[SLPVectorizer] Add initial alternate opcode support for cast instructions.

[RKSimon]

Summary

We currently only support binary instructions in the alternate opcode shuffles.

This patch is an initial attempt at adding cast instructions as well, this raises several issues that we probably want to address as we continue to generalize the alternate mechanism:

1 - Duplication of cost determination - we should probably add scalar/vector costs helper functions and get BoUpSLP::getEntryCost to use them instead of determining costs directly.
2 - Support alternate instructions with the same opcode (e.g. casts with different src types) - alternate vectorization of calls with different IntrinsicIDs will require this.
3 - Allow alternates to be a different instruction type - mixing binary/cast/call etc.
4 - Allow passthrough of unsupported alternate instructions - related to PR30787/D28907 'copyable' elements.

[dtemirbulatov]

looks good for me.

[mstorsjo]

This change broke compilation and started triggering asserts for me. See http://martin.st/temp/hb-ot-font-f498e6.cpp and .sh for repro. Will file a proper PR later when I have more time.

[mstorsjo]

Filed a proper bug report as PR38154.

[mstorsjo]

Reverted in SVN r336949.

llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp

 /// InstructionsState, the Opcode that we suppose the whole list 
 /// could be vectorized even if its structure is diverse.
 static InstructionsState getSameOpcode(ArrayRef<Value *> VL,
                                        unsigned BaseIndex = 0) {
   // Make sure these are all Instructions.
   if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); }))
     return InstructionsState(VL[BaseIndex], nullptr, nullptr);
 
+  bool IsCastOp = isa<CastInst>(VL[BaseIndex]);
   bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]);
   unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode();
   unsigned AltOpcode = Opcode;
   unsigned AltIndex = BaseIndex;
 
   // Check for one alternate opcode from another BinaryOperator.
-  // TODO - can we support other operators (casts etc.)?
+  // TODO - generalize to support all operators (calls etc.).
   for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) {
     unsigned InstOpcode = cast<Instruction>(VL[Cnt])->getOpcode();
     if (InstOpcode != Opcode && InstOpcode != AltOpcode) {
+      if (Opcode == AltOpcode && IsCastOp && isa<CastInst>(VL[Cnt])) {
+        AltOpcode = InstOpcode;
+        AltIndex = Cnt;
+        continue;
+      }
       if (Opcode == AltOpcode && IsBinOp && isa<BinaryOperator>(VL[Cnt])) {
         AltOpcode = InstOpcode;
         AltIndex = Cnt;
         continue;
       }
       return InstructionsState(VL[BaseIndex], nullptr, nullptr);
     }
   }
case Instruction::Call: {
 
       LLVM_DEBUG(dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost
                         << " (" << VecCallCost << "-" << ScalarCallCost << ")"
                         << " for " << *CI << "\n");
 
