[Bug target/105975] New: OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[Bug target/105975] New: OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

            Bug ID: 105975
           Summary: OpenMP/nvptx offloading: 'internal compiler error: in
                    maybe_legitimize_operand, at optabs.cc:7785'
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: openmp
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tschwinge at gcc dot gnu.org
                CC: jakub at gcc dot gnu.org, rsandifo at gcc dot gnu.org,
                    vries at gcc dot gnu.org
  Target Milestone: ---
            Target: nvptx

The recent commit r13-1068-g1d205dbac1e1754c01c22a31bd1688126545401e "Factor
out common internal-fn idiom" causes a class of ICEs in OpenMP/nvptx offloading
compilation: 'during RTL pass: expand', 'internal compiler error: in
maybe_legitimize_operand, at optabs.cc:7785', seen for a lot of libgomp
OpenMP/nvptx offloading test cases (with '-O1' and higher).

    0xb1b0b3 maybe_legitimize_operand
            [...]/source-gcc/gcc/optabs.cc:7785
    0xb1b0b3 maybe_legitimize_operands(insn_code, unsigned int, unsigned int,
expand_operand*)
            [...]/source-gcc/gcc/optabs.cc:7936
    0xb1b139 maybe_gen_insn(insn_code, unsigned int, expand_operand*)
            [...]/source-gcc/gcc/optabs.cc:7955
    0xb1a8b8 maybe_expand_insn(insn_code, unsigned int, expand_operand*)
            [...]/source-gcc/gcc/optabs.cc:7998
    0xb1a8b8 expand_insn(insn_code, unsigned int, expand_operand*)
            [...]/source-gcc/gcc/optabs.cc:8029
    0x95dcb3 expand_fn_using_insn
            [...]/source-gcc/gcc/internal-fn.cc:193
    0x6d3ee7 expand_call_stmt
            [...]/source-gcc/gcc/cfgexpand.cc:2737
    0x6d3ee7 expand_gimple_stmt_1
            [...]/source-gcc/gcc/cfgexpand.cc:3869

For extra entertainment: when running with '-wrapper "$GDB",-q,--args', we get
'[Inferior 1 (process [...]) exited normally]'...  (Maybe Valgrind could help? 
Unless someone directly pinpoints the issue, of course.)

I've not yet determined whether it's a latent problem just exposed by this
commit, or whether the commit itself has an issue.  It's not magically fixed by
the related subsequent commit
r13-1069-gf8baf4004ef965ce7a9edf6d2f5eb99adb15803a "Add a general mapping from
internal fns to target insns".

'gcc/internal-fn.cc':

    193    expand_insn (icode, opno, ops);

'gcc/optabs.cc':

    8026        expand_insn (enum insn_code icode, unsigned int nops,
    8027                     class expand_operand *ops)
    8028        {
    8029          if (!maybe_expand_insn (icode, nops, ops))

    7995        maybe_expand_insn (enum insn_code icode, unsigned int nops,
    7996                           class expand_operand *ops)
    7997        {
    7998          rtx_insn *pat = maybe_gen_insn (icode, nops, ops);

    7951        maybe_gen_insn (enum insn_code icode, unsigned int nops,
    7952                        class expand_operand *ops)
    7953        {
    7954          gcc_assert (nops == (unsigned int) insn_data[(int)
icode].n_generator_args);
    7955          if (!maybe_legitimize_operands (icode, 0, nops, ops))

    7935              /* Otherwise try legitimizing the operand on its own.  */
    7936              if (j == i && !maybe_legitimize_operand (icode, opno + i,
&ops[i]))

    7784            case EXPAND_OUTPUT:
    7785              gcc_assert (mode != VOIDmode);

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

--- Comment #1 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> ---
Could you point to a specific test and give command-line arguments?
I'm not set up to do an nvptx test run.

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

--- Comment #2 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Created attachment 53140
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53140&action=edit
'a-for-11.o.xz'

With commit 1d205dbac1e1754c01c22a31bd1688126545401e "Factor out common
internal-fn idiom" sources, I've run such an OpenMP/nvptx offloading
compilation ('libgomp.c-c++-common/for-11.c', '-O1') with '-v -save-temps', and
we see:

    [...]
    
