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=84370089f7625e2d699a912c6fd0f9962843f0ce;hp=96980375a58f6a6f8993fcb3c0788de49c384b50;hb=c7b746f050d33e3bb5b06707dcc52f8699858e40;hpb=911f2d55ac72cab3417983708cbb6a5fde2f13db

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 96980375..84370089 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -41,12 +41,22 @@
 #include "assem_arm.h"
 #endif
 
+#ifdef __BLACKBERRY_QNX__
+#undef __clear_cache
+#define __clear_cache(start,end) msync(start, (size_t)((void*)end - (void*)start), MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
+#elif defined(__MACH__)
+#include <libkern/OSCacheControl.h>
+#define __clear_cache mach_clear_cache
+static void __clear_cache(void *start, void *end) {
+  size_t len = (char *)end - (char *)start;
+  sys_dcache_flush(start, len);
+  sys_icache_invalidate(start, len);
+}
+#endif
+
 #define MAXBLOCK 4096
 #define MAX_OUTPUT_BLOCK_SIZE 262144
 
-int cycle_multiplier; // 100 for 1.0
-#define CLOCK_ADJUST(x) (((x) * cycle_multiplier + 50) / 100)
-
 struct regstat
 {
   signed char regmap_entry[HOST_REGS];
@@ -59,7 +69,8 @@ struct regstat
   uint64_t uu;
   u_int wasconst;
   u_int isconst;
-  uint64_t constmap[HOST_REGS];
+  u_int loadedconst;             // host regs that have constants loaded
+  u_int waswritten;              // MIPS regs that were used as store base before
 };
 
 struct ll_entry
@@ -90,7 +101,6 @@ struct ll_entry
   static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
   static uint64_t gte_rt[MAXBLOCK];
   static uint64_t gte_unneeded[MAXBLOCK];
-  static int gte_reads_flags; // gte flag read encountered
   static u_int smrv[32]; // speculated MIPS register values
   static u_int smrv_strong; // mask or regs that are likely to have correct values
   static u_int smrv_weak; // same, but somewhat less likely
@@ -108,11 +118,10 @@ struct ll_entry
   uint64_t p32[MAXBLOCK];
   uint64_t pr32[MAXBLOCK];
   signed char regmap_pre[MAXBLOCK][HOST_REGS];
-  signed char regmap[MAXBLOCK][HOST_REGS];
-  signed char regmap_entry[MAXBLOCK][HOST_REGS];
-  uint64_t constmap[MAXBLOCK][HOST_REGS];
-  struct regstat regs[MAXBLOCK];
-  struct regstat branch_regs[MAXBLOCK];
+  static uint64_t current_constmap[HOST_REGS];
+  static uint64_t constmap[MAXBLOCK][HOST_REGS];
+  static struct regstat regs[MAXBLOCK];
+  static struct regstat branch_regs[MAXBLOCK];
   signed char minimum_free_regs[MAXBLOCK];
   u_int needed_reg[MAXBLOCK];
   uint64_t requires_32bit[MAXBLOCK];
@@ -143,7 +152,13 @@ struct ll_entry
   static const u_int using_tlb=0;
 #endif
   int new_dynarec_did_compile;
+  int new_dynarec_hacks;
   u_int stop_after_jal;
+#ifndef RAM_FIXED
+  static u_int ram_offset;
+#else
+  static const u_int ram_offset=0;
+#endif
   extern u_char restore_candidate[512];
   extern int cycle_count;
 
@@ -277,6 +292,16 @@ int tracedebug=0;
 
 //#define DEBUG_CYCLE_COUNT 1
 
+#define NO_CYCLE_PENALTY_THR 12
+
+int cycle_multiplier; // 100 for 1.0
+
+static int CLOCK_ADJUST(int x)
+{
+  int s=(x>>31)|1;
+  return (x * cycle_multiplier + s * 50) / 100;
+}
+
 static void tlb_hacks()
 {
 #ifndef DISABLE_TLB
@@ -341,6 +366,7 @@ static u_int get_page(u_int vaddr)
   return page;
 }
 