       return ReuseShuffleCost + VecCallCost - ScalarCallCost;
     }
     case Instruction::ShuffleVector: {
-      assert(S.isAltShuffle() && Instruction::isBinaryOp(S.getOpcode()) &&
-             Instruction::isBinaryOp(S.getAltOpcode()) &&
+      assert(S.isAltShuffle() &&
+             ((Instruction::isBinaryOp(S.getOpcode()) &&
+               Instruction::isBinaryOp(S.getAltOpcode())) ||
+              (Instruction::isCast(S.getOpcode()) &&
+               Instruction::isCast(S.getAltOpcode()))) &&
              "Invalid Shuffle Vector Operand");
       int ScalarCost = 0;
       if (NeedToShuffleReuses) {
         for (unsigned Idx : E->ReuseShuffleIndices) {
           Instruction *I = cast<Instruction>(VL[Idx]);
-          ReuseShuffleCost -=
-              TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy);
+          ReuseShuffleCost -= TTI->getInstructionCost(
+              I, TargetTransformInfo::TCK_RecipThroughput);
         }
         for (Value *V : VL) {
           Instruction *I = cast<Instruction>(V);
-          ReuseShuffleCost +=
-              TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy);
+          ReuseShuffleCost += TTI->getInstructionCost(
+              I, TargetTransformInfo::TCK_RecipThroughput);
         }
       }
       int VecCost = 0;
       for (Value *i : VL) {
         Instruction *I = cast<Instruction>(i);
         assert(S.isOpcodeOrAlt(I) && "Unexpected main/alternate opcode");
-        ScalarCost += TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy);
+        ScalarCost += TTI->getInstructionCost(
+            I, TargetTransformInfo::TCK_RecipThroughput);
       }
       // VecCost is equal to sum of the cost of creating 2 vectors
       // and the cost of creating shuffle.
+      if (Instruction::isBinaryOp(S.getOpcode())) {
         VecCost = TTI->getArithmeticInstrCost(S.getOpcode(), VecTy);
         VecCost += TTI->getArithmeticInstrCost(S.getAltOpcode(), VecTy);
+      } else {
+        Type *Src0SclTy = S.MainOp->getOperand(0)->getType();
+        Type *Src1SclTy = S.AltOp->getOperand(0)->getType();
+        VectorType *Src0Ty = VectorType::get(Src0SclTy, VL.size());
+        VectorType *Src1Ty = VectorType::get(Src1SclTy, VL.size());
+        VecCost = TTI->getCastInstrCost(S.getOpcode(), VecTy, Src0Ty);
+        VecCost += TTI->getCastInstrCost(S.getAltOpcode(), VecTy, Src1Ty);
+      }
       VecCost += TTI->getShuffleCost(TargetTransformInfo::SK_Select, VecTy, 0);
       return ReuseShuffleCost + VecCost - ScalarCost;
     }
     default:
       llvm_unreachable("Unknown instruction");
   }
 }
 
case Instruction::Call: {
                                         E->ReuseShuffleIndices, "shuffle");
       }
       E->VectorizedValue = V;
       ++NumVectorInstructions;
       return V;
     }
     case Instruction::ShuffleVector: {
       ValueList LHSVL, RHSVL;
-      assert(S.isAltShuffle() && Instruction::isBinaryOp(S.getOpcode()) &&
-             Instruction::isBinaryOp(S.getAltOpcode()) &&
+      assert(S.isAltShuffle() &&
+             ((Instruction::isBinaryOp(S.getOpcode()) &&
+               Instruction::isBinaryOp(S.getAltOpcode())) ||
+              (Instruction::isCast(S.getOpcode()) &&
+               Instruction::isCast(S.getAltOpcode()))) &&
              "Invalid Shuffle Vector Operand");
+
+      Value *LHS, *RHS;
+      if (Instruction::isBinaryOp(S.getOpcode())) {
         reorderAltShuffleOperands(S, E->Scalars, LHSVL, RHSVL);
         setInsertPointAfterBundle(E->Scalars, S);
-
-      Value *LHS = vectorizeTree(LHSVL);
-      Value *RHS = vectorizeTree(RHSVL);
+        LHS = vectorizeTree(LHSVL);
+        RHS = vectorizeTree(RHSVL);
+      } else {
+        ValueList INVL;
+        for (Value *V : E->Scalars)
+          INVL.push_back(cast<Instruction>(V)->getOperand(0));
+        setInsertPointAfterBundle(E->Scalars, S);
+        LHS = vectorizeTree(INVL);
+      }
 
       if (E->VectorizedValue) {
         LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n");
         return E->VectorizedValue;
       }
 
-      // Create a vector of LHS op1 RHS
-      Value *V0 = Builder.CreateBinOp(
+      Value *V0, *V1;
+      if (Instruction::isBinaryOp(S.getOpcode())) {
+        V0 = Builder.CreateBinOp(
           static_cast<Instruction::BinaryOps>(S.getOpcode()), LHS, RHS);
-
-      // Create a vector of LHS op2 RHS
-      Value *V1 = Builder.CreateBinOp(
+        V1 = Builder.CreateBinOp(
           static_cast<Instruction::BinaryOps>(S.getAltOpcode()), LHS, RHS);
+      } else {
+        V0 = Builder.CreateCast(
+            static_cast<Instruction::CastOps>(S.getOpcode()), LHS, VecTy);
+        V1 = Builder.CreateCast(
+            static_cast<Instruction::CastOps>(S.getAltOpcode()), LHS, VecTy);
+      }
 
       // Create shuffle to take alternate operations from the vector.
-      // Also, gather up odd and even scalar ops to propagate IR flags to
+      // Also, gather up main and alt scalar ops to propagate IR flags to
       // each vector operation.
       ValueList OpScalars, AltScalars;
       unsigned e = E->Scalars.size();
       SmallVector<Constant *, 8> Mask(e);
       for (unsigned i = 0; i < e; ++i) {
         auto *OpInst = cast<Instruction>(E->Scalars[i]);
         assert(S.isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode");
         if (OpInst->getOpcode() == S.getAltOpcode()) {

llvm/trunk/test/Transforms/SLPVectorizer/X86/alternate-cast.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -mtriple=x86_64-unknown -basicaa -slp-vectorizer -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=SSE
 ; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=slm -basicaa -slp-vectorizer -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=SLM
 ; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=corei7-avx -basicaa -slp-vectorizer -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=AVX1
 ; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=core-avx2 -basicaa -slp-vectorizer -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=AVX2
 ; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=knl -basicaa -slp-vectorizer -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512 --check-prefix=AVX512F
 ; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=skx -basicaa -slp-vectorizer -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512 --check-prefix=AVX512BW
 
 define <8 x float> @sitofp_uitofp(<8 x i32> %a) {
-; CHECK-LABEL: @sitofp_uitofp(
-; CHECK-NEXT:    [[A0:%.*]] = extractelement <8 x i32> [[A:%.*]], i32 0
-; CHECK-NEXT:    [[A1:%.*]] = extractelement <8 x i32> [[A]], i32 1
-; CHECK-NEXT:    [[A2:%.*]] = extractelement <8 x i32> [[A]], i32 2
-; CHECK-NEXT:    [[A3:%.*]] = extractelement <8 x i32> [[A]], i32 3
-; CHECK-NEXT:    [[A4:%.*]] = extractelement <8 x i32> [[A]], i32 4
-; CHECK-NEXT:    [[A5:%.*]] = extractelement <8 x i32> [[A]], i32 5
-; CHECK-NEXT:    [[A6:%.*]] = extractelement <8 x i32> [[A]], i32 6
-; CHECK-NEXT:    [[A7:%.*]] = extractelement <8 x i32> [[A]], i32 7
-; CHECK-NEXT:    [[AB0:%.*]] = sitofp i32 [[A0]] to float
-; CHECK-NEXT:    [[AB1:%.*]] = sitofp i32 [[A1]] to float
-; CHECK-NEXT:    [[AB2:%.*]] = sitofp i32 [[A2]] to float
-; CHECK-NEXT:    [[AB3:%.*]] = sitofp i32 [[A3]] to float
-; CHECK-NEXT:    [[AB4:%.*]] = uitofp i32 [[A4]] to float
-; CHECK-NEXT:    [[AB5:%.*]] = uitofp i32 [[A5]] to float
-; CHECK-NEXT:    [[AB6:%.*]] = uitofp i32 [[A6]] to float
-; CHECK-NEXT:    [[AB7:%.*]] = uitofp i32 [[A7]] to float
