[llvm] 17677ad - [LV] Don't create WidePtrAdd recipes for scalar VFs (#169344)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Dec 1 00:12:44 PST 2025
Author: David Sherwood
Date: 2025-12-01T08:12:41Z
New Revision: 17677ad7eb2b2391d61c976887bbd2616e7d6c3e
URL: https://github.com/llvm/llvm-project/commit/17677ad7eb2b2391d61c976887bbd2616e7d6c3e
DIFF: https://github.com/llvm/llvm-project/commit/17677ad7eb2b2391d61c976887bbd2616e7d6c3e.diff
LOG: [LV] Don't create WidePtrAdd recipes for scalar VFs (#169344)
While attempting to remove the use of undef from more loop vectoriser
tests I discovered a bug where this assert was firing:
```
llvm::Constant* llvm::Constant::getSplatValue(bool) const: Assertion `this->getType()->isVectorTy() && "Only valid for vectors!"' failed.
...
#8 0x0000aaaab9e2fba4 llvm::Constant::getSplatValue
#9 0x0000aaaab9dfb844 llvm::ConstantFoldBinaryInstruction
```
This seems to be happening because we are incorrectly generating
WidePtrAdd recipes for scalar VFs. The PR fixes this by checking whether
a plan has a scalar VF only in legalizeAndOptimizeInductions.
This PR also removes the use of undef from the test `both` in
Transforms/LoopVectorize/iv_outside_user.ll, which is what started
triggering the assert.
Fixes #169334
Added:
Modified:
llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index b12f8ccc73c7e..f7281283bae81 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -769,7 +769,8 @@ static void legalizeAndOptimizeInductions(VPlan &Plan) {
// Replace wide pointer inductions which have only their scalars used by
// PtrAdd(IndStart, ScalarIVSteps (0, Step)).
if (auto *PtrIV = dyn_cast<VPWidenPointerInductionRecipe>(&Phi)) {
- if (!PtrIV->onlyScalarsGenerated(Plan.hasScalableVF()))
+ if (!Plan.hasScalarVFOnly() &&
+ !PtrIV->onlyScalarsGenerated(Plan.hasScalableVF()))
continue;
VPValue *PtrAdd = scalarizeVPWidenPointerInduction(PtrIV, Plan, Builder);
diff --git a/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll b/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
index b4fd06316a2e5..4f19a7c586bc3 100644
--- a/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
+++ b/llvm/test/Transforms/LoopVectorize/iv_outside_user.ll
@@ -152,59 +152,106 @@ for.end:
ret ptr %ptr.phi
}
-define ptr @both(i32 %k) {
-; CHECK-LABEL: define ptr @both(
-; CHECK-SAME: i32 [[K:%.*]]) {
-; CHECK-NEXT: [[ENTRY:.*]]:
-; CHECK-NEXT: [[BASE:%.*]] = getelementptr inbounds i32, ptr undef, i64 1
-; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[K]], -1
-; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
-; CHECK-NEXT: [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
-; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 2
-; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
-; CHECK: [[VECTOR_PH]]:
-; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 2
-; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
-; CHECK-NEXT: [[IND_END:%.*]] = trunc i64 [[N_VEC]] to i32
-; CHECK-NEXT: [[TMP3:%.*]] = mul i64 [[N_VEC]], 4
-; CHECK-NEXT: [[IND_END1:%.*]] = getelementptr i8, ptr [[BASE]], i64 [[TMP3]]
-; CHECK-NEXT: [[TMP4:%.*]] = mul i64 [[N_VEC]], 4
-; CHECK-NEXT: [[IND_END2:%.*]] = getelementptr i8, ptr undef, i64 [[TMP4]]
-; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
-; CHECK: [[VECTOR_BODY]]:
-; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
-; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], {{!llvm.loop ![0-9]+}}
-; CHECK: [[MIDDLE_BLOCK]]:
-; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
-; CHECK-NEXT: [[IND_ESCAPE:%.*]] = getelementptr i8, ptr [[IND_END1]], i64 -4
-; CHECK-NEXT: br i1 [[CMP_N]], label %[[FOR_END:.*]], label %[[SCALAR_PH]]
-; CHECK: [[SCALAR_PH]]:
-; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi ptr [ [[IND_END1]], %[[MIDDLE_BLOCK]] ], [ [[BASE]], %[[ENTRY]] ]
