[llvm-dev] [RFC] Enabling opt's -enable-new-pm by default with ENABLE_NEW_EXPERIMENTAL_PASS_MANAGER

[Florian Hahn via llvm-dev]

> On Jan 21, 2021, at 21:02, Arthur Eubanks via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> With many of the new PM issues ironed out, I'd like to turn on opt's -enable-new-pm flag by default when we build LLVM with ENABLE_NEW_EXPERIMENTAL_PASS_MANAGER on.
> 

That sounds like a great idea & sounds good to me, as long as all current tests pass with the new behavior and there’s a public bot that uses ENABLE_NEW_EXPERIMENTAL_PASS_MANAGER=true.

Thanks everyone involved in ironing out the remaining issues!

> -enable-new-pm essentially translates the legacy way of specifying opt passes (e.g. -instcombine) into the new way of specifying passes (e.g. -passes=instcombine), making tests that use `opt -instcombine` run under the new PM, unless specifically opted out via `-enable-new-pm=0`.
> 
> One issue is that until we enable the new PM by default, testing locally may not show issues that pop up until a bot with ENABLE_NEW_EXPERIMENTAL_PASS_MANAGER on complains after committing, since basically everybody has ENABLE_NEW_EXPERIMENTAL_PASS_MANAGER off locally. Hopefully the amount of time between conditionally turning on opt's -enable-new-pm and turning on the new PM by default everywhere is fairly small and this won't be too big of an issue. I haven't seen this discrepancy appear too often when looking at existing passing tests failing when -enable-new-pm is turned on.

I think that’s fine, the same issue exists for a lot of different options. IIUC to reproduce a failure, it should be enough to pass `-enable-new-pm` to the failing test locally?

Cheers,
Florian