-; CHECK-NEXT:    [[R0:%.*]] = insertelement <8 x float> undef, float [[AB0]], i32 0
-; CHECK-NEXT:    [[R1:%.*]] = insertelement <8 x float> [[R0]], float [[AB1]], i32 1
-; CHECK-NEXT:    [[R2:%.*]] = insertelement <8 x float> [[R1]], float [[AB2]], i32 2
-; CHECK-NEXT:    [[R3:%.*]] = insertelement <8 x float> [[R2]], float [[AB3]], i32 3
-; CHECK-NEXT:    [[R4:%.*]] = insertelement <8 x float> [[R3]], float [[AB4]], i32 4
-; CHECK-NEXT:    [[R5:%.*]] = insertelement <8 x float> [[R4]], float [[AB5]], i32 5
-; CHECK-NEXT:    [[R6:%.*]] = insertelement <8 x float> [[R5]], float [[AB6]], i32 6
-; CHECK-NEXT:    [[R7:%.*]] = insertelement <8 x float> [[R6]], float [[AB7]], i32 7
-; CHECK-NEXT:    ret <8 x float> [[R7]]
+; SSE-LABEL: @sitofp_uitofp(
+; SSE-NEXT:    [[A0:%.*]] = extractelement <8 x i32> [[A:%.*]], i32 0
+; SSE-NEXT:    [[A1:%.*]] = extractelement <8 x i32> [[A]], i32 1
+; SSE-NEXT:    [[A2:%.*]] = extractelement <8 x i32> [[A]], i32 2
+; SSE-NEXT:    [[A3:%.*]] = extractelement <8 x i32> [[A]], i32 3
+; SSE-NEXT:    [[A4:%.*]] = extractelement <8 x i32> [[A]], i32 4
+; SSE-NEXT:    [[A5:%.*]] = extractelement <8 x i32> [[A]], i32 5
+; SSE-NEXT:    [[A6:%.*]] = extractelement <8 x i32> [[A]], i32 6
+; SSE-NEXT:    [[A7:%.*]] = extractelement <8 x i32> [[A]], i32 7
+; SSE-NEXT:    [[AB0:%.*]] = sitofp i32 [[A0]] to float
+; SSE-NEXT:    [[AB1:%.*]] = sitofp i32 [[A1]] to float
+; SSE-NEXT:    [[AB2:%.*]] = sitofp i32 [[A2]] to float
+; SSE-NEXT:    [[AB3:%.*]] = sitofp i32 [[A3]] to float
+; SSE-NEXT:    [[AB4:%.*]] = uitofp i32 [[A4]] to float
+; SSE-NEXT:    [[AB5:%.*]] = uitofp i32 [[A5]] to float
+; SSE-NEXT:    [[AB6:%.*]] = uitofp i32 [[A6]] to float
+; SSE-NEXT:    [[AB7:%.*]] = uitofp i32 [[A7]] to float
+; SSE-NEXT:    [[R0:%.*]] = insertelement <8 x float> undef, float [[AB0]], i32 0
+; SSE-NEXT:    [[R1:%.*]] = insertelement <8 x float> [[R0]], float [[AB1]], i32 1
+; SSE-NEXT:    [[R2:%.*]] = insertelement <8 x float> [[R1]], float [[AB2]], i32 2
+; SSE-NEXT:    [[R3:%.*]] = insertelement <8 x float> [[R2]], float [[AB3]], i32 3
+; SSE-NEXT:    [[R4:%.*]] = insertelement <8 x float> [[R3]], float [[AB4]], i32 4
+; SSE-NEXT:    [[R5:%.*]] = insertelement <8 x float> [[R4]], float [[AB5]], i32 5
+; SSE-NEXT:    [[R6:%.*]] = insertelement <8 x float> [[R5]], float [[AB6]], i32 6
+; SSE-NEXT:    [[R7:%.*]] = insertelement <8 x float> [[R6]], float [[AB7]], i32 7
+; SSE-NEXT:    ret <8 x float> [[R7]]
+;
+; SLM-LABEL: @sitofp_uitofp(
+; SLM-NEXT:    [[A0:%.*]] = extractelement <8 x i32> [[A:%.*]], i32 0
+; SLM-NEXT:    [[A1:%.*]] = extractelement <8 x i32> [[A]], i32 1
+; SLM-NEXT:    [[A2:%.*]] = extractelement <8 x i32> [[A]], i32 2
+; SLM-NEXT:    [[A3:%.*]] = extractelement <8 x i32> [[A]], i32 3
