X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fassem_arm.c;h=c2f65ee6df827e63548d1b12d155eb02e54fd805;hp=93dbf922681f7ceaaf271c08307149df6fb957e0;hb=71e490c5930e6e5f71d1f2d5165c3a801ac46be1;hpb=2573466affd06edebd68070bf167a62ca764f2b4

diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c
index 93dbf922..c2f65ee6 100644
--- a/libpcsxcore/new_dynarec/assem_arm.c
+++ b/libpcsxcore/new_dynarec/assem_arm.c
@@ -19,15 +19,31 @@
  *   51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.          *
  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
 
+#include "../gte.h"
+#define FLAGLESS
+#include "../gte.h"
+#undef FLAGLESS
+#include "../gte_arm.h"
+#include "../gte_neon.h"
+#include "pcnt.h"
+#include "arm_features.h"
+
+#if !BASE_ADDR_FIXED
+char translation_cache[1 << TARGET_SIZE_2] __attribute__((aligned(4096)));
+#endif
+
+#ifndef __MACH__
+#define CALLER_SAVE_REGS 0x100f
+#else
+#define CALLER_SAVE_REGS 0x120f
+#endif
+
 extern int cycle_count;
 extern int last_count;
 extern int pcaddr;
 extern int pending_exception;
 extern int branch_target;
 extern uint64_t readmem_dword;
-#ifdef MUPEN64
-extern precomp_instr fake_pc;
-#endif
 extern void *dynarec_local;
 extern u_int memory_map[1048576];
 extern u_int mini_ht[32][2];
@@ -98,8 +114,6 @@ const u_int invalidate_addr_reg[16] = {
   0,
   0};
 
-#include "fpu.h"
-
 unsigned int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
 
 /* Linker */
@@ -209,7 +223,7 @@ int get_pointer(void *stub)
 u_int get_clean_addr(int addr)
 {
   int *ptr=(int *)addr;
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   ptr+=4;
   #else
   ptr+=6;
@@ -226,7 +240,7 @@ u_int get_clean_addr(int addr)
 int verify_dirty(int addr)
 {
   u_int *ptr=(u_int *)addr;
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   // get from literal pool
   assert((*ptr&0xFFFF0000)==0xe59f0000);
   u_int offset=*ptr&0xfff;
@@ -245,18 +259,6 @@ int verify_dirty(int addr)
   #endif
   if((*ptr&0xFF000000)!=0xeb000000) ptr++;
   assert((*ptr&0xFF000000)==0xeb000000); // bl instruction
-#ifndef DISABLE_TLB
-  u_int verifier=(int)ptr+((signed int)(*ptr<<8)>>6)+8; // get target of bl
-  if(verifier==(u_int)verify_code_vm||verifier==(u_int)verify_code_ds) {
-    unsigned int page=source>>12;
-    unsigned int map_value=memory_map[page];
-    if(map_value>=0x80000000) return 0;
-    while(page<((source+len-1)>>12)) {
-      if((memory_map[++page]<<2)!=(map_value<<2)) return 0;
-    }
-    source = source+(map_value<<2);
-  }
-#endif
   //printf("verify_dirty: %x %x %x\n",source,copy,len);
   return !memcmp((void *)source,(void *)copy,len);
 }
@@ -265,7 +267,7 @@ int verify_dirty(int addr)
 // guarantees that it's not dirty
 int isclean(int addr)
 {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   int *ptr=((u_int *)addr)+4;
   #else
   int *ptr=((u_int *)addr)+6;
@@ -278,10 +280,11 @@ int isclean(int addr)
   return 1;
 }
 
+// get source that block at addr was compiled from (host pointers)
 void get_bounds(int addr,u_int *start,u_int *end)
 {
   u_int *ptr=(u_int *)addr;
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   // get from literal pool
   assert((*ptr&0xFFFF0000)==0xe59f0000);
   u_int offset=*ptr&0xfff;
@@ -300,13 +303,6 @@ void get_bounds(int addr,u_int *start,u_int *end)
   #endif
   if((*ptr&0xFF000000)!=0xeb000000) ptr++;
   assert((*ptr&0xFF000000)==0xeb000000); // bl instruction
-#ifndef DISABLE_TLB
-  u_int verifier=(int)ptr+((signed int)(*ptr<<8)>>6)+8; // get target of bl
-  if(verifier==(u_int)verify_code_vm||verifier==(u_int)verify_code_ds) {
-    if(memory_map[source>>12]>=0x80000000) source = 0;
-    else source = source+(memory_map[source>>12]<<2);
-  }
-#endif
   *start=source;
   *end=source+len;
 }
@@ -477,7 +473,7 @@ void alloc_reg(struct regstat *cur,int i,signed char reg)
       }
     }
   }
-  printf("This shouldn't happen (alloc_reg)");exit(1);
+  SysPrintf("This shouldn't happen (alloc_reg)");exit(1);
 }
 
 void alloc_reg64(struct regstat *cur,int i,signed char reg)
@@ -643,7 +639,7 @@ void alloc_reg64(struct regstat *cur,int i,signed char reg)
       }
     }
   }
-  printf("This shouldn't happen");exit(1);
+  SysPrintf("This shouldn't happen");exit(1);
 }
 
 // Allocate a temporary register.  This is done without regard to
@@ -766,7 +762,7 @@ void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
       }
     }
   }
-  printf("This shouldn't happen");exit(1);
+  SysPrintf("This shouldn't happen");exit(1);
 }
 // Allocate a specific ARM register.
 void alloc_arm_reg(struct regstat *cur,int i,signed char reg,char hr)
@@ -883,7 +879,7 @@ u_int genjmp(u_int addr)
   int offset=addr-(int)out-8;
   if(offset<-33554432||offset>=33554432) {
     if (addr>2) {
-      printf("genjmp: out of range: %08x\n", offset);
+      SysPrintf("genjmp: out of range: %08x\n", offset);
       exit(1);
     }
     return 0;
@@ -990,7 +986,7 @@ void emit_movimm(u_int imm,u_int rt)
     assem_debug("mvn %s,#%d\n",regname[rt],imm);
     output_w32(0xe3e00000|rd_rn_rm(rt,0,0)|armval);
   }else if(imm<65536) {
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     assem_debug("mov %s,#%d\n",regname[rt],imm&0xFF00);
     output_w32(0xe3a00000|rd_rn_imm_shift(rt,0,imm>>8,8));
     assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF);
@@ -999,7 +995,7 @@ void emit_movimm(u_int imm,u_int rt)
     emit_movw(imm,rt);
     #endif
   }else{
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     emit_loadlp(imm,rt);
     #else
     emit_movw(imm&0x0000FFFF,rt);
@@ -1015,13 +1011,11 @@ void emit_pcreladdr(u_int rt)
 
 void emit_loadreg(int r, int hr)
 {
-#ifdef FORCE32
   if(r&64) {
-    printf("64bit load in 32bit mode!\n");
+    SysPrintf("64bit load in 32bit mode!\n");
     assert(0);
     return;
   }
-#endif
   if((r&63)==0)
     emit_zeroreg(hr);
   else {
@@ -1040,13 +1034,11 @@ void emit_loadreg(int r, int hr)
 }
 void emit_storereg(int r, int hr)
 {
-#ifdef FORCE32
   if(r&64) {
-    printf("64bit store in 32bit mode!\n");
+    SysPrintf("64bit store in 32bit mode!\n");
     assert(0);
     return;
   }
-#endif
   int addr=((int)reg)+((r&63)<<REG_SHIFT)+((r&64)>>4);
   if((r&63)==HIREG) addr=(int)&hi+((r&64)>>4);
   if((r&63)==LOREG) addr=(int)&lo+((r&64)>>4);
@@ -1143,20 +1135,32 @@ void emit_addimm(u_int rs,int imm,u_int rt)
       assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],imm);
       output_w32(0xe2800000|rd_rn_rm(rt,rs,0)|armval);
     }else if(genimm(-imm,&armval)) {
-      assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],imm);
+      assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],-imm);
       output_w32(0xe2400000|rd_rn_rm(rt,rs,0)|armval);
-    }else if(imm<0) {
-      assert(imm>-65536);
+    #ifdef HAVE_ARMV7
+    }else if(rt!=rs&&(u_int)imm<65536) {
+      emit_movw(imm&0x0000ffff,rt);
+      emit_add(rs,rt,rt);
+    }else if(rt!=rs&&(u_int)-imm<65536) {
+      emit_movw(-imm&0x0000ffff,rt);
+      emit_sub(rs,rt,rt);
+    #endif
+    }else if((u_int)-imm<65536) {
       assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],(-imm)&0xFF00);
       assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rt],(-imm)&0xFF);
       output_w32(0xe2400000|rd_rn_imm_shift(rt,rs,(-imm)>>8,8));
       output_w32(0xe2400000|rd_rn_imm_shift(rt,rt,(-imm)&0xff,0));
-    }else{
-      assert(imm<65536);
-      assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF00);
-      assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF);
-      output_w32(0xe2800000|rd_rn_imm_shift(rt,rs,imm>>8,8));
-      output_w32(0xe2800000|rd_rn_imm_shift(rt,rt,imm&0xff,0));
+    }else {
+      do {
+        int shift = (ffs(imm) - 1) & ~1;
+        int imm8 = imm & (0xff << shift);
+        genimm_checked(imm8,&armval);
+        assem_debug("add %s,%s,#0x%x\n",regname[rt],regname[rs],imm8);
+        output_w32(0xe2800000|rd_rn_rm(rt,rs,0)|armval);
+        rs = rt;
+        imm &= ~imm8;
+      }
+      while (imm != 0);
     }
   }
   else if(rs!=rt) emit_mov(rs,rt);
@@ -1263,7 +1267,7 @@ void emit_andimm(int rs,int imm,int rt)
     assem_debug("bic %s,%s,#%d\n",regname[rt],regname[rs],imm);
     output_w32(0xe3c00000|rd_rn_rm(rt,rs,0)|armval);
   }else if(imm==65535) {
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV6
     assem_debug("bic %s,%s,#FF000000\n",regname[rt],regname[rs]);
     output_w32(0xe3c00000|rd_rn_rm(rt,rs,0)|0x4FF);
     assem_debug("bic %s,%s,#00FF0000\n",regname[rt],regname[rt]);
@@ -1274,7 +1278,7 @@ void emit_andimm(int rs,int imm,int rt)
     #endif
   }else{
     assert(imm>0&&imm<65535);
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     assem_debug("mov r14,#%d\n",imm&0xFF00);
     output_w32(0xe3a00000|rd_rn_imm_shift(HOST_TEMPREG,0,imm>>8,8));
     assem_debug("add r14,r14,#%d\n",imm&0xFF);
@@ -1338,6 +1342,14 @@ void emit_lsls_imm(int rs,int imm,int rt)
   output_w32(0xe1b00000|rd_rn_rm(rt,0,rs)|(imm<<7));
 }
 
+void emit_lslpls_imm(int rs,int imm,int rt)
+{
+  assert(imm>0);
+  assert(imm<32);
+  assem_debug("lslpls %s,%s,#%d\n",regname[rt],regname[rs],imm);
+  output_w32(0x51b00000|rd_rn_rm(rt,0,rs)|(imm<<7));
+}
+
 void emit_shrimm(int rs,u_int imm,int rt)
 {
   assert(imm>0);
@@ -1388,7 +1400,7 @@ void emit_shrdimm(int rs,int rs2,u_int imm,int rt)
 
 void emit_signextend16(int rs,int rt)
 {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV6
   emit_shlimm(rs,16,rt);
   emit_sarimm(rt,16,rt);
   #else
@@ -1399,7 +1411,7 @@ void emit_signextend16(int rs,int rt)
 
 void emit_signextend8(int rs,int rt)
 {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV6
   emit_shlimm(rs,24,rt);
   emit_sarimm(rt,24,rt);
   #else
@@ -1487,20 +1499,12 @@ void emit_cmpimm(int rs,int imm)
     output_w32(0xe3700000|rd_rn_rm(0,rs,0)|armval);
   }else if(imm>0) {
     assert(imm<65536);
-    #ifdef ARMv5_ONLY
     emit_movimm(imm,HOST_TEMPREG);
-    #else
-    emit_movw(imm,HOST_TEMPREG);
-    #endif
     assem_debug("cmp %s,r14\n",regname[rs]);
     output_w32(0xe1500000|rd_rn_rm(0,rs,HOST_TEMPREG));
   }else{
     assert(imm>-65536);
-    #ifdef ARMv5_ONLY
     emit_movimm(-imm,HOST_TEMPREG);
-    #else
-    emit_movw(-imm,HOST_TEMPREG);
-    #endif
     assem_debug("cmn %s,r14\n",regname[rs]);
     output_w32(0xe1700000|rd_rn_rm(0,rs,HOST_TEMPREG));
   }
@@ -1924,6 +1928,16 @@ void emit_movzwl_indexed(int offset, int rs, int rt)
     output_w32(0xe15000b0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf));
   }
 }
+static void emit_ldrd(int offset, int rs, int rt)
+{
+  assert(offset>-256&&offset<256);
+  assem_debug("ldrd %s,%s+%d\n",regname[rt],regname[rs],offset);
+  if(offset>=0) {
+    output_w32(0xe1c000d0|rd_rn_rm(rt,rs,0)|((offset<<4)&0xf00)|(offset&0xf));
+  }else{
+    output_w32(0xe14000d0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf));
+  }
+}
 void emit_readword(int addr, int rt)
 {
   u_int offset = addr-(u_int)&dynarec_local;
@@ -2158,6 +2172,14 @@ void emit_shrcc_imm(int rs,u_int imm,int rt)
   output_w32(0x31a00000|rd_rn_rm(rt,0,rs)|0x20|(imm<<7));
 }
 
