[llvm] [VPlan] Permit more users in narrowToSingleScalars (PR #166559)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Fri Nov 7 00:56:58 PST 2025
https://github.com/artagnon updated https://github.com/llvm/llvm-project/pull/166559
>From 434b246baf9deb3455a8bba71041566776cbdcfc Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Wed, 5 Nov 2025 13:29:44 +0000
Subject: [PATCH 1/5] [VPlan] Permit more users in narrowToSingleScalars
narrowToSingleScalarRecipes can permit users that are WidenStore, or a
VPInstruction that is VectorToScalar or SingleScalar. This is an
generalization and extension of the existing code.
---
.../Transforms/Vectorize/VPlanTransforms.cpp | 12 ++--
.../AArch64/conditional-branches-cost.ll | 70 ++++++++-----------
.../extractvalue-no-scalarization-required.ll | 22 +++---
.../AArch64/sve-widen-extractvalue.ll | 8 +--
.../LoopVectorize/RISCV/dead-ops-cost.ll | 6 +-
llvm/test/Transforms/LoopVectorize/pr50686.ll | 18 ++---
6 files changed, 61 insertions(+), 75 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 48bd697397f41..6ad8ca5743114 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1419,10 +1419,14 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
// broadcasts.
if (!vputils::isSingleScalar(RepOrWidenR) ||
!all_of(RepOrWidenR->users(), [RepOrWidenR](const VPUser *U) {
- return U->usesScalars(RepOrWidenR) ||
- match(cast<VPRecipeBase>(U),
- m_CombineOr(m_ExtractLastElement(m_VPValue()),
- m_ExtractLastLanePerPart(m_VPValue())));
+ if (isa<VPWidenStoreRecipe>(U))
+ return true;
+
+ if (auto *VPI = dyn_cast<VPInstruction>(U))
+ if (VPI->isVectorToScalar() || VPI->isSingleScalar())
+ return true;
+
+ return U->usesScalars(RepOrWidenR);
}))
continue;
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/conditional-branches-cost.ll b/llvm/test/Transforms/LoopVectorize/AArch64/conditional-branches-cost.ll
index f16351720b20f..8380b41c0e9a7 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/conditional-branches-cost.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/conditional-branches-cost.ll
@@ -429,48 +429,36 @@ define i32 @header_mask_and_invariant_compare(ptr %A, ptr %B, ptr %C, ptr %D, pt
; DEFAULT-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
; DEFAULT-NEXT: br label %[[VECTOR_BODY:.*]]
; DEFAULT: [[VECTOR_BODY]]:
-; DEFAULT-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_STORE_CONTINUE37:.*]] ]
-; DEFAULT-NEXT: [[TMP9:%.*]] = load i32, ptr [[A]], align 4, !alias.scope [[META8:![0-9]+]]
-; DEFAULT-NEXT: [[BROADCAST_SPLATINSERT28:%.*]] = insertelement <4 x i32> poison, i32 [[TMP9]], i64 0
-; DEFAULT-NEXT: [[BROADCAST_SPLAT29:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT28]], <4 x i32> poison, <4 x i32> zeroinitializer
-; DEFAULT-NEXT: [[TMP19:%.*]] = load i32, ptr [[B]], align 4, !alias.scope [[META11:![0-9]+]]
-; DEFAULT-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP19]], i64 0
-; DEFAULT-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
-; DEFAULT-NEXT: [[TMP6:%.*]] = or <4 x i32> [[BROADCAST_SPLAT]], [[BROADCAST_SPLAT29]]