+; SLM-NEXT:    [[A4:%.*]] = extractelement <8 x i32> [[A]], i32 4
+; SLM-NEXT:    [[A5:%.*]] = extractelement <8 x i32> [[A]], i32 5
+; SLM-NEXT:    [[A6:%.*]] = extractelement <8 x i32> [[A]], i32 6
+; SLM-NEXT:    [[A7:%.*]] = extractelement <8 x i32> [[A]], i32 7
+; SLM-NEXT:    [[AB0:%.*]] = sitofp i32 [[A0]] to float
+; SLM-NEXT:    [[AB1:%.*]] = sitofp i32 [[A1]] to float
+; SLM-NEXT:    [[AB2:%.*]] = sitofp i32 [[A2]] to float
+; SLM-NEXT:    [[AB3:%.*]] = sitofp i32 [[A3]] to float
+; SLM-NEXT:    [[AB4:%.*]] = uitofp i32 [[A4]] to float
+; SLM-NEXT:    [[AB5:%.*]] = uitofp i32 [[A5]] to float
+; SLM-NEXT:    [[AB6:%.*]] = uitofp i32 [[A6]] to float
+; SLM-NEXT:    [[AB7:%.*]] = uitofp i32 [[A7]] to float
+; SLM-NEXT:    [[R0:%.*]] = insertelement <8 x float> undef, float [[AB0]], i32 0
+; SLM-NEXT:    [[R1:%.*]] = insertelement <8 x float> [[R0]], float [[AB1]], i32 1
+; SLM-NEXT:    [[R2:%.*]] = insertelement <8 x float> [[R1]], float [[AB2]], i32 2
+; SLM-NEXT:    [[R3:%.*]] = insertelement <8 x float> [[R2]], float [[AB3]], i32 3
+; SLM-NEXT:    [[R4:%.*]] = insertelement <8 x float> [[R3]], float [[AB4]], i32 4
+; SLM-NEXT:    [[R5:%.*]] = insertelement <8 x float> [[R4]], float [[AB5]], i32 5
+; SLM-NEXT:    [[R6:%.*]] = insertelement <8 x float> [[R5]], float [[AB6]], i32 6
+; SLM-NEXT:    [[R7:%.*]] = insertelement <8 x float> [[R6]], float [[AB7]], i32 7
+; SLM-NEXT:    ret <8 x float> [[R7]]
+;
+; AVX-LABEL: @sitofp_uitofp(
+; AVX-NEXT:    [[TMP1:%.*]] = sitofp <8 x i32> [[A:%.*]] to <8 x float>
+; AVX-NEXT:    [[TMP2:%.*]] = uitofp <8 x i32> [[A]] to <8 x float>
+; AVX-NEXT:    [[R7:%.*]] = shufflevector <8 x float> [[TMP1]], <8 x float> [[TMP2]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 13, i32 14, i32 15>
+; AVX-NEXT:    ret <8 x float> [[R7]]
+;
+; AVX512-LABEL: @sitofp_uitofp(
+; AVX512-NEXT:    [[TMP1:%.*]] = sitofp <8 x i32> [[A:%.*]] to <8 x float>
+; AVX512-NEXT:    [[TMP2:%.*]] = uitofp <8 x i32> [[A]] to <8 x float>
+; AVX512-NEXT:    [[R7:%.*]] = shufflevector <8 x float> [[TMP1]], <8 x float> [[TMP2]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 13, i32 14, i32 15>
+; AVX512-NEXT:    ret <8 x float> [[R7]]
 ;
   %a0 = extractelement <8 x i32> %a, i32 0
   %a1 = extractelement <8 x i32> %a, i32 1
   %a2 = extractelement <8 x i32> %a, i32 2
   %a3 = extractelement <8 x i32> %a, i32 3
   %a4 = extractelement <8 x i32> %a, i32 4
   %a5 = extractelement <8 x i32> %a, i32 5
   %a6 = extractelement <8 x i32> %a, i32 6
;
   %r4 = insertelement <8 x float>   %r3, float %ab4, i32 4
   %r5 = insertelement <8 x float>   %r4, float %ab5, i32 5
   %r6 = insertelement <8 x float>   %r5, float %ab6, i32 6
   %r7 = insertelement <8 x float>   %r6, float %ab7, i32 7
   ret <8 x float> %r7
 }
 
