X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=e0cff62ca622b1b5b04dd942626a76e89d4ab5f0;hp=081532981307c9078f58c38d45c090970582a868;hb=e3c6bdb5e46f72f063bb7f588da6588ac1893b17;hpb=d1e4ebd9988a9a5d9fb38b89f19e24b9ab6029d7

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 08153298..e0cff62c 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -35,13 +35,18 @@ static int sceBlock;
 #endif
 
 #include "new_dynarec_config.h"
-#include "../psxhle.h" //emulator interface
-#include "emu_if.h" //emulator interface
+#include "../psxhle.h"
+#include "../psxinterpreter.h"
+#include "../gte.h"
+#include "emu_if.h" // emulator interface
 
 #define noinline __attribute__((noinline,noclone))
 #ifndef ARRAY_SIZE
 #define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
 #endif
+#ifndef min
+#define min(a, b) ((b) < (a) ? (b) : (a))
+#endif
 
 //#define DISASM
 //#define assem_debug printf
@@ -62,9 +67,27 @@ static int sceBlock;
 #include "assem_arm64.h"
 #endif
 
+#define RAM_SIZE 0x200000
 #define MAXBLOCK 4096
 #define MAX_OUTPUT_BLOCK_SIZE 262144
 
+struct ndrc_mem
+{
+  u_char translation_cache[1 << TARGET_SIZE_2];
+  struct
+  {
+    struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
+    const void *f[2048 / sizeof(void *)];
+  } tramp;
+};
+
+#ifdef BASE_ADDR_DYNAMIC
+static struct ndrc_mem *ndrc;
+#else
+static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
+static struct ndrc_mem *ndrc = &ndrc_;
+#endif
+
 // stubs
 enum stub_type {
   CC_STUB = 1,
@@ -166,9 +189,11 @@ struct link_entry
   static char ooo[MAXBLOCK];
   static uint64_t unneeded_reg[MAXBLOCK];
   static uint64_t branch_unneeded_reg[MAXBLOCK];
-  static signed char regmap_pre[MAXBLOCK][HOST_REGS];
-  static uint64_t current_constmap[HOST_REGS];
-  static uint64_t constmap[MAXBLOCK][HOST_REGS];
+  static signed char regmap_pre[MAXBLOCK][HOST_REGS]; // pre-instruction i?
+  // contains 'real' consts at [i] insn, but may differ from what's actually
+  // loaded in host reg as 'final' value is always loaded, see get_final_value()
+  static uint32_t current_constmap[HOST_REGS];
+  static uint32_t constmap[MAXBLOCK][HOST_REGS];
   static struct regstat regs[MAXBLOCK];
   static struct regstat branch_regs[MAXBLOCK];
   static signed char minimum_free_regs[MAXBLOCK];
@@ -196,8 +221,11 @@ struct link_entry
 #endif
 
   int new_dynarec_hacks;
+  int new_dynarec_hacks_pergame;
   int new_dynarec_did_compile;
 
+  #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
+
   extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
   extern int last_count;  // last absolute target, often = next_interupt
   extern int pcaddr;
@@ -272,7 +300,7 @@ struct link_entry
 #define DJT_2 (void *)2l
 
 // asm linkage
-int new_recompile_block(int addr);
+int new_recompile_block(u_int addr);
 void *get_addr_ht(u_int vaddr);
 void invalidate_block(u_int block);
 void invalidate_addr(u_int addr);
@@ -284,9 +312,8 @@ void verify_code_ds();
 void cc_interrupt();
 void fp_exception();
 void fp_exception_ds();
-void jump_syscall_hle();
-void jump_hlecall();
-void jump_intcall();
+void jump_to_new_pc();
+void call_gteStall();
 void new_dyna_leave();
 
 // Needed by assembler
@@ -297,18 +324,22 @@ static void load_all_regs(signed char i_regmap[]);
 static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
 static void load_regs_entry(int t);
 static void load_all_consts(signed char regmap[],u_int dirty,int i);
+static u_int get_host_reglist(const signed char *regmap);
 
-static int verify_dirty(u_int *ptr);
+static int verify_dirty(const u_int *ptr);
 static int get_final_value(int hr, int i, int *value);
 static void add_stub(enum stub_type type, void *addr, void *retaddr,
   u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
-  int i, int addr_reg, struct regstat *i_regs, int ccadj, u_int reglist);
+  int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
 static void add_to_linker(void *addr, u_int target, int ext);
 static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override);
 static void *get_direct_memhandler(void *table, u_int addr,
   enum stub_type type, uintptr_t *addr_host);
+static void cop2_call_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
 static void pass_args(int a0, int a1);
+static void emit_far_jump(const void *f);
+static void emit_far_call(const void *f);
 
 static void mprotect_w_x(void *start, void *end, int is_x)
 {
@@ -337,7 +368,7 @@ static void start_tcache_write(void *start, void *end)
 
 static void end_tcache_write(void *start, void *end)
 {
-#ifdef __arm__
+#if defined(__arm__) || defined(__aarch64__)
   size_t len = (char *)end - (char *)start;
   #if   defined(__BLACKBERRY_QNX__)
   msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
@@ -347,12 +378,14 @@ static void end_tcache_write(void *start, void *end)
   sceKernelSyncVMDomain(sceBlock, start, len);
   #elif defined(_3DS)
   ctr_flush_invalidate_cache();
+  #elif defined(__aarch64__)
+  // as of 2021, __clear_cache() is still broken on arm64
+  // so here is a custom one :(
+  clear_cache_arm64(start, end);
   #else
   __clear_cache(start, end);
   #endif
   (void)len;
-#else
-  __clear_cache(start, end);
 #endif
 
   mprotect_w_x(start, end, 1);
@@ -361,8 +394,8 @@ static void end_tcache_write(void *start, void *end)
 static void *start_block(void)
 {
   u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
-  if (end > translation_cache + (1<<TARGET_SIZE_2))
-    end = translation_cache + (1<<TARGET_SIZE_2);
+  if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
+    end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
   start_tcache_write(out, end);
   return out;
 }
@@ -372,16 +405,74 @@ static void end_block(void *start)
   end_tcache_write(start, out);
 }
 
+// also takes care of w^x mappings when patching code
+static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
+
+static void mark_clear_cache(void *target)
+{
+  uintptr_t offset = (u_char *)target - ndrc->translation_cache;
+  u_int mask = 1u << ((offset >> 12) & 31);
+  if (!(needs_clear_cache[offset >> 17] & mask)) {
+    char *start = (char *)((uintptr_t)target & ~4095l);
+    start_tcache_write(start, start + 4095);
+    needs_clear_cache[offset >> 17] |= mask;
+  }
+}
+
+// Clearing the cache is rather slow on ARM Linux, so mark the areas
+// that need to be cleared, and then only clear these areas once.
+static void do_clear_cache(void)
+{
+  int i, j;
+  for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
+  {
+    u_int bitmap = needs_clear_cache[i];
+    if (!bitmap)
+      continue;
+    for (j = 0; j < 32; j++)
+    {
+      u_char *start, *end;
+      if (!(bitmap & (1<<j)))
+        continue;
+
+      start = ndrc->translation_cache + i*131072 + j*4096;
+      end = start + 4095;
+      for (j++; j < 32; j++) {
+        if (!(bitmap & (1<<j)))
+          break;
+        end += 4096;
+      }
+      end_tcache_write(start, end);
+    }
+    needs_clear_cache[i] = 0;
+  }
+}
+
 //#define DEBUG_CYCLE_COUNT 1
 
 #define NO_CYCLE_PENALTY_THR 12
 
 int cycle_multiplier; // 100 for 1.0
+int cycle_multiplier_override;
 
 static int CLOCK_ADJUST(int x)
 {
+  int m = cycle_multiplier_override
+        ? cycle_multiplier_override : cycle_multiplier;
   int s=(x>>31)|1;
-  return (x * cycle_multiplier + s * 50) / 100;
+  return (x * m + s * 50) / 100;
+}
+
+// is the op an unconditional jump?
+static int is_ujump(int i)
+{
+  return itype[i] == UJUMP || itype[i] == RJUMP
+    || (source[i] >> 16) == 0x1000; // beq r0, r0, offset // b offset
+}
+
+static int is_jump(int i)
+{
+  return itype[i] == RJUMP || itype[i] == UJUMP || itype[i] == CJUMP || itype[i] == SJUMP;
 }
 
 static u_int get_page(u_int vaddr)
@@ -529,7 +620,7 @@ void dirty_reg(struct regstat *cur,signed char reg)
   }
 }
 
-void set_const(struct regstat *cur,signed char reg,uint64_t value)
+static void set_const(struct regstat *cur, signed char reg, uint32_t value)
 {
   int hr;
   if(!reg) return;
@@ -541,7 +632,7 @@ void set_const(struct regstat *cur,signed char reg,uint64_t value)
   }
 }
 
