ext/bcmath: In the arm processor environment, NEON is used to use SIMD.

[SakiTakamachi]

Looking at the assembly code, it appears that the original bc_count_digits() was treated as inline on Arm because the amount of code was small.
This change increased the amount of code and made it no longer inline, which made it a bit slower, so I specified it as inline.

Benchmarks

When bc_count_digits() is made an inline function

1:

for ($i = 0; $i < 4000000; $i++) {
    bcadd('1.23456789', '-2.12345678', 10);
}

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/1.php
  Time (mean ± σ):     236.6 ms ±   1.2 ms    [User: 231.8 ms, System: 2.8 ms]
  Range (min … max):   234.6 ms … 239.2 ms    12 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/1.php
  Time (mean ± σ):     235.5 ms ±   1.9 ms    [User: 230.0 ms, System: 3.5 ms]
  Range (min … max):   233.2 ms … 241.0 ms    12 runs
 
Summary
  '/master/sapi/cli/php /mount/bc/1.php' ran
    1.00 ± 0.01 times faster than '/php-dev2/sapi/cli/php /mount/bc/1.php'

2:

for ($i = 0; $i < 4000000; $i++) {
    bcadd('12345678901234567890.12345678901234567890', '-212345678901234567890.12345678901234567890', 20);
}

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/2.php
  Time (mean ± σ):     321.0 ms ±   0.7 ms    [User: 315.9 ms, System: 3.1 ms]
  Range (min … max):   319.8 ms … 322.3 ms    10 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/2.php
  Time (mean ± σ):     423.4 ms ±   2.1 ms    [User: 418.1 ms, System: 3.2 ms]
  Range (min … max):   421.9 ms … 429.3 ms    10 runs
 
Summary
  '/php-dev2/sapi/cli/php /mount/bc/2.php' ran
    1.32 ± 0.01 times faster than '/master/sapi/cli/php /mount/bc/2.php'

3:

for ($i = 0; $i < 400000; $i++) {
    bcadd(str_repeat('12345678', 100), str_repeat('12345678', 100), 0);
}

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/3.php
  Time (mean ± σ):     178.0 ms ±   1.9 ms    [User: 173.1 ms, System: 3.0 ms]
  Range (min … max):   175.0 ms … 182.5 ms    16 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/3.php
  Time (mean ± σ):     572.6 ms ±   4.5 ms    [User: 566.7 ms, System: 3.6 ms]
  Range (min … max):   566.8 ms … 581.0 ms    10 runs
 
Summary
  '/php-dev2/sapi/cli/php /mount/bc/3.php' ran
    3.22 ± 0.04 times faster than '/master/sapi/cli/php /mount/bc/3.php'

If bc_count_digits() is not an inline function

1:

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/1.php
  Time (mean ± σ):     249.9 ms ±   1.0 ms    [User: 244.0 ms, System: 3.8 ms]
  Range (min … max):   248.6 ms … 251.2 ms    11 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/1.php
  Time (mean ± σ):     233.5 ms ±   1.0 ms    [User: 228.1 ms, System: 3.3 ms]
  Range (min … max):   231.1 ms … 234.6 ms    12 runs
 
Summary
  '/master/sapi/cli/php /mount/bc/1.php' ran
    1.07 ± 0.01 times faster than '/php-dev2/sapi/cli/php /mount/bc/1.php'

2:

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/2.php
  Time (mean ± σ):     332.8 ms ±   1.7 ms    [User: 327.7 ms, System: 3.0 ms]
  Range (min … max):   329.8 ms … 335.1 ms    10 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/2.php
  Time (mean ± σ):     423.5 ms ±   1.2 ms    [User: 418.5 ms, System: 2.8 ms]
  Range (min … max):   421.2 ms … 425.3 ms    10 runs
 
Summary
  '/php-dev2/sapi/cli/php /mount/bc/2.php' ran
    1.27 ± 0.01 times faster than '/master/sapi/cli/php /mount/bc/2.php'

3:

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/3.php
  Time (mean ± σ):     178.1 ms ±   0.4 ms    [User: 172.9 ms, System: 3.2 ms]
  Range (min … max):   177.5 ms … 178.7 ms    16 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/3.php
  Time (mean ± σ):     569.2 ms ±   2.3 ms    [User: 563.6 ms, System: 3.2 ms]
  Range (min … max):   566.3 ms … 573.4 ms    10 runs
 
Summary
  '/php-dev2/sapi/cli/php /mount/bc/3.php' ran
    3.20 ± 0.02 times faster than '/master/sapi/cli/php /mount/bc/3.php'

[SakiTakamachi]

Fixed a typo in the comment

[SakiTakamachi]

Ready for review

[nielsdos]

If it creates a slowdown on short inputs, do you know at what input length SIMD on NEON becomes faster than the regular loop?
I remember hearing that NEON has a large "startup overhead".

[SakiTakamachi]

@nielsdos
Measurements were taken under the condition that SIMD branches were used.
In this implementation, when SIMD can be used, it seems to be fast enough even with the minimum number of digits.

for ($i = 0; $i < 8000000; $i++) {
    $a = new BcMath\Number('1234567890123456');
}

Benchmark 1: /php-dev2/sapi/cli/php /mount/bc/0.php
  Time (mean ± σ):     301.5 ms ±   7.2 ms    [User: 295.1 ms, System: 4.2 ms]
  Range (min … max):   296.4 ms … 318.7 ms    10 runs
 
Benchmark 2: /master/sapi/cli/php /mount/bc/0.php
  Time (mean ± σ):     356.5 ms ±   2.0 ms    [User: 350.2 ms, System: 4.0 ms]
  Range (min … max):   353.8 ms … 360.6 ms    10 runs
 
Summary
  '/php-dev2/sapi/cli/php /mount/bc/0.php' ran
    1.18 ± 0.03 times faster than '/master/sapi/cli/php /mount/bc/0.php'

NEON does not have an instruction that corresponds to _mm_movemask_epi8.
Perhaps when I tried this previously, the implementation of the function equivalent to _mm_movemask_epi8 was poor, which resulted in overhead.
(In fact, I have confirmed that this time too, differences in implementation methods can result in a maximum of three times slower execution time.)

[nielsdos]

Would be interesting to know if there's a way to get the byte position without something like movemask. Anyway this is likely good enough for now.