[Bug libgcc/114131] New: std::isinf(std::float128_t) generates superfluous nan-checks

[Bug libgcc/114131] New: std::isinf(std::float128_t) generates superfluous nan-checks

[g.peterhoff@t-online.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114131

            Bug ID: 114131
           Summary: std::isinf(std::float128_t) generates superfluous
                    nan-checks
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libgcc
          Assignee: unassigned at gcc dot gnu.org
          Reporter: g.peterhoff@t-online.de
  Target Milestone: ---

please see https://godbolt.org/z/djc9q1vcv
test1(default): includes nan-checks (__unordtf2)
test2: no nan-checks, but calls __eqtf2
test3: only checks for inf (via bit_cast); no additional function calls +
branchfree. Of course, this only works if (unsigned) __int128 is available.

thx
Gero

[Bug libgcc/114131] std::isinf(std::float128_t) generates superfluous nan-checks

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114131

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-02-27
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
           Keywords|                            |missed-optimization

[Bug middle-end/114131] std::isinf(std::float128_t) generates superfluous nan-checks

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114131

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I am almost positive this is due to GCC default of trapping math here.

Adding  -fno-trapping-math GCC no longer has a check against NaN.

I think test2 can cause an fp exception to happen so it is almost definitely
not valid transformation unless with -fno-trapping-math.

As far as transforming test1 to test3, well it depends on if there are no
instructions to do the comparison and you are doing "soft" float for that type.
Basically we have:
```
  _1 = ABS_EXPR <x_3(D)>;
  _2 = _1 u<= 1.18973149535723176508575932662800701619646905264169404553e+4932;
  _4 = ~_2;
```

Which we could recognize in the middle-end to do the transformation if the FP
format is easy to handle (double double is one which is not).

[Bug middle-end/114131] std::isinf(std::float128_t) generates superfluous nan-checks

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114131

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

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

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Actually this is a fully a dup of bug 94083.

*** This bug has been marked as a duplicate of bug 94083 ***