Report processors with slow AVX

[klauspost]

Report processors with slow AVX

[klauspost]

@clausecker Definitely useful.

I will add it as part of my v2 refactor: #62

[clausecker]

As a specific real-world use case, the pospop library you contributed to could use a special kernel for Count8 that is fastest only on processors with slow AVX. However, without a way to detect this case, it is not possible to provide the kernel in a sensible manner. See clausecker/pospop#2 for details.

[klauspost]

@clausecker Added. You can see output here: https://pastebin.com/tWMEjDgL

[klauspost]

Search for AVXSLOW.

[clausecker]

Very nice! I've added a relevant Stack Overflow question in the meanwhile. I'll await the results of that. It is likely there are more processors with “slow” AVX.

[clausecker]

You might want to update the code a little following this discussion. On Sandy Bridge and Ivy Bridge, AVX is implemented using 256 bit ALUs, but loads and stores are done in two 128 bit passes. So the processors don't really have the same kind of problem as AMD processors.

[klauspost]

@clausecker Feel free to send a PR.