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=2c9130b328b97032e3478956c3fb3eb06711115f;hp=a177d881b474f7edc961c42d64e64662a088eac7;hb=ffb0b9e0f051789f97f5efdcfab0b261e62688f9;hpb=eba830cd5dc35bc900ea60d8c06ee582178280a4

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index a177d881..2c9130b3 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -21,10 +21,15 @@
 #include <stdlib.h>
 #include <stdint.h> //include for uint64_t
 #include <assert.h>
+#include <sys/mman.h>
 
 #include "emu_if.h" //emulator interface
 
-#include <sys/mman.h>
+//#define DISASM
+//#define assem_debug printf
+//#define inv_debug printf
+#define assem_debug(...)
+#define inv_debug(...)
 
 #ifdef __i386__
 #include "assem_x86.h"
@@ -80,6 +85,15 @@ struct ll_entry
   u_char dep1[MAXBLOCK];
   u_char dep2[MAXBLOCK];
   u_char lt1[MAXBLOCK];
+  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
+  static u_int smrv_strong_next; // same, but after current insn executes
+  static u_int smrv_weak_next;
   int imm[MAXBLOCK];
   u_int ba[MAXBLOCK];
   char likely[MAXBLOCK];
@@ -126,7 +140,7 @@ struct ll_entry
 #else
   static const u_int using_tlb=0;
 #endif
-  static u_int sp_in_mirror;
+  int new_dynarec_did_compile;
   u_int stop_after_jal;
   extern u_char restore_candidate[512];
   extern int cycle_count;
@@ -261,12 +275,6 @@ int tracedebug=0;
 
 //#define DEBUG_CYCLE_COUNT 1
 
-void nullf() {}
-//#define assem_debug printf
-//#define inv_debug printf
-#define assem_debug nullf
-#define inv_debug nullf
-
 static void tlb_hacks()
 {
 #ifndef DISABLE_TLB
@@ -371,6 +379,7 @@ void *get_addr(u_int vaddr)
       if(verify_dirty(head->addr)) {
         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
         invalid_code[vaddr>>12]=0;
+        inv_code_start=inv_code_end=~0;
         memory_map[vaddr>>12]|=0x40000000;
         if(vpage<2048) {
 #ifndef DISABLE_TLB
@@ -463,6 +472,7 @@ void *get_addr_32(u_int vaddr,u_int flags)
       if(verify_dirty(head->addr)) {
         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
         invalid_code[vaddr>>12]=0;
+        inv_code_start=inv_code_end=~0;
         memory_map[vaddr>>12]|=0x40000000;
         if(vpage<2048) {
 #ifndef DISABLE_TLB
@@ -842,7 +852,7 @@ void alloc_all(struct regstat *cur,int i)
   }
 }
 
-
+#ifndef FORCE32
 void div64(int64_t dividend,int64_t divisor)
 {
   lo=dividend/divisor;
@@ -953,6 +963,7 @@ uint64_t ldr_merge(uint64_t original,uint64_t loaded,u_int bits)
   else original=loaded;
   return original;
 }
+#endif
 
 #ifdef __i386__
 #include "assem_x86.c"
@@ -1127,39 +1138,10 @@ void invalidate_page(u_int page)
     head=next;
   }
 }
-void invalidate_block(u_int block)
+
+static void invalidate_block_range(u_int block, u_int first, u_int last)
 {
   u_int page=get_page(block<<12);
-  u_int vpage=get_vpage(block<<12);
-  inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
-  //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
-  u_int first,last;
-  first=last=page;
-  struct ll_entry *head;
-  head=jump_dirty[vpage];
-  //printf("page=%d vpage=%d\n",page,vpage);
-  while(head!=NULL) {
-    u_int start,end;
-    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(((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;
-        }
-      }
-#ifndef DISABLE_TLB
-      if(page<2048&&(signed int)start>=(signed int)0xC0000000&&(signed int)end>=(signed int)0xC0000000) {
-        if(((start+memory_map[start>>12]-(u_int)rdram)>>12)<=page&&((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)>=page) {
-          if((((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047)<first) first=((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047;
-          if((((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)&2047)>last) last=((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)&2047;
-        }
-      }
-#endif
-    }
-    head=head->next;
-  }
   //printf("first=%d last=%d\n",first,last);
   invalidate_page(page);
   assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
@@ -1178,9 +1160,6 @@ void invalidate_block(u_int block)
   
   // Don't trap writes
   invalid_code[block]=1;
-#ifdef PCSX
-  invalid_code[((u_int)0x80000000>>12)|page]=1;
-#endif
 #ifndef DISABLE_TLB
   // If there is a valid TLB entry for this page, remove write protect
   if(tlb_LUT_w[block]) {
@@ -1198,10 +1177,98 @@ void invalidate_block(u_int block)
   memset(mini_ht,-1,sizeof(mini_ht));
   #endif
 }
+
+void invalidate_block(u_int block)
+{
+  u_int page=get_page(block<<12);
+  u_int vpage=get_vpage(block<<12);
+  inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
+  //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
+  u_int first,last;
+  first=last=page;
+  struct ll_entry *head;
+  head=jump_dirty[vpage];
+  //printf("page=%d vpage=%d\n",page,vpage);
+  while(head!=NULL) {
+    u_int start,end;
+    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(((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;
+        }
+      }
+#ifndef DISABLE_TLB
+      if(page<2048&&(signed int)start>=(signed int)0xC0000000&&(signed int)end>=(signed int)0xC0000000) {
+        if(((start+memory_map[start>>12]-(u_int)rdram)>>12)<=page&&((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)>=page) {
+          if((((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047)<first) first=((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047;
+          if((((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)&2047)>last) last=((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)&2047;
+        }
+      }
+#endif
+    }
+    head=head->next;
+  }
+  invalidate_block_range(block,first,last);
+}
+
 void invalidate_addr(u_int addr)
 {
+#ifdef PCSX
+  //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);
+  if(page<2048) { // RAM
+    struct ll_entry *head;
+    u_int addr_min=~0, addr_max=0;
+    int mask=RAM_SIZE-1;
+    int pg1;
+    inv_code_start=addr&~0xfff;
+    inv_code_end=addr|0xfff;
+    pg1=page;
+    if (pg1>0) {
+      // must check previous page too because of spans..
+      pg1--;
+      inv_code_start-=0x1000;
+    }
+    for(;pg1<=page;pg1++) {
+      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(start<addr_min) addr_min=start;
+          if(end>addr_max) addr_max=end;
+        }
+        else if(addr<start) {
+          if(start<inv_code_end)
+            inv_code_end=start-1;
+        }
+        else {
+          if(end>inv_code_start)
+            inv_code_start=end;
+        }
+      }
+    }
+    if (addr_min!=~0) {
+      inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
+      inv_code_start=inv_code_end=~0;
+      invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
+      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)
+      return;
+  }
+#endif
   invalidate_block(addr>>12);
 }
+
 // This is called when loading a save state.
 // Anything could have changed, so invalidate everything.
 void invalidate_all_pages()
@@ -1240,6 +1307,8 @@ void add_link(u_int vaddr,void *src)
 {
   u_int page=get_page(vaddr);
   inv_debug("add_link: %x -> %x (%d)\n",(int)src,vaddr,page);
+  int *ptr=(int *)(src+4);
+  assert((*ptr&0x0fff0000)==0x059f0000);
   ll_add(jump_out+page,vaddr,src);
   //int ptr=get_pointer(src);
   //inv_debug("add_link: Pointer is to %x\n",(int)ptr);
@@ -1326,8 +1395,6 @@ void mov_alloc(struct regstat *current,int i)
 
