X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=892d8e946442f8413ec867463bc076181e672aba;hb=f9e9616e4ccc5de395bb548e95785565a42eed14;hp=00d307b3224c403383e9373807838eaf5913b099;hpb=90f98e7cf5ed4fdabf6b6ff16d6886fef9dc7bbc;p=pcsx_rearmed.git

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 00d307b3..892d8e94 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -81,7 +81,7 @@ static Jit g_jit;
 #endif
 
 #define RAM_SIZE 0x200000
-#define MAXBLOCK 4096
+#define MAXBLOCK 2048
 #define MAX_OUTPUT_BLOCK_SIZE 262144
 #define EXPIRITY_OFFSET (MAX_OUTPUT_BLOCK_SIZE * 2)
 #define PAGE_COUNT 1024
@@ -158,14 +158,14 @@ struct regstat
 {
   signed char regmap_entry[HOST_REGS];
   signed char regmap[HOST_REGS];
-  uint64_t wasdirty;
-  uint64_t dirty;
-  uint64_t u;
+  u_int wasdirty;
+  u_int dirty;
   u_int wasconst;                // before; for example 'lw r2, (r2)' wasconst is true
   u_int isconst;                 //  ... but isconst is false when r2 is known (hr)
   u_int loadedconst;             // host regs that have constants loaded
   u_int noevict;                 // can't evict this hr (alloced by current op)
   //u_int waswritten;              // MIPS regs that were used as store base before
+  uint64_t u;
 };
 
 struct ht_entry
@@ -242,8 +242,13 @@ static struct decoded_insn
   u_char is_delay_load:1; // is_load + MFC/CFC
   u_char is_exception:1;  // unconditional, also interp. fallback
   u_char may_except:1;    // might generate an exception
+  u_char ls_type:2;       // load/store type (ls_width_type)
 } dops[MAXBLOCK];
 
+enum ls_width_type {
+  LS_8 = 0, LS_16, LS_32, LS_LR
+};
+
 static struct compile_info
 {
   int imm;
@@ -338,9 +343,9 @@ static struct compile_info
 #define CCREG 36 // Cycle count
 #define INVCP 37 // Pointer to invalid_code
 //#define MMREG 38 // Pointer to memory_map
-#define ROREG 39 // ram offset (if rdram!=0x80000000)
+#define ROREG 39 // ram offset (if psxM != 0x80000000)
 #define TEMPREG 40
-#define FTEMP 40 // FPU temporary register
+#define FTEMP 40 // Load/store temporary register (was fpu)
 #define PTEMP 41 // Prefetch temporary register
 //#define TLREG 42 // TLB mapping offset
 #define RHASH 43 // Return address hash
@@ -349,7 +354,6 @@ static struct compile_info
 #define MAXREG 45
 #define AGEN1 46 // Address generation temporary register (pass5b_preallocate2)
 //#define AGEN2 47 // Address generation temporary register
-#define BTREG 50 // Branch target temporary register
 
   /* instruction types */
 #define NOP 0     // No operation
@@ -380,7 +384,6 @@ static struct compile_info
   /* branch codes */
 #define TAKEN 1
 #define NOTTAKEN 2
-#define NULLDS 3
 
 #define DJT_1 (void *)1l // no function, just a label in assem_debug log
 #define DJT_2 (void *)2l
@@ -411,9 +414,9 @@ static void invalidate_block(struct block_info *block);
 static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_);
 
