[InstCombine] Minor optimization for bswap with binary ops

[RKSimon]

Summary

Added inst combine patterns for BSWAP with and/or/xor OP:

OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )

Fixes PR15782

Since its just a one liner, I've also added BSWAP to the DAGCombiner equivalent as well:

fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))

[qcolombet]

Hi Simon,

This LGTM with a couple of minor fixes. See my inline comments.

Also, for the tests, please use opt -instnamer on the input to rename the variables. Indeed, having numbered variables makes the tests harder to update.
Moreover, I know that this tests check that we end up with no bswap. To this purpose, you added a bswap at the end of each of the computation. I believe that it complicates the test cases with no value. I think it would be best to use simpler test cases with CHECK lines, but you will have to fix all the tests.
I leave that up to you and it can be done as a subsequent commit.

Thanks,
-Quentin

lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

148 ↗

(On Diff #16619)

Use early exit.

test/Transforms/InstCombine/bswap-fold.ll

70 ↗

(On Diff #16619)

Get rid of the #0.

75 ↗

(On Diff #16619)

Add a new line.

[RKSimon]

Closed by commit rL223349 (authored by @RKSimon).

[deeprobin]

@RKSimon

Do you have an example in which productive case this optimization is useful?
No developer will intentionally create two bswaps in a row.

Or was this PR just a nice-to-have?

/cc @qcolombet
/cc @EgorBo

[xbolva00]

In D6407#3362013, @deeprobin wrote:

@RKSimon

Do you have an example in which productive case this optimization is useful?
No developer will intentionally create two bswaps in a row.

Or was this PR just a nice-to-have?

/cc @qcolombet
/cc @EgorBo

Example code in PR:

https://bugs.llvm.org/show_bug.cgi?id=15782

8 year old patch…

No developer will …

Very strong words :D

In D6407#3362013, @deeprobin wrote:

@RKSimon

Do you have an example in which productive case this optimization is useful?
No developer will intentionally create two bswaps in a row.

Or was this PR just a nice-to-have?

/cc @qcolombet
/cc @EgorBo

[deeprobin]

@xbolva00 @RKSimon

8 year old patch…

Yes I would be interested in the real-world case behind this.

No developer will …

Very strong words :D

The question is whether this is a niche optimization or whether there is some common code pattern that justifies that this is not a niche optimization.

[lebedev.ri]

In D6407#3377769, @deeprobin wrote:

@xbolva00 @RKSimon

8 year old patch…

Yes I would be interested in the real-world case behind this.

No developer will …

Very strong words :D

The question is whether this is a niche optimization or whether there is some common code pattern that justifies that this is not a niche optimization.

Perhaps it would be more productive to approach this from another angle - what prompted you to bring this up?

[RKSimon]

Sorry for not replying (I tend to assume any comments on ancient phabs are spambots).

Developers are human too - they might not intentionally do many things, but they still crop up!

IIRC (8 years...........) as well as PR15782, this particular case turned up in some ABGR -> RGBA pixel code swizzling. I can imagine it potentially happening in network 'ntohl' style ops as well.

[deeprobin]

In D6407#3377819, @lebedev.ri wrote:

In D6407#3377769, @deeprobin wrote:

@xbolva00 @RKSimon

8 year old patch…

Yes I would be interested in the real-world case behind this.

No developer will …

Very strong words :D

The question is whether this is a niche optimization or whether there is some common code pattern that justifies that this is not a niche optimization.

Perhaps it would be more productive to approach this from another angle - what prompted you to bring this up?

I generally have different scenarios in C# (even rare ones) and came up with that the .NET runtime does not optimize duplicate bswaps.
In the runtime niche optimizations are done rather rarely, so I wanted to see for what reason LLVM has this built in.

See https://github.com/dotnet/runtime/issues/66249

[deeprobin]

In D6407#3377887, @RKSimon wrote:

Sorry for not replying (I tend to assume any comments on ancient phabs are spambots).

Developers are human too - they might not intentionally do many things, but they still crop up!

All is well :D

IIRC (8 years...........) as well as PR15782, this particular case turned up in some ABGR -> RGBA pixel code swizzling. I can imagine it potentially happening in network 'ntohl' style ops as well.

