diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp index e018b351082..f25dd3582b2 100644 --- a/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -955,6 +955,19 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { ShAmt); } + // If this input is a trunc from our destination, then turn sext(trunc(x)) + // into shifts. + if (TruncInst *TI = dyn_cast(Src)) + if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { + uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); + uint32_t DestBitSize = DestTy->getScalarSizeInBits(); + + // We need to emit a shl + ashr to do the sign extend. + Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); + Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); + return BinaryOperator::CreateAShr(Res, ShAmt); + } + // If the input is a shl/ashr pair of a same constant, then this is a sign // extension from a smaller value. If we could trust arbitrary bitwidth // integers, we could turn this into a truncate to the smaller bit and then diff --git a/lib/Transforms/InstCombine/InstCombineShifts.cpp b/lib/Transforms/InstCombine/InstCombineShifts.cpp index fe91da1b6af..321c91d57d2 100644 --- a/lib/Transforms/InstCombine/InstCombineShifts.cpp +++ b/lib/Transforms/InstCombine/InstCombineShifts.cpp @@ -404,12 +404,26 @@ Instruction *InstCombiner::visitAShr(BinaryOperator &I) { if (ConstantInt *Op1C = dyn_cast(Op1)) { // If the input is a SHL by the same constant (ashr (shl X, C), C), then we - // have a sign-extend idiom. If the input value is known to already be sign - // extended enough, delete the extension. + // have a sign-extend idiom. Value *X; - if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) && - ComputeNumSignBits(X) > Op1C->getZExtValue()) - return ReplaceInstUsesWith(I, X); + if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1)))) { + // If the input value is known to already be sign extended enough, delete + // the extension. + if (ComputeNumSignBits(X) > Op1C->getZExtValue()) + return ReplaceInstUsesWith(I, X); + + // If the input is an extension from the shifted amount value, e.g. + // %x = zext i8 %A to i32 + // %y = shl i32 %x, 24 + // %z = ashr %y, 24 + // then turn this into "z = sext i8 A to i32". + if (ZExtInst *ZI = dyn_cast(X)) { + uint32_t SrcBits = ZI->getOperand(0)->getType()->getScalarSizeInBits(); + uint32_t DestBits = ZI->getType()->getScalarSizeInBits(); + if (Op1C->getZExtValue() == DestBits-SrcBits) + return new SExtInst(ZI->getOperand(0), ZI->getType()); + } + } } // See if we can turn a signed shr into an unsigned shr. diff --git a/test/Transforms/InstCombine/signext.ll b/test/Transforms/InstCombine/signext.ll index 098dde109f8..1c52b62df4e 100644 --- a/test/Transforms/InstCombine/signext.ll +++ b/test/Transforms/InstCombine/signext.ll @@ -8,8 +8,8 @@ define i32 @test1(i32 %x) { %tmp.3 = add i32 %tmp.2, 32768 ; [#uses=1] ret i32 %tmp.3 ; CHECK: @test1 -; CHECK: %sext = trunc i32 %x to i16 -; CHECK: %tmp.3 = sext i16 %sext to i32 +; CHECK: %sext1 = shl i32 %x, 16 +; CHECK: %tmp.3 = ashr i32 %sext1, 16 ; CHECK: ret i32 %tmp.3 } @@ -19,8 +19,8 @@ define i32 @test2(i32 %x) { %tmp.3 = add i32 %tmp.2, -32768 ; [#uses=1] ret i32 %tmp.3 ; CHECK: @test2 -; CHECK: %sext = trunc i32 %x to i16 -; CHECK: %tmp.3 = sext i16 %sext to i32 +; CHECK: %sext1 = shl i32 %x, 16 +; CHECK: %tmp.3 = ashr i32 %sext1, 16 ; CHECK: ret i32 %tmp.3 } @@ -50,8 +50,8 @@ define i32 @test5(i32 %x) { %tmp.3 = add i32 %tmp.2, -128 ; [#uses=1] ret i32 %tmp.3 ; CHECK: @test5 -; CHECK: %sext = trunc i32 %x to i8 -; CHECK: %tmp.3 = sext i8 %sext to i32 +; CHECK: %sext1 = shl i32 %x, 24 +; CHECK: %tmp.3 = ashr i32 %sext1, 24 ; CHECK: ret i32 %tmp.3 } @@ -65,3 +65,12 @@ define i32 @test6(i32 %x) { ; CHECK: ret i32 %tmp.4 } +define i32 @test7(i16 %P) { + %tmp.1 = zext i16 %P to i32 ; [#uses=1] + %sext1 = shl i32 %tmp.1, 16 ; [#uses=1] + %tmp.5 = ashr i32 %sext1, 16 ; [#uses=1] + ret i32 %tmp.5 +; CHECK: @test7 +; CHECK: %tmp.5 = sext i16 %P to i32 +; CHECK: ret i32 %tmp.5 +}