Implement ZEND_ARRAY_KEY_EXISTS opcode to speed up array_key_exists() #3360

Majkl578 · 2018-07-02T22:15:09Z

This change makes array_key_exists() actually faster than isset() by ~25% (tested with GCC 8, -O3, march=native, mtune=native).

addresses https://bugs.php.net/bug.php?id=76148
addresses https://bugs.php.net/bug.php?id=71239

With a synthetic benchmark:

<?php

$foo = [123 => 456, 'abc' => 'def'];

$iterations = 500000000;

$a = microtime(true);
for ($i = 0; $i < $iterations; $i++) {
    isset($foo[123]);
}
printf('isset (%dx): %.2fs%s', $iterations, microtime(true) - $a, PHP_EOL);

$b = microtime(true);
for ($i = 0; $i < $iterations; $i++) {
    array_key_exists(123, $foo);
}
printf('array_key_exists (%dx): %.2fs%s', $iterations, microtime(true) - $b, PHP_EOL);

It yields results like:
7.2 (unpatched):

isset (500000000x): 7.65s
array_key_exists (500000000x): 9.03s

7.3 + patch:

isset (500000000x): 7.60s
array_key_exists (500000000x): 5.65s

nikic

Thanks for taking this on! I've left a few implementation notes, in particular in regard to how performance may be improved.

Zend/zend_compile.c

nikic · 2018-07-03T08:26:21Z

Zend/zend_vm_def.h

+	subject = GET_OP2_ZVAL_PTR(BP_VAR_R);
+	key = GET_OP1_ZVAL_PTR(BP_VAR_R);
+
+	if (UNEXPECTED(Z_ISREF_P(subject))) {


A more efficient (and for VM code, more idiomatic) way to handle this is to check for ISREF prior to the error case and use a goto try_again pattern. See https://github.com/Majkl578/php-src/blob/2265ca4f783b9e5e3a575ebe8c3b1f21f3009015/Zend/zend_vm_def.h#L4419 for an example.

nikic · 2018-07-03T08:28:08Z

Zend/zend_vm_def.h

+
+	SAVE_OPLINE();
+
+	subject = GET_OP2_ZVAL_PTR(BP_VAR_R);


It would be preferable to use GET_OP2_ZVAL_PTR_UNDEF here and check the UNDEF case for IS_CV inside the branch that generates the type error. For an example of delayed UNDEF handling see https://github.com/Majkl578/php-src/blob/2265ca4f783b9e5e3a575ebe8c3b1f21f3009015/Zend/zend_vm_def.h#L59.

nikic · 2018-07-03T08:30:00Z

Zend/zend_vm_def.h

+			zend_error(E_WARNING, "array_key_exists(): The first argument should be either a string or an integer");
+			result = 0;
+		}
+		ZVAL_BOOL(EX_VAR(opline->result.var), result);


The code between IS_ARRAY and IS_OBJECT is duplicated here (which is actually rather dubious and will handle some cases incorrectly, but that's an existing issue and best solved by deprecating array_key_exists on objects entirely). It would be better to only extract the hashtable using Z_ARRVAL_P and Z_OBJPROP_P in these branches and have only one copy of the common key handling code afterwards.

nikic · 2018-07-03T08:33:34Z

ext/opcache/Optimizer/zend_inference.c

 		case ZEND_ASSERT_CHECK:
 		case ZEND_IN_ARRAY:
+		case ZEND_ARRAY_KEY_EXISTS:
 			UPDATE_SSA_TYPE(MAY_BE_FALSE|MAY_BE_TRUE, ssa_ops[i].result_def);


array_key_exists() can also return null for the zpp failure case.

Additionally, ARRAY_KEY_EXISTS needs to be handled in https://github.com/php/php-src/blob/master/ext/opcache/Optimizer/sccp.c#L1328. Looking at the ISSET_ISEMPTY_DIM_OBJ case should provide inspiration here.

nikic · 2018-07-03T08:35:07Z

Zend/zend_vm_def.h

+
+	FREE_OP2();
+	FREE_OP1();
+	ZEND_VM_NEXT_OPCODE_CHECK_EXCEPTION();


array_key_exists() is a function that would benefit from smart branch handling. See https://github.com/Majkl578/php-src/blob/2265ca4f783b9e5e3a575ebe8c3b1f21f3009015/Zend/zend_vm_def.h#L8085 how this is done for ZEND_IN_ARRAY. You'll also have to add it to https://github.com/php/php-src/blob/master/Zend/zend_compile.c#L2181.

