Re: Comparing lowtags: slightly inefficient assembly code

Thanks for the hint!
Can I define VOPs in my application's source code, or do I have to modify
SBCL itself?
The article you linked mentions it's possible to do it at "runtime", but
then says it's a bit "hairy".

Regarding EQ/EQL/...: The comparison that's being done, unfortunately,
cannot rely on any of these. The comparison logic is a bit intricate.

Am Mi., 30. Apr. 2025 um 22:48 Uhr schrieb Vasily Postnicov <
[email protected]>:

> It's adding the fixnum's tag by shifting left.
>
> My guess would be to write a function which translates directly to a VOP.
> You can write any assembly code you wish there. See here for an example:
> https://pvk.ca/Blog/2014/08/16/how-to-define-new-intrinsics-in-sbcl/
>
> On the other hand, I'd recommend EQ/EQL/EQUAL/EQUALP for comparing objects
> whose types aren't known to be anything more narrow than T.
>
> ср, 30 апр. 2025 г., 23:03 David Scherfgen via Sbcl-help <
> [email protected]>:
>
>> The purpose of this code is to provide an "early exit" in a function that
>> compares two objects X and Y - it's the #1 function in terms of number of
>> calls and execution time in the application (determined by deterministic
>> and statistical profiling). If X and Y have different lowtags, then the
>> function can abort immediately. I can't tell whether the extra bit shifts
>> cause significant performance issues, I'd have to compare it to the
>> "optimal" code, but that's a bit difficult in practice.
>>
>> Am Mi., 30. Apr. 2025 um 21:51 Uhr schrieb Stas Boukarev <
>> [email protected]>:
>>
>>> Is that causing you some unexpected performance issues?
>>>
>>> On Wed, Apr 30, 2025 at 10:50 PM David Scherfgen via Sbcl-help
>>> <[email protected]> wrote:
>>> >
>>> > Hello,
>>> >
>>> > I have this SBCL-specific code to determine whether two objects X and
>>> Y have the same lowtag:
>>> >
>>> > (defun same-lowtag-p (x y)
>>> >   (= (sb-kernel:lowtag-of x) (sb-kernel:lowtag-of y)))
>>> >
>>> > The relevant part of the disassembly:
>>> >
>>> > MOV RDX, R8
>>> > SHL RDX, 1
>>> > AND EDX, 30
>>> > MOV RDI, RSI
>>> > SHL RDI, 1
>>> > AND EDI, 30
>>> > CMP RDI, RDX
>>> >
>>> > It shifts the address of X and Y to the left by 1 bit, then masks with
>>> 30 (= 15 << 1), then compares.
>>> > The bit shift is unnecessary. It could mask X and Y with 15 and
>>> directly compare without shifting.
>>> >
>>> > Is there any (hacky, using internals) way to modify the Lisp function
>>> so that it results in optimal assembly without bit shifts?
>>> >
>>> > Thank you.
>>> >
>>> > Best regards
>>> > David Scherfgen
>>> > _______________________________________________
>>> > Sbcl-help mailing list
>>> > [email protected]
>>> > https://lists.sourceforge.net/lists/listinfo/sbcl-help
>>>
>> _______________________________________________
>> Sbcl-help mailing list
>> [email protected]
>> https://lists.sourceforge.net/lists/listinfo/sbcl-help
>>
>

_______________________________________________
Sbcl-help mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sbcl-help