[Bug c/103031] [12 Regression] Missing static initializer folding with -frounding-math

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103031

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-11-02
           Priority|P3                          |P1
             Status|UNCONFIRMED                 |NEW
   Target Milestone|---                         |12.0
     Ever confirmed|0                           |1
          Component|middle-end                  |c

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I suppose the frontend is calling fold_convert directly rather than
fold_build1_initializer (FLOAT_EXPR, ...).

The frontend ends up building this via

#0  0x0000000000d53d87 in convert_to_real_1 (
    type=<real_type 0x7ffff6566348 double>, expr=<integer_cst 0x7ffff6544540>, 
    fold_p=true) at /home/rguenther/src/gcc3/gcc/convert.c:319
#1  0x0000000000d53f0a in convert_to_real (
    type=<real_type 0x7ffff6566348 double>, expr=<integer_cst 0x7ffff6544540>)
    at /home/rguenther/src/gcc3/gcc/convert.c:347
#2  0x0000000000b3cb90 in convert (type=<real_type 0x7ffff6566348 double>, 
    expr=<integer_cst 0x7ffff6544540>)
    at /home/rguenther/src/gcc3/gcc/c/c-convert.c:138
#3  0x0000000000b9f602 in convert_and_check (loc=242464, 
    type=<real_type 0x7ffff6566348 double>, expr=<integer_cst 0x7ffff6544540>)
    at /home/rguenther/src/gcc3/gcc/c-family/c-common.c:1767
#4  0x0000000000b1055e in convert_for_assignment (location=242464, expr_loc=0, 
    type=<real_type 0x7ffff6566348 double>, rhs=<integer_cst 0x7ffff6544540>, 
    origtype=<tree 0x0>, errtype=ic_init, null_pointer_constant=false, 
    fundecl=<tree 0x0>, function=<tree 0x0>, parmnum=0, warnopt=0)
    at /home/rguenther/src/gcc3/gcc/c/c-typeck.c:7032
#5  0x0000000000b18c6e in digest_init (init_loc=242464, 
    type=<real_type 0x7ffff6566348 double>, init=<integer_cst 0x7ffff6544540>, 
    origtype=<tree 0x0>, null_pointer_constant=false, strict_string=true, 
    require_constant=1) at /home/rguenther/src/gcc3/gcc/c/c-typeck.c:8217
#6  0x0000000000b16df8 in store_init_value (init_loc=242464, 
    decl=<var_decl 0x7ffff7ff5c60 x>, init=<integer_cst 0x7ffff6544540>, 
--Type <RET> for more, q to quit, c to continue without paging--
    origtype=<tree 0x0>) at /home/rguenther/src/gcc3/gcc/c/c-typeck.c:7762
#7  0x0000000000ace0e7 in finish_decl (decl=<var_decl 0x7ffff7ff5c60 x>, 
    init_loc=242464, init=<integer_cst 0x7ffff6544540>, origtype=<tree 0x0>, 
    asmspec_tree=<tree 0x0>) at /home/rguenther/src/gcc3/gcc/c/c-decl.c:5363

maybe convert_for_assignment needs to do something similar as
START/END_FOLD_INIT around the convert_and_check or convert_for_assignment
is not the correct thing to call here.  It's also odd that convert_to_real
says fold_p == true to convert_to_real_1 but that function doesn't actually
fold anything (maybe that's just all dead code).  I'm quite sure that
initializer_constant_valid_p isn't the one that has to accept a FLOAT_EXPR
here.

C frontend maintainers need to sort this out.