[InstCombine] Added vector demanded bits support for SSE4A EXTRQ/INSERTQ instructions

[RKSimon]

Summary

The SSE4A instructions EXTRQ/INSERTQ only use the lower 64-bits (or less) for many of their input vector operands and all of them have undefined upper 64-bits results.

[spatel]

lib/Transforms/InstCombine/InstCombineCalls.cpp

934–935 ↗

(On Diff #34170)

Use 'auto *' for pointers? Same comment for the other autos below.

972–986 ↗

(On Diff #34170)

Isn't this identical to the extrqi case? If yes, combine cases and eliminate the duplicated code. Could also just add a 3-4 line helper function that can be used for all of the cases?

[RKSimon]

I'll get this updated - thanks for looking at this, I know SSE4A isn't the most exciting of instructions to deal with!!

lib/Transforms/InstCombine/InstCombineCalls.cpp

934–935 ↗

(On Diff #34170)

No problem.

972–986 ↗

(On Diff #34170)

I'll look into creating a helper (this is one of the most common patterns in this entire file....). I'm intending to add support for decoding EXTRQI to shuffles soon so wish to keep the actual cases separate.

[majnemer]

lib/Transforms/InstCombine/InstCombineCalls.cpp

936–937 ↗

(On Diff #34170)

I think it'd be more concise to use Op->getType()->getVectorNumElements()

[RKSimon]

Updated - thanks for the comments. If you chaps are happy with the SimplifyDemandedVectorEltsLow helper I can convert other cases to use it in future commits.

[spatel]

In D12680#247911, @RKSimon wrote:

Updated - thanks for the comments. If you chaps are happy with the SimplifyDemandedVectorEltsLow helper I can convert other cases to use it in future commits.

Can we hoist the getVectorNumElements() and associated assert into the helper as well? If yes, please do. And certainly now that we have a helper, let's use it to reduce the code bloat wherever possible.

LGTM.

[RKSimon]

In D12680#247968, @spatel wrote:

Can we hoist the getVectorNumElements() and associated assert into the helper as well? If yes, please do. And certainly now that we have a helper, let's use it to reduce the code bloat wherever possible.

Thanks Sanjay, I'll look at hoisting the getVectorNumElements() when I update the other functions to use the helper - probably over the weekend.

llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp

if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
     } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
       if (Instruction *I = SimplifyMemSet(MSI))
         return I;
     }
 
     if (Changed) return II;
   }
 
+  auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width, unsigned DemandedWidth)
+  {
+    APInt UndefElts(Width, 0);
+    APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
+    return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
+  };
+
   switch (II->getIntrinsicID()) {
   default: break;
   case Intrinsic::objectsize: {
     uint64_t Size;
     if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
       return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
     return nullptr;
   }
if (Value *V = SimplifyX86extend(*II, *Builder, false))
       return ReplaceInstUsesWith(*II, V);
     break;
 
   case Intrinsic::x86_sse41_insertps:
     if (Value *V = SimplifyX86insertps(*II, *Builder))
       return ReplaceInstUsesWith(*II, V);
     break;
 