+#ifndef PCSX
 static u_int get_vpage(u_int vaddr)
 {
   u_int vpage=(vaddr^0x80000000)>>12;
@@ -350,6 +376,13 @@ static u_int get_vpage(u_int vaddr)
   if(vpage>2048) vpage=2048+(vpage&2047);
   return vpage;
 }
+#else
+// no virtual mem in PCSX
+static u_int get_vpage(u_int vaddr)
+{
+  return get_page(vaddr);
+}
+#endif
 
 // Get address from virtual address
 // This is called from the recompiled JR/JALR instructions
@@ -590,11 +623,11 @@ void set_const(struct regstat *cur,signed char reg,uint64_t value)
   for (hr=0;hr<HOST_REGS;hr++) {
     if(cur->regmap[hr]==reg) {
       cur->isconst|=1<<hr;
-      cur->constmap[hr]=value;
+      current_constmap[hr]=value;
     }
     else if((cur->regmap[hr]^64)==reg) {
       cur->isconst|=1<<hr;
-      cur->constmap[hr]=value>>32;
+      current_constmap[hr]=value>>32;
     }
   }
 }
@@ -628,10 +661,10 @@ uint64_t get_const(struct regstat *cur,signed char reg)
   if(!reg) return 0;
   for (hr=0;hr<HOST_REGS;hr++) {
     if(cur->regmap[hr]==reg) {
-      return cur->constmap[hr];
+      return current_constmap[hr];
     }
   }
-  printf("Unknown constant in r%d\n",reg);
+  SysPrintf("Unknown constant in r%d\n",reg);
   exit(1);
 }
 
