[Reassociate] Add initial support for vector instructions.

[rob.lougher]

Summary

This patch adds initial support for vector instructions to the reassociation pass. It enables most parts of the pass to work with vectors but to keep the size of the patch small, optimization of Xor trees, canonicalization of negative constants and converting shifts to muls, etc., have been left out. This will be handled in later patches.

The patch is based on an initial patch by Chad Rosier (see http://reviews.llvm.org/D5222).

Robert Lougher
SN Systems - Sony Computer Entertainment Group

[rob.lougher]

Rebased the patch and added tests for 'or' and 'and' (previously these just showed that the operands were commuted - the tests now check that the expressions are optimized).

This is the minimum changes necessary to add reassociation of vectors. The idea is to split the changes into smaller, reviewable chunks. The differences with Chad's original patch is:

1. Xor instructions are not optimized (the code in the original patch is incomplete). This will be added later.

2. Various constant handling is missing (e.g. removal of negative constant factors). Chad's original patch extended the checks to include constant splats. Again, these will addressed in later patches.

3. CanonicalizeNegConstExpr() was added after Chad's patch. Again, this will be handled in later patches.

4. Tests that rely on other passes (gvn, instcombine) have either been modified or removed.

5. FIXME tests have been added for some of the missing optimizations.

Chad's original review has been dormant since October. Note, I emailed Chad privately to make sure he was happy with someone else continuing the work.

[rob.lougher]

ping

[spatel]

lib/Transforms/Scalar/Reassociate.cpp

2091–2093

Can you add a bit to the comment to explain?
Is this because we want to distinguish zero'ing idioms?

test/Transforms/Reassociate/fast-ReassociateVector.ll

4

I assume all of the floating point transforms apply identically to doubles as well as floats. Can you change some of these tests to use doubles so we have some coverage for those?

[rob.lougher]

Thanks for the review Sanjay.

lib/Transforms/Scalar/Reassociate.cpp

2091–2093

OK, I'll add a comment here. The only reason is that OptimizeXor (and helpers) hasn't been vectorized yet. This will be done in a later patch.

test/Transforms/Reassociate/fast-ReassociateVector.ll

4

OK. I'll update the patch.

[rob.lougher]

New version of the patch that addresses Sanjay's comments. Tests 3, 7, 9 and 11 now use double rather than float.

[mcrosier]

Thanks for taking this on, Rob. I agree that this should be committed in incremental patches.

What kind of correctness/performance testing have you done?

IIRC, I didn't see a lot of change in SPEC2000/SPEC2006 with my original patch.

[rob.lougher]

Hi Chad,

Sorry for the delay. I have ran SPEC CPU 2006 and open-source Bullet. Unfortunately, I found no performance difference at all. Sanjay also tried the n-body sim from https://github.com/tycho/nbody and didn't see any perf difference either. Apart from micro-benchmarks, the only consistent speed-up I see is in an internal benchmark of ~1%. I didn't see any performance regressions during testing.

[mcrosier]

LGTM. My past experience is that this really made no difference for my workloads (i.e., spec2000/spec2006/eembc). I just wanted to make sure we didn't have any odd regressions.

[rob.lougher]

I reverted the commit as it caused a couple of arm64 clang tests to fail. This was not expected to happen after a change to LLVM. Hopefully will be easy to fix and I'll resubmit on Monday.

[rob.lougher]

I have reapplied the patch at revision 232209. Hopefully it will be OK this time!

lib/Transforms/Scalar/Reassociate.cpp

unsigned Reassociate::getRank(Value *V) {
   // cannot be loops in the value graph that do not go through PHI nodes.
   unsigned Rank = 0, MaxRank = RankMap[I->getParent()];
   for (unsigned i = 0, e = I->getNumOperands();
        i != e && Rank != MaxRank; ++i)
     Rank = std::max(Rank, getRank(I->getOperand(i)));
 
   // If this is a not or neg instruction, do not count it for rank.  This
   // assures us that X and ~X will have the same rank.
-  Type *Ty = V->getType();
-  if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
-      (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
-       !BinaryOperator::isFNeg(I)))
+  if  (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
+       !BinaryOperator::isFNeg(I))
     ++Rank;
 
   DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
 
   return ValueRankMap[I] = Rank;
 }
 
 // Canonicalize constants to RHS.  Otherwise, sort the operands by rank.
if (isa<Constant>(RHS))
     return;
 
   if (isa<Constant>(LHS) || RHSRank < LHSRank)
     cast<BinaryOperator>(I)->swapOperands();
 }
 
 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
                                  Instruction *InsertBefore, Value *FlagsOp) {
-  if (S1->getType()->isIntegerTy())
+  if (S1->getType()->isIntOrIntVectorTy())
     return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
   else {
     BinaryOperator *Res =
         BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore);
     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
     return Res;
   }
 }
 
 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
                                  Instruction *InsertBefore, Value *FlagsOp) {
-  if (S1->getType()->isIntegerTy())
+  if (S1->getType()->isIntOrIntVectorTy())
     return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
   else {
     BinaryOperator *Res =
       BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore);
     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
     return Res;
   }
 }
 
 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
                                  Instruction *InsertBefore, Value *FlagsOp) {
-  if (S1->getType()->isIntegerTy())
+  if (S1->getType()->isIntOrIntVectorTy())
     return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
   else {
     BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
     Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags());
     return Res;
   }
 }
 
 /// LowerNegateToMultiply - Replace 0-X with X*-1.
 ///
 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
   Type *Ty = Neg->getType();
-  Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
-                                       : ConstantFP::get(Ty, -1.0);
+  Constant *NegOne = Ty->isIntOrIntVectorTy() ?
+    ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
 
   BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
   Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
   Res->takeName(Neg);
   Neg->replaceAllUsesWith(Res);
   Res->setDebugLoc(Neg->getDebugLoc());
   return Res;
 }
for (unsigned i = 0; ; ++i) {
     // hard (finding the mimimal number of multiplications needed to realize a
     // multiplication expression is NP-complete).  Whatever the reason, smart or
     // stupid, create a new node if there are none left.
     BinaryOperator *NewOp;
     if (NodesToRewrite.empty()) {
       Constant *Undef = UndefValue::get(I->getType());
       NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
                                      Undef, Undef, "", I);
-      if (NewOp->getType()->isFloatingPointTy())
+      if (NewOp->getType()->isFPOrFPVectorTy())
         NewOp->setFastMathFlags(I->getFastMathFlags());
     } else {
       NewOp = NodesToRewrite.pop_back_val();
     }
 
     DEBUG(dbgs() << "RA: " << *Op << '\n');
     Op->setOperand(0, NewOp);
     DEBUG(dbgs() << "TO: " << *Op << '\n');
if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) {
         ++NumFound;
       } while (i != Ops.size() && Ops[i].Op == TheOp);
 
       DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n');
       ++NumFactor;
 
       // Insert a new multiply.
       Type *Ty = TheOp->getType();
-      Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
-                                      : ConstantFP::get(Ty, NumFound);
+      Constant *C = Ty->isIntOrIntVectorTy() ?
+        ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
       Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
 
       // Now that we have inserted a multiply, optimize it. This allows us to
       // handle cases that require multiple factoring steps, such as this:
       // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6
       RedoInsts.insert(Mul);
 
       // If every add operand was a duplicate, return the multiply.
if (MaxOcc > 1) {
     DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n');
     ++NumFactor;
 
     // Create a new instruction that uses the MaxOccVal twice.  If we don't do
     // this, we could otherwise run into situations where removing a factor
     // from an expression will drop a use of maxocc, and this can cause
     // RemoveFactorFromExpression on successive values to behave differently.
     Instruction *DummyInst =
-        I->getType()->isIntegerTy()
+        I->getType()->isIntOrIntVectorTy()
             ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
             : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
 
     SmallVector<WeakVH, 4> NewMulOps;
     for (unsigned i = 0; i != Ops.size(); ++i) {
       // Only try to remove factors from expressions we're allowed to.
       BinaryOperator *BOp =
           isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul);

