[llvm-commits] CVS: llvm/lib/Target/X86/X86InstrFPStack.td X86InstrInfo.cpp X86InstrInfo.h X86InstrInfo.td X86InstrMMX.td X86InstrSSE.td

[Evan Cheng]

On Jun 19, 2007, at 10:31 AM, Chris Lattner wrote:

>>> It seems that simple remat works for things that:
>>>
>>> 1. have no side effects
>>> 2. have no register inputs
>>>
>>> Is there another condition?
>>
>> Can't think of any.
>>
>>>
>>> If so, simple remat can check #2 today, just by itself looking at  
>>> the
>>> machineinstr.  Given that, it comes down to how we want to
>>> represent #1.
>>
>> I don't see a better way so I guess this will be a targetinstrinfo
>> bit (true for those with side-effects).
>
> Okay, the tricky thing here is instructions that have "conditional
> side effects".  For example, all instructions marked isload/isstore/
> iscall etc should be considered to have side effects (as would
> anything with implicit definitions), but loads from constant pools
> and other special cases should not be considered to have side effects.

Calls should definitely be marked to have side-effects. To me the  
tricky cases are loads / stores. But can't we determine these from  
the operands? So, add one more condition:
3. if memory operation, no external symbol or global address inputs.

Then we don't need to mark loads / stores as having side effects.

Evan

>
> This sounds like the exact same class of problems we already have
> with trivial remat :)
>
> -Chris
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits