[llvm] [SLP][REVEC] Fix getStoreMinimumVF only accept scalar types. (PR #132181)

[Han-Kuan Chen via llvm-commits]

https://github.com/HanKuanChen created https://github.com/llvm/llvm-project/pull/132181

Fix "Element type of a VectorType must " "be an integer, floating point,
or " "pointer type.".

>From a9094e1260d98cabffe633870b840bd7039dfba2 Mon Sep 17 00:00:00 2001
From: Han-Kuan Chen <hankuan.chen at sifive.com>
Date: Thu, 20 Mar 2025 03:19:39 -0700
Subject: [PATCH 1/2] [SLP][REVEC] Pre-commit test.

---
 .../SLPVectorizer/X86/revec-getStoreMinimumVF.ll       | 10 ++++++++++
 1 file changed, 10 insertions(+)
 create mode 100644 llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll

diff --git a/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll b/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll
new file mode 100644
index 0000000000000..ed7728eb46a8a
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll
@@ -0,0 +1,10 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=x86_64-unknown-linux-gnu -passes=slp-vectorizer -S -slp-revec %s | FileCheck %s
+
+define void @test() {
+entry:
+  %0 = getelementptr i8, ptr null, i64 4
+  store <4 x i8> zeroinitializer, ptr null, align 1
+  store <4 x i8> zeroinitializer, ptr %0, align 1
+  ret void
+}

>From a52fcc0f1562315ed8f10ab36623859633f6d845 Mon Sep 17 00:00:00 2001
From: Han-Kuan Chen <hankuan.chen at sifive.com>
Date: Thu, 20 Mar 2025 03:12:17 -0700
Subject: [PATCH 2/2] [SLP][REVEC] Fix getStoreMinimumVF only accept scalar
 types.

Fix "Element type of a VectorType must " "be an integer, floating point,
or " "pointer type.".
---
 llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp  | 16 ++++++++++++----
 .../SLPVectorizer/X86/revec-getStoreMinimumVF.ll |  7 +++++++
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index a2200f283168d..ad6e0a838cc1b 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -19697,10 +19697,18 @@ bool SLPVectorizerPass::vectorizeStores(
       Type *ValueTy = StoreTy;
       if (auto *Trunc = dyn_cast<TruncInst>(Store->getValueOperand()))
         ValueTy = Trunc->getSrcTy();
-      unsigned MinVF = std::max<unsigned>(
-          2, PowerOf2Ceil(TTI->getStoreMinimumVF(
-                 R.getMinVF(DL->getTypeStoreSizeInBits(StoreTy)), StoreTy,
-                 ValueTy)));
+      // When REVEC is enabled, StoreTy and ValueTy may be FixedVectorType. But
+      // getStoreMinimumVF only support scalar type as arguments. As a result,
+      // we need to use the element type of StoreTy and ValueTy to retrieve the
+      // VF and then transform it back.
+      // Remember: VF is defined as the number we want to vectorize, not the
+      // number of elements in the final vector.
+      Type *StoreScalarTy = StoreTy->getScalarType();
+      unsigned MinVF = PowerOf2Ceil(TTI->getStoreMinimumVF(
+          R.getMinVF(DL->getTypeStoreSizeInBits(StoreScalarTy)), StoreScalarTy,
+          ValueTy->getScalarType()));
+      MinVF /= getNumElements(StoreTy);
+      MinVF = std::max<unsigned>(2, MinVF);
 
       if (MaxVF < MinVF) {
         LLVM_DEBUG(dbgs() << "SLP: Vectorization infeasible as MaxVF (" << MaxVF
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll b/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll
index ed7728eb46a8a..3aea112e9edfe 100644
--- a/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/revec-getStoreMinimumVF.ll
@@ -2,6 +2,13 @@
 ; RUN: opt -mtriple=x86_64-unknown-linux-gnu -passes=slp-vectorizer -S -slp-revec %s | FileCheck %s
 
 define void @test() {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call <8 x i8> @llvm.vector.insert.v8i8.v4i8(<8 x i8> poison, <4 x i8> zeroinitializer, i64 0)
+; CHECK-NEXT:    [[TMP1:%.*]] = call <8 x i8> @llvm.vector.insert.v8i8.v4i8(<8 x i8> [[TMP0]], <4 x i8> zeroinitializer, i64 4)
+; CHECK-NEXT:    store <8 x i8> [[TMP1]], ptr null, align 1
+; CHECK-NEXT:    ret void
+;
 entry:
   %0 = getelementptr i8, ptr null, i64 4
   store <4 x i8> zeroinitializer, ptr null, align 1