[Bug target/95730] New: GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[Bug target/95730] New: GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

            Bug ID: 95730
           Summary: GCN offloading ICEs after commit
                    fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support
                    for __builtin_bswap128"
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Keywords: openacc, openmp
          Severity: normal
          Priority: P3
         Component: target
          Assignee: ams at gcc dot gnu.org
          Reporter: tschwinge at gcc dot gnu.org
                CC: ebotcazou at gcc dot gnu.org, jules at gcc dot gnu.org
  Target Milestone: ---
            Target: gcn

For a few libgomp test cases, the recent commit
fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"
is causing ICE regressions with GCN offloading (only, apparently: not seen for
Intel MIC (emulated), or nvptx offloading):

'libgomp.fortran/pr90779.f90':

    lto1: internal compiler error: tree check: expected class 'type', have
'exceptional' (error_mark) in lto_input_tree_1, at lto-streamer-in.c:1561

    #0  internal_error (gmsgid=0x163f958 "tree check: expected class %qs, have
%qs (%s) in %s, at %s:%d") at [...]/source-gcc/gcc/diagnostic.c:1706
    #1  0x000000000055f2fd in tree_class_check_failed (node=0x7ffff67b4960,
cl=<optimized out>, file=<optimized out>, line=<optimized out>,
function=<optimized out>) at [...]/source-gcc/gcc/tree.c:9738
    #2  0x00000000009d7886 in tree_class_check (__g=<optimized out>, __l=1561,
__f=0x15d1208 "[...]/source-gcc/gcc/lto-streamer-in.c", __class=tcc_type,
__t=0x7ffff67b4960) at [...]/source-gcc/gcc/tree.h:3424
    #3  lto_input_tree_1 (ib=0x7fffffffd080, data_in=0x1dd4860, tag=<optimized
out>, hash=<optimized out>) at [...]/source-gcc/gcc/lto-streamer-in.c:1561
    #4  0x00000000005d84b6 in lto_read_decls (resolutions=..., data=0x1e3a630,
decl_data=0x7ffff7fee090) at [...]/source-gcc/gcc/lto/lto-common.c:1955
    #5  lto_file_finalize (file=0x1d8c900, order=0, file_data=0x7ffff7fee090)
at [...]/source-gcc/gcc/lto/lto-common.c:2267
    #6  lto_create_files_from_ids (order=0, count=<synthetic pointer>,
file_data=0x7ffff7fee090, file=0x1d8c900) at
[...]/source-gcc/gcc/lto/lto-common.c:2277
    #7  lto_file_read (count=<synthetic pointer>, resolution_file=<optimized
out>, file=<optimized out>) at [...]/source-gcc/gcc/lto/lto-common.c:2332
    #8  read_cgraph_and_symbols (nfiles=<optimized out>, fnames=<optimized
out>) at [...]/source-gcc/gcc/lto/lto-common.c:2784
    #9  0x00000000005b9c23 in lto_main () at [...]/source-gcc/gcc/lto/lto.c:625
    #10 0x0000000000b7f6ff in compile_file () at
[...]/source-gcc/gcc/toplev.c:458
    #11 0x000000000059302c in do_compile () at
[...]/source-gcc/gcc/toplev.c:2307
    #12 toplev::main (this=this@entry=0x7fffffffd240, argc=argc@entry=19,
argv=0x1d766e0, argv@entry=0x7fffffffd348) at
[...]/source-gcc/gcc/toplev.c:2446
    #13 0x0000000000595ac7 in main (argc=19, argv=0x7fffffffd348) at
[...]/source-gcc/gcc/main.c:39
    (gdb) frame 3
    #3  lto_input_tree_1 (ib=0x7fffffffd080, data_in=0x1dd4860, tag=<optimized
out>, hash=<optimized out>) at [...]/source-gcc/gcc/lto-streamer-in.c:1561
    1561          gcc_assert (TYPE_PRECISION (type) <=
MAX_BITSIZE_MODE_ANY_INT);
    (gdb) print type
    $1 = (tree) 0x7ffff67b4960
    (gdb) call debug_tree(type)
     <error_mark 0x7ffff67b4960>

