[Bug target/83292] New: __builtin_apply() triggers x87 stack exception on 32-bit x86

["achurch+gcc at achurch dot org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83292

            Bug ID: 83292
           Summary: __builtin_apply() triggers x87 stack exception on
                    32-bit x86
           Product: gcc
           Version: 7.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: achurch+gcc at achurch dot org
  Target Milestone: ---

Calling a function with __builtin_apply() on 32-bit x86 causes the next x87
instruction to raise an x87 stack exception, as a side effect of mixing MMX
instructions (saving %mm[0-2]) and x87 instructions (fstp after the function
call, I presume to pop possible return values off the x87 stack).

Note that it's not possible to completely work around this with -mfpmath=sse
because fild is still used to load a 64-bit integer (s/int/long long/ in the
test case below).

The problem occurs at least as far back as GCC 5.4.0; I haven't tested earlier
versions.

Editing the generated assembly to bracket the __builtin_apply() code with
fstenv/fldenv seems to fix the problem.

Testcase:
==========================
void foo(void)
{
}

int main(void)
{
    __builtin_apply(foo, __builtin_apply_args(), 0);
    volatile int i = 1;
    float f = i;
    return f==1.0f ? 0 : 1;
}
==========================