Minor optimizations to TotalSMS

main/CMakeLists.txt

@@ -21,4 +21,14 @@ idf_component_register(
         segments.lf 
 )
 
-target_compile_options(${COMPONENT_LIB} PRIVATE -Ofast -fjump-tables -ftree-switch-conversion -Wno-unused-function)
+target_compile_options(${COMPONENT_LIB} PRIVATE
+    -Ofast
+    -fjump-tables
+    -ftree-switch-conversion
+    -ftree-loop-im
+    -ftree-loop-ivcanon
+    -fivopts
+    -ftree-vectorize
+    -fvariable-expansion-in-unroller
+    -Wno-unused-function
+)

main/main.c

@@ -51,6 +51,9 @@ static uint64_t frame_time = 0;
 static uint64_t cpu_time = 0;
 static uint64_t vdp_time = 0;
 
+extern uint64_t vdp_display_enabled;
+extern uint64_t vdp_display_disabled;
+
 static xQueueHandle button_queue;
 static QueueHandle_t video_queue;
 
@@ -341,7 +344,7 @@ void main_loop() {
       continue;
     }
 
-    for (size_t i = 0; i < SMS_CPU_CLOCK / 30; i += sms.cpu.cycles) {
+    for (size_t i = 0; i < SMS_CPU_CLOCK / 60; i += sms.cpu.cycles) {
       //uint64_t cpu_start = esp_timer_get_time();
       z80_run();
       //uint64_t cpu_end = esp_timer_get_time();
@@ -383,13 +386,13 @@ void main_loop() {
       if (cpu_mhz < 3.579545) {
         set_overscan_border(0x00f0);
 
-        ESP_LOGW(TAG, "cpu_mhz: %.6f, fps: %lli, dropped: %lli, avg_frametime: %lli, avg_cputime: %lli, avg_vdptime: %lli",
-          cpu_mhz, frames, dropped_frames, frame_time / frames, cpu_time / frames, vdp_time / frames);
+        ESP_LOGW(TAG, "cpu_mhz: %.6f, fps: %lli, dropped: %lli, avg_frametime: %lli, avg_cputime: %lli, avg_vdptime: %lli, vdp_enabled: %lli, vdp_disabled: %lli",
+          cpu_mhz, frames, dropped_frames, frame_time / frames, cpu_time / frames, vdp_time / frames, vdp_display_enabled, vdp_display_disabled);
       } else {
         set_overscan_border(current_overscan_color);
 
-        ESP_LOGI(TAG, "cpu_mhz: %.6f, fps: %lli, dropped: %lli, avg_frametime: %lli, avg_cputime: %lli, avg_vdptime: %lli",
-          cpu_mhz, frames, dropped_frames, frame_time / frames, cpu_time / frames, vdp_time / frames);
+        ESP_LOGI(TAG, "cpu_mhz: %.6f, fps: %lli, dropped: %lli, avg_frametime: %lli, avg_cputime: %lli, avg_vdptime: %lli, vdp_enabled: %lli, vdp_disabled: %lli",
+          cpu_mhz, frames, dropped_frames, frame_time / frames, cpu_time / frames, vdp_time / frames, vdp_display_enabled, vdp_display_disabled);
       }
 
       frames = 0;
@@ -399,6 +402,9 @@ void main_loop() {
       cpu_time = 0;
       vdp_time = 0;
 
+      vdp_display_enabled = 0;
+      vdp_display_disabled = 0;
+
       start = esp_timer_get_time();
     }
   }

main/totalsms/bus.c

@@ -5,7 +5,6 @@
 #include <stddef.h>
 #include <stdint.h>
 #include <string.h>
-#include <assert.h>
 
 extern struct SMS_Core sms;
 
@@ -314,6 +313,7 @@ void mapper_update()
         case MAPPER_TYPE_OTHELLO:
             setup_mapper_othello();
             break;
+        default: __builtin_unreachable();
     }
 
     // map the bios is enabled (if we have it)
@@ -410,16 +410,17 @@ static void sega_mapper_fffx_write(const uint16_t addr, const uint8_t value)
                 sega_mapper_update_slot2();
             }
             break;
+        default: __builtin_unreachable();
     }
 }
 
-uint8_t SMS_read8(const uint16_t addr)
+uint8_t FORCE_INLINE SMS_read8(const uint16_t addr)
 {
     assert(sms.rmap[addr >> 10] && "NULL ptr in rmap!");
     return sms.rmap[addr >> 10][addr & 0x3FF];
 }
 
-void SMS_write8(const uint16_t addr, const uint8_t value)
+void FORCE_INLINE SMS_write8(const uint16_t addr, const uint8_t value)
 {
     // specific mapper writes
     switch (sms.cart.mapper_type)
@@ -462,21 +463,22 @@ void SMS_write8(const uint16_t addr, const uint8_t value)
         case MAPPER_TYPE_THE_CASTLE:
         case MAPPER_TYPE_OTHELLO:
             break;
+        default: __builtin_unreachable();
     }
 
     assert(sms.wmap[addr >> 10] && "NULL ptr in wmap!");
     sms.wmap[addr >> 10][addr & 0x3FF] = value;
 }
 
-uint16_t SMS_read16(const uint16_t addr)
+uint16_t FORCE_INLINE SMS_read16(const uint16_t addr)
 {
     const uint16_t lo = SMS_read8(addr + 0);
     const uint16_t hi = SMS_read8(addr + 1);
 
     return (hi << 8) | lo;
 }
 
-void SMS_write16(const uint16_t addr, const uint16_t value)
+void FORCE_INLINE SMS_write16(const uint16_t addr, const uint16_t value)
 {
     SMS_write8(addr + 0, value & 0xFF);
     SMS_write8(addr + 1, value >> 8);
@@ -634,6 +636,7 @@ static void IO_vdp_data_write(const uint8_t value)
 
             sms.vdp.addr = (sms.vdp.addr + 1) & 0x3FFF;
             break;
+        default: __builtin_unreachable();
     }
 }
 
@@ -665,6 +668,7 @@ static void IO_vdp_control_write(const uint8_t value)
             case VDP_CODE_CRAM_WRITE:
                 sms.vdp.addr = sms.vdp.control_word & 0x3FFF;
                 break;
+            default: __builtin_unreachable(); 
         }
     }
     else
@@ -685,6 +689,7 @@ static uint8_t IO_gamegear_read(const uint8_t addr)
         case 0x3: return /* 0x00; */ sms.port.gg_regs[0x3];
         case 0x4: return /* 0xFF; */ sms.port.gg_regs[0x4];
         case 0x5: return /* 0x00; */ sms.port.gg_regs[0x5];
+        default: __builtin_unreachable();
     }
 
     UNREACHABLE(0xFF);
@@ -810,6 +815,7 @@ uint8_t SMS_read_io(const uint8_t addr)
         case 0xF1: case 0xF3: case 0xF5: case 0xF7:
         case 0xF9: case 0xFB: case 0xFD: case 0xFF:
             return sms.port.b;
+        default: __builtin_unreachable(); 
     }
 
     UNREACHABLE(0xFF);
@@ -899,6 +905,7 @@ void SMS_write_io(const uint8_t addr, const uint8_t value)
         case 0xB1: case 0xB3: case 0xB5: case 0xB7:
         case 0xB9: case 0xBB: case 0xBD: case 0xBF:
             IO_vdp_control_write(value);
-            break;
+             break;
+        default: __builtin_unreachable();
     }
 }

main/totalsms/internal.h

@@ -95,6 +95,7 @@ extern "C" {
 #define ARRAY_SIZE(array) (sizeof(array) / sizeof(array[0]))
 
 #if SMS_DEBUG
+    #include <assert.h>
     #define SMS_log(...) ESP_LOGI(TAG, __VA_ARGS__)
     #define SMS_log_err(...) ESP_LOGE(TAG, __VA_ARGS__)
     #define SMS_log_fatal(...) do { ESP_LOGE(TAG, __VA_ARGS__); assert(0); } while(0)

main/totalsms/psg.c

@@ -92,7 +92,7 @@ static FORCE_INLINE void data_reg_write(const uint8_t value)
     }
 }
 
-void psg_reg_write(const uint8_t value)
+void FORCE_INLINE psg_reg_write(const uint8_t value)
 {
     psg_sync();
 

main/totalsms/sms.c

@@ -3,7 +3,6 @@
 #include "types.h"
 #include "rom_database.h"
 
-#include <assert.h>
 #include <stddef.h>
 #include <stdint.h>
 #include <string.h>

main/totalsms/vdp.c

@@ -1,7 +1,6 @@
 #include "sms.h"
 #include "internal.h"
 #include "types.h"
-#include <assert.h>
 #include <stdint.h>
 #include <string.h>
 #include <stdio.h>
@@ -1030,6 +1029,8 @@ void render_half_frame() {
   }
 }
 
+uint64_t vdp_display_enabled = 0;
+uint64_t vdp_display_disabled = 0;
 void vdp_render_frame()
 {
     // only render if display is enabled
@@ -1040,8 +1041,10 @@ void vdp_render_frame()
         {
             vdp_parse_sprites();
         }
+        ++vdp_display_disabled;
         return;
     }
+    ++vdp_display_enabled;
 
     // exit early if we have no pixels (this will break games that need sprite overflow and collision)
     if (!sms.pixels || sms.skip_frame)

main/totalsms/z80.c

@@ -5,7 +5,6 @@
 
 #include <stdint.h>
 #include <string.h>
-#include <assert.h>
 
 extern struct SMS_Core sms;
 
@@ -232,6 +231,7 @@ static FORCE_INLINE uint8_t get_r8(const uint8_t idx)
         case 0x5: return REG_L;
         case 0x6: return read8(REG_HL);
         case 0x7: return REG_A;
+        default: __builtin_unreachable();
     }
 
     UNREACHABLE(0xFF);
@@ -249,6 +249,7 @@ static FORCE_INLINE void set_r8(const uint8_t value, const uint8_t idx)
         case 0x5: REG_L = value; break;
         case 0x6: write8(REG_HL, value); break;
         case 0x7: REG_A = value; break;
+        default: __builtin_unreachable();
     }
 }
 
@@ -260,6 +261,7 @@ static FORCE_INLINE uint16_t get_r16(const uint8_t idx)
         case 0x1: return REG_DE;
         case 0x2: return REG_HL;
         case 0x3: return REG_SP;
+        default: __builtin_unreachable();
     }
 
     UNREACHABLE(0xFF);
@@ -273,6 +275,7 @@ static FORCE_INLINE void set_r16(uint16_t value, const uint8_t idx)
         case 0x1: SET_REG_DE(value); break;
         case 0x2: SET_REG_HL(value); break;
         case 0x3: REG_SP = value; break;
+        default: __builtin_unreachable();
     }
 }
 
@@ -1471,6 +1474,7 @@ static FORCE_INLINE bool _CB(const uint8_t opcode, const uint8_t value, uint8_t*
         case 0x1C: case 0x1D: case 0x1E: case 0x1F:
             *result = SET(value, 1 << ((opcode >> 3) & 0x7));
             return true;
+        default: __builtin_unreachable();
     }
 
     UNREACHABLE(false);
@@ -1512,6 +1516,7 @@ static FORCE_INLINE void execute_CB_IXIY(uint16_t ixy)
             case 0x5: REG_L = result; break;
             case 0x6: write8(addr, result); break;
             case 0x7: REG_A = result; break;
+            default: __builtin_unreachable();
         }
 
         sms.cpu.cycles += 23;
@@ -1694,14 +1699,15 @@ static FORCE_INLINE void execute_IXIY(uint8_t* ixy_hi, uint8_t* ixy_lo)
         case 0xCB: execute_CB_IXIY(pair); break;
 
         default:
-            SMS_log_fatal("UNK OP: 0xFD%02X", opcode);
+            __builtin_unreachable();
+            //SMS_log_fatal("UNK OP: 0xFD%02X", opcode);
             break;
     }
 
     #undef DISP
 }
 
-static FORCE_INLINE void execute_ED()
+__attribute__((noinline)) static void execute_ED()
 {
     const uint8_t opcode = read8(REG_PC++);
     sms.cpu.cycles += CYC_ED[opcode];
@@ -1831,7 +1837,8 @@ static FORCE_INLINE void execute_ED()
         case 0xBB: OTDR(); break;
 
         default:
-            SMS_log_fatal("UNK OP: 0xED%02X", opcode);
+            __builtin_unreachable();
+            //SMS_log_fatal("UNK OP: 0xED%02X", opcode);
             break;
     }
 }
@@ -2045,7 +2052,8 @@ static FORCE_INLINE void execute()
         case 0xFD: execute_IXIY(&REG_IYH, &REG_IYL); break;
 
         default:
-            SMS_log_fatal("UNK OP: 0x%02X", opcode);
+            __builtin_unreachable();
+            //SMS_log_fatal("UNK OP: 0x%02X", opcode);
             break;
     }
 }