 define <8 x i32> @fptosi_fptoui(<8 x float> %a) {
-; CHECK-LABEL: @fptosi_fptoui(
-; CHECK-NEXT:    [[A0:%.*]] = extractelement <8 x float> [[A:%.*]], i32 0
-; CHECK-NEXT:    [[A1:%.*]] = extractelement <8 x float> [[A]], i32 1
-; CHECK-NEXT:    [[A2:%.*]] = extractelement <8 x float> [[A]], i32 2
-; CHECK-NEXT:    [[A3:%.*]] = extractelement <8 x float> [[A]], i32 3
-; CHECK-NEXT:    [[A4:%.*]] = extractelement <8 x float> [[A]], i32 4
-; CHECK-NEXT:    [[A5:%.*]] = extractelement <8 x float> [[A]], i32 5
-; CHECK-NEXT:    [[A6:%.*]] = extractelement <8 x float> [[A]], i32 6
-; CHECK-NEXT:    [[A7:%.*]] = extractelement <8 x float> [[A]], i32 7
-; CHECK-NEXT:    [[AB0:%.*]] = fptosi float [[A0]] to i32
-; CHECK-NEXT:    [[AB1:%.*]] = fptosi float [[A1]] to i32
-; CHECK-NEXT:    [[AB2:%.*]] = fptosi float [[A2]] to i32
-; CHECK-NEXT:    [[AB3:%.*]] = fptosi float [[A3]] to i32
-; CHECK-NEXT:    [[AB4:%.*]] = fptoui float [[A4]] to i32
-; CHECK-NEXT:    [[AB5:%.*]] = fptoui float [[A5]] to i32
-; CHECK-NEXT:    [[AB6:%.*]] = fptoui float [[A6]] to i32
-; CHECK-NEXT:    [[AB7:%.*]] = fptoui float [[A7]] to i32
-; CHECK-NEXT:    [[R0:%.*]] = insertelement <8 x i32> undef, i32 [[AB0]], i32 0
-; CHECK-NEXT:    [[R1:%.*]] = insertelement <8 x i32> [[R0]], i32 [[AB1]], i32 1
-; CHECK-NEXT:    [[R2:%.*]] = insertelement <8 x i32> [[R1]], i32 [[AB2]], i32 2
-; CHECK-NEXT:    [[R3:%.*]] = insertelement <8 x i32> [[R2]], i32 [[AB3]], i32 3
-; CHECK-NEXT:    [[R4:%.*]] = insertelement <8 x i32> [[R3]], i32 [[AB4]], i32 4
-; CHECK-NEXT:    [[R5:%.*]] = insertelement <8 x i32> [[R4]], i32 [[AB5]], i32 5
-; CHECK-NEXT:    [[R6:%.*]] = insertelement <8 x i32> [[R5]], i32 [[AB6]], i32 6
-; CHECK-NEXT:    [[R7:%.*]] = insertelement <8 x i32> [[R6]], i32 [[AB7]], i32 7
-; CHECK-NEXT:    ret <8 x i32> [[R7]]
+; SSE-LABEL: @fptosi_fptoui(
+; SSE-NEXT:    [[A0:%.*]] = extractelement <8 x float> [[A:%.*]], i32 0
+; SSE-NEXT:    [[A1:%.*]] = extractelement <8 x float> [[A]], i32 1
+; SSE-NEXT:    [[A2:%.*]] = extractelement <8 x float> [[A]], i32 2
+; SSE-NEXT:    [[A3:%.*]] = extractelement <8 x float> [[A]], i32 3
+; SSE-NEXT:    [[A4:%.*]] = extractelement <8 x float> [[A]], i32 4
+; SSE-NEXT:    [[A5:%.*]] = extractelement <8 x float> [[A]], i32 5
+; SSE-NEXT:    [[A6:%.*]] = extractelement <8 x float> [[A]], i32 6
+; SSE-NEXT:    [[A7:%.*]] = extractelement <8 x float> [[A]], i32 7
+; SSE-NEXT:    [[AB0:%.*]] = fptosi float [[A0]] to i32
+; SSE-NEXT:    [[AB1:%.*]] = fptosi float [[A1]] to i32
+; SSE-NEXT:    [[AB2:%.*]] = fptosi float [[A2]] to i32
