http://cfis.savagexi.com en-us 40 Charlie Savage's Blog <p>This patch is much easier to read: <code></p> <pre> --- ruby_prof.c.orig 2006-06-09 23:06:22.000000000 +0200 +++ ruby_prof.c.new 2006-06-09 23:06:29.000000000 +0200 @@ -73,7 +73,7 @@ #define CLOCK_MODE_CLOCK 0 #define CLOCK_MODE_GETTIMEOFDAY 1 -#if defined(__GNUC__) &amp;&amp; (defined(__i386__) || defined(__powerpc__) || defined(__ppc__)) +#if defined(_WIN32) || (defined(__GNUC__) &amp;&amp; (defined(__i386__) || defined(__powerpc__) || defined(__ppc__))) #define CLOCK_MODE_CPU 2 static double cpu_frequency; #endif @@ -109,6 +109,9 @@ #ifdef CLOCK_MODE_CPU + +#if defined(__GNUC__) + static prof_clock_t cpu_get_clock() { @@ -130,6 +133,57 @@ #endif } +#elif defined(_WIN32) + +static prof_clock_t +cpu_get_clock() +{ + prof_clock_t cycles = 0; + + __asm + { + rdtsc + mov DWORD PTR cycles, eax + mov DWORD PTR [cycles + 4], edx + } + return cycles; +} + +#endif + + +/* The _WIN32 check is needed for msys (and maybe cygwin?) */ +#if defined(__GNUC__) &amp;&amp; !defined(_WIN32) + +double get_cpu_frequency() +{ + unsigned long long x, y; + + struct timespec ts; + ts.tv_sec = 0; + ts.tv_nsec = 500000000; + x = cpu_get_clock(); + nanosleep(&amp;ts, NULL); + y = cpu_get_clock(); + return (y - x) * 2; +} + +#elif defined(_WIN32) + +double get_cpu_frequency() +{ + unsigned long long x, y; + double frequency; + x = cpu_get_clock(); + /* Use the windows sleep function, not Ruby's */ + Sleep(500); + y = cpu_get_clock(); + frequency = 2*(y-x); + printf("clock speed: %f", frequency); + return frequency; +} +#endif + static double cpu_clock2sec(prof_clock_t c) { @@ -190,13 +244,13 @@ } return stack->ptr++; } - static inline prof_data_t * stack_pop(prof_stack_t *stack) { if (stack->ptr == stack->start) - rb_fatal("empty stack"); - return --stack->ptr; + return NULL; + else + return --stack->ptr; } static inline prof_data_t * @@ -459,6 +513,8 @@ now = get_clock(); data = stack_pop(stack); + if (data == NULL) + break; if (!NIL_P(klass)) { if (TYPE(klass) == T_ICLASS) { klass = RBASIC(klass)->klass; @@ -486,10 +542,10 @@ static VALUE prof_start(VALUE self) { - prof_data_t *data; - + prof_clock_t now; + if (stack_tbl) { - rb_raise(rb_eRuntimeError, "Prof.start was already called"); + rb_raise(rb_eRuntimeError, "Prof.start was already called"); } stack_tbl = stack_table_create(); class_tbl = rb_hash_new(); @@ -497,13 +553,17 @@ stack = stack_create(); stack_table_insert(stack_tbl, profiling_thread, stack); result_tbl = minfo_table_create(); + + now = get_clock(); toplevel = prof_result_create(Qnil, rb_intern("#toplevel")); + toplevel->count = 1; + toplevel->total_time = now; + toplevel->self_time = 0; + rb_add_event_hook(prof_event_hook, - RUBY_EVENT_CALL | RUBY_EVENT_RETURN | - RUBY_EVENT_C_CALL | RUBY_EVENT_C_RETURN); - data = stack_push(stack); - data->start_time = get_clock(); - data->child_cost = 0; + RUBY_EVENT_CALL | RUBY_EVENT_RETURN | + RUBY_EVENT_C_CALL | RUBY_EVENT_C_RETURN); + return Qnil; } @@ -528,19 +588,16 @@ static VALUE prof_stop(VALUE self) { - prof_data_t *data; VALUE result_list; prof_clock_t now; if (stack_tbl == NULL) { - rb_raise(rb_eRuntimeError, "Prof.start is not called yet"); + rb_raise(rb_eRuntimeError, "Prof.start is not called yet"); } + now = get_clock(); - stack_pop(stack); /* data for Prof.stop */ - data = stack_pop(stack); - toplevel->count = 1; - toplevel->total_time = now - data->start_time; - toplevel->self_time = toplevel->total_time - data->child_cost; + toplevel->total_time = now - toplevel->total_time; + st_foreach(stack_tbl, free_stack, 0); st_free_table(stack_tbl); stack_tbl = NULL; @@ -581,17 +638,8 @@ break; #ifdef CLOCK_MODE_CPU case CLOCK_MODE_CPU: - if (cpu_frequency == 0) { - unsigned long long x, y; - struct timespec ts; - - ts.tv_sec = 0; - ts.tv_nsec = 500000000; - x = cpu_get_clock(); - nanosleep(&amp;ts, NULL); - y = cpu_get_clock(); - cpu_frequency = (y - x) * 2; - } + if (cpu_frequency == 0) + cpu_frequency = get_cpu_frequency(); get_clock = cpu_get_clock; clock2sec = cpu_clock2sec; break; @@ -604,6 +652,9 @@ return val; } +#if defined(_WIN32) +__declspec(dllexport) +#endif void Init_prof() { </pre> <p></code></p> Sun, 11 Jun 2006 02:07:03 -0600 urn:uuid:8efd2ecd-3b5f-4c57-99ac-03af842495c6 http://cfis.savagexi.com/2006/06/09/porting-ruby-perf-to-windows#comment-2 <p>Hi Mauricio,</p> <p>Curious what problems you had reading the patch? Was it due to Unix line endings versus Windows, or something else?</p> <p>Great job on <a href="http://eigenclass.org/hiki.rb?rcov+0.5.0" rel="nofollow">rcov</a> - I use it almost every day.</p> Sun, 11 Jun 2006 21:11:44 -0600 urn:uuid:75fb189e-0532-4c8f-b3db-eb580807790c http://cfis.savagexi.com/2006/06/09/porting-ruby-perf-to-windows#comment-33 <p>I believe the patch is hard to read because if you look closely, the patch is &#8216;removing&#8217; and then &#8216;adding&#8217; the same lines, cancelling each other out - but taking up a whole lot of space.</p> <p>Removing the redundant lines, makes the patch much smaller like the above.</p> Wed, 14 Jun 2006 09:44:44 -0600 urn:uuid:dd862cd0-602d-4a3b-b1ee-b57c340ad368 http://cfis.savagexi.com/2006/06/09/porting-ruby-perf-to-windows#comment-34 <p>Thanks Paul - hadn&#8217;t noticed that. I&#8217;ve made some more changes in the last couple of days and will post them once they&#8217;re read.</p> Fri, 16 Jun 2006 02:27:44 -0600 urn:uuid:b32a862d-78ef-4dbc-8f94-1252c9b9d61c http://cfis.savagexi.com/2006/06/09/porting-ruby-perf-to-windows#comment-13