(original) (raw)

Hi Stefan,

I agree that this change is overdue. Thank you very much for working on it. :)

While I think this change is overdue, there is no rush to land it. I am
happy to discuss details in case you have any concerns.

I have no concerns with the intended change. Once you have left a few days for others to comment, and assuming there are no objections, I would go ahead and land this. I suspect you will see some extra failures when this hits the builders, but that's to be expected -- you may find that we need a few rounds of applying and reverting this before all the kinks can be worked out. Landing early gives us plenty of time to sort those out before the next release.

Regards,

Lang.

On Fri, Mar 19, 2021 at 3:06 AM Stefan Gränitz <stefan.graenitz@gmail.com> wrote:

Hi all

In review https://reviews.llvm.org/D98931 I am proposing to change lli's
default JIT engine to Orc. The main intention here is to get people to
realize that MCJIT is poorly maintained and that it might be
inappropriate for use with new code and tests. In fact, last week the
MLIR subproject added some new tests that use lli with MCJIT. I guess
this was unintentional; the reviews at least don't seem to discuss the
JIT engine.

Common JIT tasks are expected to work with Orc just like they did with
MCJIT. All such tests run against both engines now:
https://reviews.llvm.org/D98931?id=331802

There is a number of tests that require MCJIT-specific behavior (e.g.
interpreter mode). These were patched to explicitly pass -jit-kind=mcjit
now:
https://reviews.llvm.org/D98931?vs=331802&id=331803

While I think this change is overdue, there is no rush to land it. I am
happy to discuss details in case you have any concerns.

Cheers
Stefan

\--
https://flowcrypt.com/pub/stefan.graenitz@gmail.com