Index: lib/Transforms/InstCombine/InstCombineMulDivRem.cpp =================================================================== --- lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -1150,6 +1150,23 @@ Value *ShAmt = ConstantInt::get(Op1->getType(), Op1C->exactLogBase2()); return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName()); } + + // If the dividend is sign-extended and the constant divisor is small enough + // to fit in the source type, shrink the division to the narrower type: + // (sext X) sdiv C --> sext (X sdiv C) + Value *Op0Src; + if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) && + Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) { + + // In the general case, we need to make sure that the dividend is not the + // minimum signed value because dividing that by -1 is UB. But here, we + // know that the -1 divisor case is already handled above. + + Constant *NarrowDivisor = + ConstantExpr::getTrunc(cast(Op1), Op0Src->getType()); + Value *NarrowOp = Builder->CreateSDiv(Op0Src, NarrowDivisor); + return new SExtInst(NarrowOp, Op0->getType()); + } } if (Constant *RHS = dyn_cast(Op1)) { Index: test/Transforms/InstCombine/div.ll =================================================================== --- test/Transforms/InstCombine/div.ll +++ test/Transforms/InstCombine/div.ll @@ -409,13 +409,12 @@ ret i32 %div } -; FIXME: ; We can perform the division in the smaller type. define i32 @shrink(i8 %x) { ; CHECK-LABEL: @shrink( -; CHECK-NEXT: [[CONV:%.*]] = sext i8 %x to i32 -; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[CONV]], 127 +; CHECK-NEXT: [[TMP1:%.*]] = sdiv i8 %x, 127 +; CHECK-NEXT: [[DIV:%.*]] = sext i8 [[TMP1]] to i32 ; CHECK-NEXT: ret i32 [[DIV]] ; %conv = sext i8 %x to i32 @@ -427,8 +426,8 @@ define i32 @zap(i8 %x) { ; CHECK-LABEL: @zap( -; CHECK-NEXT: [[CONV:%.*]] = sext i8 %x to i32 -; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[CONV]], -128 +; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i8 %x, -128 +; CHECK-NEXT: [[DIV:%.*]] = zext i1 [[TMP1]] to i32 ; CHECK-NEXT: ret i32 [[DIV]] ; %conv = sext i8 %x to i32 @@ -440,8 +439,8 @@ define <3 x i32> @shrink_vec(<3 x i8> %x) { ; CHECK-LABEL: @shrink_vec( -; CHECK-NEXT: [[CONV:%.*]] = sext <3 x i8> %x to <3 x i32> -; CHECK-NEXT: [[DIV:%.*]] = sdiv <3 x i32> [[CONV]], +; CHECK-NEXT: [[TMP1:%.*]] = sdiv <3 x i8> %x, +; CHECK-NEXT: [[DIV:%.*]] = sext <3 x i8> [[TMP1]] to <3 x i32> ; CHECK-NEXT: ret <3 x i32> [[DIV]] ; %conv = sext <3 x i8> %x to <3 x i32> @@ -451,8 +450,8 @@ define <2 x i32> @zap_vec(<2 x i8> %x) { ; CHECK-LABEL: @zap_vec( -; CHECK-NEXT: [[CONV:%.*]] = sext <2 x i8> %x to <2 x i32> -; CHECK-NEXT: [[DIV:%.*]] = sdiv <2 x i32> [[CONV]], +; CHECK-NEXT: [[TMP1:%.*]] = icmp eq <2 x i8> %x, +; CHECK-NEXT: [[DIV:%.*]] = zext <2 x i1> [[TMP1]] to <2 x i32> ; CHECK-NEXT: ret <2 x i32> [[DIV]] ; %conv = sext <2 x i8> %x to <2 x i32>