+  case Intrinsic::x86_sse4a_extrq: {
+    // EXTRQ uses only the lowest 64-bits of the first 128-bit vector
+    // operands and the lowest 16-bits of the second.
+    Value *Op0 = II->getArgOperand(0);
+    Value *Op1 = II->getArgOperand(1);
+    unsigned VWidth0 = Op0->getType()->getVectorNumElements();
+    unsigned VWidth1 = Op1->getType()->getVectorNumElements();
+    assert(VWidth0 == 2 && VWidth1 == 16 && "Unexpected operand sizes");
+
+    if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
+      II->setArgOperand(0, V);
+      return II;
+    }
+    if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
+      II->setArgOperand(1, V);
+      return II;
+    }
+    break;
+  }
+
+  case Intrinsic::x86_sse4a_extrqi: {
+    // EXTRQI uses only the lowest 64-bits of the first 128-bit vector
+    // operand.
+    Value *Op = II->getArgOperand(0);
+    unsigned VWidth = Op->getType()->getVectorNumElements();
+    assert(VWidth == 2 && "Unexpected operand size");
+
+    if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
+      II->setArgOperand(0, V);
+      return II;
+    }
+    break;
+  }
+
+  case Intrinsic::x86_sse4a_insertq: {
+    // INSERTQ uses only the lowest 64-bits of the first 128-bit vector
+    // operand.
+    Value *Op = II->getArgOperand(0);
+    unsigned VWidth = Op->getType()->getVectorNumElements();
+    assert(VWidth == 2 && "Unexpected operand size");
+
+    if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
+      II->setArgOperand(0, V);
+      return II;
+    }
+    break;
+  }
+
   case Intrinsic::x86_sse4a_insertqi: {
     // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
     // ones undef
     // TODO: eventually we should lower this intrinsic to IR
     if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
       if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
         unsigned Index = CIIndex->getZExtValue();
         // From AMD documentation: "a value of zero in the field length is
if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
                     Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
                 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
               }
             }
           }
         }
       }
     }
+
+    // INSERTQI uses only the lowest 64-bits of the first two 128-bit vector
+    // operands.
+    Value *Op0 = II->getArgOperand(0);
+    Value *Op1 = II->getArgOperand(1);
+    unsigned VWidth0 = Op0->getType()->getVectorNumElements();
+    unsigned VWidth1 = Op1->getType()->getVectorNumElements();
+    assert(VWidth0 == 2 && VWidth1 == 2 && "Unexpected operand sizes");
+
+    if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
+      II->setArgOperand(0, V);
+      return II;
+    }
+
+    if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
+      II->setArgOperand(1, V);
+      return II;
+    }
     break;
   }
 
   case Intrinsic::x86_sse41_pblendvb:
   case Intrinsic::x86_sse41_blendvps:
   case Intrinsic::x86_sse41_blendvpd:
   case Intrinsic::x86_avx_blendv_ps_256:
   case Intrinsic::x86_avx_blendv_pd_256:

llvm/trunk/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

case Instruction::Xor: {
     break;
   }
   case Instruction::Select:
     // If this is a select as part of a min/max pattern, don't simplify any
     // further in case we break the structure.
     Value *LHS, *RHS;
     if (matchSelectPattern(I, LHS, RHS).Flavor != SPF_UNKNOWN)
       return nullptr;
-      
     if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask, RHSKnownZero,
                              RHSKnownOne, Depth + 1) ||
         SimplifyDemandedBits(I->getOperandUse(1), DemandedMask, LHSKnownZero,
                              LHSKnownOne, Depth + 1))
       return I;
     assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
     assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
 
case Intrinsic::x86_sse2_max_sd:
           return New;
         }
       }
 
       // Output elements are undefined if both are undefined.  Consider things
       // like undef&0.  The result is known zero, not undef.
       UndefElts &= UndefElts2;
       break;
+
+    // SSE4A instructions leave the upper 64-bits of the 128-bit result
+    // in an undefined state.
+    case Intrinsic::x86_sse4a_extrq:
+    case Intrinsic::x86_sse4a_extrqi:
+    case Intrinsic::x86_sse4a_insertq:
+    case Intrinsic::x86_sse4a_insertqi:
+      UndefElts |= APInt::getHighBitsSet(VWidth, VWidth / 2);
+      break;
     }
     break;
   }
   }
   return MadeChange ? I : nullptr;
 }

