Re: [Qemu-devel] [PATCH] target-i386: fix cmpxchg instruction emulation

[Andreas Gustafsson]

malc wrote:
> > Inefficient in what way?  The generated code only grows by a single
> > unconditional branch.
> 
> The generated code grows by a memory write

Yes, an additional store instruction is generated, but the number of
store instructions *executed* does not change.  The original code
already does a store in both the compare-equal and the compare-unequal
case, by branching to the same store instruction; with my patch, a
store is still done in both cases, but the store instruction executed
is a separate one in each.

> (which is not what the hardware does).

The hardware does an atomic read-modify-write; qemu emulates this as a
separate read and write, and like the hardware, always does the write
part whether or not the "modify" part actually modified the data.  My
patch does *not* change this in any way.

> I recall having discussion aboutit with Fabrcie (in private) and i
> blieve (and if my memory serves me) we came to the conclusion that
> there's a way forward w.r.t. to this issue i just never came around of
> implementing it, i can try to dig out the old mails and share the
> highlights with you if you are interested.

Yes, please.
-- 
Andreas Gustafsson, address@hidden