[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: gcc-4.1 with fastmath
From: |
Samuel Thibault |
Subject: |
Re: gcc-4.1 with fastmath |
Date: |
Fri, 22 Dec 2006 19:58:55 +0100 |
User-agent: |
Mutt/1.5.11 |
Hi,
Roland McGrath, le Tue 05 Dec 2006 00:18:52 -0800, a écrit :
> gnumach doesn't support newer fpu state for mmx/xmm et al.
> would need new thread state flavor for "fpx" or whatever we call it.
> cf linux asm/user.h user_i387_struct vs user_fxsr_struct
> former is what mach can handle. need to handle latter in ctx sw + user
> thread state flavor, then enable cpu support bits (cr4 bit i think).
There's a bit of work to do both in the Hurd and glibc too, because of
GNU Mach's thread_get_state(i386_FLOAT_STATE) call.
- hurd/exec/elfcore.c:fetch_thread_fpregset(): in the core dump, fpx
state would probably be needed too.
- libc/sysdeps/mach/hurd/i386/trampoline.c:_hurd_setup_sighandler(): in
the sigcontext structure, fpx state would be useful too.
Both of these are not really _required_ for now, but we need to think
about the interface as soon as now. I guess modifying the behavior of
thread_get_state(i386_FLOAT_STATE) is out of question, so should we
define a new thread_get_state(i386_FLOAT_XSTATE)?
Samuel