[Bug target/102024] New: zero width bitfields and ABIs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102024

            Bug ID: 102024
           Summary: zero width bitfields and ABIs
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
  Target Milestone: ---

My r12-2975-g32c3a75390623a0470df52af13f78baddd562981 change removed removal of
zero width unnamed bitfields from the C++ FE, the C FE has never been removing
those (the removal was certainly premature, as the newly added traits need to
see the zero width unnamed bitfields).
This unfortunately can affect ABI on several targets.
>From quick grep I think this can't affect
alpha, arc, avr, bfin, bpf, c6x, cr16, cris, csky, epiphany, fr30, frv, ft32,
gcn, h8300, lm32, m32c, m32r, m68k, mcore, microblaze, mmix, mn10300, moxie,
msp430, nds32, nios2, nvptx, or1k, pa, pdp11, pru, rl78, rx, stormy16, tilegx,
tilepro, v850, vax, visium, xtensa
are all safe, they either don't mention TYPE_FIELDS in their backends or use it
either in va_list or in structure layout code only (structure layout happens
before the zero width bitfields are removed by the FE).
TYPE_MODE should be ok too, it is just passing of arguments by value and
returning by value that could be problematic on
aarch64, arm, i386, ia64, iq2000, mips, riscv, rs6000, s390, sh and sparc
I think sh is ok, it ignores integer_zerop (DECL_SIZE (field)).

riscv apparently has old (< GCC 10) and new (>= GCC 10) ABI and the old one
didn't ignore them and made C and C++ ABI incompatible, while the new one does
ignore them.  So, my patch probably broke the 
"ABI for flattened struct with zero-length bit-fields changed in GCC 10"
warning (why isn't it guarded with -Wpsabi?), perhaps we should mark the C++
zero width bitfields with some attribute or flag so that backends could emit
correct -Wpsabi warnings?

x86_64 seems to be affected:
struct S { float a; int : 0; float b; };

__attribute__((noipa)) struct S
foo (struct S x)
{
  return x;
}

void
bar (void)
{
  struct S s = { 0.0f, 0.0f };
  foo (s);
}

With GCC 10 and C we emit:
        xorl    %edi, %edi
        jmp     foo
With C++ we emit:
        pxor    %xmm0, %xmm0
        jmp     _Z3foo1S
with current trunk
        xorl    %edi, %edi
        jmp     _Z3foo1S