https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64554
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #6)
> (In reply to charles from comment #5)
> > No you're right.
> >
> > I could add, though, that the compiler should have said something.
> >
> > Also, why did it compile just fine with -O0?
>
> Sounds like a bug (we should error immediately here IMHO). Probably
> needs to annotate the CIF codes with whether this is "fatal" (won't
> change with other inlining, IPA or LTO).
>
> OTOH the headers are also poorly designed to trigger this kind of error.
>
> Maybe we want to support __attribute__((always_inline("ISA not enabled")))
> instead to be able to print a custom error message?
The difference is that for __OPTIMIZE__ _mm_clmulepi64_si128 is implemented
as always_inline function (surrounded by #pragma GCC target("pclmul")), so for
-O1+ the inlining error is expected, such function can't be inlined into the
current function, and always_inline mandates that it is inlined (without
always_inline it would simply not be inlined, and the out of line
_mm_clmulepi64_si128 would use the pclmul ISA, while the caller would not).
For -O0, this is implemented using a macro, since the last argument must be a
constant, and the reason you get no error is that the function is const and you
don't use the result, so it is DCEd.
If you actually use the result (at -O0 it is enough to store it into say
__m128i c = _mm_clmulepi64_si128(a, b, 1); , for -O1 you'd need something
better, store into volatile var, or global var or otherwise make sure it isn't
DCEd), then you get this diagnosed:
/tmp/z.c: In function ‘main’:
/tmp/z.c:6:11: error: ‘__builtin_ia32_pclmulqdq128’ needs isa option -m32
-mpclmul
__m128i c = _mm_clmulepi64_si128(a, b, 1);
^
So I don't see any bug on the gcc side here.