-; DEFAULT-NEXT: [[TMP7:%.*]] = load i32, ptr [[C]], align 4, !alias.scope [[META13:![0-9]+]]
-; DEFAULT-NEXT: [[BROADCAST_SPLATINSERT30:%.*]] = insertelement <4 x i32> poison, i32 [[TMP7]], i64 0
-; DEFAULT-NEXT: [[BROADCAST_SPLAT31:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT30]], <4 x i32> poison, <4 x i32> zeroinitializer
-; DEFAULT-NEXT: [[TMP8:%.*]] = icmp ugt <4 x i32> [[BROADCAST_SPLAT31]], [[TMP6]]
+; DEFAULT-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_STORE_CONTINUE33:.*]] ]
+; DEFAULT-NEXT: [[TMP3:%.*]] = load i32, ptr [[A]], align 4, !alias.scope [[META8:![0-9]+]]
+; DEFAULT-NEXT: [[TMP4:%.*]] = load i32, ptr [[B]], align 4, !alias.scope [[META11:![0-9]+]]
+; DEFAULT-NEXT: [[TMP5:%.*]] = or i32 [[TMP4]], [[TMP3]]
+; DEFAULT-NEXT: [[TMP6:%.*]] = load i32, ptr [[C]], align 4, !alias.scope [[META13:![0-9]+]]
+; DEFAULT-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP6]], [[TMP5]]
+; DEFAULT-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i1> poison, i1 [[TMP7]], i64 0
+; DEFAULT-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i1> [[BROADCAST_SPLATINSERT]], <4 x i1> poison, <4 x i32> zeroinitializer
; DEFAULT-NEXT: [[TMP16:%.*]] = getelementptr i32, ptr [[D]], i64 [[INDEX]]
-; DEFAULT-NEXT: [[TMP20:%.*]] = extractelement <4 x i1> [[TMP8]], i32 0
-; DEFAULT-NEXT: br i1 [[TMP20]], label %[[PRED_STORE_IF:.*]], label %[[PRED_STORE_CONTINUE:.*]]
+; DEFAULT-NEXT: br i1 [[TMP7]], label %[[PRED_STORE_IF:.*]], label %[[PRED_STORE_CONTINUE:.*]]
; DEFAULT: [[PRED_STORE_IF]]:
-; DEFAULT-NEXT: [[TMP11:%.*]] = extractelement <4 x i32> [[TMP6]], i32 0
-; DEFAULT-NEXT: store i32 [[TMP11]], ptr [[E]], align 4, !alias.scope [[META15:![0-9]+]], !noalias [[META17:![0-9]+]]
+; DEFAULT-NEXT: store i32 [[TMP5]], ptr [[E]], align 4, !alias.scope [[META15:![0-9]+]], !noalias [[META17:![0-9]+]]
; DEFAULT-NEXT: br label %[[PRED_STORE_CONTINUE]]
; DEFAULT: [[PRED_STORE_CONTINUE]]:
-; DEFAULT-NEXT: [[TMP12:%.*]] = extractelement <4 x i1> [[TMP8]], i32 1
-; DEFAULT-NEXT: br i1 [[TMP12]], label %[[PRED_STORE_IF32:.*]], label %[[PRED_STORE_CONTINUE33:.*]]
+; DEFAULT-NEXT: br i1 [[TMP7]], label %[[PRED_STORE_IF28:.*]], label %[[PRED_STORE_CONTINUE29:.*]]
+; DEFAULT: [[PRED_STORE_IF28]]:
+; DEFAULT-NEXT: store i32 [[TMP5]], ptr [[E]], align 4, !alias.scope [[META15]], !noalias [[META17]]
+; DEFAULT-NEXT: br label %[[PRED_STORE_CONTINUE29]]
+; DEFAULT: [[PRED_STORE_CONTINUE29]]:
+; DEFAULT-NEXT: br i1 [[TMP7]], label %[[PRED_STORE_IF30:.*]], label %[[PRED_STORE_CONTINUE31:.*]]
+; DEFAULT: [[PRED_STORE_IF30]]:
+; DEFAULT-NEXT: store i32 [[TMP5]], ptr [[E]], align 4, !alias.scope [[META15]], !noalias [[META17]]
+; DEFAULT-NEXT: br label %[[PRED_STORE_CONTINUE31]]
+; DEFAULT: [[PRED_STORE_CONTINUE31]]:
+; DEFAULT-NEXT: br i1 [[TMP7]], label %[[PRED_STORE_IF32:.*]], label %[[PRED_STORE_CONTINUE33]]
; DEFAULT: [[PRED_STORE_IF32]]:
-; DEFAULT-NEXT: [[TMP13:%.*]] = extractelement <4 x i32> [[TMP6]], i32 0
-; DEFAULT-NEXT: store i32 [[TMP13]], ptr [[E]], align 4, !alias.scope [[META15]], !noalias [[META17]]
+; DEFAULT-NEXT: store i32 [[TMP5]], ptr [[E]], align 4, !alias.scope [[META15]], !noalias [[META17]]
