Add comment on lzw.decoder.suffixes performance
diff --git a/std/lzw/decode_lzw.wuffs b/std/lzw/decode_lzw.wuffs
index f7ec898..038448a 100644
--- a/std/lzw/decode_lzw.wuffs
+++ b/std/lzw/decode_lzw.wuffs
@@ -17,7 +17,13 @@
pri status "?internal error: inconsistent I/O"
// TODO: move bulk data buffers like decoder.suffixes or decoder.output into
-// the workbuf?
+// the workbuf? The first attempt at this was a performance regression for
+// decoding all but the smallest GIFs. See these git commits for numbers:
+// - 49627b4 Flatten the lzw.decoder.suffixes array
+// - f877fb2 Use the workbuf instead of lzw.decoder.suffixes
+// - 85be5b9 Delete the obsolete lzw.decoder.suffixes array
+// and the roll back has combined numbers:
+// - 3056a84 Roll back 3 recent lzw.decoder.suffixes commits
pub const decoder_workbuf_len_max_incl base.u64 = 0
pub struct decoder?(
