On Tue, Nov 03, 2020 at 05:44:35PM +0100, Borislav Petkov wrote:
> On Mon, Nov 02, 2020 at 05:18:09PM -0600, Segher Boessenkool wrote:
> > That is invalid actually: local register asm as input to an inline asm
> > should use *that* register!
> >
> > This is all correct until LRA ("reload"). Not that "movl %xmm0,$eax"
> > works, but at least it screams its head off, as it should.
>
> Screams how?
$ cat xmm1.s
movl %xmm0,%eax
$ x86_64-linux-as xmm1.s -o xmm1.o
xmm1.s: Assembler messages:
xmm1.s:1: Error: unsupported instruction `mov'
(This isn't an existing insn IIUC.)
> It builds fine without a single peep with -Wall here.
>
> Btw, that's a MOVD - not a MOVL. MOVD can do xmm -> gpr moves. And
> singlestepping it with gdb does, well, something, which is clearly
> wrong but nothing complains:
>
> => 0x555555555131 <main+12>: movd %xmm0,%eax
>
> and %xmm0 has:
>
> (gdb) p $xmm0
> $2 = {v4_float = {0.99000001, 0, 0, 0}, v2_double = {5.2627153433055495e-315,
> 0}, v16_int8 = {-92, 112, 125, 63,
> ^^^^^^^^^^
>
> so that is correct.
The original code had movl. And movl is needed for GPRs.
> and that same value goes into %r8d:
>
> mov %eax,%r8d
Which violates what is required by register asm :-(
> > Yes. But GCC doing what you should have said instead of doing what you
> > said, is not good.
>
> Oh well, should I open a low prio bug, would that help?
Sure, thanks!
> I probably should test with the latest gcc first, though...
Yeah... FWIW, I tested with
x86_64-linux-gcc (GCC) 11.0.0 20201015 (experimental)
so I doubt current ToT will have it fixed, but who knows.
Thanks,
Segher
