#10 Always use default compression window size

[Aymeric Blot]

Hi,

I have been conducting research experiments using genetic improvement, and while considering OptiPNG one of the variants we found seemed particularly interesting.
The idea was to consider a set a various images and try to see if it was possible to obtain a variant of OptiPNG that could run faster (in term of CPU instructions) while preserving the semantics of the original code (here, the image size decrease).

--- a/src/optipng/optim.c   2017-12-22 10:08:00.000000000 +0000
+++ b/src/optipng/optim.c   2020-04-29 18:41:24.680482538 +0100
@@ -1102,21 +1102,8 @@
         png_set_compression_mem_level(write_ptr, memory_level);
         png_set_compression_strategy(write_ptr, compression_strategy);
         png_set_filter(write_ptr, PNG_FILTER_TYPE_BASE, filter_table[filter]);
-        if (compression_strategy != Z_HUFFMAN_ONLY &&
-            compression_strategy != Z_RLE)
-        {
-            if (options.window_bits > 0)
-                png_set_compression_window_bits(write_ptr,
-                                                options.window_bits);
-        }
-        else
-        {
-#ifdef WBITS_8_OK
-            png_set_compression_window_bits(write_ptr, 8);
-#else
-            png_set_compression_window_bits(write_ptr, 9);
-#endif
-        }
+        if (options.window_bits > 0)
+            png_set_compression_window_bits(write_ptr, options.window_bits);

         /* Override the default libpng settings. */
         png_set_keep_unknown_chunks(write_ptr,

The attached patch removes the calls to png_set_compression_window_bits and results in a variant that is consistently faster in our dataset; in average, it uses 40% less CPU instructions and 20% less running time with no impact on image size.
The "faulty" line sets the size of the compression window to its minimal value, either 8 or 9 (256 or 512 bits), while its otherwise defaults to its maximal value (15, i.e., 32768 bits).

I am by no mean an expert, and the variant was automatically generated.
Is there a reason to prefer a small window size over a large one for Z_HUFFMAN_ONLY andZ_RLE, or is it possibly simply historical?
Additionally, why are user-provided values ignored in that case?

Regards,
Aymeric.