-; CHECK-NEXT: [[BC_RESUME_VAL2:%.*]] = phi ptr [ [[IND_END2]], %[[MIDDLE_BLOCK]] ], [ undef, %[[ENTRY]] ]
-; CHECK-NEXT: br label %[[FOR_BODY:.*]]
-; CHECK: [[FOR_BODY]]:
-; CHECK-NEXT: [[INC_PHI:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[INC:%.*]], %[[FOR_BODY]] ]
-; CHECK-NEXT: [[INC_LAG1:%.*]] = phi ptr [ [[BC_RESUME_VAL1]], %[[SCALAR_PH]] ], [ [[TMP:%.*]], %[[FOR_BODY]] ]
-; CHECK-NEXT: [[INC_LAG2:%.*]] = phi ptr [ [[BC_RESUME_VAL2]], %[[SCALAR_PH]] ], [ [[INC_LAG1]], %[[FOR_BODY]] ]
-; CHECK-NEXT: [[TMP]] = getelementptr inbounds i32, ptr [[INC_LAG1]], i64 1
-; CHECK-NEXT: [[INC]] = add nsw i32 [[INC_PHI]], 1
-; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[INC]], [[K]]
-; CHECK-NEXT: br i1 [[CMP]], label %[[FOR_END]], label %[[FOR_BODY]], {{!llvm.loop ![0-9]+}}
-; CHECK: [[FOR_END]]:
-; CHECK-NEXT: [[INC_LAG1_LCSSA:%.*]] = phi ptr [ [[INC_LAG1]], %[[FOR_BODY]] ], [ [[IND_ESCAPE]], %[[MIDDLE_BLOCK]] ]
-; CHECK-NEXT: ret ptr [[INC_LAG1_LCSSA]]
+define ptr @both(ptr %p, i32 %k) {
+; VEC-LABEL: define ptr @both(
+; VEC-SAME: ptr [[P:%.*]], i32 [[K:%.*]]) {
+; VEC-NEXT: [[ENTRY:.*]]:
+; VEC-NEXT: [[BASE:%.*]] = getelementptr inbounds i32, ptr [[P]], i64 1
+; VEC-NEXT: [[TMP0:%.*]] = add i32 [[K]], -1
+; VEC-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
+; VEC-NEXT: [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
+; VEC-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 2
+; VEC-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; VEC: [[VECTOR_PH]]:
+; VEC-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 2
+; VEC-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
+; VEC-NEXT: [[TMP3:%.*]] = trunc i64 [[N_VEC]] to i32
+; VEC-NEXT: [[TMP4:%.*]] = mul i64 [[N_VEC]], 4
+; VEC-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[BASE]], i64 [[TMP4]]
+; VEC-NEXT: br label %[[VECTOR_BODY:.*]]
+; VEC: [[VECTOR_BODY]]:
+; VEC-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; VEC-NEXT: [[POINTER_PHI:%.*]] = phi ptr [ [[BASE]], %[[VECTOR_PH]] ], [ [[PTR_IND:%.*]], %[[VECTOR_BODY]] ]
+; VEC-NEXT: [[VECTOR_GEP:%.*]] = getelementptr i8, ptr [[POINTER_PHI]], <2 x i64> <i64 0, i64 4>
+; VEC-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; VEC-NEXT: [[PTR_IND]] = getelementptr i8, ptr [[POINTER_PHI]], i64 8
+; VEC-NEXT: [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; VEC-NEXT: br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], {{!llvm.loop ![0-9]+}}
+; VEC: [[MIDDLE_BLOCK]]:
+; VEC-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <2 x ptr> [[VECTOR_GEP]], i32 1
+; VEC-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
+; VEC-NEXT: [[IND_ESCAPE:%.*]] = getelementptr i8, ptr [[TMP5]], i64 -4
+; VEC-NEXT: br i1 [[CMP_N]], label %[[FOR_END:.*]], label %[[SCALAR_PH]]
+; VEC: [[SCALAR_PH]]:
+; VEC-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP3]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; VEC-NEXT: [[BC_RESUME_VAL1:%.*]] = phi ptr [ [[TMP5]], %[[MIDDLE_BLOCK]] ], [ [[BASE]], %[[ENTRY]] ]
+; VEC-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi ptr [ [[VECTOR_RECUR_EXTRACT]], %[[MIDDLE_BLOCK]] ], [ [[BASE]], %[[ENTRY]] ]
+; VEC-NEXT: br label %[[FOR_BODY:.*]]
+; VEC: [[FOR_BODY]]:
+; VEC-NEXT: [[INC_PHI:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[INC:%.*]], %[[FOR_BODY]] ]
+; VEC-NEXT: [[INC_LAG1:%.*]] = phi ptr [ [[BC_RESUME_VAL1]], %[[SCALAR_PH]] ], [ [[TMP:%.*]], %[[FOR_BODY]] ]
+; VEC-NEXT: [[INC_LAG2:%.*]] = phi ptr [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[INC_LAG1]], %[[FOR_BODY]] ]
+; VEC-NEXT: [[TMP]] = getelementptr inbounds i32, ptr [[INC_LAG1]], i64 1
+; VEC-NEXT: [[INC]] = add nsw i32 [[INC_PHI]], 1
+; VEC-NEXT: [[CMP:%.*]] = icmp eq i32 [[INC]], [[K]]
+; VEC-NEXT: br i1 [[CMP]], label %[[FOR_END]], label %[[FOR_BODY]], {{!llvm.loop ![0-9]+}}
+; VEC: [[FOR_END]]:
