From e8075c6877d1f2e1be6c2646135912406616249c Mon Sep 17 00:00:00 2001 From: Rafael Espindola Date: Wed, 28 May 2014 15:30:40 +0000 Subject: [PATCH] InstCombine: Improvement to check if signed addition overflows. This patch implements two things: 1. If we know one number is positive and another is negative, we return true as signed addition of two opposite signed numbers will never overflow. 2. Implemented TODO : If one of the operands only has one non-zero bit, and if the other operand has a known-zero bit in a more significant place than it (not including the sign bit) the ripple may go up to and fill the zero, but won't change the sign. e.x - (x & ~4) + 1 We make sure that we are ignoring 0 at MSB. Patch by Suyog Sarda. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@209746 91177308-0d34-0410-b5e6-96231b3b80d8 --- .../InstCombine/InstCombineAddSub.cpp | 48 ++++++++++++++-- test/Transforms/InstCombine/AddOverflow.ll | 56 +++++++++++++++++++ 2 files changed, 99 insertions(+), 5 deletions(-) create mode 100644 test/Transforms/InstCombine/AddOverflow.ll diff --git a/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/lib/Transforms/InstCombine/InstCombineAddSub.cpp index c37a9cf2ef9..eca4e4a7870 100644 --- a/lib/Transforms/InstCombine/InstCombineAddSub.cpp +++ b/lib/Transforms/InstCombine/InstCombineAddSub.cpp @@ -889,11 +889,34 @@ static inline Value *dyn_castFoldableMul(Value *V, Constant *&CST) { return nullptr; } +// If one of the operands only has one non-zero bit, and if the other +// operand has a known-zero bit in a more significant place than it (not +// including the sign bit) the ripple may go up to and fill the zero, but +// won't change the sign. For example, (X & ~4) + 1. +// FIXME: Handle case where LHS has a zero before the 1 in the RHS, but also +// has one after. +static bool CheckRippleForAdd(APInt Op0KnownZero, APInt Op0KnownOne, + APInt Op1KnownZero, APInt Op1KnownOne) { + // Make sure that one of the operand has only one bit set to 1 and all other + // bit set to 0. + if ((~Op1KnownZero).countPopulation() == 1) { + int BitWidth = Op0KnownZero.getBitWidth(); + // Ignore Sign Bit. + Op0KnownZero.clearBit(BitWidth - 1); + int Op1OnePosition = BitWidth - Op1KnownOne.countLeadingZeros() - 1; + int Op0ZeroPosition = BitWidth - Op0KnownZero.countLeadingZeros() - 1; + if ((Op0ZeroPosition != (BitWidth - 1)) && + (Op0ZeroPosition >= Op1OnePosition)) + return true; + } + return false; +} /// WillNotOverflowSignedAdd - Return true if we can prove that: /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS)) /// This basically requires proving that the add in the original type would not /// overflow to change the sign bit or have a carry out. +/// TODO: Handle this for Vectors. bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) { // There are different heuristics we can use for this. Here are some simple // ones. @@ -905,14 +928,29 @@ bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) { if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1) return true; + if (IntegerType *IT = dyn_cast(LHS->getType())) { - // If one of the operands only has one non-zero bit, and if the other operand - // has a known-zero bit in a more significant place than it (not including the - // sign bit) the ripple may go up to and fill the zero, but won't change the - // sign. For example, (X & ~4) + 1. + int BitWidth = IT->getBitWidth(); + APInt LHSKnownZero(BitWidth, 0, /*isSigned*/ true); + APInt LHSKnownOne(BitWidth, 0, /*isSigned*/ true); + computeKnownBits(LHS, LHSKnownZero, LHSKnownOne); - // TODO: Implement. + APInt RHSKnownZero(BitWidth, 0, /*isSigned*/ true); + APInt RHSKnownOne(BitWidth, 0, /*isSigned*/ true); + computeKnownBits(RHS, RHSKnownZero, RHSKnownOne); + // Addition of two 2's compliment numbers having opposite signs will never + // overflow. + if ((LHSKnownOne[BitWidth - 1] && RHSKnownZero[BitWidth - 1]) || + (LHSKnownZero[BitWidth - 1] && RHSKnownOne[BitWidth - 1])) + return true; + + // Check if carry bit of addition will not cause overflow. + if (CheckRippleForAdd(LHSKnownZero, LHSKnownOne, RHSKnownZero, RHSKnownOne)) + return true; + if (CheckRippleForAdd(RHSKnownZero, RHSKnownOne, LHSKnownZero, LHSKnownOne)) + return true; + } return false; } diff --git a/test/Transforms/InstCombine/AddOverflow.ll b/test/Transforms/InstCombine/AddOverflow.ll new file mode 100644 index 00000000000..1bbd1fc59a5 --- /dev/null +++ b/test/Transforms/InstCombine/AddOverflow.ll @@ -0,0 +1,56 @@ +; RUN: opt < %s -instcombine -S | FileCheck %s + +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" +target triple = "x86_64-unknown-linux-gnu" + +; CHECK-LABEL: @ripple( +; CHECK: add nsw i16 %tmp1, 1 +define i32 @ripple(i16 signext %x) { +bb: + %tmp = sext i16 %x to i32 + %tmp1 = and i32 %tmp, -5 + %tmp2 = trunc i32 %tmp1 to i16 + %tmp3 = sext i16 %tmp2 to i32 + %tmp4 = add i32 %tmp3, 1 + ret i32 %tmp4 +} + +; CHECK-LABEL: @ripplenot( +; CHECK: add i32 %tmp3, 4 +define i32 @ripplenot(i16 signext %x) { +bb: + %tmp = sext i16 %x to i32 + %tmp1 = and i32 %tmp, -3 + %tmp2 = trunc i32 %tmp1 to i16 + %tmp3 = sext i16 %tmp2 to i32 + %tmp4 = add i32 %tmp3, 4 + ret i32 %tmp4 +} + +; CHECK-LABEL: @oppositesign( +; CHECK: add nsw i16 %tmp1, 4 +define i32 @oppositesign(i16 signext %x) { +bb: + %tmp = sext i16 %x to i32 + %tmp1 = or i32 %tmp, 32768 + %tmp2 = trunc i32 %tmp1 to i16 + %tmp3 = sext i16 %tmp2 to i32 + %tmp4 = add i32 %tmp3, 4 + ret i32 %tmp4 +} + +; CHECK-LABEL: @ripplenot_var( +; CHECK: add i32 %tmp6, %tmp7 +define i32 @ripplenot_var(i16 signext %x, i16 signext %y) { +bb: + %tmp = sext i16 %x to i32 + %tmp1 = and i32 %tmp, -5 + %tmp2 = trunc i32 %tmp1 to i16 + %tmp3 = sext i16 %y to i32 + %tmp4 = or i32 %tmp3, 2 + %tmp5 = trunc i32 %tmp4 to i16 + %tmp6 = sext i16 %tmp5 to i32 + %tmp7 = sext i16 %tmp2 to i32 + %tmp8 = add i32 %tmp6, %tmp7 + ret i32 %tmp8 +}