[InstCombine] Reassociate (C * X) * Y in foldICmpMulConstant

[junaire]

Summary

The existing logic in foldICmpMulConstant can only optimize the
constant if we have (X * Y) * C. However, reassociate pass currently
produces this form with the constant on the inner multiply, so let's
reassociate it in foldICmpMulConstant to enable further optimizations.

Related issue: https://github.com/llvm/llvm-project/issues/61538

Signed-off-by: Jun Zhang <[e-mail]>

[junaire]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2098

There's one corner case this doesn't handle, consider:
2 * X * Y > 0
2 * X will be folded into a shl. I'll appreciate it if anyone can suggest a more general way to match this.

[junaire]

dummy update to trigger CI

[nikic]

As I understand it, this is a reassociation problem. Once you reassociate the multiply from (X * C) * Y to (X * Y) * C this will automatically fold via the existing logic in foldICmpMulConstant(). I don't really want this logic to be duplicated.

It looks like reassociate currently produces this form with the constant on the inner multiply, but I'm not sure whether that's something that can be changed or not.

[junaire]

In D148210#4264628, @nikic wrote:

As I understand it, this is a reassociation problem. Once you reassociate the multiply from (X * C) * Y to (X * Y) * C this will automatically fold via the existing logic in foldICmpMulConstant(). I don't really want this logic to be duplicated.

It looks like reassociate currently produces this form with the constant on the inner multiply, but I'm not sure whether that's something that can be changed or not.

Thanks, that makes sense and solve my confusion when I wrote this. Should I try to fix the missing optimization by looking into the reassociate issue or the patch doesn't make sense at all?

[nikic]

In D148210#4264657, @junaire wrote:

In D148210#4264628, @nikic wrote:

As I understand it, this is a reassociation problem. Once you reassociate the multiply from (X * C) * Y to (X * Y) * C this will automatically fold via the existing logic in foldICmpMulConstant(). I don't really want this logic to be duplicated.

It looks like reassociate currently produces this form with the constant on the inner multiply, but I'm not sure whether that's something that can be changed or not.

Thanks, that makes sense and solve my confusion when I wrote this. Should I try to fix the missing optimization by looking into the reassociate issue or the patch doesn't make sense at all?

I took a quick look at reassociate, and I don't think it's possible to change what it does. Reversing the order is not really compatible with the ranking it performs.

In that case, I think we do want to handle this here, but we should try to extract the common code from foldICmpMulConstant(). I'm thinking that the helper function should accept the multiply constant, the icmp constant, the predicate and the OBO flags, and then return the new predicate and constant if possible. Then we'll call that function to check whether the reassociated form can fold and if so, we can actually materialize the new multiply and the new icmp.

[junaire]

In D148210#4264845, @nikic wrote:

In D148210#4264657, @junaire wrote:

In D148210#4264628, @nikic wrote:

As I understand it, this is a reassociation problem. Once you reassociate the multiply from (X * C) * Y to (X * Y) * C this will automatically fold via the existing logic in foldICmpMulConstant(). I don't really want this logic to be duplicated.

It looks like reassociate currently produces this form with the constant on the inner multiply, but I'm not sure whether that's something that can be changed or not.

Thanks, that makes sense and solve my confusion when I wrote this. Should I try to fix the missing optimization by looking into the reassociate issue or the patch doesn't make sense at all?

I took a quick look at reassociate, and I don't think it's possible to change what it does. Reversing the order is not really compatible with the ranking it performs.

In that case, I think we do want to handle this here, but we should try to extract the common code from foldICmpMulConstant(). I'm thinking that the helper function should accept the multiply constant, the icmp constant, the predicate and the OBO flags, and then return the new predicate and constant if possible. Then we'll call that function to check whether the reassociated form can fold and if so, we can actually materialize the new multiply and the new icmp.

Thanks, I'll try what you proposed.

[junaire]

Make a helper function.

