Strength reduce intrinsics with overflow into regular arithmetic operations if possible.
Some intrinsics, like s/uadd.with.overflow and umul.with.overflow, are already strength reduced.
This change adds other arithmetic intrinsics: s/usub.with.overflow, smul.with.overflow.
It completes the work on PR20194.
llvm-svn: 224417
diff --git a/llvm/lib/Transforms/InstCombine/InstCombine.h b/llvm/lib/Transforms/InstCombine/InstCombine.h
index 755de87..326bf8f 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombine.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombine.h
@@ -285,6 +285,7 @@
bool WillNotOverflowUnsignedAdd(Value *LHS, Value *RHS, Instruction *CxtI);
bool WillNotOverflowSignedSub(Value *LHS, Value *RHS, Instruction *CxtI);
bool WillNotOverflowUnsignedSub(Value *LHS, Value *RHS, Instruction *CxtI);
+ bool WillNotOverflowSignedMul(Value *LHS, Value *RHS, Instruction *CxtI);
Value *EmitGEPOffset(User *GEP);
Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 902b640..37ae797 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1008,6 +1008,51 @@
return false;
}
+/// \brief Return true if we can prove that:
+/// (mul LHS, RHS) === (mul nsw LHS, RHS)
+bool InstCombiner::WillNotOverflowSignedMul(Value *LHS, Value *RHS,
+ Instruction *CxtI) {
+ if (IntegerType *IT = dyn_cast<IntegerType>(LHS->getType())) {
+
+ // Multiplying n * m significant bits yields a result of n + m significant
+ // bits. If the total number of significant bits does not exceed the
+ // result bit width (minus 1), there is no overflow.
+ // This means if we have enough leading sign bits in the operands
+ // we can guarantee that the result does not overflow.
+ // Ref: "Hacker's Delight" by Henry Warren
+ unsigned BitWidth = IT->getBitWidth();
+
+ // Note that underestimating the number of sign bits gives a more
+ // conservative answer.
+ unsigned SignBits = ComputeNumSignBits(LHS, 0, CxtI) +
+ ComputeNumSignBits(RHS, 0, CxtI);
+
+ // First handle the easy case: if we have enough sign bits there's
+ // definitely no overflow.
+ if (SignBits > BitWidth + 1)
+ return true;
+
+ // There are two ambiguous cases where there can be no overflow:
+ // SignBits == BitWidth + 1 and
+ // SignBits == BitWidth
+ // The second case is difficult to check, therefore we only handle the
+ // first case.
+ if (SignBits == BitWidth + 1) {
+ // It overflows only when both arguments are negative and the true
+ // product is exactly the minimum negative number.
+ // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
+ // For simplicity we just check if at least one side is not negative.
+ bool LHSNonNegative, LHSNegative;
+ bool RHSNonNegative, RHSNegative;
+ ComputeSignBit(LHS, LHSNonNegative, LHSNegative, DL, 0, AT, CxtI, DT);
+ ComputeSignBit(RHS, RHSNonNegative, RHSNegative, DL, 0, AT, CxtI, DT);
+ if (LHSNonNegative || RHSNonNegative)
+ return true;
+ }
+ }
+ return false;
+}
+
// Checks if any operand is negative and we can convert add to sub.
// This function checks for following negative patterns
// ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 33af7ee..b214b55 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -427,6 +427,15 @@
return CreateOverflowTuple(II, LHS, false, /*ReUseName*/false);
}
}
+ if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) {
+ if (WillNotOverflowSignedSub(LHS, RHS, II)) {
+ return CreateOverflowTuple(II, Builder->CreateNSWSub(LHS, RHS), false);
+ }
+ } else {
+ if (WillNotOverflowUnsignedSub(LHS, RHS, II)) {
+ return CreateOverflowTuple(II, Builder->CreateNUWSub(LHS, RHS), false);
+ }
+ }
break;
}
case Intrinsic::umul_with_overflow: {
@@ -477,6 +486,12 @@
/*ReUseName*/false);
}
}
+ if (II->getIntrinsicID() == Intrinsic::smul_with_overflow) {
+ Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
+ if (WillNotOverflowSignedMul(LHS, RHS, II)) {
+ return CreateOverflowTuple(II, Builder->CreateNSWMul(LHS, RHS), false);
+ }
+ }
break;
case Intrinsic::minnum:
case Intrinsic::maxnum: {