[Bug rtl-optimization/60466] New: Support for HARD_REGNO_NREGS_HAS_PADDING and HARD_REGNO_NREGS_WITH_PADDING broken

[Bug rtl-optimization/60466] New: Support for HARD_REGNO_NREGS_HAS_PADDING and HARD_REGNO_NREGS_WITH_PADDING broken

[rsandifo at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60466

            Bug ID: 60466
           Summary: Support for HARD_REGNO_NREGS_HAS_PADDING and
                    HARD_REGNO_NREGS_WITH_PADDING broken
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rsandifo at gcc dot gnu.org

It looks like HARD_REGNO_NREGS_HAS_PADDING and HARD_REGNO_NREGS_WITH_PADDING
no longer work correctly, possibly since subreg_get_info was introduced.
I doubt anyone cares since it's only used for long double on -m32 with
-m128bit-long-double, but just recording it anyway.

Test case:

void foo (void)
{
  _Complex long double x;
  asm volatile ("" : "=r" (x));
  asm volatile ("" :: "r" (__imag__ x), "r" (__real__ x));
}

When compiled on x86_64-linux-gnu with -m32 -m128bit-long-double -O2
it fails the second of these asserts in subreg_get_info:

  /* This should always pass, otherwise we don't know how to verify
     the constraint.  These conditions may be relaxed but
     subreg_regno_offset would need to be redesigned.  */
  gcc_assert ((GET_MODE_SIZE (xmode) % GET_MODE_SIZE (ymode)) == 0);
  gcc_assert ((nregs_xmode % nregs_ymode) == 0);

nregs_xmode is set to the number of registers that there would be if
no padding were present:

      nregs_xmode = HARD_REGNO_NREGS_WITH_PADDING (xregno, xmode);

But nregs_ymode is always set to the "real" number (without padding):

  nregs_ymode = hard_regno_nregs[xregno][ymode];

nregs_xmode is used here:

  offset_adj -= subreg_lowpart_offset (ymode,
                  mode_for_size (GET_MODE_BITSIZE (xmode) / nregs_xmode,
                                 MODE_INT, 0));

where using the padded value looks correct.  The division will give
a 16-byte mode, matching ymode's size, and the lowpart offset will be 0.
But later we have:

  mode_multiple = GET_MODE_SIZE (xmode) / GET_MODE_SIZE (ymode);
  ...
  nregs_multiple = nregs_xmode / nregs_ymode;
  ...

  info->offset = (y_offset / (mode_multiple / nregs_multiple)) * nregs_ymode;

where it doesn't look correct to mix padded and unpadded register counts.

[Bug middle-end/60466] Support for HARD_REGNO_NREGS_HAS_PADDING and HARD_REGNO_NREGS_WITH_PADDING broken

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=60466

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|                            |4.1.2, 4.4.7
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-03-15

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.