Re: broken JIT support on Fedora 40 - Mailing list pgsql-hackers

From Pavel Stehule
Subject Re: broken JIT support on Fedora 40
Date
Msg-id CAFj8pRAfnqdt=efcR8-1Zw8txnt6Wa5F=7RcBX4BdqtYLyBH=g@mail.gmail.com
Whole thread Raw
In response to Re: broken JIT support on Fedora 40  (Thomas Munro <thomas.munro@gmail.com>)
List pgsql-hackers


st 10. 4. 2024 v 2:44 odesílatel Thomas Munro <thomas.munro@gmail.com> napsal:
On Tue, Apr 9, 2024 at 10:05 PM Dmitry Dolgov <9erthalion6@gmail.com> wrote:
>         +       /* In assertion builds, run the LLVM verify pass. */
>         +#ifdef USE_ASSERT_CHECKING
>         +       LLVMPassBuilderOptionsSetVerifyEach(options, true);
>         +#endif

Thanks, that seems nicer.  I think the question is whether it will
slow down build farm/CI/local meson test runs to a degree that exceeds
its value.  Another option would be to have some other opt-in macro,
like the existing #ifdef LLVM_PASS_DEBUG, for people who maintain
JIT-related stuff to turn on.

Supposing we go with USE_ASSERT_CHECKING, I have another question:

-       const char *nm = "llvm.lifetime.end.p0i8";
+       const char *nm = "llvm.lifetime.end.p0";

Was that a mistake, or did the mangling rules change in some version?
I don't currently feel inclined to go and test this on the ancient
versions we claim to support in back-branches.  Perhaps we should just
do this in master, and then it'd be limited to worrying about LLVM
versions 10-18 (see 820b5af7), which have the distinct advantage of
being available in package repositories for testing.  Or I suppose we
could back-patch, but only do it if LLVM_VERSION_MAJOR >= 10.  Or we
could do it unconditionally, and wait for ancient-LLVM build farm
animals to break if they're going to.

I pushed the illegal attribute fix though.  Thanks for the detective work!

(It crossed my mind that perhaps deform functions should have their
own template function, but if someone figures out that that's a good
idea, I think we'll *still* need that change just pushed.)

all tests passed on fc 40 without problems

Thank you

Pavel

pgsql-hackers by date:

Previous
From: Kirill Reshke
Date:
Subject: Re: psql: Greatly speed up "\d tablename" when not using regexes
Next
From: Andres Freund
Date:
Subject: Re: Issue with the PRNG used by Postgres