Refactor bcmath 2 #14110
Refactor bcmath 2 #14110
Conversation
| size_t scale_diff; | ||
| if (scale > num->n_scale && (scale_diff = scale - num->n_scale) > 8) { | ||
| memset(sptr, BCD_CHAR(0), scale_diff); | ||
| sptr += scale_diff; | ||
| } else { | ||
| for (size_t index = num->n_scale; index < scale; index++) { | ||
| *sptr++ = BCD_CHAR(0); | ||
| } |
There was a problem hiding this comment.
It's safe to always compute scale_diff, the compiler will do the optimization for you, and it will also be cleaner code.
Anyway, this part didn't make a real difference on my system, the UNEXPECTED part did however make an improvement.
To have more stable measurements, I recommend using hyperfine which is what I typically use. It runs the code multiple times and also computes a confidence interval. It can also do warmup rounds to reduce noise. I found the results from this tool to be more stable than only using the benchmark script. To use it here I would recommend to reduce the number of iterations in your benchmark script (to make sure the total runtime is below a second to avoid CPU throttling), and to remove the timing from the benchmark script and only rely on hyperfine instead.
|
After #14115 is merged, I'll check if it's still effective. |
|
I'm closing this PR because the results are no longer changing (in fact, they are a little slower). |
benchmark: #14076
before:
use UNEXPECTED:
use memset:
The difference is small, so it would be helpful if someone also could try it on.