X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=pico%2F32x%2Fmemory.c;h=b179873909e35f1e981d4cb728419df282c28321;hb=1b3f58449285f3cfcc8ae0e1aefdada44a9e9bc5;hp=902957db4847c5b72bdd559b73d5b517785e0525;hpb=97d3f47fbe0cebff925dfab2a12c047415c85346;p=picodrive.git

diff --git a/pico/32x/memory.c b/pico/32x/memory.c
index 902957d..b179873 100644
--- a/pico/32x/memory.c
+++ b/pico/32x/memory.c
@@ -7,29 +7,29 @@ struct Pico32xMem *Pico32xMem;
 
 static void bank_switch(int b);
 
-#define MSB8(x) ((x) >> 8)
-
 // poll detection
 #define POLL_THRESHOLD 6
 
 struct poll_det {
-	int addr, pc, cnt, flag;
+  u32 addr, cycles, cyc_max;
+  int cnt, flag;
 };
 static struct poll_det m68k_poll, sh2_poll[2];
 
-static int p32x_poll_detect(struct poll_det *pd, u32 a, u32 pc, int is_vdp)
+static int p32x_poll_detect(struct poll_det *pd, u32 a, u32 cycles, int is_vdp)
 {
   int ret = 0, flag = pd->flag;
 
   if (is_vdp)
     flag <<= 3;
 
-  if (a - 2 <= pd->addr && pd->addr <= a + 2 && pd->pc == pc) {
+  if (a - 2 <= pd->addr && pd->addr <= a + 2 && cycles - pd->cycles < pd->cyc_max) {
     pd->cnt++;
     if (pd->cnt > POLL_THRESHOLD) {
       if (!(Pico32x.emu_flags & flag)) {
-        elprintf(EL_32X, "%s poll addr %08x @ %06x",
-          flag == P32XF_68KPOLL ? "m68k" : (flag == P32XF_MSH2POLL ? "msh2" : "ssh2"), a, pc);
+        elprintf(EL_32X, "%s poll addr %08x, cyc %u",
+          flag & (P32XF_68KPOLL|P32XF_68KVPOLL) ? "m68k" :
+          (flag & (P32XF_MSH2POLL|P32XF_MSH2VPOLL) ? "msh2" : "ssh2"), a, cycles - pd->cycles);
         ret = 1;
       }
       Pico32x.emu_flags |= flag;
@@ -38,7 +38,7 @@ static int p32x_poll_detect(struct poll_det *pd, u32 a, u32 pc, int is_vdp)
   else
     pd->cnt = 0;
   pd->addr = a;
-  pd->pc = pc;
+  pd->cycles = cycles;
 
   return ret;
 }
@@ -47,18 +47,24 @@ static int p32x_poll_undetect(struct poll_det *pd, int is_vdp)
 {
   int ret = 0, flag = pd->flag;
   if (is_vdp)
-    flag <<= 3;
-  if (pd->cnt > POLL_THRESHOLD)
+    flag <<= 3; // VDP only
+  else
+    flag |= flag << 3; // both
+  if (Pico32x.emu_flags & flag) {
+    elprintf(EL_32X, "poll %02x -> %02x", Pico32x.emu_flags, Pico32x.emu_flags & ~flag);
     ret = 1;
-  pd->addr = pd->cnt = 0;
+  }
   Pico32x.emu_flags &= ~flag;
+  pd->addr = pd->cnt = 0;
   return ret;
 }
 
-void p32x_poll_event(int is_vdp)
+void p32x_poll_event(int cpu_mask, int is_vdp)
 {
-  p32x_poll_undetect(&sh2_poll[0], is_vdp);
-  p32x_poll_undetect(&sh2_poll[1], is_vdp);
+  if (cpu_mask & 1)
+    p32x_poll_undetect(&sh2_poll[0], is_vdp);
+  if (cpu_mask & 2)
+    p32x_poll_undetect(&sh2_poll[1], is_vdp);
 }
 
