[X86] More aggressive shuffle mask widening in combineExtractWithShuffle

Use demanded extract index to set most of the shuffle mask to undef, making it easier to widen and peek through.

llvm-svn: 351013

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 06802aa..843d287 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -33691,6 +33691,15 @@ static SDValue combineExtractWithShuffle(SDNode *N, SelectionDAG &DAG,
       scaleShuffleMask<int>(Scale, Mask, ScaledMask);
       Mask = std::move(ScaledMask);
     } else if ((Mask.size() % NumSrcElts) == 0) {
+      // Simplify Mask based on demanded element.
+      int ExtractIdx = (int)N->getConstantOperandVal(1);
+      int Scale = Mask.size() / NumSrcElts;
+      int Lo = Scale * ExtractIdx;
+      int Hi = Scale * (ExtractIdx + 1);
+      for (int i = 0, e = (int)Mask.size(); i != e; ++i)
+        if (i < Lo || Hi <= i)
+          Mask[i] = SM_SentinelUndef;
+
       SmallVector<int, 16> WidenedMask;
       while (Mask.size() > NumSrcElts &&
              canWidenShuffleElements(Mask, WidenedMask))

diff --git a/llvm/test/CodeGen/X86/and-load-fold.ll b/llvm/test/CodeGen/X86/and-load-fold.ll
index 27520f8..367ef2a 100644 (file)
--- a/llvm/test/CodeGen/X86/and-load-fold.ll
+++ b/llvm/test/CodeGen/X86/and-load-fold.ll
@@ -7,8 +7,7 @@ define i8 @foo(<4 x i8>* %V) {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; CHECK-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; CHECK-NEXT:    pextrw $2, %xmm0, %eax
+; CHECK-NEXT:    pextrw $1, %xmm0, %eax
 ; CHECK-NEXT:    andb $95, %al
 ; CHECK-NEXT:    # kill: def $al killed $al killed $eax
 ; CHECK-NEXT:    retq