[X86] Add freeze(pshufd/permilps(x,imm)) -> pshufd/permilps(freeze(x)…

…,imm) folding Add X86 isGuaranteedNotToBeUndefOrPoisonForTargetNode / canCreateUndefOrPoisonForTargetNode overrides and add X86ISD::PSHUFD/VPERMILPI handling.
llvm · Oct 23, 2022 · c175d88 · c175d88
1 parent 480e9a9
commit c175d88
Show file tree

Hide file tree

Showing 3 changed files with 48 additions and 4 deletions.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -42550,6 +42550,46 @@ SDValue X86TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode(
       Op, DemandedBits, DemandedElts, DAG, Depth);
 }
 
+bool X86TargetLowering::isGuaranteedNotToBeUndefOrPoisonForTargetNode(
+    SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
+    bool PoisonOnly, unsigned Depth) const {
+  unsigned EltsBits = Op.getScalarValueSizeInBits();
+  unsigned NumElts = DemandedElts.getBitWidth();
+
+  // TODO: Add more target shuffles.
+  switch (Op.getOpcode()) {
+  case X86ISD::PSHUFD:
+  case X86ISD::VPERMILPI: {
+    SmallVector<int, 8> Mask;
+    DecodePSHUFMask(NumElts, EltsBits, Op.getConstantOperandVal(1), Mask);
+
+    APInt DemandedSrcElts = APInt::getZero(NumElts);
+    for (unsigned I = 0; I != NumElts; ++I)
+      if (DemandedElts[I])
+        DemandedSrcElts.setBit(Mask[I]);
+
+    return DAG.isGuaranteedNotToBeUndefOrPoison(
+        Op.getOperand(0), DemandedSrcElts, PoisonOnly, Depth + 1);
+  }
+  }
+  return TargetLowering::isGuaranteedNotToBeUndefOrPoisonForTargetNode(
+      Op, DemandedElts, DAG, PoisonOnly, Depth);
+}
+
+bool X86TargetLowering::canCreateUndefOrPoisonForTargetNode(
+    SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
+    bool PoisonOnly, bool ConsiderFlags, unsigned Depth) const {
+
+  // TODO: Add more target shuffles.
+  switch (Op.getOpcode()) {
+  case X86ISD::PSHUFD:
+  case X86ISD::VPERMILPI:
+    return false;
+  }
+  return TargetLowering::canCreateUndefOrPoisonForTargetNode(
+      Op, DemandedElts, DAG, PoisonOnly, ConsiderFlags, Depth);
+}
+
 bool X86TargetLowering::isSplatValueForTargetNode(SDValue Op,
                                                   const APInt &DemandedElts,
                                                   APInt &UndefElts,

diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -1158,6 +1158,14 @@ namespace llvm {
         SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
         SelectionDAG &DAG, unsigned Depth) const override;
 
+    bool isGuaranteedNotToBeUndefOrPoisonForTargetNode(
+        SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
+        bool PoisonOnly, unsigned Depth) const override;
+
+    bool canCreateUndefOrPoisonForTargetNode(
+        SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
+        bool PoisonOnly, bool ConsiderFlags, unsigned Depth) const override;
+
     bool isSplatValueForTargetNode(SDValue Op, const APInt &DemandedElts,
                                    APInt &UndefElts,
                                    unsigned Depth) const override;

diff --git a/llvm/test/CodeGen/X86/freeze-vector.ll b/llvm/test/CodeGen/X86/freeze-vector.ll
@@ -16,8 +16,6 @@ define <4 x i32> @freeze_insert_subvector(<8 x i32> %a0) nounwind {
 define <4 x i32> @freeze_pshufd(<4 x i32> %a0) nounwind {
 ; CHECK-LABEL: freeze_pshufd:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,2,1,0]
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,2,1,0]
 ; CHECK-NEXT:    ret{{[l|q]}}
   %x = shufflevector <4 x i32> %a0, <4 x i32> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   %y = freeze <4 x i32> %x
@@ -28,8 +26,6 @@ define <4 x i32> @freeze_pshufd(<4 x i32> %a0) nounwind {
 define <4 x float> @freeze_permilps(<4 x float> %a0) nounwind {
 ; CHECK-LABEL: freeze_permilps:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,2,1,0]
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,2,1,0]
 ; CHECK-NEXT:    ret{{[l|q]}}
   %x = shufflevector <4 x float> %a0, <4 x float> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   %y = freeze <4 x float> %x