llvm dev - Aug 2021 - [llvm-reduce] Reduction to undef/poison/null?

Currently in llvm-reduce when we try to remove instructions, we'll RAUW
them with undef. But it seems nicer to have the final IR use null values
(e.g. 0, zeroinitializer, etc) rather than undef. Should we attempt to use
null values? Or why undef over poison?
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210830/b7a068f5/attachment.html>

Nicer because it's less likely to introduce new UB? Or some other reason?

I /guess/ undef because it's more vague/shows the value doesn't matter
to
some degree?

(might be worth CC'ing some of the other folks who've contributed
patches/implemented llvm-reduce - not everyone reads llvm-dev or does so
frequently enough to see relevant threads in a timely manner)

On Mon, Aug 30, 2021 at 10:50 AM Arthur Eubanks via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Currently in llvm-reduce when we try to remove instructions, we'll RAUW
> them with undef. But it seems nicer to have the final IR use null values
> (e.g. 0, zeroinitializer, etc) rather than undef. Should we attempt to use
> null values? Or why undef over poison?
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210830/8c786374/attachment.html>

What I usually end up doing when reducing by hand is to introduce a new function
parameter to replace a deleted instruction, e.g.:

 

define @f(i32 %x) {

  %y = add i32 %x, ...

...

}

=>

define @f(i32 %x, i32 %y) {

<remove instr %y>

...

}

 

Though not all users will want this sort of reduction..

 

I agree with what other said about undef/poison: avoid them at all costs when
reducing. It will just expose some other bug regarding handling of undef in
optimizers instead of the bug you wanted to track down :)

 

Nuno

 

 

From: Arthur Eubanks via llvm-dev
Sent: 30 August 2021 18:50
To: llvm-dev <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] [llvm-reduce] Reduction to undef/poison/null?

 

Currently in llvm-reduce when we try to remove instructions, we'll RAUW them
with undef. But it seems nicer to have the final IR use null values (e.g. 0,
zeroinitializer, etc) rather than undef. Should we attempt to use null values?
Or why undef over poison?

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210830/483a2f03/attachment.html>