@@ -1198,7 +1231,7 @@ void invalidate_block(u_int block)
     if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
       get_bounds((int)head->addr,&start,&end);
       //printf("start: %x end: %x\n",start,end);
-      if(page<2048&&start>=0x80000000&&end<0x80000000+RAM_SIZE) {
+      if(page<2048&&start>=(u_int)rdram&&end<(u_int)rdram+RAM_SIZE) {
         if(((start-(u_int)rdram)>>12)<=page&&((end-1-(u_int)rdram)>>12)>=page) {
           if((((start-(u_int)rdram)>>12)&2047)<first) first=((start-(u_int)rdram)>>12)&2047;
           if((((end-1-(u_int)rdram)>>12)&2047)>last) last=((end-1-(u_int)rdram)>>12)&2047;
@@ -1224,14 +1257,15 @@ void invalidate_addr(u_int addr)
   //static int rhits;
   // this check is done by the caller
   //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
-  u_int page=get_page(addr);
+  u_int page=get_vpage(addr);
   if(page<2048) { // RAM
     struct ll_entry *head;
     u_int addr_min=~0, addr_max=0;
-    int mask=RAM_SIZE-1;
+    u_int mask=RAM_SIZE-1;
+    u_int addr_main=0x80000000|(addr&mask);
     int pg1;
-    inv_code_start=addr&~0xfff;
-    inv_code_end=addr|0xfff;
+    inv_code_start=addr_main&~0xfff;
+    inv_code_end=addr_main|0xfff;
     pg1=page;
     if (pg1>0) {
       // must check previous page too because of spans..
@@ -1242,11 +1276,15 @@ void invalidate_addr(u_int addr)
       for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
         u_int start,end;
         get_bounds((int)head->addr,&start,&end);
-        if((start&mask)<=(addr&mask)&&(addr&mask)<(end&mask)) {
+        if(ram_offset) {
+          start-=ram_offset;
+          end-=ram_offset;
+        }
+        if(start<=addr_main&&addr_main<end) {
           if(start<addr_min) addr_min=start;
           if(end>addr_max) addr_max=end;
         }
-        else if(addr<start) {
+        else if(addr_main<start) {
           if(start<inv_code_end)
             inv_code_end=start-1;
         }
@@ -1263,11 +1301,11 @@ void invalidate_addr(u_int addr)
       return;
     }
     else {
-      inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);//rhits);
-    }
-    //rhits=0;
-    if(page!=0) // FIXME: don't know what's up with page 0 (Klonoa)
+      inv_code_start=(addr&~mask)|(inv_code_start&mask);
+      inv_code_end=(addr&~mask)|(inv_code_end&mask);
+      inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
       return;
+    }
   }
 #endif
   invalidate_block(addr>>12);
@@ -1967,7 +2005,7 @@ void delayslot_alloc(struct regstat *current,int i)
     case HLECALL:
     case SPAN:
       assem_debug("jump in the delay slot.  this shouldn't happen.\n");//exit(1);
-      printf("Disabled speculative precompilation\n");
+      SysPrintf("Disabled speculative precompilation\n");
       stop_after_jal=1;
       break;
     case IMM16:
@@ -2906,6 +2944,10 @@ void load_assemble(int i,struct regstat *i_regs)
         jaddr=emit_fastpath_cmp_jump(i,addr,&fastload_reg_override);
       }
     }
+    else if(ram_offset&&memtarget) {
+      emit_addimm(addr,ram_offset,HOST_TEMPREG);
+      fastload_reg_override=HOST_TEMPREG;
+    }
   }else{ // using tlb
     int x=0;
     if (opcode[i]==0x20||opcode[i]==0x24) x=3; // LB/LBU
@@ -2971,7 +3013,7 @@ void load_assemble(int i,struct regstat *i_regs)
             gen_tlb_addr_r(a,map);
             emit_movswl_indexed(x,a,tl);
           }else{
-            #ifdef RAM_OFFSET
+            #if 1 //def RAM_OFFSET
             emit_movswl_indexed(x,a,tl);
             #else
             emit_movswl_indexed((int)rdram-0x80000000+x,a,tl);
@@ -3058,7 +3100,7 @@ void load_assemble(int i,struct regstat *i_regs)
             gen_tlb_addr_r(a,map);
             emit_movzwl_indexed(x,a,tl);
           }else{
-            #ifdef RAM_OFFSET
+            #if 1 //def RAM_OFFSET
             emit_movzwl_indexed(x,a,tl);
             #else
             emit_movzwl_indexed((int)rdram-0x80000000+x,a,tl);
@@ -3215,6 +3257,10 @@ void store_assemble(int i,struct regstat *i_regs)
         jaddr=emit_fastpath_cmp_jump(i,addr,&faststore_reg_override);
       #endif
     }
+    else if(ram_offset&&memtarget) {
+      emit_addimm(addr,ram_offset,HOST_TEMPREG);
+      faststore_reg_override=HOST_TEMPREG;
+    }
   }else{ // using tlb
     int x=0;
     if (opcode[i]==0x28) x=3; // SB