+; SSE-NEXT:    [[AB3:%.*]] = fptosi float [[A3]] to i32
+; SSE-NEXT:    [[AB4:%.*]] = fptoui float [[A4]] to i32
+; SSE-NEXT:    [[AB5:%.*]] = fptoui float [[A5]] to i32
+; SSE-NEXT:    [[AB6:%.*]] = fptoui float [[A6]] to i32
+; SSE-NEXT:    [[AB7:%.*]] = fptoui float [[A7]] to i32
+; SSE-NEXT:    [[R0:%.*]] = insertelement <8 x i32> undef, i32 [[AB0]], i32 0
+; SSE-NEXT:    [[R1:%.*]] = insertelement <8 x i32> [[R0]], i32 [[AB1]], i32 1
+; SSE-NEXT:    [[R2:%.*]] = insertelement <8 x i32> [[R1]], i32 [[AB2]], i32 2
+; SSE-NEXT:    [[R3:%.*]] = insertelement <8 x i32> [[R2]], i32 [[AB3]], i32 3
+; SSE-NEXT:    [[R4:%.*]] = insertelement <8 x i32> [[R3]], i32 [[AB4]], i32 4
+; SSE-NEXT:    [[R5:%.*]] = insertelement <8 x i32> [[R4]], i32 [[AB5]], i32 5
+; SSE-NEXT:    [[R6:%.*]] = insertelement <8 x i32> [[R5]], i32 [[AB6]], i32 6
+; SSE-NEXT:    [[R7:%.*]] = insertelement <8 x i32> [[R6]], i32 [[AB7]], i32 7
+; SSE-NEXT:    ret <8 x i32> [[R7]]
+;
+; SLM-LABEL: @fptosi_fptoui(
+; SLM-NEXT:    [[A0:%.*]] = extractelement <8 x float> [[A:%.*]], i32 0
+; SLM-NEXT:    [[A1:%.*]] = extractelement <8 x float> [[A]], i32 1
+; SLM-NEXT:    [[A2:%.*]] = extractelement <8 x float> [[A]], i32 2
+; SLM-NEXT:    [[A3:%.*]] = extractelement <8 x float> [[A]], i32 3
+; SLM-NEXT:    [[A4:%.*]] = extractelement <8 x float> [[A]], i32 4
+; SLM-NEXT:    [[A5:%.*]] = extractelement <8 x float> [[A]], i32 5
+; SLM-NEXT:    [[A6:%.*]] = extractelement <8 x float> [[A]], i32 6
+; SLM-NEXT:    [[A7:%.*]] = extractelement <8 x float> [[A]], i32 7
+; SLM-NEXT:    [[AB0:%.*]] = fptosi float [[A0]] to i32
+; SLM-NEXT:    [[AB1:%.*]] = fptosi float [[A1]] to i32
+; SLM-NEXT:    [[AB2:%.*]] = fptosi float [[A2]] to i32
+; SLM-NEXT:    [[AB3:%.*]] = fptosi float [[A3]] to i32
+; SLM-NEXT:    [[AB4:%.*]] = fptoui float [[A4]] to i32
+; SLM-NEXT:    [[AB5:%.*]] = fptoui float [[A5]] to i32
+; SLM-NEXT:    [[AB6:%.*]] = fptoui float [[A6]] to i32
+; SLM-NEXT:    [[AB7:%.*]] = fptoui float [[A7]] to i32
+; SLM-NEXT:    [[R0:%.*]] = insertelement <8 x i32> undef, i32 [[AB0]], i32 0
+; SLM-NEXT:    [[R1:%.*]] = insertelement <8 x i32> [[R0]], i32 [[AB1]], i32 1
+; SLM-NEXT:    [[R2:%.*]] = insertelement <8 x i32> [[R1]], i32 [[AB2]], i32 2
+; SLM-NEXT:    [[R3:%.*]] = insertelement <8 x i32> [[R2]], i32 [[AB3]], i32 3
+; SLM-NEXT:    [[R4:%.*]] = insertelement <8 x i32> [[R3]], i32 [[AB4]], i32 4
+; SLM-NEXT:    [[R5:%.*]] = insertelement <8 x i32> [[R4]], i32 [[AB5]], i32 5
+; SLM-NEXT:    [[R6:%.*]] = insertelement <8 x i32> [[R5]], i32 [[AB6]], i32 6