[junaire]

In D148210#4264845, @nikic wrote:

In D148210#4264657, @junaire wrote:

In D148210#4264628, @nikic wrote:

As I understand it, this is a reassociation problem. Once you reassociate the multiply from (X * C) * Y to (X * Y) * C this will automatically fold via the existing logic in foldICmpMulConstant(). I don't really want this logic to be duplicated.

It looks like reassociate currently produces this form with the constant on the inner multiply, but I'm not sure whether that's something that can be changed or not.

Thanks, that makes sense and solve my confusion when I wrote this. Should I try to fix the missing optimization by looking into the reassociate issue or the patch doesn't make sense at all?

I took a quick look at reassociate, and I don't think it's possible to change what it does. Reversing the order is not really compatible with the ranking it performs.

In that case, I think we do want to handle this here, but we should try to extract the common code from foldICmpMulConstant(). I'm thinking that the helper function should accept the multiply constant, the icmp constant, the predicate and the OBO flags, and then return the new predicate and constant if possible. Then we'll call that function to check whether the reassociated form can fold and if so, we can actually materialize the new multiply and the new icmp.

I tried to unify icmp((MulC * X) * Y, C) and icmp(X * MulC, C) but it looks like adds more complexity. Because we have to extract the MulC bit by bit in case 1 in order to also get the nsw flags. And for case 2 MulC is needed in the below folds (MulC can't be zero or X * MulC == C --> X == C/MulC doesn't work anymore). What's more C < 0 is OK for case 2 but it doesn't apply to case 1. That said, though these cases have almost the same effect, they don't actually share too much common code...

[junaire]

Maybe we can just make helper function for icmp((MulC * X) * Y, C) as a special case

[junaire]

.

[junaire]

early bail out

[junaire]

Simplify

[nikic]

In D148210#4271709, @junaire wrote:

I tried to unify icmp((MulC * X) * Y, C) and icmp(X * MulC, C) but it looks like adds more complexity. Because we have to extract the MulC bit by bit in case 1 in order to also get the nsw flags. And for case 2 MulC is needed in the below folds (MulC can't be zero or X * MulC == C --> X == C/MulC doesn't work anymore). What's more C < 0 is OK for case 2 but it doesn't apply to case 1. That said, though these cases have almost the same effect, they don't actually share too much common code...

I'm not sure what you mean here. Why does MulC need to be extracted bit by bit? Don't we just need to implement these two rules on whether nuw/nsw can be preserved during reassociation? https://alive2.llvm.org/ce/z/KKtP62 https://alive2.llvm.org/ce/z/jgrYGB

[goldstein.w.n]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2049

Swap order of this conditions.

2057

match(Mul, m_c_Mul(m_Mul(m_Value(X), m_APInt(C)), m_Value(Y))

You don't need m_c_Mul when you are searching for constant being of canonicalization.

Also to avoid excessive scope depth, can you invert if the if i.e:
if(!match(...)) return nullptr

Likewise below when checking nsw/MulC s> 0

2060

the "mul" tag is probably not needed. Also I think generally you shouldn't be creating instructions unless you know its going to be used. You can still return nullptr here.

2061

ConstantInt:: -> Constant::

[bcl5980]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2056

It looks we need to limit both * to one-use.

[bcl5980]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2056

At least outer mul need one-use.

[junaire]

Try what @nikic proposed.

[junaire]

In D148210#4271835, @nikic wrote:

In D148210#4271709, @junaire wrote:

I tried to unify icmp((MulC * X) * Y, C) and icmp(X * MulC, C) but it looks like adds more complexity. Because we have to extract the MulC bit by bit in case 1 in order to also get the nsw flags. And for case 2 MulC is needed in the below folds (MulC can't be zero or X * MulC == C --> X == C/MulC doesn't work anymore). What's more C < 0 is OK for case 2 but it doesn't apply to case 1. That said, though these cases have almost the same effect, they don't actually share too much common code...