+void emit_shrne_imm(int rs,u_int imm,int rt)
+{
+  assert(imm>0);
+  assert(imm<32);
+  assem_debug("lsrne %s,%s,#%d\n",regname[rt],regname[rs],imm);
+  output_w32(0x11a00000|rd_rn_rm(rt,0,rs)|0x20|(imm<<7));
+}
+
 void emit_negmi(int rs, int rt)
 {
   assem_debug("rsbmi %s,%s,#0\n",regname[rt],regname[rs]);
@@ -2262,7 +2284,7 @@ void emit_cmov2imm_e_ne_compact(int imm1,int imm2,u_int rt)
     output_w32(0x12400000|rd_rn_rm(rt,rt,0)|armval);
   }
   else {
-    #ifdef ARMv5_ONLY
+    #ifndef HAVE_ARMV7
     emit_movimm(imm1,rt);
     add_literal((int)out,imm2);
     assem_debug("ldrne %s,pc+? [=%x]\n",regname[rt],imm2);
@@ -2553,6 +2575,14 @@ void emit_andne_imm(int rs,int imm,int rt)
   output_w32(0x12000000|rd_rn_rm(rt,rs,0)|armval);
 }
 
+void emit_addpl_imm(int rs,int imm,int rt)
+{
+  u_int armval;
+  genimm_checked(imm,&armval);
+  assem_debug("addpl %s,%s,#%d\n",regname[rt],regname[rs],imm);
+  output_w32(0x52800000|rd_rn_rm(rt,rs,0)|armval);
+}
+
 void emit_jno_unlikely(int a)
 {
   //emit_jno(a);
@@ -2560,34 +2590,40 @@ void emit_jno_unlikely(int a)
   output_w32(0x72800000|rd_rn_rm(15,15,0));
 }
 
-// Save registers before function call
-void save_regs(u_int reglist)
+static void save_regs_all(u_int reglist)
 {
-  reglist&=0x100f; // only save the caller-save registers, r0-r3, r12
+  int i;
   if(!reglist) return;
   assem_debug("stmia fp,{");
-  if(reglist&1) assem_debug("r0, ");
-  if(reglist&2) assem_debug("r1, ");
-  if(reglist&4) assem_debug("r2, ");
-  if(reglist&8) assem_debug("r3, ");
-  if(reglist&0x1000) assem_debug("r12");
+  for(i=0;i<16;i++)
+    if(reglist&(1<<i))
+      assem_debug("r%d,",i);
   assem_debug("}\n");
   output_w32(0xe88b0000|reglist);
 }
-// Restore registers after function call
-void restore_regs(u_int reglist)
+static void restore_regs_all(u_int reglist)
 {
-  reglist&=0x100f; // only restore the caller-save registers, r0-r3, r12
+  int i;
   if(!reglist) return;
   assem_debug("ldmia fp,{");
-  if(reglist&1) assem_debug("r0, ");
-  if(reglist&2) assem_debug("r1, ");
-  if(reglist&4) assem_debug("r2, ");
-  if(reglist&8) assem_debug("r3, ");
-  if(reglist&0x1000) assem_debug("r12");
+  for(i=0;i<16;i++)
+    if(reglist&(1<<i))
+      assem_debug("r%d,",i);
   assem_debug("}\n");
   output_w32(0xe89b0000|reglist);
 }
+// Save registers before function call
+static void save_regs(u_int reglist)
+{
+  reglist&=CALLER_SAVE_REGS; // only save the caller-save registers, r0-r3, r12
+  save_regs_all(reglist);
+}
+// Restore registers after function call
+static void restore_regs(u_int reglist)
+{
+  reglist&=CALLER_SAVE_REGS;
+  restore_regs_all(reglist);
+}
 
 // Write back consts using r14 so we don't disturb the other registers
 void wb_consts(signed char i_regmap[],uint64_t i_is32,u_int i_dirty,int i)
@@ -2605,12 +2641,6 @@ void wb_consts(signed char i_regmap[],uint64_t i_is32,u_int i_dirty,int i)
             emit_movimm(value,HOST_TEMPREG);
           }
           emit_storereg(i_regmap[hr],HOST_TEMPREG);
-#ifndef FORCE32
-          if((i_is32>>i_regmap[hr])&1) {
-            if(value!=-1&&value!=0) emit_sarimm(HOST_TEMPREG,31,HOST_TEMPREG);
-            emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
-          }
-#endif
         }
       }
     }
@@ -2629,12 +2659,24 @@ void literal_pool(int n)
   int i;
   for(i=0;i<literalcount;i++)
   {
+    u_int l_addr=(u_int)out;
+    int j;
+    for(j=0;j<i;j++) {
+      if(literals[j][1]==literals[i][1]) {
+        //printf("dup %08x\n",literals[i][1]);
+        l_addr=literals[j][0];
+        break;
+      }
+    }
     ptr=(u_int *)literals[i][0];
-    u_int offset=(u_int)out-(u_int)ptr-8;
+    u_int offset=l_addr-(u_int)ptr-8;
     assert(offset<4096);
     assert(!(offset&3));
     *ptr|=offset;
-    output_w32(literals[i][1]);
+    if(l_addr==(u_int)out) {
+      literals[i][0]=l_addr; // remember for dupes
+      output_w32(literals[i][1]);
+    }
   }
   literalcount=0;
 }
@@ -2651,7 +2693,7 @@ void literal_pool_jumpover(int n)
   set_jump_target(jaddr,(int)out);
 }
 
-emit_extjump2(int addr, int target, int linker)
+emit_extjump2(u_int addr, int target, int linker)
 {
   u_char *ptr=(u_char *)addr;
   assert((ptr[3]&0x0e)==0xa);
@@ -2684,14 +2726,45 @@ emit_extjump_ds(int addr, int target)
 // put rt_val into rt, potentially making use of rs with value rs_val
 static void emit_movimm_from(u_int rs_val,int rs,u_int rt_val,int rt)
 {
-  u_int xor=rs_val^rt_val;
+  u_int armval;
+  int diff;
+  if(genimm(rt_val,&armval)) {
+    assem_debug("mov %s,#%d\n",regname[rt],rt_val);
+    output_w32(0xe3a00000|rd_rn_rm(rt,0,0)|armval);
+    return;
+  }
+  if(genimm(~rt_val,&armval)) {
+    assem_debug("mvn %s,#%d\n",regname[rt],rt_val);
+    output_w32(0xe3e00000|rd_rn_rm(rt,0,0)|armval);
+    return;
+  }
+  diff=rt_val-rs_val;
+  if(genimm(diff,&armval)) {
+    assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],diff);
+    output_w32(0xe2800000|rd_rn_rm(rt,rs,0)|armval);
+    return;
+  }else if(genimm(-diff,&armval)) {
+    assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],-diff);
+    output_w32(0xe2400000|rd_rn_rm(rt,rs,0)|armval);
+    return;
+  }
+  emit_movimm(rt_val,rt);
+}
+
+// return 1 if above function can do it's job cheaply
+static int is_similar_value(u_int v1,u_int v2)
+{
   u_int xs;
-  for(xs=xor;xs!=0&&(xs&3)==0;xs>>=2)
+  int diff;
+  if(v1==v2) return 1;
+  diff=v2-v1;
+  for(xs=diff;xs!=0&&(xs&3)==0;xs>>=2)
     ;
-  if(xs<0x100)
-    emit_xorimm(rs,xor,rt);
-  else
-    emit_movimm(rt_val,rt);
+  if(xs<0x100) return 1;
+  for(xs=-diff;xs!=0&&(xs&3)==0;xs>>=2)
+    ;
+  if(xs<0x100) return 1;
+  return 0;
 }
 
 // trashes r2
@@ -2711,6 +2784,21 @@ static void pass_args(int a0, int a1)
   }
 }
 
+static void mov_loadtype_adj(int type,int rs,int rt)
+{
+  switch(type) {
+    case LOADB_STUB:  emit_signextend8(rs,rt); break;
+    case LOADBU_STUB: emit_andimm(rs,0xff,rt); break;
+    case LOADH_STUB:  emit_signextend16(rs,rt); break;
+    case LOADHU_STUB: emit_andimm(rs,0xffff,rt); break;
+    case LOADW_STUB:  if(rs!=rt) emit_mov(rs,rt); break;
+    default: assert(0);
+  }
+}
+
+#include "pcsxmem.h"
+#include "pcsxmem_inline.c"
+
 do_readstub(int n)
 {
   assem_debug("do_readstub %x\n",start+stubs[n][3]*4);
@@ -2733,7 +2821,6 @@ do_readstub(int n)
     rt=get_reg(i_regmap,rt1[i]);
   }
   assert(rs>=0);
-#ifdef PCSX
   int r,temp=-1,temp2=HOST_TEMPREG,regs_saved=0,restore_jump=0;
   reglist|=(1<<rs);
   for(r=0;r<=12;r++) {
@@ -2741,7 +2828,7 @@ do_readstub(int n)
       temp=r; break;
     }
   }
-  if(rt>=0)
+  if(rt>=0&&rt1[i]!=0)
     reglist&=~(1<<rt);
   if(temp==-1) {
     save_regs(reglist);
@@ -2787,101 +2874,14 @@ do_readstub(int n)
   emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n][6]+1),2);
   emit_call(handler);
   if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) {
-    switch(type) {
-      case LOADB_STUB:  emit_signextend8(0,rt); break;
-      case LOADBU_STUB: emit_andimm(0,0xff,rt); break;
-      case LOADH_STUB:  emit_signextend16(0,rt); break;
-      case LOADHU_STUB: emit_andimm(0,0xffff,rt); break;
-      case LOADW_STUB:  if(rt!=0) emit_mov(0,rt); break;
-    }
+    mov_loadtype_adj(type,0,rt);
   }
   if(restore_jump)
     set_jump_target(restore_jump,(int)out);
   restore_regs(reglist);
   emit_jmp(stubs[n][2]); // return address
-#else // !PCSX
-  if(addr<0) addr=rt;
-  if(addr<0&&itype[i]!=C1LS&&itype[i]!=C2LS&&itype[i]!=LOADLR) addr=get_reg(i_regmap,-1);
-  assert(addr>=0);
-  int ftable=0;
-  if(type==LOADB_STUB||type==LOADBU_STUB)
-    ftable=(int)readmemb;
-  if(type==LOADH_STUB||type==LOADHU_STUB)
-    ftable=(int)readmemh;
-  if(type==LOADW_STUB)
-    ftable=(int)readmem;
-#ifndef FORCE32
-  if(type==LOADD_STUB)
-    ftable=(int)readmemd;
-#endif
-  assert(ftable!=0);
-  emit_writeword(rs,(int)&address);
-  //emit_pusha();
-  save_regs(reglist);
-#ifndef PCSX
-  ds=i_regs!=&regs[i];
-  int real_rs=(itype[i]==LOADLR)?-1:get_reg(i_regmap,rs1[i]);
-  u_int cmask=ds?-1:(0x100f|~i_regs->wasconst);
-  if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i);
-  wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs)));
-  if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i);
-#endif
-  emit_shrimm(rs,16,1);
-  int cc=get_reg(i_regmap,CCREG);
-  if(cc<0) {
-    emit_loadreg(CCREG,2);
-  }
-  emit_movimm(ftable,0);
-  emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2);
-#ifndef PCSX
-  emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3);
-#endif
-  //emit_readword((int)&last_count,temp);
-  //emit_add(cc,temp,cc);
-  //emit_writeword(cc,(int)&Count);
-  //emit_mov(15,14);
-  emit_call((int)&indirect_jump_indexed);
-  //emit_callreg(rs);
-  //emit_readword_dualindexedx4(rs,HOST_TEMPREG,15);
-#ifndef PCSX
-  // We really shouldn't need to update the count here,
-  // but not doing so causes random crashes...
-  emit_readword((int)&Count,HOST_TEMPREG);
-  emit_readword((int)&next_interupt,2);
-  emit_addimm(HOST_TEMPREG,-2*stubs[n][6]-2,HOST_TEMPREG);
-  emit_writeword(2,(int)&last_count);
-  emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
-  if(cc<0) {
-    emit_storereg(CCREG,HOST_TEMPREG);
-  }
-#endif
-  //emit_popa();
-  restore_regs(reglist);
-  //if((cc=get_reg(regmap,CCREG))>=0) {
-  //  emit_loadreg(CCREG,cc);
-  //}
-  if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) {
-    assert(rt>=0);
-    if(type==LOADB_STUB)
-      emit_movsbl((int)&readmem_dword,rt);
-    if(type==LOADBU_STUB)
-      emit_movzbl((int)&readmem_dword,rt);
-    if(type==LOADH_STUB)
-      emit_movswl((int)&readmem_dword,rt);
-    if(type==LOADHU_STUB)
-      emit_movzwl((int)&readmem_dword,rt);
-    if(type==LOADW_STUB)
-      emit_readword((int)&readmem_dword,rt);
-    if(type==LOADD_STUB) {
-      emit_readword((int)&readmem_dword,rt);
-      if(rth>=0) emit_readword(((int)&readmem_dword)+4,rth);
-    }
-  }
-  emit_jmp(stubs[n][2]); // return address
-#endif // !PCSX
 }
 
