diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp index 661b0475cc5..9214b6bbd11 100644 --- a/lib/Transforms/Scalar/InstructionCombining.cpp +++ b/lib/Transforms/Scalar/InstructionCombining.cpp @@ -2882,13 +2882,21 @@ Instruction *InstCombiner::visitUDiv(BinaryOperator &I) { if (Instruction *Common = commonIDivTransforms(I)) return Common; - // X udiv C^2 -> X >> C - // Check to see if this is an unsigned division with an exact power of 2, - // if so, convert to a right shift. if (ConstantInt *C = dyn_cast(Op1)) { + // X udiv C^2 -> X >> C + // Check to see if this is an unsigned division with an exact power of 2, + // if so, convert to a right shift. if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2 return BinaryOperator::CreateLShr(Op0, ConstantInt::get(Op0->getType(), C->getValue().logBase2())); + + // X udiv C, where C >= signbit + if (C->getValue().isNegative()) { + Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C), + I); + return SelectInst::Create(IC, Constant::getNullValue(I.getType()), + ConstantInt::get(I.getType(), 1)); + } } // X udiv (C1 << N), where C1 is "1< X >> (N+C2) diff --git a/test/Transforms/InstCombine/2008-11-27-UDivNegative.ll b/test/Transforms/InstCombine/2008-11-27-UDivNegative.ll new file mode 100644 index 00000000000..6a6b5f38634 --- /dev/null +++ b/test/Transforms/InstCombine/2008-11-27-UDivNegative.ll @@ -0,0 +1,6 @@ +; RUN: llvm-as < %s | opt -instcombine | llvm-dis | not grep div + +define i8 @test(i8 %x) readnone nounwind { + %A = udiv i8 %x, 250 + ret i8 %A +} diff --git a/test/Transforms/InstCombine/udiv-simplify-bug-1.ll b/test/Transforms/InstCombine/udiv-simplify-bug-1.ll index 6e76bcfd753..0036760ba50 100644 --- a/test/Transforms/InstCombine/udiv-simplify-bug-1.ll +++ b/test/Transforms/InstCombine/udiv-simplify-bug-1.ll @@ -1,16 +1,11 @@ ; RUN: llvm-as < %s | opt -instcombine | llvm-dis > %t1.ll -; RUN: grep udiv %t1.ll | count 3 -; RUN: grep zext %t1.ll | count 3 +; RUN: grep udiv %t1.ll | count 2 +; RUN: grep zext %t1.ll | count 2 ; PR2274 ; The udiv instructions shouldn't be optimized away, and the ; sext instructions should be optimized to zext. -define i64 @foo(i32 %x) nounwind { - %r = udiv i32 %x, -1 - %z = sext i32 %r to i64 - ret i64 %z -} define i64 @bar(i32 %x) nounwind { %y = lshr i32 %x, 30 %r = udiv i32 %y, 3