVInLoopUnroll));
+ FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM),
+ /*UseMemorySSA=*/false));
+
FPM.addPass(WarnMissedTransformationsPass());
// Now that we are done with loop unrolling, be it either by LoopVectorizer,
// or LoopUnroll passes, some variable-offset GEP's into alloca's could have
@@ -2226,12 +2231,10 @@ PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
LPM.addPass(LoopDeletionPass());
// FIXME: Add loop interchange.
- // Unroll small loops and perform peeling.
- LPM.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
- /* OnlyWhenForced= */ !PTO.LoopUnrolling,
- PTO.ForgetAllSCEVInLoopUnroll));
// The loop passes in LPM (LoopFullUnrollPass) do not preserve MemorySSA.
// *All* loop passes must preserve it, in order to be able to use it.
+ // TODO: Explore enabling MemorySSA for loop passes after we remove the
+ // LoopFullUnrollPass.
MainFPM.addPass(
createFunctionToLoopPassAdaptor(std::move(LPM), /*UseMemorySSA=*/false));
diff --git a/llvm/test/Other/new-pm-lto-defaults.ll b/llvm/test/Other/new-pm-lto-defaults.ll
index 6d6d30298f7be..937c76e95391d 100644
--- a/llvm/test/Other/new-pm-lto-defaults.ll
+++ b/llvm/test/Other/new-pm-lto-defaults.ll
@@ -119,14 +119,15 @@
; CHECK-O23-NEXT: Running pass: LCSSAPass on foo
; CHECK-O23-NEXT: Running pass: IndVarSimplifyPass on loop
; CHECK-O23-NEXT: Running pass: LoopDeletionPass on loop
-; CHECK-O23-NEXT: Running pass: LoopFullUnrollPass on loop
; CHECK-O23-NEXT: Running pass: LoopDistributePass on foo
; CHECK-O23-NEXT: Running analysis: LoopAccessAnalysis on foo
; CHECK-O23-NEXT: Running pass: LoopVectorizePass on foo
; CHECK-O23-NEXT: Running analysis: DemandedBitsAnalysis on foo
; CHECK-O23-NEXT: Running pass: DropUnnecessaryAssumesPass on foo
; CHECK-O23-NEXT: Running pass: InferAlignmentPass on foo
-; CHECK-O23-NEXT: Running pass: LoopUnrollPass on foo
+; CHECK-O23-NEXT: Running pass: LoopSimplifyPass on foo
+; CHECK-O23-NEXT: Running pass: LCSSAPass on foo
+; CHECK-O23-NEXT: Running pass: LoopFullUnrollPass on loop
; CHECK-O23-NEXT: WarnMissedTransformationsPass on foo
; CHECK-O23-NEXT: Running pass: SROAPass on foo
; CHECK-O23-NEXT: Running pass: InstCombinePass on foo
diff --git a/llvm/test/Transforms/PhaseOrdering/loop-vectorize-bfi.ll b/llvm/test/Transforms/PhaseOrdering/loop-vectorize-bfi.ll
index baf7bf38f3351..756f537515153 100644
--- a/llvm/test/Transforms/PhaseOrdering/loop-vectorize-bfi.ll
+++ b/llvm/test/Transforms/PhaseOrdering/loop-vectorize-bfi.ll
@@ -1,6 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
; REQUIRES: riscv-registered-target
-; RUN: opt -p 'lto<O3>' -mtriple riscv64 -mattr=+v -S < %s | FileCheck %s
+; RUN: opt -p 'loop-unroll,loop-vectorize' -mtriple riscv64 -mattr=+v -S < %s | FileCheck %s
; Test that BlockFrequencyInfo is invalidated after loop passes, so it's
; accurate whenever LoopVectorize uses it. LoopVectorizer requires that
@@ -9,25 +9,29 @@
define void @f(i1 %x) !prof !0 {
; CHECK-LABEL: define void @f(
-; CHECK-SAME: i1 [[X:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {{.*}}{
-; CHECK-NEXT: [[ENTRY:.*]]:
-; CHECK-NEXT: [[DOTSCALAR:%.*]] = xor i1 [[X]], true
-; CHECK-NEXT: [[TMP1:%.*]] = insertelement <vscale x 2 x i1> poison, i1 [[DOTSCALAR]], i64 0
-; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <vscale x 2 x i1> [[TMP1]], <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
+; CHECK-SAME: i1 [[X:%.*]]) #[[ATTR0:[0-9]+]] {{.*}}{
+; CHECK-NEXT: [[ENTRY:.*:]]
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
-; CHECK-NEXT: [[EVL_BASED_IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[INDEX_EVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT: [[AVL:%.*]] = phi i64 [ 65, %[[ENTRY]] ], [ [[AVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT: [[TMP3:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 2, i1 true)
-; CHECK-NEXT: [[TMP4:%.*]] = getelementptr [8 x i8], ptr null, i64 [[EVL_BASED_IV]]
-; CHECK-NEXT: call void @llvm.vp.store.nxv2i64.p0(<vscale x 2 x i64> poison, ptr align 8 [[TMP4]], <vscale x 2 x i1> [[TMP2]], i32 [[TMP3]])
-; CHECK-NEXT: call void @llvm.vp.store.nxv2i64.p0(<vscale x 2 x i64> poison, ptr align 8 [[TMP4]], <vscale x 2 x i1> [[TMP2]], i32 [[TMP3]])
-; CHECK-NEXT: [[TMP5:%.*]] = zext nneg i32 [[TMP3]] to i64
-; CHECK-NEXT: [[INDEX_EVL_NEXT]] = add nuw i64 [[EVL_BASED_IV]], [[TMP5]]
-; CHECK-NEXT: [[AVL_NEXT]] = sub nuw nsw i64 [[AVL]], [[TMP5]]
-; CHECK-NEXT: [[TMP6:%.*]] = icmp eq i64 [[AVL_NEXT]], 0
-; CHECK-NEXT: br i1 [[TMP6]], label %[[EXIT:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP1:![0-9]+]]
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i1> poison, i1 [[X]], i64 0
+; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x i1> [[BROADCAST_SPLATINSERT]], <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
+; CHECK-NEXT: [[TMP0:%.*]] = xor <vscale x 2 x i1> [[BROADCAST_SPLAT]], splat (i1 true)
+; CHECK-NEXT: br label %[[EXIT:.*]]
; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_BODY]] ], [ [[CURRENT_ITERATION_NEXT:%.*]], %[[EXIT]] ]
+; CHECK-NEXT: [[AVL:%.*]] = phi i64 [ 65, %[[VECTOR_BODY]] ], [ [[AVL_NEXT:%.*]], %[[EXIT]] ]
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 2, i1 true)
+; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i64, ptr null, i64 [[INDEX]]
+; CHECK-NEXT: call void @llvm.vp.store.nxv2i64.p0(<vscale x 2 x i64> zeroinitializer, ptr align 8 [[TMP2]], <vscale x 2 x i1> [[TMP0]], i32 [[TMP1]])
+; CHECK-NEXT: call void @llvm.vp.store.nxv2i64.p0(<vscale x 2 x i64> zeroinitializer, ptr align 8 [[TMP2]], <vscale x 2 x i1> [[TMP0]], i32 [[TMP1]])
+; CHECK-NEXT: [[TMP3:%.*]] = zext i32 [[TMP1]] to i64
+; CHECK-NEXT: [[CURRENT_ITERATION_NEXT]] = add nuw i64 [[TMP3]], [[INDEX]]
+; CHECK-NEXT: [[AVL_NEXT]] = sub nuw i64 [[AVL]], [[TMP3]]
+; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[AVL_NEXT]], 0
+; CHECK-NEXT: br i1 [[TMP4]], label %[[EXIT1:.*]], label %[[EXIT]], !llvm.loop [[LOOP1:![0-9]+]]
+; CHECK: [[EXIT1]]:
+; CHECK-NEXT: br label %[[EXIT2:.*]]
+; CHECK: [[EXIT2]]:
; CHECK-NEXT: ret void
;
entry:
``````````
</details>
https://github.com/llvm/llvm-project/pull/197176
More information about the llvm-commits
mailing list