-#ifdef PCSX
 // return memhandler, or get directly accessable address and return 0
 u_int get_direct_memhandler(void *table,u_int addr,int type,u_int *addr_host)
 {
@@ -2908,7 +2908,6 @@ u_int get_direct_memhandler(void *table,u_int addr,int type,u_int *addr_host)
     return l2<<1;
   }
 }
-#endif
 
 inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, int adj, u_int reglist)
 {
@@ -2917,11 +2916,13 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
   int rt=get_reg(regmap,target);
   if(rs<0) rs=get_reg(regmap,-1);
   assert(rs>=0);
-#ifdef PCSX
-  u_int handler,host_addr=0;
+  u_int handler,host_addr=0,is_dynamic,far_call=0;
+  int cc=get_reg(regmap,CCREG);
+  if(pcsx_direct_read(type,addr,CLOCK_ADJUST(adj+1),cc,target?rs:-1,rt))
+    return;
   handler=get_direct_memhandler(mem_rtab,addr,type,&host_addr);
   if (handler==0) {
-    if(rt<0)
+    if(rt<0||rt1[i]==0)
       return;
     if(addr!=host_addr)
       emit_movimm_from(addr,rs,host_addr,rs);
@@ -2935,32 +2936,49 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
     }
     return;
   }
+  is_dynamic=pcsxmem_is_handler_dynamic(addr);
+  if(is_dynamic) {
+    if(type==LOADB_STUB||type==LOADBU_STUB)
+      handler=(int)jump_handler_read8;
+    if(type==LOADH_STUB||type==LOADHU_STUB)
+      handler=(int)jump_handler_read16;
+    if(type==LOADW_STUB)
+      handler=(int)jump_handler_read32;
+  }
 
   // call a memhandler
-  if(rt>=0)
+  if(rt>=0&&rt1[i]!=0)
     reglist&=~(1<<rt);
   save_regs(reglist);
   if(target==0)
     emit_movimm(addr,0);
   else if(rs!=0)
     emit_mov(rs,0);
-  int cc=get_reg(regmap,CCREG);
-  if(cc<0)
-    emit_loadreg(CCREG,2);
-  emit_readword((int)&last_count,3);
-  emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
-  emit_add(2,3,3);
-  emit_writeword(3,(int)&Count);
-
   int offset=(int)handler-(int)out-8;
   if(offset<-33554432||offset>=33554432) {
     // unreachable memhandler, a plugin func perhaps
-    emit_movimm(handler,1);
-    emit_callreg(1);
+    emit_movimm(handler,12);
+    far_call=1;
   }
+  if(cc<0)
+    emit_loadreg(CCREG,2);
+  if(is_dynamic) {
+    emit_movimm(((u_int *)mem_rtab)[addr>>12]<<1,1);
+    emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
+  }
+  else {
+    emit_readword((int)&last_count,3);
+    emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
+    emit_add(2,3,2);
+    emit_writeword(2,(int)&Count);
+  }
+
+  if(far_call)
+    emit_callreg(12);
   else
     emit_call(handler);
-  if(rt>=0) {
+
+  if(rt>=0&&rt1[i]!=0) {
     switch(type) {
       case LOADB_STUB:  emit_signextend8(0,rt); break;
       case LOADBU_STUB: emit_andimm(0,0xff,rt); break;
@@ -2971,87 +2989,6 @@ inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, i
     }
   }
   restore_regs(reglist);
-#else // if !PCSX
-  int ftable=0;
-  if(type==LOADB_STUB||type==LOADBU_STUB)
-    ftable=(int)readmemb;
-  if(type==LOADH_STUB||type==LOADHU_STUB)
-    ftable=(int)readmemh;
-  if(type==LOADW_STUB)
-    ftable=(int)readmem;
-#ifndef FORCE32
-  if(type==LOADD_STUB)
-    ftable=(int)readmemd;
-#endif
-  assert(ftable!=0);
-  if(target==0)
-    emit_movimm(addr,rs);
-  emit_writeword(rs,(int)&address);
-  //emit_pusha();
-  save_regs(reglist);
-#ifndef PCSX
-  if((signed int)addr>=(signed int)0xC0000000) {
-    // Theoretically we can have a pagefault here, if the TLB has never
-    // been enabled and the address is outside the range 80000000..BFFFFFFF
-    // Write out the registers so the pagefault can be handled.  This is
-    // a very rare case and likely represents a bug.
-    int ds=regmap!=regs[i].regmap;
-    if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i);
-    if(!ds) wb_dirtys(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty);
-    else wb_dirtys(branch_regs[i-1].regmap_entry,branch_regs[i-1].was32,branch_regs[i-1].wasdirty);
-  }
-#endif
-  //emit_shrimm(rs,16,1);
-  int cc=get_reg(regmap,CCREG);
-  if(cc<0) {
-    emit_loadreg(CCREG,2);
-  }
-  //emit_movimm(ftable,0);
-  emit_movimm(((u_int *)ftable)[addr>>16],0);
-  //emit_readword((int)&last_count,12);
-  emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
-#ifndef PCSX
-  if((signed int)addr>=(signed int)0xC0000000) {
-    // Pagefault address
-    int ds=regmap!=regs[i].regmap;
-    emit_movimm(start+i*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3);
-  }
-#endif
-  //emit_add(12,2,2);
-  //emit_writeword(2,(int)&Count);
-  //emit_call(((u_int *)ftable)[addr>>16]);
-  emit_call((int)&indirect_jump);
-#ifndef PCSX
-  // We really shouldn't need to update the count here,
-  // but not doing so causes random crashes...
-  emit_readword((int)&Count,HOST_TEMPREG);
-  emit_readword((int)&next_interupt,2);
-  emit_addimm(HOST_TEMPREG,-CLOCK_ADJUST(adj+1),HOST_TEMPREG);
-  emit_writeword(2,(int)&last_count);
-  emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
-  if(cc<0) {
-    emit_storereg(CCREG,HOST_TEMPREG);
-  }
-#endif
-  //emit_popa();
-  restore_regs(reglist);
-  if(rt>=0) {
-    if(type==LOADB_STUB)
-      emit_movsbl((int)&readmem_dword,rt);
-    if(type==LOADBU_STUB)
-      emit_movzbl((int)&readmem_dword,rt);
-    if(type==LOADH_STUB)
-      emit_movswl((int)&readmem_dword,rt);
-    if(type==LOADHU_STUB)
-      emit_movzwl((int)&readmem_dword,rt);
-    if(type==LOADW_STUB)
-      emit_readword((int)&readmem_dword,rt);
-    if(type==LOADD_STUB) {
-      emit_readword((int)&readmem_dword,rt);
-      if(rth>=0) emit_readword(((int)&readmem_dword)+4,rth);
-    }
-  }
-#endif // !PCSX
 }
 
 do_writestub(int n)
@@ -3077,7 +3014,6 @@ do_writestub(int n)
   }
   assert(rs>=0);
   assert(rt>=0);
-#ifdef PCSX
   int rtmp,temp=-1,temp2=HOST_TEMPREG,regs_saved=0,restore_jump=0,ra;
   int reglist2=reglist|(1<<rs)|(1<<rt);
   for(rtmp=0;rtmp<=12;rtmp++) {
@@ -3136,81 +3072,7 @@ do_writestub(int n)
     set_jump_target(restore_jump,(int)out);
   restore_regs(reglist);
   ra=stubs[n][2];
-  if(!restore_jump) ra+=4*3; // skip invcode check
   emit_jmp(ra);
-#else // if !PCSX
-  if(addr<0) addr=get_reg(i_regmap,-1);
-  assert(addr>=0);
-  int ftable=0;
-  if(type==STOREB_STUB)
-    ftable=(int)writememb;
-  if(type==STOREH_STUB)
-    ftable=(int)writememh;
-  if(type==STOREW_STUB)
-    ftable=(int)writemem;
-#ifndef FORCE32
-  if(type==STORED_STUB)
-    ftable=(int)writememd;
-#endif
-  assert(ftable!=0);
-  emit_writeword(rs,(int)&address);
-  //emit_shrimm(rs,16,rs);
-  //emit_movmem_indexedx4(ftable,rs,rs);
-  if(type==STOREB_STUB)
-    emit_writebyte(rt,(int)&byte);
-  if(type==STOREH_STUB)
-    emit_writehword(rt,(int)&hword);
-  if(type==STOREW_STUB)
-    emit_writeword(rt,(int)&word);
-  if(type==STORED_STUB) {
-#ifndef FORCE32
-    emit_writeword(rt,(int)&dword);
-    emit_writeword(r?rth:rt,(int)&dword+4);
-#else
-    printf("STORED_STUB\n");
-#endif
-  }
-  //emit_pusha();
-  save_regs(reglist);
-#ifndef PCSX
-  ds=i_regs!=&regs[i];
-  int real_rs=get_reg(i_regmap,rs1[i]);
-  u_int cmask=ds?-1:(0x100f|~i_regs->wasconst);
-  if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i);
-  wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs)));
-  if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i);
-#endif
-  emit_shrimm(rs,16,1);
-  int cc=get_reg(i_regmap,CCREG);
-  if(cc<0) {
-    emit_loadreg(CCREG,2);
-  }
-  emit_movimm(ftable,0);
-  emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2);
-#ifndef PCSX
-  emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3);
-#endif
-  //emit_readword((int)&last_count,temp);
-  //emit_addimm(cc,2*stubs[n][5]+2,cc);
-  //emit_add(cc,temp,cc);
-  //emit_writeword(cc,(int)&Count);
-  emit_call((int)&indirect_jump_indexed);
-  //emit_callreg(rs);
-  emit_readword((int)&Count,HOST_TEMPREG);
-  emit_readword((int)&next_interupt,2);
-  emit_addimm(HOST_TEMPREG,-2*stubs[n][6]-2,HOST_TEMPREG);
-  emit_writeword(2,(int)&last_count);
-  emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
-  if(cc<0) {
-    emit_storereg(CCREG,HOST_TEMPREG);
-  }
-  //emit_popa();
-  restore_regs(reglist);
-  //if((cc=get_reg(regmap,CCREG))>=0) {
-  //  emit_loadreg(CCREG,cc);
-  //}
-  emit_jmp(stubs[n][2]); // return address
-#endif // !PCSX
 }
 
