[PATCH] D38154: [PassManager] Run global opts after the inliner

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc added a comment.

In https://reviews.llvm.org/D38154#894312, @andjo403 wrote:

> so it is possible that the large improvement in compile time is due to we broke the pipeline and due to this less optimization was possible/executed.


That's one possibility. There also is quite likely an issue Rust is hitting that the later positioning of these passes doesn't help with. It's hard or impossible to tell which is dominant here....

The only thing I'm confident in is what Eli said: this positioning *does* cripple the intended design of the core LLVM optimization pipeline and will severely hurt inlining and optimization quality on some code patterns. (Others, of course, will be largely unimpacted.)


Repository:
  rL LLVM

https://reviews.llvm.org/D38154