Mercurial > hg > xemacs-beta
view src/profile.c @ 853:2b6fa2618f76
[xemacs-hg @ 2002-05-28 08:44:22 by ben]
merge my stderr-proc ws
make-docfile.c: Fix places where we forget to check for EOF.
code-init.el: Don't use CRLF conversion by default on process output. CMD.EXE and
friends work both ways but Cygwin programs don't like the CRs.
code-process.el, multicast.el, process.el: Removed.
Improvements to call-process-internal:
-- allows a buffer to be specified for input and stderr output
-- use it on all systems
-- implement C-g as documented
-- clean up and comment
call-process-region uses new call-process facilities; no temp file.
remove duplicate funs in process.el.
comment exactly how coding systems work and fix various problems.
open-multicast-group now does similar coding-system frobbing to
open-network-stream.
dumped-lisp.el, faces.el, msw-faces.el: Fix some hidden errors due to code not being defined at the right time.
xemacs.mak: Add -DSTRICT.
================================================================
ALLOW SEPARATION OF STDOUT AND STDERR IN PROCESSES
================================================================
Standard output and standard error can be processed separately in
a process. Each can have its own buffer, its own mark in that buffer,
and its filter function. You can specify a separate buffer for stderr
in `start-process' to get things started, or use the new primitives:
set-process-stderr-buffer
process-stderr-buffer
process-stderr-mark
set-process-stderr-filter
process-stderr-filter
Also, process-send-region takes a 4th optional arg, a buffer.
Currently always uses a pipe() under Unix to read the error output.
(#### Would a PTY be better?)
sysdep.h, sysproc.h, unexfreebsd.c, unexsunos4.c, nt.c, emacs.c, callproc.c, symsinit.h, sysdep.c, Makefile.in.in, process-unix.c: Delete callproc.c. Move child_setup() to process-unix.c.
wait_for_termination() now only needed on a few really old systems.
console-msw.h, event-Xt.c, event-msw.c, event-stream.c, event-tty.c, event-unixoid.c, events.h, process-nt.c, process-unix.c, process.c, process.h, procimpl.h: Rewrite the process methods to handle a separate channel for
error input. Create Lstreams for reading in the error channel.
Many process methods need change. In general the changes are
fairly clear as they involve duplicating what's used for reading
the normal stdout and changing for stderr -- although tedious,
as such changes are required throughout the entire process code.
Rewrote the code that reads process output to do two loops, one
for stdout and one for stderr.
gpmevent.c, tooltalk.c: set_process_filter takes an argument for stderr.
================================================================
NEW ERROR-TRAPPING MECHANISM
================================================================
Totally rewrite error trapping code to be unified and support more
features. Basic function is call_trapping_problems(), which lets
you specify, by means of flags, what sorts of problems you want
trapped. these can include
-- quit
-- errors
-- throws past the function
-- creation of "display objects" (e.g. buffers)
-- deletion of already-existing "display objects" (e.g. buffers)
-- modification of already-existing buffers
-- entering the debugger
-- gc
-- errors->warnings (ala suspended errors)
etc. All other error funs rewritten in terms of this one.
Various older mechanisms removed or rewritten.
window.c, insdel.c, console.c, buffer.c, device.c, frame.c: When creating a display object, added call to
note_object_created(), for use with trapping_problems mechanism.
When deleting, call check_allowed_operation() and note_object
deleted().
The trapping-problems code records the objects created since the
call-trapping-problems began. Those objects can be deleted, but
none others (i.e. previously existing ones).
bytecode.c, cmdloop.c: internal_catch takes another arg.
eval.c: Add long comments describing the "five lists" used to maintain
state (backtrace, gcpro, specbind, etc.) in the Lisp engine.
backtrace.h, eval.c: Implement trapping-problems mechanism, eliminate old mechanisms or
redo in terms of new one.
frame.c, gutter.c: Flush out the concept of "critical display section", defined by
the in_display() var. Use an internal_bind() to get it reset,
rather than just doing it at end, because there may be a non-local
exit.
event-msw.c, event-stream.c, console-msw.h, device.c, dialog-msw.c, frame.c, frame.h, intl.c, toolbar.c, menubar-msw.c, redisplay.c, alloc.c, menubar-x.c: Make use of new trapping-errors stuff and rewrite code based on
old mechanisms.
glyphs-widget.c, redisplay.h: Protect calling Lisp in redisplay.
insdel.c: Protect hooks against deleting existing buffers.
frame-msw.c: Use EQ, not EQUAL in hash tables whose keys are just numbers.
