APPROVE Windows-21-4 Advisory, of course, it's your call. Andy> I plan to apply this to the windows branch and compile Andy> binaries without syntax-table support. Go ahead, with two caveats: (1) Please change all the references (especially the configure option) from "syntax table support" to "syntax [text] property support". We still support syntax tables, of course. What Matt's patch added is consulting tables attached to extents rather than buffer-local ones. (2) This patch still suffers from semantic redundancy and (the existing) bad design/implementation as mentioned before. And I'd still like to aim for a run-time rather than build-time implementation of the option. If I don't get time to work on it myself, then I'll give up on it, but for the moment I'm not going to put this patch in the 21.4 mainline. Of course I'd appreciate your advice on implementation. (Of course I've considered that given that Ben has already reworked 21.5, this may not be worth doing in 21.4. However, I think it's a worthwhile investment for me personally as I plan to do some other stuff with syntax and with syntax-property-like ideas in the future.) -- Institute of Policy and Planning Sciences http://turnbull.sk.tsukuba.ac.jp University of Tsukuba Tennodai 1-1-1 Tsukuba 305-8573 JAPAN My nostalgia for Icon makes me forget about any of the bad things. I don't have much nostalgia for Perl, so its faults I remember. Scott Gilbert c.l.py

QUERY Andy> Ok, attached is the patch I actually committed. These two hunks at least are regressions. `BUF_PT (buf)' does _not_ correspond to `pos_byte'. I don't yet know which is correct, but Ben specifically made the change from the !SYNTAX_TEXT_PROPERTIES form to the SYNTAX_TEXT_PROPERTIES forms---so I would bet on the form you just compiled _out_ of windows-21-4-9-2 .... Index: src/search.c =================================================================== RCS file: /pack/xemacscvs/XEmacs/xemacs/src/search.c,v retrieving revision 1.17.2.1.4.2 diff -u -r1.17.2.1.4.2 search.c --- src/search.c 2002/08/24 04:47:52 1.17.2.1.4.2 +++ src/search.c 2002/09/02 20:34:47 ＠＠ -896,15 +913,23 ＠＠ if (pos < limit) while (fastmap[(unsigned char) syntax_code_spec +#ifndef SYNTAX_TEXT_PROPERTIES + [(int) SYNTAX (syntax_table, + BUF_FETCH_CHAR + (buf, BUF_PT (buf)))]]) +#else /* SYNTAX_TEXT_PROPERTIES */ [(int) SYNTAX_FROM_CACHE (syntax_table, BI_BUF_FETCH_CHAR (buf, pos_byte))]]) +#endif /* SYNTAX_TEXT_PROPERTIES */ { pos++; INC_BYTIND (buf, pos_byte); if (pos >= limit) break; +#ifdef SYNTAX_TEXT_PROPERTIES UPDATE_SYNTAX_CACHE_FORWARD (pos); +#endif /* SYNTAX_TEXT_PROPERTIES */ } } else ＠＠ -914,12 +939,20 ＠＠ Bufpos savepos = pos_byte; pos--; DEC_BYTIND (buf, pos_byte); +#ifdef SYNTAX_TEXT_PROPERTIES UPDATE_SYNTAX_CACHE_BACKWARD (pos); +#endif /* SYNTAX_TEXT_PROPERTIES */ if (!fastmap[(unsigned char) syntax_code_spec +#ifndef SYNTAX_TEXT_PROPERTIES + [(int) SYNTAX (syntax_table, + BUF_FETCH_CHAR + (buf, BUF_PT (buf)))]]) +#else /* SYNTAX_TEXT_PROPERTIES */ [(int) SYNTAX_FROM_CACHE (syntax_table, BI_BUF_FETCH_CHAR (buf, pos_byte))]]) +#endif /* SYNTAX_TEXT_PROPERTIES */ { pos++; pos_byte = savepos; -- Institute of Policy and Planning Sciences http://turnbull.sk.tsukuba.ac.jp University of Tsukuba Tennodai 1-1-1 Tsukuba 305-8573 JAPAN My nostalgia for Icon makes me forget about any of the bad things. I don't have much nostalgia for Perl, so its faults I remember. Scott Gilbert c.l.py

Andy> My previous patch did not work since I had auto-generated Andy> it. This one has had some careful hand-editing and removes Andy> the bogosity. Cosmetic changes have been re-instated. On the one example I have where there's a gross difference between your patch and vanilla 21.4.9, disabling lookup-syntax-properties gives me about a 4:1 speedup in a --with-syntax-text-properties build (in a r21-4-9-3-windows tree with no debugging or error-checking). There's no subjectively perceptible difference with the default build. This is also the case for my usual build (nearly vanilla 21.4.9 --debug --error-checking=all --use-union-type), except that all times are about 40% longer. The 4:1 ratio for lookup-syntax-properties t vs nil holds, though. I thought you said setting lookup-syntax-properties nil didn't help? -- Institute of Policy and Planning Sciences http://turnbull.sk.tsukuba.ac.jp University of Tsukuba Tennodai 1-1-1 Tsukuba 305-8573 JAPAN My nostalgia for Icon makes me forget about any of the bad things. I don't have much nostalgia for Perl, so its faults I remember. Scott Gilbert c.l.py

Andy> It didn't make any difference to indentation IIRC. Font-lock Andy> maybe. Oh. Well, I have yet to see anything that makes a difference to indentation if I turn off font lock. The one reproducible recipe I have with a really clear difference (in TraverseSchema.java from Xerces, splitting line 8942, the declaration of reportSchemaError, after the comma, "good" is 14 seconds, "bad" is 65 seconds) shows no difference unless I enable fontification of large files. Then toggling either lookup-syntax-properties or font-lock-mode gives a big win on newline-and-indent. I haven't tried with any of the lazy-lock packages, but I don't think that will tell us much anyway. Interestingly enough, deletion of the whitespace is invariably much less costly (never more than about 2 seconds, although subjectively it seems slower when font-lock is active). -- Institute of Policy and Planning Sciences http://turnbull.sk.tsukuba.ac.jp University of Tsukuba Tennodai 1-1-1 Tsukuba 305-8573 JAPAN My nostalgia for Icon makes me forget about any of the bad things. I don't have much nostalgia for Perl, so its faults I remember. Scott Gilbert c.l.py

Andy> You mean it gets better when font-lock is turned on? That Andy> seems weird. Sorry, I mean "if both font-lock and lookup-syntax-properties are _on_, then turning off either one gives a factor-of-5 speedup". Narrowing the buffer also helps a lot, if the buffer is already fontified. But the TraverseSchema example still takes more than 10 seconds in a no-error-checking 21-4-9-3-windows (built on Linux), hardly less than my full-error-checking 21-4-9 version. Andy> 2 seconds per character deletion is waaaaaaaay too costly. Right. I just don't have a test case available where I can see more speedup from r21-4-9-3-windows than I can get from M-: (setq lookup-syntax-properties nil). -- Institute of Policy and Planning Sciences http://turnbull.sk.tsukuba.ac.jp University of Tsukuba Tennodai 1-1-1 Tsukuba 305-8573 JAPAN My nostalgia for Icon makes me forget about any of the bad things. I don't have much nostalgia for Perl, so its faults I remember. Scott Gilbert c.l.py