llvm/trunk/test/Transforms/InstCombine/x86-sse4a.ll

-; RUN: opt < %s -instcombine -S | FileCheck %s
-
-; We should optimize these two redundant insertqi into one
-; CHECK: define <2 x i64> @testInsertTwice(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertTwice(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 32)
-; CHECK-NOT: insertqi
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 32)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 32)
-  ret <2 x i64> %2
-}
-
-; The result of this insert is the second arg, since the top 64 bits of
-; the result are undefined, and we copy the bottom 64 bits from the
-; second arg
-; CHECK: define <2 x i64> @testInsert64Bits(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsert64Bits(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: ret <2 x i64> %i
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 64, i8 0)
-  ret <2 x i64> %1
-}
-
-; Test the several types of ranges and ordering that exist for two insertqi
-; CHECK: define <2 x i64> @testInsertContainedRange(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertContainedRange(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
-; CHECK: ret <2 x i64> %[[RES]]
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 16)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertContainedRange_2(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertContainedRange_2(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
-; CHECK: ret <2 x i64> %[[RES]]
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 16)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertOverlappingRange(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertOverlappingRange(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
-; CHECK: ret <2 x i64> %[[RES]]
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 16)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertOverlappingRange_2(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertOverlappingRange_2(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
-; CHECK: ret <2 x i64> %[[RES]]
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 16)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertAdjacentRange(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertAdjacentRange(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
-; CHECK: ret <2 x i64> %[[RES]]
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertAdjacentRange_2(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertAdjacentRange_2(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
-; CHECK: ret <2 x i64> %[[RES]]
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 32)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertDisjointRange(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertDisjointRange(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
-; CHECK: tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testInsertDisjointRange_2(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testInsertDisjointRange_2(<2 x i64> %v, <2 x i64> %i) {
-; CHECK:  tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
-; CHECK:  tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
-  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
-  ret <2 x i64> %2
-}
-
-; CHECK: define <2 x i64> @testZeroLength(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testZeroLength(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: ret <2 x i64> %i
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 0, i8 0)
-  ret <2 x i64> %1
-}
-
-; CHECK: define <2 x i64> @testUndefinedInsertq_1(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testUndefinedInsertq_1(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: ret <2 x i64> undef
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 0, i8 16)
-  ret <2 x i64> %1
-}
-
-; CHECK: define <2 x i64> @testUndefinedInsertq_2(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testUndefinedInsertq_2(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: ret <2 x i64> undef
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 32)
-  ret <2 x i64> %1
-}
-
-; CHECK: define <2 x i64> @testUndefinedInsertq_3(<2 x i64> %v, <2 x i64> %i)
-define <2 x i64> @testUndefinedInsertq_3(<2 x i64> %v, <2 x i64> %i) {
-; CHECK: ret <2 x i64> undef
-  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 64, i8 16)
-  ret <2 x i64> %1
-}
-
-; CHECK: declare <2 x i64> @llvm.x86.sse4a.insertqi
-declare <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64>, <2 x i64>, i8, i8) nounwind
+; RUN: opt < %s -instcombine -S | FileCheck %s
+
+; We should optimize these two redundant insertqi into one
+; CHECK: define <2 x i64> @testInsertTwice(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertTwice(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 32)
+; CHECK-NOT: insertqi
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 32)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 32)
+  ret <2 x i64> %2
+}
+
+; The result of this insert is the second arg, since the top 64 bits of
+; the result are undefined, and we copy the bottom 64 bits from the
+; second arg
+; CHECK: define <2 x i64> @testInsert64Bits(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsert64Bits(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: ret <2 x i64> %i
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 64, i8 0)
+  ret <2 x i64> %1
+}
+
+; Test the several types of ranges and ordering that exist for two insertqi
+; CHECK: define <2 x i64> @testInsertContainedRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertContainedRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 16)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertContainedRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertContainedRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 16)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertOverlappingRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertOverlappingRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 16)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertOverlappingRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertOverlappingRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 16)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertAdjacentRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertAdjacentRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertAdjacentRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertAdjacentRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 32)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertDisjointRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertDisjointRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+; CHECK: tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertDisjointRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertDisjointRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK:  tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+; CHECK:  tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testZeroLength(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testZeroLength(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: ret <2 x i64> %i
