[llvm-commits] [llvm] r169182 - in /llvm/trunk: lib/Transforms/InstCombine/InstCombine.h lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp test/Transforms/InstCombine/2010-11-01-lshr-mask.ll test/Transforms/InstCombine/shift.ll test/Transforms/InstCombine/signext.ll

[Eli Friedman]

On Mon, Dec 10, 2012 at 3:39 PM, Shuxin Yang <shuxin.llvm at gmail.com> wrote:
> Hi, Eli:
>
>   Thank you for the feedback. See the interleaving text.
>
>
> On 12/10/12 2:38 PM, Eli Friedman wrote:
>>
>> On Mon, Dec 3, 2012 at 4:04 PM, Shuxin Yang <shuxin.llvm at gmail.com> wrote:
>>>
>>> +  Value *VarX = Shr->getOperand(0);
>>> +  Type *Ty = VarX->getType();
>>> +
>>> +  APInt BitMask1(Ty->getIntegerBitWidth(), (uint64_t)-1);
>>> +  APInt BitMask2(Ty->getIntegerBitWidth(), (uint64_t)-1);
>>
>> APInt::getAllOnesValue().  (As-is, this isn't right for bit-widths >= 64.)
>
> You aright. APInt::getAllOnesValue() was the function I were looking for.
> I didn't find one. So, I try to use following function as alternative. But
> somehow
> I misread the default value of 3rd parameter. I will fix the problem and
> test it
> tomorrow.
>
> APInt BitMask1(Ty->getIntegerBitWidth(), (uint64_t)-1, *TRUE*);

OK.

>>
>>> +    BinaryOperator *New;
>>> +    if (ShrAmt < ShlAmt) {
>>> +      Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt -
>>> ShrAmt);
>>> +      New = BinaryOperator::CreateShl(VarX, Amt);
>>> +      BinaryOperator *Orig = cast<BinaryOperator>(Shl);
>>> +      New->setHasNoSignedWrap(Orig->hasNoSignedWrap());
>>> +      New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap());
>>> +    } else {
>>> +      Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt -
>>> ShlAmt);
>>> +      New = BinaryOperator::CreateLShr(VarX, Amt);
>>> +    }
>>
>> Can we propagate the "exact" bit from the original shr?
>
> I'm not sure this question. Could you please elaborate?

If the shift is e.g. "lshr exact i32 %x, 3", can we make the new shift
"exact" as well?  (See
http://llvm.org/docs/LangRef.html#lshr-instruction .)

-Eli