 void shiftimm_alloc(struct regstat *current,int i)
 {
-  clear_const(current,rs1[i]);
-  clear_const(current,rt1[i]);
   if(opcode2[i]<=0x3) // SLL/SRL/SRA
   {
     if(rt1[i]) {
@@ -1336,8 +1403,21 @@ void shiftimm_alloc(struct regstat *current,int i)
       alloc_reg(current,i,rt1[i]);
       current->is32|=1LL<<rt1[i];
       dirty_reg(current,rt1[i]);
+      if(is_const(current,rs1[i])) {
+        int v=get_const(current,rs1[i]);
+        if(opcode2[i]==0x00) set_const(current,rt1[i],v<<imm[i]);
+        if(opcode2[i]==0x02) set_const(current,rt1[i],(u_int)v>>imm[i]);
+        if(opcode2[i]==0x03) set_const(current,rt1[i],v>>imm[i]);
+      }
+      else clear_const(current,rt1[i]);
     }
   }
+  else
+  {
+    clear_const(current,rs1[i]);
+    clear_const(current,rt1[i]);
+  }
+
   if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
   {
     if(rt1[i]) {
@@ -2623,7 +2703,7 @@ void shiftimm_assemble(int i,struct regstat *i_regs)
       t=get_reg(i_regs->regmap,rt1[i]);
       s=get_reg(i_regs->regmap,rs1[i]);
       //assert(t>=0);
-      if(t>=0){
+      if(t>=0&&!((i_regs->isconst>>t)&1)){
         if(rs1[i]==0)
         {
           emit_zeroreg(t);
@@ -2817,23 +2897,7 @@ void load_assemble(int i,struct regstat *i_regs)
       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
       #endif
       {
-        #ifdef PCSX
-        if(sp_in_mirror&&rs1[i]==29) {
-          emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
-          emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
-          fastload_reg_override=HOST_TEMPREG;
-        }
-        else
-        #endif
-        emit_cmpimm(addr,RAM_SIZE);
-        jaddr=(int)out;
-        #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
-        // Hint to branch predictor that the branch is unlikely to be taken
-        if(rs1[i]>=28)
-          emit_jno_unlikely(0);
-        else
-        #endif
-        emit_jno(0);
+        jaddr=emit_fastpath_cmp_jump(i,addr,&fastload_reg_override);
       }
     }
   }else{ // using tlb
@@ -3118,14 +3182,7 @@ void store_assemble(int i,struct regstat *i_regs)
   else addr=s;
   if(!using_tlb) {
     if(!c) {
-      #ifdef PCSX
-      if(sp_in_mirror&&rs1[i]==29) {
-        emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
-        emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
-        faststore_reg_override=HOST_TEMPREG;
-      }
-      else
-      #endif
+      #ifndef PCSX
       #ifdef R29_HACK
       // Strmnnrmn's speed hack
       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
@@ -3148,6 +3205,9 @@ void store_assemble(int i,struct regstat *i_regs)
         #endif
         emit_jno(0);
       }
+      #else
+        jaddr=emit_fastpath_cmp_jump(i,addr,&faststore_reg_override);
+      #endif
     }
   }else{ // using tlb
     int x=0;
@@ -3224,6 +3284,14 @@ void store_assemble(int i,struct regstat *i_regs)
     }
     type=STORED_STUB;
   }
+#ifdef PCSX
+  if(jaddr) {
+    // PCSX store handlers don't check invcode again
+    reglist|=1<<addr;
+    add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
+    jaddr=0;
+  }
+#endif
   if(!using_tlb) {
     if(!c||memtarget) {
       #ifdef DESTRUCTIVE_SHIFT
@@ -3731,6 +3799,7 @@ void c2ls_assemble(int i,struct regstat *i_regs)
   int memtarget=0,c=0;
   int jaddr2=0,jaddr3,type;
   int agr=AGEN1+(i&1);
+  int fastio_reg_override=0;
   u_int hr,reglist=0;
   u_int copr=(source[i]>>16)&0x1f;
   s=get_reg(i_regs->regmap,rs1[i]);
@@ -3772,22 +3841,24 @@ void c2ls_assemble(int i,struct regstat *i_regs)
   }
   else {
     if(!c) {
-      emit_cmpimm(offset||c||s<0?ar:s,RAM_SIZE);
-      jaddr2=(int)out;
-      emit_jno(0);
+      jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
     }
     if (opcode[i]==0x32) { // LWC2
       #ifdef HOST_IMM_ADDR32
       if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl);
       else
       #endif
-      emit_readword_indexed(0,ar,tl);
+      int a=ar;
+      if(fastio_reg_override) a=fastio_reg_override;
+      emit_readword_indexed(0,a,tl);
     }
     if (opcode[i]==0x3a) { // SWC2
       #ifdef DESTRUCTIVE_SHIFT
       if(!offset&&!c&&s>=0) emit_mov(s,ar);
       #endif
-      emit_writeword_indexed(tl,0,ar);
+      int a=ar;
+      if(fastio_reg_override) a=fastio_reg_override;
+      emit_writeword_indexed(tl,0,a);
     }
   }
   if(jaddr2)
@@ -3892,6 +3963,7 @@ void intcall_assemble(int i,struct regstat *i_regs)
 
