Re: [Caml-list] Compiler bug?

[Dmitry Bely <dmitry.bely@gmail.com>]

On Wed, Aug 8, 2012 at 10:22 PM, Jesper Louis Andersen
<jesper.louis.andersen@gmail.com> wrote:
> On Wed, Aug 8, 2012 at 8:03 PM, Alain Frisch <alain@frisch.fr> wrote:
>
>> I'm not sure, someone else would have to reply!  I guess that these
>> registers are supposed to be preserved by the callee in the x64 ABI (and
>> obviously, they don't hold pointers to OCaml values, so they don't have to
>> be tracked by the GC).

Yes, they should not be tracked by GC, but they may hold some FP
values (can they across GC call?). Actually for x64 FPU registers are
saved; see amd64.S/amd64nt.asm. But not for x86. I'm curious why.

> Also, the GC itself will not be using Floating Point code at all, so
> we can save a lot by not saving/restoring these values on the stack.

This is not entirely true; GC also triggers pending signals processing
that can modify FPU state and calls C library fuctions (e.g. malloc)
that can do anything inside.

> It is akin to what the Linux kernel does: trap on the first FP
> instruction and only then do work to save/restore the FP context - but
> here we know a priori we never will access FP. By doing so we can cut
> the time to the GC call down - and reap the benefits by having a
> faster GC cycle.

- Dmitry Bely