Otherwise we run into stickiness in redisplay because
internal_equal() can QUIT.
================================================================
SIGNAL, C-G CHANGES
================================================================
Here we change the way that C-g interacts with event reading. The
idea is that a C-g occurring while we're reading a user event
should be read as C-g, but elsewhere should be a QUIT. The former
code did all sorts of bizarreness -- requiring that no QUIT occurs
anywhere in event-reading code (impossible to enforce given the
stuff called or Lisp code invoked), and having some weird system
involving enqueue/dequeue of a C-g and interaction with Vquit_flag
-- and it didn't work.
Now, we simply enclose all code where we want C-g read as an event
with {begin/end}_dont_check_for_quit(). This completely turns off
the mechanism that checks (and may remove or alter) C-g in the
read-ahead queues, so we just get the C-g normal.
Signal.c documents this very carefully.
cmdloop.c: Correct use of dont_check_for_quit to new scheme, remove old
out-of-date comments.
event-stream.c: Fix C-g handling to actually work.
device-x.c: Disable quit checking when err out.
signal.c: Cleanup. Add large descriptive comment.
process-unix.c, process-nt.c, sysdep.c: Use QUIT instead of REALLY_QUIT.
It's not necessary to use REALLY_QUIT and just confuses the issue.
lisp.h: Comment quit handlers.
================================================================
CONS CHANGES
================================================================
free_cons() now takes a Lisp_Object not the result of XCONS().
car and cdr have been renamed so that they don't get used directly;
go through XCAR(), XCDR() instead.
alloc.c, dired.c, editfns.c, emodules.c, fns.c, glyphs-msw.c, glyphs-x.c, glyphs.c, keymap.c, minibuf.c, search.c, eval.c, lread.c, lisp.h: Correct free_cons calling convention: now takes Lisp_Object,
not Lisp_Cons
chartab.c: Eliminate direct use of ->car, ->cdr, should be black box.
callint.c: Rewrote using EXTERNAL_LIST_LOOP to avoid use of Lisp_Cons.
================================================================
USE INTERNAL-BIND-*
================================================================
eval.c: Cleanups of these funs.
alloc.c, fileio.c, undo.c, specifier.c, text.c, profile.c, lread.c, redisplay.c, menubar-x.c, macros.c: Rewrote to use internal_bind_int() and internal_bind_lisp_object()
in place of whatever varied and cumbersome mechanisms were
formerly there.
================================================================
SPECBIND SANITY
================================================================
backtrace.h: - Improved comments
backtrace.h, bytecode.c, eval.c: Add new mechanism check_specbind_stack_sanity() for sanity
checking code each time the catchlist or specbind stack change.
Removed older prototype of same mechanism.
================================================================
MISC
================================================================
lisp.h, insdel.c, window.c, device.c, console.c, buffer.c: Fleshed out authorship.
device-msw.c: Correct bad Unicode-ization.
print.c: Be more careful when not initialized or in fatal error handling.
search.c: Eliminate running_asynch_code, an FSF holdover.
alloc.c: Added comments about gc-cons-threshold.
dialog-x.c: Use begin_gc_forbidden() around code to build up a widget value
tree, like in menubar-x.c.
gui.c: Use Qunbound not Qnil as the default for
gethash.
lisp-disunion.h, lisp-union.h: Added warnings on use of VOID_TO_LISP().
lisp.h: Use ERROR_CHECK_STRUCTURES to turn on
ERROR_CHECK_TRAPPING_PROBLEMS and ERROR_CHECK_TYPECHECK
lisp.h: Add assert_with_message.
lisp.h: Add macros for gcproing entire arrays. (You could do this before
but it required manual twiddling the gcpro structure.)
lisp.h: Add prototypes for new functions defined elsewhere.
