[php-src] PR #22231: perf: move EG() and CG() in ZTS builds into __thread storage
| From: | henderkes | Date: | Fri, 05 Jun 2026 13:00:32 +0000 |
| Subject: | [php-src] PR #22231: perf: move EG() and CG() in ZTS builds into __thread storage | ||
| Groups: | php.git-pulls | ||
| Request: | Send a blank email to git-pulls+get-36305@lists.php.net to get a copy of this message | ||
Pull Request: https://github.com/php/php-src/pull/22231
Author: henderkes
So, I was chasing what caused [the much higher instruction count on aarch64 for clang vs
gcc](https://github.com/shivammathur/homebrew-php/issues/5161#issuecomment-4608440602) and while
figuring that out, got side tracked to figuring out that ZTS builds also have a ton of extra
instructions compared to NTS builds. Turned out it was mostly just from constantly accessing
executor and compiler globals when running phpstan. Each access became pseudo assembly:
```asm
mrs _tsrm_ls_cache
add cache offset low bits
add cache offset high bits
load relative base address
load accessed global
```
Here we're slashing one extra load and one add by moving EG and CG onto actual thread storage.
Benchmarks are using phpstan on full laravel codebase (full symfony codebase showed the same, but I
added the LE optimization after and don't want to rerun 6 hours of benchmarks) and
phoronix-test-suite phpbench.
aarch64:
- ~80% fewer extra instructions
- Ampere A1 cloud machine, ~0.15% variance in wall clock between runs
<table><thead><tr><th>build</th><th>PHPStan
cyc</th><th>instr</th><th>wall</th><th>phpbench</th><th>phpbench
JIT</th></tr></thead><tbody><tr><td>gcc-nts</td><td>90.67B</td><td>108.22B</td><td>32.0s</td><td>715719</td><td>116467
1</td></tr><tr><td>gcc-zts</td><td>94.44B</td><td>116.28B</td><td>33.2s</td><td>662475</td><td>1111162</td></tr><tr><t
d>d>gcc-zts-patch</td><td>91.61B</td><td>109.66B</td><td>32.3s</td><td>704490</td><td>1146721</td></tr><tr><td>clang-nts
</td><td>95.20B</td><td>113.29B</td><td>33.4s</td><td>709077</td><td>1118672</td></tr><tr><td>clang-zts</td><td>97.92B
</td><td>120.82B</td><td>34.4s</td><td>665074</td><td>1069905</td></tr><tr><td>clang-zts-patch</td><td>93.40B</td><td>
114.72B</td><td>33.1s</td><td>696929</td><td>1101997</td></tr></tbody></table>
x86_64:
- ~35% fewer extra instructions
- no cycle count on x86_64 because lack of hardware counter support in WSL
- i7 12650h with locked cpu frequency, take the wall times with a grain of salt
now as the minified html I pasted at the beginning20:00Claude responded:
<table><thead><tr><th>build</th><th>PHPStan
cyc</th><th>instr</th><th>wall</th><th>phpbench</th><th>phpbench
JIT</th></tr></thead><tbody><tr><td>gcc-nts</td><t…<table><thead><tr><th>build</th><th>PHPStan
cyc</th><th>instr</th><th>wall</th><th>phpbench</th><th>phpbench
JIT</th></tr></thead><tbody><tr><td>gcc-nts</td><td>n/a</td><td>11.01B</td><td>113.91s</td><td>1065687</td><td>1793058</td></tr><tr><td>gcc-zts</td><td>n/a</td><td>11.23B</td><td>115.47s</td><td>1033552</td><td>1767360</td></tr><tr><td>gcc-zts-patch</td><td>n/a</td><td>11.15B</td><td>116.34s</td><td>1005077</td><td>1713823</td></tr><tr><td>clang-nts</td><td>n/a</td><td>11.38B</td><td>112.92s</td><td>1125988</td><td>1795096</td></tr><tr><td>clang-zts</td><td>n/a</td><td>11.64B</td><td>114.67s</td><td>1087697</td><td>1757513</td></tr><tr><td>clang-zts-patch</td><td>n/a</td><td>11..53B</td><td>112.22s</td><td>1145744</td><td>1793294</td></tr></tbody></table>
### LLM disclosure
claude
- adjusted the JIT
- transformed the raw bash results into these (hopefully beautiful) html tables