[VectorCombine] foldShuffleOfCastops - don't restrict to oneuse but compare total costs instead

Some casts (especially bitcasts but others as well) are incredibly cheap (or free), so don't limit the shuffle(cast(x),cast(y)) -> cast(shuffle(x,y)) to oneuse cases, but instead compare the total before/after costs of possibly repeating some casts.

Author: RKSimon

llvm/lib/Transforms/Vectorize/VectorCombine.cpp

@@ -1491,8 +1491,7 @@ bool VectorCombine::foldShuffleOfBinops(Instruction &I) {
 bool VectorCombine::foldShuffleOfCastops(Instruction &I) {
   Value *V0, *V1;
   ArrayRef<int> OldMask;
-  if (!match(&I, m_Shuffle(m_OneUse(m_Value(V0)), m_OneUse(m_Value(V1)),
-                           m_Mask(OldMask))))
+  if (!match(&I, m_Shuffle(m_Value(V0), m_Value(V1), m_Mask(OldMask))))
     return false;
 
   auto *C0 = dyn_cast<CastInst>(V0);
@@ -1551,11 +1550,13 @@ bool VectorCombine::foldShuffleOfCastops(Instruction &I) {
   // Try to replace a castop with a shuffle if the shuffle is not costly.
   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
 
-  InstructionCost OldCost =
+  InstructionCost CostC0 =
       TTI.getCastInstrCost(C0->getOpcode(), CastDstTy, CastSrcTy,
-                           TTI::CastContextHint::None, CostKind) +
+                           TTI::CastContextHint::None, CostKind);
+  InstructionCost CostC1 =
       TTI.getCastInstrCost(C1->getOpcode(), CastDstTy, CastSrcTy,
                            TTI::CastContextHint::None, CostKind);
+  InstructionCost OldCost = CostC0 + CostC1;
   OldCost +=
       TTI.getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, CastDstTy,
                          OldMask, CostKind, 0, nullptr, std::nullopt, &I);
@@ -1564,6 +1565,10 @@ bool VectorCombine::foldShuffleOfCastops(Instruction &I) {
       TargetTransformInfo::SK_PermuteTwoSrc, CastSrcTy, NewMask, CostKind);
   NewCost += TTI.getCastInstrCost(Opcode, ShuffleDstTy, NewShuffleDstTy,
                                   TTI::CastContextHint::None, CostKind);
+  if (!C0->hasOneUse())
+    NewCost += CostC0;
+  if (!C1->hasOneUse())
+    NewCost += CostC1;
 
   LLVM_DEBUG(dbgs() << "Found a shuffle feeding two casts: " << I
                     << "\n  OldCost: " << OldCost << " vs NewCost: " << NewCost

llvm/test/Transforms/VectorCombine/X86/shuffle-of-casts.ll

@@ -224,14 +224,14 @@ define <16 x i16> @concat_bitcast_v4i32_v16i16(<4 x i32> %a0, <4 x i32> %a1) {
   ret <16 x i16> %r
 }
 
-; negative - multiuse
+; multiuse - ensure cost of any duplicated casts are worth it
 
 define <8 x i16> @concat_trunc_v4i32_v8i16_multiuse(<4 x i32> %a0, <4 x i32> %a1, ptr %a2) {
 ; CHECK-LABEL: define <8 x i16> @concat_trunc_v4i32_v8i16_multiuse(
 ; CHECK-SAME: <4 x i32> [[A0:%.*]], <4 x i32> [[A1:%.*]], ptr [[A2:%.*]]) #[[ATTR0]] {
 ; CHECK-NEXT:    [[X0:%.*]] = trunc <4 x i32> [[A0]] to <4 x i16>
-; CHECK-NEXT:    [[X1:%.*]] = trunc <4 x i32> [[A1]] to <4 x i16>
-; CHECK-NEXT:    [[R:%.*]] = shufflevector <4 x i16> [[X0]], <4 x i16> [[X1]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i32> [[A0]], <4 x i32> [[A1]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT:    [[R:%.*]] = trunc <8 x i32> [[TMP1]] to <8 x i16>
 ; CHECK-NEXT:    store <4 x i16> [[X0]], ptr [[A2]], align 8
 ; CHECK-NEXT:    ret <8 x i16> [[R]]
 ;
@@ -242,6 +242,24 @@ define <8 x i16> @concat_trunc_v4i32_v8i16_multiuse(<4 x i32> %a0, <4 x i32> %a1
   ret <8 x i16> %r
 }
 
+; negative - multiuse - ensure cost of any duplicated casts are worth it
+
+define <16 x i8> @concat_trunc_v8i64_v16i8_multiuse(<8 x i64> %a0, <8 x i64> %a1, ptr %a2) {
+; CHECK-LABEL: define <16 x i8> @concat_trunc_v8i64_v16i8_multiuse(
+; CHECK-SAME: <8 x i64> [[A0:%.*]], <8 x i64> [[A1:%.*]], ptr [[A2:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[X0:%.*]] = trunc <8 x i64> [[A0]] to <8 x i8>
+; CHECK-NEXT:    [[X1:%.*]] = trunc <8 x i64> [[A1]] to <8 x i8>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <8 x i8> [[X0]], <8 x i8> [[X1]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 4, i32 15>
+; CHECK-NEXT:    store <8 x i8> [[X0]], ptr [[A2]], align 8
+; CHECK-NEXT:    ret <16 x i8> [[R]]
+;
+  %x0 = trunc <8 x i64> %a0 to <8 x i8>
+  %x1 = trunc <8 x i64> %a1 to <8 x i8>
+  %r = shufflevector <8 x i8> %x0, <8 x i8> %x1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 4, i32 15>
+  store <8 x i8> %x0, ptr %a2
+  ret <16 x i8> %r
+}
+
 ; negative - bitcasts (unscalable higher element count)
 
 define <16 x i16> @revpair_bitcast_v4i32_v16i16(<4 x i32> %a0, <4 x i32> %a1) {