[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: speeding up `wc -m`

From: Pádraig Brady
Subject: Re: speeding up `wc -m`
Date: Mon, 21 May 2018 19:26:08 -0700
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.8.0

On 21/05/18 11:00, Bruno Haible wrote:
> Hi Pádraig,
>> $ yes áááááááááááááááááááá | head -n100000 > mbc.txt
>> $ yes 12345678901234567890 | head -n100000 > num.txt
>> ===== Before ====
>> $ time src/wc -m < mbc.txt
>> 2100000
>> real    0m0.186s
>> $ time src/wc -m < num.txt
>> 2100000
>> real    0m0.056s
> Here's my take on improving this. I'm attaching draft patches that have
> this effect on the timings:
> * On glibc:
>              num     mbc
>   Before    0.056   0.152
>   After     0.057   0.089
>   -------
>   Speedup    1.0     1.7
>   factor
> * On macOS 10.13:
>              num     mbc
>   Before    0.153   0.229
>   After     0.042   0.112
>   -------
>   Speedup    3.6     2.0
>   factor
> Basically, the two problems that the profiling found were:
>   * It is pointless to call locale_charset repeatedly, because the
>     locale won't change while 'wc' is running.
>   * glibc has a slow mbrtowc() implementation for UTF-8 locales.
> Both problems can be addressed with the "abstract factory" design patterns.
> Namely, instead of using the generic 'wcwidth'/'mbrtowc' function each
> time, let the program produce an optimized 'wcwidth'/'mbrtowc' function
> [pointer] once, and then call this optimized function pointer repeatedly
> for each character.
> While at it, let me also do the same for the initialization of an mbstate_t,
> because on macOS the mbstate_t is 128 bytes long but only the first 12 bytes
> actually matter.
> This factory of function pointers side-steps the portability problems of
> 'locale_t'.
> Notes:
>   - When you use these new gnulib modules, you are programming against an API
>     that is very similar to POSIX, but not exactly POSIX.
>   - The platform-specific #ifs have to be adjusted, by the help of configure
>     tests.
>   - mbrtowc-factory needs a unit test (for which I have a draft).

Wow thanks for doing all that.

It's also worth noting that using wcwidth-factory etc.
will always use the replacement routines on utf8,
which would be a disadvantage on code size and divergence.

Another disadvantage is the change in API which is a bit awkward,
and would need tweaking elsewhere in coreutils to take advantage
of the speedup.

The first locale_charset() issue at least could be dealt with internally
to gnulib with a simple gnulib level config.  The attached is a proposed
solution to the charset issue, that would just require depending
on the wchar-single gnulib module to indicate locales don't change across calls.

Now if utf8_mbrtowc() is about 4.5 and 2.3 times faster than mbrtowc()
on glibc and macOS respectively, it's probably worthwhile to replace 
Though it would be cool to do that behind the wchar-single setting
which could setup the appropriate calls/pointers internally to rpl_mbrtowc()?

As for the reduced mbstate settings, it might not be worth
complicating the interface for this perf gain?


Attachment: gnulib-wchar-single.patch
Description: Text Data

reply via email to

[Prev in Thread] Current Thread [Next in Thread]