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

[Paul Khuong]

stassats wrote:
> +(define-vop (fixnum-len)
[...]
> + (:args (arg :scs (any-reg) :target res))
> + (:arg-types tagged-num)
> + (:results (res :scs (unsigned-reg)))
> + (:result-types unsigned-num)
> + (:generator 25
> + (move res arg)
> + (when (> n-fixnum-tag-bits 1)
> + (inst shr res (1- n-fixnum-tag-bits)))
Shouldn't that be an arithmetic shift?
> + (if (sc-is res unsigned-reg)
> + (inst test res res)
> + (inst cmp res 0))
> + (inst jmp :ge POS)
What's the SC test for?
> + (inst not res)
That's probably more quickly computed as:

mov mask, res
sar mask, (1- n-word-bits)
xor res, mask

Paul Khuong

------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1
_______________________________________________
Sbcl-devel mailing list
Sbcl-...@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sbcl-devel

[Stas Boukarev]

Paul Khuong <p...@pvk.ca> writes:

> stassats wrote:
>> +(define-vop (fixnum-len)
> [...]
>> + (:args (arg :scs (any-reg) :target res))
>> + (:arg-types tagged-num)
>> + (:results (res :scs (unsigned-reg)))
>> + (:result-types unsigned-num)
>> + (:generator 25
>> + (move res arg)
>> + (when (> n-fixnum-tag-bits 1)
>> + (inst shr res (1- n-fixnum-tag-bits)))
> Shouldn't that be an arithmetic shift?

Right, I should get a build where n-fixnum-tag-bits > 1.

>> + (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.

>> + (inst not res)
> That's probably more quickly computed as:
>
> mov mask, res
> sar mask, (1- n-word-bits)
> xor res, mask

That uses an additional register, is 2 bytes larger and is only slightly
faster in my tests, not worth it.

--
With best regards, Stas.

[Lutz Euler]

Hi,

Stas wrote:

> Paul Khuong <p...@pvk.ca> 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.

Just grep for "smaller instruction" in src/compiler/x86-64/*.lisp.

It seemed so far not important or advantageous to me (as a person who
added a few of these) to abstract that pattern into a macro or function.

Regards,

Lutz

[Paul Khuong]

Lutz Euler wrote:
> Stas wrote:
>
>> Paul Khuong<p...@pvk.ca> 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).

Paul Khuong

[Lutz Euler]

Paul Khuong wrote:
> Lutz Euler wrote:
> > Stas wrote:
> >
> >> Paul Khuong<p...@pvk.ca> 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