@@ -3259,7 +3305,8 @@ void store_assemble(int i,struct regstat *i_regs)
         gen_tlb_addr_w(a,map);
         emit_writehword_indexed(tl,x,a);
       }else
-        emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
+        //emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
+        emit_writehword_indexed(tl,x,a);
     }
     type=STOREH_STUB;
   }
@@ -3298,7 +3345,7 @@ void store_assemble(int i,struct regstat *i_regs)
     jaddr=0;
   }
 #endif
-  if(!using_tlb) {
+  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
     if(!c||memtarget) {
       #ifdef DESTRUCTIVE_SHIFT
       // The x86 shift operation is 'destructive'; it overwrites the
@@ -3321,10 +3368,24 @@ void store_assemble(int i,struct regstat *i_regs)
       #endif
     }
   }
+  u_int addr_val=constmap[i][s]+offset;
   if(jaddr) {
     add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
   } else if(c&&!memtarget) {
-    inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
+    inline_writestub(type,i,addr_val,i_regs->regmap,rs2[i],ccadj[i],reglist);
+  }
+  // basic current block modification detection..
+  // not looking back as that should be in mips cache already
+  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
+    if(i_regs->regmap==regs[i].regmap) {
+      load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i);
+      wb_dirtys(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty);
+      emit_movimm(start+i*4+4,0);
+      emit_writeword(0,(int)&pcaddr);
+      emit_jmp((int)do_interrupt);
+    }
   }
   //if(opcode[i]==0x2B || opcode[i]==0x3F)
   //if(opcode[i]==0x2B || opcode[i]==0x28)
@@ -3572,7 +3633,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
   }
   if(!c||!memtarget)
     add_stub(STORELR_STUB,jaddr,(int)out,i,(int)i_regs,temp,ccadj[i],reglist);
-  if(!using_tlb) {
+  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
     #ifdef RAM_OFFSET
     int map=get_reg(i_regs->regmap,ROREG);
     if(map<0) map=HOST_TEMPREG;
@@ -3750,7 +3811,7 @@ void c1ls_assemble(int i,struct regstat *i_regs)
     emit_writedword_indexed_tlb(th,tl,0,offset||c||s<0?temp:s,map,temp);
     type=STORED_STUB;
   }
-  if(!using_tlb) {
+  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
     if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
       #ifndef DESTRUCTIVE_SHIFT
       temp=offset||c||s<0?ar:s;
@@ -3849,6 +3910,10 @@ void c2ls_assemble(int i,struct regstat *i_regs)
     if(!c) {
       jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
     }
+    else if(ram_offset&&memtarget) {
+      emit_addimm(ar,ram_offset,HOST_TEMPREG);
+      fastio_reg_override=HOST_TEMPREG;
+    }
     if (opcode[i]==0x32) { // LWC2
       #ifdef HOST_IMM_ADDR32
       if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl);
@@ -3869,7 +3934,8 @@ void c2ls_assemble(int i,struct regstat *i_regs)
   }
   if(jaddr2)
     add_stub(type,jaddr2,(int)out,i,ar,(int)i_regs,ccadj[i],reglist);
-  if (opcode[i]==0x3a) { // SWC2
+  if(opcode[i]==0x3a) // SWC2
+  if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
 #if defined(HOST_IMM8)
     int ir=get_reg(i_regs->regmap,INVCP);
     assert(ir>=0);
@@ -4019,7 +4085,7 @@ void ds_assemble(int i,struct regstat *i_regs)
     case CJUMP:
     case SJUMP:
     case FJUMP:
-      printf("Jump in the delay slot.  This is probably a bug.\n");
+      SysPrintf("Jump in the delay slot.  This is probably a bug.\n");
   }
   is_delayslot=0;
 }
@@ -4257,6 +4323,7 @@ void address_generation(int i,struct regstat *i_regs,signed char entry[])
                  (using_tlb&&((signed int)constmap[i][rs]+offset)>=(signed int)0xC0000000))
               #endif
               emit_movimm(constmap[i][rs]+offset,ra);
+              regs[i].loadedconst|=1<<ra;
             }
           } // else did it in the previous cycle
         } // else load_consts already did it
@@ -4317,6 +4384,7 @@ void address_generation(int i,struct regstat *i_regs,signed char entry[])
              (using_tlb&&((signed int)constmap[i+1][rs]+offset)>=(signed int)0xC0000000))
           #endif
           emit_movimm(constmap[i+1][rs]+offset,ra);
+          regs[i+1].loadedconst|=1<<ra;
         }
       }
       else if(rs1[i+1]==0) {
@@ -4385,22 +4453,51 @@ int get_final_value(int hr, int i, int *value)
 // Load registers with known constants
 void load_consts(signed char pre[],signed char regmap[],int is32,int i)
 {
-  int hr;
+  int hr,hr2;
+  // propagate loaded constant flags
+  if(i==0||bt[i])
+    regs[i].loadedconst=0;
+  else {
+    for(hr=0;hr<HOST_REGS;hr++) {
+      if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
+         &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
+      {
+        regs[i].loadedconst|=1<<hr;
+      }
+    }
+  }
   // Load 32-bit regs
   for(hr=0;hr<HOST_REGS;hr++) {
     if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
       //if(entry[hr]!=regmap[hr]) {
-      if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
+      if(!((regs[i].loadedconst>>hr)&1)) {
         if(((regs[i].isconst>>hr)&1)&&regmap[hr]<64&&regmap[hr]>0) {
-          int value;
+          int value,similar=0;
           if(get_final_value(hr,i,&value)) {
-            if(value==0) {
+            // see if some other register has similar value
+            for(hr2=0;hr2<HOST_REGS;hr2++) {
+              if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
+                if(is_similar_value(value,constmap[i][hr2])) {
+                  similar=1;
+                  break;
+                }
+              }
+            }
+            if(similar) {
+              int value2;
+              if(get_final_value(hr2,i,&value2)) // is this needed?
+                emit_movimm_from(value2,hr2,value,hr);
+              else
+                emit_movimm(value,hr);
+            }
+            else if(value==0) {
               emit_zeroreg(hr);
             }
             else {
               emit_movimm(value,hr);
             }
           }
+          regs[i].loadedconst|=1<<hr;
         }
       }
     }
@@ -4881,7 +4978,7 @@ void ds_assemble_entry(int i)
     case CJUMP:
     case SJUMP:
     case FJUMP:
-      printf("Jump in the delay slot.  This is probably a bug.\n");
+      SysPrintf("Jump in the delay slot.  This is probably a bug.\n");
   }
   store_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
   load_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
@@ -4899,6 +4996,7 @@ void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
   int count;
   int jaddr;
   int idle=0;
+  int t=0;
   if(itype[i]==RJUMP)
   {
     *adj=0;
@@ -4906,7 +5004,7 @@ void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
   //if(ba[i]>=start && ba[i]<(start+slen*4))
   if(internal_branch(branch_regs[i].is32,ba[i]))
   {
-    int t=(ba[i]-start)>>2;
+    t=(ba[i]-start)>>2;
     if(is_ds[t]) *adj=-1; // Branch into delay slot adds an extra cycle
     else *adj=ccadj[t];
   }
@@ -4925,7 +5023,14 @@ void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
     emit_jmp(0);
   }
   else if(*adj==0||invert) {
-    emit_addimm_and_set_flags(CLOCK_ADJUST(count+2),HOST_CCREG);
+    int cycles=CLOCK_ADJUST(count+2);
+    // faster loop HACK
+    if (t&&*adj) {
+      int rel=t-i;
+      if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
+        cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
+    }
+    emit_addimm_and_set_flags(cycles,HOST_CCREG);
     jaddr=(int)out;
     emit_jns(0);
   }
@@ -5151,7 +5256,7 @@ void do_ccstub(int n)
       }
       emit_writeword(r,(int)&pcaddr);
     }
-    else {printf("Unknown branch type in do_ccstub\n");exit(1);}
+    else {SysPrintf("Unknown branch type in do_ccstub\n");exit(1);}
   }
   // Update cycle count
   assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
@@ -6708,7 +6813,7 @@ static void pagespan_ds()
     case CJUMP:
     case SJUMP:
     case FJUMP:
-      printf("Jump in the delay slot.  This is probably a bug.\n");
+      SysPrintf("Jump in the delay slot.  This is probably a bug.\n");
   }
   int btaddr=get_reg(regs[0].regmap,BTREG);
   if(btaddr<0) {
@@ -6737,16 +6842,20 @@ void unneeded_registers(int istart,int iend,int r)
 {
   int i;
   uint64_t u,uu,gte_u,b,bu,gte_bu;
-  uint64_t temp_u,temp_uu,temp_gte_u;
+  uint64_t temp_u,temp_uu,temp_gte_u=0;
   uint64_t tdep;
+  uint64_t gte_u_unknown=0;
+  if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
+    gte_u_unknown=~0ll;
   if(iend==slen-1) {
     u=1;uu=1;
+    gte_u=gte_u_unknown;
   }else{
     u=unneeded_reg[iend+1];
     uu=unneeded_reg_upper[iend+1];
     u=1;uu=1;
+    gte_u=gte_unneeded[iend+1];
   }
-  gte_u=temp_gte_u=0;
 
