[PATCH] D116692: [SimplifyCFG] Tail-merging all blocks with `unreachable` terminator, final take

Mon Jan 17 16:03:14 PST 2022

aeubanks added a comment.

In D116692#3249159 <https://reviews.llvm.org/D116692#3249159>, @lebedev.ri wrote:

> In D116692#3232485 <https://reviews.llvm.org/D116692#3232485>, @nikic wrote:
>
>> It looks like the new version still has the large code size regressions (9% on mafft, 3% on 7zip). I understand that some code size increase is expected (and intended), but I don't think a particularly good case for the tradeoff has been made yet (in terms of where / how much performance this is buying for more code size). Though maybe I missed this in previous discussion threads.
>
> How do we know that whatever compile time benchmark we see regresses is a reliable indicator in this regard?
> I think this is yet another irresolvable clash between the optimization and compilation time/size.
> -O3 does not mean "please quickly give me minimal code", there's -Os/-Oz for that.
> IOW if you indent to block a patch, could you please actually do so, not just waguley imply so?

As somebody without much context for this patch reading the patch summary, there's no "why" for this patch. The llvm-compile-time-tracker numbers are all negative in various aspects. I think nikic's question is do you have metrics showing that this actually helps performance on any code?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D116692/new/

https://reviews.llvm.org/D116692