Use single-step CPU call for 17% performance boost

Makefile

@@ -1,7 +1,7 @@
 CC = gcc
 LD = $(CC)
-PFLAGS = -g
-#PFLAGS = -O3
+#PFLAGS = -pg
+PFLAGS += -O3
 #PFLAGS += -DDEBUG_MAPPER
 #PFLAGS += -DDEBUG_RENDER
 #PFLAGS += -DDEBUG_PPU -DDEBUG_VRAM -DDEBUG_OAM

src/nes.c

@@ -76,26 +76,23 @@ void nes_reset(nes* sys) {
     nes_apu_reset(&sys->apu);
 }
 
-nes_ppu_Result nes_run(nes* sys, int master_cycles, int* run) {
-    int cpu_run = 0;
-    int cpu_cycles = (master_cycles + (nes_clock_cpu_div - 1)) /
-                     nes_clock_cpu_div;
-    int status = e6502_run(&sys->cpu, cpu_cycles, &cpu_run, 0);
-
-    master_cycles = cpu_run * nes_clock_cpu_div;
-    int ppu_cycles = master_cycles / nes_clock_ppu_div;
-    nes_ppu_Result result = nes_ppu_run(&sys->ppu, ppu_cycles);
-
-    if (result == ppu_Result_VBlank_On) {
-        e6502_set_nmi(&sys->cpu, 1);
-    } else if (result == ppu_Result_VBlank_Off) {
-        e6502_set_nmi(&sys->cpu, 0);
-    }
+nes_ppu_Result nes_step(nes* sys, int* run) {
+    nes_ppu_Result result = ppu_Result_Halt;
+
+    int cpu_run = e6502_step(&sys->cpu);
 
-    if (run) *run = master_cycles;
+    if (cpu_run > 0) {
+        int master_cycles = cpu_run * nes_clock_cpu_div;
+        int ppu_cycles = master_cycles / nes_clock_ppu_div;
+        result = nes_ppu_run(&sys->ppu, ppu_cycles);
+
+        if (result == ppu_Result_VBlank_On) {
+            e6502_set_nmi(&sys->cpu, 1);
+        } else if (result == ppu_Result_VBlank_Off) {
+            e6502_set_nmi(&sys->cpu, 0);
+        }
 
-    if (status < 0) {
-        result = ppu_Result_Halt;
+        if (run) *run = master_cycles;
     }
 
     return result;

src/nes.h

@@ -59,7 +59,7 @@ int nes_init(nes*);
 
 void nes_reset(nes*);
 
-nes_ppu_Result nes_run(nes*, int cycles, int* run);
+nes_ppu_Result nes_step(nes*, int* run);
 
 
 #endif // NES_H_

src/nese.c

@@ -97,7 +97,7 @@ int main(int argc, char* argv[]) {
 //        int last_frame_rendered = -1;
         for (int i = 0; i < n_loops && status == 0; ++i) {
             int run = 0;
-            nes_ppu_Result result = nes_run(&sys, 1U, &run);
+            nes_ppu_Result result = nes_step(&sys, &run);
             total_cycles += run;
 /*
             float us_run = (    run * 1000. * 1000. *