 // Needed by assembler
-static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
-static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
-static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
+static void wb_register(signed char r, const signed char regmap[], u_int dirty);
+static void wb_dirtys(const signed char i_regmap[], u_int i_dirty);
+static void wb_needed_dirtys(const signed char i_regmap[], u_int i_dirty, int addr);
 static void load_all_regs(const signed char i_regmap[]);
 static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
 static void load_regs_entry(int t);
@@ -1054,12 +1057,8 @@ static void lsn(u_char hsn[], int i)
   if(dops[i].itype==C2LS) {
     hsn[FTEMP]=0;
   }
-  // Load L/R also uses FTEMP as a temporary register
-  if(dops[i].itype==LOADLR) {
-    hsn[FTEMP]=0;
-  }
-  // Also SWL/SWR/SDL/SDR
-  if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
+  // Load/store L/R also uses FTEMP as a temporary register
+  if (dops[i].itype == LOADLR || dops[i].itype == STORELR) {
     hsn[FTEMP]=0;
   }
   // Don't remove the miniht registers
@@ -1236,7 +1235,11 @@ static const struct {
   FUNCNAME(do_memhandler_post),
 #endif
 #ifdef DRC_DBG
+# ifdef __aarch64__
+  FUNCNAME(do_insn_cmp_arm64),
+# else
   FUNCNAME(do_insn_cmp),
+# endif
 #endif
 };
 
@@ -2290,7 +2293,7 @@ static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
 }
 
 // Write out a single register
-static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
+static void wb_register(signed char r, const signed char regmap[], u_int dirty)
 {
   int hr;
   for(hr=0;hr<HOST_REGS;hr++) {
@@ -2300,6 +2303,7 @@ static void wb_register(signed char r, const signed char regmap[], uint64_t dirt
           assert(regmap[hr]<64);
           emit_storereg(r,hr);
         }
+        break;
       }
     }
   }
@@ -2837,11 +2841,11 @@ static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
     // alignment check
     u_int op = dops[i].opcode;
     int mask = ((op & 0x37) == 0x21 || op == 0x25) ? 1 : 3; // LH/SH/LHU
-    void *jaddr;
+    void *jaddr2;
     emit_testimm(addr, mask);
-    jaddr = out;
+    jaddr2 = out;
     emit_jne(0);
-    add_stub_r(ALIGNMENT_STUB, jaddr, out, i, addr, i_regs, ccadj_, 0);
+    add_stub_r(ALIGNMENT_STUB, jaddr2, out, i, addr, i_regs, ccadj_, 0);
   }
 
   if(type==MTYPE_8020) { // RAM 80200000+ mirror
@@ -4491,7 +4495,7 @@ static void address_generation(int i, const struct regstat *i_regs, signed char
     {
       int offset = cinfo[i].imm;
       int add_offset = offset != 0;
-      int c=(i_regs->wasconst>>rs)&1;
+      int c = rs >= 0 && ((i_regs->wasconst >> rs) & 1);
       if(dops[i].rs1==0) {
         // Using r0 as a base address
         assert(ra >= 0);
@@ -4700,7 +4704,8 @@ static void load_all_consts(const signed char regmap[], u_int dirty, int i)
 }
 
 // Write out all dirty registers (except cycle count)
-static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
+#ifndef wb_dirtys
+static void wb_dirtys(const signed char i_regmap[], u_int i_dirty)
 {
   int hr;
   for(hr=0;hr<HOST_REGS;hr++) {
@@ -4716,10 +4721,11 @@ static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
     }
   }
 }
+#endif
 
 // Write out dirty registers that we need to reload (pair with load_needed_regs)
 // This writes the registers not written by store_regs_bt
-static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
+static void wb_needed_dirtys(const signed char i_regmap[], u_int i_dirty, int addr)
 {
   int hr;
   int t=(addr-start)>>2;
@@ -4740,6 +4746,7 @@ static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int
 }
 
 // Load all registers (except cycle count)
+#ifndef load_all_regs
 static void load_all_regs(const signed char i_regmap[])
 {
   int hr;
@@ -4756,48 +4763,31 @@ static void load_all_regs(const signed char i_regmap[])
     }
   }
 }
+#endif
 