-void clear_const(struct regstat *cur,signed char reg)
+static void clear_const(struct regstat *cur, signed char reg)
 {
   int hr;
   if(!reg) return;
@@ -552,7 +643,7 @@ void clear_const(struct regstat *cur,signed char reg)
   }
 }
 
-int is_const(struct regstat *cur,signed char reg)
+static int is_const(struct regstat *cur, signed char reg)
 {
   int hr;
   if(reg<0) return 0;
@@ -564,7 +655,8 @@ int is_const(struct regstat *cur,signed char reg)
   }
   return 0;
 }
-uint64_t get_const(struct regstat *cur,signed char reg)
+
+static uint32_t get_const(struct regstat *cur, signed char reg)
 {
   int hr;
   if(!reg) return 0;
@@ -590,7 +682,7 @@ void lsn(u_char hsn[], int i, int *preferred_reg)
       j=slen-i-1;
       break;
     }
-    if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
+    if (is_ujump(i+j))
     {
       // Don't go past an unconditonal jump
       j++;
@@ -638,7 +730,7 @@ void lsn(u_char hsn[], int i, int *preferred_reg)
     // TODO: preferred register based on backward branch
   }
   // Delay slot should preferably not overwrite branch conditions or cycle count
-  if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP)) {
+  if (i > 0 && is_jump(i-1)) {
     if(rs1[i-1]) if(hsn[rs1[i-1]]>1) hsn[rs1[i-1]]=1;
     if(rs2[i-1]) if(hsn[rs2[i-1]]>1) hsn[rs2[i-1]]=1;
     hsn[CCREG]=1;
@@ -673,7 +765,7 @@ int needed_again(int r, int i)
   int b=-1;
   int rn=10;
 
-  if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
+  if (i > 0 && is_ujump(i-1))
   {
     if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
       return 0; // Don't need any registers if exiting the block
@@ -684,7 +776,7 @@ int needed_again(int r, int i)
       j=slen-i-1;
       break;
     }
-    if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
+    if (is_ujump(i+j))
     {
       // Don't go past an unconditonal jump
       j++;
@@ -740,7 +832,7 @@ int loop_reg(int i, int r, int hr)
       j=slen-i-1;
       break;
     }
-    if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
+    if (is_ujump(i+j))
     {
       // Don't go past an unconditonal jump
       j++;
@@ -832,13 +924,15 @@ static const struct {
   FUNCNAME(jump_handler_write16),
   FUNCNAME(jump_handler_write32),
   FUNCNAME(invalidate_addr),
-  FUNCNAME(verify_code),
-  FUNCNAME(jump_hlecall),
-  FUNCNAME(jump_syscall_hle),
+  FUNCNAME(jump_to_new_pc),
+  FUNCNAME(call_gteStall),
   FUNCNAME(new_dyna_leave),
   FUNCNAME(pcsx_mtc0),
   FUNCNAME(pcsx_mtc0_ds),
   FUNCNAME(do_insn_cmp),
+#ifdef __arm__
+  FUNCNAME(verify_code),
+#endif
 };
 
 static const char *func_name(const void *a)
@@ -866,6 +960,48 @@ static const char *func_name(const void *a)
 #include "assem_arm64.c"
 #endif
 
+static void *get_trampoline(const void *f)
+{
+  size_t i;
+
+  for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
+    if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
+      break;
+  }
+  if (i == ARRAY_SIZE(ndrc->tramp.f)) {
+    SysPrintf("trampoline table is full, last func %p\n", f);
+    abort();
+  }
+  if (ndrc->tramp.f[i] == NULL) {
+    start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
+    ndrc->tramp.f[i] = f;
+    end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
+  }
+  return &ndrc->tramp.ops[i];
+}
+
+static void emit_far_jump(const void *f)
+{
+  if (can_jump_or_call(f)) {
+    emit_jmp(f);
+    return;
+  }
+
+  f = get_trampoline(f);
+  emit_jmp(f);
+}
+
+static void emit_far_call(const void *f)
+{
+  if (can_jump_or_call(f)) {
+    emit_call(f);
+    return;
+  }
+
+  f = get_trampoline(f);
+  emit_call(f);
+}
+
 // Add virtual address mapping to linked list
 void ll_add(struct ll_entry **head,int vaddr,void *addr)
 {
@@ -993,9 +1129,7 @@ static void ll_kill_pointers(struct ll_entry *head,uintptr_t addr,int shift)
     {
       inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
       void *host_addr=find_extjump_insn(head->addr);
-      #if defined(__arm__) || defined(__aarch64__)
-        mark_clear_cache(host_addr);
-      #endif
+      mark_clear_cache(host_addr);
       set_jump_target(host_addr, head->addr);
     }
     head=head->next;
@@ -1021,9 +1155,7 @@ static void invalidate_page(u_int page)
   while(head!=NULL) {
     inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
     void *host_addr=find_extjump_insn(head->addr);
-    #if defined(__arm__) || defined(__aarch64__)
-      mark_clear_cache(host_addr);
-    #endif
+    mark_clear_cache(host_addr);
     set_jump_target(host_addr, head->addr);
     next=head->next;
     free(head);
@@ -1046,9 +1178,7 @@ static void invalidate_block_range(u_int block, u_int first, u_int last)
   for(first=page+1;first<last;first++) {
     invalidate_page(first);
   }
-  #if defined(__arm__) || defined(__aarch64__)
-    do_clear_cache();
-  #endif
+  do_clear_cache();
 
   // Don't trap writes
   invalid_code[block]=1;
@@ -1145,7 +1275,7 @@ void invalidate_addr(u_int addr)
 
 // This is called when loading a save state.
 // Anything could have changed, so invalidate everything.
-void invalidate_all_pages()
+void invalidate_all_pages(void)
 {
   u_int page;
   for(page=0;page<4096;page++)
@@ -1158,6 +1288,7 @@ void invalidate_all_pages()
   #ifdef USE_MINI_HT
   memset(mini_ht,-1,sizeof(mini_ht));
   #endif
+  do_clear_cache();
 }
 
 static void do_invstub(int n)
@@ -1167,7 +1298,7 @@ static void do_invstub(int n)
   set_jump_target(stubs[n].addr, out);
   save_regs(reglist);
   if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
-  emit_call(invalidate_addr);
+  emit_far_call(invalidate_addr);
   restore_regs(reglist);
   emit_jmp(stubs[n].retaddr); // return address
 }
@@ -1616,7 +1747,7 @@ static void imm16_alloc(struct regstat *current,int i)
     else clear_const(current,rt1[i]);
   }
   else {
-    set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
+    set_const(current,rt1[i],imm[i]<<16); // LUI
   }
   dirty_reg(current,rt1[i]);
 }
@@ -1796,19 +1927,19 @@ void cop0_alloc(struct regstat *current,int i)
   minimum_free_regs[i]=HOST_REGS;
 }
 
-static void cop12_alloc(struct regstat *current,int i)
+static void cop2_alloc(struct regstat *current,int i)
 {
-  alloc_reg(current,i,CSREG); // Load status
-  if(opcode2[i]<3) // MFC1/CFC1
+  if (opcode2[i] < 3) // MFC2/CFC2
   {
+    alloc_cc(current,i); // for stalls
+    dirty_reg(current,CCREG);
     if(rt1[i]){
       clear_const(current,rt1[i]);
       alloc_reg(current,i,rt1[i]);
       dirty_reg(current,rt1[i]);
     }
-    alloc_reg_temp(current,i,-1);
   }
-  else if(opcode2[i]>3) // MTC1/CTC1
+  else if (opcode2[i] > 3) // MTC2/CTC2
   {
     if(rs1[i]){
       clear_const(current,rs1[i]);
@@ -1818,13 +1949,15 @@ static void cop12_alloc(struct regstat *current,int i)
       current->u&=~1LL;
       alloc_reg(current,i,0);
     }
-    alloc_reg_temp(current,i,-1);
   }
+  alloc_reg_temp(current,i,-1);
   minimum_free_regs[i]=1;
 }
 
 void c2op_alloc(struct regstat *current,int i)
 {
+  alloc_cc(current,i); // for stalls
+  dirty_reg(current,CCREG);
   alloc_reg_temp(current,i,-1);
 }
 
@@ -1881,8 +2014,9 @@ void delayslot_alloc(struct regstat *current,int i)
       cop0_alloc(current,i);
       break;
     case COP1:
+      break;
     case COP2:
-      cop12_alloc(current,i);
+      cop2_alloc(current,i);
       break;
     case C1LS:
       c1ls_alloc(current,i);
@@ -1948,7 +2082,7 @@ static void add_stub(enum stub_type type, void *addr, void *retaddr,
 }
 