I'm not sure what you mean here. Why does MulC need to be extracted bit by bit? Don't we just need to implement these two rules on whether nuw/nsw can be preserved during reassociation? https://alive2.llvm.org/ce/z/KKtP62 https://alive2.llvm.org/ce/z/jgrYGB

Thanks, I think I get your ideas. Note I didn't add folds for nuw since I found the existing logic works great for it (https://alive2.llvm.org/ce/z/r_oHlt)
Note we should also reassociate for shl since X * Power_Of_N will be fold to X << N, I left it as a TODO. But I'm OK with that if you want to do it in this patch :)

[junaire]

Fix comments

[junaire]

handle shl cuz why not

[junaire]

Rebase

[junaire]

clang-format

[goldstein.w.n]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2071

1 << C, not C << C?

llvm/test/Transforms/InstCombine/icmp.ll

5345

Add some tests for the shl case. Better yet split adding the shl logic to a follow up patch.

[bcl5980]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2043

I'm not sure what are you going to do now. If you just want to reassociate C to outer mul, there is no extra condition. https://alive2.llvm.org/ce/z/v-f92Q
The code should be simple:

BinaryOperator *InnerMul;
Value *Y;
Constant *InnerC;
if (match(Mul, m_OneUse(m_c_Mul(m_BinOp(InnerMul), m_Value(Y)))) &&
    match(InnerMul, m_Mul(m_Value(X), m_Constant(InnerC)))) {
  Value *NewInner = Builder.CreateMul(X, Y);
  Value *NewMul = Builder.CreateMul(NewInner, InnerC);
  return new ICmpInst(Pred, NewMul, Mul->getOperand(1));
}

If you want to keep nsw/nuw flags, you can set the flag after create NewMul like:

if (!InnerC->isZeroValue() && Mul->hasNoUnsignedWrap() &&
    InnerMul->hasNoUnsignedWrap()) {
  cast<BinaryOperator>(NewInner)->setHasNoUnsignedWrap();
  cast<BinaryOperator>(NewMul)->setHasNoUnsignedWrap();
} else {
  Constant *Zero = Constant::getNullValue(InnerC->getType());
  Constant *InnerCSleZero =
      ConstantExpr::getCompare(ICmpInst::ICMP_SLE, InnerC, Zero);
  if (InnerCSleZero->isZeroValue() && Mul->hasNoSignedWrap() &&
      InnerMul->hasNoSignedWrap()) {
    cast<BinaryOperator>(NewInner)->setHasNoSignedWrap();
    cast<BinaryOperator>(NewMul)->setHasNoSignedWrap();
  }
}

2052

Do we really need llvm:: here? There is already using namespace llvm in this file at line 30.

[junaire]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2043

IIUC we want to reassociate (X * C) * Y to (X * Y) * C if they both has nsw/nuw flags so they can be folded by the existing logic in foldICmpMulConstant.

I'm not sure what are you going to do now. If you just want to reassociate C to outer mul, there is no extra condition. https://alive2.llvm.org/ce/z/v-f92Q
The code should be simple:

BinaryOperator *InnerMul;
Value *Y;
Constant *InnerC;
if (match(Mul, m_OneUse(m_c_Mul(m_BinOp(InnerMul), m_Value(Y)))) &&
    match(InnerMul, m_Mul(m_Value(X), m_Constant(InnerC)))) {
  Value *NewInner = Builder.CreateMul(X, Y);
  Value *NewMul = Builder.CreateMul(NewInner, InnerC);
  return new ICmpInst(Pred, NewMul, Mul->getOperand(1));
}

The issue is that if the inner mul's constant operand is a power of 2, it will be folded to an shl. like:

%mul = shl nsw i8 %a, i8 1
%mul2 = mul nsw i8 %b, %mul

Then the pattern match you proposed doesn't work anymore.

If you want to keep nsw/nuw flags, you can set the flag after create NewMul like:

if (!InnerC->isZeroValue() && Mul->hasNoUnsignedWrap() &&
    InnerMul->hasNoUnsignedWrap()) {
  cast<BinaryOperator>(NewInner)->setHasNoUnsignedWrap();
  cast<BinaryOperator>(NewMul)->setHasNoUnsignedWrap();
} else {
  Constant *Zero = Constant::getNullValue(InnerC->getType());
  Constant *InnerCSleZero =
      ConstantExpr::getCompare(ICmpInst::ICMP_SLE, InnerC, Zero);
  if (InnerCSleZero->isZeroValue() && Mul->hasNoSignedWrap() &&
      InnerMul->hasNoSignedWrap()) {
    cast<BinaryOperator>(NewInner)->setHasNoSignedWrap();
    cast<BinaryOperator>(NewMul)->setHasNoSignedWrap();
  }
}

2052

Do we really need llvm:: here? There is already using namespace llvm in this file at line 30.

