@solardiz I have no idea what formats use this? My only testing was adding a temprary #warning showing that the code was indeed compiled in, then a full -test=0 -format=cpu. And bots seem happy. Consequently I don't know if there was a boost or regression...

I have no idea what formats use this?

In scalar builds, the md5crypt format uses this. Also some/most/all(?) dynamic formats that use MD5 will probably use this.

Looking at the dynamic formats stuff, quite possibly this is also frequently used even in SIMD builds, in cases where SIMD-enabled MD5 is somehow unusable/unused by a given dynamic format.

The main md5crypt format will have X2 in its algorithm name. Unfortunately, it looks like this reporting isn't propagated into the dynamic formats.

My only testing was adding a temprary #warning showing that the code was indeed compiled in

Was that in scalar or SIMD builds?

My only testing was adding a temprary #warning showing that the code was indeed compiled in

Was that in scalar or SIMD builds?

SIMD, and I too think the dynamic formats use this a lot.

I fixed step 48 now but testing with md5crypt indicate a regression (not bc step 48 but the whole thing) with 1-2% on M1 macbook. I can't remember whether the original change in the same file was ever benchmarked. Very slight regression (or just noise) seen on raw-md5. This should be benchmarked on some actual non-SIMD arch though.

EDIT: Obviously non-simd build now

md5crypt .. regression .. with 1-2% on M1

3% regression on super single thread or 8 threads, but 0-0.5% boost running 32 threads.

Besides speeds, I'd also look at size MD5_std.o.

This should be benchmarked on some actual non-SIMD arch though.

I suggest you try on cfarm SPARC machines https://portal.cfarm.net/machines/list/ as we don't have SIMD code for SPARC. On the other hand, the newer ones of those have MD5 instructions used by OpenSSL, so e.g. our md5crypt-long performs better than our md5crypt there.

Oh, also they have POWER7, where we broke our SIMD when we started requiring POWER8+ because of our use of 64-bit vector elements. We should ideally fix that separately, but meanwhile scalar code is used on POWER7.