   for (i=iend;i>=istart;i--)
   {
@@ -6761,7 +6870,7 @@ void unneeded_registers(int istart,int iend,int r)
         // Branch out of this block, flush all regs
         u=1;
         uu=1;
-        gte_u=0;
+        gte_u=gte_u_unknown;
         /* Hexagon hack 
         if(itype[i]==UJUMP&&rt1[i]==31)
         {
@@ -6807,7 +6916,7 @@ void unneeded_registers(int istart,int iend,int r)
           {
             u=1;
             uu=1;
-            gte_u=0;
+            gte_u=gte_u_unknown;
           }
         }
       }
@@ -6850,7 +6959,7 @@ void unneeded_registers(int istart,int iend,int r)
             {
               temp_u=1;
               temp_uu=1;
-              temp_gte_u=0;
+              temp_gte_u=gte_u_unknown;
             }
           }
           tdep=(~temp_uu>>rt1[i])&1;
@@ -6872,7 +6981,7 @@ void unneeded_registers(int istart,int iend,int r)
           }else{
             unneeded_reg[(ba[i]-start)>>2]=1;
             unneeded_reg_upper[(ba[i]-start)>>2]=1;
-            gte_unneeded[(ba[i]-start)>>2]=0;
+            gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
           }
         } /*else*/ if(1) {
           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
@@ -6977,6 +7086,8 @@ void unneeded_registers(int istart,int iend,int r)
     uu&=~(1LL<<us1[i]);
     uu&=~(1LL<<us2[i]);
     gte_u&=~gte_rs[i];
+    if(gte_rs[i]&&rt1[i]&&(unneeded_reg[i+1]&(1ll<<rt1[i])))
+      gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
     // Source-target dependencies
     uu&=~(tdep<<dep1[i]);
     uu&=~(tdep<<dep2[i]);
@@ -7911,7 +8022,6 @@ void new_dynarec_clear_full()
   literalcount=0;
   stop_after_jal=0;
   inv_code_start=inv_code_end=~0;
-  gte_reads_flags=0;
   // TLB
 #ifndef DISABLE_TLB
   using_tlb=0;
@@ -7931,10 +8041,16 @@ void new_dynarec_init()
 {
   printf("Init new dynarec\n");
   out=(u_char *)BASE_ADDR;
+#if BASE_ADDR_FIXED
   if (mmap (out, 1<<TARGET_SIZE_2,
             PROT_READ | PROT_WRITE | PROT_EXEC,
             MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
-            -1, 0) <= 0) {printf("mmap() failed\n");}
+            -1, 0) <= 0) {SysPrintf("mmap() failed\n");}
+#else
+  // not all systems allow execute in data segment by default
+  if (mprotect(out, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
+    SysPrintf("mprotect() failed\n");
+#endif
 #ifdef MUPEN64
   rdword=&readmem_dword;
   fake_pc.f.r.rs=&readmem_dword;
@@ -7988,17 +8104,24 @@ void new_dynarec_init()
 #endif
   tlb_hacks();
   arch_init();
+#ifndef RAM_FIXED
+  ram_offset=(u_int)rdram-0x80000000;
+#endif
+  if (ram_offset!=0)
+    SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
 }
 