/home/thomas/tmp/source/gcc/build/trace-slim-omp/build-gcc-offload-nvptx-none/gcc/lto1
-quiet -dumpbase ./a.xnvptx-none.mkoffload -m64 -mgomp -O1 -version
-fno-openacc -fno-pie -fcf-protection=none -foffload-abi=lp64 -fopenmp
@./a.xnvptx-none.mkoffload.args.0 -o ./a.xnvptx-none.mkoffload.s
    [...]
    during RTL pass: expand
    source-gcc/libgomp/testsuite/libgomp.c-c++-common/for-2.h: In function
‘f20_tpf_simd_static._omp_fn.1’:
    source-gcc/libgomp/testsuite/libgomp.c-c++-common/for-2.h:343:9: internal
compiler error: in maybe_legitimize_operand, at optabs.cc:7785
    [...]

    $ cat < a.xnvptx-none.mkoffload.args.0
    a-for-11.o

If you now build a nvptx offloading compiler (same commit
1d205dbac1e1754c01c22a31bd1688126545401e sources):

    $ ../source-gcc/configure --target=nvptx-none
--enable-as-accelerator-for=x86_64-pc-linux-gnu --enable-languages=lto
--enable-checking=yes,extra,rtl
    $ make -j16 all-gcc

..., you ought to be able to reproduce the ICE (using the 'a-for-11.o' I've
attached to this PR, 'xz'-compressed):

    $ gcc/lto1 -quiet -dumpbase ./a.xnvptx-none.mkoffload -m64 -mgomp -O1
-version -fno-openacc -fno-pie -fcf-protection=none -foffload-abi=lp64 -fopenmp
a-for-11.o -o ./a.xnvptx-none.mkoffload.s
    [...]
    during RTL pass: expand
    ../source-gcc/libgomp/testsuite/libgomp.c-c++-common/for-2.h: In function
‘f20_tpf_simd_static._omp_fn.1’:
    ../source-gcc/libgomp/testsuite/libgomp.c-c++-common/for-2.h:343:9:
internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785
    [...]

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

--- Comment #3 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Created attachment 53141
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53141&action=edit
'a.xnvptx-none.mkoffload.s.xz'

For cross-checking: if I now temporarily 'git revert
1d205dbac1e1754c01c22a31bd1688126545401e', and rebuilt things, I get the same
'a-for-11.o' file (with just some randomness in '.section' names; different
issue), and the nvptx 'gcc/lto1' command produces 'a.xnvptx-none.mkoffload.s'
as attached ('xz'-compressed).


(I'll soon be off of work for a few days.)

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rsandifo at gcc dot gnu.org
   Last reconfirmed|                            |2022-06-15

--- Comment #4 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> ---
Thanks for the reproducer.

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The trunk branch has been updated by Richard Sandiford <rsandifo@gcc.gnu.org>:

https://gcc.gnu.org/g:9d2fe6d427b37755410bb4eb4ecce8382f2ebfc5

commit r13-1111-g9d2fe6d427b37755410bb4eb4ecce8382f2ebfc5
Author: Richard Sandiford <richard.sandiford@arm.com>
Date:   Wed Jun 15 17:40:09 2022 +0100

    Revert recent internal-fn changes [PR105975]

    The recent internal-fn âclean-upsâ triggered problems on nvptx
    because some of the omp_simt_* patterns had modeless operands.
    I wondered about adapting expand_fn_using_insn to cope with that,
    but then the problem becomes: what should the mode of operand 0
    be when there is no lhs?  The answer depends on the target insn.
    For GOMP_SIMT_ENTER_ALLOC the answer was: use Pmode.
    For GOMP_SIMT_ORDERED_PRED and others the answer was: elide the call.
    (However, GOMP_SIMT_ORDERED_PRED doesn't seem to have ECF_* flags
    that would normally allow it to be dropped at the gimple level.)