+; VEC-NEXT: [[INC_LAG1_LCSSA:%.*]] = phi ptr [ [[INC_LAG1]], %[[FOR_BODY]] ], [ [[IND_ESCAPE]], %[[MIDDLE_BLOCK]] ]
+; VEC-NEXT: ret ptr [[INC_LAG1_LCSSA]]
+;
+; INTERLEAVE-LABEL: define ptr @both(
+; INTERLEAVE-SAME: ptr [[P:%.*]], i32 [[K:%.*]]) {
+; INTERLEAVE-NEXT: [[ENTRY:.*]]:
+; INTERLEAVE-NEXT: [[BASE:%.*]] = getelementptr inbounds i32, ptr [[P]], i64 1
+; INTERLEAVE-NEXT: [[TMP0:%.*]] = add i32 [[K]], -1
+; INTERLEAVE-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
+; INTERLEAVE-NEXT: [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
+; INTERLEAVE-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 2
+; INTERLEAVE-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; INTERLEAVE: [[VECTOR_PH]]:
+; INTERLEAVE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 2
+; INTERLEAVE-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
+; INTERLEAVE-NEXT: [[TMP3:%.*]] = trunc i64 [[N_VEC]] to i32
+; INTERLEAVE-NEXT: [[TMP6:%.*]] = mul i64 [[N_VEC]], 4
+; INTERLEAVE-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[BASE]], i64 [[TMP6]]
+; INTERLEAVE-NEXT: br label %[[VECTOR_BODY:.*]]
+; INTERLEAVE: [[VECTOR_BODY]]:
+; INTERLEAVE-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; INTERLEAVE-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 4
+; INTERLEAVE-NEXT: [[TMP8:%.*]] = add i64 [[OFFSET_IDX]], 4
+; INTERLEAVE-NEXT: [[NEXT_GEP1:%.*]] = getelementptr i8, ptr [[BASE]], i64 [[TMP8]]
+; INTERLEAVE-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; INTERLEAVE-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; INTERLEAVE-NEXT: br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], {{!llvm.loop ![0-9]+}}
+; INTERLEAVE: [[MIDDLE_BLOCK]]:
+; INTERLEAVE-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
+; INTERLEAVE-NEXT: [[IND_ESCAPE:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i64 -4
+; INTERLEAVE-NEXT: br i1 [[CMP_N]], label %[[FOR_END:.*]], label %[[SCALAR_PH]]
+; INTERLEAVE: [[SCALAR_PH]]:
+; INTERLEAVE-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP3]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; INTERLEAVE-NEXT: [[BC_RESUME_VAL1:%.*]] = phi ptr [ [[NEXT_GEP]], %[[MIDDLE_BLOCK]] ], [ [[BASE]], %[[ENTRY]] ]
+; INTERLEAVE-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi ptr [ [[NEXT_GEP1]], %[[MIDDLE_BLOCK]] ], [ [[BASE]], %[[ENTRY]] ]
+; INTERLEAVE-NEXT: br label %[[FOR_BODY:.*]]
+; INTERLEAVE: [[FOR_BODY]]:
+; INTERLEAVE-NEXT: [[INC_PHI:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[INC:%.*]], %[[FOR_BODY]] ]
+; INTERLEAVE-NEXT: [[INC_LAG1:%.*]] = phi ptr [ [[BC_RESUME_VAL1]], %[[SCALAR_PH]] ], [ [[TMP:%.*]], %[[FOR_BODY]] ]
+; INTERLEAVE-NEXT: [[INC_LAG2:%.*]] = phi ptr [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[INC_LAG1]], %[[FOR_BODY]] ]
+; INTERLEAVE-NEXT: [[TMP]] = getelementptr inbounds i32, ptr [[INC_LAG1]], i64 1
+; INTERLEAVE-NEXT: [[INC]] = add nsw i32 [[INC_PHI]], 1
+; INTERLEAVE-NEXT: [[CMP:%.*]] = icmp eq i32 [[INC]], [[K]]
+; INTERLEAVE-NEXT: br i1 [[CMP]], label %[[FOR_END]], label %[[FOR_BODY]], {{!llvm.loop ![0-9]+}}
+; INTERLEAVE: [[FOR_END]]:
+; INTERLEAVE-NEXT: [[INC_LAG1_LCSSA:%.*]] = phi ptr [ [[INC_LAG1]], %[[FOR_BODY]] ], [ [[IND_ESCAPE]], %[[MIDDLE_BLOCK]] ]
+; INTERLEAVE-NEXT: ret ptr [[INC_LAG1_LCSSA]]
;
entry:
- %base = getelementptr inbounds i32, ptr undef, i64 1
+ %base = getelementptr inbounds i32, ptr %p, i64 1
br label %for.body
for.body:
%inc.phi = phi i32 [ 0, %entry ], [ %inc, %for.body ]
%inc.lag1 = phi ptr [ %base, %entry ], [ %tmp, %for.body]
- %inc.lag2 = phi ptr [ undef, %entry ], [ %inc.lag1, %for.body]
+ %inc.lag2 = phi ptr [ %base, %entry ], [ %inc.lag1, %for.body]
%tmp = getelementptr inbounds i32, ptr %inc.lag1, i64 1
%inc = add nsw i32 %inc.phi, 1
%cmp = icmp eq i32 %inc, %k
More information about the llvm-commits
mailing list