Profiler improvements: more-accurate timers, overflow checks.

* profiler.c: Don't include stdio.h, limits.h, sys/time.h,
signal.h, setjmp.h.  Include systime.h instead.
(saturated_add): New function.
(record_backtrace, current_sample_interval): Use EMACS_INT, not size_t.
(record_backtrace, handle_profiler_signal): Saturate on fixnum overflow.
(profiler_timer, profiler_timer_ok) [HAVE_TIMER_SETTIME]:
New static vars.
(enum profiler_cpu_running): New enumn.
(profiler_cpu_running): Now of that enum type, not bool.
All uses changed to store the new value.
(handle_profiler_signal): Rename from sigprof_handler_1,
for consistency with other handlers.  Do not check whether
cpu_log is a hash-table if garbage collecting, since it
doesn't matter in that case.
(deliver_profiler_signal): Rename from sigprof_handler,
for consistency with other handlers.
(setup_cpu_timer): New function, with much of what used to be in
Fprofiler_cpu_start.  Check for out-of-range argument.
Prefer timer_settime if available, and prefer
thread cputime clocks, then process cputime clocks, then
monotonic clocks, to the old realtime clock.  Use make_timeval
to round more-correctly when falling back to setitimer.
(Fprofiler_cpu_start): Use it.
(Fprofiler_cpu_stop): Prefer timer_settime if available.
Don't assume that passing NULL as the 2nd argument of setitimer
is the same as passing a pointer to all-zero storage.
Ignore SIGPROF afterwards.
(malloc_probe): Saturate at MOST_POSITIVE_FIXNUM.
* sysdep.c (emacs_sigaction_init): Also mask out SIGPROF in
non-fatal signal handlers.  Ignore SIGPROF on startup.
* syssignal.h (PROFILER_CPU_SUPPORT): Define this macro here, not
in profiler.c, since sysdep.c now uses it.

1 files changed, 34 insertions, 0 deletions

diff --git a/src/ChangeLog b/src/ChangeLog
index fd4319eb4b0..97f955e7136 100644
--- a/src/ChangeLog
+++ b/src/ChangeLog
@@ -1,5 +1,39 @@
 2012-09-30  Paul Eggert  <eggert@cs.ucla.edu>
 
+        Profiler improvements: more-accurate timers, overflow checks.
+        * profiler.c: Don't include stdio.h, limits.h, sys/time.h,
+        signal.h, setjmp.h.  Include systime.h instead.
+        (saturated_add): New function.
+        (record_backtrace, current_sample_interval): Use EMACS_INT, not size_t.
+        (record_backtrace, handle_profiler_signal): Saturate on fixnum overflow.
+        (profiler_timer, profiler_timer_ok) [HAVE_TIMER_SETTIME]:
+        New static vars.
+        (enum profiler_cpu_running): New enumn.
+        (profiler_cpu_running): Now of that enum type, not bool.
+        All uses changed to store the new value.
+        (handle_profiler_signal): Rename from sigprof_handler_1,
+        for consistency with other handlers.  Do not check whether
+        cpu_log is a hash-table if garbage collecting, since it
+        doesn't matter in that case.
+        (deliver_profiler_signal): Rename from sigprof_handler,
+        for consistency with other handlers.
+        (setup_cpu_timer): New function, with much of what used to be in
+        Fprofiler_cpu_start.  Check for out-of-range argument.
+        Prefer timer_settime if available, and prefer
+        thread cputime clocks, then process cputime clocks, then
+        monotonic clocks, to the old realtime clock.  Use make_timeval
+        to round more-correctly when falling back to setitimer.
+        (Fprofiler_cpu_start): Use it.
+        (Fprofiler_cpu_stop): Prefer timer_settime if available.
+        Don't assume that passing NULL as the 2nd argument of setitimer
+        is the same as passing a pointer to all-zero storage.
+        Ignore SIGPROF afterwards.
+        (malloc_probe): Saturate at MOST_POSITIVE_FIXNUM.
+        * sysdep.c (emacs_sigaction_init): Also mask out SIGPROF in
+        non-fatal signal handlers.  Ignore SIGPROF on startup.
+        * syssignal.h (PROFILER_CPU_SUPPORT): Define this macro here, not
+        in profiler.c, since sysdep.c now uses it.
+
         * sysdep.c (handle_fatal_signal): Bump backtrace size to 40.
         Suggested by Eli Zaretskii in
         <http://lists.gnu.org/archive/html/emacs-devel/2012-09/msg00796.html>.