On Fri, Sep 4, 2026 at 10:19 AM, Dora, Sunil Kumar wrote:

I am reducing a testcase and will report it to llvm, will post the link here.
If a fix lands I will request a release/23.x backport.

Hi Khem,

I bisected llvm between 22.1.8 and 23.1.0. The problem comes from the new DFAJumpThreading pass
which is on by default in clang 23. The full file takes more than 90 minutes where clang 22 needs about
2 seconds.

I reported it upstream with a reduced testcase:
https://github.com/llvm/llvm-project/issues/221391

There is also a better workaround than the -disable-lsr I posted earlier. Turning off just the new pass keeps
LSR and behaves the same as clang 22:

CFLAGS:append:toolchain-clang = " -mllvm -enable-dfa-jump-thread=false"

I sent this as v2 of your dovecot patch on openembedded-devel. Once the llvm fix lands I will ask for a
release/23.x backport, then we can drop the workaround.

Thanks,
Sunil Dora