search

iains / gcc-darwin-arm64

GCC master branch for Darwin with experimental support for Arm64. Currently GCC-15.0.0 [October 2024]
GNU General Public License v2.0
269 stars 34 forks source link

Fortran compiler seg-fault with -O2 and -ftree-loop-vectorize or -ftree-slp-vectorize #44

Closed erykoff closed 3 years ago

erykoff commented 3 years ago

Somewhere between commit https://github.com/iains/gcc-darwin-arm64/commits/62cfbb7cf8a06c2ec07d3343bfdc67ca8123239e and https://github.com/iains/gcc-darwin-arm64/commits/02d30eeb2534a593216564c07e94f4d49f0b679e (corresponding to builds 17 and 18 on conda-forge, https://github.com/conda-forge/gfortran_impl_osx-64-feedstock/commit/9656879871750f281983f16c31bf68894257695e ), the fortran compiler can segmentation fault on certain source files.

Specifically, https://github.com/scipy/scipy/blob/master/scipy/special/specfun/specfun.f will cause the compiler to segmentation fault if compiled with -O2 -ftree-loop-vectorize or -O2 -ftree-slp-vectorize.

iains commented 3 years ago

thanks for the report, however it does not reproduce for me with a build from the current sources:

$ /opt/iains/aarch64-apple-darwin20/gcc-11-0-1/bin/gfortran --version GNU Fortran (GCC) 11.0.1 20210313 (experimental) [master-wip-apple-si revision r11-7708-g3d132a12726b]

$ /opt/iains/aarch64-apple-darwin20/gcc-11-0-1/bin/gfortran /src-local/specfun.f -O2 -ftree-slp-vectorize -S $ echo $? 0

$ /opt/iains/aarch64-apple-darwin20/gcc-11-0-1/bin/gfortran /src-local/specfun.f -O2 -ftree-loop-vectorize -S $ echo $? 0

erykoff commented 3 years ago

Ah! It may be fixed on master, sorry I wasn't more thorough in my report. I was having trouble building and only had easy access to the latest hash released on conda-forge.

isuruf commented 3 years ago

Thanks @erykoff for opening this issue and @iains for confirming. @erykoff, it looks like the segfault occurs only in our osx-x86_64->osx-arm64 cross compiler and not the native compiler.

erykoff commented 3 years ago

With further investigation, it seems that this happens on the cross-compiler (at least) with the addition of -march=armv8.3-a which is set with the conda forge FFLAGS.

iains commented 3 years ago

FWIW, I build a x86_64 X aarch64 every week on x86-64-darwin20 (and an aarch64 target/host built on x86_64 so a "native" or "canadian" cross). here:

$ /opt/iains/x86_64-apple-darwin20/gcc-11-0-1-asi/bin/aarch64-apple-darwin20-gfortran --version GNU Fortran (GCC) 11.0.1 20210313 (experimental) Copyright (C) 2021 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ /opt/iains/x86_64-apple-darwin20/gcc-11-0-1-asi/bin/aarch64-apple-darwin20-gfortran /source/test/fortran/specfun.f -O2 -ftree-loop-vectorize -S -w

seems OK there too.

isuruf commented 3 years ago

@erykoff, I can reproduce with conda-forge compilers with -march=armv8.3-a -O2 -ftree-loop-vectorize. I'll try updating to latest master.

iains commented 3 years ago

adding -march=armv8.3-a makes no difference to either my native or cross compiler (at least with -O2 -ftree-loop-vectorize)

Of course, this is an experimental branch :)

isuruf commented 3 years ago

Of course, this is an experimental branch :)

Of course and many thanks for doing the work to port gcc to aaarch64-apple-darwin. We @conda-forge and all our users appreciate it.

isuruf commented 3 years ago

Latest commit fixes this segfault. This issue can be closed.

isuruf commented 3 years ago

With e330c4193fcc9dfd030d14b309b573753aa0afb0, there's a new segfault. compiling https://raw.githubusercontent.com/Reference-LAPACK/lapack/master/BLAS/SRC/cgemv.f with gfortran cgemv.f -O2 -march=armv8.3-a -ftree-vectorize -fPIC -fno-stack-protector -c errors with,

during GIMPLE pass: slp
cgemv.f:157:22:

  157 |       SUBROUTINE CGEMV(TRANS,M,N,ALPHA,A,LDA,X,INCX,BETA,Y,INCY)
      |                      ^
internal compiler error: in vect_slp_analyze_node_operations_1, at tree-vect-slp.c:3727
Please submit a full bug report,
with preprocessed source if appropriate.
See <https://gcc.gnu.org/bugs/> for instructions.

whereas gfortran cgemv.f -O2 -march=armv8.2-a -ftree-vectorize -fPIC -fno-stack-protector -c doesn't

iains commented 3 years ago

I see this also on an x86_64 => aarch64 cross - here's the backtrace.

