[llvm] [X86] getFauxShuffleMask - move INSERT_SUBVECTOR(SRC0, EXTRACT_SUBVECTOR(SRC1)) matching behind common one use bitcast checks (PR #134227)

[Simon Pilgrim via llvm-commits]

https://github.com/RKSimon updated https://github.com/llvm/llvm-project/pull/134227

>From 25e806f96b410d5a0d75ab67e793714762e628f8 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Thu, 3 Apr 2025 11:13:25 +0100
Subject: [PATCH 1/2] [X86] getFauxShuffleMask - move INSERT_SUBVECTOR(SRC0,
 EXTRACT_SUBVECTOR(SRC1)) matching behind common one use bitcast checks

No need to ignore one use checks for the INSERT_SUBVECTOR(SRC0, EXTRACT_SUBVECTOR(SRC1)) fold

Noticed while working on the #133947 regressions
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 29 +++++++++++++------------
 1 file changed, 15 insertions(+), 14 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 34716929f61f1..79cfe77669137 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6185,13 +6185,20 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
     }
     if (!N->isOnlyUserOf(Sub.getNode()))
       return false;
-    SDValue SubBC = peekThroughBitcasts(Sub);
+
+    SmallVector<int, 64> SubMask;
+    SmallVector<SDValue, 2> SubInputs;
+    SDValue SubSrc = peekThroughOneUseBitcasts(Sub);
+    EVT SubSrcVT = SubSrc.getValueType();
+    if (!SubSrcVT.isVector())
+      return false;
+
     // Handle INSERT_SUBVECTOR(SRC0, EXTRACT_SUBVECTOR(SRC1)).
-    if (SubBC.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
-        SubBC.getOperand(0).getValueSizeInBits() == NumSizeInBits) {
-      uint64_t ExtractIdx = SubBC.getConstantOperandVal(1);
-      SDValue SubBCSrc = SubBC.getOperand(0);
-      unsigned NumSubSrcBCElts = SubBCSrc.getValueType().getVectorNumElements();
+    if (SubSrc.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+        SubSrc.getOperand(0).getValueSizeInBits() == NumSizeInBits) {
+      uint64_t ExtractIdx = SubSrc.getConstantOperandVal(1);
+      SDValue SubSrcSrc = SubSrc.getOperand(0);
+      unsigned NumSubSrcBCElts = SubSrcSrc.getValueType().getVectorNumElements();
       unsigned MaxElts = std::max(NumElts, NumSubSrcBCElts);
       assert((MaxElts % NumElts) == 0 && (MaxElts % NumSubSrcBCElts) == 0 &&
              "Subvector valuetype mismatch");
@@ -6205,17 +6212,11 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
         Mask[InsertIdx + i] = (SrcIsUndef ? 0 : MaxElts) + ExtractIdx + i;
       if (!SrcIsUndef)
         Ops.push_back(Src);
-      Ops.push_back(SubBCSrc);
+      Ops.push_back(SubSrcSrc);
       return true;
     }
-    // Handle INSERT_SUBVECTOR(SRC0, SHUFFLE(SRC1)).
-    SmallVector<int, 64> SubMask;
-    SmallVector<SDValue, 2> SubInputs;
-    SDValue SubSrc = peekThroughOneUseBitcasts(Sub);
-    EVT SubSrcVT = SubSrc.getValueType();
-    if (!SubSrcVT.isVector())
-      return false;
 
+    // Handle INSERT_SUBVECTOR(SRC0, SHUFFLE(SRC1)).
     APInt SubDemand = APInt::getAllOnes(SubSrcVT.getVectorNumElements());
     if (!getTargetShuffleInputs(SubSrc, SubDemand, SubInputs, SubMask, DAG,
                                 Depth + 1, ResolveKnownElts))

>From 656ebdfa693acdda2bb88eb3a0f3bfa60c30d077 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Thu, 3 Apr 2025 12:23:05 +0100
Subject: [PATCH 2/2] clang-format

---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 79cfe77669137..8271d9c486650 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6198,12 +6198,13 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
         SubSrc.getOperand(0).getValueSizeInBits() == NumSizeInBits) {
       uint64_t ExtractIdx = SubSrc.getConstantOperandVal(1);
       SDValue SubSrcSrc = SubSrc.getOperand(0);
-      unsigned NumSubSrcBCElts = SubSrcSrc.getValueType().getVectorNumElements();
-      unsigned MaxElts = std::max(NumElts, NumSubSrcBCElts);
-      assert((MaxElts % NumElts) == 0 && (MaxElts % NumSubSrcBCElts) == 0 &&
+      unsigned NumSubSrcSrcElts =
+          SubSrcSrc.getValueType().getVectorNumElements();
+      unsigned MaxElts = std::max(NumElts, NumSubSrcSrcElts);
+      assert((MaxElts % NumElts) == 0 && (MaxElts % NumSubSrcSrcElts) == 0 &&
              "Subvector valuetype mismatch");
       InsertIdx *= (MaxElts / NumElts);
-      ExtractIdx *= (MaxElts / NumSubSrcBCElts);
+      ExtractIdx *= (MaxElts / NumSubSrcSrcElts);
       NumSubElts *= (MaxElts / NumElts);
       bool SrcIsUndef = Src.isUndef();
       for (int i = 0; i != (int)MaxElts; ++i)
@@ -6231,10 +6232,11 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
 
     if (SubMask.size() != NumSubElts) {
       assert(((SubMask.size() % NumSubElts) == 0 ||
-              (NumSubElts % SubMask.size()) == 0) && "Illegal submask scale");
+              (NumSubElts % SubMask.size()) == 0) &&
+             "Illegal submask scale");
       if ((NumSubElts % SubMask.size()) == 0) {
         int Scale = NumSubElts / SubMask.size();
-        SmallVector<int,64> ScaledSubMask;
+        SmallVector<int, 64> ScaledSubMask;
         narrowShuffleMaskElts(Scale, SubMask, ScaledSubMask);
         SubMask = ScaledSubMask;
       } else {