 // SH2 faking
@@ -108,9 +114,13 @@ static void dma_68k2sh2_do(void)
   if (dmac0->tcr0 != *dreqlen)
     elprintf(EL_32X|EL_ANOMALY, "tcr0 and dreq len differ: %d != %d", dmac0->tcr0, *dreqlen);
 
+  // HACK: assume bus is busy and SH2 is halted
+  // XXX: use different mechanism for this, not poll det
+  Pico32x.emu_flags |= P32XF_MSH2POLL; // id ? P32XF_SSH2POLL : P32XF_MSH2POLL;
+
   for (i = 0; i < Pico32x.dmac_ptr && dmac0->tcr0 > 0; i++) {
     extern void p32x_sh2_write16(u32 a, u32 d, int id);
-      elprintf(EL_32X|EL_ANOMALY, "dmaw [%08x] %04x, left %d", dmac0->dar0, Pico32x.dmac_fifo[i], *dreqlen);
+      elprintf(EL_32X, "dmaw [%08x] %04x, left %d", dmac0->dar0, Pico32x.dmac_fifo[i], *dreqlen);
     p32x_sh2_write16(dmac0->dar0, Pico32x.dmac_fifo[i], 0);
     dmac0->dar0 += 2;
     dmac0->tcr0--;
@@ -121,8 +131,10 @@ static void dma_68k2sh2_do(void)
   Pico32x.regs[6 / 2] &= ~P32XS_FULL;
   if (*dreqlen == 0)
     Pico32x.regs[6 / 2] &= ~P32XS_68S; // transfer complete
-  if (dmac0->tcr0 == 0)
+  if (dmac0->tcr0 == 0) {
     dmac0->chcr0 |= 2; // DMA has ended normally
+    p32x_poll_undetect(&sh2_poll[0], 0);
+  }
 }
 
 // ------------------------------------------------------------------
@@ -132,19 +144,17 @@ static u32 p32x_reg_read16(u32 a)
 {
   a &= 0x3e;
 
+  if (a == 2) // INTM, INTS
+    return ((Pico32x.sh2irqi[0] & P32XI_CMD) >> 4) | ((Pico32x.sh2irqi[1] & P32XI_CMD) >> 3);
 #if 0
   if ((a & 0x30) == 0x20)
     return sh2_comm_faker(a);
 #else
-  if ((a & 0x30) == 0x20 && p32x_poll_detect(&m68k_poll, a, SekPc, 0)) {
+  if ((a & 0x30) == 0x20 && p32x_poll_detect(&m68k_poll, a, SekCyclesDoneT(), 0)) {
     SekEndRun(16);
   }
 #endif
-#ifdef FAKE_SH2
-  // fake only slave for now
-  if (a == 0x24 || a == 0x26)
-    return sh2_comm_faker(a);
-#endif
+
   if ((a & 0x30) == 0x30)
     return p32x_pwm_read16(a);
 
@@ -171,27 +181,41 @@ static void p32x_reg_write8(u32 a, u32 d)
   switch (a) {
     case 0: // adapter ctl
       r[0] = (r[0] & 0x83) | ((d << 8) & P32XS_FM);
-      break;
+      return;
     case 3: // irq ctl
       if ((d & 1) && !(Pico32x.sh2irqi[0] & P32XI_CMD)) {
         Pico32x.sh2irqi[0] |= P32XI_CMD;
         p32x_update_irls();
+        SekEndRun(16);
       }
       if ((d & 2) && !(Pico32x.sh2irqi[1] & P32XI_CMD)) {
         Pico32x.sh2irqi[1] |= P32XI_CMD;
         p32x_update_irls();
+        SekEndRun(16);
       }
-      break;
+      return;
     case 5: // bank
       d &= 7;
       if (r[4 / 2] != d) {
         r[4 / 2] = d;
         bank_switch(d);
       }
-      break;
+      return;
     case 7: // DREQ ctl
       r[6 / 2] = (r[6 / 2] & P32XS_FULL) | (d & (P32XS_68S|P32XS_DMA|P32XS_RV));
-      break;
+      return;
+    case 0x1b: // TV
+      r[0x1a / 2] = d;
+      return;
+  }
+
+  if ((a & 0x30) == 0x20) {
+    u8 *r8 = (u8 *)r;
+    r8[a ^ 1] = d;
+    if (p32x_poll_undetect(&sh2_poll[0], 0) || p32x_poll_undetect(&sh2_poll[1], 0))
+      // if some SH2 is busy waiting, it needs to see the result ASAP
+      SekEndRun(16);
+    return;
   }
 }
 
