rustc SIGILL since qsort_r patches

[Andrew Pinski]

On Sat, Nov 4, 2023 at 6:10 PM Andrew Pinski <pinskia@gmail.com> wrote:
>
> On Sat, Nov 4, 2023 at 5:55 PM Cristian Rodríguez <cristian@rodriguez.im> wrote:
> >
> > Hi:
> >
> > After quite some head scratching I found the recent qsort_r patches triggers a bug "somewhere" that makes rustc LTO end in sigkill (ud2)
> >
> > happens compiling anything in release mode with any toolchain (stable, beta, nightly)
> >
> > #0  0x000015129e40516a in int llvm::array_pod_sort_comparator<llvm::BranchFolder::MergePotentialsElt>(void const*, void const*) () from /home/crrodriguez/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/bin/../lib/../lib/libLLVM-17-rust-1.75.0-nightly.so
> >
> > #1  0x00001512a0043779 in __GI___qsort_r (pbase=0x15128c90bd80, total_elems=<optimized out>, size=0x10, cmp=<optimized out>, arg=<optimized out>) at qsort.c:335
> >
> > #2  0x000015129e4df6e6 in llvm::BranchFolder::TryTailMergeBlocks(llvm::MachineBasicBlock*, llvm::MachineBasicBlock*, unsigned int) () from /home/crrodriguez/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/bin/../lib/../lib/libLLVM-17-rust-1.75.0-nightly.so
> >
>
> I think this is a bug in LLVM's
> BranchFolder::MergePotentialsElt::operator< code:
> See: https://llvm.org/doxygen/BranchFolding_8cpp_source.html
>
> Specifically it has the following comment:
> // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
> 489 // an object with itself.
> 490#ifndef _GLIBCXX_DEBUG
> 491 llvm_unreachable("Predecessor appears twice");
> 492#else
> 493 return false;
> 494#endif
>
> Well glibc's qsort implementation will also now compare with itself.

Also qsort comparison is not exactly < but rather <=>  which
array_pod_sort_comparator tries to emulate but
BranchFolder::MergePotentialsElt::operator< specifically calls
unreachable if 2 are things are equal.

>
> Thanks,
> Andrew
>
> > git head == bad
> > 2.38   == good
> >
> > Main interest is to know if I am not alone and if anyone else has observed this problem so far.
> > Thanks..