    So these instructions seem to be special enough that they need
    their own code after all.  This patch reverts the second patch
    and most of the first.  The only part retained from the first
    is splitting expand_fn_using_insn out of expand_direct_optab_fn,
    since I think expand_fn_using_insn could still be useful in future.

    gcc/
            PR middle-end/105975
            Revert everything apart from the expand_fn_using_insn and
            expand_direct_optab_fn changes from:

            * internal-fn.def (DEF_INTERNAL_INSN_FN): New macro.
            (GOMP_SIMT_ENTER_ALLOC, GOMP_SIMT_EXIT, GOMP_SIMT_LANE)
            (GOMP_SIMT_LAST_LANE, GOMP_SIMT_ORDERED_PRED, GOMP_SIMT_VOTE_ANY)
            (GOMP_SIMT_XCHG_BFLY, GOMP_SIMT_XCHG_IDX): Use it.
            * internal-fn.h (direct_internal_fn_info::directly_mapped): New
            member variable.
            (direct_internal_fn_info::vectorizable): Reduce to 1 bit.
            (direct_internal_fn_p): Also return true for internal functions
            that map directly to instructions defined target-insns.def.
            (direct_internal_fn): Adjust comment accordingly.
            * internal-fn.cc (direct_insn, optab1, optab2, vectorizable_optab1)
            (vectorizable_optab2): New local macros.
            (not_direct): Initialize directly_mapped.
            (mask_load_direct, load_lanes_direct, mask_load_lanes_direct)
            (gather_load_direct, len_load_direct, mask_store_direct)
            (store_lanes_direct, mask_store_lanes_direct, vec_cond_mask_direct)
            (vec_cond_direct, scatter_store_direct, len_store_direct)
            (vec_set_direct, unary_direct, binary_direct, ternary_direct)
            (cond_unary_direct, cond_binary_direct, cond_ternary_direct)
            (while_direct, fold_extract_direct, fold_left_direct)
            (mask_fold_left_direct, check_ptrs_direct): Use the macros above.
            (expand_GOMP_SIMT_ENTER_ALLOC, expand_GOMP_SIMT_EXIT): Delete
            (expand_GOMP_SIMT_LANE, expand_GOMP_SIMT_LAST_LANE): Likewise;
            (expand_GOMP_SIMT_ORDERED_PRED, expand_GOMP_SIMT_VOTE_ANY):
Likewise.
            (expand_GOMP_SIMT_XCHG_BFLY, expand_GOMP_SIMT_XCHG_IDX): Likewise.
            (direct_internal_fn_types): Handle functions that map to
instructions
            defined in target-insns.def.
            (direct_internal_fn_types): Likewise.
            (direct_internal_fn_supported_p): Likewise.
            (internal_fn_expanders): Likewise.

            (expand_fn_using_insn): New function,
            split out and adapted from...
            (expand_direct_optab_fn): ...here.
            (expand_GOMP_SIMT_ENTER_ALLOC): Use it.
            (expand_GOMP_SIMT_EXIT): Likewise.
            (expand_GOMP_SIMT_LANE): Likewise.
            (expand_GOMP_SIMT_LAST_LANE): Likewise.
            (expand_GOMP_SIMT_ORDERED_PRED): Likewise.
            (expand_GOMP_SIMT_VOTE_ANY): Likewise.
            (expand_GOMP_SIMT_XCHG_BFLY): Likewise.
            (expand_GOMP_SIMT_XCHG_IDX): Likewise.

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #6 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> ---
Should be fixed now.

[Bug target/105975] OpenMP/nvptx offloading: 'internal compiler error: in maybe_legitimize_operand, at optabs.cc:7785'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105975

--- Comment #7 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
(In reply to rsandifo@gcc.gnu.org from comment #4)
> Thanks for the reproducer.

Glad that this approach did work.


(In reply to rsandifo@gcc.gnu.org from comment #6)
> Should be fixed now.

ACK, thanks.  And sorry that you had to revert most of your changes!  But I'm
not familiar enough with these OpenMP/nvptx 'GOMP_SIMT_[...]' codes to easily
come up with an idea about how to make this work in your refactored code.