[Bug target/115161] [15 Regression] highway-1.0.7 miscompilation of some SSE2 intrinsics

["amonakov at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115161

--- Comment #20 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #19)
> If we guarantee that we never constant fold FIX/UNSIGNED_FIX with
> -ftrapping-math (we shouldn't, as the exceptions should be raised), then
> using FIX/UNSIGNED_FIX is ok in
> flag_trapping_math guarded patterns even if the intrinsics have some
> specific value they want in those cases, because they will never be folded.

No, this sounds illogical. The compiler may fold out-of-range conversions just
fine, it just needs to preserve the original operation (with its result now
dead), or use any other opportunity to ensure that exceptions are still raised.
Specifically, it is okay to transform

float f;
int i;

f = 0x0.8p+33f;
i = (int) f;

to

f = 0x0.8p+33f;
dummy = (int) f;
i = INT_MAX; // or whatever

The "known" value of 'i' may then facilitate further folding.