Faster index range checks

[serhiy-storchaka]

BPO	28397
Nosy	@rhettinger, @vstinner, @skrah, @socketpair, @serhiy-storchaka, @sir-sigurd
PRs	Micro-optimize list index range checks #9784
Files	check_index.patch check_index.cocci issue28397-2.c

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2016-10-11.14:25:25.607>
created_at = <Date 2016-10-09.12:32:05.573>
labels = ['extension-modules', 'interpreter-core', '3.7', 'performance']
title = 'Faster index range checks'
updated_at = <Date 2018-10-10.12:03:32.312>
user = 'https://github.com/serhiy-storchaka'

bugs.python.org fields:

activity = <Date 2018-10-10.12:03:32.312>
actor = 'serhiy.storchaka'
assignee = 'none'
closed = True
closed_date = <Date 2016-10-11.14:25:25.607>
closer = 'serhiy.storchaka'
components = ['Extension Modules', 'Interpreter Core']
creation = <Date 2016-10-09.12:32:05.573>
creator = 'serhiy.storchaka'
dependencies = []
files = ['45034', '45035', '45055']
hgrepos = []
issue_num = 28397
keywords = ['patch']
message_count = 22.0
messages = ['278355', '278397', '278408', '278410', '278469', '278471', '278472', '278473', '278474', '278475', '278476', '278477', '278478', '278479', '278481', '278484', '278485', '278486', '278489', '278491', '278492', '327467']
nosy_count = 6.0
nosy_names = ['rhettinger', 'vstinner', 'skrah', 'socketpair', 'serhiy.storchaka', 'sir-sigurd']
pr_nums = ['9784']
priority = 'normal'
resolution = 'rejected'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue28397'
versions = ['Python 3.7']

[serhiy-storchaka]

The expression for testing that some index is in half-open range from 0 to limit can be written as

index >= 0 && index < limit

or as

(size_t)index < (size_t)limit

The latter form generates simpler machine code. This is idiomatic code, it is used in many C and C++ libraries (including C++ stdlib implementations). It already is used in CPython (in deque implementation).

Proposed patch rewrites index range checks in more efficient way. The patch was generated automatically by coccinelle script, and then manually cleaned up.

[rhettinger]

Don't change the code in the collections module. While semantically valid, the change obfuscates the code. The meaning of maxlen < 0 is that there is no maximum length. I don't want this test hidden; otherwise, I would have changed it long ago as was done elsewhere in the module where it made sense.

Elsewhere, consider making a macro with clear name and comment (as was done with NEEDS_TRIM) in the collections module. Otherwise, you're just leaving behind constipated and tricky code with no indication of why a signed variable is being coerced to unsigned.

[skrah]

Serhiy, could you please take out _decimal.c?

I thought we had been through that before. :) :) :)

[skrah]

The same applies to memoryview.c and _testbuffer.c -- I doubt that there's any measurable speed benefit and the clarity is reduced (I also don't want macros there).

[skrah]

In the following program, with gcc-5.3 doit() is significantly faster than doit2() in 64-bit Linux:

================================================================

#include <stdint.h>

int
doit(int64_t index, int64_t nitems)
{
    return index < 0 || index >= nitems;
}

int
doit2(int64_t index, int64_t nitems)
{
    return (uint64_t)index >= (uint64_t)nitems;
}

int
main(void)
{
    int count, i;

    for (i = 0; i < 1000000000; i++) {
        count += doit(832921, i);
    }

    return count;
}

================================================================

[skrah]

The difference in favor of doit() is even more pronounced with this
loop (also sorry for the uninitialized variable, but that does not
make a difference for benchmarking):

=====================================

for (i = 0; i < 10000; i++) {
        for (j = 0; j < 10000; j++) {
            count += doit(i, j);
        }
    }

======================================

[socketpair]

I have tested. performace differs in about of two times.
gcc -O3

0.22 nanoseconds per comparison.
vs
0.58 nanoseconds per comparison.

Does it cost a time that we spent to discuss here ?

[skrah]

Which version is faster in your tests?

[socketpair]

Much more conveniet way is to use unsigned variables in appropriate places.

[socketpair]

oh shi....doit() i.e.

return index < 0 || index >= nitems;

is faster!