 inline_writestub(int type, int i, u_int addr, signed char regmap[], int target, int adj, u_int reglist)
@@ -3220,7 +3082,6 @@ inline_writestub(int type, int i, u_int addr, signed char regmap[], int target,
   int rt=get_reg(regmap,target);
   assert(rs>=0);
   assert(rt>=0);
-#ifdef PCSX
   u_int handler,host_addr=0;
   handler=get_direct_memhandler(mem_wtab,addr,type,&host_addr);
   if (handler==0) {
@@ -3249,82 +3110,6 @@ inline_writestub(int type, int i, u_int addr, signed char regmap[], int target,
   if(cc<0)
     emit_storereg(CCREG,2);
   restore_regs(reglist);
-#else // if !pcsx
-  int ftable=0;
-  if(type==STOREB_STUB)
-    ftable=(int)writememb;
-  if(type==STOREH_STUB)
-    ftable=(int)writememh;
-  if(type==STOREW_STUB)
-    ftable=(int)writemem;
-#ifndef FORCE32
-  if(type==STORED_STUB)
-    ftable=(int)writememd;
-#endif
-  assert(ftable!=0);
-  emit_writeword(rs,(int)&address);
-  //emit_shrimm(rs,16,rs);
-  //emit_movmem_indexedx4(ftable,rs,rs);
-  if(type==STOREB_STUB)
-    emit_writebyte(rt,(int)&byte);
-  if(type==STOREH_STUB)
-    emit_writehword(rt,(int)&hword);
-  if(type==STOREW_STUB)
-    emit_writeword(rt,(int)&word);
-  if(type==STORED_STUB) {
-#ifndef FORCE32
-    emit_writeword(rt,(int)&dword);
-    emit_writeword(target?rth:rt,(int)&dword+4);
-#else
-    printf("STORED_STUB\n");
-#endif
-  }
-  //emit_pusha();
-  save_regs(reglist);
-#ifndef PCSX
-  // rearmed note: load_all_consts prevents BIOS boot, some bug?
-  if((signed int)addr>=(signed int)0xC0000000) {
-    // Theoretically we can have a pagefault here, if the TLB has never
-    // been enabled and the address is outside the range 80000000..BFFFFFFF
-    // Write out the registers so the pagefault can be handled.  This is
-    // a very rare case and likely represents a bug.
-    int ds=regmap!=regs[i].regmap;
-    if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i);
-    if(!ds) wb_dirtys(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty);
-    else wb_dirtys(branch_regs[i-1].regmap_entry,branch_regs[i-1].was32,branch_regs[i-1].wasdirty);
-  }
-#endif
-  //emit_shrimm(rs,16,1);
-  int cc=get_reg(regmap,CCREG);
-  if(cc<0) {
-    emit_loadreg(CCREG,2);
-  }
-  //emit_movimm(ftable,0);
-  emit_movimm(((u_int *)ftable)[addr>>16],0);
-  //emit_readword((int)&last_count,12);
-  emit_addimm(cc<0?2:cc,CLOCK_ADJUST(adj+1),2);
-#ifndef PCSX
-  if((signed int)addr>=(signed int)0xC0000000) {
-    // Pagefault address
-    int ds=regmap!=regs[i].regmap;
-    emit_movimm(start+i*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3);
-  }
-#endif
-  //emit_add(12,2,2);
-  //emit_writeword(2,(int)&Count);
-  //emit_call(((u_int *)ftable)[addr>>16]);
-  emit_call((int)&indirect_jump);
-  emit_readword((int)&Count,HOST_TEMPREG);
-  emit_readword((int)&next_interupt,2);
-  emit_addimm(HOST_TEMPREG,-CLOCK_ADJUST(adj+1),HOST_TEMPREG);
-  emit_writeword(2,(int)&last_count);
-  emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
-  if(cc<0) {
-    emit_storereg(CCREG,HOST_TEMPREG);
-  }
-  //emit_popa();
-  restore_regs(reglist);
-#endif
 }
 
 do_unalignedwritestub(int n)
@@ -3367,14 +3152,6 @@ do_unalignedwritestub(int n)
   emit_writeword(temp2,(int)&address);
 
   save_regs(reglist);
-#ifndef PCSX
-  ds=i_regs!=&regs[i];
-  real_rs=get_reg(i_regmap,rs1[i]);
-  u_int cmask=ds?-1:(0x100f|~i_regs->wasconst);
-  if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i);
-  wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs)));
-  if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i);
-#endif
   emit_shrimm(addr,16,1);
   int cc=get_reg(i_regmap,CCREG);
   if(cc<0) {
@@ -3382,10 +3159,6 @@ do_unalignedwritestub(int n)
   }
   emit_movimm((u_int)readmem,0);
   emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2);
-#ifndef PCSX
-  // pagefault address
-  emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3);
-#endif
   emit_call((int)&indirect_jump_indexed);
   restore_regs(reglist);
 
@@ -3448,12 +3221,9 @@ do_invstub(int n)
 int do_dirty_stub(int i)
 {
   assem_debug("do_dirty_stub %x\n",start+i*4);
-  u_int addr=(int)start<(int)0xC0000000?(u_int)source:(u_int)start;
-  #ifdef PCSX
-  addr=(u_int)source;
-  #endif
+  u_int addr=(u_int)source;
   // Careful about the code output here, verify_dirty needs to parse it.
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_loadlp(addr,1);
   emit_loadlp((int)copy,2);
   emit_loadlp(slen*4,3);
@@ -3476,7 +3246,7 @@ int do_dirty_stub(int i)
 void do_dirty_stub_ds()
 {
   // Careful about the code output here, verify_dirty needs to parse it.
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_loadlp((int)start<(int)0xC0000000?(int)source:(int)start,1);
   emit_loadlp((int)copy,2);
   emit_loadlp(slen*4,3);
@@ -3512,101 +3282,12 @@ do_cop1stub(int n)
   emit_jmp(ds?(int)fp_exception_ds:(int)fp_exception);
 }
 
-#ifndef DISABLE_TLB
-
-/* TLB */
-
-int do_tlb_r(int s,int ar,int map,int x,int a,int shift,int c,u_int addr)
-{
-  if(c) {
-    if((signed int)addr>=(signed int)0xC0000000) {
-      // address_generation already loaded the const
-      emit_readword_dualindexedx4(FP,map,map);
-    }
-    else
-      return -1; // No mapping
-  }
-  else {
-    assert(s!=map);
-    emit_movimm(((int)memory_map-(int)&dynarec_local)>>2,map);
-    emit_addsr12(map,s,map);
-    // Schedule this while we wait on the load
-    //if(x) emit_xorimm(s,x,ar);
-    if(shift>=0) emit_shlimm(s,3,shift);
-    if(~a) emit_andimm(s,a,ar);
-    emit_readword_dualindexedx4(FP,map,map);
-  }
-  return map;
-}
-int do_tlb_r_branch(int map, int c, u_int addr, int *jaddr)
-{
-  if(!c||(signed int)addr>=(signed int)0xC0000000) {
-    emit_test(map,map);
-    *jaddr=(int)out;
-    emit_js(0);
-  }
-  return map;
-}
-
-int gen_tlb_addr_r(int ar, int map) {
-  if(map>=0) {
-    assem_debug("add %s,%s,%s lsl #2\n",regname[ar],regname[ar],regname[map]);
-    output_w32(0xe0800100|rd_rn_rm(ar,ar,map));
-  }
-}
-
-int do_tlb_w(int s,int ar,int map,int x,int c,u_int addr)
-{
-  if(c) {
-    if(addr<0x80800000||addr>=0xC0000000) {
-      // address_generation already loaded the const
-      emit_readword_dualindexedx4(FP,map,map);
-    }
-    else
-      return -1; // No mapping
-  }
-  else {
-    assert(s!=map);
-    emit_movimm(((int)memory_map-(int)&dynarec_local)>>2,map);
-    emit_addsr12(map,s,map);
-    // Schedule this while we wait on the load
-    //if(x) emit_xorimm(s,x,ar);
-    emit_readword_dualindexedx4(FP,map,map);
-  }
-  return map;
-}
-int do_tlb_w_branch(int map, int c, u_int addr, int *jaddr)
-{
-  if(!c||addr<0x80800000||addr>=0xC0000000) {
-    emit_testimm(map,0x40000000);
-    *jaddr=(int)out;
-    emit_jne(0);
-  }
-}
-
-int gen_tlb_addr_w(int ar, int map) {
-  if(map>=0) {
-    assem_debug("add %s,%s,%s lsl #2\n",regname[ar],regname[ar],regname[map]);
-    output_w32(0xe0800100|rd_rn_rm(ar,ar,map));
-  }
-}
-
-// Generate the address of the memory_map entry, relative to dynarec_local
-generate_map_const(u_int addr,int reg) {
-  //printf("generate_map_const(%x,%s)\n",addr,regname[reg]);
-  emit_movimm((addr>>12)+(((u_int)memory_map-(u_int)&dynarec_local)>>2),reg);
-}
-
-#else
-
-static int do_tlb_r() { return 0; }
-static int do_tlb_r_branch() { return 0; }
-static int gen_tlb_addr_r() { return 0; }
-static int do_tlb_w() { return 0; }
-static int do_tlb_w_branch() { return 0; }
-static int gen_tlb_addr_w() { return 0; }
-
-#endif // DISABLE_TLB
+static int do_tlb_r(int a, ...) { return 0; }
+static int do_tlb_r_branch(int a, ...) { return 0; }
+static int gen_tlb_addr_r(int a, ...) { return 0; }
+static int do_tlb_w(int a, ...) { return 0; }
+static int do_tlb_w_branch(int a, ...) { return 0; }
+static int gen_tlb_addr_w(int a, ...) { return 0; }
 
 /* Special assem */
 
@@ -3720,7 +3401,6 @@ void shift_assemble_arm(int i,struct regstat *i_regs)
   }
 }
 
-#ifdef PCSX
 static void speculate_mov(int rs,int rt)
 {
   if(rt!=0) {
@@ -3838,13 +3518,10 @@ static int get_ptr_mem_type(u_int a)
     return MTYPE_A000;
   return MTYPE_8000;
 }
-#endif
 
 static int emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
 {
   int jaddr,type=0;
-
-#ifdef PCSX
   int mr=rs1[i];
   if(((smrv_strong|smrv_weak)>>mr)&1) {
     type=get_ptr_mem_type(smrv[mr]);
@@ -3872,12 +3549,17 @@ static int emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
     type=0;
   }
   else if(type==MTYPE_1F80) { // scratchpad
-    emit_addimm(addr,-0x1f800000,HOST_TEMPREG);
-    emit_cmpimm(HOST_TEMPREG,0x1000);
-    jaddr=(int)out;
-    emit_jc(0);
+    if (psxH == (void *)0x1f800000) {
+      emit_addimm(addr,-0x1f800000,HOST_TEMPREG);
+      emit_cmpimm(HOST_TEMPREG,0x1000);
+      jaddr=(int)out;
+      emit_jc(0);
+    }
+    else {
+      // do usual RAM check, jump will go to the right handler
+      type=0;
+    }
   }
-#endif
 
   if(type==0)
   {
@@ -3890,6 +3572,10 @@ static int emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
     else
     #endif
       emit_jno(0);
+    if(ram_offset!=0) {
+      emit_addimm(addr,ram_offset,HOST_TEMPREG);
+      addr=*addr_reg_override=HOST_TEMPREG;
+    }
   }
 