+; SLM-NEXT:    [[R7:%.*]] = insertelement <8 x i32> [[R6]], i32 [[AB7]], i32 7
+; SLM-NEXT:    ret <8 x i32> [[R7]]
+;
+; AVX-LABEL: @fptosi_fptoui(
+; AVX-NEXT:    [[A0:%.*]] = extractelement <8 x float> [[A:%.*]], i32 0
+; AVX-NEXT:    [[A1:%.*]] = extractelement <8 x float> [[A]], i32 1
+; AVX-NEXT:    [[A2:%.*]] = extractelement <8 x float> [[A]], i32 2
+; AVX-NEXT:    [[A3:%.*]] = extractelement <8 x float> [[A]], i32 3
+; AVX-NEXT:    [[A4:%.*]] = extractelement <8 x float> [[A]], i32 4
+; AVX-NEXT:    [[A5:%.*]] = extractelement <8 x float> [[A]], i32 5
+; AVX-NEXT:    [[A6:%.*]] = extractelement <8 x float> [[A]], i32 6
+; AVX-NEXT:    [[A7:%.*]] = extractelement <8 x float> [[A]], i32 7
+; AVX-NEXT:    [[AB0:%.*]] = fptosi float [[A0]] to i32
+; AVX-NEXT:    [[AB1:%.*]] = fptosi float [[A1]] to i32
+; AVX-NEXT:    [[AB2:%.*]] = fptosi float [[A2]] to i32
+; AVX-NEXT:    [[AB3:%.*]] = fptosi float [[A3]] to i32
+; AVX-NEXT:    [[AB4:%.*]] = fptoui float [[A4]] to i32
+; AVX-NEXT:    [[AB5:%.*]] = fptoui float [[A5]] to i32
+; AVX-NEXT:    [[AB6:%.*]] = fptoui float [[A6]] to i32
+; AVX-NEXT:    [[AB7:%.*]] = fptoui float [[A7]] to i32
+; AVX-NEXT:    [[R0:%.*]] = insertelement <8 x i32> undef, i32 [[AB0]], i32 0
+; AVX-NEXT:    [[R1:%.*]] = insertelement <8 x i32> [[R0]], i32 [[AB1]], i32 1
+; AVX-NEXT:    [[R2:%.*]] = insertelement <8 x i32> [[R1]], i32 [[AB2]], i32 2
+; AVX-NEXT:    [[R3:%.*]] = insertelement <8 x i32> [[R2]], i32 [[AB3]], i32 3
+; AVX-NEXT:    [[R4:%.*]] = insertelement <8 x i32> [[R3]], i32 [[AB4]], i32 4
+; AVX-NEXT:    [[R5:%.*]] = insertelement <8 x i32> [[R4]], i32 [[AB5]], i32 5
+; AVX-NEXT:    [[R6:%.*]] = insertelement <8 x i32> [[R5]], i32 [[AB6]], i32 6
+; AVX-NEXT:    [[R7:%.*]] = insertelement <8 x i32> [[R6]], i32 [[AB7]], i32 7
+; AVX-NEXT:    ret <8 x i32> [[R7]]
+;
+; AVX512-LABEL: @fptosi_fptoui(
+; AVX512-NEXT:    [[TMP1:%.*]] = fptosi <8 x float> [[A:%.*]] to <8 x i32>
+; AVX512-NEXT:    [[TMP2:%.*]] = fptoui <8 x float> [[A]] to <8 x i32>
+; AVX512-NEXT:    [[R7:%.*]] = shufflevector <8 x i32> [[TMP1]], <8 x i32> [[TMP2]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 13, i32 14, i32 15>
+; AVX512-NEXT:    ret <8 x i32> [[R7]]
 ;
   %a0 = extractelement <8 x float> %a, i32 0
   %a1 = extractelement <8 x float> %a, i32 1
   %a2 = extractelement <8 x float> %a, i32 2
   %a3 = extractelement <8 x float> %a, i32 3
   %a4 = extractelement <8 x float> %a, i32 4
   %a5 = extractelement <8 x float> %a, i32 5
   %a6 = extractelement <8 x float> %a, i32 6