criterion.rs
criterion.rs copied to clipboard
Diferents results for same assembler
I have been observing strange behavior in the benchmarks for a long time. The same code, identical assembler, gives different results depending on this built by hand or by derivation.
I know that it is ridiculous to think about it, but an example is the benchmarks of yarte. The solution by hand with fmt::Display
is slower than the solution with derivation, but the assembler generated is totally the same. https://github.com/botika/yarte/blob/master/benches/src/all.rs, use [email protected] or lower for reproduce.
Or the best is this one where it marks the same time doing nothing as doing something. https://github.com/Kogia-sima/sailfish/pull/29#issuecomment-662886879
Although it is ridiculous I have come to the conclusion that it is because of the derivation. But it looks like something deeper.
I don't know but people take it as a reference and since they don't look at the code and since they are the majority, you already know how Trump is.
Hi @botika. Think you could isolate the minimal amount of code that displays this behavior?