 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
-  int i, int addr_reg, struct regstat *i_regs, int ccadj, u_int reglist)
+  int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
 {
   add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
 }
@@ -2056,10 +2190,11 @@ static void alu_assemble(int i,struct regstat *i_regs)
           s2l=get_reg(i_regs->regmap,rs2[i]);
           if(rs2[i]==0) // rx<r0
           {
-            assert(s1l>=0);
-            if(opcode2[i]==0x2a) // SLT
+            if(opcode2[i]==0x2a&&rs1[i]!=0) { // SLT
+              assert(s1l>=0);
               emit_shrimm(s1l,31,t);
-            else // SLTU (unsigned can not be less than zero)
+            }
+            else // SLTU (unsigned can not be less than zero, 0<0)
               emit_zeroreg(t);
           }
           else if(rs1[i]==0) // r0<rx
@@ -2361,11 +2496,44 @@ void shiftimm_assemble(int i,struct regstat *i_regs)
 }
 
 #ifndef shift_assemble
-void shift_assemble(int i,struct regstat *i_regs)
+static void shift_assemble(int i,struct regstat *i_regs)
 {
-  printf("Need shift_assemble for this architecture.\n");
-  abort();
+  signed char s,t,shift;
+  if (rt1[i] == 0)
+    return;
+  assert(opcode2[i]<=0x07); // SLLV/SRLV/SRAV
+  t = get_reg(i_regs->regmap, rt1[i]);
+  s = get_reg(i_regs->regmap, rs1[i]);
+  shift = get_reg(i_regs->regmap, rs2[i]);
+  if (t < 0)
+    return;
+
+  if(rs1[i]==0)
+    emit_zeroreg(t);
+  else if(rs2[i]==0) {
+    assert(s>=0);
+    if(s!=t) emit_mov(s,t);
+  }
+  else {
+    host_tempreg_acquire();
+    emit_andimm(shift,31,HOST_TEMPREG);
+    switch(opcode2[i]) {
+    case 4: // SLLV
+      emit_shl(s,HOST_TEMPREG,t);
+      break;
+    case 6: // SRLV
+      emit_shr(s,HOST_TEMPREG,t);
+      break;
+    case 7: // SRAV
+      emit_sar(s,HOST_TEMPREG,t);
+      break;
+    default:
+      assert(0);
+    }
+    host_tempreg_release();
+  }
 }
+
 #endif
 
 enum {
@@ -2430,7 +2598,7 @@ static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
   else if(type==MTYPE_1F80) { // scratchpad
     if (psxH == (void *)0x1f800000) {
       host_tempreg_acquire();
-      emit_addimm(addr,-0x1f800000,HOST_TEMPREG);
+      emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
       emit_cmpimm(HOST_TEMPREG,0x1000);
       host_tempreg_release();
       jaddr=out;
@@ -2491,20 +2659,45 @@ static void *get_direct_memhandler(void *table, u_int addr,
   }
 }
 
-static void load_assemble(int i,struct regstat *i_regs)
+static u_int get_host_reglist(const signed char *regmap)
+{
+  u_int reglist = 0, hr;
+  for (hr = 0; hr < HOST_REGS; hr++) {
+    if (hr != EXCLUDE_REG && regmap[hr] >= 0)
+      reglist |= 1 << hr;
+  }
+  return reglist;
+}
+
+static u_int reglist_exclude(u_int reglist, int r1, int r2)
+{
+  if (r1 >= 0)
+    reglist &= ~(1u << r1);
+  if (r2 >= 0)
+    reglist &= ~(1u << r2);
+  return reglist;
+}
+
+// find a temp caller-saved register not in reglist (so assumed to be free)
+static int reglist_find_free(u_int reglist)
+{
+  u_int free_regs = ~reglist & CALLER_SAVE_REGS;
+  if (free_regs == 0)
+    return -1;
+  return __builtin_ctz(free_regs);
+}
+
+static void load_assemble(int i, const struct regstat *i_regs)
 {
   int s,tl,addr;
   int offset;
   void *jaddr=0;
   int memtarget=0,c=0;
   int fastio_reg_override=-1;
-  u_int hr,reglist=0;
+  u_int reglist=get_host_reglist(i_regs->regmap);
   tl=get_reg(i_regs->regmap,rt1[i]);
   s=get_reg(i_regs->regmap,rs1[i]);
   offset=imm[i];
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
   if(s>=0) {
     c=(i_regs->wasconst>>s)&1;
@@ -2631,14 +2824,87 @@ static void load_assemble(int i,struct regstat *i_regs)
 }
 
 #ifndef loadlr_assemble
-void loadlr_assemble(int i,struct regstat *i_regs)
+static void loadlr_assemble(int i, const struct regstat *i_regs)
 {
-  printf("Need loadlr_assemble for this architecture.\n");
-  abort();
+  int s,tl,temp,temp2,addr;
+  int offset;
+  void *jaddr=0;
+  int memtarget=0,c=0;
+  int fastio_reg_override=-1;
+  u_int reglist=get_host_reglist(i_regs->regmap);
+  tl=get_reg(i_regs->regmap,rt1[i]);
+  s=get_reg(i_regs->regmap,rs1[i]);
+  temp=get_reg(i_regs->regmap,-1);
+  temp2=get_reg(i_regs->regmap,FTEMP);
+  addr=get_reg(i_regs->regmap,AGEN1+(i&1));
+  assert(addr<0);
+  offset=imm[i];
+  reglist|=1<<temp;
+  if(offset||s<0||c) addr=temp2;
+  else addr=s;
+  if(s>=0) {
+    c=(i_regs->wasconst>>s)&1;
+    if(c) {
+      memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
+    }
+  }
+  if(!c) {
+    emit_shlimm(addr,3,temp);
+    if (opcode[i]==0x22||opcode[i]==0x26) {
+      emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
+    }else{
+      emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
+    }
+    jaddr=emit_fastpath_cmp_jump(i,temp2,&fastio_reg_override);
+  }
+  else {
+    if(ram_offset&&memtarget) {
+      host_tempreg_acquire();
+      emit_addimm(temp2,ram_offset,HOST_TEMPREG);
+      fastio_reg_override=HOST_TEMPREG;
+    }
+    if (opcode[i]==0x22||opcode[i]==0x26) {
+      emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
+    }else{
+      emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
+    }
+  }
+  if (opcode[i]==0x22||opcode[i]==0x26) { // LWL/LWR
+    if(!c||memtarget) {
+      int a=temp2;
+      if(fastio_reg_override>=0) a=fastio_reg_override;
+      emit_readword_indexed(0,a,temp2);
+      if(fastio_reg_override==HOST_TEMPREG) host_tempreg_release();
+      if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj[i],reglist);
+    }
+    else
+      inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj[i],reglist);
+    if(rt1[i]) {
+      assert(tl>=0);
+      emit_andimm(temp,24,temp);
+      if (opcode[i]==0x22) // LWL
+        emit_xorimm(temp,24,temp);
+      host_tempreg_acquire();
+      emit_movimm(-1,HOST_TEMPREG);
+      if (opcode[i]==0x26) {
+        emit_shr(temp2,temp,temp2);
+        emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
+      }else{
+        emit_shl(temp2,temp,temp2);
+        emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
+      }
+      host_tempreg_release();
+      emit_or(temp2,tl,tl);
+    }
+    //emit_storereg(rt1[i],tl); // DEBUG
+  }
+  if (opcode[i]==0x1A||opcode[i]==0x1B) { // LDL/LDR
+    assert(0);
+  }
 }
 #endif
 
-void store_assemble(int i,struct regstat *i_regs)
+void store_assemble(int i, const struct regstat *i_regs)
 {
   int s,tl;
   int addr,temp;
@@ -2648,7 +2914,7 @@ void store_assemble(int i,struct regstat *i_regs)
   int memtarget=0,c=0;
   int agr=AGEN1+(i&1);
   int fastio_reg_override=-1;
-  u_int hr,reglist=0;
+  u_int reglist=get_host_reglist(i_regs->regmap);
   tl=get_reg(i_regs->regmap,rs2[i]);
   s=get_reg(i_regs->regmap,rs1[i]);
   temp=get_reg(i_regs->regmap,agr);
@@ -2662,9 +2928,6 @@ void store_assemble(int i,struct regstat *i_regs)
   }
   assert(tl>=0);
   assert(temp>=0);
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
   if(offset||s<0||c) addr=temp;
   else addr=s;
@@ -2715,7 +2978,7 @@ void store_assemble(int i,struct regstat *i_regs)
     add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
     jaddr=0;
   }
-  if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
+  if(!(i_regs->waswritten&(1<<rs1[i])) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
     if(!c||memtarget) {
       #ifdef DESTRUCTIVE_SHIFT
       // The x86 shift operation is 'destructive'; it overwrites the
@@ -2746,7 +3009,7 @@ void store_assemble(int i,struct regstat *i_regs)
   }
   // basic current block modification detection..
   // not looking back as that should be in mips cache already
-  // (note: doesn't seem to trigger, migh be broken)
+  // (see Spyro2 title->attract mode)
   if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
     SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
     assert(i_regs->regmap==regs[i].regmap); // not delay slot
@@ -2756,13 +3019,13 @@ void store_assemble(int i,struct regstat *i_regs)
       emit_movimm(start+i*4+4,0);
       emit_writeword(0,&pcaddr);
       emit_addimm(HOST_CCREG,2,HOST_CCREG);
-      emit_call(get_addr_ht);
+      emit_far_call(get_addr_ht);
       emit_jmpreg(0);
     }
   }
 }
 
