[llvm-dev] After InstCombine pass, expected output got changed

[Florian Hahn via llvm-dev]

> On Aug 20, 2021, at 13:45, Rohit Aggarwal <cs18mtech11030 at iith.ac.in> wrote:
> 
> Hi Florian,
> 
> 
> 
> On Thu, Aug 19, 2021 at 3:46 PM Florian Hahn <florian_hahn at apple.com <mailto:florian_hahn at apple.com>> wrote:
> 
> 
>> On 19 Aug 2021, at 11:01, Rohit Aggarwal via llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>> 
>> 
>> Hi All,
>> 
>> Thanks Florian for response.
>> 
>>  Adding more details to the issue:
>> I revert the commit  [ConstantFold] Allow propagation of poison for and/or i1 <https://github.com/llvm/llvm-project/commit/2fd3037ac615643fe8058292d2b89bb19a49cb2f#> then it is working fine.
>> In this commit, a check is removed to stop the constantfold of instruction to poison value incase of integer ADD or OR instruction if either of operands are poison.
>> Using both release branch and commit reverted branch, a poison value is generated by a shift left instruction due to overflow (UB is present) 
> 
> Ok so it sounds like this is the reason for the different output and the transformation is correct.
>     >> Yes, this poison which is generated is handled differently in llvm-trunk and reverted commit code down the pipeline. 
>  
>> Now in the release branch, this poison value is propagated down to 'or' inst. and  is constantfolds %or = or  true poison  -> poison  
> 
> 
> where as in the reverted code,  %or = or  true poison  -> true
>   >> Here is the main reason, where the problem started 
>> In the release branch, the poison value propagates to its uses
>> There is a store instruction which now has a poison value
>> In visitStoreInst, store instruction got removed due to poison value
>> But in revert branch, store is not getting remove as it has defined value
Sure, the change in instcombine propagates poison more aggressively. But as far as I can tell, this is in line with the semantics specified in LangRef and the transformation is correct. The reason for the change in output seems to be that instcombine was taught to exploit the undefined behavior in this case.

If this change in output is causing any issues, I’d suggest you check the source file from which the LLVM IR has been generated for any undefined behavior and fix it.

If the input source file is well defined, I’d suggest to check where the poison value causing issues comes from.

Cheers,
Florian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210820/a18ae5ad/attachment.html>