32x: built-in BIOS; reset handling; 68k memhandler split

git-svn-id: file:///home/notaz/opt/svn/PicoDrive@801 be3aeb3a-fb24-0410-a615-afba39da0efa

cpu/sh2mame/sh2pico.c

@@ -41,20 +41,11 @@ void p32x_sh2_write32(unsigned int a, unsigned int d, int id);
 
 void sh2_reset(SH2 *sh2)
 {
-	int save_is_slave;
-	void *save_irqcallback;
-
-	save_irqcallback = sh2->irq_callback;
-	save_is_slave = sh2->is_slave;
-
-	memset(sh2, 0, sizeof(SH2));
-
-	sh2->is_slave = save_is_slave;
-	sh2->irq_callback = save_irqcallback;
-
 	sh2->pc = RL(0);
 	sh2->r[15] = RL(4);
 	sh2->sr = I;
+	sh2->vbr = 0;
+	sh2->pending_int_irq = 0;
 }
 
 static void sh2_do_irq(SH2 *sh2, int level, int vector)

pico/32x/32x.c

@@ -1,8 +1,8 @@
 #include "../pico_int.h"
 #include "../sound/ym2612.h"
 
-SH2 sh2s[2];
 struct Pico32x Pico32x;
+SH2 sh2s[2];
 
 static void sh2_irq_cb(int id, int level)
 {
@@ -39,15 +39,12 @@ void Pico32xStartup(void)
   elprintf(EL_STATUS|EL_32X, "32X startup");
 
   PicoAHW |= PAHW_32X;
-  PicoMemSetup32x();
-
   sh2_init(&msh2, 0);
   msh2.irq_callback = sh2_irq_cb;
-  sh2_reset(&msh2);
-
   sh2_init(&ssh2, 1);
   ssh2.irq_callback = sh2_irq_cb;
-  sh2_reset(&ssh2);
+
+  PicoMemSetup32x();
 
   if (!Pico.m.pal)
     Pico32x.vdp_regs[0] |= P32XV_nPAL;
@@ -58,6 +55,54 @@ void Pico32xStartup(void)
   emu_32x_startup();
 }
 
+#define HWSWAP(x) (((x) << 16) | ((x) >> 16))
+void p32x_reset_sh2s(void)
+{
+  elprintf(EL_32X, "sh2 reset");
+
+  sh2_reset(&msh2);
+  sh2_reset(&ssh2);
+
+  // if we don't have BIOS set, perform it's work here.
+  // MSH2
+  if (p32x_bios_m == NULL) {
+    unsigned int idl_src, idl_dst, idl_size; // initial data load
+    unsigned int vbr;
+
+    // initial data
+    idl_src = HWSWAP(*(unsigned int *)(Pico.rom + 0x3d4)) & ~0xf0000000;
+    idl_dst = HWSWAP(*(unsigned int *)(Pico.rom + 0x3d8)) & ~0xf0000000;
+    idl_size= HWSWAP(*(unsigned int *)(Pico.rom + 0x3dc));
+    if (idl_size > Pico.romsize || idl_src + idl_size > Pico.romsize ||
+        idl_size > 0x40000 || idl_dst + idl_size > 0x40000 || (idl_src & 3) || (idl_dst & 3)) {
+      elprintf(EL_STATUS|EL_ANOMALY, "32x: invalid initial data ptrs: %06x -> %06x, %06x",
+        idl_src, idl_dst, idl_size);
+    }
+    else
+      memcpy(Pico32xMem->sdram + idl_dst, Pico.rom + idl_src, idl_size);
+
+    // GBR/VBR
+    vbr = HWSWAP(*(unsigned int *)(Pico.rom + 0x3e8));
+    sh2_set_gbr(0, 0x20004000);
+    sh2_set_vbr(0, vbr);
+
+    // checksum and M_OK
+    Pico32x.regs[0x28 / 2] = *(unsigned short *)(Pico.rom + 0x18e);
+    // program will set M_OK
+  }
+
+  // SSH2
+  if (p32x_bios_s == NULL) {
+    unsigned int vbr;
+
+    // GBR/VBR
+    vbr = HWSWAP(*(unsigned int *)(Pico.rom + 0x3ec));
+    sh2_set_gbr(1, 0x20004000);
+    sh2_set_vbr(1, vbr);
+    // program will set S_OK
+  }
+}
+
 void Pico32xInit(void)
 {
 }
@@ -66,7 +111,7 @@ void PicoPower32x(void)
 {
   memset(&Pico32x, 0, sizeof(Pico32x));
 
-  Pico32x.regs[0] = 0x0082; // SH2 reset?
+  Pico32x.regs[0] = P32XS_REN|P32XS_nRES; // verified
   Pico32x.vdp_regs[0x0a/2] = P32XV_VBLK|P32XV_HBLK|P32XV_PEN;
   Pico32x.sh2_regs[0] = P32XS2_ADEN;
 }
@@ -82,8 +127,11 @@ void PicoUnload32x(void)
 
 void PicoReset32x(void)
 {
-  extern int p32x_csum_faked;
+  if (PicoAHW & PAHW_32X) {
-  p32x_csum_faked = 0; // tmp
+    Pico32x.sh2irqs |= P32XI_VRES;
+    p32x_update_irls();
+    p32x_poll_event(3, 0);
+  }
 }
 
 static void p32x_start_blank(void)
@@ -129,6 +177,8 @@ static __inline void run_m68k(int cyc)
   while (SekCycleCnt < SekCycleAim) { \
     slice = SekCycleCnt; \
     run_m68k(SekCycleAim - SekCycleCnt); \
+    if (!(Pico32x.regs[0] & P32XS_nRES)) \
+      continue; /* SH2s reseting */ \
     slice = SekCycleCnt - slice; /* real count from 68k */ \
     if (SekCycleCnt < SekCycleAim) \
       elprintf(EL_32X, "slice %d", slice); \

pico/32x/memory.c

@@ -1,3 +1,22 @@
+/*
+ * Register map:
+ * a15100 F....... R.....EA  F.....AC N...VHMP 4000 // Fm Ren nrEs Aden Cart heN V H cMd Pwm
+ * a15102 ........ ......SM  ?                 4002 // intS intM
+ * a15104 ........ ......10  ........ hhhhhhhh 4004 // bk1 bk0 Hint
+ * a15106 F....... .....SDR  UE...... .....SDR 4006 // Full 68S Dma Rv fUll[fb] Empt[fb]
+ * a15108           (32bit DREQ src)           4008
+ * a1510c           (32bit DREQ dst)           400c
+ * a15110          llllllll llllll00           4010 // DREQ Len
+ * a15112           (16bit FIFO reg)           4012
+ * a15114 ?                  (16bit VRES clr)  4014
+ * a15116 ?                  (16bit Vint clr)  4016
+ * a15118 ?                  (16bit Hint clr)  4018
+ * a1511a ........ .......C  (16bit CMD clr)   401a // Cm
+ * a1511c ?                  (16bit PWM clr)   401c
+ * a1511e ?                  ?                 401e
+ * a15120            (16 bytes comm)           2020
+ * a15130                 (PWM)                2030
+ */
 #include "../pico_int.h"
 #include "../memory.h"
 
@@ -10,6 +29,7 @@
 
 static const char str_mars[] = "MARS";
 
+void *p32x_bios_g, *p32x_bios_m, *p32x_bios_s;
 struct Pico32xMem *Pico32xMem;
 
 static void bank_switch(int b);
@@ -187,18 +207,14 @@ static void p32x_reg_write8(u32 a, u32 d)
   // for things like bset on comm port
   m68k_poll.cnt = 0;
 
-  if (a == 1 && !(r[0] & 1)) {
-    r[0] |= 1;
-    Pico32xStartup();
-    return;
-  }
-
-  if (!(r[0] & 1))
-    return;
-
   switch (a) {
     case 0: // adapter ctl
-      r[0] = (r[0] & 0x83) | ((d << 8) & P32XS_FM);
+      r[0] = (r[0] & ~P32XS_FM) | ((d << 8) & P32XS_FM);
+      return;
+    case 1: // adapter ctl, RES bit writeable
+      if ((d ^ r[0]) & d & P32XS_nRES)
+        p32x_reset_sh2s();
+      r[0] = (r[0] & ~P32XS_nRES) | (d & P32XS_nRES);
       return;
     case 3: // irq ctl
       if ((d & 1) && !(Pico32x.sh2irqi[0] & P32XI_CMD)) {
@@ -249,7 +265,9 @@ static void p32x_reg_write16(u32 a, u32 d)
 
   switch (a) {
     case 0x00: // adapter ctl
-      r[0] = (r[0] & 0x83) | (d & P32XS_FM);
+      if ((d ^ r[0]) & d & P32XS_nRES)
+        p32x_reset_sh2s();
+      r[0] = (r[0] & ~(P32XS_FM|P32XS_nRES)) | (d & (P32XS_FM|P32XS_nRES));
       return;
     case 0x10: // DREQ len
       r[a / 2] = d & ~3;
@@ -591,8 +609,10 @@ static void sh2_peripheral_write32(u32 a, u32 d, int id)
 }
 
 // ------------------------------------------------------------------
-// default 32x handlers
+// 32x handlers
-u32 PicoRead8_32x(u32 a)
+
+// after ADEN
+static u32 PicoRead8_32x_on(u32 a)
 {
   u32 d = 0;
   if ((a & 0xffc0) == 0x5100) { // a15100
@@ -600,8 +620,8 @@ u32 PicoRead8_32x(u32 a)
     goto out_16to8;
   }
 
-  if (!(Pico32x.regs[0] & 1))
+  if ((a & 0xfc00) != 0x5000)
-    goto no_vdp;
+    return PicoRead8_io(a);
 
   if ((a & 0xfff0) == 0x5180) { // a15180
     d = p32x_vdp_read16(a);
@@ -613,7 +633,6 @@ u32 PicoRead8_32x(u32 a)
     goto out_16to8;
   }
 
-no_vdp:
   if ((a & 0xfffc) == 0x30ec) { // a130ec
     d = str_mars[a & 3];
     goto out;
@@ -633,7 +652,7 @@ out:
   return d;
 }
 
-u32 PicoRead16_32x(u32 a)
+static u32 PicoRead16_32x_on(u32 a)
 {
   u32 d = 0;
   if ((a & 0xffc0) == 0x5100) { // a15100
@@ -641,8 +660,8 @@ u32 PicoRead16_32x(u32 a)
     goto out;
   }
 
-  if (!(Pico32x.regs[0] & 1))
+  if ((a & 0xfc00) != 0x5000)
-    goto no_vdp;
+    return PicoRead16_io(a);
 
   if ((a & 0xfff0) == 0x5180) { // a15180
     d = p32x_vdp_read16(a);
@@ -654,7 +673,110 @@ u32 PicoRead16_32x(u32 a)
     goto out;
   }
 
-no_vdp:
+  if ((a & 0xfffc) == 0x30ec) { // a130ec
+    d = !(a & 2) ? ('M'<<8)|'A' : ('R'<<8)|'S';
+    goto out;
+  }
+
+  elprintf(EL_UIO, "m68k unmapped r16 [%06x] @%06x", a, SekPc);
+  return d;
+
+out:
+  elprintf(EL_32X, "m68k 32x r16 [%06x] %04x @%06x", a, d, SekPc);
+  return d;
+}
+
+static void PicoWrite8_32x_on(u32 a, u32 d)
+{
+  if ((a & 0xfc00) == 0x5000)
+    elprintf(EL_32X, "m68k 32x w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
+
+  if ((a & 0xffc0) == 0x5100) { // a15100
+    p32x_reg_write8(a, d);
+    return;
+  }
+
+  if ((a & 0xfc00) != 0x5000) {
+    PicoWrite8_io(a, d);
+    return;
+  }
+
+  if ((a & 0xfff0) == 0x5180) { // a15180
+    p32x_vdp_write8(a, d);
+    return;
+  }
+
+  // TODO: verify
+  if ((a & 0xfe00) == 0x5200) { // a15200
+    elprintf(EL_32X|EL_ANOMALY, "m68k 32x PAL w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
+    ((u8 *)Pico32xMem->pal)[(a & 0x1ff) ^ 1] = d;
+    Pico32x.dirty_pal = 1;
+    return;
+  }
+
+  elprintf(EL_UIO, "m68k unmapped w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
+}
+
+static void PicoWrite16_32x_on(u32 a, u32 d)
+{
+  if ((a & 0xfc00) == 0x5000)
+    elprintf(EL_UIO, "m68k 32x w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
+
+  if ((a & 0xffc0) == 0x5100) { // a15100
+    p32x_reg_write16(a, d);
+    return;
+  }
+
+  if ((a & 0xfc00) != 0x5000) {
+    PicoWrite16_io(a, d);
+    return;
+  }
+
+  if ((a & 0xfff0) == 0x5180) { // a15180
+    p32x_vdp_write16(a, d);
+    return;
+  }
+
+  if ((a & 0xfe00) == 0x5200) { // a15200
+    Pico32xMem->pal[(a & 0x1ff) / 2] = d;
+    Pico32x.dirty_pal = 1;
+    return;
+  }
+
+  elprintf(EL_UIO, "m68k unmapped w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
+}
+
+// before ADEN
+u32 PicoRead8_32x(u32 a)
+{
+  u32 d = 0;
+  if ((a & 0xffc0) == 0x5100) { // a15100
+    // regs are always readable
+    d = ((u8 *)Pico32x.regs)[(a & 0x3f) ^ 1];
+    goto out;
+  }
+
+  if ((a & 0xfffc) == 0x30ec) { // a130ec
+    d = str_mars[a & 3];
+    goto out;
+  }
+
+  elprintf(EL_UIO, "m68k unmapped r8  [%06x] @%06x", a, SekPc);
+  return d;
+
+out:
+  elprintf(EL_32X, "m68k 32x r8  [%06x]   %02x @%06x", a, d, SekPc);
+  return d;
+}
+
+u32 PicoRead16_32x(u32 a)
+{
+  u32 d = 0;
+  if ((a & 0xffc0) == 0x5100) { // a15100
+    d = Pico32x.regs[(a & 0x3f) / 2];
+    goto out;
+  }
+
   if ((a & 0xfffc) == 0x30ec) { // a130ec
     d = !(a & 2) ? ('M'<<8)|'A' : ('R'<<8)|'S';
     goto out;
@@ -670,62 +792,60 @@ out:
 
 void PicoWrite8_32x(u32 a, u32 d)
 {
-  if ((a & 0xfc00) == 0x5000)
-    elprintf(EL_32X, "m68k 32x w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
-
   if ((a & 0xffc0) == 0x5100) { // a15100
-    p32x_reg_write8(a, d);
+    u16 *r = Pico32x.regs;
+
+    elprintf(EL_32X, "m68k 32x w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
+    a &= 0x3f;
+    if (a == 1) {
+      if ((d ^ r[0]) & d & P32XS_ADEN) {
+        Pico32xStartup();
+        r[0] &= ~P32XS_nRES; // causes reset if specified by this write
+        r[0] |= P32XS_ADEN;
+        p32x_reg_write8(a, d); // forward for reset processing
+      }
       return;
     }
 
-  if (!(Pico32x.regs[0] & 1))
+    // allow only COMM for now
-    goto no_vdp;
+    if ((a & 0x30) == 0x20) {
-
+      u8 *r8 = (u8 *)r;
-  if ((a & 0xfff0) == 0x5180) { // a15180
+      r8[a ^ 1] = d;
-    p32x_vdp_write8(a, d);
+    }
     return;
   }
 