-void storelr_assemble(int i,struct regstat *i_regs)
+static void storelr_assemble(int i, const struct regstat *i_regs)
 {
   int s,tl;
   int temp;
@@ -2772,7 +3035,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
   void *done0, *done1, *done2;
   int memtarget=0,c=0;
   int agr=AGEN1+(i&1);
-  u_int hr,reglist=0;
+  u_int reglist=get_host_reglist(i_regs->regmap);
   tl=get_reg(i_regs->regmap,rs2[i]);
   s=get_reg(i_regs->regmap,rs1[i]);
   temp=get_reg(i_regs->regmap,agr);
@@ -2785,9 +3048,6 @@ void storelr_assemble(int i,struct regstat *i_regs)
     }
   }
   assert(tl>=0);
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
   assert(temp>=0);
   if(!c) {
     emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
@@ -2802,7 +3062,8 @@ void storelr_assemble(int i,struct regstat *i_regs)
       emit_jmp(0);
     }
   }
-  emit_addimm_no_flags(ram_offset,temp);
+  if(ram_offset)
+    emit_addimm_no_flags(ram_offset,temp);
 
   if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
     assert(0);
@@ -2819,15 +3080,11 @@ void storelr_assemble(int i,struct regstat *i_regs)
   if (opcode[i]==0x2A) { // SWL
     emit_writeword_indexed(tl,0,temp);
   }
-  if (opcode[i]==0x2E) { // SWR
+  else if (opcode[i]==0x2E) { // SWR
     emit_writebyte_indexed(tl,3,temp);
   }
-  if (opcode[i]==0x2C) { // SDL
-    assert(0);
-  }
-  if (opcode[i]==0x2D) { // SDR
+  else
     assert(0);
-  }
   done0=out;
   emit_jmp(0);
   // 1
@@ -2840,16 +3097,10 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_writebyte_indexed(tl,1,temp);
     if(rs2[i]) emit_rorimm(tl,8,tl);
   }
-  if (opcode[i]==0x2E) { // SWR
+  else if (opcode[i]==0x2E) { // SWR
     // Write two lsb into two most significant bytes
     emit_writehword_indexed(tl,1,temp);
   }
-  if (opcode[i]==0x2C) { // SDL
-    assert(0);
-  }
-  if (opcode[i]==0x2D) { // SDR
-    assert(0);
-  }
   done1=out;
   emit_jmp(0);
   // 2
@@ -2863,19 +3114,13 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_writehword_indexed(tl,-2,temp);
     if(rs2[i]) emit_rorimm(tl,16,tl);
   }
-  if (opcode[i]==0x2E) { // SWR
+  else if (opcode[i]==0x2E) { // SWR
     // Write 3 lsb into three most significant bytes
     emit_writebyte_indexed(tl,-1,temp);
     if(rs2[i]) emit_rorimm(tl,8,tl);
     emit_writehword_indexed(tl,0,temp);
     if(rs2[i]) emit_rorimm(tl,24,tl);
   }
-  if (opcode[i]==0x2C) { // SDL
-    assert(0);
-  }
-  if (opcode[i]==0x2D) { // SDR
-    assert(0);
-  }
   done2=out;
   emit_jmp(0);
   // 3
@@ -2886,28 +3131,16 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_writebyte_indexed(tl,-3,temp);
     if(rs2[i]) emit_rorimm(tl,8,tl);
   }
-  if (opcode[i]==0x2E) { // SWR
+  else if (opcode[i]==0x2E) { // SWR
     // Write entire word
     emit_writeword_indexed(tl,-3,temp);
   }
-  if (opcode[i]==0x2C) { // SDL
-    assert(0);
-  }
-  if (opcode[i]==0x2D) { // SDR
-    assert(0);
-  }
   set_jump_target(done0, out);
   set_jump_target(done1, out);
   set_jump_target(done2, out);
-  if (opcode[i]==0x2C) { // SDL
-    assert(0);
-  }
-  if (opcode[i]==0x2D) { // SDR
-    assert(0);
-  }
   if(!c||!memtarget)
     add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj[i],reglist);
-  if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
+  if(!(i_regs->waswritten&(1<<rs1[i])) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
     emit_addimm_no_flags(-ram_offset,temp);
     #if defined(HOST_IMM8)
     int ir=get_reg(i_regs->regmap,INVCP);
@@ -2964,7 +3197,7 @@ static void cop0_assemble(int i,struct regstat *i_regs)
         emit_storereg(CCREG,HOST_CCREG);
         emit_loadreg(rs1[i],1);
         emit_movimm(copr,0);
-        emit_call(pcsx_mtc0_ds);
+        emit_far_call(pcsx_mtc0_ds);
         emit_loadreg(rs1[i],s);
         return;
       }
@@ -2973,14 +3206,12 @@ static void cop0_assemble(int i,struct regstat *i_regs)
       emit_movimm(0,HOST_TEMPREG);
       emit_writeword(HOST_TEMPREG,&pending_exception);
     }
-    //else if(copr==12&&is_delayslot) emit_call((int)MTC0_R12);
-    //else
     if(s==HOST_CCREG)
       emit_loadreg(rs1[i],1);
     else if(s!=1)
       emit_mov(s,1);
     emit_movimm(copr,0);
-    emit_call(pcsx_mtc0);
+    emit_far_call(pcsx_mtc0);
     if(copr==9||copr==11||copr==12||copr==13) {
       emit_readword(&Count,HOST_CCREG);
       emit_readword(&next_interupt,HOST_TEMPREG);
@@ -2997,7 +3228,7 @@ static void cop0_assemble(int i,struct regstat *i_regs)
       emit_jeq(0);
       emit_readword(&pcaddr, 0);
       emit_addimm(HOST_CCREG,2,HOST_CCREG);
-      emit_call(get_addr_ht);
+      emit_far_call(get_addr_ht);
       emit_jmpreg(0);
       set_jump_target(jaddr, out);
     }
@@ -3057,7 +3288,131 @@ static void do_cop1stub(int n)
   if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
   emit_movimm(start+(i-ds)*4,EAX); // Get PC
   emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right?  There should probably be an extra cycle...
-  emit_jmp(ds?fp_exception_ds:fp_exception);
+  emit_far_jump(ds?fp_exception_ds:fp_exception);
+}
+
+static int cop2_is_stalling_op(int i, int *cycles)
+{
+  if (opcode[i] == 0x3a) { // SWC2
+    *cycles = 0;
+    return 1;
+  }
+  if (itype[i] == COP2 && (opcode2[i] == 0 || opcode2[i] == 2)) { // MFC2/CFC2
+    *cycles = 0;
+    return 1;
+  }
+  if (itype[i] == C2OP) {
+    *cycles = gte_cycletab[source[i] & 0x3f];
+    return 1;
+  }
+  // ... what about MTC2/CTC2/LWC2?
+  return 0;
+}
+
+#if 0
+static void log_gte_stall(int stall, u_int cycle)
+{
+  if ((u_int)stall <= 44)
+    printf("x    stall %2d %u\n", stall, cycle + last_count);
+ if (cycle + last_count > 1215348544) exit(1);
+}
+
+static void emit_log_gte_stall(int i, int stall, u_int reglist)
+{
+  save_regs(reglist);
+  if (stall > 0)
+    emit_movimm(stall, 0);
+  else
+    emit_mov(HOST_TEMPREG, 0);
+  emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
+  emit_far_call(log_gte_stall);
+  restore_regs(reglist);
+}
+#endif
+
+static void cop2_call_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
+{
+  int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
+  int rtmp = reglist_find_free(reglist);
+
+  if (HACK_ENABLED(NDHACK_GTE_NO_STALL))
+    return;
+  //assert(get_reg(i_regs->regmap, CCREG) == HOST_CCREG);
+  if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
+    // happens occasionally... cc evicted? Don't bother then
+    //printf("no cc %08x\n", start + i*4);
+    return;
+  }
+  if (!bt[i]) {
+    for (j = i - 1; j >= 0; j--) {
+      //if (is_ds[j]) break;
+      if (cop2_is_stalling_op(j, &other_gte_op_cycles) || bt[j])
+        break;
+    }
+  }
+  cycles_passed = CLOCK_ADJUST(ccadj[i] - ccadj[j]);
+  if (other_gte_op_cycles >= 0)
+    stall = other_gte_op_cycles - cycles_passed;
+  else if (cycles_passed >= 44)
+    stall = 0; // can't stall
+  if (stall == -MAXBLOCK && rtmp >= 0) {
+    // unknown stall, do the expensive runtime check
+    assem_debug("; cop2_call_stall_check\n");
+#if 0 // too slow
+    save_regs(reglist);
+    emit_movimm(gte_cycletab[op], 0);
+    emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
+    emit_far_call(call_gteStall);
+    restore_regs(reglist);
+#else
+    host_tempreg_acquire();
+    emit_readword(&psxRegs.gteBusyCycle, rtmp);
+    emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
+    emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
+    emit_cmpimm(HOST_TEMPREG, 44);
+    emit_cmovb_reg(rtmp, HOST_CCREG);
+    //emit_log_gte_stall(i, 0, reglist);
+    host_tempreg_release();
+#endif
+  }
+  else if (stall > 0) {
+    //emit_log_gte_stall(i, stall, reglist);
+    emit_addimm(HOST_CCREG, stall, HOST_CCREG);
+  }
+
+  // save gteBusyCycle, if needed
+  if (gte_cycletab[op] == 0)
+    return;
+  other_gte_op_cycles = -1;
+  for (j = i + 1; j < slen; j++) {
+    if (cop2_is_stalling_op(j, &other_gte_op_cycles))
+      break;
+    if (is_jump(j)) {
+      // check ds
+      if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
+        j++;
+      break;
+    }
+  }
+  if (other_gte_op_cycles >= 0)
+    // will handle stall when assembling that op
+    return;
+  cycles_passed = CLOCK_ADJUST(ccadj[min(j, slen -1)] - ccadj[i]);
+  if (cycles_passed >= 44)
+    return;
+  assem_debug("; save gteBusyCycle\n");
+  host_tempreg_acquire();
+#if 0
+  emit_readword(&last_count, HOST_TEMPREG);
+  emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
+  emit_addimm(HOST_TEMPREG, CLOCK_ADJUST(ccadj[i]), HOST_TEMPREG);
+  emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
+  emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
+#else
+  emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + gte_cycletab[op], HOST_TEMPREG);
+  emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
+#endif
+  host_tempreg_release();
 }
 