; DEFAULT-NEXT: br label %[[PRED_STORE_CONTINUE33]]
; DEFAULT: [[PRED_STORE_CONTINUE33]]:
-; DEFAULT-NEXT: [[TMP14:%.*]] = extractelement <4 x i1> [[TMP8]], i32 2
-; DEFAULT-NEXT: br i1 [[TMP14]], label %[[PRED_STORE_IF34:.*]], label %[[PRED_STORE_CONTINUE35:.*]]
-; DEFAULT: [[PRED_STORE_IF34]]:
-; DEFAULT-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP6]], i32 0
-; DEFAULT-NEXT: store i32 [[TMP15]], ptr [[E]], align 4, !alias.scope [[META15]], !noalias [[META17]]
-; DEFAULT-NEXT: br label %[[PRED_STORE_CONTINUE35]]
-; DEFAULT: [[PRED_STORE_CONTINUE35]]:
-; DEFAULT-NEXT: [[TMP21:%.*]] = extractelement <4 x i1> [[TMP8]], i32 3
-; DEFAULT-NEXT: br i1 [[TMP21]], label %[[PRED_STORE_IF36:.*]], label %[[PRED_STORE_CONTINUE37]]
-; DEFAULT: [[PRED_STORE_IF36]]:
-; DEFAULT-NEXT: [[TMP22:%.*]] = extractelement <4 x i32> [[TMP6]], i32 0
-; DEFAULT-NEXT: store i32 [[TMP22]], ptr [[E]], align 4, !alias.scope [[META15]], !noalias [[META17]]
-; DEFAULT-NEXT: br label %[[PRED_STORE_CONTINUE37]]
-; DEFAULT: [[PRED_STORE_CONTINUE37]]:
-; DEFAULT-NEXT: call void @llvm.masked.store.v4i32.p0(<4 x i32> zeroinitializer, ptr align 4 [[TMP16]], <4 x i1> [[TMP8]]), !alias.scope [[META19:![0-9]+]], !noalias [[META20:![0-9]+]]
+; DEFAULT-NEXT: call void @llvm.masked.store.v4i32.p0(<4 x i32> zeroinitializer, ptr align 4 [[TMP16]], <4 x i1> [[BROADCAST_SPLAT]]), !alias.scope [[META19:![0-9]+]], !noalias [[META20:![0-9]+]]
; DEFAULT-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
; DEFAULT-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; DEFAULT-NEXT: br i1 [[TMP18]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
@@ -660,16 +648,16 @@ define void @low_trip_count_fold_tail_scalarized_store(ptr %dst) {
; COMMON-NEXT: store i8 6, ptr [[TMP6]], align 1
; COMMON-NEXT: br label %[[PRED_STORE_CONTINUE12]]
; COMMON: [[PRED_STORE_CONTINUE12]]:
-; COMMON-NEXT: br i1 false, label %[[PRED_STORE_IF13:.*]], label %[[EXIT:.*]]
+; COMMON-NEXT: br i1 false, label %[[PRED_STORE_IF13:.*]], label %[[PRED_STORE_CONTINUE14:.*]]
; COMMON: [[PRED_STORE_IF13]]:
; COMMON-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[DST]], i64 7
; COMMON-NEXT: store i8 7, ptr [[TMP7]], align 1
-; COMMON-NEXT: br label %[[EXIT]]
+; COMMON-NEXT: br label %[[PRED_STORE_CONTINUE14]]
+; COMMON: [[PRED_STORE_CONTINUE14]]:
+; COMMON-NEXT: br label %[[MIDDLE_BLOCK:.*]]
+; COMMON: [[MIDDLE_BLOCK]]:
+; COMMON-NEXT: br label %[[EXIT:.*]]
; COMMON: [[EXIT]]:
-; COMMON-NEXT: br label %[[SCALAR_PH:.*]]
-; COMMON: [[SCALAR_PH]]:
-; COMMON-NEXT: br label %[[EXIT1:.*]]
-; COMMON: [[EXIT1]]:
; COMMON-NEXT: ret void
;
entry:
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll b/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
index 5970608794b55..bea34e29e3530 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
@@ -16,7 +16,7 @@
; CM: vector.ph:
; CM: CLONE ir<%a> = extractvalue ir<%sv>
; CM: CLONE ir<%b> = extractvalue ir<%sv>
-; CM: WIDEN ir<%add> = add ir<%a>, ir<%b>
+; CM: CLONE ir<%add> = add ir<%a>, ir<%b>
; CM: Successor(s): vector loop
; CM: LV: Scalar loop costs: 5.
@@ -30,23 +30,22 @@ define void @test1(ptr %dst, {i64, i64} %sv) {
; FORCED-NEXT: br label %[[VECTOR_PH:.*]]
; FORCED: [[VECTOR_PH]]:
; FORCED-NEXT: [[TMP0:%.*]] = extractvalue { i64, i64 } [[SV]], 0
-; FORCED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i64> poison, i64 [[TMP0]], i64 0
-; FORCED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i64> [[BROADCAST_SPLATINSERT]], <2 x i64> poison, <2 x i32> zeroinitializer
; FORCED-NEXT: [[TMP4:%.*]] = extractvalue { i64, i64 } [[SV]], 1
-; FORCED-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <2 x i64> poison, i64 [[TMP4]], i64 0
+; FORCED-NEXT: [[TMP5:%.*]] = add i64 [[TMP0]], [[TMP4]]
+; FORCED-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <2 x i64> poison, i64 [[TMP5]], i64 0
; FORCED-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <2 x i64> [[BROADCAST_SPLATINSERT1]], <2 x i64> poison, <2 x i32> zeroinitializer
-; FORCED-NEXT: [[TMP1:%.*]] = add <2 x i64> [[BROADCAST_SPLAT]], [[BROADCAST_SPLAT2]]
; FORCED-NEXT: br label %[[VECTOR_BODY:.*]]
; FORCED: [[VECTOR_BODY]]:
; FORCED-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; FORCED-NEXT: [[TMP2:%.*]] = getelementptr i64, ptr [[DST]], i32 [[INDEX]]
-; FORCED-NEXT: store <2 x i64> [[TMP1]], ptr [[TMP2]], align 4
+; FORCED-NEXT: store <2 x i64> [[BROADCAST_SPLAT2]], ptr [[TMP2]], align 4
; FORCED-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
; FORCED-NEXT: [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
; FORCED-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; FORCED: [[MIDDLE_BLOCK]]:
-; FORCED-NEXT: br [[EXIT:label %.*]]
-; FORCED: [[SCALAR_PH:.*:]]
+; FORCED-NEXT: br label %[[EXIT:.*]]
+; FORCED: [[EXIT]]:
+; FORCED-NEXT: ret void
;
entry:
br label %loop.body
@@ -99,10 +98,11 @@ define void @test_getVectorCallCost(ptr %dst, {float, float} %sv) {
; FORCED-NEXT: store <2 x float> [[TMP2]], ptr [[TMP1]], align 4
; FORCED-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
; FORCED-NEXT: [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
-; FORCED-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; FORCED-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
; FORCED: [[MIDDLE_BLOCK]]:
-; FORCED-NEXT: br [[EXIT:label %.*]]
-; FORCED: [[SCALAR_PH:.*:]]
+; FORCED-NEXT: br label %[[EXIT:.*]]
+; FORCED: [[EXIT]]:
+; FORCED-NEXT: ret void
;
entry:
br label %loop.body
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
index 0c6a490ddf4ba..eceda0897b174 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-widen-extractvalue.ll
@@ -17,17 +17,15 @@ define void @widen_extractvalue(ptr %dst, {i64, i64} %sv) #0 {
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 1000, [[TMP3]]
; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 1000, [[N_MOD_VF]]
; CHECK-NEXT: [[EXTRACT0:%.*]] = extractvalue { i64, i64 } [[SV]], 0
-; CHECK-NEXT: [[DOTSPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[EXTRACT0]], i64 0
-; CHECK-NEXT: [[DOTSPLAT2:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
; CHECK-NEXT: [[TMP10:%.*]] = extractvalue { i64, i64 } [[SV]], 1
-; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP10]], i64 0
+; CHECK-NEXT: [[TMP6:%.*]] = add i64 [[EXTRACT0]], [[TMP10]]
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP6]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
-; CHECK-NEXT: [[TMP7:%.*]] = add <vscale x 2 x i64> [[DOTSPLAT2]], [[BROADCAST_SPLAT2]]