author | ben |
---|---|
date | Tue, 28 May 2002 08:45:36 +0000 |
parents | 6728e641994e |
children | 37bdd24225ef |
line wrap: on
line source
/* Why the hell is XEmacs so fucking slow? Copyright (C) 1996 Ben Wing. Copyright (C) 1998 Free Software Foundation, Inc. This file is part of XEmacs. XEmacs is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2, or (at your option) any later version. XEmacs is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with XEmacs; see the file COPYING. If not, write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */ #include <config.h> #include "lisp.h" #include "backtrace.h" #include "bytecode.h" #include "elhash.h" #include "hash.h" #include "syssignal.h" #include "systime.h" #ifndef HAVE_SETITIMER #error Sorry charlie. We need a scalpel and all we have is a lawnmower. #endif /* We implement our own profiling scheme so that we can determine things like which Lisp functions are occupying the most time. Any standard OS-provided profiling works on C functions, which is somewhat useless. The basic idea is simple. We set a profiling timer using setitimer (ITIMER_PROF), which generates a SIGPROF every so often. (This runs not in real time but rather when the process is executing or the system is running on behalf of the process.) When the signal goes off, we see what we're in, and add 1 to the count associated with that function. It would be nice to use the Lisp allocation mechanism etc. to keep track of the profiling information, but we can't because that's not safe, and trying to make it safe would be much more work than it's worth. Jan 1998: In addition to this, I have added code to remember call counts of Lisp funcalls. The profile_increase_call_count() function is called from Ffuncall(), and serves to add data to Vcall_count_profile_table. This mechanism is much simpler and independent of the SIGPROF-driven one. It uses the Lisp allocation mechanism normally, since it is not called from a handler. It may even be useful to provide a way to turn on only one profiling mechanism, but I haven't done so yet. --hniksic */ static struct hash_table *big_profile_table; Lisp_Object Vcall_count_profile_table; Fixnum default_profiling_interval; int profiling_active; /* The normal flag in_display is used as a critical-section flag and is not set the whole time we're in redisplay. */ int profiling_redisplay_flag; static Lisp_Object QSin_redisplay; static Lisp_Object QSin_garbage_collection; static Lisp_Object QSprocessing_events_at_top_level; static Lisp_Object QSunknown; /* We use inside_profiling to prevent the handler from writing to the table while another routine is operating on it. We also set inside_profiling in case the timeout between signal calls is short enough to catch us while we're already in there. */ static volatile int inside_profiling; /* Increase the value of OBJ in Vcall_count_profile_table hash table. If the hash table is nil, create it first. */ void profile_increase_call_count (Lisp_Object obj) { Lisp_Object count; if (NILP (Vcall_count_profile_table)) Vcall_count_profile_table = make_lisp_hash_table (100, HASH_TABLE_NON_WEAK, HASH_TABLE_EQ); count = Fgethash (obj, Vcall_count_profile_table, Qzero); if (!INTP (count)) count = Qzero; Fputhash (obj, make_int (1 + XINT (count)), Vcall_count_profile_table); } static SIGTYPE sigprof_handler (int signo) { /* Don't do anything if we are shutting down, or are doing a maphash or clrhash on the table. */ if (!inside_profiling && !preparing_for_armageddon) { Lisp_Object fun; /* If something below causes an error to be signaled, we'll not correctly reset this flag. But we'll be in worse shape than that anyways, since we'll longjmp back to the last condition case. */ inside_profiling = 1; if (profiling_redisplay_flag) fun = QSin_redisplay; else if (gc_in_progress) fun = QSin_garbage_collection; else if (backtrace_list) { fun = *backtrace_list->function; if (!SYMBOLP (fun) && !COMPILED_FUNCTIONP (fun) && !SUBRP (fun) && !CONSP (fun)) fun = QSunknown; } else fun = QSprocessing_events_at_top_level; { /* #### see comment about memory allocation in start-profiling. Allocating memory in a signal handler is BAD BAD BAD. If you are using the non-mmap rel-alloc code, you might lose because of this. Even worse, if the memory allocation fails, the `error' generated whacks everything hard. */ long count; const void *vval; if (gethash (LISP_TO_VOID (fun), big_profile_table, &vval)) count = (long) vval; else count = 0; count++; vval = (const void *) count; puthash (LISP_TO_VOID (fun), (void *) vval, big_profile_table); } inside_profiling = 0; } } DEFUN ("start-profiling", Fstart_profiling, 0, 1, 