Re: PATCH: PR gas/1874: mwait/monitor don't work in 64bit

["H. J. Lu" <hjl at lucon dot org>]

On Thu, Nov 17, 2005 at 10:30:05AM +0100, Jan Beulich wrote:
> >Here is the information I got from our chip people. The address size
> >override prefix is honored for "monitor" in both 16 and 32bit modes.
> >In 64bit mode, RCX and RDX are used for "monitor", RAX and RCX are
> >used for "mwait". This patch implements it.
> 
> Indeed, monitor and mwait GP-fault on non-zero RCX (but zero ECX) in
> 64-bit mode (regardless of address size override). In addition to what
> you say (but consistent with what your new tests implement) monitor also
> honors the address size override in 64-bit mode. Further, even in 16-bit
> mode (and thus I suppose also in 32-bit mode), use of the operand size
> override causes a (presumably #UD) fault (the manual mentions this only
> for 64-bit mode).
> 
> Bottom line is the documentation really doesn't match actual behavior.
> 
> However, I also have doubts that in 16-bit mode the address size
> handling is as you say: Using a value of 0x80000000 for eax doesn't
> produce a #GP fault regardless of the use of an address size override.


Ooops. I meant to say "The address size override prefix is honored for
monitor in both 32bit and 64bit modes." BTW, the current processsor
only checks edx/32bit for monitor and eax/32bit for mwait in 64bit
mode. But it isn't guaranteed for future processors.

> 
> Regarding the patch, I don't think hard-coding the address size
> override for 16-bit operation of monitor is appropriate - under .code16
> this is not going to produce the desired effect (and likewise is the
> opposite 32-bit operation in 16-bit mode then missing the override). The
> base problem here is that monitor's address operand, when explicitly
> specified, should take the form of a memory operand rather than a plain
> register (and note that the SDM doesn't indicate any particular syntax
> for how the operands should be specified). Once that is done, the
> address size prefixing will come out right implicitly, and the
> restriction to only use %ax/%eax/%rax could be easily added to the code
> in i386_index_check that already deals with AMD's SVME instructions that
> have the same restriction (except that they also ignore segment
> overrides, which monitor doesn't).

I will check.

> 
> Further, for the disassembly part I'd really like to see the explicit
> addr16/addr32 displaying to go away; these should only be output on
> instructions where none of the operands indicates the address size
> used.

Well, the Intel mode only displays "monitor/mwait" then.


H.J.