[Bug c++/104052] New: Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[Bug c++/104052] New: Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

            Bug ID: 104052
           Summary: Modules ICE on powerpc64le-linux with
                    -mabi=ieeelongdouble
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
  Target Milestone: ---

In gcc configured --with-long-double-format=ieee (but also just when running
make check-g++ RUNTESTFLAGS='--target_board=unix/-mabi=ieeelongdouble
modules.exp'
I'm seeing all the tests ICE like:
gcc/testsuite/g++.dg/modules/global-3_a.C:5:8: internal compiler error: in
type_node, at cp/module.cc:8721
0x10218fcb trees_out::type_node(tree_node*)
        ../../gcc/cp/module.cc:8721
0x10216f57 trees_out::tree_node(tree_node*)
        ../../gcc/cp/module.cc:9049
0x10218ecb trees_out::type_node(tree_node*)
        ../../gcc/cp/module.cc:8715
0x10216f57 trees_out::tree_node(tree_node*)
        ../../gcc/cp/module.cc:9049
0x10217b0f trees_out::core_vals(tree_node*)
        ../../gcc/cp/module.cc:6003
0x1021880f trees_out::tree_node_vals(tree_node*)
        ../../gcc/cp/module.cc:7062
0x1021894f trees_out::fn_parms_init(tree_node*)
        ../../gcc/cp/module.cc:9981
0x10215923 trees_out::decl_value(tree_node*, depset*)
        ../../gcc/cp/module.cc:7626
0x1024ab47 depset::hash::find_dependencies(module_state*)
        ../../gcc/cp/module.cc:13128
0x1024beef module_state::write(elf_out*, cpp_reader*)
        ../../gcc/cp/module.cc:17618
0x112c25cb finish_module_processing(cpp_reader*)
        ../../gcc/cp/module.cc:19874
0x10911f93 c_parse_final_cleanups()
        ../../gcc/cp/decl2.c:5311
0x11274c23 c_common_parse_file()
        ../../gcc/c-family/c-opts.c:1260
where the innermost type_node is called on
 <real_type 0x7fffe82f0b28 __ibm128 type_6 IF
    size <integer_cst 0x7fffe8131230 type <integer_type 0x7fffe82200a8
bitsizetype> constant 128>
    unit-size <integer_cst 0x7fffe8131248 type <integer_type 0x7fffe8220000
sizetype> constant 16>
    align:128 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7fffe82f0b28 precision:128
    pointer_to_this <pointer_type 0x7fffdb9680a0> reference_to_this
<reference_type 0x7fffddba1848>>
But I don't really see type_node called on other REAL_TYPEs, e.g. with
-mabi=ibmlongdouble

[Bug c++/104052] Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu.org

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
          ibm128_float_type_node = make_node (REAL_TYPE);
          TYPE_PRECISION (ibm128_float_type_node) = 128;
          SET_TYPE_MODE (ibm128_float_type_node, IFmode);
          layout_type (ibm128_float_type_node);


  long_double_type_node = make_node (REAL_TYPE);
  TYPE_PRECISION (long_double_type_node) = LONG_DOUBLE_TYPE_SIZE;
  layout_type (long_double_type_node);

I don't see any difference there either.
I don't see any reference to long_double_type_node in module.cc either. Nor
TYPE_MODE.

[Bug c++/104052] Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I bet we don't reach this because modules.cc has:
/* Global trees.  */
static const std::pair<tree *, unsigned> global_tree_arys[] =
  {
    std::pair<tree *, unsigned> (sizetype_tab, stk_type_kind_last),
    std::pair<tree *, unsigned> (integer_types, itk_none),
    std::pair<tree *, unsigned> (global_trees, TI_MODULE_HWM),
    std::pair<tree *, unsigned> (c_global_trees, CTI_MODULE_HWM),
    std::pair<tree *, unsigned> (cp_global_trees, CPTI_MODULE_HWM),
    std::pair<tree *, unsigned> (NULL, 0)
  };
and special cases those somehow.  So e.g. long_double_type_node and others
#define float_type_node                 global_trees[TI_FLOAT_TYPE]
#define double_type_node                global_trees[TI_DOUBLE_TYPE]
#define long_double_type_node           global_trees[TI_LONG_DOUBLE_TYPE]
or
#define float128_type_node              global_trees[TI_FLOAT128_TYPE]
all go through that.  But modules.cc doesn't know about types the backend
creates
like __ibm128.  __int24 etc. is probably ok due to integer_types
containing
  itk_intN_0,
  itk_unsigned_intN_0,
  itk_intN_1,
  itk_unsigned_intN_1,
  itk_intN_2,
  itk_unsigned_intN_2,
  itk_intN_3,
  itk_unsigned_intN_3,
So, one way through this would be to reserve one or more global_trees for the
extra backend floating point types.

[Bug c++/104052] Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
config/aarch64/aarch64-builtins.c:  aarch64_fp16_type_node = make_node
(REAL_TYPE);
config/aarch64/aarch64-builtins.c:  aarch64_bf16_type_node = make_node
(REAL_TYPE);
config/arm/arm-builtins.c:  arm_bf16_type_node = make_node (REAL_TYPE);
config/arm/arm-builtins.c:  arm_fp16_type_node = make_node (REAL_TYPE);
config/csky/csky.c:  static tree csky_floatHF_type_node = make_node
(REAL_TYPE);
config/i386/i386-builtins.c:      ix86_float16_type_node = make_node
(REAL_TYPE);
config/i386/i386-builtins.c:      float80_type_node = make_node (REAL_TYPE);
config/ia64/ia64.c:  fpreg_type = make_node (REAL_TYPE);
config/ia64/ia64.c:      float80_type = make_node (REAL_TYPE);
config/rs6000/rs6000-call.c:      ibm128_float_type_node = make_node
(REAL_TYPE);

is all I could find, now the question is what from these types are already in
some global_trees.  But the upper bound for target specific floating point
types will be most likely either 2 or 1.

[Bug c++/104052] Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
There are other types it messes up with too. E.g. the opaque types (PR 98688).
I noticed vector types are not handled correctly either. It assumes all vector
types of the same size are the same which is definitely not true for aarch64.

[Bug c++/104052] Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2022-01-16
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Seems like when register_builtin_type is called, there should be another vector
holding those types and handling those in module.cc too. That would fix this
and the vector type issues I think.

[Bug c++/104052] Modules ICE on powerpc64le-linux with -mabi=ieeelongdouble

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104052

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |DUPLICATE
             Status|NEW                         |RESOLVED

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Dup.

*** This bug has been marked as a duplicate of bug 98645 ***