Thanks, fixed.

2071

1 << C, not C << C?

good catch, fixed. Thx.

llvm/test/Transforms/InstCombine/icmp.ll

5345

I thought I already added tests for mul(shl(X, C), Y)? I don't have a strong opinion about whether we should delay shl support, I think that's easy to do?

[junaire]

Address comments.

[junaire]

fix nit.

[bcl5980]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2043

SHL is not the main thing I discuss here, we can support shl similar to mul. What I mean is do we really need to limit the transform to nsw/nuw. Always hoist C to outer mul should be a more general solution I think.

2093

should be

return new ICmpInst(Pred, NewMul, Mul->getOperand(1));

[junaire]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2043

Yeah, we can always hoist C to outer mul but does that have any benefit? We hoist it here cuz we know we can eliminate it later in foldICmpMulConstant. Can we do the same thing for muls that doesn't have nsw/nuw? I took a quick look at foldICmpMulConstant and believe the answer it no. Correct me if I'm wrong.

2093

Why return it? It doesn't make sense to me. I thought what we want to do is reassociate (X * C) * Y to (X * Y) * C so the later code can recognize the pattern and fold it. Thus we reuse the existing logic.

[bcl5980]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2043

That is not for foldICmpMulConstant only. Constant may match more patterns generally. For example, In foldICmpBinOp:
icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
@nikic What do you think about this?

2093

After return you still can reuse existing. I'm not sure if currect code will break something or not but I don't think we have similar usage in instcombine.

[junaire]

Always reassociate despite we have OBO flags or not.

[junaire]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2043

OK! I updated the patch and added a test when we don't have nsw.

2093

Hi @bcl5980, thanks for your suggestion, I appreciate it. However, after I tried to replace replaceInstUsesWith(*Mul, NewMul); to return new ICmpInst(Pred, NewMul, Mul->getOperand(1));, the folds in my tests stop working. I took another deep look and I think your suggestion should be right, let me try to investigate why... (any more comments are welcome

[junaire]

Simplify

[junaire]

replace replaceInstUsesWith(*Mul, NewMul); with return new ICmpInst(Pred, NewMul, ConstantInt::get(Mul->getType(), C));

[junaire]

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2093

So that's new ICmpInst(Pred, NewMul, ConstantInt::get(Mul->getType(), C));, not return new ICmpInst(Pred, NewMul, Mul->getOperand(1));. Fixed, thanks!

[goldstein.w.n]

llvm/test/Transforms/InstCombine/icmp.ll

4970–4971

I don't see any tests with nuw or for the comination of the flags.

[junaire]

Add tests for nuw

[junaire]

llvm/test/Transforms/InstCombine/icmp.ll

4970–4971

Thanks, I added more tests. PTAL :)

[junaire]

The review process seems to be stuck. What's your take on the patch? @nikic

[junaire]

gentle ping :)

[junaire]

Rebase + ping :)

[junaire]

Ooops, updated a wrong diff, sorry

[nikic]

The reassociation code itself looks basically fine. The problem is how it's used. There is two ways we can do this:

1. Say that InstCombine is always going to reassociate the constant to the outside, in opposition to -reassociate. (Is there precedent for this?)
2. Factor out the logic for "will this fold" from the logic for "do the fold" in foldICmpMulConstant(). Then only "virtually" reassociate and only perform the transform if it's actually going to simplify.

The current implementation is in between those two: It only reassociates if there's a comparison, but there's no actual guarantee that the resulting reassociation will actually fold (unless I'm missing something).

Option 2 is what I was suggesting before. I'm open to considering option 1, but I don't full understand the implications. I think one problem could be that this breaks up a multiply that would otherwise be loop invariant.

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

2047

IRBuilderBase &

2050

It looks like this variable declaration can be moved into the if block below.

2053

I'd probably accept CI as APInt here as well, and move creation of the ConstantInt into this helper.

2073

Unnecessary newline.

2078

Should check here that the shift is valid, otherwise we'll assert. (It will usually get simplified, but is not guaranteed.)

[junaire]

Been busy working on other topics so just address some nit issues to indicate the patch isn't dead yet

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
     Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
     auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
     return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
   }
 
   return nullptr;
 }
 
+// Reassociate (C * X) * Y to (X * Y) * C to enable further
+// optimizations.
+static Value *tryReassociateMulConstant(ICmpInst::Predicate Pred,
+                                        BinaryOperator *Mul,
+                                        InstCombiner::BuilderTy &Builder) {
+  BinaryOperator *Inner;
+  Value *X, *Y;
+  const APInt *C;
+
+  auto Reassociate = [&](Value *X, Value *Y, Constant *CI) -> llvm::Value * {
+    if (Mul->hasNoSignedWrap() && Inner->hasNoSignedWrap() &&
+        C->isStrictlyPositive()) {
+      Value *NewInner = Builder.CreateNSWMul(X, Y);
+      return Builder.CreateNSWMul(NewInner, CI);
+    }
+    return nullptr;
+  };
+
+  if (match(Mul, m_OneUse(m_c_Mul(m_BinOp(Inner), m_Value(Y))))) {
+
+    if (match(Inner, m_Mul(m_Value(X), m_APInt(C))))
+      if (Value *NewMul =
+              Reassociate(X, Y, ConstantInt::get(Mul->getType(), *C)))
+        return NewMul;
+
+    // X * Power_of_N will be fold to X << N, recognize the pattern.
+    if (match(Inner, m_Shl(m_Value(X), m_APInt(C))))
+      if (Value *NewMul =
+              Reassociate(X, Y, ConstantInt::get(Mul->getType(), C->shl(*C))))
+        return NewMul;
+  }
+  return nullptr;
+}
+
 /// Fold icmp (mul X, Y), C.
 Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
                                                    BinaryOperator *Mul,
                                                    const APInt &C) {
   ICmpInst::Predicate Pred = Cmp.getPredicate();
   Type *MulTy = Mul->getType();
   Value *X = Mul->getOperand(0);
 
   // If there's no overflow:
   // X * X == 0 --> X == 0
   // X * X != 0 --> X != 0
   if (Cmp.isEquality() && C.isZero() && X == Mul->getOperand(1) &&
       (Mul->hasNoUnsignedWrap() || Mul->hasNoSignedWrap()))
     return new ICmpInst(Pred, X, ConstantInt::getNullValue(MulTy));
 
+  if (Value *NewMul = tryReassociateMulConstant(Pred, Mul, Builder))
+    replaceInstUsesWith(*Mul, NewMul);
+
   const APInt *MulC;
   if (!match(Mul->getOperand(1), m_APInt(MulC)))
     return nullptr;
 
   // If this is a test of the sign bit and the multiply is sign-preserving with
   // a constant operand, use the multiply LHS operand instead:
   // (X * +MulC) < 0 --> X < 0
   // (X * -MulC) < 0 --> X > 0
   if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
     if (MulC->isNegative())
       Pred = ICmpInst::getSwappedPredicate(Pred);
     return new ICmpInst(Pred, X, ConstantInt::getNullValue(MulTy));

llvm/test/Transforms/InstCombine/icmp.ll

 ;
   %b = or i8 %a, 24
   call void @use_i8(i8 %b)
   %cmp = icmp sgt i8 %b, 0
   ret i1 %cmp
 }
 
