diff --git a/device/device.c b/device/device.c
index 0eda15e..89924e8 100644
--- a/device/device.c
+++ b/device/device.c
@@ -151,10 +151,11 @@ void device_destroy(struct cen64_device *device, const char *cart_path) {
 
     uint32_t i;
     for (i = 0; i < 8 * 1024 * 1024; i++) {
-      uint64_t sample = get_profile_sample(device->vr4300, i);
-      if (sample < 10)
+      const uint64_t sample = get_profile_sample(device->vr4300, i);
+      const uint64_t l1d_sample = get_profile_sample(device->vr4300, i + (8 * 1024 * 1024));
+      if (sample < 10 && l1d_sample < 10)
         continue;
-      fprintf(f, "%x %lu\n", i + 0x80000000, sample);
+      fprintf(f, "%x %lu %lu\n", i + 0x80000000, sample, l1d_sample);
     }
 
     fclose(f);
diff --git a/util/cen64-profile2callgrind.cpp b/util/cen64-profile2callgrind.cpp
index 601d78f..d078b73 100644
--- a/util/cen64-profile2callgrind.cpp
+++ b/util/cen64-profile2callgrind.cpp
@@ -130,11 +130,11 @@ int main(int argc, char **argv) {
 
 	fprintf(out, "# callgrind format\n");
 	fprintf(out, "cmd: %s\n", argv[2]);
-	fprintf(out, "events: instructions\n\n");
+	fprintf(out, "events: instructions l1d-misses\n\n");
 
 	fprintf(out, "ob=%s\n\n", argv[2]);
 
-	uint64_t summary = 0;
+	uint64_t summary = 0, l1d_summary = 0;
 
 	// We have a list of functions, now turn the samples into C line info
 	f = fopen(argv[1], "r");
@@ -142,11 +142,12 @@ int main(int argc, char **argv) {
 
 	while (fgets(buf, PATH_MAX, f)) {
 		uint32_t addr;
-		uint64_t num;
-		if (sscanf(buf, "%x %lu", &addr, &num) != 2)
+		uint64_t num, l1d_num;
+		if (sscanf(buf, "%x %lu %lu", &addr, &num, &l1d_num) != 3)
 			die("Malformed profile file\n");
 
 		summary += num;
+		l1d_summary += l1d_num;
 
 		found = false;
 		bfd_map_over_sections(bin, findsym, &addr);
@@ -161,20 +162,20 @@ int main(int argc, char **argv) {
 			}
 
 			fprintf(out, "fn=%s\n", funcname);
-			fprintf(out, "%u %lu\n\n", lineno, num);
+			fprintf(out, "%u %lu %lu\n\n", lineno, num, l1d_num);
 		} else {
 			fprintf(out, "fl=??\n");
 			map<uint32_t, string>::const_iterator it = funcs.lower_bound(addr);
 			it--;
 			fprintf(out, "fn=%s\n", it->second.c_str());
 
-			fprintf(out, "0 %lu\n\n", num);
+			fprintf(out, "0 %lu %lu\n\n", num, l1d_num);
 		}
 	}
 
 	bfd_close(bin);
 
-	fprintf(out, "totals: %lu\n", summary);
+	fprintf(out, "totals: %lu %lu\n", summary, l1d_summary);
 
 	fclose(f);
 	fclose(out);
diff --git a/vr4300/cpu.c b/vr4300/cpu.c
index 019fb59..81a269a 100644
--- a/vr4300/cpu.c
+++ b/vr4300/cpu.c
@@ -66,7 +66,7 @@ int vr4300_init(struct vr4300 *vr4300, struct bus_controller *bus, bool profilin
   vr4300->mi_regs[MI_INIT_MODE_REG] = 0x80;
 
   if (profiling)
-    vr4300->profile_samples = calloc(8 * 1024 * 1024, sizeof(uint64_t));
+    vr4300->profile_samples = calloc(2 * 8 * 1024 * 1024, sizeof(uint64_t));
   else
     vr4300->profile_samples = NULL;
 
diff --git a/vr4300/pipeline.c b/vr4300/pipeline.c
index 368f9e1..e085485 100644
--- a/vr4300/pipeline.c
+++ b/vr4300/pipeline.c
@@ -298,6 +298,12 @@ static int vr4300_dc_stage(struct vr4300 *vr4300) {
           request->paddr = paddr;
           exdc_latch->cached = cached;
 
+          if (vr4300->profile_samples) {
+            uint32_t idx = exdc_latch->common.pc - 0x80000000;
+            idx &= (8 * 1024 * 1024) - 1;
+            vr4300->profile_samples[idx + (8 * 1024 * 1024)]++;
+          }
+
           // Miss: stall for one cycle, then move to the DCM phase.
           vr4300->pipeline.cycles_to_stall = 0;
           vr4300->regs[PIPELINE_CYCLE_TYPE] = 6;
@@ -316,6 +322,13 @@ static int vr4300_dc_stage(struct vr4300 *vr4300) {
 
         request->paddr = paddr;
         exdc_latch->cached = cached;
+
+        if (vr4300->profile_samples) {
+          uint32_t idx = exdc_latch->common.pc - 0x80000000;
+          idx &= (8 * 1024 * 1024) - 1;
+          vr4300->profile_samples[idx + (8 * 1024 * 1024)]++;
+        }
+
         VR4300_DCM(vr4300);
         return 1;
       }