 static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
@@ -3089,22 +3444,7 @@ static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
       break;
     case 28:
     case 29:
-      emit_readword(&reg_cop2d[9],temp);
-      emit_testimm(temp,0x8000); // do we need this?
-      emit_andimm(temp,0xf80,temp);
-      emit_andne_imm(temp,0,temp);
-      emit_shrimm(temp,7,tl);
-      emit_readword(&reg_cop2d[10],temp);
-      emit_testimm(temp,0x8000);
-      emit_andimm(temp,0xf80,temp);
-      emit_andne_imm(temp,0,temp);
-      emit_orrshr_imm(temp,2,tl);
-      emit_readword(&reg_cop2d[11],temp);
-      emit_testimm(temp,0x8000);
-      emit_andimm(temp,0xf80,temp);
-      emit_andne_imm(temp,0,temp);
-      emit_orrshl_imm(temp,3,tl);
-      emit_writeword(tl,&reg_cop2d[copr]);
+      c2op_mfc2_29_assemble(tl,temp);
       break;
     default:
       emit_readword(&reg_cop2d[copr],tl);
@@ -3136,8 +3476,7 @@ static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
       emit_writeword(sl,&reg_cop2d[28]);
       break;
     case 30:
-      emit_movs(sl,temp);
-      emit_mvnmi(temp,temp);
+      emit_xorsar_imm(sl,sl,31,temp);
 #if defined(HAVE_ARMV5) || defined(__aarch64__)
       emit_clz(temp,temp);
 #else
@@ -3159,7 +3498,7 @@ static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
   }
 }
 
-static void c2ls_assemble(int i,struct regstat *i_regs)
+static void c2ls_assemble(int i, const struct regstat *i_regs)
 {
   int s,tl;
   int ar;
@@ -3169,7 +3508,7 @@ static void c2ls_assemble(int i,struct regstat *i_regs)
   enum stub_type type;
   int agr=AGEN1+(i&1);
   int fastio_reg_override=-1;
-  u_int hr,reglist=0;
+  u_int reglist=get_host_reglist(i_regs->regmap);
   u_int copr=(source[i]>>16)&0x1f;
   s=get_reg(i_regs->regmap,rs1[i]);
   tl=get_reg(i_regs->regmap,FTEMP);
@@ -3177,9 +3516,6 @@ static void c2ls_assemble(int i,struct regstat *i_regs)
   assert(rs1[i]>0);
   assert(tl>=0);
 
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
   if(i_regs->regmap[HOST_CCREG]==CCREG)
     reglist&=~(1<<HOST_CCREG);
 
@@ -3197,7 +3533,8 @@ static void c2ls_assemble(int i,struct regstat *i_regs)
   assert(ar>=0);
 
   if (opcode[i]==0x3a) { // SWC2
-    cop2_get_dreg(copr,tl,HOST_TEMPREG);
+    cop2_call_stall_check(0, i, i_regs, reglist_exclude(reglist, tl, -1));
+    cop2_get_dreg(copr,tl,-1);
     type=STOREW_STUB;
   }
   else
@@ -3235,7 +3572,7 @@ static void c2ls_assemble(int i,struct regstat *i_regs)
   if(jaddr2)
     add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj[i],reglist);
   if(opcode[i]==0x3a) // SWC2
-  if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
+  if(!(i_regs->waswritten&(1<<rs1[i])) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
 #if defined(HOST_IMM8)
     int ir=get_reg(i_regs->regmap,INVCP);
     assert(ir>=0);
@@ -3258,10 +3595,18 @@ static void c2ls_assemble(int i,struct regstat *i_regs)
   }
 }
 
-static void cop2_assemble(int i,struct regstat *i_regs)
+static void cop2_assemble(int i, const struct regstat *i_regs)
 {
-  u_int copr=(source[i]>>11)&0x1f;
-  signed char temp=get_reg(i_regs->regmap,-1);
+  u_int copr = (source[i]>>11) & 0x1f;
+  signed char temp = get_reg(i_regs->regmap, -1);
+
+  if (opcode2[i] == 0 || opcode2[i] == 2) { // MFC2/CFC2
+    if (!HACK_ENABLED(NDHACK_GTE_NO_STALL)) {
+      signed char tl = get_reg(i_regs->regmap, rt1[i]);
+      u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), tl, temp);
+      cop2_call_stall_check(0, i, i_regs, reglist);
+    }
+  }
   if (opcode2[i]==0) { // MFC2
     signed char tl=get_reg(i_regs->regmap,rt1[i]);
     if(tl>=0&&rt1[i]!=0)
@@ -3291,14 +3636,7 @@ static void cop2_assemble(int i,struct regstat *i_regs)
         emit_signextend16(sl,temp);
         break;
       case 31:
-        //value = value & 0x7ffff000;
-        //if (value & 0x7f87e000) value |= 0x80000000;
-        emit_shrimm(sl,12,temp);
-        emit_shlimm(temp,12,temp);
-        emit_testimm(temp,0x7f000000);
-        emit_testeqimm(temp,0x00870000);
-        emit_testeqimm(temp,0x0000e000);
-        emit_orrne_imm(temp,0x80000000,temp);
+        c2op_ctc2_31_assemble(sl,temp);
         break;
       default:
         temp=sl;
@@ -3309,6 +3647,90 @@ static void cop2_assemble(int i,struct regstat *i_regs)
   }
 }
 
