Improve hc_matchfinder_longest_match() performance on Apple Silicon

[andrews05]

I was comparing performance on two MacBooks and was surprised at some of the results. I used the included benchmark program to compare the following hardware: 2015 MacBook Pro: 2.2 GHz Quad-Core Intel Core i7 2021 MacBook Pro: 8-core Apple M1 Pro

Level	Intel	M1 Pro	Difference
1	70ms	43ms	63%
2	107ms	59ms	81%
3	103ms	60ms	72%
4	108ms	61ms	77%
5	112ms	64ms	75%
6	120ms	74ms	62%
7	157ms	110ms	43%
8	316ms	270ms	17%
9	443ms	429ms	3%
10	2062ms	1340ms	54%
11	4481ms	3142ms	43%
12	9192ms	7309ms	26%

At level 9, M1 Pro is only 3% faster than a 6-year older Intel! I tried profiling it with xctrace and as best I can tell the performance hit comes from load_u32_unaligned (I can attach the trace output if that would be helpful). I can confirm that UNALIGNED_ACCESS_IS_FAST is set, but beyond that I haven't been able to work out why there's an issue. Do you have any ideas, or is it really the case that the Intel hardware is simply better at this?

[ebiggers]

At level 9 the bottleneck is usually the inner loop in hc_matchfinder_longest_match(). This is one very specific thing, so it is plausible that a CPU that is faster in general is not faster at this specific thing. This inner loop is also not very friendly to CPUs in general, as it doesn't allow for much instruction pipelining. But it's hard to improve.

[andrews05]

Yeah, it is hc_matchfinder_longest_match, but load_u32_unaligned seems to be taking around 1/3rd of the time within that function on the M1.

Intel:

M1:

[ebiggers]

It doesn't really make sense to talk about just the load_u32_unaligned part of hc_matchfinder_longest_match, since load_u32_unaligned is always inlined and compiles down to a single load. The fact that you are seeing a difference in the time accounted to load_u32_unaligned is probably just a side effect of whatever profiling tool you are using.