 define i1 @mul_mul_sge_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_sge_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], -1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sge i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_sge_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_sge_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], <i8 -1, i8 -1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], <i8 -1, i8 -1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sge <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_sge_one(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_sge_one(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sge i8 %mul2, 1
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_sge_one_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_sge_one_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sge <2 x i8> %mul2, <i8 1, i8 1>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_slt_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_slt_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp slt i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_slt_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_slt_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp slt <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_slt_one(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_slt_one(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp slt i8 %mul2, 1
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_slt_one_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_slt_one_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], <i8 1, i8 1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], <i8 1, i8 1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp slt <2 x i8> %mul2, <i8 1, i8 1>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_sle_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_sle_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sle i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_sle_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_sle_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], <i8 1, i8 1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], <i8 1, i8 1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sle <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_sle_negone(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_sle_negone(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sle i8 %mul2, -1
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_sle_negone_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_sle_negone_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sle <2 x i8> %mul2, <i8 -1, i8 -1>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_sgt_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_sgt_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sgt i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_sgt_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_sgt_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sgt <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_sgt_negone(i8 %a, i8 %b) {
 ; CHECK-LABEL: @mul_mul_sgt_negone(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i8 [[A:%.*]], 42
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], -1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = mul nsw i8 %a, 42
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sgt i8 %mul2, -1
   ret i1 %cmp
 }
 
 define <2 x i1> @mul_mul_sgt_negone_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @mul_mul_sgt_negone_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = mul nsw <2 x i8> [[A:%.*]], <i8 42, i8 42>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], <i8 -1, i8 -1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], <i8 -1, i8 -1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = mul nsw <2 x i8> %a, <i8 42, i8 42>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sgt <2 x i8> %mul2, <i8 -1, i8 -1>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_sge_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_sge_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], -1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sge i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_sge_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_sge_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], <i8 -1, i8 -1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], <i8 -1, i8 -1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sge <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_sge_one(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_sge_one(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sge i8 %mul2, 1
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_sge_one_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_sge_one_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sge <2 x i8> %mul2, <i8 1, i8 1>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_slt_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_slt_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp slt i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_slt_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_slt_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp slt <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_slt_one(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_slt_one(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp slt i8 %mul2, 1
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_slt_one_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_slt_one_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], <i8 1, i8 1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], <i8 1, i8 1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp slt <2 x i8> %mul2, <i8 1, i8 1>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_sle_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_sle_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sle i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_sle_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_sle_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], <i8 1, i8 1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], <i8 1, i8 1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sle <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_sle_negone(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_sle_negone(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sle i8 %mul2, -1
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_sle_negone_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_sle_negone_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sle <2 x i8> %mul2, <i8 -1, i8 -1>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_sgt_zero(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_sgt_zero(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sgt i8 %mul2, 0
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_sgt_zero_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_sgt_zero_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sgt <2 x i8> %mul2, <i8 0, i8 0>
   ret <2 x i1> %cmp
 }
 
 define i1 @shl_mul_sgt_negone(i8 %a, i8 %b) {
 ; CHECK-LABEL: @shl_mul_sgt_negone(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i8 [[A:%.*]], 1
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw i8 [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[MUL2]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw i8 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[TMP1]], -1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %mul = shl nsw i8 %a, 1
   %mul2 = mul nsw i8 %b, %mul
   %cmp = icmp sgt i8 %mul2, -1
   ret i1 %cmp
 }
 
 define <2 x i1> @shl_mul_sgt_negone_vec(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @shl_mul_sgt_negone_vec(
-; CHECK-NEXT:    [[MUL:%.*]] = shl nsw <2 x i8> [[A:%.*]], <i8 1, i8 1>
-; CHECK-NEXT:    [[MUL2:%.*]] = mul nsw <2 x i8> [[MUL]], [[B:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[MUL2]], <i8 -1, i8 -1>
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <2 x i8> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <2 x i8> [[TMP1]], <i8 -1, i8 -1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %mul = shl nsw <2 x i8> %a, <i8 1, i8 1>
   %mul2 = mul nsw <2 x i8> %b, %mul
   %cmp = icmp sgt <2 x i8> %mul2, <i8 -1, i8 -1>
   ret <2 x i1> %cmp
 }
 
 define i1 @mul_mul_sge_zero_multiuse(i8 %a, i8 %b) {