'libgomp.oacc-c-c++-common/firstprivate-mappings-1.c':

    lto1: internal compiler error: tree check: expected class 'type', have
'exceptional' (error_mark) in lto_fixup_prevailing_type, at
lto/lto-common.c:993

    #0  internal_error (gmsgid=0x163f958 "tree check: expected class %qs, have
%qs (%s) in %s, at %s:%d") at [...]/source-gcc/gcc/diagnostic.c:1706
    #1  0x000000000055f2fd in tree_class_check_failed (node=0x7ffff67b4960,
cl=<optimized out>, file=<optimized out>, line=<optimized out>,
function=<optimized out>) at [...]/source-gcc/gcc/tree.c:9738
    #2  0x00000000005d7724 in tree_class_check (__g=<optimized out>, __l=993,
__f=0x1559cc8 "[...]/source-gcc/gcc/lto/lto-common.c", __class=tcc_type,
__t=<optimized out>) at [...]/source-gcc/gcc/tree.h:3424
    #3  lto_fixup_prevailing_type (t=0x7ffff6883b28) at
[...]/source-gcc/gcc/lto/lto-common.c:993
    #4  process_new_tree (t=t@entry=0x7ffff6883b28, hm=hm@entry=0x7fffffffd0b0,
index=index@entry=467, total=total@entry=0x7fffffffd048,
data_in=data_in@entry=0x1e37560) at [...]/source-gcc/gcc/lto/lto-common.c:1817
    #5  0x00000000005d8e24 in lto_read_decls (resolutions=..., data=0x1e44f50,
decl_data=0x7ffff7fee090) at [...]/source-gcc/gcc/lto/lto-common.c:1938
    #6  lto_file_finalize (file=0x1d8bc30, order=0, file_data=0x7ffff7fee090)
at [...]/source-gcc/gcc/lto/lto-common.c:2267
    #7  lto_create_files_from_ids (order=0, count=<synthetic pointer>,
file_data=0x7ffff7fee090, file=0x1d8bc30) at
[...]/source-gcc/gcc/lto/lto-common.c:2277
    #8  lto_file_read (count=<synthetic pointer>, resolution_file=<optimized
out>, file=<optimized out>) at [...]/source-gcc/gcc/lto/lto-common.c:2332
    #9  read_cgraph_and_symbols (nfiles=<optimized out>, fnames=<optimized
out>) at [...]/source-gcc/gcc/lto/lto-common.c:2784
    #10 0x00000000005b9c23 in lto_main () at [...]/source-gcc/gcc/lto/lto.c:625
    #11 0x0000000000b7f6ff in compile_file () at
[...]/source-gcc/gcc/toplev.c:458
    #12 0x000000000059302c in do_compile () at
[...]/source-gcc/gcc/toplev.c:2307
    #13 toplev::main (this=this@entry=0x7fffffffd250, argc=argc@entry=18,
argv=0x1d76a30, argv@entry=0x7fffffffd358) at
[...]/source-gcc/gcc/toplev.c:2446
    #14 0x0000000000595ac7 in main (argc=18, argv=0x7fffffffd358) at
[...]/source-gcc/gcc/main.c:39
    (gdb) frame 3
    #3  lto_fixup_prevailing_type (t=0x7ffff6883b28) at
[...]/source-gcc/gcc/lto/lto-common.c:993
    993           TYPE_NEXT_PTR_TO (t) = TYPE_POINTER_TO (TREE_TYPE (t));
    (gdb) call debug_tree(t)
     <pointer_type 0x7ffff6883b28 type <error_mark 0x7ffff67b4960>
        unsigned DI
        size <integer_cst 0x7ffff67b4978 type <integer_type 0x7ffff67b50a8
bitsizetype> constant 64>
        unit-size <integer_cst 0x7ffff67b4990 type <integer_type 0x7ffff67b5000
sizetype> constant 8>
        align:64 warn_if_not_align:0 symtab:0 alias-set -1 structural-equality>