 void new_dynarec_cleanup()
 {
   int n;
-  if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0) {printf("munmap() failed\n");}
+  #if BASE_ADDR_FIXED
+  if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0) {SysPrintf("munmap() failed\n");}
+  #endif
   for(n=0;n<4096;n++) ll_clear(jump_in+n);
   for(n=0;n<4096;n++) ll_clear(jump_out+n);
   for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
   #ifdef ROM_COPY
-  if (munmap (ROM_COPY, 67108864) < 0) {printf("munmap() failed\n");}
+  if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
   #endif
 }
 
@@ -8095,7 +8218,7 @@ int new_recompile_block(int addr)
   }
 #endif
   else {
-    printf("Compile at bogus memory address: %x \n", (int)addr);
+    SysPrintf("Compile at bogus memory address: %x \n", (int)addr);
     exit(1);
   }
 
@@ -8424,7 +8547,7 @@ int new_recompile_block(int addr)
       case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
 #endif
       default: strcpy(insn[i],"???"); type=NI;
-        printf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
+        SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
         break;
     }
     itype[i]=type;
@@ -8621,12 +8744,7 @@ int new_recompile_block(int addr)
         {
           case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
           case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
-          case 0x02: gte_rs[i]=1ll<<(gr+32); // CFC2
-            if(gr==31&&!gte_reads_flags) {
-              assem_debug("gte flag read encountered @%08x\n",addr + i*4);
-              gte_reads_flags=1;
-            }
-            break;
+          case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
           case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
         }
         break;
@@ -8651,8 +8769,15 @@ int new_recompile_block(int addr)
         rs2[i]=0;
         rt1[i]=0;
         rt2[i]=0;
-        gte_rt[i]=1ll<<63; // every op changes flags
-        // TODO: other regs?
+        gte_rs[i]=gte_reg_reads[source[i]&0x3f];
+        gte_rt[i]=gte_reg_writes[source[i]&0x3f];
+        gte_rt[i]|=1ll<<63; // every op changes flags
+        if((source[i]&0x3f)==GTE_MVMVA) {
+          int v = (source[i] >> 15) & 3;
+          gte_rs[i]&=~0xe3fll;
+          if(v==3) gte_rs[i]|=0xe00ll;
+          else gte_rs[i]|=3ll<<(v*2);
+        }
         break;
       case FLOAT:
       case FCONV:
@@ -8697,7 +8822,7 @@ int new_recompile_block(int addr)
       // branch in delay slot?
       if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
         // don't handle first branch and call interpreter if it's hit
-        printf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
+        SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
         do_in_intrp=1;
       }
       // basic load delay detection
@@ -8705,14 +8830,14 @@ int new_recompile_block(int addr)
         int t=(ba[i-1]-start)/4;
         if(0 <= t && t < i &&(rt1[i]==rs1[t]||rt1[i]==rs2[t])&&itype[t]!=CJUMP&&itype[t]!=SJUMP) {
           // jump target wants DS result - potential load delay effect
-          printf("load delay @%08x (%08x)\n", addr + i*4, addr);
+          SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
           do_in_intrp=1;
           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))) {
           // v0 overwrite like this is a sign of trouble, bail out
-          printf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
+          SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
           do_in_intrp=1;
         }
       }
@@ -8759,7 +8884,7 @@ int new_recompile_block(int addr)
     // Stop if we're compiling junk
     if(itype[i]==NI&&opcode[i]==0x11) {
       done=stop_after_jal=1;
-      printf("Disabled speculative precompilation\n");
+      SysPrintf("Disabled speculative precompilation\n");
     }
   }
   slen=i;
@@ -8786,6 +8911,7 @@ int new_recompile_block(int addr)
   dirty_reg(&current,CCREG);
   current.isconst=0;
   current.wasconst=0;
+  current.waswritten=0;
   int ds=0;
   int cc=0;
   int hr=-1;
@@ -8814,6 +8940,7 @@ int new_recompile_block(int addr)
         if(current.regmap[hr]==0) current.regmap[hr]=-1;
       }
       current.isconst=0;
+      current.waswritten=0;
     }
     if(i>1)
     {
@@ -8878,6 +9005,7 @@ int new_recompile_block(int addr)
     regs[i].wasconst=current.isconst;
     regs[i].was32=current.is32;
     regs[i].wasdirty=current.dirty;
+    regs[i].loadedconst=0;
     #if defined(DESTRUCTIVE_WRITEBACK) && !defined(FORCE32)
     // To change a dirty register from 32 to 64 bits, we must write
     // it out during the previous cycle (for branches, 2 cycles)
@@ -8973,7 +9101,7 @@ int new_recompile_block(int addr)
         current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
         current.u|=1;
         current.uu|=1;
-      } else { printf("oops, branch at end of block with no delay slot\n");exit(1); }
+      } else { SysPrintf("oops, branch at end of block with no delay slot\n");exit(1); }
     }
     is_ds[i]=ds;
     if(ds) {
@@ -9440,6 +9568,14 @@ int new_recompile_block(int addr)
       }
       memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
     }
+
+    if(i>0&&(itype[i-1]==STORE||itype[i-1]==STORELR||(itype[i-1]==C2LS&&opcode[i-1]==0x3a))&&(u_int)imm[i-1]<0x800)
+      current.waswritten|=1<<rs1[i-1];
+    current.waswritten&=~(1<<rt1[i]);
+    current.waswritten&=~(1<<rt2[i]);
+    if((itype[i]==STORE||itype[i]==STORELR||(itype[i]==C2LS&&opcode[i]==0x3a))&&(u_int)imm[i]>=0x800)
+      current.waswritten&=~(1<<rs1[i]);
+
     /* Branch post-alloc */
     if(i>0)
     {
@@ -9460,7 +9596,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].is32|=1LL<<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(current_constmap));
           break;
         case RJUMP:
           memcpy(&branch_regs[i-1],&current,sizeof(current));
@@ -9485,7 +9621,7 @@ 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(current_constmap));
           break;
         case CJUMP:
           if((opcode[i-1]&0x3E)==4) // BEQ/BNE
@@ -9521,7 +9657,7 @@ 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(current_constmap));
           }
           else
           if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
@@ -9554,7 +9690,7 @@ 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(current_constmap));
           }
           else
           // Alloc the delay slot in case the branch is taken
@@ -9620,7 +9756,7 @@ 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(current_constmap));
           }
           else
           // Alloc the delay slot in case the branch is taken
@@ -9734,8 +9870,13 @@ int new_recompile_block(int addr)
     {
       cc=0;
     }
-#ifdef PCSX
-    else if(/*itype[i]==LOAD||*/itype[i]==STORE||itype[i]==C1LS) // load causes weird timing issues
+#if defined(PCSX) && !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)
     }
@@ -9754,7 +9895,7 @@ int new_recompile_block(int addr)
       regs[i].is32=current.is32;
       regs[i].dirty=current.dirty;
       regs[i].isconst=current.isconst;
-      memcpy(constmap[i],current.constmap,sizeof(current.constmap));
+      memcpy(constmap[i],current_constmap,sizeof(current_constmap));
     }
     for(hr=0;hr<HOST_REGS;hr++) {
       if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
@@ -9764,6 +9905,7 @@ int new_recompile_block(int addr)
       }
     }
     if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
+    regs[i].waswritten=current.waswritten;
   }
   
