Revert "Revert "InstCombine: Improvement to check if signed addition overflows.""

This reverts commit r209762, bringing back r209746. It was not responsible for the libc++ build failure

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@209776 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Rafael Espindola 2014-05-28 21:43:52 +00:00
parent 836475b2e2
commit c81cf72ef3
2 changed files with 99 additions and 5 deletions

View File

@ -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<IntegerType>(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;
}

View File

@ -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
}