Re: broken JIT support on Fedora 40

Dmitry Dolgov <9erthalion6@gmail.com>

From: Dmitry Dolgov <9erthalion6@gmail.com>
To: Thomas Munro <thomas.munro@gmail.com>
Cc: Daniel Gustafsson <daniel@yesql.se>, Pavel Stehule <pavel.stehule@gmail.com>, Robert Haas <robertmhaas@gmail.com>, PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Date: 2024-03-17T20:02:08Z
Lists: pgsql-hackers
> On Fri, Mar 15, 2024 at 01:54:38PM +1300, Thomas Munro wrote:
> For me it seems that the LLVMRunPasses() call, new in
>
> commit 76200e5ee469e4a9db5f9514b9d0c6a31b496bff
> Author: Thomas Munro <tmunro@postgresql.org>
> Date:   Wed Oct 18 22:15:54 2023 +1300
>
>     jit: Changes for LLVM 17.
>
> is reaching code that segfaults inside libLLVM, specifically in
> llvm::InlineFunction(llvm::CallBase&, llvm::InlineFunctionInfo&, bool,
> llvm::AAResults*, bool, llvm::Function*).  First obvious question
> would be: is that NULL argument still acceptable?  Perhaps it wants
> our LLVMTargetMachineRef there:
>
>        err = LLVMRunPasses(module, passes, NULL, options);
>
> But then when we see what is does with that argument, it arrives at a
> place that apparently accepts nullptr.
>
> https://github.com/llvm/llvm-project/blob/6b2bab2839c7a379556a10287034bd55906d7094/llvm/lib/Passes/PassBuilderBindings.cpp#L56
> https://github.com/llvm/llvm-project/blob/6b2bab2839c7a379556a10287034bd55906d7094/llvm/include/llvm/Passes/PassBuilder.h#L124
>
> Hrmph.  Might need an assertion build to learn more.  I'll try to look
> again next week or so.

Looks like I can reproduce this as well, libLLVM crashes when reaching
AddReturnAttributes inside InlineFunction, when trying to access
operands of the return instruction. I think, for whatever reason, the
latest LLVM doesn't like (i.e. do not expect this when performing
inlining pass) return instructions that do not have a return value, and
this is what happens in the outblock of deform function we generate
(slot_compile_deform).

For verification, I've modified the deform.outblock to call LLVMBuildRet
instead of LLVMBuildRetVoid and this seems to help -- inline and deform
stages are still performed as before, but nothing crashes. But of course
it doesn't sound right that inlining pass cannot process such code.
Unfortunately I don't see any obvious change in the recent LLVM history
that would justify this outcome, might be a genuine bug, will
investigate further.



Commits

  1. Run LLVM verify pass on IR in assert builds.

  2. Fix illegal attribute propagation in LLVM JIT.

  3. Track LLVM 18 changes.

  4. jit: Changes for LLVM 17.