Re: [Sbcl-devel] [Sbcl-commits] master: Micro-optimize integer-length on fixnums on x86-64.

[<lut...@fr...>]

Paul Khuong wrote:
> Lutz Euler wrote:
> > Stas wrote:
> >
> >> Paul Khuong<pv...@pv...>  writes:
> >>
> >>> stassats wrote:
> > [...]
> >>>> +    (if (sc-is res unsigned-reg)
> >>>> +        (inst test res res)
> >>>> +        (inst cmp res 0))
> >>>> +    (inst jmp :ge POS)
> >>> What's the SC test for?
> >> That was copied from the original instruction.
> >
> > this kind of SC test is quite common: to test if a value is 0, if it is
> > in a register, "test" can be used, which has a shorter encoding than
> > "cmp". "cmp" on the other hand can take one operand in memory, too.
> 
> Right, but the VOP wants an unsigned-reg for res (without any load-if 
> condition).

You are right. Sorry, I didn't expect anything wrong in that direction
so didn't look hard enough.
The SC-test is unnecessary; one can always use the TEST instruction.
The original VOP SIGNED-BYTE-64-LEN has the same unnecessary test.
The change from "always use cmp" to "test the sc" there was commit
16028d14234d2acd0e6a3145a7364f2a52eabf63, which affects x86, too.

Greetings,

Lutz