-  // TODO: verify
-  if ((a & 0xfe00) == 0x5200) { // a15200
-    elprintf(EL_32X|EL_ANOMALY, "m68k 32x PAL w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
-    ((u8 *)Pico32xMem->pal)[(a & 0x1ff) ^ 1] = d;
-    Pico32x.dirty_pal = 1;
-    return;
-  }
-
-no_vdp:
   elprintf(EL_UIO, "m68k unmapped w8  [%06x]   %02x @%06x", a, d & 0xff, SekPc);
 }
 
 void PicoWrite16_32x(u32 a, u32 d)
 {
-  if ((a & 0xfc00) == 0x5000)
-    elprintf(EL_UIO, "m68k 32x w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
-
   if ((a & 0xffc0) == 0x5100) { // a15100
-    p32x_reg_write16(a, d);
+    u16 *r = Pico32x.regs;
+
+    elprintf(EL_UIO, "m68k 32x w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
+    a &= 0x3e;
+    if (a == 0) {
+      if ((d ^ r[0]) & d & P32XS_ADEN) {
+        Pico32xStartup();
+        r[0] &= ~P32XS_nRES; // causes reset if specified by this write
+        r[0] |= P32XS_ADEN;
+        p32x_reg_write16(a, d); // forward for reset processing
+      }
       return;
     }
 
-  if (!(Pico32x.regs[0] & 1))
+    // allow only COMM for now
-    goto no_vdp;
+    if ((a & 0x30) == 0x20)
-
+      r[a / 2] = d;
-  if ((a & 0xfff0) == 0x5180) { // a15180
-    p32x_vdp_write16(a, d);
     return;
   }
 
-  if ((a & 0xfe00) == 0x5200) { // a15200
-    Pico32xMem->pal[(a & 0x1ff) / 2] = d;
-    Pico32x.dirty_pal = 1;
-    return;
-  }
-
-no_vdp:
   elprintf(EL_UIO, "m68k unmapped w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
 }
 
+// -----------------------------------------------------------------
+
 // hint vector is writeable
 static void PicoWrite8_hint(u32 a, u32 d)
 {
@@ -1017,25 +1137,62 @@ void p32x_sh2_write32(u32 a, u32 d, int id)
   p32x_sh2_write16(a + 2, d, id);
 }
 
+static const u16 msh2_code[] = {
+  // trap instructions
+  0xaffe, // bra <self>
+  0x0009, // nop
+  // have to wait a bit until m68k initial program finishes clearing stuff
+  // to avoid races with game SH2 code, like in Tempo
+  0xd004, // mov.l   @(_m_ok,pc), r0
+  0xd105, // mov.l   @(_cnt,pc), r1
+  0xd205, // mov.l   @(_start,pc), r2
+  0x71ff, // add     #-1, r1
+  0x4115, // cmp/pl  r1
+  0x89fc, // bt      -2
+  0xc208, // mov.l   r0, @(h'20,gbr)
+  0x6822, // mov.l   @r2, r8
+  0x482b, // jmp     @r8
+  0x0009, // nop
+  ('M'<<8)|'_', ('O'<<8)|'K',
+  0x0001, 0x0000,
+  0x2200, 0x03e0  // master start pointer in ROM
+};
+
+static const u16 ssh2_code[] = {
+  0xaffe, // bra <self>
+  0x0009, // nop
+  // code to wait for master, in case authentic master BIOS is used
+  0xd104, // mov.l   @(_m_ok,pc), r1
+  0xd206, // mov.l   @(_start,pc), r2
+  0xc608, // mov.l   @(h'20,gbr), r0
+  0x3100, // cmp/eq  r0, r1
+  0x8bfc, // bf      #-2
+  0xd003, // mov.l   @(_s_ok,pc), r0
+  0xc209, // mov.l   r0, @(h'24,gbr)
+  0x6822, // mov.l   @r2, r8
+  0x482b, // jmp     @r8
+  0x0009, // nop
+  ('M'<<8)|'_', ('O'<<8)|'K',
+  ('S'<<8)|'_', ('O'<<8)|'K',
+  0x2200, 0x03e4  // slave start pointer in ROM
+};
+
 #define HWSWAP(x) (((x) << 16) | ((x) >> 16))
-void PicoMemSetup32x(void)
+static void get_bios(void)
 {
-  unsigned short *ps;
+  u16 *ps;
-  unsigned int *pl;
+  u32 *pl;
-  unsigned int rs;
   int i;
 
