[PATCH] D103009: [DSE] Transform memset + malloc --> calloc (PR25892)

[Dávid Bolvanský via Phabricator via llvm-commits]

xbolva00 added a comment.

In D103009#2954430 <https://reviews.llvm.org/D103009#2954430>, @alexfh wrote:

> This commit seems to cause memory usage (rss) increase in MariaDB's mysqld by a factor of 4. Looking back into the history, I found that the previous commit here caused the same regression, but we quickly picked up the revert and moved on. Now I'm trying to isolate the problem, but it's taking time.
> So far, my sole hypothesis is that malloc + memset of a smaller size can still be converted to calloc. But I have no evidence so far. Naive attempts to synthetically reproduce this didn't work. Maybe this only happens when some UB is in place, but again, I have nothing to back this up.
>
> Given this is a quite serious regression, can we roll this back while I'm investigating?

You should be able to use flag -fno-builtin-calloc to disable this transformation.

This transformation is correct and valid; GCC does it as well. No reason to revert, not justified. You should check asm diffs w and w/o patch for any surprises.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103009/new/

https://reviews.llvm.org/D103009