[llvm] [VPlan] Support VPWidenCastRecipe in narrowToSingleScalarRecipes. (PR #141080)

[Mel Chen via llvm-commits]

https://github.com/Mel-Chen updated https://github.com/llvm/llvm-project/pull/141080

>From d0cd55038232154cde278346f1438e713ccf3fb1 Mon Sep 17 00:00:00 2001
From: Mel Chen <mel.chen at sifive.com>
Date: Mon, 19 May 2025 03:34:40 -0700
Subject: [PATCH] patch widen-cast in isSingleScalar

---
 llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp    | 2 +-
 llvm/lib/Transforms/Vectorize/VPlanUtils.h           | 5 +++++
 llvm/test/Transforms/LoopVectorize/X86/cost-model.ll | 4 ++--
 3 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 34633cd748eb1..4655e911ab3a1 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1175,7 +1175,7 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
            vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
     for (VPRecipeBase &R : make_early_inc_range(reverse(*VPBB))) {
       auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
-      if (!RepR && !isa<VPWidenRecipe>(&R))
+      if (!RepR && !isa<VPWidenRecipe, VPWidenCastRecipe>(&R))
         continue;
       if (RepR && (RepR->isSingleScalar() || RepR->isPredicated()))
         continue;
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUtils.h b/llvm/lib/Transforms/Vectorize/VPlanUtils.h
index 28c1a6af2570b..22f8086e7bdde 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUtils.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanUtils.h
@@ -75,6 +75,11 @@ inline bool isSingleScalar(const VPValue *VPV) {
     return PreservesUniformity(WidenR->getOpcode()) &&
            all_of(WidenR->operands(), isSingleScalar);
   }
+  if (auto *CastR = dyn_cast<VPWidenCastRecipe>(VPV)) {
+    return PreservesUniformity(CastR->getOpcode()) &&
+           all_of(CastR->operands(), isSingleScalar);
+  }
+
   if (auto *VPI = dyn_cast<VPInstruction>(VPV))
     return VPI->isSingleScalar() || VPI->isVectorToScalar() ||
            (PreservesUniformity(VPI->getOpcode()) &&
diff --git a/llvm/test/Transforms/LoopVectorize/X86/cost-model.ll b/llvm/test/Transforms/LoopVectorize/X86/cost-model.ll
index 7c42c3d9cd52e..fca0caa5cce31 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/cost-model.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/cost-model.ll
@@ -378,8 +378,8 @@ define void @multi_exit(ptr %dst, ptr %src.1, ptr %src.2, i64 %A, i64 %B) #0 {
 ; CHECK-NEXT:    [[TMP15:%.*]] = icmp eq <2 x i64> [[BROADCAST_SPLAT]], zeroinitializer
 ; CHECK-NEXT:    [[TMP16:%.*]] = icmp ne <2 x i64> [[BROADCAST_SPLAT10]], zeroinitializer
 ; CHECK-NEXT:    [[TMP17:%.*]] = and <2 x i1> [[TMP16]], [[TMP15]]
-; CHECK-NEXT:    [[TMP18:%.*]] = zext <2 x i1> [[TMP17]] to <2 x i8>
-; CHECK-NEXT:    [[TMP19:%.*]] = extractelement <2 x i8> [[TMP18]], i32 1
+; CHECK-NEXT:    [[TMP18:%.*]] = extractelement <2 x i1> [[TMP17]], i32 0
+; CHECK-NEXT:    [[TMP19:%.*]] = zext i1 [[TMP18]] to i8
 ; CHECK-NEXT:    store i8 [[TMP19]], ptr [[DST]], align 1, !alias.scope [[META10:![0-9]+]], !noalias [[META12:![0-9]+]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]