Thank you very much. I think that helps me already.

[lebedev.ri]

In D6407#3378725, @deeprobin wrote:

In D6407#3377819, @lebedev.ri wrote:

In D6407#3377769, @deeprobin wrote:

@xbolva00 @RKSimon

8 year old patch…

Yes I would be interested in the real-world case behind this.

No developer will …

Very strong words :D

The question is whether this is a niche optimization or whether there is some common code pattern that justifies that this is not a niche optimization.

Perhaps it would be more productive to approach this from another angle - what prompted you to bring this up?

I generally have different scenarios in C# (even rare ones) and came up with that the .NET runtime does not optimize duplicate bswaps.
In the runtime niche optimizations are done rather rarely, so I wanted to see for what reason LLVM has this built in.

See https://github.com/dotnet/runtime/issues/66249

That issue is about bswap(bswap(x)) --> x, this patch was for logic(bswap(x), bswap(y)) --> bswap(logic(x, y)).

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
 
   // Bail early if none of these transforms apply.
   if (N0.getNode()->getNumOperands() == 0) return SDValue();
 
   // For each of OP in AND/OR/XOR:
   // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
   // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
   // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
+  // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
   // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
   //
   // do not sink logical op inside of a vector extend, since it may combine
   // into a vsetcc.
   EVT Op0VT = N0.getOperand(0).getValueType();
   if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
        N0.getOpcode() == ISD::SIGN_EXTEND ||
+       N0.getOpcode() == ISD::BSWAP ||
        // Avoid infinite looping with PromoteIntBinOp.
        (N0.getOpcode() == ISD::ANY_EXTEND &&
         (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
        (N0.getOpcode() == ISD::TRUNCATE &&
         (!TLI.isZExtFree(VT, Op0VT) ||
          !TLI.isTruncateFree(Op0VT, VT)) &&
         TLI.isTypeLegal(Op0VT))) &&
       !VT.isVector() &&

llvm/trunk/lib/Transforms/InstCombine/InstCombine.h

private:
   /// SimplifyDemandedBits knows about.  See if the instruction has any
   /// properties that allow us to simplify its operands.
   bool SimplifyDemandedInstructionBits(Instruction &Inst);
 
   Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
                                     APInt &UndefElts, unsigned Depth = 0);
 
   Value *SimplifyVectorOp(BinaryOperator &Inst);
+  Value *SimplifyBSwap(BinaryOperator &Inst);
 
   // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
   // which has a PHI node as operand #0, see if we can fold the instruction
   // into the PHI (which is only possible if all operands to the PHI are
   // constants).
   //
   Instruction *FoldOpIntoPhi(Instruction &I);
 

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

static Value *getFCmpValue(bool isordered, unsigned code,
   case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
   case 7:
     if (!isordered) return ConstantInt::getTrue(LHS->getContext());
     Pred = FCmpInst::FCMP_ORD; break;
   }
   return Builder->CreateFCmp(Pred, LHS, RHS);
 }
 
