[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH] Make vector-map and vector-for-each in (rnrs base) fast
From: |
Linus Björnstam |
Subject: |
Re: [PATCH] Make vector-map and vector-for-each in (rnrs base) fast |
Date: |
Fri, 19 Feb 2021 10:08:43 +0100 |
User-agent: |
Cyrus-JMAP/3.5.0-alpha0-141-gf094924a34-fm-20210210.001-gf094924a |
As suspected: the eq? optimization does nothing.
I will write some tests for (r6rs base) to test the multi-vector behaviour of
this. Currently only (vector-map proc vec1) is tested in there. That comes
whenever I get my next computer time. It might be weeks, though.
Also regarding (r6rs base): the two maps in guile seem compatible: (ice-9
boot-9) and (rnrs base) map both seem to have the same limitations (only proper
lists of the same length), but the boot-9 one seems faster with worse error
messages - probably due to some offloading to C.
Wouldn't one definition be a better choice, and if so: which one?
--
Linus Björnstam
On Thu, 18 Feb 2021, at 08:34, Linus Björnstam wrote:
> Hi there!
>
> I was spelunking through the guile source tree and found (rnrs base).
> The vector-map and vector-for-each in there are horribly inefficient.
> They are doing (list->vector (apply map PROC (map vector->list
> vectors))), which means it spends quite some time checking for circular
> references.
>
> This fixes that. The speedup is surprisingly small, considering we pass
> through the elements 2 fewer times and don't chase pointers through
> memory trying to find cycles. Anywhere from 30 to 300% depending on how
> the stars are aligned on things like (vector-map + vec vec)
>
> One potential speedup we could do is using eq? to compare numbers, but
> I don't know how well fixnums in guile overlap size_t, regardless of
> how realistic such a limitation would be. If I change the behaviour of
> vector-map to go back-to-front (order is unspecified in r6rs) we can
> easily do (eq? -1 index) as a stop condition to avoid any eventual
> overhead of type checking with =. (If those are not elided, which I
> suspect might be the case. ). I did not look at that, since I have too
> little computer time these days.
>
> As an added bonus, this speeds up quicksort.scm in ecraven's benchmarks
> by a little.
> --
> Linus Björnstam
> Attachments:
> * 0001-Write-a-proper-vector-map-and-vector-for-each-for-rn.patch