 // Load all current registers also needed by next instruction
 static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
 {
+  signed char regmap_sel[HOST_REGS];
   int hr;
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(hr!=EXCLUDE_REG) {
-      if(get_reg(next_regmap,i_regmap[hr])>=0) {
-        if(i_regmap[hr]==0) {
-          emit_zeroreg(hr);
-        }
-        else
-        if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
-        {
-          emit_loadreg(i_regmap[hr],hr);
-        }
-      }
-    }
+  for (hr = 0; hr < HOST_REGS; hr++) {
+    regmap_sel[hr] = -1;
+    if (hr != EXCLUDE_REG)
+      if (next_regmap[hr] == i_regmap[hr] || get_reg(next_regmap, i_regmap[hr]) >= 0)
+        regmap_sel[hr] = i_regmap[hr];
   }
+  load_all_regs(regmap_sel);
 }
 
 // Load all regs, storing cycle count if necessary
 static void load_regs_entry(int t)
 {
-  int hr;
   if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
   else if(cinfo[t].ccadj) emit_addimm(HOST_CCREG,-cinfo[t].ccadj,HOST_CCREG);
   if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
     emit_storereg(CCREG,HOST_CCREG);
   }
-  // Load 32-bit regs
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
-      if(regs[t].regmap_entry[hr]==0) {
-        emit_zeroreg(hr);
-      }
-      else if(regs[t].regmap_entry[hr]!=CCREG)
-      {
-        emit_loadreg(regs[t].regmap_entry[hr],hr);
-      }
-    }
-  }
+  load_all_regs(regs[t].regmap_entry);
 }
 
 // Store dirty registers prior to branch
@@ -4942,6 +4932,8 @@ static void drc_dbg_emit_do_cmp(int i, int ccadj_)
   extern void do_insn_cmp();
   //extern int cycle;
   u_int hr, reglist = get_host_reglist(regs[i].regmap);
+  reglist |= get_host_reglist(regs[i].regmap_entry);
+  reglist &= DRC_DBG_REGMASK;
 
   assem_debug("//do_insn_cmp %08x\n", start+i*4);
   save_regs(reglist);
@@ -5090,11 +5082,7 @@ static void do_ccstub(int n)
   assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
   set_jump_target(stubs[n].addr, out);
   int i=stubs[n].b;
-  if(stubs[n].d==NULLDS) {
-    // Delay slot instruction is nullified ("likely" branch)
-    wb_dirtys(regs[i].regmap,regs[i].dirty);
-  }
-  else if(stubs[n].d!=TAKEN) {
+  if (stubs[n].d != TAKEN) {
     wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
   }
   else {
@@ -5259,10 +5247,6 @@ static void do_ccstub(int n)
   }else if(stubs[n].d==NOTTAKEN) {
     if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
     else load_all_regs(branch_regs[i].regmap);
-  }else if(stubs[n].d==NULLDS) {
-    // Delay slot instruction is nullified ("likely" branch)
-    if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
-    else load_all_regs(regs[i].regmap);
   }else{
     load_all_regs(branch_regs[i].regmap);
   }
@@ -6145,7 +6129,7 @@ void disassemble_inst(int i)
     #ifndef REGMAP_PRINT
     return;
     #endif
-    printf("D: %"PRIx64"  WD: %"PRIx64"  U: %"PRIx64"  hC: %x  hWC: %x  hLC: %x\n",
+    printf("D: %x  WD: %x  U: %"PRIx64"  hC: %x  hWC: %x  hLC: %x\n",
       regs[i].dirty, regs[i].wasdirty, unneeded_reg[i],
       regs[i].isconst, regs[i].wasconst, regs[i].loadedconst);
     print_regmap("pre:   ", regmap_pre[i]);
@@ -6302,7 +6286,7 @@ void new_dynarec_init(void)
 #endif
   arch_init();
   new_dynarec_test();
-  ram_offset=(uintptr_t)rdram-0x80000000;
+  ram_offset = (uintptr_t)psxM - 0x80000000;
   if (ram_offset!=0)
     SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
   SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
@@ -6339,12 +6323,13 @@ void new_dynarec_cleanup(void)
 