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i64, ptr [[DST]], i32 [[INDEX]]
-; CHECK-NEXT: store <vscale x 2 x i64> [[TMP7]], ptr [[TMP8]], align 8
+; CHECK-NEXT: store <vscale x 2 x i64> [[BROADCAST_SPLAT2]], ptr [[TMP8]], align 8
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP3]]
; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP9]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/dead-ops-cost.ll b/llvm/test/Transforms/LoopVectorize/RISCV/dead-ops-cost.ll
index f25b86d3b20c2..b81637f50989d 100644
--- a/llvm/test/Transforms/LoopVectorize/RISCV/dead-ops-cost.ll
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/dead-ops-cost.ll
@@ -293,9 +293,9 @@ define void @test_phi_in_latch_redundant(ptr %dst, i32 %a) {
; CHECK-NEXT: [[ENTRY:.*:]]
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
-; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[A]], i64 0
+; CHECK-NEXT: [[TMP0:%.*]] = xor i32 [[A]], -1
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP0]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
-; CHECK-NEXT: [[TMP19:%.*]] = xor <vscale x 4 x i32> [[BROADCAST_SPLAT]], splat (i32 -1)
; CHECK-NEXT: [[TMP6:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
; CHECK-NEXT: [[TMP7:%.*]] = mul <vscale x 4 x i64> [[TMP6]], splat (i64 9)
; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 4 x i64> zeroinitializer, [[TMP7]]
@@ -309,7 +309,7 @@ define void @test_phi_in_latch_redundant(ptr %dst, i32 %a) {
; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP9]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr i32, ptr [[DST]], <vscale x 4 x i64> [[VEC_IND]]
-; CHECK-NEXT: call void @llvm.vp.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[TMP19]], <vscale x 4 x ptr> align 4 [[TMP16]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP8]])
+; CHECK-NEXT: call void @llvm.vp.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT]], <vscale x 4 x ptr> align 4 [[TMP16]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP8]])
; CHECK-NEXT: [[AVL_NEXT]] = sub nuw i64 [[AVL]], [[TMP5]]
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[BROADCAST_SPLAT2]]
; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i64 [[AVL_NEXT]], 0
diff --git a/llvm/test/Transforms/LoopVectorize/pr50686.ll b/llvm/test/Transforms/LoopVectorize/pr50686.ll
index 878fbec452220..be9110ce0093a 100644
--- a/llvm/test/Transforms/LoopVectorize/pr50686.ll
+++ b/llvm/test/Transforms/LoopVectorize/pr50686.ll
@@ -18,20 +18,16 @@ define void @m(ptr nocapture %p, ptr nocapture %p2, i32 %q) {
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[P2]], align 4, !alias.scope [[META0:![0-9]+]]
-; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP1]], i64 0
-; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
-; CHECK-NEXT: [[TMP2:%.*]] = sub nsw <4 x i32> zeroinitializer, [[BROADCAST_SPLAT]]
-; CHECK-NEXT: [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX9_1]], align 4, !alias.scope [[META0]]
-; CHECK-NEXT: [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <4 x i32> poison, i32 [[TMP3]], i64 0
-; CHECK-NEXT: [[BROADCAST_SPLAT3:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT2]], <4 x i32> poison, <4 x i32> zeroinitializer
-; CHECK-NEXT: [[TMP4:%.*]] = sub nsw <4 x i32> [[TMP2]], [[BROADCAST_SPLAT3]]
-; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[ARRAYIDX9_2]], align 4, !alias.scope [[META0]]
+; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[P2]], align 4, !alias.scope [[META0:![0-9]+]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub nsw i32 0, [[TMP0]]
+; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[ARRAYIDX9_1]], align 4, !alias.scope [[META0]]
+; CHECK-NEXT: [[TMP3:%.*]] = sub nsw i32 [[TMP1]], [[TMP2]]
+; CHECK-NEXT: [[TMP4:%.*]] = load i32, ptr [[ARRAYIDX9_2]], align 4, !alias.scope [[META0]]
+; CHECK-NEXT: [[TMP5:%.*]] = sub nsw i32 [[TMP3]], [[TMP4]]
; CHECK-NEXT: [[BROADCAST_SPLATINSERT4:%.*]] = insertelement <4 x i32> poison, i32 [[TMP5]], i64 0
; CHECK-NEXT: [[BROADCAST_SPLAT5:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT4]], <4 x i32> poison, <4 x i32> zeroinitializer
-; CHECK-NEXT: [[TMP6:%.*]] = sub nsw <4 x i32> [[TMP4]], [[BROADCAST_SPLAT5]]
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[P]], i64 [[INDEX]]
-; CHECK-NEXT: store <4 x i32> [[TMP6]], ptr [[TMP7]], align 4, !alias.scope [[META3:![0-9]+]], !noalias [[META0]]
+; CHECK-NEXT: store <4 x i32> [[BROADCAST_SPLAT5]], ptr [[TMP7]], align 4, !alias.scope [[META3:![0-9]+]], !noalias [[META0]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
; CHECK-NEXT: [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 60
; CHECK-NEXT: br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
>From 4b0b89674174e82b73859df86b0fc8fe15dbffd0 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Wed, 5 Nov 2025 18:17:35 +0000
Subject: [PATCH 2/5] [VPlan] Forbid things like ComputeReductionResult
---
llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 6ad8ca5743114..50bfcecd066f9 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1423,7 +1423,10 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
return true;
if (auto *VPI = dyn_cast<VPInstruction>(U))
- if (VPI->isVectorToScalar() || VPI->isSingleScalar())
+ if (VPI->isSingleScalar() ||
+ VPI->getOpcode() == VPInstruction::ExtractLastElement ||
+ VPI->getOpcode() == VPInstruction::ExtractLastLanePerPart ||
+ VPI->getOpcode() == VPInstruction::ExtractPenultimateElement)
return true;
return U->usesScalars(RepOrWidenR);
>From 77093ba4e0e154f539e608dc71e55c98e5514217 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Wed, 5 Nov 2025 18:29:16 +0000
Subject: [PATCH 3/5] [VPlan] Increase clarity around WidenStore condition
(NFC)
---
llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 50bfcecd066f9..eac744703591a 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1419,8 +1419,9 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
// broadcasts.
if (!vputils::isSingleScalar(RepOrWidenR) ||
!all_of(RepOrWidenR->users(), [RepOrWidenR](const VPUser *U) {
- if (isa<VPWidenStoreRecipe>(U))
- return true;
+ if (auto *Store = dyn_cast<VPWidenStoreRecipe>(U))
+ if (vputils::isSingleScalar(Store->getStoredValue()))
+ return true;
if (auto *VPI = dyn_cast<VPInstruction>(U))
if (VPI->isSingleScalar() ||
>From 90215ac25fd75e613137f69b28b313aed029b727 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Thu, 6 Nov 2025 10:02:13 +0000
Subject: [PATCH 4/5] [VPlan] Rebase, add assert
---
.../Transforms/Vectorize/VPlanTransforms.cpp | 10 +++++++---
...first-order-recurrence-with-uniform-ops.ll | 20 ++++++-------------
2 files changed, 13 insertions(+), 17 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index eac744703591a..d7c8b5732b459 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1419,9 +1419,13 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
// broadcasts.