0, /* Start profiling, with profile queries every MICROSECS. If MICROSECS is nil or omitted, the value of `default-profiling-interval' is used. You can retrieve the recorded profiling info using `get-profiling-info'. Starting and stopping profiling does not clear the currently recorded info. Thus you can start and stop as many times as you want and everything will be properly accumulated. */ (microsecs)) { /* This function can GC */ int msecs; struct itimerval foo; /* #### The hash code can safely be called from a signal handler except when it has to grow the hash table. In this case, it calls realloc(), which is not (in general) re-entrant. We'll just be sleazy and make the table large enough that it (hopefully) won't need to be realloc()ed. */ if (!big_profile_table) big_profile_table = make_hash_table (10000); if (NILP (microsecs)) msecs = default_profiling_interval; else { CHECK_NATNUM (microsecs); msecs = XINT (microsecs); } if (msecs <= 0) msecs = 1000; set_timeout_signal (SIGPROF, sigprof_handler); foo.it_value.tv_sec = 0; foo.it_value.tv_usec = msecs; EMACS_NORMALIZE_TIME (foo.it_value); foo.it_interval = foo.it_value; profiling_active = 1; inside_profiling = 0; qxe_setitimer (ITIMER_PROF, &foo, 0); return Qnil; } DEFUN ("stop-profiling", Fstop_profiling, 0, 0, 0, /* Stop profiling. */ ()) { /* This function does not GC */ struct itimerval foo; foo.it_value.tv_sec = 0; foo.it_value.tv_usec = 0; foo.it_interval = foo.it_value; qxe_setitimer (ITIMER_PROF, &foo, 0); profiling_active = 0; set_timeout_signal (SIGPROF, fatal_error_signal); return Qnil; } struct get_profiling_info_closure { Lisp_Object accum; }; static int get_profiling_info_maphash (const void *void_key, void *void_val, void *void_closure) { /* This function does not GC */ Lisp_Object key; struct get_profiling_info_closure *closure = (struct get_profiling_info_closure *) void_closure; EMACS_INT val; key = VOID_TO_LISP (void_key); val = (EMACS_INT) void_val; closure->accum = Fcons (Fcons (key, make_int (val)), closure->accum); return 0; } DEFUN ("get-profiling-info", Fget_profiling_info, 0, 0, 0, /* Return the profiling info as an alist. */ ()) { /* This function does not GC */ struct get_profiling_info_closure closure; closure.accum = Qnil; if (big_profile_table) { int count = internal_bind_int ((int *) &inside_profiling, 1); maphash (get_profiling_info_maphash, big_profile_table, &closure); unbind_to (count); } return closure.accum; } static int mark_profiling_info_maphash (const void *void_key, void *void_val, void *void_closure) { Lisp_Object key; key = VOID_TO_LISP (void_key); mark_object (key); return 0; } void mark_profiling_info (void) { /* This function does not GC */ if (big_profile_table) { inside_profiling = 1; maphash (mark_profiling_info_maphash, big_profile_table, 0); inside_profiling = 0; } } DEFUN ("clear-profiling-info", Fclear_profiling_info, 0, 0, "", /* Clear out the recorded profiling info. */ ()) { /* This function does not GC */ if (big_profile_table) { inside_profiling = 1; clrhash (big_profile_table); inside_profiling = 0; } if (!NILP (Vcall_count_profile_table)) Fclrhash (Vcall_count_profile_table); return Qnil; } DEFUN ("profiling-active-p", Fprofiling_active_p, 0, 0, 0, /* Return non-nil if profiling information is currently being recorded. */ ()) { return profiling_active ? Qt : Qnil; } void syms_of_profile (void) { DEFSUBR (Fstart_profiling); DEFSUBR (Fstop_profiling); DEFSUBR (Fget_profiling_info); DEFSUBR (Fclear_profiling_info); DEFSUBR (Fprofiling_active_p); } void vars_of_profile (void) { DEFVAR_INT ("default-profiling-interval", &default_profiling_interval /* Default CPU time in microseconds between profiling sampling. Used when the argument to `start-profiling' is nil or omitted. Note that the time in question is CPU time (when the program is executing or the kernel is executing on behalf of the program) and not real time. */ ); default_profiling_interval = 1000; DEFVAR_LISP ("call-count-profile-table", &Vcall_count_profile_table /* The table where call-count information is stored by the profiling primitives. This is a hash table whose keys are funcallable objects, and whose values are their call counts (integers). */ ); Vcall_count_profile_table = Qnil; inside_profiling = 0; QSin_redisplay = build_msg_string ("(in redisplay)"); staticpro (&QSin_redisplay); QSin_garbage_collection = build_msg_string ("(in garbage collection)"); staticpro (&QSin_garbage_collection); QSunknown = build_msg_string ("(unknown)"); staticpro (&QSunknown); QSprocessing_events_at_top_level = build_msg_string ("(processing events at top level)"); staticpro (&QSprocessing_events_at_top_level); }