-  Pico32xMem = calloc(1, sizeof(*Pico32xMem));
+  // M68K ROM
-  if (Pico32xMem == NULL) {
+  if (p32x_bios_g != NULL) {
-    elprintf(EL_STATUS, "OOM");
+    elprintf(EL_STATUS|EL_32X, "32x: using supplied 68k BIOS");
-    return;
+    Byteswap(Pico32xMem->m68k_rom, p32x_bios_g, 0x100);
   }
-
+  else {
-  dmac0 = (void *)&Pico32xMem->sh2_peri_regs[0][0x180 / 4];
-
     // generate 68k ROM
-  ps = (unsigned short *)Pico32xMem->m68k_rom;
+    ps = (u16 *)Pico32xMem->m68k_rom;
-  pl = (unsigned int *)Pico32xMem->m68k_rom;
+    pl = (u32 *)ps;
     for (i = 1; i < 0xc0/4; i++)
       pl[i] = HWSWAP(0x880200 + (i - 1) * 6);
 
@@ -1050,41 +1207,67 @@ void PicoMemSetup32x(void)
 #else
     ps[0xfe/2] = 0x4e75; // rts
 #endif
-
+  }
-  // fill remaining mem with ROM
+  // fill remaining m68k_rom page with game ROM
   memcpy(Pico32xMem->m68k_rom + 0x100, Pico.rom + 0x100, sizeof(Pico32xMem->m68k_rom) - 0x100);
 
-  // 32X ROM
+  // MSH2
-  // TODO: move
+  if (p32x_bios_m != NULL) {
-  {
+    elprintf(EL_STATUS|EL_32X, "32x: using supplied master SH2 BIOS");
-    FILE *f = fopen("32X_M_BIOS.BIN", "rb");
+    Byteswap(Pico32xMem->sh2_rom_m, p32x_bios_m, sizeof(Pico32xMem->sh2_rom_m));
-    int i;
-    if (f == NULL) {
-      printf("missing 32X_M_BIOS.BIN\n");
-      exit(1);
   }
-    fread(Pico32xMem->sh2_rom_m, 1, sizeof(Pico32xMem->sh2_rom_m), f);
+  else {
-    fclose(f);
+    pl = (u32 *)Pico32xMem->sh2_rom_m;
-    f = fopen("32X_S_BIOS.BIN", "rb");
+
-    if (f == NULL) {
+    // fill exception vector table to our trap address
-      printf("missing 32X_S_BIOS.BIN\n");
+    for (i = 0; i < 128; i++)
-      exit(1);
+      pl[i] = HWSWAP(0x200);
+
+    // startup code
+    memcpy(Pico32xMem->sh2_rom_m + 0x200, msh2_code, sizeof(msh2_code));
+
+    // reset SP
+    pl[1] = pl[3] = HWSWAP(0x6040000);
+    // start
+    pl[0] = pl[2] = HWSWAP(0x204);
   }
-    fread(Pico32xMem->sh2_rom_s, 1, sizeof(Pico32xMem->sh2_rom_s), f);
+
-    fclose(f);
+  // SSH2
-    // byteswap
+  if (p32x_bios_s != NULL) {
-    for (i = 0; i < sizeof(Pico32xMem->sh2_rom_m); i += 2) {
+    elprintf(EL_STATUS|EL_32X, "32x: using supplied slave SH2 BIOS");
-      int t = Pico32xMem->sh2_rom_m[i];
+    Byteswap(Pico32xMem->sh2_rom_s, p32x_bios_s, sizeof(Pico32xMem->sh2_rom_s));
-      Pico32xMem->sh2_rom_m[i] = Pico32xMem->sh2_rom_m[i + 1];
-      Pico32xMem->sh2_rom_m[i + 1] = t;
   }
-    for (i = 0; i < sizeof(Pico32xMem->sh2_rom_s); i += 2) {
+  else {
-      int t = Pico32xMem->sh2_rom_s[i];
+    pl = (u32 *)Pico32xMem->sh2_rom_s;
-      Pico32xMem->sh2_rom_s[i] = Pico32xMem->sh2_rom_s[i + 1];
+
-      Pico32xMem->sh2_rom_s[i + 1] = t;
+    // fill exception vector table to our trap address
+    for (i = 0; i < 128; i++)
+      pl[i] = HWSWAP(0x200);
+
+    // startup code
+    memcpy(Pico32xMem->sh2_rom_s + 0x200, ssh2_code, sizeof(ssh2_code));
+
+    // reset SP
+    pl[1] = pl[3] = HWSWAP(0x603f800);
+    // start
+    pl[0] = pl[2] = HWSWAP(0x204);
   }
+}
+
+void PicoMemSetup32x(void)
+{
+  unsigned int rs;
+
+  Pico32xMem = calloc(1, sizeof(*Pico32xMem));
+  if (Pico32xMem == NULL) {
+    elprintf(EL_STATUS, "OOM");
+    return;
   }
 
+  dmac0 = (void *)&Pico32xMem->sh2_peri_regs[0][0x180 / 4];
+
+  get_bios();
+
   // cartridge area becomes unmapped
   // XXX: we take the easy way and don't unmap ROM,
   // so that we can avoid handling the RV bit.
@@ -1110,6 +1293,12 @@ void PicoMemSetup32x(void)
   // 32X ROM (banked)
   bank_switch(0);
 
+  // SYS regs
+  cpu68k_map_set(m68k_read8_map,   0xa10000, 0xa1ffff, PicoRead8_32x_on, 1);
+  cpu68k_map_set(m68k_read16_map,  0xa10000, 0xa1ffff, PicoRead16_32x_on, 1);
+  cpu68k_map_set(m68k_write8_map,  0xa10000, 0xa1ffff, PicoWrite8_32x_on, 1);
+  cpu68k_map_set(m68k_write16_map, 0xa10000, 0xa1ffff, PicoWrite16_32x_on, 1);
+
   // setup poll detector
   m68k_poll.flag = P32XF_68KPOLL;
   m68k_poll.cyc_max = 64;

pico/cart.c

@@ -378,23 +378,21 @@ int pm_close(pm_file *fp)
   return ret;
 }
 
-
+// byteswap, data needs to be int aligned, src can match dst
-static void Byteswap(unsigned char *data,int len)
+void Byteswap(void *dst, const void *src, int len)
 {
-  int i=0;
+  const unsigned int *ps = src;
+  unsigned int *pd = dst;
+  int i, m;
 
-  if (len<2) return; // Too short
+  if (len < 2)
+    return;
 
-  do
+  m = 0x00ff00ff;
-  {
+  for (i = 0; i < len / 4; i++) {
-    unsigned short *pd=(unsigned short *)(data+i);
+    unsigned int t = ps[i];
-    int value=*pd; // Get 2 bytes
+    pd[i] = ((t & m) << 8) | ((t & ~m) >> 8);
-
-    value=(value<<8)|(value>>8); // Byteswap it
-    *pd=(unsigned short)value; // Put 2b ytes
-    i+=2;
   }
-  while (i+2<=len);
 }
 