if (!vputils::isSingleScalar(RepOrWidenR) ||
!all_of(RepOrWidenR->users(), [RepOrWidenR](const VPUser *U) {
- if (auto *Store = dyn_cast<VPWidenStoreRecipe>(U))
- if (vputils::isSingleScalar(Store->getStoredValue()))
- return true;
+ if (auto *Store = dyn_cast<VPWidenStoreRecipe>(U)) {
+ // The assert must hold as we checked the RepOrWidenR operand
+ // against vputils::isSingleScalar.
+ assert(RepOrWidenR == Store->getAddr() ||
+ vputils::isSingleScalar(Store->getStoredValue()));
+ return true;
+ }
if (auto *VPI = dyn_cast<VPInstruction>(U))
if (VPI->isSingleScalar() ||
diff --git a/llvm/test/Transforms/LoopVectorize/first-order-recurrence-with-uniform-ops.ll b/llvm/test/Transforms/LoopVectorize/first-order-recurrence-with-uniform-ops.ll
index 8a579734a06e1..372876c5faac6 100644
--- a/llvm/test/Transforms/LoopVectorize/first-order-recurrence-with-uniform-ops.ll
+++ b/llvm/test/Transforms/LoopVectorize/first-order-recurrence-with-uniform-ops.ll
@@ -134,22 +134,18 @@ define i16 @for_phi_removed(ptr %src) {
; UNROLL-NO-IC: [[VECTOR_BODY]]:
; UNROLL-NO-IC-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; UNROLL-NO-IC-NEXT: [[TMP0:%.*]] = load i32, ptr [[SRC]], align 4
-; UNROLL-NO-IC-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP0]], i64 0
-; UNROLL-NO-IC-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
-; UNROLL-NO-IC-NEXT: [[TMP1:%.*]] = icmp eq <4 x i32> [[BROADCAST_SPLAT]], zeroinitializer
-; UNROLL-NO-IC-NEXT: [[TMP4:%.*]] = extractelement <4 x i1> [[TMP1]], i32 0
-; UNROLL-NO-IC-NEXT: [[TMP2:%.*]] = select i1 [[TMP4]], <4 x i16> splat (i16 1), <4 x i16> zeroinitializer
+; UNROLL-NO-IC-NEXT: [[TMP1:%.*]] = icmp eq i32 [[TMP0]], 0
+; UNROLL-NO-IC-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i16 1, i16 0
; UNROLL-NO-IC-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
; UNROLL-NO-IC-NEXT: [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 104
; UNROLL-NO-IC-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; UNROLL-NO-IC: [[MIDDLE_BLOCK]]:
-; UNROLL-NO-IC-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP2]], i32 3
; UNROLL-NO-IC-NEXT: br label %[[SCALAR_PH:.*]]
; UNROLL-NO-IC: [[SCALAR_PH]]:
; UNROLL-NO-IC-NEXT: br label %[[LOOP:.*]]
; UNROLL-NO-IC: [[LOOP]]:
; UNROLL-NO-IC-NEXT: [[IV:%.*]] = phi i16 [ 104, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
-; UNROLL-NO-IC-NEXT: [[P:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
+; UNROLL-NO-IC-NEXT: [[P:%.*]] = phi i16 [ [[TMP2]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
; UNROLL-NO-IC-NEXT: [[L:%.*]] = load i32, ptr [[SRC]], align 4
; UNROLL-NO-IC-NEXT: [[C:%.*]] = icmp eq i32 [[L]], 0
; UNROLL-NO-IC-NEXT: [[SEL]] = select i1 [[C]], i16 1, i16 0
@@ -200,22 +196,18 @@ define i16 @for_phi_removed(ptr %src) {
; SINK-AFTER: [[VECTOR_BODY]]:
; SINK-AFTER-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; SINK-AFTER-NEXT: [[TMP0:%.*]] = load i32, ptr [[SRC]], align 4
-; SINK-AFTER-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP0]], i64 0
-; SINK-AFTER-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
