From: Paul Eggert <eggert@cs.ucla.edu>
To: Adhemerval Zanella <adhemerval.zanella@linaro.org>,
libc-alpha@sourceware.org
Subject: Re: [PATCH 6/7] stdlib: Optimization qsort{_r} swap implementation
Date: Mon, 22 Jan 2018 08:27:00 -0000 [thread overview]
Message-ID: <f1472674-e8f1-77d6-acd2-069f7fa89ce0@cs.ucla.edu> (raw)
In-Reply-To: <1516298002-4618-7-git-send-email-adhemerval.zanella@linaro.org>
Adhemerval Zanella wrote:
> +static inline bool
> +check_alignment (const void *base, size_t align)
> +{
> + return _STRING_ARCH_unaligned || ((uintptr_t)base % (align - 1)) == 0;
> +}
Surely the '(align - 1)' was supposed to be 'align'. Has this been tested on an
architecture that does not allow unaligned access?
> +static inline void
> +swap_generic (void *a, void *b, size_t size)
Why is this inline? It's used only as a function pointer, and the other
functions so used are not declared inline.
> +static inline swap_t
> +select_swap_func (const void *base, size_t size)
> +{
> + if (size == 4 && check_alignment (base, 4))
> + return swap_u32;
> + else if (size == 8 && check_alignment (base, 8))
> + return swap_u64;
> + return swap_generic;
> +}
The conditions aren't portable enough. Use something like this instead for
swap_u32, and similarly for swap_u64.
if (size == sizeof (uint32_t) && check_alignment (base, alignof (uint32_t)))
return swap_u32;
> +static void
> +swap_u32 (void *a, void *b, size_t size)
The pointer arguments should be declared 'void *restrict'. This can help GCC
generate better code. Similarly for the other swap functions.
> + uint32_t tmp = *(uint32_t*) a;
> + *(uint32_t*) a = *(uint32_t*) b;
> + *(uint32_t*) b = tmp;
It's nicer to avoid casts when possible, as is the case here and elsewhere. This
is because casts are too powerful in C. Something like this, say:
uint32_t *ua = a, *ub = b, tmp = *ua;
*ua = *ub, *ub = tmp;
> + unsigned char tmp[128];
Why 128? A comment seems called for.
> +static inline void
> +swap_generic (void *a, void *b, size_t size)
> +{
> + unsigned char tmp[128];
> + do
> + {
> + size_t s = size > sizeof (tmp) ? sizeof (tmp) : size;
> + memcpy (tmp, a, s);
> + a = __mempcpy (a, b, s);
> + b = __mempcpy (b, tmp, s);
> + size -= s;
> + }
> + while (size > 0);
> +}
On my platform (GCC 7.2.1 20170915 (Red Hat 7.2.1-2) x86-64) this inlined the
memcpy but not the mempcpy calls. How about something like this instead? It
should let the compiler do a better job of block-move-style operations in the
loop. If mempcpy is inlined for you, feel free to substitute it for two of the
loop's calls to memcpy.
static void
swap_generic (void *restrict va, void *restrict vb, size_t size)
{
char *a = va, *b = vb;
enum { n = 128 }; /* Why 128? */
unsigned char tmp[n];
while (size >= n)
{
memcpy (tmp, a, n);
memcpy (a, b, n);
memcpy (b, tmp, n);
a += n;
b += n;
size -= n;
}
memcpy (tmp, a, size);
memcpy (a, b, size);
memcpy (b, tmp, size);
}
next prev parent reply other threads:[~2018-01-22 8:27 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2018-01-18 17:53 [PATCH 0/7] Refactor qsort implementation Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 5/7] stdlib: Remove use of mergesort on qsort Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 2/7] support: Add Mersenne Twister pseudo-random number generator Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 4/7] stdlib: Add more qsort{_r} coverage Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 7/7] stdlib: Remove undefined behavior from qsort implementation Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 3/7] benchtests: Add bench-qsort Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 1/7] stdlib: Adjust tst-qsort{2} to libsupport Adhemerval Zanella
2018-01-18 17:53 ` [PATCH 6/7] stdlib: Optimization qsort{_r} swap implementation Adhemerval Zanella
2018-01-22 8:27 ` Paul Eggert [this message]
2018-01-22 10:55 ` Adhemerval Zanella
2018-01-22 13:46 ` Alexander Monakov
2018-01-22 15:23 ` Adhemerval Zanella
2018-01-22 17:15 ` Paul Eggert
2018-01-22 17:48 ` Adhemerval Zanella
2018-01-22 18:29 ` Paul Eggert
2018-01-22 19:33 ` Adhemerval Zanella
2018-01-23 6:04 ` Paul Eggert
2018-01-23 18:28 ` Adhemerval Zanella
2018-01-23 23:37 ` Paul Eggert
2018-01-24 10:47 ` Adhemerval Zanella
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=f1472674-e8f1-77d6-acd2-069f7fa89ce0@cs.ucla.edu \
--to=eggert@cs.ucla.edu \
--cc=adhemerval.zanella@linaro.org \
--cc=libc-alpha@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).