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=cafa4a425726151f9fffd4281275da3e12528238;hb=c7b746f050d33e3bb5b06707dcc52f8699858e40;hpb=bdeade4633d41d76e0c22b3810241bbf7cb5a8a3

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index cafa4a42..84370089 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -41,6 +41,19 @@
 #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
 
@@ -141,6 +154,11 @@ struct ll_entry
   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;
 
@@ -274,6 +292,8 @@ 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)
@@ -644,7 +664,7 @@ uint64_t get_const(struct regstat *cur,signed char reg)
       return current_constmap[hr];
     }
   }
-  printf("Unknown constant in r%d\n",reg);
+  SysPrintf("Unknown constant in r%d\n",reg);
   exit(1);
 }
 
@@ -1211,7 +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;
@@ -1241,10 +1261,11 @@ void invalidate_addr(u_int 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..
@@ -1255,11 +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;
         }
@@ -1276,6 +1301,8 @@ void invalidate_addr(u_int addr)
       return;
     }
     else {
+      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;
     }
@@ -1978,7 +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:
@@ -2917,6 +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
@@ -2982,7 +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);
@@ -3069,7 +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);
@@ -3226,6 +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
@@ -3270,7 +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;
   }
@@ -3341,7 +3377,7 @@ void store_assemble(int i,struct regstat *i_regs)
   // basic current block modification detection..
   // not looking back as that should be in mips cache already
   if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
-    printf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*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);
@@ -3874,6 +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);
@@ -4045,7 +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;
 }
@@ -4938,7 +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);
@@ -4956,6 +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;
@@ -4963,7 +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];
   }
@@ -4982,7 +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);
   }
@@ -5208,7 +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);
@@ -6765,7 +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) {
@@ -7993,15 +8041,15 @@ void new_dynarec_init()
 {
   printf("Init new dynarec\n");
   out=(u_char *)BASE_ADDR;
-#ifdef BASE_ADDR_FIXED
+#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)
-    printf("mprotect() failed\n");
+    SysPrintf("mprotect() failed\n");
 #endif
 #ifdef MUPEN64
   rdword=&readmem_dword;
@@ -8056,19 +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;
-  #ifdef BASE_ADDR_FIXED
-  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
 }
 
@@ -8165,7 +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);
   }
 
@@ -8494,7 +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;
@@ -8769,7 +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
@@ -8777,14 +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;
         }
       }
@@ -8831,7 +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;
@@ -9048,7 +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) {
@@ -9823,7 +9876,7 @@ int new_recompile_block(int addr)
       // 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 causes weird timing issues
+    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)
     }
@@ -10114,7 +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;