 /// \brief Build a tree of multiplies, computing the product of Ops.
 static Value *buildMultiplyTree(IRBuilder<> &Builder,
                                 SmallVectorImpl<Value*> &Ops) {
   if (Ops.size() == 1)
     return Ops.back();
 
   Value *LHS = Ops.pop_back_val();
   do {
-    if (LHS->getType()->isIntegerTy())
+    if (LHS->getType()->isIntOrIntVectorTy())
       LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
     else
       LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
   } while (!Ops.empty());
 
   return LHS;
 }
 
if (Instruction *Res = canonicalizeNegConstExpr(I))
     I = Res;
 
   // Commute binary operators, to canonicalize the order of their operands.
   // This can potentially expose more CSE opportunities, and makes writing other
   // transformations simpler.
   if (I->isCommutative())
     canonicalizeOperands(I);
 
-  // Don't optimize vector instructions.
-  if (I->getType()->isVectorTy())
+  // TODO: We should optimize vector Xor instructions, but they are
+  // currently unsupported.
+  if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor)
     return;
 
   // Don't optimize floating point instructions that don't have unsafe algebra.
   if (I->getType()->isFloatingPointTy() && !I->hasUnsafeAlgebra())
     return;
 
   // Do not reassociate boolean (i1) expressions.  We want to preserve the
   // original order of evaluation for short-circuited comparisons that
void Reassociate::OptimizeInst(Instruction *I) {
   if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd &&
       cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub)
     return;
 
   ReassociateExpression(BO);
 }
 
 void Reassociate::ReassociateExpression(BinaryOperator *I) {
-  assert(!I->getType()->isVectorTy() &&
-         "Reassociation of vector instructions is not supported.");
-
   // First, walk the expression tree, linearizing the tree, collecting the
   // operand information.
   SmallVector<RepeatedValue, 8> Tree;
   MadeChange |= LinearizeExprTree(I, Tree);
   SmallVector<ValueEntry, 8> Ops;
   Ops.reserve(Tree.size());
   for (unsigned i = 0, e = Tree.size(); i != e; ++i) {
     RepeatedValue E = Tree[i];