+static void do_unalignedwritestub(int n)
+{
+  assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
+  literal_pool(256);
+  set_jump_target(stubs[n].addr, out);
+
+  int i=stubs[n].a;
+  struct regstat *i_regs=(struct regstat *)stubs[n].c;
+  int addr=stubs[n].b;
+  u_int reglist=stubs[n].e;
+  signed char *i_regmap=i_regs->regmap;
+  int temp2=get_reg(i_regmap,FTEMP);
+  int rt;
+  rt=get_reg(i_regmap,rs2[i]);
+  assert(rt>=0);
+  assert(addr>=0);
+  assert(opcode[i]==0x2a||opcode[i]==0x2e); // SWL/SWR only implemented
+  reglist|=(1<<addr);
+  reglist&=~(1<<temp2);
+
+#if 1
+  // don't bother with it and call write handler
+  save_regs(reglist);
+  pass_args(addr,rt);
+  int cc=get_reg(i_regmap,CCREG);
+  if(cc<0)
+    emit_loadreg(CCREG,2);
+  emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n].d+1),2);
+  emit_far_call((opcode[i]==0x2a?jump_handle_swl:jump_handle_swr));
+  emit_addimm(0,-CLOCK_ADJUST((int)stubs[n].d+1),cc<0?2:cc);
+  if(cc<0)
+    emit_storereg(CCREG,2);
+  restore_regs(reglist);
+  emit_jmp(stubs[n].retaddr); // return address
+#else
+  emit_andimm(addr,0xfffffffc,temp2);
+  emit_writeword(temp2,&address);
+
+  save_regs(reglist);
+  emit_shrimm(addr,16,1);
+  int cc=get_reg(i_regmap,CCREG);
+  if(cc<0) {
+    emit_loadreg(CCREG,2);
+  }
+  emit_movimm((u_int)readmem,0);
+  emit_addimm(cc<0?2:cc,2*stubs[n].d+2,2);
+  emit_call((int)&indirect_jump_indexed);
+  restore_regs(reglist);
+
+  emit_readword(&readmem_dword,temp2);
+  int temp=addr; //hmh
+  emit_shlimm(addr,3,temp);
+  emit_andimm(temp,24,temp);
+  if (opcode[i]==0x2a) // SWL
+    emit_xorimm(temp,24,temp);
+  emit_movimm(-1,HOST_TEMPREG);
+  if (opcode[i]==0x2a) { // SWL
+    emit_bic_lsr(temp2,HOST_TEMPREG,temp,temp2);
+    emit_orrshr(rt,temp,temp2);
+  }else{
+    emit_bic_lsl(temp2,HOST_TEMPREG,temp,temp2);
+    emit_orrshl(rt,temp,temp2);
+  }
+  emit_readword(&address,addr);
+  emit_writeword(temp2,&word);
+  //save_regs(reglist); // don't need to, no state changes
+  emit_shrimm(addr,16,1);
+  emit_movimm((u_int)writemem,0);
+  //emit_call((int)&indirect_jump_indexed);
+  emit_mov(15,14);
+  emit_readword_dualindexedx4(0,1,15);
+  emit_readword(&Count,HOST_TEMPREG);
+  emit_readword(&next_interupt,2);
+  emit_addimm(HOST_TEMPREG,-2*stubs[n].d-2,HOST_TEMPREG);
+  emit_writeword(2,&last_count);
+  emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
+  if(cc<0) {
+    emit_storereg(CCREG,HOST_TEMPREG);
+  }
+  restore_regs(reglist);
+  emit_jmp(stubs[n].retaddr); // return address
+#endif
+}
+
 #ifndef multdiv_assemble
 void multdiv_assemble(int i,struct regstat *i_regs)
 {
@@ -3333,43 +3755,44 @@ static void mov_assemble(int i,struct regstat *i_regs)
   }
 }
 
-static void syscall_assemble(int i,struct regstat *i_regs)
+// call interpreter, exception handler, things that change pc/regs/cycles ...
+static void call_c_cpu_handler(int i, const struct regstat *i_regs, u_int pc, void *func)
 {
   signed char ccreg=get_reg(i_regs->regmap,CCREG);
   assert(ccreg==HOST_CCREG);
   assert(!is_delayslot);
   (void)ccreg;
-  emit_movimm(start+i*4,EAX); // Get PC
-  emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right?  There should probably be an extra cycle...
-  emit_jmp(jump_syscall_hle); // XXX
+
+  emit_movimm(pc,3); // Get PC
+  emit_readword(&last_count,2);
+  emit_writeword(3,&psxRegs.pc);
+  emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
+  emit_add(2,HOST_CCREG,2);
+  emit_writeword(2,&psxRegs.cycle);
+  emit_far_call(func);
+  emit_far_jump(jump_to_new_pc);
+}
+
+static void syscall_assemble(int i,struct regstat *i_regs)
+{
+  emit_movimm(0x20,0); // cause code
+  emit_movimm(0,1);    // not in delay slot
+  call_c_cpu_handler(i,i_regs,start+i*4,psxException);
 }
 
 static void hlecall_assemble(int i,struct regstat *i_regs)
 {
-  extern void psxNULL();
-  signed char ccreg=get_reg(i_regs->regmap,CCREG);
-  assert(ccreg==HOST_CCREG);
-  assert(!is_delayslot);
-  (void)ccreg;
-  emit_movimm(start+i*4+4,0); // Get PC
+  void *hlefunc = psxNULL;
   uint32_t hleCode = source[i] & 0x03ffffff;
-  if (hleCode >= ARRAY_SIZE(psxHLEt))
-    emit_movimm((uintptr_t)psxNULL,1);
-  else
-    emit_movimm((uintptr_t)psxHLEt[hleCode],1);
-  emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
-  emit_jmp(jump_hlecall);
+  if (hleCode < ARRAY_SIZE(psxHLEt))
+    hlefunc = psxHLEt[hleCode];
+
+  call_c_cpu_handler(i,i_regs,start+i*4+4,hlefunc);
 }
 
 static void intcall_assemble(int i,struct regstat *i_regs)
 {
-  signed char ccreg=get_reg(i_regs->regmap,CCREG);
-  assert(ccreg==HOST_CCREG);
-  assert(!is_delayslot);
-  (void)ccreg;
-  emit_movimm(start+i*4,0); // Get PC
-  emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
-  emit_jmp(jump_intcall);
+  call_c_cpu_handler(i,i_regs,start+i*4,execI);
 }
 
 static void speculate_mov(int rs,int rt)
@@ -4075,19 +4498,34 @@ static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
 static void drc_dbg_emit_do_cmp(int i)
 {
   extern void do_insn_cmp();
-  extern int cycle;
-  u_int hr,reglist=0;
+  //extern int cycle;
+  u_int hr, reglist = get_host_reglist(regs[i].regmap);
 
-  for(hr=0;hr<HOST_REGS;hr++)
-    if(regs[i].regmap[hr]>=0) reglist|=1<<hr;
+  assem_debug("//do_insn_cmp %08x\n", start+i*4);
   save_regs(reglist);
+  // write out changed consts to match the interpreter
+  if (i > 0 && !bt[i]) {
+    for (hr = 0; hr < HOST_REGS; hr++) {
+      int reg = regs[i-1].regmap[hr];
+      if (hr == EXCLUDE_REG || reg < 0)
+        continue;
+      if (!((regs[i-1].isconst >> hr) & 1))
+        continue;
+      if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
+        continue;
+      emit_movimm(constmap[i-1][hr],0);
+      emit_storereg(reg, 0);
+    }
+  }
   emit_movimm(start+i*4,0);
   emit_writeword(0,&pcaddr);
-  emit_call(do_insn_cmp);
+  emit_far_call(do_insn_cmp);
   //emit_readword(&cycle,0);
   //emit_addimm(0,2,0);
   //emit_writeword(0,&cycle);
+  (void)get_reg2;
   restore_regs(reglist);
+  assem_debug("\\\\do_insn_cmp\n");
 }
 #else
 #define drc_dbg_emit_do_cmp(x)
@@ -4214,11 +4652,13 @@ void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
   else if(*adj==0||invert) {
     int cycles=CLOCK_ADJUST(count+2);
     // faster loop HACK
+#if 0
     if (t&&*adj) {
       int rel=t-i;
       if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
         cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
     }
+#endif
     emit_addimm_and_set_flags(cycles,HOST_CCREG);
     jaddr=out;
     emit_jns(0);
@@ -4413,7 +4853,7 @@ static void do_ccstub(int n)
   // Update cycle count
   assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
   if(stubs[n].a) emit_addimm(HOST_CCREG,CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
-  emit_call(cc_interrupt);
+  emit_far_call(cc_interrupt);
   if(stubs[n].a) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
   if(stubs[n].d==TAKEN) {
     if(internal_branch(ba[i]))
@@ -4667,6 +5107,9 @@ static void cjump_assemble(int i,struct regstat *i_regs)
   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
   if(i>(ba[i]-start)>>2) invert=1;
   #endif
+  #ifdef __aarch64__
+  invert=1; // because of near cond. branches
+  #endif
 
   if(ooo[i]) {
     s1l=get_reg(branch_regs[i].regmap,rs1[i]);
@@ -4956,6 +5399,9 @@ static void sjump_assemble(int i,struct regstat *i_regs)
   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
   if(i>(ba[i]-start)>>2) invert=1;
   #endif
+  #ifdef __aarch64__
+  invert=1; // because of near cond. branches
+  #endif
 