+/// \brief Transform BITWISE_OP(BSWAP(A),BSWAP(B)) to BSWAP(BITWISE_OP(A, B))
+/// \param I Binary operator to transform.
+/// \return Pointer to node that must replace the original binary operator, or
+///         null pointer if no transformation was made.
+Value *InstCombiner::SimplifyBSwap(BinaryOperator &I) {
+  IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
+
+  // Can't do vectors.
+  if (I.getType()->isVectorTy()) return nullptr;
+
+  // Can only do bitwise ops.
+  unsigned Op = I.getOpcode();
+  if (Op != Instruction::And && Op != Instruction::Or &&
+      Op != Instruction::Xor)
+    return nullptr;
+
+  Value *OldLHS = I.getOperand(0);
+  Value *OldRHS = I.getOperand(1);
+  ConstantInt *ConstLHS = dyn_cast<ConstantInt>(OldLHS);
+  ConstantInt *ConstRHS = dyn_cast<ConstantInt>(OldRHS);
+  IntrinsicInst *IntrLHS = dyn_cast<IntrinsicInst>(OldLHS);
+  IntrinsicInst *IntrRHS = dyn_cast<IntrinsicInst>(OldRHS);
+  bool IsBswapLHS = (IntrLHS && IntrLHS->getIntrinsicID() == Intrinsic::bswap);
+  bool IsBswapRHS = (IntrRHS && IntrRHS->getIntrinsicID() == Intrinsic::bswap);
+
+  if (!IsBswapLHS && !IsBswapRHS)
+    return nullptr;
+
+  if (!IsBswapLHS && !ConstLHS)
+    return nullptr;
+
+  if (!IsBswapRHS && !ConstRHS)
+    return nullptr;
+
+  /// OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
+  /// OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
+  Value *NewLHS = IsBswapLHS ? IntrLHS->getOperand(0) :
+                  Builder->getInt(ConstLHS->getValue().byteSwap());
+
+  Value *NewRHS = IsBswapRHS ? IntrRHS->getOperand(0) :
+                  Builder->getInt(ConstRHS->getValue().byteSwap());
+
+  Value *BinOp = nullptr;
+  if (Op == Instruction::And)
+    BinOp = Builder->CreateAnd(NewLHS, NewRHS);
+  else if (Op == Instruction::Or)
+    BinOp = Builder->CreateOr(NewLHS, NewRHS);
+  else //if (Op == Instruction::Xor)
+    BinOp = Builder->CreateXor(NewLHS, NewRHS);
+
+  Module *M = I.getParent()->getParent()->getParent();
+  Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
+  return Builder->CreateCall(F, BinOp);
+}
+
 // OptAndOp - This handles expressions of the form ((val OP C1) & C2).  Where
 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'.  Op is
 // guaranteed to be a binary operator.
 Instruction *InstCombiner::OptAndOp(Instruction *Op,
                                     ConstantInt *OpRHS,
                                     ConstantInt *AndRHS,
                                     BinaryOperator &TheAnd) {
   Value *X = Op->getOperand(0);
Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
   if (Value *V = SimplifyUsingDistributiveLaws(I))
     return ReplaceInstUsesWith(I, V);
 
   // See if we can simplify any instructions used by the instruction whose sole
   // purpose is to compute bits we don't care about.
   if (SimplifyDemandedInstructionBits(I))
     return &I;
 
+  if (Value *V = SimplifyBSwap(I))
+    return ReplaceInstUsesWith(I, V);
+
   if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
     const APInt &AndRHSMask = AndRHS->getValue();
 
     // Optimize a variety of ((val OP C1) & C2) combinations...
     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
       Value *Op0LHS = Op0I->getOperand(0);
       Value *Op0RHS = Op0I->getOperand(1);
       switch (Op0I->getOpcode()) {
Instruction *InstCombiner::visitOr(BinaryOperator &I) {
   if (Value *V = SimplifyUsingDistributiveLaws(I))
     return ReplaceInstUsesWith(I, V);
 
   // See if we can simplify any instructions used by the instruction whose sole
   // purpose is to compute bits we don't care about.
   if (SimplifyDemandedInstructionBits(I))
     return &I;
 
+  if (Value *V = SimplifyBSwap(I))
+    return ReplaceInstUsesWith(I, V);
+
   if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
     ConstantInt *C1 = nullptr; Value *X = nullptr;
     // (X & C1) | C2 --> (X | C2) & (C1|C2)
     // iff (C1 & C2) == 0.
     if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
         (RHS->getValue() & C1->getValue()) != 0 &&
         Op0->hasOneUse()) {
       Value *Or = Builder->CreateOr(X, RHS);
Instruction *InstCombiner::visitXor(BinaryOperator &I) {
   if (Value *V = SimplifyUsingDistributiveLaws(I))
     return ReplaceInstUsesWith(I, V);
 
   // See if we can simplify any instructions used by the instruction whose sole
   // purpose is to compute bits we don't care about.
   if (SimplifyDemandedInstructionBits(I))
     return &I;
 