@@ -271,6 +295,11 @@ static void p32x_vdp_write8(u32 a, u32 d)
       if ((r[0] ^ d) & P32XV_PRI)
         Pico32x.dirty_pal = 1;
       r[0] = (r[0] & P32XV_nPAL) | (d & 0xff);
+      if ((d & 3) == 3)
+        elprintf(EL_32X|EL_ANOMALY, "TODO: mode3");
+      break;
+    case 0x05: // fill len
+      r[4 / 2] = d & 0xff;
       break;
     case 0x0b:
       d &= 1;
@@ -287,6 +316,24 @@ static void p32x_vdp_write8(u32 a, u32 d)
 
 static void p32x_vdp_write16(u32 a, u32 d)
 {
+  a &= 0x0e;
+  if (a == 6) { // fill start
+    Pico32x.vdp_regs[6 / 2] = d;
+    return;
+  }
+  if (a == 8) { // fill data
+    u16 *dram = Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1];
+    int len = Pico32x.vdp_regs[4 / 2] + 1;
+    a = Pico32x.vdp_regs[6 / 2];
+    while (len--) {
+      dram[a] = d;
+      a = (a & 0xff00) | ((a + 1) & 0xff);
+    }
+    Pico32x.vdp_regs[6 / 2] = a;
+    Pico32x.vdp_regs[8 / 2] = d;
+    return;
+  }
+
   p32x_vdp_write8(a | 1, d);
 }
 
@@ -300,7 +347,9 @@ static u32 p32x_sh2reg_read16(u32 a, int cpuid)
 
   switch (a) {
     case 0x00: // adapter/irq ctl
-      return (r[0] & P32XS_FM) | P32XS2_ADEN | Pico32x.sh2irq_mask[cpuid];
+      return (r[0] & P32XS_FM) | Pico32x.sh2_regs[0] | Pico32x.sh2irq_mask[cpuid];
+    case 0x04: // H count
+      return Pico32x.sh2_regs[4 / 2];
     case 0x10: // DREQ len
       return r[a / 2];
   }
@@ -310,7 +359,7 @@ static u32 p32x_sh2reg_read16(u32 a, int cpuid)
     return r[a / 2];
   // comm port
   if ((a & 0x30) == 0x20) {
-    if (p32x_poll_detect(&sh2_poll[cpuid], a, sh2_pc(cpuid), 0))
+    if (p32x_poll_detect(&sh2_poll[cpuid], a, ash2_cycles_done(), 0))
       ash2_end_run(8);
     return r[a / 2];
   }
@@ -325,9 +374,28 @@ static u32 p32x_sh2reg_read16(u32 a, int cpuid)
 static void p32x_sh2reg_write8(u32 a, u32 d, int cpuid)
 {
   a &= 0xff;
-  if (a == 1) {
-    Pico32x.sh2irq_mask[cpuid] = d & 0x0f;
-    p32x_update_irls();
+  switch (a) {
+    case 0: // FM
+      Pico32x.regs[0] &= ~P32XS_FM;
+      Pico32x.regs[0] |= (d << 8) & P32XS_FM;
+      return;
+    case 1: // 
+      Pico32x.sh2irq_mask[cpuid] = d & 0x8f;
+      Pico32x.sh2_regs[0] &= ~0x80;
+      Pico32x.sh2_regs[0] |= d & 0x80;
+      p32x_update_irls();
+      return;
+    case 5: // H count
+      Pico32x.sh2_regs[4 / 2] = d & 0xff;
+      return;
+  }
+
+  if ((a & 0x30) == 0x20) {
+    u8 *r8 = (u8 *)Pico32x.regs;
+    r8[a ^ 1] = d;
+    p32x_poll_undetect(&m68k_poll, 0);
+    p32x_poll_undetect(&sh2_poll[cpuid ^ 1], 0);
+    return;
   }
 }
 
