[X86][SSE] SimplifyMultipleUseDemandedBits - Add PEXTR/PINSR B+W handling

This adds SimplifyMultipleUseDemandedBitsForTargetNode X86 support and uses it to allow us to peek through vector insertions to avoid dependencies on entire insertion chains.

llvm-svn: 367570

3 files changed, 39 insertions, 8 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 6fe906de8cf..c2be03f8d90 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -34577,6 +34577,11 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
                                KnownVec, TLO, Depth + 1))
         return true;
 
+      if (SDValue V = SimplifyMultipleUseDemandedBits(
+              Vec, DemandedVecBits, DemandedVecElts, TLO.DAG, Depth + 1))
+        return TLO.CombineTo(
+            Op, TLO.DAG.getNode(Opc, SDLoc(Op), VT, V, Op.getOperand(1)));
+
       Known = KnownVec.zext(BitWidth, true);
       return false;
     }
@@ -34678,6 +34683,28 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
       Op, OriginalDemandedBits, OriginalDemandedElts, Known, TLO, Depth);
 }
 
+SDValue X86TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode(
+    SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
+    SelectionDAG &DAG, unsigned Depth) const {
+  unsigned Opc = Op.getOpcode();
+  switch (Opc) {
+  case X86ISD::PINSRB:
+  case X86ISD::PINSRW: {
+    // If we don't demand the inserted element, return the base vector.
+    SDValue Vec = Op.getOperand(0);
+    auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2));
+    MVT VecVT = Vec.getSimpleValueType();
+    if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements()) &&
+        !DemandedElts[CIdx->getZExtValue()])
+      return Vec;
+     break;
+  }
+  }
+
+  return TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode(
+      Op, DemandedBits, DemandedElts, DAG, Depth);
+}
+
 /// Check if a vector extract from a target-specific shuffle of a load can be
 /// folded into a single element load.
 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
index dd11cc4497a..625b42d3515 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -920,6 +920,10 @@ namespace llvm {
                                            TargetLoweringOpt &TLO,
                                            unsigned Depth) const override;
 
+    SDValue SimplifyMultipleUseDemandedBitsForTargetNode(
+        SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
+        SelectionDAG &DAG, unsigned Depth) const override;
+
     const Constant *getTargetConstantFromLoad(LoadSDNode *LD) const override;
 
     SDValue unwrapAddress(SDValue N) const override;
diff --git a/llvm/test/CodeGen/X86/promote-vec3.ll b/llvm/test/CodeGen/X86/promote-vec3.ll
index cb4e99b3b18..b66570e8aaa 100644
--- a/llvm/test/CodeGen/X86/promote-vec3.ll
+++ b/llvm/test/CodeGen/X86/promote-vec3.ll
@@ -8,14 +8,14 @@
 define <3 x i16> @zext_i8(<3 x i8>) {
 ; SSE3-LABEL: zext_i8:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; SSE3-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; SSE3-NEXT:    movzbl {{[0-9]+}}(%esp), %edx
-; SSE3-NEXT:    movd %edx, %xmm0
-; SSE3-NEXT:    pinsrw $1, %ecx, %xmm0
-; SSE3-NEXT:    pinsrw $2, %eax, %xmm0
+; SSE3-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
+; SSE3-NEXT:    movd %eax, %xmm0
 ; SSE3-NEXT:    pextrw $0, %xmm0, %eax
+; SSE3-NEXT:    pinsrw $1, %edx, %xmm0
 ; SSE3-NEXT:    pextrw $1, %xmm0, %edx
+; SSE3-NEXT:    pinsrw $2, %ecx, %xmm0
 ; SSE3-NEXT:    pextrw $2, %xmm0, %ecx
 ; SSE3-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE3-NEXT:    # kill: def $dx killed $dx killed $edx
@@ -27,9 +27,9 @@ define <3 x i16> @zext_i8(<3 x i8>) {
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
 ; SSE41-NEXT:    pinsrb $0, {{[0-9]+}}(%esp), %xmm0
 ; SSE41-NEXT:    pinsrb $4, {{[0-9]+}}(%esp), %xmm0
+; SSE41-NEXT:    pextrw $2, %xmm0, %edx
 ; SSE41-NEXT:    pinsrb $8, {{[0-9]+}}(%esp), %xmm0
 ; SSE41-NEXT:    movd %xmm0, %eax
-; SSE41-NEXT:    pextrw $2, %xmm0, %edx
 ; SSE41-NEXT:    pextrw $4, %xmm0, %ecx
 ; SSE41-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE41-NEXT:    # kill: def $dx killed $dx killed $edx
@@ -41,10 +41,10 @@ define <3 x i16> @zext_i8(<3 x i8>) {
 ; AVX-32-NEXT:    vpxor %xmm0, %xmm0, %xmm0
 ; AVX-32-NEXT:    vpinsrb $0, {{[0-9]+}}(%esp), %xmm0, %xmm0
 ; AVX-32-NEXT:    vpinsrb $4, {{[0-9]+}}(%esp), %xmm0, %xmm0
-; AVX-32-NEXT:    vpinsrb $8, {{[0-9]+}}(%esp), %xmm0, %xmm0
-; AVX-32-NEXT:    vmovd %xmm0, %eax
+; AVX-32-NEXT:    vpinsrb $8, {{[0-9]+}}(%esp), %xmm0, %xmm1
 ; AVX-32-NEXT:    vpextrw $2, %xmm0, %edx
-; AVX-32-NEXT:    vpextrw $4, %xmm0, %ecx
+; AVX-32-NEXT:    vmovd %xmm1, %eax
+; AVX-32-NEXT:    vpextrw $4, %xmm1, %ecx
 ; AVX-32-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX-32-NEXT:    # kill: def $dx killed $dx killed $edx
 ; AVX-32-NEXT:    # kill: def $cx killed $cx killed $ecx