Re: [GIT PULL] x86 fixes

[Linus Torvalds]

On Wed, Aug 19, 2015 at 3:00 AM, H. Peter Anvin <hpa@xxxxxxxxx> wrote:
>
> And I bet if CPUID actually reported the right thing it probably would work
> okay.  As I said, I tested this under Qemu which reported an accurate (lack
> of) CPUID for a 486SX.

While I agree that cpuid is a problem for FPU emulation on modern
CPU's, if this is due to "fucomip" then I think it's just that modern
distributions (and not-so-modern ones, for that matter) are compiled
with i686 support, so gcc just generates fucomip directly. So it's
just "plain FPU" code (no mmx, nothing like that), but it still fails.

The "set regular integer flags instructions" versions of floating
point compares are some of the bigger improvements to the legacy i87
instruction set, because the sequences to do FP compares without them
are just insane. I forget the details, but it's something like "store
status word to ax, then use sahf to get it into the flags register".
Crazy crazy crap. So it's no wonder that gcc wants to use a i686-only
instruction even for just regular FP code if at all possible.

I suspect it shouldn't be that hard to add f[u]compi[p] support to the
emulator.

But I also expect that most modern distributions are likely fairly
eager to use mmx etc, which sounds like a major pain to emulate.

                Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/