 // Interleve a 16k block and byteswap
@@ -524,7 +522,7 @@ int PicoCartLoad(pm_file *f,unsigned char **prom,unsigned int *psize,int is_sms)
       elprintf(EL_STATUS, "SMD format detected.");
       DecodeSmd(rom,size); size-=0x200; // Decode and byteswap SMD
     }
-    else Byteswap(rom,size); // Just byteswap
+    else Byteswap(rom, rom, size); // Just byteswap
   }
   else
   {
@@ -621,9 +619,9 @@ static unsigned int rom_crc32(void)
   elprintf(EL_STATUS, "caclulating CRC32..");
 
   // have to unbyteswap for calculation..
-  Byteswap(Pico.rom, Pico.romsize);
+  Byteswap(Pico.rom, Pico.rom, Pico.romsize);
   crc = crc32(0, Pico.rom, Pico.romsize);
-  Byteswap(Pico.rom, Pico.romsize);
+  Byteswap(Pico.rom, Pico.rom, Pico.romsize);
   return crc;
 }
 

pico/pico.h

@@ -37,6 +37,10 @@ extern void emu_video_mode_change(int start_line, int line_count, int is_32cols)
 // this must switch to 32bpp mode
 extern void emu_32x_startup(void);
 
+// optional 32X BIOS, should be left NULL if not used
+// must be 256, 2048, 1024 bytes
+extern void *p32x_bios_g, *p32x_bios_m, *p32x_bios_s;
+
 // Pico.c
 #define POPT_EN_FM          (1<< 0) // 00 000x
 #define POPT_EN_PSG         (1<< 1)

pico/pico_int.h

@@ -252,6 +252,11 @@ extern SH2 sh2s[2];
 #define sh2_vbr(c)    (c) ? ssh2.vbr : msh2.vbr
 #define sh2_sr(c)     (c) ? ssh2.sr : msh2.sr
 
+#define sh2_set_gbr(c, v) \
+  { if (c) ssh2.gbr = v; else msh2.gbr = v; }
+#define sh2_set_vbr(c, v) \
+  { if (c) ssh2.vbr = v; else msh2.vbr = v; }
+
 // ---------------------------------------------------------
 
 // main oscillator clock which controls timing
@@ -416,6 +421,9 @@ typedef struct
 
 // 32X
 #define P32XS_FM    (1<<15)
+#define P32XS_REN   (1<< 7)
+#define P32XS_nRES  (1<< 1)
+#define P32XS_ADEN  (1<< 0)
 #define P32XS2_ADEN (1<< 9)
 #define P32XS_FULL  (1<< 7) // DREQ FIFO full
 #define P32XS_68S   (1<< 2)
@@ -514,6 +522,7 @@ extern areaseek *areaSeek;
 extern areaclose *areaClose;
 
 // cart.c
+void Byteswap(void *dst, const void *src, int len);
 extern void (*PicoCartMemSetup)(void);
 extern void (*PicoCartUnloadHook)(void);
 
@@ -680,6 +689,7 @@ void Pico32xStartup(void);
 void PicoUnload32x(void);
 void PicoFrame32x(void);
 void p32x_update_irls(void);
+void p32x_reset_sh2s(void);
 
 // 32x/memory.c
 struct Pico32xMem *Pico32xMem;

platform/common/emu.c

@@ -1299,6 +1299,19 @@ void emu_init(void)
 	char path[512];
 	int pos;
 
+#if 0
+	// FIXME: handle through menu, etc
+	FILE *f;
+	f = fopen("32X_M_BIOS.BIN", "rb");
+	p32x_bios_m = malloc(2048);
+	fread(p32x_bios_m, 1, 2048, f);
+	fclose(f);
+	f = fopen("32X_S_BIOS.BIN", "rb");
+	p32x_bios_s = malloc(1024);
+	fread(p32x_bios_s, 1, 1024, f);
+	fclose(f);
+#endif
+
 	/* make dirs for saves */
 	pos = plat_get_root_dir(path, sizeof(path) - 4);
 	mkdir_path(path, pos, "mds");