Re: [Sbcl-devel] [Sbcl-commits] master: Three microoptimizations to numeric tag testing for 1-bit fixnum tag.

[Stas B. <sta...@gm...>]

"Douglas Katzman" <sn...@us...> writes:

> The branch "master" has been updated in SBCL:
>        via  46e70b8e713a8677882ed675115b01c82d823255 (commit)
>       from  d55a3540322e28841b7712986887830c6a24d32e (commit)
>
> - Log -----------------------------------------------------------------
> commit 46e70b8e713a8677882ed675115b01c82d823255
> Author: Douglas Katzman <do...@go...>
> Date:   Tue May 6 08:14:29 2014 -0400
>
>     Three microoptimizations to numeric tag testing for 1-bit fixnum tag.
>     
>     Each of the following might save 1 byte on x86-64, depending on
>     register allocation.
>     
>     - Instead of "LEA EAX, [Rxx - other-pointer-lowtag]" use full-width RAX.
>       At worst this adds a REX prefix, but might not if Rxx needed it anyway.
>       The payback is in loading widetag from BYTE PTR [RAX].
>       After the LEA we never need a displacement, which either breaks even
>       by saving 1 byte, or wins - nets a 1 byte savings if Rxx had required
>       a REX prefix.
>       Note: RAX is an alias pointer, but safe because of GC conservativism.
>     
>     These two depend on AL having the shortest encoding:
>     - Do the fixnum lowtag test only after loading RAX by checking for
>       a 1 bit instead of looking for a 0 bit in the original value.
>     - Similarly for SHORT-FLOAT, do the immediate tag test on AL
>       compensating for the biased value in the widetag field.
The GC is conservative, but this will make it harder to make it precise.

-- 
With best regards, Stas.