Re: Build regressions/improvements in v3.4-rc7

On Wed, May 16, 2012 at 5:46 PM, James Bottomley
[off-list ref] wrote:

On Wed, 2012-05-16 at 10:47 +0200, Geert Uytterhoeven wrote:

quoted

On Wed, May 16, 2012 at 10:30 AM, Geert Uytterhoeven
[off-list ref] wrote:
=C2=A0 + lib/mpi/generic_mpih-mul1.c: error: inconsistent operand
constraints in an 'asm': =C2=A0=3D> 50:70
=C2=A0 + lib/mpi/generic_mpih-mul2.c: error: inconsistent operand
constraints in an 'asm': =C2=A0=3D> 49:70
=C2=A0 + lib/mpi/generic_mpih-mul3.c: error: inconsistent operand
constraints in an 'asm': =C2=A0=3D> 49:70
=C2=A0 + lib/mpi/mpih-div.c: error: inconsistent operand constraints in =

an

quoted

'asm': =C2=A0=3D> 135:122, 135:371, 97:122, 106:121, 106:370, 97:371

parisc-allmodconfig

Wow, lib/mpi/ is a complete horror: it's full of hand crafted asm code.
The error in this case appears to be that umul_ppm() is implemented as
an xmpyu instruction. =C2=A0That's a floating point instruction. =C2=A0We
deliberately compile the kernel with floating point disabled because we
don't want to save and restore the floating point register file on each
context switch, hence the operand constraints are unsatisfiable.

It appears to be completely untested on non-x86 and to have been
imported via the security tree ... what are we supposed to do with this?
I thought the general principle was that asm code was really supposed to
be confined to the arch directories?

James

Yes.. it is suxx.
But it is only longlong.h and inherited from gpg..
And now there is a stuff such as /include/math-emu/soft-fp and
<stdlib/longlong.h>
I planned to look to that but now I will start doing it right away.

- Dmitry