test/Transforms/Reassociate/fast-ReassociateVector.ll

 ; RUN: opt < %s -reassociate -S | FileCheck %s
 
-; Canonicalize operands, but don't optimize floating point vector operations.
-define <4 x float> @test1() {
-; CHECK-LABEL: test1
-; CHECK-NEXT: %tmp1 = fsub fast <4 x float> zeroinitializer, zeroinitializer
-; CHECK-NEXT: %tmp2 = fmul fast <4 x float> %tmp1, zeroinitializer
-
-  %tmp1 = fsub fast <4 x float> zeroinitializer, zeroinitializer
-  %tmp2 = fmul fast <4 x float> zeroinitializer, %tmp1
-  ret <4 x float> %tmp2
-}
-
-; Commute integer vector operations.
-define <2 x i32> @test2(<2 x i32> %x, <2 x i32> %y) {
-; CHECK-LABEL: test2
-; CHECK-NEXT: %tmp1 = add <2 x i32> %x, %y
-; CHECK-NEXT: %tmp2 = add <2 x i32> %x, %y
-; CHECK-NEXT: %tmp3 = add <2 x i32> %tmp1, %tmp2
-
-  %tmp1 = add <2 x i32> %x, %y
-  %tmp2 = add <2 x i32> %y, %x
-  %tmp3 = add <2 x i32> %tmp1, %tmp2
-  ret <2 x i32> %tmp3
-}
-
-define <2 x i32> @test3(<2 x i32> %x, <2 x i32> %y) {
-; CHECK-LABEL: test3
-; CHECK-NEXT: %tmp1 = mul <2 x i32> %x, %y
-; CHECK-NEXT: %tmp2 = mul <2 x i32> %x, %y
-; CHECK-NEXT: %tmp3 = mul <2 x i32> %tmp1, %tmp2
-
-  %tmp1 = mul <2 x i32> %x, %y
-  %tmp2 = mul <2 x i32> %y, %x
-  %tmp3 = mul <2 x i32> %tmp1, %tmp2
-  ret <2 x i32> %tmp3
-}
-
-define <2 x i32> @test4(<2 x i32> %x, <2 x i32> %y) {
-; CHECK-LABEL: test4
-; CHECK-NEXT: %tmp1 = and <2 x i32> %x, %y
-; CHECK-NEXT: %tmp2 = and <2 x i32> %x, %y
-; CHECK-NEXT: %tmp3 = and <2 x i32> %tmp1, %tmp2
+; Check that a*c+b*c is turned into (a+b)*c
+define <4 x float> @test1(<4 x float> %a, <4 x float> %b, <4 x float> %c) {
+; CHECK-LABEL: @test1
+; CHECK-NEXT: %tmp = fadd fast <4 x float> %b, %a
+; CHECK-NEXT: %tmp1 = fmul fast <4 x float> %tmp, %c
+; CHECK-NEXT: ret <4 x float> %tmp1
+
+  %mul = fmul fast <4 x float> %a, %c
+  %mul1 = fmul fast <4 x float> %b, %c
+  %add = fadd fast <4 x float> %mul, %mul1
+  ret <4 x float> %add
+}
+
+; Check that a*a*b+a*a*c is turned into a*(a*(b+c)).
+define <2 x float> @test2(<2 x float> %a, <2 x float> %b, <2 x float> %c) {
+; CHECK-LABEL: @test2
+; CHECK-NEXT: fadd fast <2 x float> %c, %b
+; CHECK-NEXT: fmul fast <2 x float> %a, %tmp2
+; CHECK-NEXT: fmul fast <2 x float> %tmp3, %a
+; CHECK-NEXT: ret <2 x float>
+
+  %t0 = fmul fast <2 x float> %a, %b
+  %t1 = fmul fast <2 x float> %a, %t0
+  %t2 = fmul fast <2 x float> %a, %c
+  %t3 = fmul fast <2 x float> %a, %t2
+  %t4 = fadd fast <2 x float> %t1, %t3
+  ret <2 x float> %t4
+}
+
+; Check that a*b+a*c+d is turned into a*(b+c)+d.
+define <2 x double> @test3(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %d) {
+; CHECK-LABEL: @test3
+; CHECK-NEXT: fadd fast <2 x double> %c, %b
+; CHECK-NEXT: fmul fast <2 x double> %tmp, %a
+; CHECK-NEXT: fadd fast <2 x double> %tmp1, %d
+; CHECK-NEXT: ret <2 x double>
+
+  %t0 = fmul fast <2 x double> %a, %b
+  %t1 = fmul fast <2 x double> %a, %c
+  %t2 = fadd fast <2 x double> %t1, %d
+  %t3 = fadd fast <2 x double> %t0, %t2
+  ret <2 x double> %t3
+}
+
+; No fast-math.
+define <2 x float> @test4(<2 x float> %A) {
+; CHECK-LABEL: @test4
+; CHECK-NEXT: %X = fadd <2 x float> %A, <float 1.000000e+00, float 1.000000e+00>
+; CHECK-NEXT: %Y = fadd <2 x float> %A, <float 1.000000e+00, float 1.000000e+00>
+; CHECK-NEXT: %R = fsub <2 x float> %X, %Y
+; CHECK-NEXT: ret <2 x float> %R
+
+  %X = fadd <2 x float> %A, < float 1.000000e+00, float 1.000000e+00 >
+  %Y = fadd <2 x float> %A, < float 1.000000e+00, float 1.000000e+00 >
+  %R = fsub <2 x float> %X, %Y
+  ret <2 x float> %R
+}
+
+; Check 47*X + 47*X -> 94*X.
+define <2 x float> @test5(<2 x float> %X) {
+; CHECK-LABEL: @test5
+; CHECK-NEXT: fmul fast <2 x float> %X, <float 9.400000e+01, float 9.400000e+01>
+; CHECK-NEXT: ret <2 x float>
+
+  %Y = fmul fast <2 x float> %X, <float 4.700000e+01, float 4.700000e+01>
+  %Z = fadd fast <2 x float> %Y, %Y
+  ret <2 x float> %Z
+}
+
+; Check X+X+X -> 3*X.
+define <2 x float> @test6(<2 x float> %X) {
+; CHECK-LABEL: @test6
+; CHECK-NEXT: fmul fast <2 x float> %X, <float 3.000000e+00, float 3.000000e+00>
+; CHECK-NEXT: ret <2 x float>
+
