[VectorCombine] Relax vector type constraint on bitop(bitcast, bitcast)

Author: XChy

llvm/lib/Transforms/Vectorize/VectorCombine.cpp

@@ -870,38 +870,37 @@ bool VectorCombine::foldBitOpOfCastops(Instruction &I) {
   if (LHSSrc->getType() != RHSSrc->getType())
     return false;
 
-  // Only handle vector types with integer elements
-  auto *SrcVecTy = dyn_cast<FixedVectorType>(LHSSrc->getType());
-  auto *DstVecTy = dyn_cast<FixedVectorType>(I.getType());
-  if (!SrcVecTy || !DstVecTy)
+  auto *SrcTy = LHSSrc->getType();
+  auto *DstTy = I.getType();
+  // Only handle vector types with integer elements if the cast is not bitcast
+  if (CastOpcode != Instruction::BitCast &&
+      (!isa<FixedVectorType>(SrcTy) || !isa<FixedVectorType>(DstTy)))
     return false;
 
-  if (!SrcVecTy->getScalarType()->isIntegerTy() ||
-      !DstVecTy->getScalarType()->isIntegerTy())
+  if (!SrcTy->getScalarType()->isIntegerTy() ||
+      !DstTy->getScalarType()->isIntegerTy())
     return false;
 
   // Cost Check :
   // OldCost = bitlogic + 2*casts
   // NewCost = bitlogic + cast
 
   // Calculate specific costs for each cast with instruction context
-  InstructionCost LHSCastCost =
-      TTI.getCastInstrCost(CastOpcode, DstVecTy, SrcVecTy,
-                           TTI::CastContextHint::None, CostKind, LHSCast);
-  InstructionCost RHSCastCost =
-      TTI.getCastInstrCost(CastOpcode, DstVecTy, SrcVecTy,
-                           TTI::CastContextHint::None, CostKind, RHSCast);
+  InstructionCost LHSCastCost = TTI.getCastInstrCost(
+      CastOpcode, DstTy, SrcTy, TTI::CastContextHint::None, CostKind, LHSCast);
+  InstructionCost RHSCastCost = TTI.getCastInstrCost(
+      CastOpcode, DstTy, SrcTy, TTI::CastContextHint::None, CostKind, RHSCast);
 
   InstructionCost OldCost =
-      TTI.getArithmeticInstrCost(BinOp->getOpcode(), DstVecTy, CostKind) +
+      TTI.getArithmeticInstrCost(BinOp->getOpcode(), DstTy, CostKind) +
       LHSCastCost + RHSCastCost;
 
   // For new cost, we can't provide an instruction (it doesn't exist yet)
   InstructionCost GenericCastCost = TTI.getCastInstrCost(
-      CastOpcode, DstVecTy, SrcVecTy, TTI::CastContextHint::None, CostKind);
+      CastOpcode, DstTy, SrcTy, TTI::CastContextHint::None, CostKind);
 
   InstructionCost NewCost =
-      TTI.getArithmeticInstrCost(BinOp->getOpcode(), SrcVecTy, CostKind) +
+      TTI.getArithmeticInstrCost(BinOp->getOpcode(), SrcTy, CostKind) +
       GenericCastCost;
 
   // Account for multi-use casts using specific costs

llvm/test/Transforms/VectorCombine/X86/bitop-of-castops.ll

@@ -510,10 +510,9 @@ define <16 x i1> @xor_bitcast_i16_to_v16i1_constant(i16 %a) {
 
 define <16 x i1> @xor_bitcast_i16_to_v16i1(i16 %a, i16 %b) {
 ; CHECK-LABEL: @xor_bitcast_i16_to_v16i1(
-; CHECK-NEXT:    [[BC1:%.*]] = bitcast i16 [[A:%.*]] to <16 x i1>
-; CHECK-NEXT:    [[BC2:%.*]] = bitcast i16 [[B:%.*]] to <16 x i1>
-; CHECK-NEXT:    [[OR:%.*]] = xor <16 x i1> [[BC1]], [[BC2]]
-; CHECK-NEXT:    ret <16 x i1> [[OR]]
+; CHECK-NEXT:    [[B:%.*]] = xor i16 [[A:%.*]], [[B1:%.*]]
+; CHECK-NEXT:    [[BC2:%.*]] = bitcast i16 [[B]] to <16 x i1>
+; CHECK-NEXT:    ret <16 x i1> [[BC2]]
 ;
   %bc1 = bitcast i16 %a to <16 x i1>
   %bc2 = bitcast i16 %b to <16 x i1>