'libgomp.oacc-c++/firstprivate-mappings-1.C':

    lto1: internal compiler error: tree check: expected class 'type', have
'exceptional' (error_mark) in lto_fixup_prevailing_type, at
lto/lto-common.c:993

    #0  internal_error (gmsgid=0x163f958 "tree check: expected class %qs, have
%qs (%s) in %s, at %s:%d") at [...]/source-gcc/gcc/diagnostic.c:1706
    #1  0x000000000055f2fd in tree_class_check_failed (node=0x7ffff67b4960,
cl=<optimized out>, file=<optimized out>, line=<optimized out>,
function=<optimized out>) at [...]/source-gcc/gcc/tree.c:9738
    #2  0x00000000005d7849 in tree_class_check (__g=<optimized out>, __l=999,
__f=0x1559cc8 "[...]/source-gcc/gcc/lto/lto-common.c", __class=tcc_type,
__t=<optimized out>) at [...]/source-gcc/gcc/tree.h:3424
    #3  lto_fixup_prevailing_type (t=0x7ffff687fbd0) at
[...]/source-gcc/gcc/lto/lto-common.c:999
    #4  process_new_tree (t=t@entry=0x7ffff687fbd0, hm=hm@entry=0x7fffffffd0b0,
index=index@entry=293, total=total@entry=0x7fffffffd048,
data_in=data_in@entry=0x1e37560) at [...]/source-gcc/gcc/lto/lto-common.c:1817
    #5  0x00000000005d8e24 in lto_read_decls (resolutions=..., data=0x1e44f50,
decl_data=0x7ffff7fee090) at [...]/source-gcc/gcc/lto/lto-common.c:1938
    #6  lto_file_finalize (file=0x1d8bc30, order=0, file_data=0x7ffff7fee090)
at [...]/source-gcc/gcc/lto/lto-common.c:2267
    #7  lto_create_files_from_ids (order=0, count=<synthetic pointer>,
file_data=0x7ffff7fee090, file=0x1d8bc30) at
[...]/source-gcc/gcc/lto/lto-common.c:2277
    #8  lto_file_read (count=<synthetic pointer>, resolution_file=<optimized
out>, file=<optimized out>) at [...]/source-gcc/gcc/lto/lto-common.c:2332
    #9  read_cgraph_and_symbols (nfiles=<optimized out>, fnames=<optimized
out>) at [...]/source-gcc/gcc/lto/lto-common.c:2784
    #10 0x00000000005b9c23 in lto_main () at [...]/source-gcc/gcc/lto/lto.c:625
    #11 0x0000000000b7f6ff in compile_file () at
[...]/source-gcc/gcc/toplev.c:458
    #12 0x000000000059302c in do_compile () at
[...]/source-gcc/gcc/toplev.c:2307
    #13 toplev::main (this=this@entry=0x7fffffffd250, argc=argc@entry=18,
argv=0x1d76a30, argv@entry=0x7fffffffd358) at
[...]/source-gcc/gcc/toplev.c:2446
    #14 0x0000000000595ac7 in main (argc=18, argv=0x7fffffffd358) at
[...]/source-gcc/gcc/main.c:39
    (gdb) frame 3
    #3  lto_fixup_prevailing_type (t=0x7ffff687fbd0) at
[...]/source-gcc/gcc/lto/lto-common.c:999
    999           TYPE_NEXT_REF_TO (t) = TYPE_REFERENCE_TO (TREE_TYPE (t));
    (gdb) call debug_tree(t)
     <reference_type 0x7ffff687fbd0 type <error_mark 0x7ffff67b4960>
        unsigned DI
        size <integer_cst 0x7ffff67b4978 type <integer_type 0x7ffff67b50a8
bitsizetype> constant 64>
        unit-size <integer_cst 0x7ffff67b4990 type <integer_type 0x7ffff67b5000
sizetype> constant 8>
        align:64 warn_if_not_align:0 symtab:0 alias-set -1 structural-equality>

This is with '--enable-checking=yes,extra,rtl'.

