idle loops adjusted

git-svn-id: file:///home/notaz/opt/svn/PicoDrive@549 be3aeb3a-fb24-0410-a615-afba39da0efa

Pico/Pico.c

@@ -171,16 +171,16 @@ int PicoReset(void)
   if (Pico.m.dma_xfers == 0 && !(PicoOpt&POPT_DIS_VDP_FIFO))
     Pico.m.dma_xfers = rand() & 0x1fff;
 
+  SekFinishIdleDet();
+
   if (PicoAHW & PAHW_MCD) {
     PicoResetMCD();
     return 0;
   }
-  else {
+
   // reinit, so that checksum checks pass
-    SekFinishIdleDet();
   if (!(PicoOpt & POPT_DIS_IDLE_DET))
     SekInitIdleDet();
-  }
 
   // reset sram state; enable sram access by default if it doesn't overlap with ROM
   Pico.m.sram_reg=sram_reg&0x14;
@@ -315,6 +315,21 @@ void PicoFrame(void)
 {
   Pico.m.frame_count++;
 
+#if 0
+  if ((Pico.m.frame_count & 0x3f) == 0)
+  {
+    extern int idlehit_addrs[], idlehit_counts[];
+    int i;
+    printf("--\n");
+    for (i = 0; i < 128 && idlehit_addrs[i] != 0; i++) {
+      if (idlehit_counts[i] != 0) {
+        printf("%06x %i %i\n", idlehit_addrs[i], idlehit_counts[i],  idlehit_counts[i] >> 6);
+        idlehit_counts[i] = 0;
+      }
+    }
+  }
+#endif
+
   if (PicoAHW & PAHW_MCD) {
     PicoFrameMCD();
     return;

Pico/Sek.c

@@ -199,6 +199,25 @@ static int *idledet_addrs = NULL;
 static int idledet_count = 0, idledet_bads = 0;
 int idledet_start_frame = 0;
 
+#if 0
+#define IDLE_STATS 1
+unsigned int idlehit_addrs[128], idlehit_counts[128];
+
+void SekRegisterIdleHit(unsigned int pc)
+{
+  int i;
+  for (i = 0; i < 127 && idlehit_addrs[i]; i++) {
+    if (idlehit_addrs[i] == pc) {
+      idlehit_counts[i]++;
+      return;
+    }
+  }
+  idlehit_addrs[i] = pc;
+  idlehit_counts[i] = 1;
+  idlehit_addrs[i+1] = 0;
+}
+#endif
+
 void SekInitIdleDet(void)
 {
   void *tmp = realloc(idledet_addrs, 0x200*4);
@@ -210,6 +229,9 @@ void SekInitIdleDet(void)
     idledet_addrs = tmp;
   idledet_count = idledet_bads = 0;
   idledet_start_frame = Pico.m.frame_count + 360;
+#ifdef IDLE_STATS
+  idlehit_addrs[0] = 0;
+#endif
 
 #ifdef EMU_C68K
   CycloneInitIdle();
@@ -224,6 +246,10 @@ int SekIsIdleCode(unsigned short *dst, int bytes)
   // printf("SekIsIdleCode %04x %i\n", *dst, bytes);
   switch (bytes)
   {
+    case 2:
+      if ((*dst & 0xf000) != 0x6000)     // not another branch
+        return 1;
+      break;
     case 4:
       if (  (*dst & 0xfff8) == 0x4a10 || // tst.b ($aX)      // there should be no need to wait
             (*dst & 0xfff8) == 0x4a28 || // tst.b ($xxxx,a0) // for byte change anywhere
@@ -261,13 +287,20 @@ int SekIsIdleCode(unsigned short *dst, int bytes)
   return 0;
 }
 