@@ -349,11 +417,18 @@ static void p32x_sh2reg_write16(u32 a, u32 d, int cpuid)
   }
 
   switch (a) {
+    case 0: // FM
+      Pico32x.regs[0] &= ~P32XS_FM;
+      Pico32x.regs[0] |= d & P32XS_FM;
+      break;
     case 0x14: Pico32x.sh2irqs &= ~P32XI_VRES; goto irls;
     case 0x16: Pico32x.sh2irqs &= ~P32XI_VINT; goto irls;
     case 0x18: Pico32x.sh2irqs &= ~P32XI_HINT; goto irls;
     case 0x1a: Pico32x.sh2irqi[cpuid] &= ~P32XI_CMD; goto irls;
-    case 0x1c: Pico32x.sh2irqs &= ~P32XI_PWM;  goto irls;
+    case 0x1c:
+      Pico32x.sh2irqs &= ~P32XI_PWM;
+      p32x_pwm_irq_check(0);
+      goto irls;
   }
 
   p32x_sh2reg_write8(a | 1, d, cpuid);
@@ -363,7 +438,23 @@ irls:
   p32x_update_irls();
 }
 
-static u32 sh2_peripheral_read(u32 a, int id)
+// ------------------------------------------------------------------
+// SH2 internal peripherals
+static u32 sh2_peripheral_read8(u32 a, int id)
+{
+  u8 *r = (void *)Pico32xMem->sh2_peri_regs[id];
+  u32 d;
+
+  a &= 0x1ff;
+  d = r[a];
+  if (a == 4)
+    d = 0x84; // SCI SSR
+
+  elprintf(EL_32X, "%csh2 peri r8  [%08x]       %02x @%06x", id ? 's' : 'm', a | ~0x1ff, d, sh2_pc(id));
+  return d;
+}
+
+static u32 sh2_peripheral_read32(u32 a, int id)
 {
   u32 d;
   a &= 0x1fc;
@@ -373,31 +464,42 @@ static u32 sh2_peripheral_read(u32 a, int id)
   return d;
 }
 
-static void sh2_peripheral_write(u32 a, u32 d, int id)
+static void sh2_peripheral_write8(u32 a, u32 d, int id)
 {
-  unsigned int *r = Pico32xMem->sh2_peri_regs[id];
+  u8 *r = (void *)Pico32xMem->sh2_peri_regs[id];
+  elprintf(EL_32X, "%csh2 peri w8  [%08x]       %02x @%06x", id ? 's' : 'm', a, d, sh2_pc(id));
+
+  a &= 0x1ff;
+  r[a] = d;
+}
+
+static void sh2_peripheral_write32(u32 a, u32 d, int id)
+{
+  u32 *r = Pico32xMem->sh2_peri_regs[id];
   elprintf(EL_32X, "%csh2 peri w32 [%08x] %08x @%06x", id ? 's' : 'm', a, d, sh2_pc(id));
 
   a &= 0x1fc;
   r[a / 4] = d;
 