 void ds_assemble(int i,struct regstat *i_regs)
 {
+  speculate_register_values(i);
   is_delayslot=1;
   switch(itype[i]) {
     case ALU:
@@ -6650,8 +6722,8 @@ static void pagespan_ds()
 void unneeded_registers(int istart,int iend,int r)
 {
   int i;
-  uint64_t u,uu,b,bu;
-  uint64_t temp_u,temp_uu;
+  uint64_t u,uu,gte_u,b,bu,gte_bu;
+  uint64_t temp_u,temp_uu,temp_gte_u;
   uint64_t tdep;
   if(iend==slen-1) {
     u=1;uu=1;
@@ -6660,6 +6732,8 @@ void unneeded_registers(int istart,int iend,int r)
     uu=unneeded_reg_upper[iend+1];
     u=1;uu=1;
   }
+  gte_u=temp_gte_u=0;
+
   for (i=iend;i>=istart;i--)
   {
     //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
@@ -6673,6 +6747,7 @@ void unneeded_registers(int istart,int iend,int r)
         // Branch out of this block, flush all regs
         u=1;
         uu=1;
+        gte_u=0;
         /* Hexagon hack 
         if(itype[i]==UJUMP&&rt1[i]==31)
         {
@@ -6704,17 +6779,21 @@ void unneeded_registers(int istart,int iend,int r)
         uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
         uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
         u|=1;uu|=1;
+        gte_u|=gte_rt[i+1];
+        gte_u&=~gte_rs[i+1];
         // If branch is "likely" (and conditional)
         // then we skip the delay slot on the fall-thru path
         if(likely[i]) {
           if(i<slen-1) {
             u&=unneeded_reg[i+2];
             uu&=unneeded_reg_upper[i+2];
+            gte_u&=gte_unneeded[i+2];
           }
           else
           {
             u=1;
             uu=1;
+            gte_u=0;
           }
         }
       }
@@ -6728,10 +6807,12 @@ void unneeded_registers(int istart,int iend,int r)
           {
             // Unconditional branch
             temp_u=1;temp_uu=1;
+            temp_gte_u=0;
           } else {
             // Conditional branch (not taken case)
             temp_u=unneeded_reg[i+2];
             temp_uu=unneeded_reg_upper[i+2];
+            temp_gte_u&=gte_unneeded[i+2];
           }
           // Merge in delay slot
           tdep=(~temp_uu>>rt1[i+1])&1;
@@ -6741,17 +6822,21 @@ void unneeded_registers(int istart,int iend,int r)
           temp_uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
           temp_uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
           temp_u|=1;temp_uu|=1;
+          temp_gte_u|=gte_rt[i+1];
+          temp_gte_u&=~gte_rs[i+1];
           // If branch is "likely" (and conditional)
           // then we skip the delay slot on the fall-thru path
           if(likely[i]) {
             if(i<slen-1) {
               temp_u&=unneeded_reg[i+2];
               temp_uu&=unneeded_reg_upper[i+2];
+              temp_gte_u&=gte_unneeded[i+2];
             }
             else
             {
               temp_u=1;
               temp_uu=1;
+              temp_gte_u=0;
             }
           }
           tdep=(~temp_uu>>rt1[i])&1;
@@ -6761,8 +6846,11 @@ void unneeded_registers(int istart,int iend,int r)
           temp_uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
           temp_uu&=~((tdep<<dep1[i])|(tdep<<dep2[i]));
