Adam Stylinski

Somewhat interesting is at level 6 the decompression speed is higher than level 3, but perhaps that speaks to having an inflate that is working more in inflate_fast rather than...

Heh, the sorting by size rather than throughput is throwing me off a bit. Looks like we don't do too much worse (in terms of compression throughput) than libdeflate at...

The adler32 checksum should be fixed at least for Altivec. Which compiler are you using? The POWER9 fix tests the GCC version and applies a manual inline asm call if...

> Why libdeflate is faster than vanilla zlib for decompression: https://github.com/ebiggers/libdeflate/blob/02dfa32da3ee3982c66278e714d2e21276dfb67b/lib/deflate_decompress.c#L32-L43 > Word accesses rather than byte accesses when copying matches Pretty sure we do this, at least. > Word...

If you change that macro to always be true does that fix things? If so we may be missing an adjustment for endianness when using the intrinsic.

My guess, and I don't have power9 systems to verify this but, gcc's fix for endianness inverted the need to switch between trailing and leading zero. Perhaps gcc flips which...

Either that or that fix for gcc for endianness was actually swapping the emitted instruction based on endianness.

I didn't write the test but this is clearly unsigned subtraction wrapping around. There should probably be termination after compr_len is 0 or compr_len should be subtracted by a clamped...

I'm half tempted to capture the function pointers at initialization time for comparison and crash and burn with a scary message if they differ but...if people have been screwing this...

I guess at the very least the warning that prints out here will help catch with a wide net projects that are using this functionality incorrectly. We also sparingly call...