-; SINK-AFTER-NEXT: [[TMP1:%.*]] = icmp eq <4 x i32> [[BROADCAST_SPLAT]], zeroinitializer
-; SINK-AFTER-NEXT: [[TMP4:%.*]] = extractelement <4 x i1> [[TMP1]], i32 0
-; SINK-AFTER-NEXT: [[TMP2:%.*]] = select i1 [[TMP4]], <4 x i16> splat (i16 1), <4 x i16> zeroinitializer
+; SINK-AFTER-NEXT: [[TMP1:%.*]] = icmp eq i32 [[TMP0]], 0
+; SINK-AFTER-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i16 1, i16 0
; SINK-AFTER-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
; SINK-AFTER-NEXT: [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 108
; SINK-AFTER-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; SINK-AFTER: [[MIDDLE_BLOCK]]:
-; SINK-AFTER-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP2]], i32 3
; SINK-AFTER-NEXT: br label %[[SCALAR_PH:.*]]
; SINK-AFTER: [[SCALAR_PH]]:
; SINK-AFTER-NEXT: br label %[[LOOP:.*]]
; SINK-AFTER: [[LOOP]]:
; SINK-AFTER-NEXT: [[IV:%.*]] = phi i16 [ 108, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
-; SINK-AFTER-NEXT: [[P:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
+; SINK-AFTER-NEXT: [[P:%.*]] = phi i16 [ [[TMP2]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
; SINK-AFTER-NEXT: [[L:%.*]] = load i32, ptr [[SRC]], align 4
; SINK-AFTER-NEXT: [[C:%.*]] = icmp eq i32 [[L]], 0
; SINK-AFTER-NEXT: [[SEL]] = select i1 [[C]], i16 1, i16 0
>From bd76cb59c934c912be43fa096623ac7669e4ed3b Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Fri, 7 Nov 2025 08:54:34 +0000
Subject: [PATCH 5/5] [VPlan] Address review
---
.../Transforms/Vectorize/VPlanTransforms.cpp | 21 ++++++++++++-------
1 file changed, 14 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index d7c8b5732b459..6726a6e71488d 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1420,19 +1420,26 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
if (!vputils::isSingleScalar(RepOrWidenR) ||
!all_of(RepOrWidenR->users(), [RepOrWidenR](const VPUser *U) {
if (auto *Store = dyn_cast<VPWidenStoreRecipe>(U)) {
- // The assert must hold as we checked the RepOrWidenR operand
- // against vputils::isSingleScalar.
+ // VPWidenStore doesn't have users, and stores are always
+ // profitable to widen: hence, permitting single-scalar stored
+ // values is an important leaf condition. The assert must hold as
+ // we checked the RepOrWidenR operand against
+ // vputils::isSingleScalar.
assert(RepOrWidenR == Store->getAddr() ||
vputils::isSingleScalar(Store->getStoredValue()));
return true;
}
- if (auto *VPI = dyn_cast<VPInstruction>(U))
- if (VPI->isSingleScalar() ||
- VPI->getOpcode() == VPInstruction::ExtractLastElement ||
- VPI->getOpcode() == VPInstruction::ExtractLastLanePerPart ||
- VPI->getOpcode() == VPInstruction::ExtractPenultimateElement)
+ if (auto *VPI = dyn_cast<VPInstruction>(U)) {
+ switch (VPI->getOpcode()) {
+ case VPInstruction::ExtractLastElement:
+ case VPInstruction::ExtractLastLanePerPart:
+ case VPInstruction::ExtractPenultimateElement:
return true;
+ default:
+ return false;
+ }
+ }
return U->usesScalars(RepOrWidenR);
}))
More information about the llvm-commits
mailing list