           temp_u|=1;temp_uu|=1;
+          temp_gte_u|=gte_rt[i];
+          temp_gte_u&=~gte_rs[i];
           unneeded_reg[i]=temp_u;
           unneeded_reg_upper[i]=temp_uu;
+          gte_unneeded[i]=temp_gte_u;
           // Only go three levels deep.  This recursion can take an
           // excessive amount of time if there are a lot of nested loops.
           if(r<2) {
@@ -6770,6 +6858,7 @@ 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;
           }
         } /*else*/ if(1) {
           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
@@ -6777,6 +6866,7 @@ void unneeded_registers(int istart,int iend,int r)
             // Unconditional branch
             u=unneeded_reg[(ba[i]-start)>>2];
             uu=unneeded_reg_upper[(ba[i]-start)>>2];
+            gte_u=gte_unneeded[(ba[i]-start)>>2];
             branch_unneeded_reg[i]=u;
             branch_unneeded_reg_upper[i]=uu;
         //u=1;
@@ -6791,10 +6881,13 @@ void unneeded_registers(int istart,int iend,int r)
             uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
             uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
             u|=1;uu|=1;
+            gte_u|=gte_rt[i+1];
+            gte_u&=~gte_rs[i+1];
           } else {
             // Conditional branch
             b=unneeded_reg[(ba[i]-start)>>2];
             bu=unneeded_reg_upper[(ba[i]-start)>>2];
+            gte_bu=gte_unneeded[(ba[i]-start)>>2];
             branch_unneeded_reg[i]=b;
             branch_unneeded_reg_upper[i]=bu;
         //b=1;
@@ -6809,20 +6902,25 @@ void unneeded_registers(int istart,int iend,int r)
             bu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
             bu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
             b|=1;bu|=1;
+            gte_bu|=gte_rt[i+1];
+            gte_bu&=~gte_rs[i+1];
             // If branch is "likely" then we skip the
             // delay slot on the fall-thru path
             if(likely[i]) {
               u=b;
               uu=bu;
+              gte_u=gte_bu;
               if(i<slen-1) {
                 u&=unneeded_reg[i+2];
                 uu&=unneeded_reg_upper[i+2];
+                gte_u&=gte_unneeded[i+2];
         //u=1;
         //uu=1;
               }
             } else {
               u&=b;
               uu&=bu;
+              gte_u&=gte_bu;
         //u=1;
         //uu=1;
             }
@@ -6858,11 +6956,13 @@ void unneeded_registers(int istart,int iend,int r)
     u|=1LL<<rt2[i];
     uu|=1LL<<rt1[i];
     uu|=1LL<<rt2[i];
+    gte_u|=gte_rt[i];
     // Accessed registers are needed
     u&=~(1LL<<rs1[i]);
     u&=~(1LL<<rs2[i]);
     uu&=~(1LL<<us1[i]);
     uu&=~(1LL<<us2[i]);
+    gte_u&=~gte_rs[i];
     // Source-target dependencies
     uu&=~(tdep<<dep1[i]);
     uu&=~(tdep<<dep2[i]);
@@ -6871,6 +6971,7 @@ void unneeded_registers(int istart,int iend,int r)
     // Save it
     unneeded_reg[i]=u;
     unneeded_reg_upper[i]=uu;
+    gte_unneeded[i]=gte_u;
     /*
     printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
     printf("U:");
@@ -7686,6 +7787,7 @@ void clean_registers(int istart,int iend,int wr)
   }
 }
 
