Adler32 Hash; clang 7.0 is 1.3x slower than gcc 7.3

[nigeltao]

Bugzilla Link	35567
Version	7.0
OS	Linux
Attachments	Adler32 C implementation
CC	@adibiagio,@topperc,@RKSimon,@rotateright

Extended Description

adler32-standalone.c (attached, also at https://github.com/google/wuffs/blob/master/script/adler32-standalone.c) is some straight C code to calculate the standard Adler32 hash, used by the zlib format.

Measuring throughput on a recent Debian Testing system (clang 5.0 and gcc 7.2), Intel x86_64 Broadwell, suggests that clang's generated code is 1.3x slower.

$ clang-5.0 -O3 adler32-standalone.c; ./a.out
2311 MiB/s, clang 5.0.0 (tags/RELEASE_500/rc2)
$ gcc -O3 adler32-standalone.c; ./a.out
3052 MiB/s, gcc 7.2.1 20171025

The core loop (manually unrolled 8 times) is:

while (p < end0) {
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
s1 += ((uint32_t)(*p));
s2 += s1;
p++;
}

[RKSimon]

This might be to do with gcc being better at reducing integer multiplies for UREM:

https://godbolt.org/g/G7umh3

Clang ends up using a lot of i64 muls while gcc reduces them all to i32

[nigeltao]

UREM isn't part of the inner-most loop. As I said in the OP, the core loop is the "while (p < end0) { etc }" loop.

https://godbolt.org/g/bQgtKx is probably a better codegen comparison. gcc is certainly doing something more sophisticated wrt ordering loads and adds, and using more of the available registers.

[nigeltao]

clang-7 (released September 2018) is still 1.3x slower than gcc.

$ clang-7 -O3 adler32-standalone.c; ./a.out
2292 MiB/s, clang 7.0.0 (tags/RELEASE_700/final)
$ gcc -O3 adler32-standalone.c; ./a.out
3034 MiB/s, gcc 7.3.0

[tcShadowWalker]

It seems to me this issue is resolved in recent clang versions. (At least in clang-12, maybe earlier)

$ gcc-12 -o bin -O3 adler32.c && ./bin
    3005 MiB/s, gcc 12.3.0
$ clang-12 -o bin -O3 adler32.c && ./bin 
    3020 MiB/s, clang 12.0.1

I tried with clang-12, clang-13, clang-14 and clang-18, all giving similar results, at 3000 MiB/s.
On the gcc side, I tried gcc-11, gcc-12 and gcc-13, also all giving similar results.