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=362e3c62a95c3b8e1a39d012121b621227131c5e;hb=ffb0b9e0f051789f97f5efdcfab0b261e62688f9;hpb=b96d3df7d4272503a287d243c09173eb6c842609

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 362e3c62..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"
@@ -84,6 +89,11 @@ struct ll_entry
   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];
@@ -130,7 +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;
@@ -265,12 +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
@@ -848,7 +852,7 @@ void alloc_all(struct regstat *cur,int i)
   }
 }
 
-
+#ifndef FORCE32
 void div64(int64_t dividend,int64_t divisor)
 {
   lo=dividend/divisor;
@@ -959,6 +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"
@@ -1390,8 +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]) {
@@ -1400,8 +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]) {
@@ -2687,7 +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);
@@ -2881,23 +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
@@ -3182,14 +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)
@@ -3212,6 +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;
@@ -3803,6 +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]);
@@ -3844,22 +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)
@@ -3964,6 +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:
@@ -7787,6 +7787,7 @@ void clean_registers(int istart,int iend,int wr)
   }
 }
 
+#ifdef DISASM
   /* disassembly */
 void disassemble_inst(int i)
 {
@@ -7875,6 +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
@@ -7898,7 +7902,6 @@ void new_dynarec_clear_full()
 #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
@@ -8007,12 +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?
@@ -10894,9 +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:");
@@ -11112,6 +11111,7 @@ int new_recompile_block(int addr)
 #endif
     }
   }
+#endif // DISASM
 
   /* Pass 8 - Assembly */
   linkcount=0;stubcount=0;
@@ -11144,12 +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))
       {