   return jaddr;
@@ -3941,6 +3627,10 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs)
       jaddr=emit_fastpath_cmp_jump(i,temp2,&fastload_reg_override);
     }
     else {
+      if(ram_offset&&memtarget) {
+        emit_addimm(temp2,ram_offset,HOST_TEMPREG);
+        fastload_reg_override=HOST_TEMPREG;
+      }
       if (opcode[i]==0x22||opcode[i]==0x26) {
         emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
       }else{
@@ -3959,7 +3649,6 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs)
     map=get_reg(i_regs->regmap,TLREG);
     assert(map>=0);
     reglist&=~(1<<map);
-    map=do_tlb_r(addr,temp2,map,0,a,c?-1:temp,c,constmap[i][s]+offset);
     if(c) {
       if (opcode[i]==0x22||opcode[i]==0x26) {
         emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
@@ -3967,7 +3656,6 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs)
         emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
       }
     }
-    do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr);
   }
   if (opcode[i]==0x22||opcode[i]==0x26) { // LWL/LWR
     if(!c||memtarget) {
@@ -4054,23 +3742,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
     char copr=(source[i]>>11)&0x1f;
     //assert(t>=0); // Why does this happen?  OOT is weird
     if(t>=0&&rt1[i]!=0) {
-#ifdef MUPEN64
-      emit_addimm(FP,(int)&fake_pc-(int)&dynarec_local,0);
-      emit_movimm((source[i]>>11)&0x1f,1);
-      emit_writeword(0,(int)&PC);
-      emit_writebyte(1,(int)&(fake_pc.f.r.nrd));
-      if(copr==9) {
-        emit_readword((int)&last_count,ECX);
-        emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
-        emit_add(HOST_CCREG,ECX,HOST_CCREG);
-        emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
-        emit_writeword(HOST_CCREG,(int)&Count);
-      }
-      emit_call((int)MFC0);
-      emit_readword((int)&readmem_dword,t);
-#else
       emit_readword((int)&reg_cop0+copr*4,t);
-#endif
     }
   }
   else if(opcode2[i]==4) // MTC0
@@ -4078,16 +3750,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
     signed char s=get_reg(i_regs->regmap,rs1[i]);
     char copr=(source[i]>>11)&0x1f;
     assert(s>=0);
-#ifdef MUPEN64
-    emit_writeword(s,(int)&readmem_dword);
-    wb_register(rs1[i],i_regs->regmap,i_regs->dirty,i_regs->is32);
-    emit_addimm(FP,(int)&fake_pc-(int)&dynarec_local,0);
-    emit_movimm((source[i]>>11)&0x1f,1);
-    emit_writeword(0,(int)&PC);
-    emit_writebyte(1,(int)&(fake_pc.f.r.nrd));
-#else
     wb_register(rs1[i],i_regs->regmap,i_regs->dirty,i_regs->is32);
-#endif
     if(copr==9||copr==11||copr==12||copr==13) {
       emit_readword((int)&last_count,HOST_TEMPREG);
       emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
@@ -4100,7 +3763,6 @@ void cop0_assemble(int i,struct regstat *i_regs)
     // The interrupt must be taken immediately, because a subsequent
     // instruction might disable interrupts again.
     if(copr==12||copr==13) {
-#ifdef PCSX
       if (is_delayslot) {
         // burn cycles to cause cc_interrupt, which will
         // reschedule next_interupt. Relies on CCREG from above.
@@ -4108,13 +3770,12 @@ void cop0_assemble(int i,struct regstat *i_regs)
         emit_writeword(HOST_CCREG,(int)&last_count);
         emit_movimm(0,HOST_CCREG);
         emit_storereg(CCREG,HOST_CCREG);
-        if(s!=1)
-          emit_mov(s,1);
+        emit_loadreg(rs1[i],1);
         emit_movimm(copr,0);
         emit_call((int)pcsx_mtc0_ds);
+        emit_loadreg(rs1[i],s);
         return;
       }
-#endif
       emit_movimm(start+i*4+4,HOST_TEMPREG);
       emit_writeword(HOST_TEMPREG,(int)&pcaddr);
       emit_movimm(0,HOST_TEMPREG);
@@ -4122,57 +3783,34 @@ void cop0_assemble(int i,struct regstat *i_regs)
     }
     //else if(copr==12&&is_delayslot) emit_call((int)MTC0_R12);
     //else
-#ifdef PCSX
-    if(s!=1)
+    if(s==HOST_CCREG)
+      emit_loadreg(rs1[i],1);
+    else if(s!=1)
       emit_mov(s,1);
     emit_movimm(copr,0);
     emit_call((int)pcsx_mtc0);
-#else
-    emit_call((int)MTC0);
-#endif
     if(copr==9||copr==11||copr==12||copr==13) {
       emit_readword((int)&Count,HOST_CCREG);
-      emit_readword((int)&next_interupt,ECX);
+      emit_readword((int)&next_interupt,HOST_TEMPREG);
       emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
-      emit_sub(HOST_CCREG,ECX,HOST_CCREG);
-      emit_writeword(ECX,(int)&last_count);
+      emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
+      emit_writeword(HOST_TEMPREG,(int)&last_count);
       emit_storereg(CCREG,HOST_CCREG);
     }
     if(copr==12||copr==13) {
       assert(!is_delayslot);
       emit_readword((int)&pending_exception,14);
+      emit_test(14,14);
+      emit_jne((int)&do_interrupt);
     }
     emit_loadreg(rs1[i],s);
     if(get_reg(i_regs->regmap,rs1[i]|64)>=0)
       emit_loadreg(rs1[i]|64,get_reg(i_regs->regmap,rs1[i]|64));
-    if(copr==12||copr==13) {
-      emit_test(14,14);
-      emit_jne((int)&do_interrupt);
-    }
     cop1_usable=0;
   }
   else
   {
     assert(opcode2[i]==0x10);
-#ifndef DISABLE_TLB
-    if((source[i]&0x3f)==0x01) // TLBR
-      emit_call((int)TLBR);
-    if((source[i]&0x3f)==0x02) // TLBWI
-      emit_call((int)TLBWI_new);
-    if((source[i]&0x3f)==0x06) { // TLBWR
-      // The TLB entry written by TLBWR is dependent on the count,
-      // so update the cycle count
-      emit_readword((int)&last_count,ECX);
-      if(i_regs->regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
-      emit_add(HOST_CCREG,ECX,HOST_CCREG);
-      emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
-      emit_writeword(HOST_CCREG,(int)&Count);
-      emit_call((int)TLBWR_new);
-    }
-    if((source[i]&0x3f)==0x08) // TLBP
-      emit_call((int)TLBP);
-#endif
-#ifdef PCSX
     if((source[i]&0x3f)==0x10) // RFE
     {
       emit_readword((int)&Status,0);
@@ -4181,15 +3819,6 @@ void cop0_assemble(int i,struct regstat *i_regs)
       emit_orrshr_imm(1,2,0);
       emit_writeword(0,(int)&Status);
     }
-#else
-    if((source[i]&0x3f)==0x18) // ERET
-    {
-      int count=ccadj[i];
-      if(i_regs->regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
-      emit_addimm(HOST_CCREG,CLOCK_ADJUST(count),HOST_CCREG); // TODO: Should there be an extra cycle here?
-      emit_jmp((int)jump_eret);
-    }
-#endif
   }
 }
 
@@ -4271,7 +3900,16 @@ static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
     case 30:
       emit_movs(sl,temp);
       emit_mvnmi(temp,temp);
+#ifdef HAVE_ARMV5
       emit_clz(temp,temp);
+#else
+      emit_movs(temp,HOST_TEMPREG);
+      emit_movimm(0,temp);
+      emit_jeq((int)out+4*4);
+      emit_addpl_imm(temp,1,temp);
+      emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
+      emit_jns((int)out-2*4);
+#endif
       emit_writeword(sl,(int)&reg_cop2d[30]);
       emit_writeword(temp,(int)&reg_cop2d[31]);
       break;
@@ -4334,36 +3972,162 @@ void cop2_assemble(int i,struct regstat *i_regs)
   }
 }
 
-void c2op_assemble(int i,struct regstat *i_regs)
+static void c2op_prologue(u_int op,u_int reglist)
+{
+  save_regs_all(reglist);
+#ifdef PCNT
+  emit_movimm(op,0);
+  emit_call((int)pcnt_gte_start);
+#endif
+  emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0); // cop2 regs
+}
+
+static void c2op_epilogue(u_int op,u_int reglist)
+{
+#ifdef PCNT
+  emit_movimm(op,0);
+  emit_call((int)pcnt_gte_end);
+#endif
+  restore_regs_all(reglist);
+}
+
+static void c2op_call_MACtoIR(int lm,int need_flags)
+{
+  if(need_flags)
+    emit_call((int)(lm?gteMACtoIR_lm1:gteMACtoIR_lm0));
+  else
+    emit_call((int)(lm?gteMACtoIR_lm1_nf:gteMACtoIR_lm0_nf));
+}
+
+static void c2op_call_rgb_func(void *func,int lm,int need_ir,int need_flags)
+{
+  emit_call((int)func);
+  // func is C code and trashes r0
+  emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0);
+  if(need_flags||need_ir)
+    c2op_call_MACtoIR(lm,need_flags);
+  emit_call((int)(need_flags?gteMACtoRGB:gteMACtoRGB_nf));
+}
+
+static void c2op_assemble(int i,struct regstat *i_regs)
 {
   signed char temp=get_reg(i_regs->regmap,-1);
   u_int c2op=source[i]&0x3f;
-  u_int hr,reglist=0;
-  int need_flags;
+  u_int hr,reglist_full=0,reglist;
+  int need_flags,need_ir;
   for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
+    if(i_regs->regmap[hr]>=0) reglist_full|=1<<hr;
   }
-  if(i==0||itype[i-1]!=C2OP)
-    save_regs(reglist);
+  reglist=reglist_full&CALLER_SAVE_REGS;
 
   if (gte_handlers[c2op]!=NULL) {
-    int cc=get_reg(i_regs->regmap,CCREG);
-    emit_movimm(source[i],1); // opcode
-    if (cc>=0&&gte_cycletab[c2op])
-      emit_addimm(cc,gte_cycletab[c2op]/2,cc); // XXX: could just adjust ccadj?
-    emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0); // cop2 regs
-    emit_writeword(1,(int)&psxRegs.code);
     need_flags=!(gte_unneeded[i+1]>>63); // +1 because of how liveness detection works
-    assem_debug("gte unneeded %016llx, need_flags %d\n",gte_unneeded[i+1],need_flags);
-#ifdef ARMv5_ONLY
-    // let's take more risk here
-    need_flags=need_flags&&gte_reads_flags;
+    need_ir=(gte_unneeded[i+1]&0xe00)!=0xe00;
+    assem_debug("gte op %08x, unneeded %016llx, need_flags %d, need_ir %d\n",
+      source[i],gte_unneeded[i+1],need_flags,need_ir);
+    if(new_dynarec_hacks&NDHACK_GTE_NO_FLAGS)
+      need_flags=0;
+    int shift = (source[i] >> 19) & 1;
+    int lm = (source[i] >> 10) & 1;
+    switch(c2op) {
+#ifndef DRC_DBG
+      case GTE_MVMVA: {
+#ifdef HAVE_ARMV5
+        int v  = (source[i] >> 15) & 3;
+        int cv = (source[i] >> 13) & 3;
+        int mx = (source[i] >> 17) & 3;
+        reglist=reglist_full&(CALLER_SAVE_REGS|0xf0); // +{r4-r7}
+        c2op_prologue(c2op,reglist);
+        /* r4,r5 = VXYZ(v) packed; r6 = &MX11(mx); r7 = &CV1(cv) */
+        if(v<3)
+          emit_ldrd(v*8,0,4);
+        else {
+          emit_movzwl_indexed(9*4,0,4);  // gteIR
+          emit_movzwl_indexed(10*4,0,6);
+          emit_movzwl_indexed(11*4,0,5);
+          emit_orrshl_imm(6,16,4);
+        }
+        if(mx<3)
+          emit_addimm(0,32*4+mx*8*4,6);
+        else
+          emit_readword((int)&zeromem_ptr,6);
+        if(cv<3)
+          emit_addimm(0,32*4+(cv*8+5)*4,7);
+        else
+          emit_readword((int)&zeromem_ptr,7);
+#ifdef __ARM_NEON__
+        emit_movimm(source[i],1); // opcode
+        emit_call((int)gteMVMVA_part_neon);
+        if(need_flags) {
+          emit_movimm(lm,1);
+          emit_call((int)gteMACtoIR_flags_neon);
+        }
+#else
+        if(cv==3&&shift)
+          emit_call((int)gteMVMVA_part_cv3sh12_arm);
+        else {
+          emit_movimm(shift,1);
+          emit_call((int)(need_flags?gteMVMVA_part_arm:gteMVMVA_part_nf_arm));
+        }
+        if(need_flags||need_ir)
+          c2op_call_MACtoIR(lm,need_flags);
 #endif
-    emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
+#else /* if not HAVE_ARMV5 */
+        c2op_prologue(c2op,reglist);
+        emit_movimm(source[i],1); // opcode
+        emit_writeword(1,(int)&psxRegs.code);
+        emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
+#endif
+        break;
+      }
+      case GTE_OP:
+        c2op_prologue(c2op,reglist);
+        emit_call((int)(shift?gteOP_part_shift:gteOP_part_noshift));
+        if(need_flags||need_ir) {
+          emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0);
+          c2op_call_MACtoIR(lm,need_flags);
+        }
+        break;
+      case GTE_DPCS:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteDPCS_part_shift:gteDPCS_part_noshift,lm,need_ir,need_flags);
+        break;
+      case GTE_INTPL:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteINTPL_part_shift:gteINTPL_part_noshift,lm,need_ir,need_flags);
+        break;
+      case GTE_SQR:
+        c2op_prologue(c2op,reglist);
+        emit_call((int)(shift?gteSQR_part_shift:gteSQR_part_noshift));
+        if(need_flags||need_ir) {
+          emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0);
+          c2op_call_MACtoIR(lm,need_flags);
+        }
+        break;
+      case GTE_DCPL:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(gteDCPL_part,lm,need_ir,need_flags);
+        break;
+      case GTE_GPF:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteGPF_part_shift:gteGPF_part_noshift,lm,need_ir,need_flags);
+        break;
+      case GTE_GPL:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteGPL_part_shift:gteGPL_part_noshift,lm,need_ir,need_flags);
+        break;
+#endif
+      default:
+        c2op_prologue(c2op,reglist);
+#ifdef DRC_DBG
+        emit_movimm(source[i],1); // opcode
+        emit_writeword(1,(int)&psxRegs.code);
+#endif
+        emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
+        break;
+    }
+    c2op_epilogue(c2op,reglist);
   }
-
-  if(i>=slen-1||itype[i+1]!=C2OP)
-    restore_regs(reglist);
 }
 