Majkl578 · 2018-07-03T20:55:30Z

@nikic Thanks for your review 🥇, I'll try to address it later this week.

KalleZ · 2018-07-03T21:01:45Z

Also you should just revert the change to NEWS, usually the merger of the PR will do this part =)

cmb69 · 2018-07-03T22:08:11Z

Also you should just revert the change to NEWS, usually the merger of the PR will do this part =)

For this PR an (additional) entry in UPGRADING appears to be appropriate (UPGRADING usually serves as base for the migration guide in the PHP manual, but NEWS is likely to be overlooked, and this improvement seems to be quite noteworthy).

laruence · 2018-07-04T03:35:35Z

Although isset and array_key_exists behavior slightly different, but I think in most cases, array_key_exists can be repalced with isset..

thus, I doubt such optimization is worthy to do...

nikic · 2018-07-04T08:20:05Z

@laruence array_key_exists() is common in foundational libraries that need to support null values. These are often very hot code paths and currently the usual way to handle them is if (isset($array[$key]) || \array_key_exists($key, $array)), because directly calling array_key_exists() is too slow.

nikic · 2018-07-04T10:26:59Z

@cmb69 Currently we usually do not include performance changes in UPGRADING, unless they come with some observable behavior change. But I think it's a good idea to start doing this, maybe under a new "Performance Improvements" section. We do tend to have a fair bit of improvements in each release, but they aren't really called out anywhere.

cmb69 · 2018-07-04T10:46:05Z

@nikic Even if we won't document performance changes, I still think this very change should be documented, since it can affect how userland code will be written. You already gave a typical example which could be improved wrt. this PR.

hikari-no-yume · 2018-07-04T21:40:04Z

By the way, it would be nice to have a comment above the code for the normal function array_key_exists that mentions this special case. I added one for all the special cases there were some time ago (2015 maybe?), because I had modified a special-cased function's behaviour (in the actual function, not the compiler) and spent some time confused as to why the change wasn't taking effect.

Majkl578 · 2018-08-08T14:43:44Z

@nikic I adressed all your comments, except one: SCCP for this new opcode, since it's a bit beyond my current knowledge of PHP internals. May I ask you or anyone else to help me with this one? Thanks!

Majkl578 · 2018-08-08T14:50:57Z

gcc-8 -O0

isset (100000000x): 3.57s
array_key_exists (100000000x): 3.33s

gcc-8 -O3

isset (100000000x): 1.49s
array_key_exists (100000000x): 1.34s

@nikic Can you please another round of review? Thanks.

nikic · 2018-08-08T15:15:12Z

/cc @cmb69 for whether this can still land in 7.3. It introduces a new opcode, not sure what our policy is there. I guess as it may require further extension changes (xdebug support etc) probably too late?

nikic · 2018-08-08T15:18:18Z

Zend/zend_vm_def.h

+
+ZEND_VM_C_LABEL(try_again):
+
+	if (UNEXPECTED(Z_ISREF_P(subject)) || UNEXPECTED(Z_ISREF_P(key))) {


This isn't quite what I had in mind regarding the reference handling. The idea is to do something like this:

ZEND_VM_C_LABEL(try_again_subject): if (EXPECTED(Z_TYPE_P(subject) == IS_ARRAY)) { ht = Z_ARRVAL_P(subject); } else if (UNEXPECTED(Z_TYPE_P(subject) == IS_OBJECT)) { ht = Z_OBJPROP_P(subject); } else if (Z_TYPE_P(Z_ISREF_P(subject)) { subject = Z_REFVAL_P(subject); ZEND_VM_C_GOTO(try_again_subject); } else { if (OP1_TYPE == IS_CV && UNEXPECTED(Z_TYPE_INFO_P(key) == IS_UNDEF)) { key = GET_OP1_UNDEF_CV(key, BP_VAR_R); } if (OP2_TYPE == IS_CV && UNEXPECTED(Z_TYPE_INFO_P(subject) == IS_UNDEF)) { subject = GET_OP2_UNDEF_CV(subject, BP_VAR_R); } zend_internal_type_error(EX_USES_STRICT_TYPES(), "array_key_exists() expects parameter 2 to be array, %s given", zend_get_type_by_const(Z_TYPE_P(subject))); FREE_OP2(); FREE_OP1(); ZEND_VM_SMART_BRANCH(result, 0); ZVAL_NULL(EX_VAR(opline->result.var)); ZEND_VM_NEXT_OPCODE_CHECK_EXCEPTION(); }

The point is that the check for the reference is only done after the most likely case (subject is an array) is already handled, so that rare cases (references) don't have a runtime impact on common ones (array).

Hah, got it!

nikic · 2018-08-08T15:20:55Z

Zend/zend_vm_def.h

+		zend_internal_type_error(EX_USES_STRICT_TYPES(), "array_key_exists() expects parameter 2 to be array, %s given", zend_get_type_by_const(Z_TYPE_P(subject)));
+		FREE_OP2();
+		FREE_OP1();
+		ZEND_VM_SMART_BRANCH(result, 0);


I believe result is uninitialized here. It should be fine to just drop the smart branch line here, as this is a rare error case anyway.

Zend/zend_vm_def.h

cmb69 · 2018-08-08T15:41:01Z

Regarding PHP 7.3: I cannot assess how many extensions and SAPIs (phpdbg?) would be affected, and how much work it would be to update them, but given that Xdebug is not ready for PHP 7.3 yet, and the long UPGRADING.INTERNALS it might indeed be better to delay this to PHP 7.4. Otherwise this issue should be discussed on internals@.

Majkl578 · 2018-08-08T16:32:55Z

I only quickly checked code of XDebug and PHPDBG and found no usages of i.e. ZEND_COUNT or ZEND_IN_ARRAY (which is basically the same thing). Of course I don't know if there are any indirect implications for these tools, but it doesn't seem so, unlike i.e. new syntax.

Majkl578 · 2018-08-08T17:12:49Z

http://news.php.net/php.internals/103064

Majkl578 · 2018-08-09T15:59:25Z

(Waiting on author label can be removed.)

Majkl578 · 2018-08-14T01:38:35Z

Alright, retargeted to master/7.4.

Majkl578 · 2018-12-24T15:33:32Z

I'll handle the feedback over next few days. 👍

Majkl578 · 2018-12-26T17:40:38Z

@dstogov Adressed your comments and also rebased again. 👍

dstogov · 2018-12-26T19:55:42Z

Zend/zend_vm_def.h

+		result = zend_hash_exists_ind(ht, ZSTR_EMPTY_ALLOC());
+	} else if (Z_TYPE_P(key) <= IS_NULL) {
+		if (OP1_TYPE == IS_CV && UNEXPECTED(Z_TYPE_P(key) == IS_UNDEF)) {
+			ZVAL_UNDEFINED_OP1();


Must be OP2_TYPE ans ZVAL_UNDEFINED_OP2.

It seems, you changed OP1->OP2 in wrong place. I'll take care about this and merge.

Oh, my bad. Everything is fine.

Key is OP1 and haystack is OP2 so it should be fine.
https://github.com/php/php-src/compare/9d438180daf644b9937ba4177c27b7e9e95e05cf..f4731326d340d9f67d608201a95ef182af223db2#diff-3054389ad750ce9a9f5895cd6d27800fR6400

yeah. My mistake. It's already merged into master.

nikic · 2018-12-26T21:28:13Z

I've added the perf improvement section for UPGRADING which I mentioned above: f1c0e67

Majkl578 · 2018-12-26T21:52:05Z

Thanks @nikic @dstogov @cmb69!

