* module/statprof.scm (<state>): Instead of a boolean count-calls?,
treat the presence of a call-counts hash table as indicating a need to
count calls. That hash table maps callees to call counts. A "callee"
is either the IP of the entry of a program, the symbolic name of a
primitive, or the identity of a non-program.
New members "buffer" and "buffer-pos" replace "procedure-data".
We try to avoid analyzing things at runtime, instead just recording
the stack traces into a buffer. This will let us do smarter things
when post-processing.
(fresh-buffer, expand-buffer): New helpers.
(fresh-profiler-state): Adapt to <state> changes.
(sample-stack-procs): Instead of updating the procedure-data
table (which no longer exists), instead trace the stack into the
buffer.
(count-call): Update to update the call-counts table instead of the
procedure-data table.
(statprof-start, statprof-start): Adapt to call-counts change.
(call-data): Move lower in the file. Add "name" and "printable"
members, and no longer store a proc.
(source->string, program-debug-info-printable, addr->pdi)
(addr->printable): New helpers.
(stack-samples->procedure-data): New procedure to process stack trace
buffer into a hash table of the same format as the old procedure-data
table.
(statprof-fold-call-data, statprof-proc-call-data): Use
stack-samples->procedure-data instead of procedure-data.
(statprof-call-data->stats): Adapt to count-calls change.
(statprof-display, statprof-display-anomalies): Adapt.
* module/system/vm/program.scm: Export primitive?. Primitive
program-code doesn't map uniquely to the primitive, which may be of
interest to various meta-level utilities like statprof.
* test-suite/tests/statprof.test ("statistical sample counts within
expected range"): Increase number of calls, as the computer speed
increases and VM/compiler speed increases have meant that we get fewer
samples than before. Also, compare the maximum deviation to the
square root of the expected value. I don't actually know what the
right statistical check is here, but this is closer to an "all points
fall within a standard deviation" than our previous 30% check. Print
a nicer warning when the check fails.
* module/statprof.scm (statprof-start, statprof-stop): Take optional
state arg.
(statprof-reset): Return no values.
(statprof): Take port keyword arg. Since statprof-reset is now the
same as parameterizing profiler-state, there's no need to call
statprof-reset. Pass the state argument explicitly to statprof-start,
statprof-stop, and statprof-display.
* module/statprof.scm (<state>): Add field for the previous SIGPROF
handler.
(statprof-start, statprof-stop, statprof-reset): Instead of setting
the SIGPROF handler in statprof-reset, set it when the profiler
becomes active, and actually restore it when the profiler becomes
inactive.
* module/statprof.scm (statprof-display, statprof-display-anomalies)
(statprof-accumulated-time, statprof-sample-count)
(statprof-fetch-stacks, statprof-fetch-call-tree): Take optional state
argument.
(statprof-display-anomolies): Deprecate this mis-spelling.
(statprof): Just compute usecs for the period.
* doc/ref/posix.texi (Signals): Fix the documentation for setitimer; it
was wrong.
* libguile/scmsigs.c (pack_tv): New helper. Allow usecs >= 1e6.
(unpack_tv): New helper.
(scm_setitimer): Use the new helpers.
* test-suite/tests/signals.test: Add setitimer tests.
* module/texinfo.scm (read-char-data): Preserve newlines in @example and
similar environments in the case when the next line starts with an @.
* test-suite/tests/texinfo.test ("test-texinfo->stexinfo"): Add a test.
* module/statprof.scm (gcprof): No need to reset in gcprof; the fresh
profiler state and the parameterize handle that. Fix mistaken
set-vm-trace-level! as well.
* module/statprof.scm (<state>): The sampling frequency is actually a
period; label it as such, and express in microseconds instead of as a
pair. Likewise for remaining-prof-time.
(fresh-profiler-state): Adapt.
(reset-sigprof-timer): New helper.
(profile-signal-handler): Use the new helper.
(statprof-start): Use the new helper.
(statprof-stop): Here too.
(statprof-reset): Adapt to <state> change.
(gcprof): Set remaining prof time to 0.
* module/language/cps/simplify.scm (prune-continuations): Prune
continuations as a post-pass with a fresh DFG. Using a
pre-eta-conversion DFG as we were doing before missed some cases.
* lib/jit_x86.c: Rewrite previous patch to inline save/restore
because clobbering %ebx in x86 is treated as an error
(jit_x86.c:239:5: error: PIC register clobbered by 'ebx' in 'asm').
* module/statprof.scm (profile-signal-handler): Don't bother detecting
if we were in a count-call call or not; it doesn't matter, and we
should accumulate time in any case.
* module/statprof.scm (fresh-profiler-state): accumulated-time and
gc-time-taken are in jiffies, not seconds, so they are exact.
