basic working implementation with manual calls

Makefile

@@ -2,9 +2,10 @@
 # libprof Makefile
 
 cflags := -m3 -mb -ffreestanding -nostdlib -fstrict-volatile-bitfields -Wall \
-          -Wextra -Os -std=c11
+          -Wextra -Os -I .
-target := sh3eb-elf
+target ?= sh3eb-elf
 lib    := libprof.a
+header := libprof.h
 
 prefix := $(shell $(target)-gcc -print-search-dirs | grep install \
                   | sed 's/install: //')
@@ -42,3 +43,4 @@ distclean: clean
 
 install:
 	cp $(lib) $(prefix)
+	cp $(header) $(prefix)/include

libprof.c

@@ -1,2 +1,73 @@
 #include <gint/timer.h>
+#include <gint/clock.h>
 #include <gint/mpu/tmu.h>
+#include <gint/std/stdlib.h>
+#include <gint/hardware.h>
+
+#include <libprof.h>
+
+/* Recursion depth of each function currently being executed */
+uint8_t *prof_rec = NULL;
+/* Time that has elapsed within each function; the value for a given function
+   is only relevant when it is not executing, due to optimizations */
+uint32_t *prof_elapsed = NULL;
+/* Timer counter */
+uint32_t volatile *prof_tcnt = NULL;
+/* Timer ID */
+static int prof_timer;
+
+/* prof_init(): Initialize the profiler's data and timer */
+int prof_init(int n, int timer)
+{
+	if((unsigned)timer >= 3) return 1;
+
+	prof_rec     = malloc(n * sizeof *prof_rec);
+	prof_elapsed = malloc(n * sizeof *prof_elapsed);
+
+	int status = timer_setup(timer, 0xffffffff, timer_Po_4, NULL, NULL);
+
+	if(!prof_rec || !prof_elapsed || status < 0)
+	{
+		prof_quit();
+		return 1;
+	}
+
+	/* Fix the configuration done by gint by disabling the interrupt */
+	if(isSH3())
+	{
+		SH7705_TMU.TMU[timer].TCR.UNIE = 0;
+		prof_tcnt = &SH7705_TMU.TMU[timer].TCNT;
+	}
+	else
+	{
+		SH7305_TMU.TMU[timer].TCR.UNIE = 0;
+		prof_tcnt = &SH7305_TMU.TMU[timer].TCNT;
+	}
+
+	timer_start(timer);
+	prof_timer = timer;
+
+	return 0;
+}
+
+/* prof_quit(): Free the profiler's data and timer */
+void prof_quit(void)
+{
+	timer_stop(prof_timer);
+
+	free(prof_rec);
+	free(prof_time);
+}
+
+//---
+//	Post-measurement analysis
+//---
+
+/* prof_time(): Time spent in a given context, in microseconds */
+uint32_t prof_time(int ctx)
+{
+	int Pphi = clock_freq()->Pphi_f;
+	uint64_t time = prof_elapsed[ctx];
+
+	return (time * 4 * 1000000) / Pphi;
+}

libprof.h

@@ -0,0 +1,75 @@
+//---
+//	libprof: A manual profiling library for gint
+//---
+
+#ifndef LIBPROF_LIBPROF
+#define LIBPROF_LIBPROF
+
+#include <stdint.h>
+
+//---
+//	Initialization
+//---
+
+/* prof_init(): Initialize the profiler's data and timer
+
+   Initializes [prof_rec] and [prof_time] (see below) with enough elements to
+   hold all the context IDs. Context IDs should be numbered from 0 to [n-1];
+   due to speed requirements array bounds are not checked so be careful.
+
+   Also starts a timer to count time. The timer ID must be set to 0, 1 or 2 as
+   the standard TMU is the most tweakable and precise. libprof automatically
+   selects an accurate timer configuration.
+
+   @n      Number of different contexts (functions) that will be measured
+   @timer  Timer ID, see <gint/timer.h> to select one
+   Returns non-zero if a setup error occurs. */
+int prof_init(int n, int timer);
+
+/* prof_quit(): Free the profiler's data and timer */
+void prof_quit(void);
+
+//---
+//	Runtime time measurement
+//---
+
+/* Recursion depth of each function currently being executed */
+extern uint8_t *prof_rec;
+/* Time that has elapsed within each function; the value for a given function
+   is only relevant when it is not executing, due to optimizations */
+extern uint32_t *prof_elapsed;
+/* Timer counter */
+extern uint32_t volatile *prof_tcnt;
+
+/* prof_enter(): Start counting time for a function
+   This macro should be called at the start of the context scope. If the
+   function was already executing then the deepest instance in the stack is
+   used instead of creating a new counter. */
+#define prof_enter(ctx) {					\
+	if(!prof_rec[ctx]++) prof_elapsed[ctx] += *prof_tcnt;	\
+}
+
+/* prof_leave(): Stop counting time for a function
+   This should be called at the end of the context scope; it only actually
+   stops if there is no deeper instance of the context in the stack. If there
+   are not as exactly as many prof_leave()'s as prof_enter()'s then the
+   resulting time measure will not be relevant at all. */
+#define prof_leave(ctx) {					\
+	if(!--prof_rec[ctx]) prof_elapsed[ctx] -= *prof_tcnt;	\
+}
+
+/* prof_clear(): Clear a context's counter
+   This operation is defined only if the context is not being profiled. */
+#define prof_clear(ctx) {		\
+	prof_elapsed[ctx] = 0;		\
+}
+
+//---
+//	Post-measurement analysis
+//---
+
+/* prof_time(): Time spent in a given context, in microseconds
+   Should only be called when the context is not currently executing. */
+uint32_t prof_time(int ctx);
+
+#endif /* LIBPROF_LIBPROF */