 void cop1_unusable(int i,struct regstat *i_regs)
@@ -4379,624 +4143,23 @@ void cop1_unusable(int i,struct regstat *i_regs)
 
 void cop1_assemble(int i,struct regstat *i_regs)
 {
-#ifndef DISABLE_COP1
-  // Check cop1 unusable
-  if(!cop1_usable) {
-    signed char rs=get_reg(i_regs->regmap,CSREG);
-    assert(rs>=0);
-    emit_testimm(rs,0x20000000);
-    int jaddr=(int)out;
-    emit_jeq(0);
-    add_stub(FP_STUB,jaddr,(int)out,i,rs,(int)i_regs,is_delayslot,0);
-    cop1_usable=1;
-  }
-  if (opcode2[i]==0) { // MFC1
-    signed char tl=get_reg(i_regs->regmap,rt1[i]);
-    if(tl>=0) {
-      emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],tl);
-      emit_readword_indexed(0,tl,tl);
-    }
-  }
-  else if (opcode2[i]==1) { // DMFC1
-    signed char tl=get_reg(i_regs->regmap,rt1[i]);
-    signed char th=get_reg(i_regs->regmap,rt1[i]|64);
-    if(tl>=0) {
-      emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],tl);
-      if(th>=0) emit_readword_indexed(4,tl,th);
-      emit_readword_indexed(0,tl,tl);
-    }
-  }
-  else if (opcode2[i]==4) { // MTC1
-    signed char sl=get_reg(i_regs->regmap,rs1[i]);
-    signed char temp=get_reg(i_regs->regmap,-1);
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    emit_writeword_indexed(sl,0,temp);
-  }
-  else if (opcode2[i]==5) { // DMTC1
-    signed char sl=get_reg(i_regs->regmap,rs1[i]);
-    signed char sh=rs1[i]>0?get_reg(i_regs->regmap,rs1[i]|64):sl;
-    signed char temp=get_reg(i_regs->regmap,-1);
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-    emit_writeword_indexed(sh,4,temp);
-    emit_writeword_indexed(sl,0,temp);
-  }
-  else if (opcode2[i]==2) // CFC1
-  {
-    signed char tl=get_reg(i_regs->regmap,rt1[i]);
-    if(tl>=0) {
-      u_int copr=(source[i]>>11)&0x1f;
-      if(copr==0) emit_readword((int)&FCR0,tl);
-      if(copr==31) emit_readword((int)&FCR31,tl);
-    }
-  }
-  else if (opcode2[i]==6) // CTC1
-  {
-    signed char sl=get_reg(i_regs->regmap,rs1[i]);
-    u_int copr=(source[i]>>11)&0x1f;
-    assert(sl>=0);
-    if(copr==31)
-    {
-      emit_writeword(sl,(int)&FCR31);
-      // Set the rounding mode
-      //FIXME
-      //char temp=get_reg(i_regs->regmap,-1);
-      //emit_andimm(sl,3,temp);
-      //emit_fldcw_indexed((int)&rounding_modes,temp);
-    }
-  }
-#else
   cop1_unusable(i, i_regs);
-#endif
 }
 
 void fconv_assemble_arm(int i,struct regstat *i_regs)
 {
-#ifndef DISABLE_COP1
-  signed char temp=get_reg(i_regs->regmap,-1);
-  assert(temp>=0);
-  // Check cop1 unusable
-  if(!cop1_usable) {
-    signed char rs=get_reg(i_regs->regmap,CSREG);
-    assert(rs>=0);
-    emit_testimm(rs,0x20000000);
-    int jaddr=(int)out;
-    emit_jeq(0);
-    add_stub(FP_STUB,jaddr,(int)out,i,rs,(int)i_regs,is_delayslot,0);
-    cop1_usable=1;
-  }
-  
-  #if(defined(__VFP_FP__) && !defined(__SOFTFP__)) 
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0d) { // trunc_w_s
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    emit_flds(temp,15);
-    emit_ftosizs(15,15); // float->int, truncate
-    if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f))
-      emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-    emit_fsts(15,temp);
-    return;
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0d) { // trunc_w_d
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-    emit_vldr(temp,7);
-    emit_ftosizd(7,13); // double->int, truncate
-    emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-    emit_fsts(13,temp);
-    return;
-  }
-  
-  if(opcode2[i]==0x14&&(source[i]&0x3f)==0x20) { // cvt_s_w
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    emit_flds(temp,13);
-    if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f))
-      emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-    emit_fsitos(13,15);
-    emit_fsts(15,temp);
-    return;
-  }
-  if(opcode2[i]==0x14&&(source[i]&0x3f)==0x21) { // cvt_d_w
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    emit_flds(temp,13);
-    emit_readword((int)&reg_cop1_double[(source[i]>>6)&0x1f],temp);
-    emit_fsitod(13,7);
-    emit_vstr(7,temp);
-    return;
-  }
-  
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x21) { // cvt_d_s
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    emit_flds(temp,13);
-    emit_readword((int)&reg_cop1_double[(source[i]>>6)&0x1f],temp);
-    emit_fcvtds(13,7);
-    emit_vstr(7,temp);
-    return;
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x20) { // cvt_s_d
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-    emit_vldr(temp,7);
-    emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-    emit_fcvtsd(7,13);
-    emit_fsts(13,temp);
-    return;
-  }
-  #endif
-  
-  // C emulation code
-  
-  u_int hr,reglist=0;
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
-  save_regs(reglist);
-  
-  if(opcode2[i]==0x14&&(source[i]&0x3f)==0x20) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_s_w);
-  }
-  if(opcode2[i]==0x14&&(source[i]&0x3f)==0x21) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_d_w);
-  }
-  if(opcode2[i]==0x15&&(source[i]&0x3f)==0x20) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_s_l);
-  }
-  if(opcode2[i]==0x15&&(source[i]&0x3f)==0x21) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_d_l);
-  }
-  
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x21) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_d_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x24) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_w_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x25) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_l_s);
-  }
-  
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x20) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_s_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x24) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_w_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x25) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)cvt_l_d);
-  }
-  
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x08) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)round_l_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x09) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)trunc_l_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0a) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)ceil_l_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0b) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)floor_l_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0c) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)round_w_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0d) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)trunc_w_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0e) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)ceil_w_s);
-  }
-  if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0f) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)floor_w_s);
-  }
-  
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x08) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)round_l_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x09) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)trunc_l_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0a) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)ceil_l_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0b) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)floor_l_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0c) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)round_w_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0d) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)trunc_w_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0e) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)ceil_w_d);
-  }
-  if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0f) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    emit_call((int)floor_w_d);
-  }
-  
-  restore_regs(reglist);
-#else
   cop1_unusable(i, i_regs);
-#endif
 }
 #define fconv_assemble fconv_assemble_arm
 
 void fcomp_assemble(int i,struct regstat *i_regs)
 {
-#ifndef DISABLE_COP1
-  signed char fs=get_reg(i_regs->regmap,FSREG);
-  signed char temp=get_reg(i_regs->regmap,-1);
-  assert(temp>=0);
-  // Check cop1 unusable
-  if(!cop1_usable) {
-    signed char cs=get_reg(i_regs->regmap,CSREG);
-    assert(cs>=0);
-    emit_testimm(cs,0x20000000);
-    int jaddr=(int)out;
-    emit_jeq(0);
-    add_stub(FP_STUB,jaddr,(int)out,i,cs,(int)i_regs,is_delayslot,0);
-    cop1_usable=1;
-  }
-  
-  if((source[i]&0x3f)==0x30) {
-    emit_andimm(fs,~0x800000,fs);
-    return;
-  }
-  
-  if((source[i]&0x3e)==0x38) {
-    // sf/ngle - these should throw exceptions for NaNs
-    emit_andimm(fs,~0x800000,fs);
-    return;
-  }
-  
-  #if(defined(__VFP_FP__) && !defined(__SOFTFP__)) 
-  if(opcode2[i]==0x10) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    emit_readword((int)&reg_cop1_simple[(source[i]>>16)&0x1f],HOST_TEMPREG);
-    emit_orimm(fs,0x800000,fs);
-    emit_flds(temp,14);
-    emit_flds(HOST_TEMPREG,15);
-    emit_fcmps(14,15);
-    emit_fmstat();
-    if((source[i]&0x3f)==0x31) emit_bicvc_imm(fs,0x800000,fs); // c_un_s
-    if((source[i]&0x3f)==0x32) emit_bicne_imm(fs,0x800000,fs); // c_eq_s
-    if((source[i]&0x3f)==0x33) {emit_bicne_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ueq_s
-    if((source[i]&0x3f)==0x34) emit_biccs_imm(fs,0x800000,fs); // c_olt_s
-    if((source[i]&0x3f)==0x35) {emit_biccs_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ult_s 
-    if((source[i]&0x3f)==0x36) emit_bichi_imm(fs,0x800000,fs); // c_ole_s
-    if((source[i]&0x3f)==0x37) {emit_bichi_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ule_s
-    if((source[i]&0x3f)==0x3a) emit_bicne_imm(fs,0x800000,fs); // c_seq_s
-    if((source[i]&0x3f)==0x3b) emit_bicne_imm(fs,0x800000,fs); // c_ngl_s
-    if((source[i]&0x3f)==0x3c) emit_biccs_imm(fs,0x800000,fs); // c_lt_s
-    if((source[i]&0x3f)==0x3d) emit_biccs_imm(fs,0x800000,fs); // c_nge_s
-    if((source[i]&0x3f)==0x3e) emit_bichi_imm(fs,0x800000,fs); // c_le_s
-    if((source[i]&0x3f)==0x3f) emit_bichi_imm(fs,0x800000,fs); // c_ngt_s
-    return;
-  }
-  if(opcode2[i]==0x11) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-    emit_readword((int)&reg_cop1_double[(source[i]>>16)&0x1f],HOST_TEMPREG);
-    emit_orimm(fs,0x800000,fs);
-    emit_vldr(temp,6);
-    emit_vldr(HOST_TEMPREG,7);
-    emit_fcmpd(6,7);
-    emit_fmstat();
-    if((source[i]&0x3f)==0x31) emit_bicvc_imm(fs,0x800000,fs); // c_un_d
-    if((source[i]&0x3f)==0x32) emit_bicne_imm(fs,0x800000,fs); // c_eq_d
-    if((source[i]&0x3f)==0x33) {emit_bicne_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ueq_d
-    if((source[i]&0x3f)==0x34) emit_biccs_imm(fs,0x800000,fs); // c_olt_d
-    if((source[i]&0x3f)==0x35) {emit_biccs_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ult_d
-    if((source[i]&0x3f)==0x36) emit_bichi_imm(fs,0x800000,fs); // c_ole_d
-    if((source[i]&0x3f)==0x37) {emit_bichi_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ule_d
-    if((source[i]&0x3f)==0x3a) emit_bicne_imm(fs,0x800000,fs); // c_seq_d
-    if((source[i]&0x3f)==0x3b) emit_bicne_imm(fs,0x800000,fs); // c_ngl_d
-    if((source[i]&0x3f)==0x3c) emit_biccs_imm(fs,0x800000,fs); // c_lt_d
-    if((source[i]&0x3f)==0x3d) emit_biccs_imm(fs,0x800000,fs); // c_nge_d
-    if((source[i]&0x3f)==0x3e) emit_bichi_imm(fs,0x800000,fs); // c_le_d
-    if((source[i]&0x3f)==0x3f) emit_bichi_imm(fs,0x800000,fs); // c_ngt_d
-    return;
-  }
-  #endif
-  
-  // C only
-  
-  u_int hr,reglist=0;
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
-  reglist&=~(1<<fs);
-  save_regs(reglist);
-  if(opcode2[i]==0x10) {
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_simple[(source[i]>>16)&0x1f],ARG2_REG);
-    if((source[i]&0x3f)==0x30) emit_call((int)c_f_s);
-    if((source[i]&0x3f)==0x31) emit_call((int)c_un_s);
-    if((source[i]&0x3f)==0x32) emit_call((int)c_eq_s);
-    if((source[i]&0x3f)==0x33) emit_call((int)c_ueq_s);
-    if((source[i]&0x3f)==0x34) emit_call((int)c_olt_s);
-    if((source[i]&0x3f)==0x35) emit_call((int)c_ult_s);
-    if((source[i]&0x3f)==0x36) emit_call((int)c_ole_s);
-    if((source[i]&0x3f)==0x37) emit_call((int)c_ule_s);
-    if((source[i]&0x3f)==0x38) emit_call((int)c_sf_s);
-    if((source[i]&0x3f)==0x39) emit_call((int)c_ngle_s);
-    if((source[i]&0x3f)==0x3a) emit_call((int)c_seq_s);
-    if((source[i]&0x3f)==0x3b) emit_call((int)c_ngl_s);
-    if((source[i]&0x3f)==0x3c) emit_call((int)c_lt_s);
-    if((source[i]&0x3f)==0x3d) emit_call((int)c_nge_s);
-    if((source[i]&0x3f)==0x3e) emit_call((int)c_le_s);
-    if((source[i]&0x3f)==0x3f) emit_call((int)c_ngt_s);
-  }
-  if(opcode2[i]==0x11) {
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    emit_readword((int)&reg_cop1_double[(source[i]>>16)&0x1f],ARG2_REG);
-    if((source[i]&0x3f)==0x30) emit_call((int)c_f_d);
-    if((source[i]&0x3f)==0x31) emit_call((int)c_un_d);
-    if((source[i]&0x3f)==0x32) emit_call((int)c_eq_d);
-    if((source[i]&0x3f)==0x33) emit_call((int)c_ueq_d);
-    if((source[i]&0x3f)==0x34) emit_call((int)c_olt_d);
-    if((source[i]&0x3f)==0x35) emit_call((int)c_ult_d);
-    if((source[i]&0x3f)==0x36) emit_call((int)c_ole_d);
-    if((source[i]&0x3f)==0x37) emit_call((int)c_ule_d);
-    if((source[i]&0x3f)==0x38) emit_call((int)c_sf_d);
-    if((source[i]&0x3f)==0x39) emit_call((int)c_ngle_d);
-    if((source[i]&0x3f)==0x3a) emit_call((int)c_seq_d);
-    if((source[i]&0x3f)==0x3b) emit_call((int)c_ngl_d);
-    if((source[i]&0x3f)==0x3c) emit_call((int)c_lt_d);
-    if((source[i]&0x3f)==0x3d) emit_call((int)c_nge_d);
-    if((source[i]&0x3f)==0x3e) emit_call((int)c_le_d);
-    if((source[i]&0x3f)==0x3f) emit_call((int)c_ngt_d);
-  }
-  restore_regs(reglist);
-  emit_loadreg(FSREG,fs);
-#else
   cop1_unusable(i, i_regs);
-#endif
 }
 