 static u_int *get_source_start(u_int addr, u_int *limit)
 {
-  if (addr < 0x00200000 ||
-    (0xa0000000 <= addr && addr < 0xa0200000))
+  if (addr < 0x00800000
+      || (0x80000000 <= addr && addr < 0x80800000)
+      || (0xa0000000 <= addr && addr < 0xa0800000))
   {
     // used for BIOS calls mostly?
-    *limit = (addr&0xa0000000)|0x00200000;
-    return (u_int *)(rdram + (addr&0x1fffff));
+    *limit = (addr & 0xa0600000) + 0x00200000;
+    return (u_int *)(psxM + (addr & 0x1fffff));
   }
   else if (!Config.HLE && (
     /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
@@ -6358,10 +6343,6 @@ static u_int *get_source_start(u_int addr, u_int *limit)
     *limit = (addr & 0xfff00000) | 0x80000;
     return (u_int *)((u_char *)psxR + (addr&0x7ffff));
   }
-  else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
-    *limit = (addr & 0x80600000) + 0x00200000;
-    return (u_int *)(rdram + (addr&0x1fffff));
-  }
   return NULL;
 }
 
@@ -6553,6 +6534,7 @@ static void force_intcall(int i)
 static void disassemble_one(int i, u_int src)
 {
     unsigned int type, op, op2, op3;
+    enum ls_width_type ls_type = LS_32;
     memset(&dops[i], 0, sizeof(dops[i]));
     memset(&cinfo[i], 0, sizeof(cinfo[i]));
     cinfo[i].ba = -1;
@@ -6680,20 +6662,20 @@ static void disassemble_one(int i, u_int src)
       case 0x13: set_mnemonic(i, "COP3");
         op2 = (src >> 21) & 0x1f;
         break;
-      case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
-      case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
-      case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
-      case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
-      case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
-      case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
-      case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
-      case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
-      case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
-      case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
-      case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
-      case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
-      case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
-      case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
+      case 0x20: set_mnemonic(i, "LB"); type=LOAD; ls_type = LS_8; break;
+      case 0x21: set_mnemonic(i, "LH"); type=LOAD; ls_type = LS_16; break;
+      case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; ls_type = LS_LR; break;
+      case 0x23: set_mnemonic(i, "LW"); type=LOAD; ls_type = LS_32; break;
+      case 0x24: set_mnemonic(i, "LBU"); type=LOAD; ls_type = LS_8; break;
+      case 0x25: set_mnemonic(i, "LHU"); type=LOAD; ls_type = LS_16; break;
+      case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; ls_type = LS_LR; break;
+      case 0x28: set_mnemonic(i, "SB"); type=STORE; ls_type = LS_8; break;
+      case 0x29: set_mnemonic(i, "SH"); type=STORE; ls_type = LS_16; break;
+      case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; ls_type = LS_LR; break;
+      case 0x2B: set_mnemonic(i, "SW"); type=STORE; ls_type = LS_32; break;
+      case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; ls_type = LS_LR; break;
+      case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; ls_type = LS_32; break;
+      case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; ls_type = LS_32; break;
       case 0x3B:
         if (Config.HLE && (src & 0x03ffffff) < ARRAY_SIZE(psxHLEt)) {
           set_mnemonic(i, "HLECALL");
@@ -6705,8 +6687,9 @@ static void disassemble_one(int i, u_int src)
     }
     if (type == INTCALL)
       SysPrintf("NI %08x @%08x (%08x)\n", src, start + i*4, start);
-    dops[i].itype=type;
-    dops[i].opcode2=op2;
+    dops[i].itype = type;
+    dops[i].opcode2 = op2;
+    dops[i].ls_type = ls_type;
     /* Get registers/immediates */
     dops[i].use_lt1=0;
     gte_rs[i]=gte_rt[i]=0;
@@ -6852,6 +6835,7 @@ static void disassemble_one(int i, u_int src)
 static noinline void pass1_disassemble(u_int pagelimit)
 {
   int i, j, done = 0, ni_count = 0;
+  int ds_next = 0;
 
