diff --git a/m1/perf/sched.c b/m1/perf/sched.c
index 8d09e0c..49c6ed6 100644
--- a/m1/perf/sched.c
+++ b/m1/perf/sched.c
@@ -31,6 +31,8 @@
 #include <hw/pfpu.h>
 
 
+#define	REG_STATS
+
 #ifdef DEBUG
 #define	Dprintf printf
 #else
@@ -88,6 +90,9 @@ static struct sched_ctx {
 	struct list unscheduled; /* unscheduled insns */
 	struct list waiting;	/* insns waiting to be scheduled */
 	struct list ready[PFPU_PROGSIZE]; /* insns ready at nth cycle */
+#ifdef REG_STATS
+	int max_regs, curr_regs;	/* allocation statistics */
+#endif
 } *sc;
 
 
@@ -211,6 +216,13 @@ static int alloc_reg(struct insn *setter)
 	reg = list_pop(&sc->unallocated);
 	if (!reg)
 		abort();
+
+#ifdef REG_STATS
+	sc->curr_regs++;
+	if (sc->curr_regs > sc->max_regs)
+		sc->max_regs = sc->curr_regs;
+#endif
+
 	reg->vm_reg = vm_reg;
 	pfpu_reg = reg-sc->pfpu_regs;
 
@@ -238,6 +250,10 @@ static void put_reg(int vm_reg)
 
 	Dprintf("  free reg %d\n", regs[vm_idx].pfpu_reg);
 
+#ifdef REG_STATS
+	sc->curr_regs--;
+#endif
+
 	/*
 	 * Prepend so that register numbers stay small and bugs reveal
 	 * themselves more rapidly.
@@ -488,7 +504,13 @@ int gfpus_schedule(struct fpvm_fragment *frag, unsigned int *code,
 	init_scheduler_context(frag, reg);
 	memset(code, 0, PFPU_PROGSIZE*sizeof(*code));
 	res = schedule(code);
+
+#ifdef REG_STATS
+	printf("regs: %d/%d\n", sc->curr_regs, sc->max_regs);
+#endif
+
 	free(sc->regs);
+	free(sc);
 	if (res < 0)
 		return res;
 	if (frag->vector_mode)