   switch (a) {
-    // division unit:
+    // division unit (TODO: verify):
     case 0x104: // DVDNT: divident L, starts divide
       elprintf(EL_32X, "%csh2 divide %08x / %08x", id ? 's' : 'm', d, r[0x100 / 4]);
       if (r[0x100 / 4]) {
-        r[0x118 / 4] = r[0x110 / 4] = d % r[0x100 / 4];
-        r[0x11c / 4] = r[0x114 / 4] = d / r[0x100 / 4];
+        signed int divisor = r[0x100 / 4];
+                       r[0x118 / 4] = r[0x110 / 4] = (signed int)d % divisor;
+        r[0x104 / 4] = r[0x11c / 4] = r[0x114 / 4] = (signed int)d / divisor;
       }
       break;
     case 0x114:
       elprintf(EL_32X, "%csh2 divide %08x%08x / %08x @%08x",
         id ? 's' : 'm', r[0x110 / 4], d, r[0x100 / 4], sh2_pc(id));
       if (r[0x100 / 4]) {
-        long long divident = (long long)r[0x110 / 4] << 32 | d;
+        signed long long divident = (signed long long)r[0x110 / 4] << 32 | d;
+        signed int divisor = r[0x100 / 4];
         // XXX: undocumented mirroring to 0x118,0x11c?
-        r[0x118 / 4] = r[0x110 / 4] = divident % r[0x100 / 4];
-        r[0x11c / 4] = r[0x114 / 4] = divident / r[0x100 / 4];
+        r[0x118 / 4] = r[0x110 / 4] = divident % divisor;
+        r[0x11c / 4] = r[0x114 / 4] = divident / divisor;
       }
       break;
   }
@@ -406,6 +508,10 @@ static void sh2_peripheral_write(u32 a, u32 d, int id)
     elprintf(EL_32X, "sh2 DMA %08x -> %08x, cnt %d, chcr %04x @%06x",
       dmac0->sar0, dmac0->dar0, dmac0->tcr0, dmac0->chcr0, sh2_pc(id));
     dmac0->tcr0 &= 0xffffff;
+
+    // HACK: assume 68k starts writing soon and end the timeslice
+    ash2_end_run(16);
+
     // DREQ is only sent after first 4 words are written.
     // we do multiple of 4 words to avoid messing up alignment
     if (dmac0->sar0 == 0x20004012 && Pico32x.dmac_ptr && (Pico32x.dmac_ptr & 3) == 0) {
@@ -614,38 +720,41 @@ u32 p32x_sh2_read8(u32 a, int id)
   if (id == 1 && a < sizeof(Pico32xMem->sh2_rom_s))
     return Pico32xMem->sh2_rom_s[a ^ 1];
 
-  if ((a & 0x0ffc0000) == 0x06000000)
+  if ((a & 0xdffc0000) == 0x06000000)
     return Pico32xMem->sdram[(a & 0x3ffff) ^ 1];
 
-  if ((a & 0x0fc00000) == 0x02000000)
+  if ((a & 0xdfc00000) == 0x02000000)
     if ((a & 0x003fffff) < Pico.romsize)
       return Pico.rom[(a & 0x3fffff) ^ 1];
 
   if ((a & ~0xfff) == 0xc0000000)
     return Pico32xMem->data_array[id][(a & 0xfff) ^ 1];
 
-  if ((a & 0x0ffe0000) == 0x04000000) {
+  if ((a & 0xdffe0000) == 0x04000000) {
     u8 *dram = (u8 *)Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1];
     return dram[(a & 0x1ffff) ^ 1];
   }
 
-  if ((a & 0x0fffff00) == 0x4000) {
+  if ((a & 0xdfffff00) == 0x4000) {
     d = p32x_sh2reg_read16(a, id);
     goto out_16to8;
   }
 
-  if ((a & 0x0fffff00) == 0x4100) {
+  if ((a & 0xdfffff00) == 0x4100) {
     d = p32x_vdp_read16(a);
-    if (p32x_poll_detect(&sh2_poll[id], a, sh2_pc(id), 1))
+    if (p32x_poll_detect(&sh2_poll[id], a, ash2_cycles_done(), 1))
       ash2_end_run(8);
     goto out_16to8;
   }
 
-  if ((a & 0x0fffff00) == 0x4200) {
+  if ((a & 0xdfffff00) == 0x4200) {
     d = Pico32xMem->pal[(a & 0x1ff) / 2];
     goto out_16to8;
   }
 
