[PATCH] D147740: [NFC][llvm-profdata] Refactoring Sample Profile Reader to increase FDO build speed

David Li via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Tue Apr 25 22:39:28 PDT 2023


davidxl added a comment.

In D147740#4297883 <https://reviews.llvm.org/D147740#4297883>, @wenlei wrote:

> In D147740#4250599 <https://reviews.llvm.org/D147740#4250599>, @davidxl wrote:
>
>> In D147740#4250545 <https://reviews.llvm.org/D147740#4250545>, @aeubanks wrote:
>>
>>> as somebody with very little context for this, is it possible to use faster hashing algorithms?
>>
>> faster hashing may help, but I think the intention of the patch is to eliminate unnecessary computation completely.
>
> Might be a bit off topic, but we have seen MD5 hashing in the top hottest functions in profile of llvm itself when compiling workloads with compact binary profile in the past.
>
> It'd be great if we can transition to cheaper hash (e.g. xxhash) for all profile format. We need hash with all distribution, but we don't need cryptographic hash.
>
> I think it's doable to switch profile format to use xxhash, with a flag in header to tell reader what hash to expect. But agreed that it's orthogonal to this patch.

Agree -- it is worth exploring further after this round of changes.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147740/new/

https://reviews.llvm.org/D147740



More information about the llvm-commits mailing list