+  %Y = fadd fast <2 x float> %X ,%X
+  %Z = fadd fast <2 x float> %Y, %X
+  ret <2 x float> %Z
+}
+
+; Check 127*W+50*W -> 177*W.
+define <2 x double> @test7(<2 x double> %W) {
+; CHECK-LABEL: @test7
+; CHECK-NEXT: fmul fast <2 x double> %W, <double 1.770000e+02, double 1.770000e+02>
+; CHECK-NEXT: ret <2 x double>
+
+  %X = fmul fast <2 x double> %W, <double 127.0, double 127.0>
+  %Y = fmul fast <2 x double> %W, <double 50.0, double 50.0>
+  %Z = fadd fast <2 x double> %Y, %X
+  ret <2 x double> %Z
+}
+
+; Check X*12*12 -> X*144.
+define <2 x float> @test8(<2 x float> %arg) {
+; CHECK-LABEL: @test8
+; CHECK: fmul fast <2 x float> %arg, <float 1.440000e+02, float 1.440000e+02>
+; CHECK-NEXT: ret <2 x float> %tmp2
+
+  %tmp1 = fmul fast <2 x float> <float 1.200000e+01, float 1.200000e+01>, %arg
+  %tmp2 = fmul fast <2 x float> %tmp1, <float 1.200000e+01, float 1.200000e+01>
+  ret <2 x float> %tmp2
+}
+
+; Check (b+(a+1234))+-a -> b+1234.
+define <2 x double> @test9(<2 x double> %b, <2 x double> %a) {
+; CHECK-LABEL: @test9
+; CHECK: fadd fast <2 x double> %b, <double 1.234000e+03, double 1.234000e+03>
+; CHECK-NEXT: ret <2 x double>
+
+  %1 = fadd fast <2 x double> %a, <double 1.234000e+03, double 1.234000e+03>
+  %2 = fadd fast <2 x double> %b, %1
+  %3 = fsub fast <2 x double> <double 0.000000e+00, double 0.000000e+00>, %a
+  %4 = fadd fast <2 x double> %2, %3
+  ret <2 x double> %4
+}
+
+; Check -(-(z*40)*a) -> a*40*z.
+define <2 x float> @test10(<2 x float> %a, <2 x float> %b, <2 x float> %z) {
+; CHECK-LABEL: @test10
+; CHECK: fmul fast <2 x float> %a, <float 4.000000e+01, float 4.000000e+01>
+; CHECK-NEXT: fmul fast <2 x float> %e, %z
+; CHECK-NEXT: ret <2 x float>
+
+  %d = fmul fast <2 x float> %z, <float 4.000000e+01, float 4.000000e+01>
+  %c = fsub fast <2 x float> <float 0.000000e+00, float 0.000000e+00>, %d
+  %e = fmul fast <2 x float> %a, %c
+  %f = fsub fast <2 x float> <float 0.000000e+00, float 0.000000e+00>, %e
+  ret <2 x float> %f
+}
+
+; Check x*y+y*x -> x*y*2.
+define <2 x double> @test11(<2 x double> %x, <2 x double> %y) {
+; CHECK-LABEL: @test11
+; CHECK-NEXT: %factor = fmul fast <2 x double> %y, <double 2.000000e+00, double 2.000000e+00>
+; CHECK-NEXT: %tmp1 = fmul fast <2 x double> %factor, %x
+; CHECK-NEXT: ret <2 x double> %tmp1
+
+  %1 = fmul fast <2 x double> %x, %y
+  %2 = fmul fast <2 x double> %y, %x
+  %3 = fadd fast <2 x double> %1, %2
+  ret <2 x double> %3
+}
+
+; FIXME: shifts should be converted to mul to assist further reassociation.
+define <2 x i64> @test12(<2 x i64> %b, <2 x i64> %c) {
+; CHECK-LABEL: @test12
+; CHECK-NEXT:  %mul = mul <2 x i64> %c, %b
+; CHECK-NEXT:  %shl = shl <2 x i64> %mul, <i64 5, i64 5>
+; CHECK-NEXT:  ret <2 x i64> %shl
+
+  %mul = mul <2 x i64> %c, %b
+  %shl = shl <2 x i64> %mul, <i64 5, i64 5>
+  ret <2 x i64> %shl
+}
+
+; FIXME: expressions with a negative const should be canonicalized to assist
+; further reassociation.
+; We would expect (-5*b)+a -> a-(5*b) but only the constant operand is commuted.
+define <4 x float> @test13(<4 x float> %a, <4 x float> %b) {
+; CHECK-LABEL: @test13
+; CHECK-NEXT:  %mul = fmul fast <4 x float> %b, <float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00>
+; CHECK-NEXT:  %add = fadd fast <4 x float> %mul, %a
+; CHECK-NEXT:  ret <4 x float> %add
+
+  %mul = fmul fast <4 x float> <float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00>, %b
+  %add = fadd fast <4 x float> %mul, %a
+  ret <4 x float> %add
+}
+
+; Break up subtract to assist further reassociation.
+; Check a+b-c -> a+b+-c.
+define <2 x i64> @test14(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c) {
+; CHECK-LABEL: @test14
+; CHECK-NEXT: %add = add <2 x i64> %b, %a
+; CHECK-NEXT: %c.neg = sub <2 x i64> zeroinitializer, %c
+; CHECK-NEXT: %sub = add <2 x i64> %add, %c.neg
+; CHECK-NEXT: ret <2 x i64> %sub
+
+  %add = add <2 x i64> %b, %a
+  %sub = sub <2 x i64> %add, %c
+  ret <2 x i64> %sub
+}
+
+define <2 x i32> @test15(<2 x i32> %x, <2 x i32> %y) {
+; CHECK-LABEL: test15
+; CHECK-NEXT: %tmp3 = and <2 x i32> %y, %x
+; CHECK-NEXT: ret <2 x i32> %tmp3
 
