bug#72689: 31.0.50; Proposal to improve string-pixel-width

[David Ponce via "Bug reports for GNU Emacs, the Swiss army knife of text editors" <bug-gnu-emacs@gnu.org>]

On 18/08/2024 6:40 AM, Eli Zaretskii wrote:
>> Date: Sun, 18 Aug 2024 00:03:22 +0200
>> From:  David Ponce via "Bug reports for GNU Emacs,
>>   the Swiss army knife of text editors" <bug-gnu-emacs@gnu.org>
>>
>> The function string-pixel-width is essential for calculating pixel
>> dimensions, especially for UI components. And in this context this
>> function can be called very often when the display is refreshed.
>>
>> I propose the attached patch to make string-pixel-width faster while
>> using less memory, as shown by the following results of a basic
>> benchmark run in emacs -Q to compare the current implementation and this
>> proposal:
>>
>> ;; Basic benchmark
>> (let* ((text1 (make-string 1000 ?x))
>>          (text2 (propertize text1 'line-prefix "XXXX ")))
>>     (list
>>      (string-pixel-width text1)
>>      (string-pixel-width text2)
>>      (progn (garbage-collect)
>>             (benchmark-run 10000 (string-pixel-width text1)))
>>      (progn (garbage-collect)
>>             (benchmark-run 10000 (string-pixel-width text2)))
>>      ;;(insert text "\n")
>>      ))
>>
>> ;; Result with current implementation (4 run):
>> (12000 12000 (1.854707611 17 0.120106147) (1.884129905 17 0.12258791599999996))
>>
>> (12000 12000 (1.846544798 17 0.12243524500000003) (1.8822177530000002 17 0.12349287399999997))
>>
>> (12000 12000 (1.851244125 17 0.12162041699999998) (1.860517709 17 0.12352999599999998))
>>
>> (12000 12000 (1.8542218929999998 17 0.12164553900000001) (1.856302462 17 0.122891689))
>>
>> ;; Result with proposed implementation (4 run):
>> (12000 12000 (1.698974522 0 0.0) (1.727446 2 0.014782505999999973))
>>
>> (12000 12000 (1.701800124 0 0.0) (1.728024111 2 0.014718454999999908))
>>
>> (12000 12000 (1.6850850800000001 0 0.0) (1.732370238 2 0.014801913000000111))
>>
>> (12000 12000 (1.7356390130000001 0 0.0) (1.7858915800000001 2 0.014816158000000135))
>>
>>   From my observations, the new implementation is around 8% faster, and
>> trigger less GC.  When there is no line-prefix property to
>> remove, the new implementation doesn't trigger any GC after 10000 runs.
>> Otherwise, only 2 GC are triggered instead of 17.
>>
>> Maybe this proposal might be of interest, or at least provide some ideas
>> for improvement.
> 
> Thanks.  The idea SGTM, but I think the implementation needs to cater
> for the case where more than one execution thread performs this job
> "in parallel" (however improbable this could sound), so we need to be
> able to detect when this buffer is "busy".  The simplest way is to use
> some boolean buffer-local variable, which will be set non-nil when the
> function starts using the buffer and reset to nil when the function is
> done with its job.

Thanks.  Your point about "parallelism" is quite relevant.  However
I'm not sure I understand your suggestion to introduce a buffer-local
boolean variable.  In particular, what to do when the buffer-local
flag is set when entering the function?  Wait until the flag is
reset?  Create another buffer?

Perhaps for parallelism, using temporary buffers might be a better
approach?  But I have no idea how much stress would be put on Emacs if
thousands of temporary buffers were created/freed during a session?

Could you elaborate further?