[clang] [Clang][Inliner] Enable module-local optimizer hints by default (PR #238529)
Kuba Mracek via cfe-commits
cfe-commits at lists.llvm.org
Sat Nov 7 03:05:37 PST 2026
kubamracek wrote:
> re-check compile time impact in general
This is a fresh LLVM Compile Time Tracker result: https://llvm-compile-time-tracker.com/compare.php?from=71653467fa83249f7f748eb05347950f91c46f26&to=8bb77011b5687e90474f11a7536d5d969c59af73&stat=instructions:u
There is the one heavy outlier in llvm-test-suite, tramp3d-v4, which we already discovered in the previous PR. I think I'd want to discuss that one separately (the structure of this benchmark is... unusual and inherently makes it very heavy on this new optimization). The reported geomeans are heavily skewed because of tramp3d-v4, If I exclude it, the geomeans are between -0.16% and +0.07% for -O3 and/or LTO.
For -O0 the deltas are basically 0 for -O0 (as expected).
The compile time measurements that are beyond a 0.10% change are a small subset:
```
| stage1-O3: | |
| - 7zip | 194177M -> 194450M (+0.14%) |
| - beman-execution | 7686M -> 7705M (+0.24%) |
| stage1-ReleaseThinLTO: | |
| - kimwitu++ | 43950M -> 43251M (-1.59%) |
| stage1-ReleaseLTO-g: | |
| - Bullet | 99088M -> 99263M (+0.18%) |
| - 7zip | 240291M -> 240544M (+0.11%) |
| stage1-aarch64-O3: | |
| - kimwitu++ | 36578M -> 36425M (-0.42%) |
| - Bullet | 98775M -> 98872M (+0.10%) |
| - 7zip | 201448M -> 201749M (+0.15%) |
| stage2-O3: | |
| - Bullet | 83506M -> 83664M (+0.19%) |
| - 7zip | 173713M -> 173958M (+0.14%) |
```
I would tend to think that (again, with the exlusion of tramp3d-v4) this is reasonable and somewhat expected, we're inlining in some more cases, and in exchange we're getting more optimizations. There's a case of a compile-time reduction, too (kimwitu++).
https://github.com/llvm/llvm-project/pull/238529
More information about the cfe-commits
mailing list