   //if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
   //assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
@@ -5519,7 +5965,7 @@ void unneeded_registers(int istart,int iend,int r)
   uint64_t u,gte_u,b,gte_b;
   uint64_t temp_u,temp_gte_u=0;
   uint64_t gte_u_unknown=0;
-  if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
+  if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
     gte_u_unknown=~0ll;
   if(iend==slen-1) {
     u=1;
@@ -5570,7 +6016,7 @@ void unneeded_registers(int istart,int iend,int r)
         bt[(ba[i]-start)>>2]=1;
         if(ba[i]<=start+i*4) {
           // Backward branch
-          if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
+          if(is_ujump(i))
           {
             // Unconditional branch
             temp_u=1;
@@ -5615,7 +6061,7 @@ void unneeded_registers(int istart,int iend,int r)
             gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
           }
         } /*else*/ if(1) {
-          if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
+          if (is_ujump(i))
           {
             // Unconditional branch
             u=unneeded_reg[(ba[i]-start)>>2];
@@ -5725,7 +6171,7 @@ void clean_registers(int istart,int iend,int wr)
       if(ba[i]<start || ba[i]>=(start+slen*4))
       {
         // Branch out of this block, flush all regs
-        if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
+        if (is_ujump(i))
         {
           // Unconditional branch
           will_dirty_i=0;
@@ -5805,7 +6251,7 @@ void clean_registers(int istart,int iend,int wr)
         // Internal branch
         if(ba[i]<=start+i*4) {
           // Backward branch
-          if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
+          if (is_ujump(i))
           {
             // Unconditional branch
             temp_will_dirty=0;
@@ -5902,7 +6348,7 @@ void clean_registers(int istart,int iend,int wr)
         }
         /*else*/ if(1)
         {
-          if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
+          if (is_ujump(i))
           {
             // Unconditional branch
             will_dirty_i=0;
@@ -6069,7 +6515,7 @@ void clean_registers(int istart,int iend,int wr)
         regs[i].dirty&=wont_dirty_i;
         if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
         {
-          if(i<iend-1&&itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
+          if (i < iend-1 && !is_ujump(i)) {
             for(r=0;r<HOST_REGS;r++) {
               if(r!=EXCLUDE_REG) {
                 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
@@ -6247,7 +6693,7 @@ static void new_dynarec_test(void)
   ((volatile u_int *)out)[0]++; // make cache dirty
 
   for (i = 0; i < ARRAY_SIZE(ret); i++) {
-    out = translation_cache;
+    out = ndrc->translation_cache;
     beginning = start_block();
     emit_movimm(DRC_TEST_VAL + i, 0); // test
     emit_ret();
@@ -6261,15 +6707,15 @@ static void new_dynarec_test(void)
     SysPrintf("test passed.\n");
   else
     SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
-  out = translation_cache;
+  out = ndrc->translation_cache;
 }
 
 // clear the state completely, instead of just marking
 // things invalid like invalidate_all_pages() does
-void new_dynarec_clear_full()
+void new_dynarec_clear_full(void)
 {
   int n;
-  out = translation_cache;
+  out = ndrc->translation_cache;
   memset(invalid_code,1,sizeof(invalid_code));
   memset(hash_table,0xff,sizeof(hash_table));
   memset(mini_ht,-1,sizeof(mini_ht));
@@ -6287,34 +6733,28 @@ void new_dynarec_clear_full()
   for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
 }
 
-void new_dynarec_init()
+void new_dynarec_init(void)
 {
   SysPrintf("Init new dynarec\n");
 
-  // allocate/prepare a buffer for translation cache
-  // see assem_arm.h for some explanation
-#if   defined(BASE_ADDR_FIXED)
-  if (mmap(translation_cache, 1 << TARGET_SIZE_2,
-            PROT_READ | PROT_WRITE | PROT_EXEC,
-            MAP_PRIVATE | MAP_ANONYMOUS,
-            -1, 0) != translation_cache) {
-    SysPrintf("mmap() failed: %s\n", strerror(errno));
-    SysPrintf("disable BASE_ADDR_FIXED and recompile\n");
-    abort();
-  }
-#elif defined(BASE_ADDR_DYNAMIC)
+#ifdef BASE_ADDR_DYNAMIC
   #ifdef VITA
   sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
   if (sceBlock < 0)
     SysPrintf("sceKernelAllocMemBlockForVM failed\n");
-  int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&translation_cache);
+  int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
   if (ret < 0)
     SysPrintf("sceKernelGetMemBlockBase failed\n");
   #else
-  translation_cache = mmap (NULL, 1 << TARGET_SIZE_2,
+  uintptr_t desired_addr = 0;
+  #ifdef __ELF__
+  extern char _end;
+  desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
+  #endif
+  ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
             PROT_READ | PROT_WRITE | PROT_EXEC,
             MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
-  if (translation_cache == MAP_FAILED) {
+  if (ndrc == MAP_FAILED) {
     SysPrintf("mmap() failed: %s\n", strerror(errno));
     abort();
   }
@@ -6322,11 +6762,12 @@ void new_dynarec_init()
 #else
   #ifndef NO_WRITE_EXEC
   // not all systems allow execute in data segment by default
-  if (mprotect(translation_cache, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
+  if (mprotect(ndrc, sizeof(ndrc->translation_cache) + sizeof(ndrc->tramp.ops),
+               PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
     SysPrintf("mprotect() failed: %s\n", strerror(errno));
   #endif
 #endif
-  out = translation_cache;
+  out = ndrc->translation_cache;
   cycle_multiplier=200;
   new_dynarec_clear_full();
 #ifdef HOST_IMM8
@@ -6342,15 +6783,15 @@ void new_dynarec_init()
     SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
 }
 
-void new_dynarec_cleanup()
+void new_dynarec_cleanup(void)
 {
   int n;
-#if defined(BASE_ADDR_FIXED) || defined(BASE_ADDR_DYNAMIC)
+#ifdef BASE_ADDR_DYNAMIC
   #ifdef VITA
   sceKernelFreeMemBlock(sceBlock);
   sceBlock = -1;
   #else
-  if (munmap(translation_cache, 1<<TARGET_SIZE_2) < 0)
+  if (munmap(ndrc, sizeof(*ndrc)) < 0)
     SysPrintf("munmap() failed\n");
   #endif
 #endif
@@ -6364,16 +6805,25 @@ void new_dynarec_cleanup()
 
 static u_int *get_source_start(u_int addr, u_int *limit)
 {
+  if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
+    cycle_multiplier_override = 0;
+
   if (addr < 0x00200000 ||
-    (0xa0000000 <= addr && addr < 0xa0200000)) {
+    (0xa0000000 <= addr && addr < 0xa0200000))
+  {
     // used for BIOS calls mostly?
     *limit = (addr&0xa0000000)|0x00200000;
     return (u_int *)(rdram + (addr&0x1fffff));
   }
   else if (!Config.HLE && (
     /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
-    (0xbfc00000 <= addr && addr < 0xbfc80000))) {
-    // BIOS
+    (0xbfc00000 <= addr && addr < 0xbfc80000)))
+  {
+    // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
+    // but timings in PCSX are too tied to the interpreter's BIAS
+    if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
+      cycle_multiplier_override = 200;
+
     *limit = (addr & 0xfff00000) | 0x80000;
     return (u_int *)((u_char *)psxR + (addr&0x7ffff));
   }
@@ -6484,7 +6934,7 @@ void new_dynarec_load_blocks(const void *save, int size)
   memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
 }
 
-int new_recompile_block(int addr)
+int new_recompile_block(u_int addr)
 {
   u_int pagelimit = 0;
   u_int state_rflags = 0;
@@ -6513,7 +6963,7 @@ int new_recompile_block(int addr)
     invalid_code[start>>12]=0;
     emit_movimm(start,0);
     emit_writeword(0,&pcaddr);
-    emit_jmp(new_dyna_leave);
+    emit_far_jump(new_dyna_leave);
     literal_pool(0);
     end_block(beginning);
     ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
@@ -6956,7 +7406,7 @@ int new_recompile_block(int addr)
     else if(type==CJUMP||type==SJUMP)
       ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
     else ba[i]=-1;
-    if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP)) {
+    if (i > 0 && is_jump(i-1)) {
       int do_in_intrp=0;
       // branch in delay slot?
       if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP) {
@@ -6974,7 +7424,7 @@ int new_recompile_block(int addr)
           bt[t+1]=1; // expected return from interpreter
         }
         else if(i>=2&&rt1[i-2]==2&&rt1[i]==2&&rs1[i]!=2&&rs2[i]!=2&&rs1[i-1]!=2&&rs2[i-1]!=2&&
-              !(i>=3&&(itype[i-3]==RJUMP||itype[i-3]==UJUMP||itype[i-3]==CJUMP||itype[i-3]==SJUMP))) {
+              !(i>=3&&is_jump(i-3))) {
           // v0 overwrite like this is a sign of trouble, bail out
           SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
           do_in_intrp=1;
@@ -6990,7 +7440,7 @@ int new_recompile_block(int addr)
       }
     }
     /* Is this the end of the block? */
-    if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
+    if (i > 0 && is_ujump(i-1)) {
       if(rt1[i-1]==0) { // Continue past subroutine call (JAL)
         done=2;
       }
@@ -7357,8 +7807,9 @@ int new_recompile_block(int addr)
           cop0_alloc(&current,i);
           break;
         case COP1:
+          break;
         case COP2:
-          cop12_alloc(&current,i);
+          cop2_alloc(&current,i);
           break;
         case C1LS:
           c1ls_alloc(&current,i);
@@ -7451,7 +7902,7 @@ int new_recompile_block(int addr)
             dirty_reg(&branch_regs[i-1],31);
           }
           memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
-          memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+          memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
           break;
         case RJUMP:
           memcpy(&branch_regs[i-1],&current,sizeof(current));
@@ -7472,7 +7923,7 @@ int new_recompile_block(int addr)
           }
           #endif
           memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
-          memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+          memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
           break;
         case CJUMP:
           if((opcode[i-1]&0x3E)==4) // BEQ/BNE
@@ -7499,7 +7950,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].isconst=0;
             branch_regs[i-1].wasconst=0;
             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
-            memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+            memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
           }
           else
           if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
@@ -7524,7 +7975,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].isconst=0;
             branch_regs[i-1].wasconst=0;
             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
-            memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+            memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
           }
           else
           // Alloc the delay slot in case the branch is taken
@@ -7578,7 +8029,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].isconst=0;
             branch_regs[i-1].wasconst=0;
             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
-            memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+            memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
           }
           else
           // Alloc the delay slot in case the branch is taken
@@ -7602,7 +8053,7 @@ int new_recompile_block(int addr)
           break;
       }
 
