[CostModel][X86] getVectorInstrCost - correctly cost v4f32 insertelement into index 0

This is just the MOVSS instruction (SSE41 INSERTPS is still necessary for index != 0)

This exposed an issue in VectorCombine::foldInsExtFNeg - we need to use the more general SK_PermuteTwoSrc shuffle kind to allow getShuffleCost to match other shuffle kinds (not just SK_Select).

Author: RKSimon

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

@@ -4804,9 +4804,12 @@ InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
     MVT MScalarTy = LT.second.getScalarType();
     auto IsCheapPInsrPExtrInsertPS = [&]() {
       // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets.
+      // Inserting f32 into index0 is just movss.
       // Also, assume insertps is relatively cheap on all >= SSE41 targets.
       return (MScalarTy == MVT::i16 && ST->hasSSE2()) ||
              (MScalarTy.isInteger() && ST->hasSSE41()) ||
+             (MScalarTy == MVT::f32 && ST->hasSSE1() && Index == 0 &&
+              Opcode == Instruction::InsertElement) ||
              (MScalarTy == MVT::f32 && ST->hasSSE41() &&
               Opcode == Instruction::InsertElement);
     };

llvm/lib/Transforms/Vectorize/VectorCombine.cpp

@@ -705,7 +705,8 @@ bool VectorCombine::foldInsExtFNeg(Instruction &I) {
 
   InstructionCost NewCost =
       TTI.getArithmeticInstrCost(Instruction::FNeg, VecTy, CostKind) +
-      TTI.getShuffleCost(TargetTransformInfo::SK_Select, VecTy, Mask, CostKind);
+      TTI.getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, VecTy, Mask,
+                         CostKind);
 
   bool NeedLenChg = SrcVecTy->getNumElements() != NumElts;
   // If the lengths of the two vectors are not equal,

llvm/test/Analysis/CostModel/X86/vector-insert-value.ll

@@ -76,58 +76,58 @@ define i32 @insert_double(i32 %arg, double %val, <2 x double> %src128, <4 x doub
 define i32 @insert_float(i32 %arg, float %val, <2 x float> %src64, <4 x float> %src128, <8 x float> %src256, <16 x float> %src512) {
 ; SSE2-LABEL: 'insert_float'
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v2f32_a = insertelement <2 x float> %src64, float %val, i32 %arg
-; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2f32_0 = insertelement <2 x float> %src64, float %val, i32 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v2f32_0 = insertelement <2 x float> %src64, float %val, i32 0
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2f32_1 = insertelement <2 x float> %src64, float %val, i32 1
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v4f32_a = insertelement <4 x float> %src128, float %val, i32 %arg
-; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4f32_0 = insertelement <4 x float> %src128, float %val, i32 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v4f32_0 = insertelement <4 x float> %src128, float %val, i32 0
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4f32_3 = insertelement <4 x float> %src128, float %val, i32 3
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %v8f32_a = insertelement <8 x float> %src256, float %val, i32 %arg
-; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_0 = insertelement <8 x float> %src256, float %val, i32 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v8f32_0 = insertelement <8 x float> %src256, float %val, i32 0
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_3 = insertelement <8 x float> %src256, float %val, i32 3
-; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_4 = insertelement <8 x float> %src256, float %val, i32 4
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v8f32_4 = insertelement <8 x float> %src256, float %val, i32 4
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_7 = insertelement <8 x float> %src256, float %val, i32 7
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %v16f32_a = insertelement <16 x float> %src512, float %val, i32 %arg
-; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_0 = insertelement <16 x float> %src512, float %val, i32 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v16f32_0 = insertelement <16 x float> %src512, float %val, i32 0
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_3 = insertelement <16 x float> %src512, float %val, i32 3
-; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_8 = insertelement <16 x float> %src512, float %val, i32 8
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v16f32_8 = insertelement <16 x float> %src512, float %val, i32 8
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_15 = insertelement <16 x float> %src512, float %val, i32 15
 ; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; SSE3-LABEL: 'insert_float'
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v2f32_a = insertelement <2 x float> %src64, float %val, i32 %arg
-; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2f32_0 = insertelement <2 x float> %src64, float %val, i32 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v2f32_0 = insertelement <2 x float> %src64, float %val, i32 0
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2f32_1 = insertelement <2 x float> %src64, float %val, i32 1
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v4f32_a = insertelement <4 x float> %src128, float %val, i32 %arg
-; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4f32_0 = insertelement <4 x float> %src128, float %val, i32 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v4f32_0 = insertelement <4 x float> %src128, float %val, i32 0
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4f32_3 = insertelement <4 x float> %src128, float %val, i32 3
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %v8f32_a = insertelement <8 x float> %src256, float %val, i32 %arg
-; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_0 = insertelement <8 x float> %src256, float %val, i32 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v8f32_0 = insertelement <8 x float> %src256, float %val, i32 0
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_3 = insertelement <8 x float> %src256, float %val, i32 3
-; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_4 = insertelement <8 x float> %src256, float %val, i32 4
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v8f32_4 = insertelement <8 x float> %src256, float %val, i32 4
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_7 = insertelement <8 x float> %src256, float %val, i32 7
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %v16f32_a = insertelement <16 x float> %src512, float %val, i32 %arg
-; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_0 = insertelement <16 x float> %src512, float %val, i32 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v16f32_0 = insertelement <16 x float> %src512, float %val, i32 0
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_3 = insertelement <16 x float> %src512, float %val, i32 3
-; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_8 = insertelement <16 x float> %src512, float %val, i32 8
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v16f32_8 = insertelement <16 x float> %src512, float %val, i32 8
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_15 = insertelement <16 x float> %src512, float %val, i32 15
 ; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; SSSE3-LABEL: 'insert_float'
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v2f32_a = insertelement <2 x float> %src64, float %val, i32 %arg
-; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2f32_0 = insertelement <2 x float> %src64, float %val, i32 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v2f32_0 = insertelement <2 x float> %src64, float %val, i32 0
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2f32_1 = insertelement <2 x float> %src64, float %val, i32 1
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v4f32_a = insertelement <4 x float> %src128, float %val, i32 %arg
-; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4f32_0 = insertelement <4 x float> %src128, float %val, i32 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v4f32_0 = insertelement <4 x float> %src128, float %val, i32 0
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4f32_3 = insertelement <4 x float> %src128, float %val, i32 3
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %v8f32_a = insertelement <8 x float> %src256, float %val, i32 %arg
-; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_0 = insertelement <8 x float> %src256, float %val, i32 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v8f32_0 = insertelement <8 x float> %src256, float %val, i32 0
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_3 = insertelement <8 x float> %src256, float %val, i32 3
-; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_4 = insertelement <8 x float> %src256, float %val, i32 4
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v8f32_4 = insertelement <8 x float> %src256, float %val, i32 4
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v8f32_7 = insertelement <8 x float> %src256, float %val, i32 7
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %v16f32_a = insertelement <16 x float> %src512, float %val, i32 %arg
-; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_0 = insertelement <16 x float> %src512, float %val, i32 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v16f32_0 = insertelement <16 x float> %src512, float %val, i32 0
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_3 = insertelement <16 x float> %src512, float %val, i32 3
-; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_8 = insertelement <16 x float> %src512, float %val, i32 8
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v16f32_8 = insertelement <16 x float> %src512, float %val, i32 8
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16f32_15 = insertelement <16 x float> %src512, float %val, i32 15
 ; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;