-int SekRegisterIdlePatch(unsigned int pc, int oldop, int newop)
+int SekRegisterIdlePatch(unsigned int pc, int oldop, int newop, void *ctx)
 {
-#ifdef EMU_C68K
+  int is_main68k = 1;
-  pc -= PicoCpuCM68k.membase;
+#if   defined(EMU_C68K)
+  struct Cyclone *cyc = ctx;
+  is_main68k = cyc == &PicoCpuCM68k;
+  pc -= cyc->membase;
+#elif defined(EMU_F68K)
+  is_main68k = ctx == &PicoCpuFM68k;
 #endif
   pc &= ~0xff000000;
-  elprintf(EL_IDLE, "idle: patch %06x %04x %04x #%i", pc, oldop, newop, idledet_count);
+  elprintf(EL_IDLE, "idle: patch %06x %04x %04x %c %c #%i", pc, oldop, newop,
+    (newop&0x200)?'n':'y', is_main68k?'m':'s', idledet_count);
+
   if (pc > Pico.romsize && !(PicoAHW & PAHW_SVP)) {
     if (++idledet_bads > 128) return 2; // remove detector
     return 1; // don't patch

cpu/Cyclone/tools/idle.s

@@ -22,8 +22,8 @@ patch_desc_table:
   .word (0x75f8<<16) | 0x67f8, idle_detector_bcc8, idle_beq, Op6701  @ beq.s
   .word (0x75f6<<16) | 0x67f6, idle_detector_bcc8, idle_beq, Op6701  @ beq.s
   .word (0x75f2<<16) | 0x67f2, idle_detector_bcc8, idle_beq, Op6701  @ beq.s
-  .word (0x7dfe<<16) | 0x60fe, idle_detector_dead, idle_bra, Op6001  @ bra.s
+  .word (0x7dfe<<16) | 0x60fe, idle_detector_bcc8, idle_bra, Op6001  @ bra.s
-  .word (0x7dfc<<16) | 0x60fc, idle_detector_dead, idle_bra, Op6001  @ bra.s
+  .word (0x7dfc<<16) | 0x60fc, idle_detector_bcc8, idle_bra, Op6001  @ bra.s
 
 
 .text
@@ -90,10 +90,12 @@ cfi_loop:
 
 
 .macro inc_counter cond
-@    ldr     r0, =idle_hit_counter
+@    ldr\cond r0, [r7, #0x60]
-@    ldr     r1, [r0]
+@    mov     r11,lr
-@    add     r1, r1, #1
+@    sub     r0, r4, r0
-@    str\cond r1, [r0]
+@    sub     r0, r0, #2
+@    bl\cond SekRegisterIdleHit
+@    mov     lr, r11
 .endm
 
 idle_bra:
@@ -139,43 +141,36 @@ idle_detector_bcc8:
     and     r2, r8, #0x00ff
     orr     r2, r2, #0x7100
     orreq   r2, r2, #0x0200
-    tst     r8, #0x0100           @ 67xx (beq)?
+    mov     r0, r8, lsr #8
-    orrne   r2, r2, #0x0400
+    cmp     r0, #0x66
+    orrgt   r2, r2, #0x0400       @ 67xx (beq)
+    orrlt   r2, r2, #0x0c00       @ 60xx (bra)
 
     @ r2 = patch_opcode
     sub     r0, r4, #2
     ldrh    r1, [r0]
     mov     r11,r2
+    mov     r3, r7
     bl      SekRegisterIdlePatch
     cmp     r0, #1                @ 0 - ok to patch, 1 - no patch, 2 - remove detector
     strlth  r11,[r4, #-2]
     ble     exit_detector
 
     @ remove detector from Cyclone
-    tst     r8, #0x0100           @ 67xx (beq)?
+    mov     r0, r8, lsr #8
+    cmp     r0, #0x66
+    ldrlt   r1, =Op6001
     ldreq   r1, =Op6601
-    ldrne   r1, =Op6701
+    ldrgt   r1, =Op6701
 
     ldr     r3, =CycloneJumpTab
     str     r1, [r3, r8, lsl #2]
     bx      r1
 
 exit_detector:
-    tst     r8, #0x0100           @ 67xx (beq)?
+    mov     r0, r8, lsr #8
+    cmp     r0, #0x66
+    blt     Op6001
     beq     Op6601
     b       Op6701
 
-
-idle_detector_dead:
-    @ patch without further questions
-    and     r2, r8, #0x00ff
-    orr     r2, r2, #0x7d00
-    sub     r0, r4, #2
-    ldrh    r1, [r0]
-    mov     r11,r2
-    bl      SekRegisterIdlePatch
-    strh    r11,[r4, #-2]
-    b       Op6001
-
-.pool
-

cpu/fame/famec.c

@@ -5041,8 +5041,8 @@ idle_install:
 	INSTALL_IDLE(0x75f8, 0x67f8, idle_detector_bcc8, 0x6701_idle, 0x6701);
 	INSTALL_IDLE(0x75f6, 0x67f6, idle_detector_bcc8, 0x6701_idle, 0x6701);
 	INSTALL_IDLE(0x75f2, 0x67f2, idle_detector_bcc8, 0x6701_idle, 0x6701);
-	INSTALL_IDLE(0x7dfe, 0x60fe, idle_detector_dead, 0x6001_idle, 0x6001);
+	INSTALL_IDLE(0x7dfe, 0x60fe, idle_detector_bcc8, 0x6001_idle, 0x6001);
-	INSTALL_IDLE(0x7dfc, 0x60fc, idle_detector_dead, 0x6001_idle, 0x6001);
+	INSTALL_IDLE(0x7dfc, 0x60fc, idle_detector_bcc8, 0x6001_idle, 0x6001);
 	return 0;
 
 idle_remove:

cpu/fame/famec_opcodes.h

@@ -40013,7 +40013,7 @@ RET(8)
 
 
 extern int SekIsIdleCode(unsigned short *dst, int bytes);
-extern int SekRegisterIdlePatch(unsigned int pc, int oldop, int newop);
+extern int SekRegisterIdlePatch(unsigned int pc, int oldop, int newop, void *ctx);
 
 OPCODE(idle_detector_bcc8)
 {
@@ -40032,18 +40032,22 @@ OPCODE(idle_detector_bcc8)
 	ret = SekIsIdleCode(dest_pc, bytes);
 	newop = (Opcode & 0xfe) | 0x7100;
 	if (!ret) newop |= 0x200;
-	if (Opcode & 0x0100) newop |= 0x400; // beq
+	if (  Opcode & 0x0100)  newop |= 0x400; // beq
+	if (!(Opcode & 0x0f00)) newop |= 0xc00; // bra
 
-	ret = SekRegisterIdlePatch(GET_PC - 2, Opcode, newop);
+	ret = SekRegisterIdlePatch(GET_PC - 2, Opcode, newop, &m68kcontext);
 	switch (ret)
 	{
 		case 0: PC[-1] = newop; break;
 		case 1: break;
-		case 2: JumpTable[Opcode] = (Opcode & 0x0100) ? CAST_OP(0x6701) : CAST_OP(0x6601); break;
+		case 2: JumpTable[Opcode] = (Opcode & 0x0f00) ?
+				((Opcode & 0x0100) ? CAST_OP(0x6701) : CAST_OP(0x6601)) :
+				CAST_OP(0x6001); break;
 	}
 
 end:
-	cond_true = (Opcode & 0x0100) ? !flag_NotZ : flag_NotZ; // beq?
+	if ((Opcode & 0xff00) == 0x6000) cond_true = 1;
+	else cond_true = (Opcode & 0x0100) ? !flag_NotZ : flag_NotZ; // beq?
 	if (cond_true)
 	{
 		PC = dest_pc;
@@ -40052,14 +40056,4 @@ end:
 RET(8)
 }
 
-OPCODE(idle_detector_dead)
-{
-	// patch without further questions
-	int newop = 0x7d00 | (Opcode & 0xff);
-	PC[-1] = newop;
-	SekRegisterIdlePatch(GET_PC - 2, Opcode, newop);
-
-	PC += ((s8)(Opcode & 0xFE)) >> 1;
-RET(10)
-}
 #endif // PICODRIVE_HACK