   /* Pass 4 - Cull unused host registers */
@@ -10025,7 +10167,7 @@ int new_recompile_block(int addr)
                 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
                 if(regs[i].regmap[hr]<64||!((regs[i].was32>>(regs[i].regmap[hr]&63))&1))
                 {
-                  printf("fail: %x (%d %d!=%d)\n",start+i*4,hr,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]);
                   assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
                 }
                 regmap_pre[i+1][hr]=-1;
@@ -11462,7 +11604,7 @@ 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((int)out>BASE_ADDR+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE) out=(u_char *)BASE_ADDR;
+  if((u_int)out>(u_int)BASE_ADDR+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE) out=(u_char *)BASE_ADDR;
   
   // Trap writes to any of the pages we compiled
   for(i=start>>12;i<=(start+slen*4)>>12;i++) {
@@ -11490,11 +11632,11 @@ int new_recompile_block(int addr)
   
   /* Pass 10 - Free memory by expiring oldest blocks */
   
-  int end=((((int)out-BASE_ADDR)>>(TARGET_SIZE_2-16))+16384)&65535;
+  int end=((((int)out-(int)BASE_ADDR)>>(TARGET_SIZE_2-16))+16384)&65535;
   while(expirep!=end)
   {
     int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
-    int base=BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
+    int base=(int)BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
     inv_debug("EXP: Phase %d\n",expirep);
     switch((expirep>>11)&3)
     {