[socketpair]

Without optimisation in compiler (-O0) speed is the same.

[socketpair]

-O2 -- the same speed too!
-O3 really helps.

[socketpair]

$ gcc --version
gcc (Ubuntu 5.4.0-6ubuntu1~16.04.2) 5.4.0 20160609

[skrah]

That matches my results as well:

-O2 gives about the same speed, with -O3 doit() has a huge advantage.

I'm not sure this is an optimization at all.

[serhiy-storchaka]

In your example functions are inlined. If prohibit inlining, the second function is faster.

$ gcc -O3 -o issue28397 issue28397-2.c 
$ time ./issue28397 0

real 0m8.097s
user 0m7.992s
sys 0m0.012s
$ time ./issue28397 1

real 0m5.467s
user 0m5.436s
sys 0m0.024s

[socketpair]

$ gcc -O3 -DDOIT=doit ./zzz.c -o zzz && time ./zzz

real 0m1.675s
user 0m1.672s
sys 0m0.000s

$ gcc -O3 -DDOIT=doit2 ./zzz.c -o zzz && time ./zzz

real 0m1.657s
user 0m1.656s
sys 0m0.000s

====================================================

#include <stdint.h>

static int __attribute__((noinline)) doit(int64_t index, int64_t nitems)
{
    return index < 0 || index >= nitems;
}

static int __attribute__((noinline)) doit2(int64_t index, int64_t nitems)
{
    return (uint64_t)index >= (uint64_t)nitems;
}

int main(void)
{
    int count=0, i;

    for (i = 0; i < 1000000000; i++) {
        count += DOIT(832921, i);
    }

    return count;
}

[skrah]

On 64-bit Linux there's no difference:

$ ./usr/bin/gcc -O3 -o issue28397-2 issue28397-2.c
$  time ./issue28397-2 0

real 0m2.486s
user 0m2.424s
sys 0m0.014s
$ time ./issue28397-2 1

real 0m2.433s
user 0m2.422s
sys 0m0.008s

Also, most of the time "index < 0 || index >= nitems" *is* inlined,
and it was at least three times faster here.

I guess the general point is that such micro-optimizations are
unpredictable on modern architectures and modern compilers.

Note that the fast inlined version used SSE instructions.

[vstinner]

Serhiy: "The latter form generates simpler machine code."

Since this issue is an optimization, can you please provide results of Python benchmarks? Maybe run the performance benchmark suite? I just released performance 0.3 with new benchmarks and less bugs! ;-)

http://github.com/python/performance

[serhiy-storchaka]

Unlikely this optimization have measurable affect on benchmarks.

I opened this issue because this optimization already was applied to deque (bpo-23553), and it is easy to write a script for applying it to all code. But since there are doubts about this optimization, I withdraw my patch.

[vstinner]

Serhiy Storchaka: "I opened this issue because this optimization already was applied to deque (bpo-23553)"

Ah, change 1e89094998b2 written by Raymond Hettinger last year, Raymond who wrote (msg278397): "Don't change the code in the collections module. While semantically valid, the change obfuscates the code." :-)

To stay consistent, I suggest to revert the useless micro-optimization 1e89094998b2. Python uses Py_ssize_t instead of size_t to support "i >= 0" checks", it's a deliberate choice to catch bugs.

[serhiy-storchaka]

Raymond extracted his optimization into separate function and commented it.

[serhiy-storchaka]

See also PR 9784 where Raymond shown assempler code generated for two variants.

There is the similar difference on 64-bit platform with GCC 7.3:

$ gcc -O3 -o issue28397 issue28397-2.c 
$ time ./issue28397-2 0

real 0m2,740s
user 0m2,739s
sys 0m0,000s
$ time ./issue28397-2 1

real 0m2,449s
user 0m2,449s
sys 0m0,000s

But with GCC 8.2 there is not a difference.

$ time ./issue28397-2 0

real 0m2,498s
user 0m2,498s
sys 0m0,000s
$ time ./issue28397-2 1

real 0m2,500s
user 0m2,496s
sys 0m0,004s

Both versions generate the same code for tested functions, there are only minor differences in the main() function (some independent instructions are reordered). I don't know what is the cause of such difference.