Make opcache replace the result with false if the array argument is known to be empty. This may be useful when a codebase has placeholders, e.g. `if (!in_array($method, self::ALLOWED_METHODS)) { return; }` In zend_inference.c: In php 8, array_key_exists will throw a TypeError instead of returning null. I didn't see any discussion of this optimization (for/against) after a quick search on github, e.g. phpGH-3360 Potential future optimizations: - convert `in_array($needle, ['only one element'], true)` to `===`? (or `==` for strict=false) - When the number of elements is less than 4, switch to looping instead of hash lookup. (exact threshold for better performance to be determined) Also support looping for `in_array($value, [false, 'str', 2.5], true/false)`

…#3360

Make opcache replace the result with false if the array argument is known to be empty. This may be useful when a codebase has placeholders, e.g. `if (!in_array($method, self::ALLOWED_METHODS)) { return; }` In zend_inference.c: In php 8, array_key_exists will throw a TypeError instead of returning null. I didn't see any discussion of this optimization (for/against) after a quick search on github, e.g. phpGH-3360 Potential future optimizations: - convert `in_array($needle, ['only one element'], true)` to `===`? (or `==` for strict=false) - When the number of elements is less than 4, switch to looping instead of hash lookup. (exact threshold for better performance to be determined) Also support looping for `in_array($value, [false, 'str', 2.5], true/false)`

Besides being faster (since php/php-src#3360), array_key_exists is also safer as it narrows down key existence to arrays only and performs type-checking on both parameters.

Make opcache replace the result with false if the array argument is known to be empty. This may be useful when a codebase has placeholders, e.g. `if (!in_array($method, self::ALLOWED_METHODS)) { return; }` In zend_inference.c: In php 8, array_key_exists will throw a TypeError instead of returning null. I didn't see any discussion of this optimization (for/against) after a quick search on github, e.g. phpGH-3360 Potential future optimizations: - convert `in_array($needle, ['only one element'], true)` to `===`? (or `==` for strict=false) - When the number of elements is less than 4, switch to looping instead of hash lookup. (exact threshold for better performance to be determined) Also support looping for `in_array($value, [false, 'str', 2.5], true/false)`

This is pretty “hot” code (compare also change I906d5b3f2f, commit 9c89543), and by using \array_key_exists() instead of array_key_exists(), we inform PHP that we’re looking for the built-in function and not a potential \MediaWiki\Config\array_key_exists() or \MediaWiki\array_key_exists() function. This in turn lets opcache compile it to a custom ARRAY_KEY_EXISTS opcode (see PHP PR #3360 [1] and commits f5044a12dd [2] + 0bbfebb6d9 [3]). Ori Livneh benchmarked a version of this change and noticed a ~40% improvement on getAll() for enwiki, from 0.952 ms to 0.688 ms. [1]: php/php-src#3360 [2]: php/php-src@f5044a12dd [3]: php/php-src@0bbfebb6d9 Change-Id: Id4e932e7c9512eba6a9843486a6539b0165a1b61

nikic requested changes Jul 3, 2018

View reviewed changes

nikic added the Waiting on Author label Jul 7, 2018

Majkl578 force-pushed the ZEND_ARRAY_KEY_EXISTS-opcode branch from 2265ca4 to 6a76ae4 Compare August 8, 2018 14:38

Majkl578 changed the base branch from master to PHP-7.3 August 8, 2018 14:39

Majkl578 force-pushed the ZEND_ARRAY_KEY_EXISTS-opcode branch from 6a76ae4 to 4ff60bb Compare August 8, 2018 14:40

Majkl578 force-pushed the ZEND_ARRAY_KEY_EXISTS-opcode branch from 4ff60bb to 16c5a7b Compare August 8, 2018 15:03

nikic reviewed Aug 8, 2018

View reviewed changes

Zend/zend_vm_def.h Show resolved Hide resolved

Majkl578 force-pushed the ZEND_ARRAY_KEY_EXISTS-opcode branch from 16c5a7b to bc6db1f Compare August 8, 2018 15:32

nikic added Waiting on Review and removed Waiting on Author labels Aug 9, 2018

Majkl578 changed the base branch from PHP-7.3 to master August 14, 2018 01:37

Majkl578 force-pushed the ZEND_ARRAY_KEY_EXISTS-opcode branch from bc6db1f to c9a918b Compare August 14, 2018 01:38

