PPC: support closure cfunctions (aborts during codegen)

[vchuravy]

From worker 46: running testset ccall...

      From worker 46:
      From worker 46:   signal (6): Aborted
      From worker 46:   in expression starting at /autofs/nccsopen-svm1_home/vchuravy/julia/test/ccall.jl:837
      From worker 46:   gsignal at /lib64/libc.so.6 (unknown line)
      From worker 46:   abort at /lib64/libc.so.6 (unknown line)
      From worker 46:   unknown function (ip: 0x7fff843a83d7)
      From worker 46:   jl_get_cfunction_trampoline at /autofs/nccsopen-svm1_home/vchuravy/julia/src/runtime_ccall.cpp:322
      From worker 46:   Allocations: 11091552 (Pool: 11089407; Big: 2145); GC: 26

Upstream:

• https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98952
• https://bugs.llvm.org/show_bug.cgi?id=49004

[vchuravy]

Turns out the abort is not due to lack of support in LLVM, but part of the implementation of __init_trampoline

#27174 (comment)

[vchuravy]

So in r4 we are passing 48 which stems from https://github.com/llvm/llvm-project/blob/ffc3e800c65ee58166255ff897f8b7e6d850ddda/llvm/lib/Target/PowerPC/PPCISelLowering.cpp#L3485

0000000000008658 <__trampoline_setup>:
    8658:       03 00 4c 3c     addis   r2,r12,3
    865c:       a8 f6 42 38     addi    r2,r2,-2392
    8660:       fd ff e2 3c     addis   r7,r2,-3
    8664:       30 09 e7 38     addi    r7,r7,2352
    8668:       20 00 00 39     li      r8,32
    866c:       00 20 88 7c     cmpw    cr1,r8,r4
    8670:       fe e8 84 54     rlwinm  r4,r4,29,3,31
    8674:       f8 ff 23 39     addi    r9,r3,-8
    8678:       a6 03 89 7c     mtctr   r4
    867c:       38 00 84 41     blt     cr1,86b4 <__trampoline_setup+0x5c>
    8680:       09 00 47 e9     ldu     r10,8(r7)
    8684:       09 00 49 f9     stdu    r10,8(r9)
    8688:       f8 ff 00 42     bdnz    8680 <__trampoline_setup+0x28>
    868c:       10 00 a3 f8     std     r5,16(r3)
    8690:       18 00 c3 f8     std     r6,24(r3)
    8694:       a6 03 89 7c     mtctr   r4
    8698:       ac 1f 00 7c     icbi    0,r3
    869c:       ac 18 00 7c     dcbf    0,r3
    86a0:       08 00 63 38     addi    r3,r3,8
    86a4:       f4 ff 00 42     bdnz    8698 <__trampoline_setup+0x40>
    86a8:       ac 04 00 7c     hwsync
    86ac:       2c 01 00 4c     isync
    86b0:       20 00 80 4e     blr
    86b4:       bd a9 ff 4b     bl      3070 <00000017.plt_call.abort@@GLIBC_2.17>

I am slightly confused since

li      r8,32
cmpw cr1,r8,r4
blt     cr1,86b4 <__trampoline_setup+0x5c>

Means branch to <__trampoline_setup+0x5c> if r8 is less then r4. With r4=48 and r8=32 that is true, therefore we jump to abort.

I took https://github.com/gcc-mirror/gcc/blob/02551aa99960ff6b0d9a44d234b8f6c9b230788c/libgcc/config/rs6000/tramp.S#L159
to mean that we check whether the provided trampoline buffer is big enough, for that the check would be the wrong way around.

[vchuravy]

This is fixed on recent glibc.