-      if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
+      if (is_ujump(i-1))
       {
         if(rt1[i-1]==31) // JAL/JALR
         {
@@ -7651,12 +8102,10 @@ int new_recompile_block(int addr)
 #if !defined(DRC_DBG)
     else if(itype[i]==C2OP&&gte_cycletab[source[i]&0x3f]>2)
     {
-      // GTE runs in parallel until accessed, divide by 2 for a rough guess
-      cc+=gte_cycletab[source[i]&0x3f]/2;
-    }
-    else if(/*itype[i]==LOAD||itype[i]==STORE||*/itype[i]==C1LS) // load,store causes weird timing issues
-    {
-      cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
+      // this should really be removed since the real stalls have been implemented,
+      // but doing so causes sizeable perf regression against the older version
+      u_int gtec = gte_cycletab[source[i] & 0x3f];
+      cc += HACK_ENABLED(NDHACK_GTE_NO_STALL) ? gtec/2 : 2;
     }
     else if(i>1&&itype[i]==STORE&&itype[i-1]==STORE&&itype[i-2]==STORE&&!bt[i])
     {
@@ -7664,7 +8113,8 @@ int new_recompile_block(int addr)
     }
     else if(itype[i]==C2LS)
     {
-      cc+=4;
+      // same as with C2OP
+      cc += HACK_ENABLED(NDHACK_GTE_NO_STALL) ? 4 : 2;
     }
 #endif
     else
@@ -7675,7 +8125,7 @@ int new_recompile_block(int addr)
     if(!is_ds[i]) {
       regs[i].dirty=current.dirty;
       regs[i].isconst=current.isconst;
-      memcpy(constmap[i],current_constmap,sizeof(current_constmap));
+      memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
     }
     for(hr=0;hr<HOST_REGS;hr++) {
       if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
@@ -7716,7 +8166,7 @@ int new_recompile_block(int addr)
         }
       }
       // Conditional branch may need registers for following instructions
-      if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
+      if (!is_ujump(i))
       {
         if(i<slen-2) {
           nr|=needed_reg[i+2];
@@ -7817,7 +8267,7 @@ int new_recompile_block(int addr)
            (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
            (regs[i].regmap[hr]&63)!=PTEMP && (regs[i].regmap[hr]&63)!=CCREG)
         {
-          if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
+          if (!is_ujump(i))
           {
             if(likely[i]) {
               regs[i].regmap[hr]=-1;
@@ -7861,7 +8311,7 @@ int new_recompile_block(int addr)
             {
               branch_regs[i].regmap[hr]=-1;
               branch_regs[i].regmap_entry[hr]=-1;
-              if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
+              if (!is_ujump(i))
               {
                 if(!likely[i]&&i<slen-2) {
                   regmap_pre[i+2][hr]=-1;
@@ -7891,7 +8341,7 @@ int new_recompile_block(int addr)
             {
               if(i<slen-1&&!is_ds[i]) {
                 assert(regs[i].regmap[hr]<64);
-                if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]!=-1)
+                if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
                 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
                 {
                   SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
@@ -7906,8 +8356,8 @@ int new_recompile_block(int addr)
             }
           }
         }
-      }
-    }
+      } // if needed
+    } // for hr
   }
 
   /* Pass 5 - Pre-allocate registers */
@@ -8052,7 +8502,7 @@ int new_recompile_block(int addr)
                       branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
                       branch_regs[i].wasconst&=~(1<<hr);
                       branch_regs[i].isconst&=~(1<<hr);
-                      if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
+                      if (!is_ujump(i)) {
                         regmap_pre[i+2][hr]=f_regmap[hr];
                         regs[i+2].wasdirty&=~(1<<hr);
                         regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
@@ -8073,7 +8523,7 @@ int new_recompile_block(int addr)
                       branch_regs[k].dirty&=~(1<<hr);
                       branch_regs[k].wasconst&=~(1<<hr);
                       branch_regs[k].isconst&=~(1<<hr);
-                      if(itype[k]!=RJUMP&&itype[k]!=UJUMP&&(source[k]>>16)!=0x1000) {
+                      if (!is_ujump(k)) {
                         regmap_pre[k+2][hr]=f_regmap[hr];
                         regs[k+2].wasdirty&=~(1<<hr);
                       }
@@ -8095,7 +8545,7 @@ int new_recompile_block(int addr)
                   //printf("no-match due to different register\n");
                   break;
                 }
-                if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
+                if (is_ujump(j))
                 {
                   // Stop on unconditional branch
                   break;
@@ -8533,7 +8983,7 @@ int new_recompile_block(int addr)
   void *instr_addr0_override = NULL;
 
   if (start == 0x80030000) {
-    // nasty hack for fastbios thing
+    // nasty hack for the fastbios thing
     // override block entry to this code
     instr_addr0_override = out;
     emit_movimm(start,0);
@@ -8543,7 +8993,12 @@ int new_recompile_block(int addr)
     emit_writeword(0,&pcaddr);
     emit_writeword(0,&address);
     emit_cmp(0,1);
+    #ifdef __aarch64__
+    emit_jeq(out + 4*2);
+    emit_far_jump(new_dyna_leave);
+    #else
     emit_jne(new_dyna_leave);
+    #endif
   }
   for(i=0;i<slen;i++)
   {
@@ -8556,7 +9011,7 @@ int new_recompile_block(int addr)
     } else {
       speculate_register_values(i);
       #ifndef DESTRUCTIVE_WRITEBACK
-      if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
+      if (i < 2 || !is_ujump(i-2))
       {
         wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
       }
@@ -8567,7 +9022,7 @@ int new_recompile_block(int addr)
       }
       #endif
       // write back
-      if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
+      if (i < 2 || !is_ujump(i-2))
       {
         wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
         loop_preload(regmap_pre[i],regs[i].regmap_entry);
@@ -8659,17 +9114,17 @@ int new_recompile_block(int addr)
         case SPAN:
           pagespan_assemble(i,&regs[i]);break;
       }
-      if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
+      if (is_ujump(i))
         literal_pool(1024);
       else
         literal_pool_jumpover(256);
     }
   }
-  //assert(itype[i-2]==UJUMP||itype[i-2]==RJUMP||(source[i-2]>>16)==0x1000);
+  //assert(is_ujump(i-2));
   // If the block did not end with an unconditional branch,
   // add a jump to the next instruction.
   if(i>1) {
-    if(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000&&itype[i-1]!=SPAN) {
+    if(!is_ujump(i-2)&&itype[i-1]!=SPAN) {
       assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP);
       assert(i==slen);
       if(itype[i-2]!=CJUMP&&itype[i-2]!=SJUMP) {
@@ -8812,8 +9267,8 @@ int new_recompile_block(int addr)
 
   // If we're within 256K of the end of the buffer,
   // start over from the beginning. (Is 256K enough?)
-  if (out > translation_cache+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE)
-    out = translation_cache;
+  if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
+    out = ndrc->translation_cache;
 
   // Trap writes to any of the pages we compiled
   for(i=start>>12;i<=(start+slen*4)>>12;i++) {
@@ -8830,11 +9285,11 @@ int new_recompile_block(int addr)
 
   /* Pass 10 - Free memory by expiring oldest blocks */
 
-  int end=(((out-translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
+  int end=(((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
   while(expirep!=end)
   {
     int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
-    uintptr_t base=(uintptr_t)translation_cache+((expirep>>13)<<shift); // Base address of this block
+    uintptr_t base=(uintptr_t)ndrc->translation_cache+((expirep>>13)<<shift); // Base address of this block
     inv_debug("EXP: Phase %d\n",expirep);
     switch((expirep>>11)&3)
     {
@@ -8872,10 +9327,8 @@ int new_recompile_block(int addr)
         break;
       case 3:
         // Clear jump_out
-        #if defined(__arm__) || defined(__aarch64__)
         if((expirep&2047)==0)
           do_clear_cache();
-        #endif
         ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
         ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
         break;