By the way. This check is still not quite right.
// Loads from stubs of global addresses are side effect free.
if (Reg != 0 && MRegisterInfo::isVirtualRegister(Reg) &&
In dynamic-no-pic mode, Reg can be 0. So it should be Reg == 0 || ...
Evan
On Jan 5, 2008, at 3:37 PM, Bill Wendling wrote:
> On Jan 5, 2008, at 2:49 AM, Evan Cheng wrote:
>> On Jan 5, 2008, at 1:18 AM, Bill Wendling <[EMAIL PROTECTED]> wrote:
>>
>>> --- llvm/trunk/lib/Target/X86/X86InstrInfo.cpp (original)
>>> +++ llvm/trunk/lib/Target/X86/X86InstrInfo.cpp Sat Jan 5 03:18:04
>>> 2008
>>> @@ -171,12 +171,15 @@
>>> case X86::MOV32rm:
>>> if (MI->getOperand(1).isRegister()) {
>>> unsigned Reg = MI->getOperand(1).getReg();
>>> + const X86Subtarget &ST = TM.getSubtarget<X86Subtarget>();
>>>
>>> // Loads from global addresses which aren't redefined in the
>>> function are
>>> // side effect free.
>>> if (Reg != 0 &&
>>
>> I assume caller ensure this operant is side effect free?
>
> It ensures that all of the operands are loop-invariant. Reg is
> defined outside of the loop at this point.
>
>> Can you change this to accept a list of operands which should be
>> checked? For
>> LICM, that means non-invariant operands.
>>
> I'm not sure why this is necessary. Like I said, all of the operands
> should be loop-invariant at this point...
>
>> Also, please fix the comment.
>>
> Okay.
>
> -bw
> _______________________________________________
> llvm-commits mailing list
> llvm-commits@cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
_______________________________________________
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