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 0, i8 0)
+  ret <2 x i64> %1
+}
+
+; CHECK: define <2 x i64> @testUndefinedInsertq_1(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testUndefinedInsertq_1(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 0, i8 16)
+  ret <2 x i64> %1
+}
+
+; CHECK: define <2 x i64> @testUndefinedInsertq_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testUndefinedInsertq_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 32)
+  ret <2 x i64> %1
+}
+
+; CHECK: define <2 x i64> @testUndefinedInsertq_3(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testUndefinedInsertq_3(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 64, i8 16)
+  ret <2 x i64> %1
+}
+
+;
+; Vector Demanded Bits
+;
+
+define <2 x i64> @test_extrq_arg0(<2 x i64> %x, <16 x i8> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_extrq_arg0
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %x, <16 x i8> %y)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %x, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %1, <16 x i8> %y) nounwind
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_extrq_arg1(<2 x i64> %x, <16 x i8> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_extrq_arg1
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %x, <16 x i8> %y)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <16 x i8> %y, <16 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %x, <16 x i8> %1) nounwind
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_extrq_args01(<2 x i64> %x, <16 x i8> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_extrq_args01
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %x, <16 x i8> %y)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %x, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = shufflevector <16 x i8> %y, <16 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
+  %3 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %1, <16 x i8> %2) nounwind
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @test_extrq_ret(<2 x i64> %x, <16 x i8> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_extrq_ret
+; CHECK-NEXT: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64> %x, <16 x i8> %y) nounwind
+  %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <2 x i32> <i32 1, i32 1>
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_extrqi_arg0(<2 x i64> %x) nounwind uwtable ssp {
+; CHECK-LABEL: @test_extrqi_arg0
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64> %x, i8 3, i8 2)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %x, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64> %1, i8 3, i8 2)
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_extrqi_ret(<2 x i64> %x) nounwind uwtable ssp {
+; CHECK-LABEL: @test_extrqi_ret
+; CHECK-NEXT: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64> %x, i8 3, i8 2) nounwind
+  %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <2 x i32> <i32 1, i32 1>
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_insertq_arg0(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_insertq_arg0
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.insertq(<2 x i64> %x, <2 x i64> %y)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %x, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertq(<2 x i64> %1, <2 x i64> %y) nounwind
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_insertq_ret(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_insertq_ret
+; CHECK-NEXT: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertq(<2 x i64> %x, <2 x i64> %y) nounwind
+  %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <2 x i32> <i32 1, i32 1>
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_insertqi_arg0(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_insertqi_arg0
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %x, <2 x i64> %y, i8 3, i8 2)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %x, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %y, i8 3, i8 2) nounwind
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_insertqi_arg1(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_insertqi_arg1
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %x, <2 x i64> %y, i8 3, i8 2)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %y, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %x, <2 x i64> %1, i8 3, i8 2) nounwind
+  ret <2 x i64> %2
+}
+
+define <2 x i64> @test_insertqi_args01(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_insertqi_args01
+; CHECK-NEXT: %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %x, <2 x i64> %y, i8 3, i8 2)
+; CHECK-NEXT: ret <2 x i64> %1
+  %1 = shufflevector <2 x i64> %x, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %2 = shufflevector <2 x i64> %y, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
+  %3 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %2, i8 3, i8 2) nounwind
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @test_insertqi_ret(<2 x i64> %x, <2 x i64> %y) nounwind uwtable ssp {
+; CHECK-LABEL: @test_insertqi_ret
+; CHECK-NEXT: ret <2 x i64> undef
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %x, <2 x i64> %y, i8 3, i8 2) nounwind
+  %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <2 x i32> <i32 1, i32 1>
+  ret <2 x i64> %2
+}
+
+; CHECK: declare <2 x i64> @llvm.x86.sse4a.extrq
+declare <2 x i64> @llvm.x86.sse4a.extrq(<2 x i64>, <16 x i8>) nounwind
+
+; CHECK: declare <2 x i64> @llvm.x86.sse4a.extrqi
+declare <2 x i64> @llvm.x86.sse4a.extrqi(<2 x i64>, i8, i8) nounwind
+
+; CHECK: declare <2 x i64> @llvm.x86.sse4a.insertq
+declare <2 x i64> @llvm.x86.sse4a.insertq(<2 x i64>, <2 x i64>) nounwind
+
+; CHECK: declare <2 x i64> @llvm.x86.sse4a.insertqi
+declare <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64>, <2 x i64>, i8, i8) nounwind