+  if (Value *V = SimplifyBSwap(I))
+    return ReplaceInstUsesWith(I, V);
+
   // Is this a ~ operation?
   if (Value *NotOp = dyn_castNotVal(&I)) {
     if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
       if (Op0I->getOpcode() == Instruction::And ||
           Op0I->getOpcode() == Instruction::Or) {
         // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
         // ~(~X | Y) === (X & ~Y) - De Morgan's Law
         if (dyn_castNotVal(Op0I->getOperand(1)))

llvm/trunk/test/Transforms/InstCombine/bswap-fold.ll

-; RUN: opt < %s -instcombine -S | not grep call.*bswap
+; RUN: opt < %s -instcombine -S | FileCheck %s
 
 define i1 @test1(i16 %tmp2) {
+; CHECK-LABEL: @test1
+; CHECK-NEXT:  %tmp = icmp eq i16 %tmp2, 256
+; CHECK-NEXT:  ret i1 %tmp
         %tmp10 = call i16 @llvm.bswap.i16( i16 %tmp2 )
         %tmp = icmp eq i16 %tmp10, 1
         ret i1 %tmp
 }
 
 define i1 @test2(i32 %tmp) {
+; CHECK-LABEL: @test2
+; CHECK-NEXT:  %tmp.upgrd.1 = icmp eq i32 %tmp, 16777216
+; CHECK-NEXT:  ret i1 %tmp.upgrd.1
         %tmp34 = tail call i32 @llvm.bswap.i32( i32 %tmp )
         %tmp.upgrd.1 = icmp eq i32 %tmp34, 1
         ret i1 %tmp.upgrd.1
 }
 
-declare i32 @llvm.bswap.i32(i32)
-
 define i1 @test3(i64 %tmp) {
+; CHECK-LABEL: @test3
+; CHECK-NEXT:  %tmp.upgrd.2 = icmp eq i64 %tmp, 72057594037927936
+; CHECK-NEXT:  ret i1 %tmp.upgrd.2
         %tmp34 = tail call i64 @llvm.bswap.i64( i64 %tmp )
         %tmp.upgrd.2 = icmp eq i64 %tmp34, 1
         ret i1 %tmp.upgrd.2
 }
 
-declare i64 @llvm.bswap.i64(i64)
-
-declare i16 @llvm.bswap.i16(i16)
-
 ; rdar://5992453
 ; A & 255
 define i32 @test4(i32 %a) nounwind  {
-entry:
-  %tmp2 = tail call i32 @llvm.bswap.i32( i32 %a )  
+; CHECK-LABEL: @test4
+; CHECK-NEXT:  %tmp2 = and i32 %a, 255
+; CHECK-NEXT:  ret i32 %tmp2
+  %tmp2 = tail call i32 @llvm.bswap.i32( i32 %a )
   %tmp4 = lshr i32 %tmp2, 24
   ret i32 %tmp4
 }
 
 ; A
-define i32 @test5(i32 %a) nounwind  {
-entry:
+define i32 @test5(i32 %a) nounwind {
+; CHECK-LABEL: @test5
+; CHECK-NEXT:  ret i32 %a
   %tmp2 = tail call i32 @llvm.bswap.i32( i32 %a )
   %tmp4 = tail call i32 @llvm.bswap.i32( i32 %tmp2 )
   ret i32 %tmp4
 }
 
 ; a >> 24
-define i32 @test6(i32 %a) nounwind  {
-entry:
-  %tmp2 = tail call i32 @llvm.bswap.i32( i32 %a )  
+define i32 @test6(i32 %a) nounwind {
+; CHECK-LABEL: @test6
+; CHECK-NEXT:  %tmp2 = lshr i32 %a, 24
+; CHECK-NEXT   ret i32 %tmp4
+  %tmp2 = tail call i32 @llvm.bswap.i32( i32 %a )
   %tmp4 = and i32 %tmp2, 255
   ret i32 %tmp4
 }
 