+#ifdef DISASM
   /* disassembly */
 void disassemble_inst(int i)
 {
@@ -7774,6 +7876,9 @@ void disassemble_inst(int i)
         printf (" %x: %s\n",start+i*4,insn[i]);
     }
 }
+#else
+static void disassemble_inst(int i) {}
+#endif // DISASM
 
 // clear the state completely, instead of just marking
 // things invalid like invalidate_all_pages() does
@@ -7791,11 +7896,12 @@ void new_dynarec_clear_full()
   pending_exception=0;
   literalcount=0;
   stop_after_jal=0;
+  inv_code_start=inv_code_end=~0;
+  gte_reads_flags=0;
   // TLB
 #ifndef DISABLE_TLB
   using_tlb=0;
 #endif
-  sp_in_mirror=0;
   for(n=0;n<524288;n++) // 0 .. 0x7FFFFFFF
     memory_map[n]=-1;
   for(n=524288;n<526336;n++) // 0x80000000 .. 0x807FFFFF
@@ -7904,12 +8010,8 @@ int new_recompile_block(int addr)
   //rlist();
   start = (u_int)addr&~3;
   //assert(((u_int)addr&1)==0);
+  new_dynarec_did_compile=1;
 #ifdef PCSX
-  if(!sp_in_mirror&&(signed int)(psxRegs.GPR.n.sp&0xffe00000)>0x80200000&&
-     0x10000<=psxRegs.GPR.n.sp&&(psxRegs.GPR.n.sp&~0xe0e00000)<RAM_SIZE) {
-    printf("SP hack enabled (%08x), @%08x\n", psxRegs.GPR.n.sp, psxRegs.pc);
-    sp_in_mirror=1;
-  }
   if (Config.HLE && start == 0x80001000) // hlecall
   {
     // XXX: is this enough? Maybe check hleSoftCall?
@@ -7919,6 +8021,7 @@ int new_recompile_block(int addr)
     emit_movimm(start,0);
     emit_writeword(0,(int)&pcaddr);
     emit_jmp((int)new_dyna_leave);
+    literal_pool(0);
 #ifdef __arm__
     __clear_cache((void *)beginning,out);
 #endif
@@ -8282,11 +8385,14 @@ int new_recompile_block(int addr)
 #endif
 #ifdef PCSX
       case 0x12: strcpy(insn[i],"COP2"); type=NI;
-        // note: COP MIPS-1 encoding differs from MIPS32
         op2=(source[i]>>21)&0x1f;
-        if (source[i]&0x3f) {
+        //if (op2 & 0x10) {
+        if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
           if (gte_handlers[source[i]&0x3f]!=NULL) {
-            snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
+            if (gte_regnames[source[i]&0x3f]!=NULL)
+              strcpy(insn[i],gte_regnames[source[i]&0x3f]);
+            else
+              snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
             type=C2OP;
           }
         }
@@ -8314,6 +8420,7 @@ int new_recompile_block(int addr)
     us2[i]=0;
     dep1[i]=0;
     dep2[i]=0;
+    gte_rs[i]=gte_rt[i]=0;
     switch(type) {
       case LOAD:
         rs1[i]=(source[i]>>21)&0x1f;
@@ -8477,7 +8584,6 @@ int new_recompile_block(int addr)
         if(op2==16) if((source[i]&0x3f)==0x18) rs2[i]=CCREG; // ERET
         break;
       case COP1:
-      case COP2:
         rs1[i]=0;
         rs2[i]=0;
         rt1[i]=0;
@@ -8487,6 +8593,28 @@ int new_recompile_block(int addr)
         if(op2==5) us1[i]=rs1[i]; // DMTC1
         rs2[i]=CSREG;
         break;
+      case COP2:
+        rs1[i]=0;
+        rs2[i]=0;
+        rt1[i]=0;
+        rt2[i]=0;
+        if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC2/CFC2
+        if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC2/CTC2
+        rs2[i]=CSREG;
+        int gr=(source[i]>>11)&0x1F;
+        switch(op2)
+        {
+          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 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
+        }
+        break;
       case C1LS:
         rs1[i]=(source[i]>>21)&0x1F;
         rs2[i]=CSREG;
@@ -8500,6 +8628,16 @@ int new_recompile_block(int addr)
         rt1[i]=0;
         rt2[i]=0;
         imm[i]=(short)source[i];
+        if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
+        else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
+        break;
+      case C2OP:
+        rs1[i]=0;
+        rs2[i]=0;
+        rt1[i]=0;
+        rt2[i]=0;
+        gte_rt[i]=1ll<<63; // every op changes flags
+        // TODO: other regs?
         break;
       case FLOAT:
       case FCONV:
@@ -10755,9 +10893,9 @@ int new_recompile_block(int addr)
   if(itype[slen-1]==SPAN) {
     bt[slen-1]=1; // Mark as a branch target so instruction can restart after exception
   }
-  
+
+#ifdef DISASM
   /* Debug/disassembly */
-  if((void*)assem_debug==(void*)printf) 
   for(i=0;i<slen;i++)
   {
     printf("U:");
@@ -10973,6 +11111,7 @@ int new_recompile_block(int addr)
 #endif
     }
   }
+#endif // DISASM
 
   /* Pass 8 - Assembly */
   linkcount=0;stubcount=0;
@@ -11005,12 +11144,13 @@ int new_recompile_block(int addr)
   for(i=0;i<slen;i++)
   {
     //if(ds) printf("ds: ");
-    if((void*)assem_debug==(void*)printf) disassemble_inst(i);
+    disassemble_inst(i);
     if(ds) {
       ds=0; // Skip delay slot
       if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
       instr_addr[i]=0;
     } else {
+      speculate_register_values(i);
       #ifndef DESTRUCTIVE_WRITEBACK
       if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
       {
@@ -11323,11 +11463,14 @@ int new_recompile_block(int addr)
     }
 #endif
   }
+  inv_code_start=inv_code_end=~0;
 #ifdef PCSX
-  // PCSX maps all RAM mirror invalid_code tests to 0x80000000..0x80000000+RAM_SIZE
+  // for PCSX we need to mark all mirrors too
   if(get_page(start)<(RAM_SIZE>>12))
     for(i=start>>12;i<=(start+slen*4)>>12;i++)
-      invalid_code[((u_int)0x80000000>>12)|i]=0;
+      invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
+      invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
+      invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
 #endif
   
   /* Pass 10 - Free memory by expiring oldest blocks */