 void float_assemble(int i,struct regstat *i_regs)
 {
-#ifndef DISABLE_COP1
-  signed char temp=get_reg(i_regs->regmap,-1);
-  assert(temp>=0);
-  // Check cop1 unusable
-  if(!cop1_usable) {
-    signed char cs=get_reg(i_regs->regmap,CSREG);
-    assert(cs>=0);
-    emit_testimm(cs,0x20000000);
-    int jaddr=(int)out;
-    emit_jeq(0);
-    add_stub(FP_STUB,jaddr,(int)out,i,cs,(int)i_regs,is_delayslot,0);
-    cop1_usable=1;
-  }
-  
-  #if(defined(__VFP_FP__) && !defined(__SOFTFP__)) 
-  if((source[i]&0x3f)==6) // mov
-  {
-    if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-      if(opcode2[i]==0x10) {
-        emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-        emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],HOST_TEMPREG);
-        emit_readword_indexed(0,temp,temp);
-        emit_writeword_indexed(temp,0,HOST_TEMPREG);
-      }
-      if(opcode2[i]==0x11) {
-        emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-        emit_readword((int)&reg_cop1_double[(source[i]>>6)&0x1f],HOST_TEMPREG);
-        emit_vldr(temp,7);
-        emit_vstr(7,HOST_TEMPREG);
-      }
-    }
-    return;
-  }
-  
-  if((source[i]&0x3f)>3)
-  {
-    if(opcode2[i]==0x10) {
-      emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-      emit_flds(temp,15);
-      if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-        emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-      }
-      if((source[i]&0x3f)==4) // sqrt
-        emit_fsqrts(15,15);
-      if((source[i]&0x3f)==5) // abs
-        emit_fabss(15,15);
-      if((source[i]&0x3f)==7) // neg
-        emit_fnegs(15,15);
-      emit_fsts(15,temp);
-    }
-    if(opcode2[i]==0x11) {
-      emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-      emit_vldr(temp,7);
-      if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-        emit_readword((int)&reg_cop1_double[(source[i]>>6)&0x1f],temp);
-      }
-      if((source[i]&0x3f)==4) // sqrt
-        emit_fsqrtd(7,7);
-      if((source[i]&0x3f)==5) // abs
-        emit_fabsd(7,7);
-      if((source[i]&0x3f)==7) // neg
-        emit_fnegd(7,7);
-      emit_vstr(7,temp);
-    }
-    return;
-  }
-  if((source[i]&0x3f)<4)
-  {
-    if(opcode2[i]==0x10) {
-      emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],temp);
-    }
-    if(opcode2[i]==0x11) {
-      emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],temp);
-    }
-    if(((source[i]>>11)&0x1f)!=((source[i]>>16)&0x1f)) {
-      if(opcode2[i]==0x10) {
-        emit_readword((int)&reg_cop1_simple[(source[i]>>16)&0x1f],HOST_TEMPREG);
-        emit_flds(temp,15);
-        emit_flds(HOST_TEMPREG,13);
-        if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-          if(((source[i]>>16)&0x1f)!=((source[i]>>6)&0x1f)) {
-            emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-          }
-        }
-        if((source[i]&0x3f)==0) emit_fadds(15,13,15);
-        if((source[i]&0x3f)==1) emit_fsubs(15,13,15);
-        if((source[i]&0x3f)==2) emit_fmuls(15,13,15);
-        if((source[i]&0x3f)==3) emit_fdivs(15,13,15);
-        if(((source[i]>>16)&0x1f)==((source[i]>>6)&0x1f)) {
-          emit_fsts(15,HOST_TEMPREG);
-        }else{
-          emit_fsts(15,temp);
-        }
-      }
-      else if(opcode2[i]==0x11) {
-        emit_readword((int)&reg_cop1_double[(source[i]>>16)&0x1f],HOST_TEMPREG);
-        emit_vldr(temp,7);
-        emit_vldr(HOST_TEMPREG,6);
-        if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-          if(((source[i]>>16)&0x1f)!=((source[i]>>6)&0x1f)) {
-            emit_readword((int)&reg_cop1_double[(source[i]>>6)&0x1f],temp);
-          }
-        }
-        if((source[i]&0x3f)==0) emit_faddd(7,6,7);
-        if((source[i]&0x3f)==1) emit_fsubd(7,6,7);
-        if((source[i]&0x3f)==2) emit_fmuld(7,6,7);
-        if((source[i]&0x3f)==3) emit_fdivd(7,6,7);
-        if(((source[i]>>16)&0x1f)==((source[i]>>6)&0x1f)) {
-          emit_vstr(7,HOST_TEMPREG);
-        }else{
-          emit_vstr(7,temp);
-        }
-      }
-    }
-    else {
-      if(opcode2[i]==0x10) {
-        emit_flds(temp,15);
-        if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-          emit_readword((int)&reg_cop1_simple[(source[i]>>6)&0x1f],temp);
-        }
-        if((source[i]&0x3f)==0) emit_fadds(15,15,15);
-        if((source[i]&0x3f)==1) emit_fsubs(15,15,15);
-        if((source[i]&0x3f)==2) emit_fmuls(15,15,15);
-        if((source[i]&0x3f)==3) emit_fdivs(15,15,15);
-        emit_fsts(15,temp);
-      }
-      else if(opcode2[i]==0x11) {
-        emit_vldr(temp,7);
-        if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) {
-          emit_readword((int)&reg_cop1_double[(source[i]>>6)&0x1f],temp);
-        }
-        if((source[i]&0x3f)==0) emit_faddd(7,7,7);
-        if((source[i]&0x3f)==1) emit_fsubd(7,7,7);
-        if((source[i]&0x3f)==2) emit_fmuld(7,7,7);
-        if((source[i]&0x3f)==3) emit_fdivd(7,7,7);
-        emit_vstr(7,temp);
-      }
-    }
-    return;
-  }
-  #endif
-  
-  u_int hr,reglist=0;
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
-  }
-  if(opcode2[i]==0x10) { // Single precision
-    save_regs(reglist);
-    emit_readword((int)&reg_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG);
-    if((source[i]&0x3f)<4) {
-      emit_readword((int)&reg_cop1_simple[(source[i]>>16)&0x1f],ARG2_REG);
-      emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG3_REG);
-    }else{
-      emit_readword((int)&reg_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG);
-    }
-    switch(source[i]&0x3f)
-    {
-      case 0x00: emit_call((int)add_s);break;
-      case 0x01: emit_call((int)sub_s);break;
-      case 0x02: emit_call((int)mul_s);break;
-      case 0x03: emit_call((int)div_s);break;
-      case 0x04: emit_call((int)sqrt_s);break;
-      case 0x05: emit_call((int)abs_s);break;
-      case 0x06: emit_call((int)mov_s);break;
-      case 0x07: emit_call((int)neg_s);break;
-    }
-    restore_regs(reglist);
-  }
-  if(opcode2[i]==0x11) { // Double precision
-    save_regs(reglist);
-    emit_readword((int)&reg_cop1_double[(source[i]>>11)&0x1f],ARG1_REG);
-    if((source[i]&0x3f)<4) {
-      emit_readword((int)&reg_cop1_double[(source[i]>>16)&0x1f],ARG2_REG);
-      emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG3_REG);
-    }else{
-      emit_readword((int)&reg_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG);
-    }
-    switch(source[i]&0x3f)
-    {
-      case 0x00: emit_call((int)add_d);break;
-      case 0x01: emit_call((int)sub_d);break;
-      case 0x02: emit_call((int)mul_d);break;
-      case 0x03: emit_call((int)div_d);break;
-      case 0x04: emit_call((int)sqrt_d);break;
-      case 0x05: emit_call((int)abs_d);break;
-      case 0x06: emit_call((int)mov_d);break;
-      case 0x07: emit_call((int)neg_d);break;
-    }
-    restore_regs(reglist);
-  }
-#else
   cop1_unusable(i, i_regs);
-#endif
 }
 