 ; PR5284
 define i16 @test7(i32 %A) {
-  %B = tail call i32 @llvm.bswap.i32(i32 %A) nounwind 
+; CHECK-LABEL: @test7
+; CHECK-NEXT:  %1 = lshr i32 %A, 16
+; CHECK-NEXT:  %D = trunc i32 %1 to i16
+; CHECK-NEXT   ret i16 %D
+  %B = tail call i32 @llvm.bswap.i32(i32 %A) nounwind
   %C = trunc i32 %B to i16
   %D = tail call i16 @llvm.bswap.i16(i16 %C) nounwind
   ret i16 %D
 }
 
 define i16 @test8(i64 %A) {
+; CHECK-LABEL: @test8
+; CHECK-NEXT:  %1 = lshr i64 %A, 48
+; CHECK-NEXT:  %D = trunc i64 %1 to i16
+; CHECK-NEXT   ret i16 %D
   %B = tail call i64 @llvm.bswap.i64(i64 %A) nounwind 
   %C = trunc i64 %B to i16
   %D = tail call i16 @llvm.bswap.i16(i16 %C) nounwind
   ret i16 %D
 }
 
 ; Misc: Fold bswap(undef) to undef.
 define i64 @foo() {
+; CHECK-LABEL: @foo
+; CHECK-NEXT: ret i64 undef
   %a = call i64 @llvm.bswap.i64(i64 undef)
   ret i64 %a
 }
+
+; PR15782
+; Fold: OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
+; Fold: OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
+define i16 @bs_and16i(i16 %a, i16 %b) #0 {
+; CHECK-LABEL: @bs_and16i
+; CHECK-NEXT:  %1 = and i16 %a, 4391
+; CHECK-NEXT:  %2 = call i16 @llvm.bswap.i16(i16 %1)
+; CHECK-NEXT:  ret i16 %2
+  %1 = tail call i16 @llvm.bswap.i16(i16 %a)
+  %2 = and i16 %1, 10001
+  ret i16 %2
+}
+
+define i16 @bs_and16(i16 %a, i16 %b) #0 {
+; CHECK-LABEL: @bs_and16
+; CHECK-NEXT:  %1 = and i16 %a, %b
+; CHECK-NEXT:  %2 = call i16 @llvm.bswap.i16(i16 %1)
+; CHECK-NEXT:  ret i16 %2
+  %tmp1 = tail call i16 @llvm.bswap.i16(i16 %a)
+  %tmp2 = tail call i16 @llvm.bswap.i16(i16 %b)
+  %tmp3 = and i16 %tmp1, %tmp2
+  ret i16 %tmp3
+}
+
+define i16 @bs_or16(i16 %a, i16 %b) #0 {
+; CHECK-LABEL: @bs_or16
+; CHECK-NEXT:  %1 = or i16 %a, %b
+; CHECK-NEXT:  %2 = call i16 @llvm.bswap.i16(i16 %1)
+; CHECK-NEXT:  ret i16 %2
+  %tmp1 = tail call i16 @llvm.bswap.i16(i16 %a)
+  %tmp2 = tail call i16 @llvm.bswap.i16(i16 %b)
+  %tmp3 = or i16 %tmp1, %tmp2
+  ret i16 %tmp3
+}
+
+define i16 @bs_xor16(i16 %a, i16 %b) #0 {
+; CHECK-LABEL: @bs_xor16
+; CHECK-NEXT:  %1 = xor i16 %a, %b
+; CHECK-NEXT:  %2 = call i16 @llvm.bswap.i16(i16 %1)
+; CHECK-NEXT:  ret i16 %2
+  %tmp1 = tail call i16 @llvm.bswap.i16(i16 %a)
+  %tmp2 = tail call i16 @llvm.bswap.i16(i16 %b)
+  %tmp3 = xor i16 %tmp1, %tmp2
+  ret i16 %tmp3
+}
+
+define i32 @bs_and32i(i32 %a, i32 %b) #0 {
+; CHECK-LABEL: @bs_and32i
+; CHECK-NEXT:  %1 = and i32 %a, -1585053440
+; CHECK-NEXT:  %2 = call i32 @llvm.bswap.i32(i32 %1)
+; CHECK-NEXT:  ret i32 %2