   %tmp1 = and <2 x i32> %x, %y
   %tmp2 = and <2 x i32> %y, %x
   %tmp3 = and <2 x i32> %tmp1, %tmp2
   ret <2 x i32> %tmp3
 }
 
-define <2 x i32> @test5(<2 x i32> %x, <2 x i32> %y) {
-; CHECK-LABEL: test5
-; CHECK-NEXT: %tmp1 = or <2 x i32> %x, %y
-; CHECK-NEXT: %tmp2 = or <2 x i32> %x, %y
-; CHECK-NEXT: %tmp3 = or <2 x i32> %tmp1, %tmp2
+define <2 x i32> @test16(<2 x i32> %x, <2 x i32> %y) {
+; CHECK-LABEL: test16
+; CHECK-NEXT: %tmp3 = or <2 x i32> %y, %x
+; CHECK-NEXT: ret <2 x i32> %tmp3
 
   %tmp1 = or <2 x i32> %x, %y
   %tmp2 = or <2 x i32> %y, %x
   %tmp3 = or <2 x i32> %tmp1, %tmp2
   ret <2 x i32> %tmp3
 }
 
-define <2 x i32> @test6(<2 x i32> %x, <2 x i32> %y) {
-; CHECK-LABEL: test6
+; FIXME: Optimize vector xor.  Currently only commute operands.
+define <2 x i32> @test17(<2 x i32> %x, <2 x i32> %y) {
+; CHECK-LABEL: test17
 ; CHECK-NEXT: %tmp1 = xor <2 x i32> %x, %y
 ; CHECK-NEXT: %tmp2 = xor <2 x i32> %x, %y
 ; CHECK-NEXT: %tmp3 = xor <2 x i32> %tmp1, %tmp2
 
   %tmp1 = xor <2 x i32> %x, %y
   %tmp2 = xor <2 x i32> %y, %x
   %tmp3 = xor <2 x i32> %tmp1, %tmp2
   ret <2 x i32> %tmp3
 }