 void multdiv_assemble_arm(int i,struct regstat *i_regs)
@@ -5053,9 +4216,16 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
         emit_negmi(remainder,remainder); // .. remainder for div0 case (will be negated back after jump)
         emit_movs(d2,HOST_TEMPREG);
         emit_jeq((int)out+52); // Division by zero
-        emit_negmi(HOST_TEMPREG,HOST_TEMPREG);
+        emit_negsmi(HOST_TEMPREG,HOST_TEMPREG);
+#ifdef HAVE_ARMV5
         emit_clz(HOST_TEMPREG,quotient);
         emit_shl(HOST_TEMPREG,quotient,HOST_TEMPREG);
+#else
+        emit_movimm(0,quotient);
+        emit_addpl_imm(quotient,1,quotient);
+        emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
+        emit_jns((int)out-2*4);
+#endif
         emit_orimm(quotient,1<<31,quotient);
         emit_shr(quotient,quotient,quotient);
         emit_cmp(remainder,HOST_TEMPREG);
@@ -5082,9 +4252,17 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
         emit_movimm(0xffffffff,quotient); // div0 case
         emit_test(d2,d2);
         emit_jeq((int)out+40); // Division by zero
+#ifdef HAVE_ARMV5
         emit_clz(d2,HOST_TEMPREG);
         emit_movimm(1<<31,quotient);
         emit_shl(d2,HOST_TEMPREG,d2);
+#else
+        emit_movimm(0,HOST_TEMPREG);
+        emit_addpl_imm(HOST_TEMPREG,1,HOST_TEMPREG);
+        emit_lslpls_imm(d2,1,d2);
+        emit_jns((int)out-2*4);
+        emit_movimm(1<<31,quotient);
+#endif
         emit_shr(quotient,HOST_TEMPREG,quotient);
         emit_cmp(remainder,d2);
         emit_subcs(remainder,d2,remainder);
@@ -5094,183 +4272,7 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
       }
     }
     else // 64-bit
-#ifndef FORCE32
-    {
-      if(opcode2[i]==0x1C) // DMULT
-      {
-        assert(opcode2[i]!=0x1C);
-        signed char m1h=get_reg(i_regs->regmap,rs1[i]|64);
-        signed char m1l=get_reg(i_regs->regmap,rs1[i]);
-        signed char m2h=get_reg(i_regs->regmap,rs2[i]|64);
-        signed char m2l=get_reg(i_regs->regmap,rs2[i]);
-        assert(m1h>=0);
-        assert(m2h>=0);
-        assert(m1l>=0);
-        assert(m2l>=0);
-        emit_pushreg(m2h);
-        emit_pushreg(m2l);
-        emit_pushreg(m1h);
-        emit_pushreg(m1l);
-        emit_call((int)&mult64);
-        emit_popreg(m1l);
-        emit_popreg(m1h);
-        emit_popreg(m2l);
-        emit_popreg(m2h);
-        signed char hih=get_reg(i_regs->regmap,HIREG|64);
-        signed char hil=get_reg(i_regs->regmap,HIREG);
-        if(hih>=0) emit_loadreg(HIREG|64,hih);
-        if(hil>=0) emit_loadreg(HIREG,hil);
-        signed char loh=get_reg(i_regs->regmap,LOREG|64);
-        signed char lol=get_reg(i_regs->regmap,LOREG);
-        if(loh>=0) emit_loadreg(LOREG|64,loh);
-        if(lol>=0) emit_loadreg(LOREG,lol);
-      }
-      if(opcode2[i]==0x1D) // DMULTU
-      {
-        signed char m1h=get_reg(i_regs->regmap,rs1[i]|64);
-        signed char m1l=get_reg(i_regs->regmap,rs1[i]);
-        signed char m2h=get_reg(i_regs->regmap,rs2[i]|64);
-        signed char m2l=get_reg(i_regs->regmap,rs2[i]);
-        assert(m1h>=0);
-        assert(m2h>=0);
-        assert(m1l>=0);
-        assert(m2l>=0);
-        save_regs(0x100f);
-        if(m1l!=0) emit_mov(m1l,0);
-        if(m1h==0) emit_readword((int)&dynarec_local,1);
-        else if(m1h>1) emit_mov(m1h,1);
-        if(m2l<2) emit_readword((int)&dynarec_local+m2l*4,2);
-        else if(m2l>2) emit_mov(m2l,2);
-        if(m2h<3) emit_readword((int)&dynarec_local+m2h*4,3);
-        else if(m2h>3) emit_mov(m2h,3);
-        emit_call((int)&multu64);
-        restore_regs(0x100f);
-        signed char hih=get_reg(i_regs->regmap,HIREG|64);
-        signed char hil=get_reg(i_regs->regmap,HIREG);
-        signed char loh=get_reg(i_regs->regmap,LOREG|64);
-        signed char lol=get_reg(i_regs->regmap,LOREG);
-        /*signed char temp=get_reg(i_regs->regmap,-1);
-        signed char rh=get_reg(i_regs->regmap,HIREG|64);
-        signed char rl=get_reg(i_regs->regmap,HIREG);
-        assert(m1h>=0);
-        assert(m2h>=0);
-        assert(m1l>=0);
-        assert(m2l>=0);
-        assert(temp>=0);
-        //emit_mov(m1l,EAX);
-        //emit_mul(m2l);
-        emit_umull(rl,rh,m1l,m2l);
-        emit_storereg(LOREG,rl);
-        emit_mov(rh,temp);
-        //emit_mov(m1h,EAX);
-        //emit_mul(m2l);
-        emit_umull(rl,rh,m1h,m2l);
-        emit_adds(rl,temp,temp);
-        emit_adcimm(rh,0,rh);
-        emit_storereg(HIREG,rh);
-        //emit_mov(m2h,EAX);
-        //emit_mul(m1l);
-        emit_umull(rl,rh,m1l,m2h);
-        emit_adds(rl,temp,temp);
-        emit_adcimm(rh,0,rh);
-        emit_storereg(LOREG|64,temp);
-        emit_mov(rh,temp);
-        //emit_mov(m2h,EAX);
-        //emit_mul(m1h);
-        emit_umull(rl,rh,m1h,m2h);
-        emit_adds(rl,temp,rl);
-        emit_loadreg(HIREG,temp);
-        emit_adcimm(rh,0,rh);
-        emit_adds(rl,temp,rl);
-        emit_adcimm(rh,0,rh);
-        // DEBUG
-        /*
-        emit_pushreg(m2h);
-        emit_pushreg(m2l);
-        emit_pushreg(m1h);
-        emit_pushreg(m1l);
-        emit_call((int)&multu64);
-        emit_popreg(m1l);
-        emit_popreg(m1h);
-        emit_popreg(m2l);
-        emit_popreg(m2h);
-        signed char hih=get_reg(i_regs->regmap,HIREG|64);
-        signed char hil=get_reg(i_regs->regmap,HIREG);
-        if(hih>=0) emit_loadreg(HIREG|64,hih);  // DEBUG
-        if(hil>=0) emit_loadreg(HIREG,hil);  // DEBUG
-        */
-        // Shouldn't be necessary
-        //char loh=get_reg(i_regs->regmap,LOREG|64);
-        //char lol=get_reg(i_regs->regmap,LOREG);
-        //if(loh>=0) emit_loadreg(LOREG|64,loh);
-        //if(lol>=0) emit_loadreg(LOREG,lol);
-      }
-      if(opcode2[i]==0x1E) // DDIV
-      {
-        signed char d1h=get_reg(i_regs->regmap,rs1[i]|64);
-        signed char d1l=get_reg(i_regs->regmap,rs1[i]);
-        signed char d2h=get_reg(i_regs->regmap,rs2[i]|64);
-        signed char d2l=get_reg(i_regs->regmap,rs2[i]);
-        assert(d1h>=0);
-        assert(d2h>=0);
-        assert(d1l>=0);
-        assert(d2l>=0);
-        save_regs(0x100f);
-        if(d1l!=0) emit_mov(d1l,0);
-        if(d1h==0) emit_readword((int)&dynarec_local,1);
-        else if(d1h>1) emit_mov(d1h,1);
-        if(d2l<2) emit_readword((int)&dynarec_local+d2l*4,2);
-        else if(d2l>2) emit_mov(d2l,2);
-        if(d2h<3) emit_readword((int)&dynarec_local+d2h*4,3);
-        else if(d2h>3) emit_mov(d2h,3);
-        emit_call((int)&div64);
-        restore_regs(0x100f);
-        signed char hih=get_reg(i_regs->regmap,HIREG|64);
-        signed char hil=get_reg(i_regs->regmap,HIREG);
-        signed char loh=get_reg(i_regs->regmap,LOREG|64);
-        signed char lol=get_reg(i_regs->regmap,LOREG);
-        if(hih>=0) emit_loadreg(HIREG|64,hih);
-        if(hil>=0) emit_loadreg(HIREG,hil);
-        if(loh>=0) emit_loadreg(LOREG|64,loh);
-        if(lol>=0) emit_loadreg(LOREG,lol);
-      }
-      if(opcode2[i]==0x1F) // DDIVU
-      {
-      //u_int hr,reglist=0;
-      //for(hr=0;hr<HOST_REGS;hr++) {
-      //  if(i_regs->regmap[hr]>=0 && (i_regs->regmap[hr]&62)!=HIREG) reglist|=1<<hr;
-      //}
-        signed char d1h=get_reg(i_regs->regmap,rs1[i]|64);
-        signed char d1l=get_reg(i_regs->regmap,rs1[i]);
-        signed char d2h=get_reg(i_regs->regmap,rs2[i]|64);
-        signed char d2l=get_reg(i_regs->regmap,rs2[i]);
-        assert(d1h>=0);
-        assert(d2h>=0);
-        assert(d1l>=0);
-        assert(d2l>=0);
-        save_regs(0x100f);
-        if(d1l!=0) emit_mov(d1l,0);
-        if(d1h==0) emit_readword((int)&dynarec_local,1);
-        else if(d1h>1) emit_mov(d1h,1);
-        if(d2l<2) emit_readword((int)&dynarec_local+d2l*4,2);
-        else if(d2l>2) emit_mov(d2l,2);
-        if(d2h<3) emit_readword((int)&dynarec_local+d2h*4,3);
-        else if(d2h>3) emit_mov(d2h,3);
-        emit_call((int)&divu64);
-        restore_regs(0x100f);
-        signed char hih=get_reg(i_regs->regmap,HIREG|64);
-        signed char hil=get_reg(i_regs->regmap,HIREG);
-        signed char loh=get_reg(i_regs->regmap,LOREG|64);
-        signed char lol=get_reg(i_regs->regmap,LOREG);
-        if(hih>=0) emit_loadreg(HIREG|64,hih);
-        if(hil>=0) emit_loadreg(HIREG,hil);
-        if(loh>=0) emit_loadreg(LOREG|64,loh);
-        if(lol>=0) emit_loadreg(LOREG,lol);
-      }
-    }
-#else
-    assert(0);
-#endif
+      assert(0);
   }
   else
   {
@@ -5315,7 +4317,7 @@ void do_miniht_jump(int rs,int rh,int ht) {
 }
 
 void do_miniht_insert(u_int return_address,int rt,int temp) {
-  #ifdef ARMv5_ONLY
+  #ifndef HAVE_ARMV7
   emit_movimm(return_address,rt); // PC into link register
   add_to_linker((int)out,return_address,1);
   emit_pcreladdr(temp);
@@ -5331,32 +4333,6 @@ void do_miniht_insert(u_int return_address,int rt,int temp) {
   #endif
 }
 
-// Sign-extend to 64 bits and write out upper half of a register
-// This is useful where we have a 32-bit value in a register, and want to
-// keep it in a 32-bit register, but can't guarantee that it won't be read
-// as a 64-bit value later.
-void wb_sx(signed char pre[],signed char entry[],uint64_t dirty,uint64_t is32_pre,uint64_t is32,uint64_t u,uint64_t uu)
-{
-#ifndef FORCE32
-  if(is32_pre==is32) return;
-  int hr,reg;
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(hr!=EXCLUDE_REG) {
-      //if(pre[hr]==entry[hr]) {
-        if((reg=pre[hr])>=0) {
-          if((dirty>>hr)&1) {
-            if( ((is32_pre&~is32&~uu)>>reg)&1 ) {
-              emit_sarimm(hr,31,HOST_TEMPREG);
-              emit_storereg(reg|64,HOST_TEMPREG);
-            }
-          }
-        }
-      //}
-    }
-  }
-#endif
-}
-
 void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t is32_pre,uint64_t u,uint64_t uu)
 {
   //if(dirty_pre==dirty) return;
@@ -5454,7 +4430,7 @@ void do_clear_cache()
       for(j=0;j<32;j++) 
       {
         if(bitmap&(1<<j)) {
-          start=BASE_ADDR+i*131072+j*4096;
+          start=(u_int)BASE_ADDR+i*131072+j*4096;
           end=start+4095;
           j++;
           while(j<32) {
@@ -5474,13 +4450,7 @@ void do_clear_cache()
 }
 
 // CPU-architecture-specific initialization
-void arch_init() {
-#ifndef DISABLE_COP1
-  rounding_modes[0]=0x0<<22; // round
-  rounding_modes[1]=0x3<<22; // trunc
-  rounding_modes[2]=0x1<<22; // ceil
-  rounding_modes[3]=0x2<<22; // floor
-#endif
+static void arch_init() {
 }
 
 // vim:shiftwidth=2:expandtab