I have not assessed whether there are other code paths that may fail in similar
ways.

Note that commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for
__builtin_bswap128" didn't actually touch the LTO streaming code.  From a very
quick look, I cannot tell why/where gcn vs. nvptx back ends would be different
in 128-bit data type handling, which I presume is where this stray 'error_mark'
is coming from, indirectly.

Testing current commit 1b76341441da344a0738edffdfab5d6307311ff2, I'm confirming
that reverting commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for
__builtin_bswap128", and as a pre-requisite commit
0bd4508f86889995f6864bf50773bd6859951007 "Fix bits/bytes confusion in latest
change to expand_unop" gets us rid of the ICE regressions.

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[ams at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

--- Comment #1 from Andrew Stubbs <ams at gcc dot gnu.org> ---
GCN uses TImode for a few special purposes, but lacks real TImode support.

(Basically, it allows TImode loads and stores for the SLP fake vectorization,
and there's one instruction that needs two DImode values in consecutive
registers.)

Probably this is a GCN-specific issue, and we should be finding another way to
handle the quasi-128-bit values.

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[ebotcazou at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-06-17
             Status|UNCONFIRMED                 |NEW

--- Comment #2 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
The error_mark_node very likely comes from:

DEF_PRIMITIVE_TYPE (BT_UINT128, uint128_type_node
                                ? uint128_type_node
                                : error_mark_node)

but this is a standard idiom in builtin-types.def so rather strange indeed.

I gather that targetm.scalar_mode_supported_p (TImode) returns false for GCN?

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[ams at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

--- Comment #3 from Andrew Stubbs <ams at gcc dot gnu.org> ---
The GCN port does not define a scalar_mode_supported, and I think the default
definition is allowing TImode (as long long int). As I said, the SLP
fake-vector load/store use it fine as a substitute for V4SI or V2DI.

Probably this is a bug we should fix. I've been planning to implement
additional vector sizes, so I can fold this fix into that work and not regress
anything in SLP (hopefully).

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[ams at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

--- Comment #4 from Andrew Stubbs <ams at gcc dot gnu.org> ---
In fact default_scalar_mode_supported_p does return *false* for TImode (because
LONG_LONG_TYPE_SIZE == 64, and BITS_PER_WORD == 32).

Therefore int128_t does not exist, as far as users are concerned. I'm not sure
how that translates to the GCC internals, but trying to use TImode for anything
other than moves is not going to work on GCN.

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[ebotcazou at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

--- Comment #5 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
> Therefore int128_t does not exist, as far as users are concerned. I'm not
> sure how that translates to the GCC internals, but trying to use TImode for
> anything other than moves is not going to work on GCN.

Yes, the code precisely makes sure it is not used in this case, but it looks
like something causes it to be nevertheless streamed.

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

Thomas Schwinge <tschwinge at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |11.0
             Status|NEW                         |WAITING

--- Comment #6 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
The "lto1: internal compiler error"s that I had originally mentioned here got
resolved by recent commit g:8d0b2b33748014ee57973c1d7bc9fd7706bb3da9 "amdgcn:
Enable TImode".

Does anything else remain to be done here?

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.0                        |11.2

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 11.1 has been released, retargeting bugs to GCC 11.2.

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.2                        |11.3

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 11.2 is being released, retargeting bugs to GCC 11.3

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.3                        |11.4

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 11.3 is being released, retargeting bugs to GCC 11.4.

[Bug target/95730] GCN offloading ICEs after commit fe7ebef7fe4f9acb79658ed9db0749b07efc3105 "Add support for __builtin_bswap128"

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95730

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.4                        |11.5

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 11.4 is being released, retargeting bugs to GCC 11.5.