I think the next thing to check is whether this fires for a non-Darwin case - this is not part of the compiler I've touched with my port.

    frame #3: 0x00000001017da973 f951`::__second_sect_of_vect_slp_analyze_node_operations() at tree-vect-slp.c:3727:3
    frame #4: 0x00000001011e1d86 f951`::vect_slp_analyze_node_operations(vinfo=0x0000000144d1b210, node=<unavailable>, node_instance=0x0000000144d6e580, visited_set=0x00007ffeefbff030, visited_vec=<unavailable>, cost_vec=0x00007ffeefbff060) at tree-vect-slp.c:3913:46
    frame #5: 0x00000001011e1d86 f951`::vect_slp_analyze_node_operations(vinfo=0x0000000144d1b210, node=<unavailable>, node_instance=0x0000000144d6e580, visited_set=0x00007ffeefbff030, visited_vec=<unavailable>, cost_vec=0x00007ffeefbff060) at tree-vect-slp.c:3913:46
    frame #6: 0x00000001011e1d86 f951`::vect_slp_analyze_node_operations(vinfo=0x0000000144d1b210, node=<unavailable>, node_instance=0x0000000144d6e580, visited_set=0x00007ffeefbff030, visited_vec=<unavailable>, cost_vec=0x00007ffeefbff060) at tree-vect-slp.c:3913:46
    frame #7: 0x00000001011e1d86 f951`::vect_slp_analyze_node_operations(vinfo=0x0000000144d1b210, node=<unavailable>, node_instance=0x0000000144d6e580, visited_set=0x00007ffeefbff030, visited_vec=<unavailable>, cost_vec=0x00007ffeefbff060) at tree-vect-slp.c:3913:46
    frame #8: 0x00000001011e1d86 f951`::vect_slp_analyze_node_operations(vinfo=0x0000000144d1b210, node=<unavailable>, node_instance=0x0000000144d6e580, visited_set=0x00007ffeefbff030, visited_vec=<unavailable>, cost_vec=0x00007ffeefbff060) at tree-vect-slp.c:3913:46
    frame #9: 0x00000001011e1d86 f951`::vect_slp_analyze_node_operations(vinfo=0x0000000144d1b210, node=<unavailable>, node_instance=0x0000000144d6e580, visited_set=0x00007ffeefbff030, visited_vec=<unavailable>, cost_vec=0x00007ffeefbff060) at tree-vect-slp.c:3913:46
    frame #10: 0x00000001011e4af6 f951`vect_slp_analyze_operations(vinfo=<unavailable>) at tree-vect-slp.c:4120:45
    frame #11: 0x00000001011e88a3 f951`::vect_slp_region(bbs=<unavailable>, datarefs=<unavailable>, dataref_groups=0x00007ffeefbff318, n_stmts=746) at tree-vect-slp.c:4899:36
    frame #12: 0x00000001011ea630 f951`::vect_slp_bbs(bbs=vec<basic_block_def*, va_heap, vl_ptr> @ r14) at tree-vect-slp.c:5095:26
    frame #13: 0x00000001011eaa5d f951`vect_slp_function(fun=0x00000001437e3000) at tree-vect-slp.c:5181:23
    frame #14: 0x00000001011ee92a f951`pass_slp_vectorize::execute(this=<unavailable>, fun=0x00000001437e3000) const at tree-vectorizer.c:1450:21
    frame #15: 0x0000000100db0ba2 f951`execute_one_pass(pass=0x0000000143512650) at passes.c:2567:30
    frame #16: 0x0000000100db14b3 f951`::execute_pass_list_1(pass=0x0000000143512650) at passes.c:2656:28
    frame #17: 0x0000000100db14c5 f951`::execute_pass_list_1(pass=0x0000000143511bd0) at passes.c:2657:22
    frame #18: 0x0000000100db14c5 f951`::execute_pass_list_1(pass=0x0000000143510910) at passes.c:2657:22
    frame #19: 0x0000000100db14f2 f951`execute_pass_list(fn=0x00000001437e3000, pass=<unavailable>) at passes.c:2667:23
    frame #20: 0x0000000100895861 f951`cgraph_node::expand(this=0x00000001437f4000) at cgraphunit.c:1830:21
    frame #21: 0x00000001008969cf f951`symbol_table::compile() at cgraphunit.c:1998:17
    frame #22: 0x00000001008967bd f951`symbol_table::compile(this=<unavailable>)
    frame #23: 0x0000000100899d76 f951`symbol_table::finalize_compilation_unit() [inlined] symbol_table::compile(this=<unavailable>) at cgraphunit.c:2275:3
    frame #24: 0x0000000100899d65 f951`symbol_table::finalize_compilation_unit(this=0x0000000143606000)
    frame #25: 0x0000000100ed3e11 f951`::compile_file() at toplev.c:482:41
    frame #26: 0x0000000101817270 f951`toplev::main(int, char**) at toplev.c:2201:24
    frame #27: 0x0000000101816f3e f951`toplev::main(this=0x00007ffeefbff73e, argc=<unavailable>, argv=<unavailable>)
    frame #28: 0x0000000101819481 f951`main(argc=24, argv=0x00007ffeefbff770) at main.c:39:22
iains commented 3 years ago

As suspected, not specific to the Darwin port.

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99807 Bug 99807 - [11 Regression] ICE in vect_slp_analyze_node_operations_1, at tree-vect-slp.c:3727

This is fixed on master, the fix was 29 March, so will be pulled in on the next rebase (end of this week).

iains commented 3 years ago

my local testing shows this fixed with the just-pushed branch. Please check and close this if so ...

isuruf commented 3 years ago

Yes, it works for me. Thanks. @erykoff, we can close this issue.