+  if ((a & 0xfffffe00) == 0xfffffe00)
+    return sh2_peripheral_read8(a, id);
+
   elprintf(EL_UIO, "%csh2 unmapped r8  [%08x]       %02x @%06x",
     id ? 's' : 'm', a, d, sh2_pc(id));
   return d;
@@ -670,32 +779,34 @@ u32 p32x_sh2_read16(u32 a, int id)
   if (id == 1 && a < sizeof(Pico32xMem->sh2_rom_s))
     return *(u16 *)(Pico32xMem->sh2_rom_s + a);
 
-  if ((a & 0x0ffc0000) == 0x06000000)
+  if ((a & 0xdffc0000) == 0x06000000)
     return ((u16 *)Pico32xMem->sdram)[(a & 0x3ffff) / 2];
 
-  if ((a & 0x0fc00000) == 0x02000000)
+  if ((a & 0xdfc00000) == 0x02000000)
     if ((a & 0x003fffff) < Pico.romsize)
       return ((u16 *)Pico.rom)[(a & 0x3fffff) / 2];
 
   if ((a & ~0xfff) == 0xc0000000)
     return ((u16 *)Pico32xMem->data_array[id])[(a & 0xfff) / 2];
 
-  if ((a & 0x0ffe0000) == 0x04000000)
+  if ((a & 0xdffe0000) == 0x04000000)
     return Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1][(a & 0x1ffff) / 2];
 
-  if ((a & 0x0fffff00) == 0x4000) {
+  if ((a & 0xdfffff00) == 0x4000) {
     d = p32x_sh2reg_read16(a, id);
+    if (!(EL_LOGMASK & EL_PWM) && (a & 0x30) == 0x30) // hide PWM
+      return d;
     goto out;
   }
 
-  if ((a & 0x0fffff00) == 0x4100) {
+  if ((a & 0xdfffff00) == 0x4100) {
     d = p32x_vdp_read16(a);
-    if (p32x_poll_detect(&sh2_poll[id], a, sh2_pc(id), 1))
+    if (p32x_poll_detect(&sh2_poll[id], a, ash2_cycles_done(), 1))
       ash2_end_run(8);
     goto out;
   }
 
-  if ((a & 0x0fffff00) == 0x4200) {
+  if ((a & 0xdfffff00) == 0x4200) {
     d = Pico32xMem->pal[(a & 0x1ff) / 2];
     goto out;
   }
@@ -713,7 +824,7 @@ out:
 u32 p32x_sh2_read32(u32 a, int id)
 {
   if ((a & 0xfffffe00) == 0xfffffe00)
-    return sh2_peripheral_read(a, id);
+    return sh2_peripheral_read32(a, id);
 
 //  elprintf(EL_UIO, "sh2 r32 [%08x] %08x @%06x", a, d, ash2_pc());
   return (p32x_sh2_read16(a, id) << 16) | p32x_sh2_read16(a + 2, id);
@@ -721,19 +832,22 @@ u32 p32x_sh2_read32(u32 a, int id)
 
 void p32x_sh2_write8(u32 a, u32 d, int id)
 {
-  if ((a & 0x0ffffc00) == 0x4000)
+  if ((a & 0xdffffc00) == 0x4000)
     elprintf(EL_32X, "%csh2 w8  [%08x]       %02x @%06x",
       id ? 's' : 'm', a, d & 0xff, sh2_pc(id));
 
-  if ((a & 0x0ffc0000) == 0x06000000) {
+  if ((a & 0xdffc0000) == 0x06000000) {
     Pico32xMem->sdram[(a & 0x3ffff) ^ 1] = d;
     return;
   }
 
-  if ((a & 0x0ffe0000) == 0x04000000) {
-    u8 *dram = (u8 *)Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1];
-    dram[(a & 0x1ffff) ^ 1] = d;
-    return;
+  if ((a & 0xdffc0000) == 0x04000000) {
+    u8 *dram;
+    if (!(a & 0x20000) || d) {
+      dram = (u8 *)Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1];
+      dram[(a & 0x1ffff) ^ 1] = d;
+      return;
+    }
   }
 
   if ((a & ~0xfff) == 0xc0000000) {
@@ -741,27 +855,36 @@ void p32x_sh2_write8(u32 a, u32 d, int id)
     return;
   }
 