+  %tmp1 = tail call i32 @llvm.bswap.i32(i32 %a)
+  %tmp2 = and i32 %tmp1, 100001
+  ret i32 %tmp2
+}
+
+define i32 @bs_and32(i32 %a, i32 %b) #0 {
+; CHECK-LABEL: @bs_and32
+; CHECK-NEXT:  %1 = and i32 %a, %b
+; CHECK-NEXT:  %2 = call i32 @llvm.bswap.i32(i32 %1)
+; CHECK-NEXT:  ret i32 %2
+  %tmp1 = tail call i32 @llvm.bswap.i32(i32 %a)
+  %tmp2 = tail call i32 @llvm.bswap.i32(i32 %b)
+  %tmp3 = and i32 %tmp1, %tmp2
+  ret i32 %tmp3
+}
+
+define i32 @bs_or32(i32 %a, i32 %b) #0 {
+; CHECK-LABEL: @bs_or32
+; CHECK-NEXT:  %1 = or i32 %a, %b
+; CHECK-NEXT:  %2 = call i32 @llvm.bswap.i32(i32 %1)
+; CHECK-NEXT:  ret i32 %2
+  %tmp1 = tail call i32 @llvm.bswap.i32(i32 %a)
+  %tmp2 = tail call i32 @llvm.bswap.i32(i32 %b)
+  %tmp3 = or i32 %tmp1, %tmp2
+  ret i32 %tmp3
+}
+
+define i32 @bs_xor32(i32 %a, i32 %b) #0 {
+; CHECK-LABEL: @bs_xor32
+; CHECK-NEXT:  %1 = xor i32 %a, %b
+; CHECK-NEXT:  %2 = call i32 @llvm.bswap.i32(i32 %1)
+; CHECK-NEXT:  ret i32 %2
+  %tmp1 = tail call i32 @llvm.bswap.i32(i32 %a)
+  %tmp2 = tail call i32 @llvm.bswap.i32(i32 %b)
+  %tmp3 = xor i32 %tmp1, %tmp2
+  ret i32 %tmp3
+}
+
+define i64 @bs_and64i(i64 %a, i64 %b) #0 {
+; CHECK-LABEL: @bs_and64i
+; CHECK-NEXT:  %1 = and i64 %a, 129085117527228416
+; CHECK-NEXT:  %2 = call i64 @llvm.bswap.i64(i64 %1)
+; CHECK-NEXT:  ret i64 %2
+  %tmp1 = tail call i64 @llvm.bswap.i64(i64 %a)
+  %tmp2 = and i64 %tmp1, 1000000001
+  ret i64 %tmp2
+}
+
+define i64 @bs_and64(i64 %a, i64 %b) #0 {
+; CHECK-LABEL: @bs_and64
+; CHECK-NEXT:  %1 = and i64 %a, %b
+; CHECK-NEXT:  %2 = call i64 @llvm.bswap.i64(i64 %1)
+; CHECK-NEXT:  ret i64 %2
+  %tmp1 = tail call i64 @llvm.bswap.i64(i64 %a)
+  %tmp2 = tail call i64 @llvm.bswap.i64(i64 %b)
+  %tmp3 = and i64 %tmp1, %tmp2
+  ret i64 %tmp3
+}
+
+define i64 @bs_or64(i64 %a, i64 %b) #0 {
+; CHECK-LABEL: @bs_or64
+; CHECK-NEXT:  %1 = or i64 %a, %b
+; CHECK-NEXT:  %2 = call i64 @llvm.bswap.i64(i64 %1)
+; CHECK-NEXT:  ret i64 %2
+  %tmp1 = tail call i64 @llvm.bswap.i64(i64 %a)
+  %tmp2 = tail call i64 @llvm.bswap.i64(i64 %b)
+  %tmp3 = or i64 %tmp1, %tmp2
+  ret i64 %tmp3
+}
+
+define i64 @bs_xor64(i64 %a, i64 %b) #0 {
+; CHECK-LABEL: @bs_xor64
+; CHECK-NEXT:  %1 = xor i64 %a, %b
+; CHECK-NEXT:  %2 = call i64 @llvm.bswap.i64(i64 %1)
+; CHECK-NEXT:  ret i64 %2
+  %tmp1 = tail call i64 @llvm.bswap.i64(i64 %a)
+  %tmp2 = tail call i64 @llvm.bswap.i64(i64 %b)
+  %tmp3 = xor i64 %tmp1, %tmp2
+  ret i64 %tmp3
+}
+
+declare i16 @llvm.bswap.i16(i16)
+declare i32 @llvm.bswap.i32(i32)
+declare i64 @llvm.bswap.i64(i64)