Reinstate this optimization to fold icmp of xor when possible. Don't try to
turn icmp eq a+x, b+x into icmp eq a, b if a+x or b+x has other uses. This
may have been increasing register pressure leading to the bzip2 slowdown.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63487 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp
index 8b9c599..a1cb3dc 100644
--- a/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -6095,18 +6095,40 @@
if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
- Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
- I.isEquality()) {
+ Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
switch (Op0I->getOpcode()) {
default: break;
case Instruction::Add:
case Instruction::Sub:
case Instruction::Xor:
- // a+x icmp eq/ne b+x --> a icmp b
- return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
- Op1I->getOperand(0));
+ if (I.isEquality()) {
+ // a+x icmp eq/ne b+x --> a icmp b
+ return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
+ Op1I->getOperand(0));
+ } else {
+ // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
+ if (CI->getValue().isSignBit()) {
+ ICmpInst::Predicate Pred = I.isSignedPredicate()
+ ? I.getUnsignedPredicate()
+ : I.getSignedPredicate();
+ return new ICmpInst(Pred, Op0I->getOperand(0),
+ Op1I->getOperand(0));
+ } else if ((~CI->getValue()).isSignBit()) {
+ ICmpInst::Predicate Pred = I.isSignedPredicate()
+ ? I.getUnsignedPredicate()
+ : I.getSignedPredicate();
+ Pred = I.getSwappedPredicate(Pred);
+ return new ICmpInst(Pred, Op0I->getOperand(0),
+ Op1I->getOperand(0));
+ }
+ }
+ }
break;
case Instruction::Mul:
+ if (!I.isEquality())
+ break;
+
if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
// a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
// Mask = -1 >> count-trailing-zeros(Cst).
@@ -6425,6 +6447,29 @@
else
return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
}
+
+ if (LHSI->hasOneUse()) {
+ // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
+ if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
+ const APInt &SignBit = XorCST->getValue();
+ ICmpInst::Predicate Pred = ICI.isSignedPredicate()
+ ? ICI.getUnsignedPredicate()
+ : ICI.getSignedPredicate();
+ return new ICmpInst(Pred, LHSI->getOperand(0),
+ ConstantInt::get(RHSV ^ SignBit));
+ }
+
+ // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
+ if (!ICI.isEquality() && (~XorCST->getValue()).isSignBit()) {
+ const APInt &NotSignBit = XorCST->getValue();
+ ICmpInst::Predicate Pred = ICI.isSignedPredicate()
+ ? ICI.getUnsignedPredicate()
+ : ICI.getSignedPredicate();
+ Pred = ICI.getSwappedPredicate(Pred);
+ return new ICmpInst(Pred, LHSI->getOperand(0),
+ ConstantInt::get(RHSV ^ NotSignBit));
+ }
+ }
}
break;
case Instruction::And: // (icmp pred (and X, AndCST), RHS)