-  if ((a & 0x0fffff00) == 0x4100) {
+  if ((a & 0xdfffff00) == 0x4100) {
     p32x_vdp_write8(a, d);
     return;
   }
 
-  if ((a & 0x0fffff00) == 0x4000) {
+  if ((a & 0xdfffff00) == 0x4000) {
     p32x_sh2reg_write8(a, d, id);
     return;
   }
 
+  if ((a & 0xfffffe00) == 0xfffffe00) {
+    sh2_peripheral_write8(a, d, id);
+    return;
+  }
+
   elprintf(EL_UIO, "%csh2 unmapped w8  [%08x]       %02x @%06x",
     id ? 's' : 'm', a, d & 0xff, sh2_pc(id));
 }
 
 void p32x_sh2_write16(u32 a, u32 d, int id)
 {
-  if ((a & 0x0ffffc00) == 0x4000)
+  if ((a & 0xdffffc00) == 0x4000 && ((EL_LOGMASK & EL_PWM) || (a & 0x30) != 0x30)) // hide PWM
     elprintf(EL_32X, "%csh2 w16 [%08x]     %04x @%06x",
       id ? 's' : 'm', a, d & 0xffff, sh2_pc(id));
 
-  if ((a & 0x0ffc0000) == 0x06000000) {
+  // ignore "Associative purge space"
+  if ((a & 0xf8000000) == 0x40000000)
+    return;
+
+  if ((a & 0xdffc0000) == 0x06000000) {
     ((u16 *)Pico32xMem->sdram)[(a & 0x3ffff) / 2] = d;
     return;
   }
@@ -771,23 +894,32 @@ void p32x_sh2_write16(u32 a, u32 d, int id)
     return;
   }
 
-  if ((a & 0x0ffe0000) == 0x04000000) {
-    Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1][(a & 0x1ffff) / 2] = d;
+  if ((a & 0xdffc0000) == 0x04000000) {
+    u16 *pd = &Pico32xMem->dram[(Pico32x.vdp_regs[0x0a/2] & P32XV_FS) ^ 1][(a & 0x1ffff) / 2];
+    if (!(a & 0x20000)) {
+      *pd = d;
+      return;
+    }
+    // overwrite
+    if (!(d & 0xff00)) d |= *pd & 0xff00;
+    if (!(d & 0x00ff)) d |= *pd & 0x00ff;
+    *pd = d;
     return;
   }
 
-  if ((a & 0x0fffff00) == 0x4100) {
+  if ((a & 0xdfffff00) == 0x4100) {
+    sh2_poll[id].cnt = 0; // for poll before VDP accesses
     p32x_vdp_write16(a, d);
     return;
   }
 
-  if ((a & 0x0ffffe00) == 0x4200) {
+  if ((a & 0xdffffe00) == 0x4200) {
     Pico32xMem->pal[(a & 0x1ff) / 2] = d;
     Pico32x.dirty_pal = 1;
     return;
   }
 
-  if ((a & 0x0fffff00) == 0x4000) {
+  if ((a & 0xdfffff00) == 0x4000) {
     p32x_sh2reg_write16(a, d, id);
     return;
   }
@@ -799,7 +931,7 @@ void p32x_sh2_write16(u32 a, u32 d, int id)
 void p32x_sh2_write32(u32 a, u32 d, int id)
 {
   if ((a & 0xfffffe00) == 0xfffffe00) {
-    sh2_peripheral_write(a, d, id);
+    sh2_peripheral_write32(a, d, id);
     return;
   }
 
@@ -902,7 +1034,10 @@ void PicoMemSetup32x(void)
 
   // setup poll detector
   m68k_poll.flag = P32XF_68KPOLL;
+  m68k_poll.cyc_max = 64;
   sh2_poll[0].flag = P32XF_MSH2POLL;
+  sh2_poll[0].cyc_max = 16;
   sh2_poll[1].flag = P32XF_SSH2POLL;
+  sh2_poll[1].cyc_max = 16;
 }