llvm dev - Jan 2008 - [LLVMdev] alias analysis during code generation

Hello,

Is it somehow possible to use alias analyser data during code
generation, after instruction selection? I would need to run
the AA using MachineOperands, not Values. Is this in any way
in LLVM possible at the moment?

We have a post-pass instruction scheduler which would benefit
from AA results, but the AA interface uses Values.


-- 
Pekka

Pekka Jääskeläinen schrieb:
> Is it somehow possible to use alias analyser data during code
> generation, after instruction selection? I would need to run
> the AA using MachineOperands, not Values. Is this in any way
> in LLVM possible at the moment?
i've been working on this (also because of a post-pass scheduler).

i managed to hand the source values (values passed to the AA) over to
the machine instruction level. unfortunately the "codegen prepare"
pass
rewrites GEP llvm instructions and confuses the AA. some work needs to
be done inside the AA to get reasonable results.

i did not have time to prepare and submit a patch for llvm.

florian

On Jan 17, 2008, at 2:07 AM, Florian Brandner wrote:
> Pekka Jääskeläinen schrieb:
>> Is it somehow possible to use alias analyser data during code
>> generation, after instruction selection? I would need to run
>> the AA using MachineOperands, not Values. Is this in any way
>> in LLVM possible at the moment?
>
> i've been working on this (also because of a post-pass scheduler).
>
> i managed to hand the source values (values passed to the AA) over to
> the machine instruction level. unfortunately the "codegen
prepare"
> pass
> rewrites GEP llvm instructions and confuses the AA. some work needs to
> be done inside the AA to get reasonable results.
>
Right, currently AA information might get lost during translation from  
llvm to machineinstrs and during various passes that operate on them.  
This is not a hard problem, but very tedious.

> i did not have time to prepare and submit a patch for llvm.
It would be nice if you can contribute it back even if it's  
incomplete. :-)

Evan

>
>
> florian
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev