Merge branch 'master' into new_apu

Core/display.c

@@ -250,9 +250,6 @@ void GB_palette_changed(GB_gameboy_t *gb, bool background_palette, uint8_t index
 
 static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
 {
-    uint8_t previous_stat_interrupt_line = gb->stat_interrupt_line;
-    gb->stat_interrupt_line = false;
-    
     if (!(gb->io_registers[GB_IO_LCDC] & 0x80)) {
         /* LCD is disabled, state is constant */
         
@@ -261,6 +258,7 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
         gb->io_registers[GB_IO_LY] = 0;
         gb->io_registers[GB_IO_STAT] &= ~3;
         gb->io_registers[GB_IO_STAT] |= 4;
+        gb->effective_scx = gb->io_registers[GB_IO_SCX];
         if (gb->hdma_on_hblank) {
             gb->hdma_on_hblank = false;
             gb->hdma_on = false;
@@ -293,23 +291,37 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
        these findings.
        Todo: Investigate what causes the difference between our findings */
     uint8_t stat_delay = gb->cgb_double_speed? 2 : 4; // (gb->cgb_mode? 0 : 4);
-    /* Todo: This is correct for DMG. Is it correct for the 3 CGB modes (DMG/single/double)?*/
+    /* Todo: Is this correct for DMG mode CGB? */
-    uint8_t scx_delay = ((gb->effective_scx & 7) + atomic_increase - 1) & ~(atomic_increase - 1);
+    uint8_t scx_delay = gb->effective_scx;
+    if (gb->cgb_double_speed) {
+        scx_delay = (scx_delay + 1) & ~1 & 7;
+    }
+    else {
+        scx_delay = (scx_delay + (gb->first_scanline ? 2 : 0)) & ~3 & 7;
+    }
+    
     /* Todo: These are correct for DMG, DMG-mode CGB, and single speed CGB. Is is correct for double speed CGB? */
     uint8_t oam_blocking_rush = gb->cgb_double_speed? 2 : 4;
     uint8_t vram_blocking_rush = gb->is_cgb? 0 : 4;
     
     for (; cycles; cycles -= atomic_increase) {
-        gb->delayed_interrupts &= ~3;
+        bool dmg_future_stat = false;
+        gb->io_registers[GB_IO_IF] |= gb->future_interrupts & 3;
+        gb->future_interrupts &= ~3;
+        
+        bool previous_stat_interrupt_line = gb->stat_interrupt_line;
+        gb->stat_interrupt_line = false;
+        
         gb->display_cycles += atomic_increase;
         /* The very first line is 4 clocks shorter when the LCD turns on. Verified on SGB2, CGB in CGB mode and
          CGB in double speed mode. */
-        if (gb->first_scanline && gb->display_cycles >= LINE_LENGTH - 4) {
+        if (gb->first_scanline && gb->display_cycles >= LINE_LENGTH - 8) {
             gb->first_scanline = false;
             gb->display_cycles += 4;
         }
         bool should_compare_ly = true;
         uint8_t ly_for_comparison = gb->io_registers[GB_IO_LY] = gb->display_cycles / LINE_LENGTH;
+        bool just_entered_hblank = false;
         
         
         /* Handle cycle completion. STAT's initial value depends on model and mode */
@@ -343,10 +355,11 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
         else if (gb->display_cycles == LINES * LINE_LENGTH + stat_delay) {
             gb->io_registers[GB_IO_STAT] &= ~3;
             gb->io_registers[GB_IO_STAT] |= 1;
-            gb->io_registers[GB_IO_IF] |= 1;
             if (gb->is_cgb) {
-                /* See comment on STAT interrupt at the end of the loop */
+                gb->future_interrupts |= 1;
-                gb->delayed_interrupts |= 1;
+            }
+            else {
+                gb->io_registers[GB_IO_IF] |= 1;
             }
             
             /* Entering VBlank state triggers the OAM interrupt. In CGB, it happens 4 cycles earlier */
@@ -382,17 +395,19 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
             else if (gb->display_cycles == MODE2_LENGTH) {
                 gb->io_registers[GB_IO_STAT] &= ~3;
                 gb->io_registers[GB_IO_STAT] |= 3;
+                gb->effective_scx = gb->io_registers[GB_IO_SCX];
                 gb->oam_read_blocked = true;
                 gb->vram_read_blocked = true;
                 gb->oam_write_blocked = true;
                 gb->vram_write_blocked = true;
             }
-            else if (gb->display_cycles == MODE2_LENGTH + MODE3_LENGTH) {
+            else if (gb->display_cycles == MODE2_LENGTH + MODE3_LENGTH + scx_delay) {
                 gb->io_registers[GB_IO_STAT] &= ~3;
                 gb->oam_read_blocked = false;
                 gb->vram_read_blocked = false;
                 gb->oam_write_blocked = false;
                 gb->vram_write_blocked = false;
+                just_entered_hblank = true;
             }
         }
         
@@ -440,6 +455,7 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
                 gb->previous_lcdc_x = - (gb->effective_scx & 0x7);
             }
             else if (position_in_line == MODE2_LENGTH + MODE3_LENGTH + stat_delay + scx_delay) {
+                just_entered_hblank = true;
                 gb->io_registers[GB_IO_STAT] &= ~3;
                 gb->oam_read_blocked = false;
                 gb->vram_read_blocked = false;
@@ -541,15 +557,23 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
         
         if (!gb->stat_interrupt_line) {
             switch (gb->io_registers[GB_IO_STAT] & 3) {
-                case 0: gb->stat_interrupt_line = gb->io_registers[GB_IO_STAT] & 8; break;
+                case 0:
+                    gb->stat_interrupt_line = (gb->io_registers[GB_IO_STAT] & 8);
+                    if (!gb->cgb_double_speed && just_entered_hblank && ((gb->effective_scx + (gb->first_scanline ? 2 : 0)) & 3) == 3) {
+                        gb->stat_interrupt_line = false;
+                    }
+                    else if (just_entered_hblank && ((gb->effective_scx + (gb->first_scanline ? 2 : 0)) & 3) != 0) {
+                        dmg_future_stat = true;
+                    }
+                    break;
                 case 1: gb->stat_interrupt_line = gb->io_registers[GB_IO_STAT] & 0x10; break;
                 case 2: gb->stat_interrupt_line = gb->io_registers[GB_IO_STAT] & 0x20; break;
             }
             
-            /* Use requested a LY=LYC interrupt and the LY=LYC bit is on */
+            /* User requested a LY=LYC interrupt and the LY=LYC bit is on */
             if ((gb->io_registers[GB_IO_STAT] & 0x44) == 0x44) {
                 gb->stat_interrupt_line = true;
-            }
+                dmg_future_stat = false;
             }
         }
         
@@ -562,12 +586,12 @@ static void update_display_state(GB_gameboy_t *gb, uint8_t cycles)
         }
         
         if (gb->stat_interrupt_line && !previous_stat_interrupt_line) {
+            if (gb->is_cgb || dmg_future_stat) {
+                gb->future_interrupts |= 2;
+            }
+            else {
                 gb->io_registers[GB_IO_IF] |= 2;
-        if (gb->is_cgb) {
+            }
-            /* On CGB, the STAT interrupt is not aligned to a T-Cycle, therefore it is only effective the next T-Cycle
-               Todo: verify on DMG mode CGB. This was only tested on LYC STAT interrupts, should be tested on others
-               as well. */
-            gb->delayed_interrupts |= 2;
         }
     }
     

Core/gb.h

@@ -331,10 +331,8 @@ struct GB_gameboy_internal_s {
         GB_PADDING(uint16_t, serial_cycles);
         uint16_t serial_cycles; /* This field changed its meaning in v0.10 */
         uint16_t serial_length;
-        uint8_t delayed_interrupts; /* When an interrupt occurs while not aligned to a T-cycle, it must be "delayed" */
+        uint8_t future_interrupts; /* Interrupts can occur in any T-cycle. Some timings result in different interrupt
-        bool dont_delay_timer_interrupt; /* If the timer glitch causes a TIMA overflow, it causes the timer to overflow
+                                     timing when the CPU is in halt mode, and might also affect the DI instruction. */
-                                            with different timing, so the triggered interrupt is not delayed.
-                                            Todo: needs test ROM. */
     );
 
     /* APU */

Core/memory.c

@@ -136,7 +136,7 @@ static uint8_t read_high_memory(GB_gameboy_t *gb, uint16_t addr)
     if (addr < 0xFF80) {
         switch (addr & 0xFF) {
             case GB_IO_IF:
-                return gb->io_registers[GB_IO_IF] | 0xE0;
+                return gb->io_registers[GB_IO_IF] | 0xE0 | gb->future_interrupts;
             case GB_IO_TAC:
                 return gb->io_registers[GB_IO_TAC] | 0xF8;
             case GB_IO_STAT:
@@ -414,8 +414,9 @@ static void write_high_memory(GB_gameboy_t *gb, uint16_t addr, uint8_t value)
             case GB_IO_WX:
                 GB_window_related_write(gb, addr & 0xFF, value);
                 break;
-            case GB_IO_SCX:
             case GB_IO_IF:
+                gb->future_interrupts = 0;
+            case GB_IO_SCX:
             case GB_IO_SCY:
             case GB_IO_LYC:
             case GB_IO_BGP:

Core/timing.c

@@ -82,15 +82,13 @@ static void GB_ir_run(GB_gameboy_t *gb)
 
 static void advance_tima_state_machine(GB_gameboy_t *gb)
 {
-    gb->delayed_interrupts &= ~4;
+    gb->io_registers[GB_IO_IF] |= gb->future_interrupts & 4;
+    gb->future_interrupts &= ~4;
     if (gb->tima_reload_state == GB_TIMA_RELOADED) {
         gb->tima_reload_state = GB_TIMA_RUNNING;
     }
     else if (gb->tima_reload_state == GB_TIMA_RELOADING) {
-        gb->io_registers[GB_IO_IF] |= 4;
+        gb->future_interrupts |= 4;
-        if (!gb->dont_delay_timer_interrupt) {
-            gb->delayed_interrupts |= 4; // Timer interrupt is not aligned to a T-cycle and therefore is effective only the next one.
-        }
         gb->tima_reload_state = GB_TIMA_RELOADED;
     }
 }
@@ -157,7 +155,6 @@ static void increase_tima(GB_gameboy_t *gb)
 {
     gb->io_registers[GB_IO_TIMA]++;
     if (gb->io_registers[GB_IO_TIMA] == 0) {
-        gb->dont_delay_timer_interrupt = false;
         gb->io_registers[GB_IO_TIMA] = gb->io_registers[GB_IO_TMA];
         gb->tima_reload_state = GB_TIMA_RELOADING;
     }
@@ -201,7 +198,6 @@ void GB_emulate_timer_glitch(GB_gameboy_t *gb, uint8_t old_tac, uint8_t new_tac)
         /* And now either the timer must be disabled, or the new bit used for overflow testing be 0. */
         if (!(new_tac & 4) || gb->div_cycles & (new_clocks >> 1)) {
             increase_tima(gb);
-            gb->dont_delay_timer_interrupt = true;
         }
     }
 }

Core/z80_cpu.c

@@ -1337,7 +1337,13 @@ static GB_opcode_t *opcodes[256] = {
 void GB_cpu_run(GB_gameboy_t *gb)
 {
     gb->vblank_just_occured = false;
-    uint8_t interrupt_queue = gb->interrupt_enable & gb->io_registers[GB_IO_IF] & 0x1F  & ~gb->delayed_interrupts;
+    uint8_t interrupt_queue = gb->interrupt_enable & gb->io_registers[GB_IO_IF] & 0x1F;
+    if (!gb->halted) {
+        interrupt_queue |= gb->future_interrupts & gb->interrupt_enable;
+    }
+
+    gb->io_registers[GB_IO_IF] |= gb->future_interrupts;
+    gb->future_interrupts = 0;
     
     if (interrupt_queue) {
         gb->halted = false;