Files
wuffs/std/jpeg
Nigel Tao 8d9d6c1292 Performance mystery: load-bearing "no-op continue"
There is no semantic difference in this commit. There's no further code
after the if-else chain in the loop body. Nonetheless, adding a "no-op
continue" impacts performance: bad for clang and good for gcc.

Profiling suggests we spend less than 5% overall in
wuffs_jpeg__decoder__fill_bitstream.

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               241MB/s ± 0%  229MB/s ± 0%  -5.02%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  213MB/s ± 0%  205MB/s ± 0%  -4.13%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   386MB/s ± 0%  375MB/s ± 0%  -2.88%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11              341MB/s ± 0%  326MB/s ± 0%  -4.58%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         356MB/s ± 0%  339MB/s ± 0%  -4.90%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         484MB/s ± 0%  475MB/s ± 0%  -1.96%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            364MB/s ± 0%  343MB/s ± 0%  -5.83%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 234MB/s ± 0%  245MB/s ± 0%  +4.68%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    245MB/s ± 0%  254MB/s ± 0%  +3.35%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     500MB/s ± 0%  514MB/s ± 0%  +2.81%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                417MB/s ± 0%  434MB/s ± 0%  +4.18%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           442MB/s ± 0%  459MB/s ± 0%  +3.76%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           477MB/s ± 0%  486MB/s ± 0%  +1.99%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              450MB/s ± 0%  467MB/s ± 0%  +3.70%  (p=0.008 n=5+5)

It is a mystery why this happens. Note though that the "continue" in the
Wuffs code becomes a "goto" in the generated C code, because (in the
general case) Wuffs has "labeled continue/break", which can jump out of
a not-inner-most loop, and C99 does not. Wuffs' "generate C code" thus
uses a C "goto". A follow-up commit may try a C "continue" instead.

To reproduce:

----
$ git clone https://github.com/google/wuffs.git
$ cd wuffs
$ gcc --version | head -n 1
gcc (Debian 10.2.1-6) 10.2.1 20210110

$ # Check out this commit.
$ git checkout --quiet THE_HASH_OF_THIS_COMMIT
$ gcc -O3 -Wall -std=c99 test/c/std/jpeg.c -DWUFFS_MIMIC -ljpeg
$ ./a.out -bench -focus=wuffs_jpeg_decode_77k_24bpp | grep MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  176988 ns/op	     433.927 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  176849 ns/op	     434.268 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  176800 ns/op	     434.389 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  176844 ns/op	     434.279 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  176869 ns/op	     434.218 MB/s

$ # Check out the parent commit.
$ git checkout --quiet THE_HASH_OF_THIS_COMMIT^
$ gcc -O3 -Wall -std=c99 test/c/std/jpeg.c -DWUFFS_MIMIC -ljpeg
$ ./a.out -bench -focus=wuffs_jpeg_decode_77k_24bpp | grep MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  183848 ns/op	     417.734 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  183870 ns/op	     417.684 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  183833 ns/op	     417.770 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  183873 ns/op	     417.679 MB/s
Benchmarkwuffs_jpeg_decode_77k_24bpp/gcc10	    3000	  183843 ns/op	     417.746 MB/s
----
2023-06-23 17:02:14 +10:00
..

JPEG

JPEG (Joint Photographic Experts Group) is a lossy image compression format for opaque, still images.