Majkl578 and others added 2 commits December 26, 2018 18:21

Implement ZEND_ARRAY_KEY_EXISTS opcode to speed up array_key_exists()

6ce76d5

Add SCCP support for ZEND_ARRAY_KEY_EXISTS

f473132

Majkl578 force-pushed the ZEND_ARRAY_KEY_EXISTS-opcode branch from 9d43818 to f473132 Compare December 26, 2018 17:39

dstogov reviewed Dec 26, 2018

View reviewed changes

nikic closed this Dec 26, 2018

Majkl578 deleted the ZEND_ARRAY_KEY_EXISTS-opcode branch December 26, 2018 21:52

fabpot mentioned this pull request Apr 29, 2019

Optimize context access on PHP 7.4 twigphp/Twig#2985

Merged

KennedyTedesco mentioned this pull request Aug 14, 2019

[6.x] PHP 7.4 support laravel/framework#29482

Merged

mainframe98 mentioned this pull request Oct 1, 2019

Performance improvements (+4% to speed in Magento 2.3.2) wikimedia/less.php#14

Merged

gharlan mentioned this pull request Oct 22, 2019

[YAML] Improve performance of YAML parser symfony/symfony#34058

Merged

TysonAndre mentioned this pull request Nov 17, 2019

Optimize array_key_exists/in_array for empty array #4925

Closed

nicolas-grekas mentioned this pull request Jan 23, 2020

[7.x] Refactor route caching laravel/framework#31188

Merged

rob006 mentioned this pull request Feb 11, 2020

Made getValue consistent to setValue. yiisoft/arrays#18

Merged

germanow added a commit to germanow/arrays that referenced this pull request Feb 12, 2020

Removed isset() micro optimization. It is no longer valid php/php-src…

6090e8f

…#3360

ihor-sviziev mentioned this pull request Sep 17, 2020

Speedup static content deploy for specific theme magento/magento2#30090

Merged

5 tasks

jmikola referenced this pull request Aug 29, 2023

Implement ZEND_ARRAY_KEY_EXISTS opcode to speed up array_key_exists()

f5044a1

kiatng mentioned this pull request Jun 18, 2024

UPS Rest API: handling cases where a single service is returned by UPS OpenMage/magento-lts#4044

Merged


		ZEND_VM_C_LABEL(try_again):

		if (UNEXPECTED(Z_ISREF_P(subject)) \|\| UNEXPECTED(Z_ISREF_P(key))) {

Implement ZEND_ARRAY_KEY_EXISTS opcode to speed up array_key_exists() #3360

Implement ZEND_ARRAY_KEY_EXISTS opcode to speed up array_key_exists() #3360

Uh oh!

Conversation

Majkl578 commented Jul 2, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Majkl578 commented Jul 3, 2018

Uh oh!

KalleZ commented Jul 3, 2018

Uh oh!

cmb69 commented Jul 3, 2018

Uh oh!

laruence commented Jul 4, 2018

Uh oh!

nikic commented Jul 4, 2018

Uh oh!

nikic commented Jul 4, 2018

Uh oh!

cmb69 commented Jul 4, 2018

Uh oh!

hikari-no-yume commented Jul 4, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Majkl578 commented Aug 8, 2018

Uh oh!

Majkl578 commented Aug 8, 2018

Uh oh!

nikic commented Aug 8, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nikic Aug 8, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cmb69 commented Aug 8, 2018

Uh oh!

Majkl578 commented Aug 8, 2018

Uh oh!

Majkl578 commented Aug 8, 2018

Uh oh!

Majkl578 commented Aug 9, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Majkl578 commented Aug 14, 2018

Uh oh!

Majkl578 commented Dec 24, 2018

Uh oh!

Majkl578 commented Dec 26, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nikic commented Dec 26, 2018

Uh oh!

Majkl578 commented Jul 2, 2018 •

edited

Loading

hikari-no-yume commented Jul 4, 2018 •

edited

Loading

nikic Aug 8, 2018 •

edited

Loading

Majkl578 commented Aug 9, 2018 •

edited

Loading