(statprof-accumulated-time): Divide by 1.0 so that we get a flonum.
Also refactor use of assq to get the gc-time-taken.
* libguile/vm-engine.c (BR_UNARY, BR_BINARY, BR_ARITHMETIC): A jump with
a zero offset is also a backward branch, in the sense that it's not a
forward branch.
("br"): We forgot to VM_HANDLE_INTERRUPTS here on backwards branches.
Oops!
* module/language/cps/simplify.scm (compute-eta-reductions): Avoid
trying to eta-reduce a jump-to-self, as in (let lp () (lp)). This
caused the compiler to hang.
* module/statprof.scm (get-call-data, sample-stack-procs): Take the
state as an argument.
(profile-signal-handler, count-call, statprof-proc-call-data)
(gcprof): Adapt.
* module/statprof.scm (statprof-reset, statprof-fold-call-data):
(statprof-proc-call-data, statprof-accumulated-time):
(statprof-sample-count): Refactor some things to use statprof-active?
instead of checking the profile level manually.
* module/statprof.scm (fresh-profiler-state): New helper.
(ensure-profiler-state): Use it.
(accumulate-time): No need to add 0.0 here.
(statprof-reset): Create a new state instead of mutating the existing
one.
* module/statprof.scm (existing-profiler-state): New helper, gets the
profiler state or fails if there is no state.
(sample-stack-procs, profile-signal-handler, count-call)
(statprof-fold-call-data, statprof-proc-call-data)
(statprof-call-data->stats, statprof-display)
(statprof-display-anomolies, statprof-accumulated-time)
(statprof-sample-count, statprof-fetch-stacks)
(statprof-fetch-call-tree): Use the new helper.
(statprof-active?): Don't create a state if there isn't one already.
* module/statprof.scm (<state>, profiler-state, ensure-profiler-state):
A mostly-mechanical refactor to encapsulate profiler state in a
parameter and a record instead of global variables.
* libguile/threads.c (thread_mark): There is a window in which the
thread has a handle but doesn't yet have the set of pointerless
freelists, so don't unconditionally dereference
t->pointerless_freelists.
* libguile/vm.c:
* libguile/vm-engine.c: Change the one use of
VM_ENABLE_PARANOID_ASSERTIONS to ASSERT. That becomes the one use of
VM_ENABLE_ASSERTIONS, so disable that too.
* libguile/vm-engine.c (ALLOC_FRAME): Fold CHECK_OVERFLOW into this
routine, and rework to not extend vp->sp until the stack has been
expanded.
* libguile/vm.c (vm_increase_sp): Likewise, don't extend vp->sp until
the stack has expanded.
(vm_expand_stack): Rework to take the new stack pointer as an
argument, and also to update vp->sp_max_since_gc and vp->sp.
* module/ice-9/boot-9.scm (catch): Signal an early error if the handler
or pre-unwind handler types aren't right. This is more important than
it was, given that we dispatch on type now when finding matching catch
clauses.
* libguile/vm.c (vm_expand_stack): Use the standard
scm_report_stack_overflow to signal stack overflow. This will avoid
running pre-unwind handlers.
* libguile/throw.h: Move scm_report_stack_overflow here.
* libguile/throw.c (catch): Define a version of catch in C.
(throw_without_pre_unwind): New helper. Besides serving as the
pre-boot "throw" binding, it allows stack overflow to throw without
running pre-unwind handlers.
(scm_catch, scm_catch_with_pre_unwind_handler)
(scm_with_throw_handler): Use the new catch in C.
(scm_report_stack_overflow): Moved from stackchk.c; throws an
unwind-only exception.
* libguile/stackchk.h:
* libguile/stackchk.c: Remove the scm_report_stack_overflow bits.
* libguile/throw.c (scm_init_throw): Define %exception-handler here.
* module/ice-9/boot-9.scm (%eh): Use the incoming %exception-handler,
and then delete it. This way we should be able to do unwind-only
exceptions from C.
* module/ice-9/boot-9.scm: Reimplement catch, throw, and
with-throw-handler in such a way that the exception handler is
threaded not through the exception-handling closures, but through a
data structure in the exception-handler fluid. This will allow us to
do unwind-only exception dispatch on stack overflow.
* lib/jit_x86.c: Rewrite incorrect inline assembly that could
truncate a variable in a callee save register. Now it simply
tells gcc that the register is clobbered, instead of using a
*32 bit* swap with a temporary variable. The problem only
happens when compiling with optimization.
* module/ice-9/boot-9.scm: Move error-handling initialization after
psyntax initialization. Only "throw" is used before psyntax, and both
throw and catch have pre-boot variants in C.
