Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

[Stefan Monnier]

> +      ;; TODO: Once :linear-search is implemented for `make-hash-table'
> +      ;; set it to `t' for cond forms with a small number of cases.

Indeed, we can probably substantially improve our hash-table performance
for small hash tables (not only in terms of CPU use but also in terms of
memory use).

But rather than a :linear-search keyword to make-hash-table, can't we
just test the value of `count` and use a linear search if it's small?


        Stefan

Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

[Clément Pit-Claudel]

On 2017-02-06 10:35, Stefan Monnier wrote:
> Indeed, we can probably substantially improve our hash-table performance
> for small hash tables (not only in terms of CPU use but also in terms of
> memory use).
> 
> But rather than a :linear-search keyword to make-hash-table, can't we
> just test the value of `count` and use a linear search if it's small?

This would be great.  I currently have quite a bit of hash-table-heavy code with many relatively small hash tables, but in a few rare cases they can grow large, so alists don't cut it.  A linear-search keyword wouldn't really help, because I can't tell beforehand which ones will be large.

Clément.

Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

[Vibhav Pant]

On Mon, Feb 6, 2017 at 9:05 PM, Stefan Monnier <monnier@iro.umontreal.ca> wrote:
> But rather than a :linear-search keyword to make-hash-table, can't we
> just test the value of `count` and use a linear search if it's small?

Indeed. That should take less to code to implement too.

-- 
Vibhav Pant
vibhavp@gmail.com

Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

[Stefan Monnier]

>> But rather than a :linear-search keyword to make-hash-table, can't we
>> just test the value of `count` and use a linear search if it's small?

Nitpick: actually, it should be based on the hash-table's *size* rather
than its *count*.


        Stefan

Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

[Stefan Monnier]

>>> But rather than a :linear-search keyword to make-hash-table, can't we
>>> just test the value of `count` and use a linear search if it's small?
> Nitpick: actually, it should be based on the hash-table's *size* rather
> than its *count*.

For me, the place where small hash tables showed up is in cl-generic,
where many of them have just a single entry.
For byte-switch, optimizing the single-entry case is not important (you
wouldn't want to use byte-switch for it), tho.


        Stefan

Re: [Emacs-diffs] feature/byte-switch c4316a2: ; bytecomp.el (byte-compile-cond-jump-table): Add TODO note

[Vibhav Pant]

[-- Attachment #1: Type: text/plain, Size: 253 bytes --]

On 06-Feb-2017 11:25 PM, "Stefan Monnier" <monnier@iro.umontreal.ca> wrote:

For byte-switch, optimizing the single-entry case is not important (you
wouldn't want to use byte-switch for it), tho.


byte-compile-cond-jump-table does check for that, yes.

[-- Attachment #2: Type: text/html, Size: 718 bytes --]