   for (i = 0; !done; i++)
   {
@@ -6859,6 +6843,7 @@ static noinline void pass1_disassemble(u_int pagelimit)
     unsigned int type, op, op2;
 
     disassemble_one(i, source[i]);
+    dops[i].is_ds = ds_next; ds_next = 0;
     type = dops[i].itype;
     op = dops[i].opcode;
     op2 = dops[i].opcode2;
@@ -6891,6 +6876,7 @@ static noinline void pass1_disassemble(u_int pagelimit)
     dops[i].is_store = type == STORE || type == STORELR || op == 0x3a; // SWC2
     dops[i].is_exception = type == SYSCALL || type == HLECALL || type == INTCALL;
     dops[i].may_except = dops[i].is_exception || (type == ALU && (op2 == 0x20 || op2 == 0x22)) || op == 8;
+    ds_next = dops[i].is_jump;
 
     if (((op & 0x37) == 0x21 || op == 0x25) // LH/SH/LHU
         && ((cinfo[i].imm & 1) || Config.PreciseExceptions))
@@ -6995,7 +6981,8 @@ static noinline void pass1_disassemble(u_int pagelimit)
       // Don't recompile stuff that's already compiled
       if(check_addr(start+i*4+4)) done=1;
       // Don't get too close to the limit
-      if(i>MAXBLOCK/2) done=1;
+      if (i > MAXBLOCK - 64)
+        done = 1;
     }
     if (dops[i].itype == HLECALL)
       stop = 1;
@@ -7015,7 +7002,8 @@ static noinline void pass1_disassemble(u_int pagelimit)
     //assert(i<MAXBLOCK-1);
     if(start+i*4==pagelimit-4) done=1;
     assert(start+i*4<pagelimit);
-    if (i==MAXBLOCK-1) done=1;
+    if (i == MAXBLOCK - 2)
+      done = 1;
     // Stop if we're compiling junk
     if (dops[i].itype == INTCALL && (++ni_count > 8 || dops[i].opcode == 0x11)) {
       done=stop_after_jal=1;
@@ -7178,6 +7166,31 @@ static noinline void pass2_unneeded_regs(int istart,int iend,int r)
   }
 }
 
+static noinline void pass2a_unneeded_other(void)
+{
+  int i, j;
+  for (i = 0; i < slen; i++)
+  {
+    // remove redundant alignment checks
+    if (dops[i].may_except && (dops[i].is_load || dops[i].is_store)
+        && dops[i].rt1 != dops[i].rs1 && !dops[i].is_ds)
+    {
+      int base = dops[i].rs1, lsb = cinfo[i].imm, ls_type = dops[i].ls_type;
+      int mask = ls_type == LS_32 ? 3 : 1;
+      lsb &= mask;
+      for (j = i + 1; j < slen; j++) {
+        if (dops[j].bt || dops[j].is_jump)
+          break;
+        if ((dops[j].is_load || dops[j].is_store) && dops[j].rs1 == base
+            && dops[j].ls_type == ls_type && (cinfo[j].imm & mask) == lsb)
+          dops[j].may_except = 0;
+        if (dops[j].rt1 == base)
+          break;
+      }
+    }
+  }
+}
+
 static noinline void pass3_register_alloc(u_int addr)
 {
   struct regstat current; // Current register allocations/status
@@ -7203,7 +7216,6 @@ static noinline void pass3_register_alloc(u_int addr)
     dops[1].bt=1;
     ds=1;
     unneeded_reg[0]=1;
-    current.regmap[HOST_BTREG]=BTREG;
   }
 
   for(i=0;i<slen;i++)
@@ -7243,7 +7255,7 @@ static noinline void pass3_register_alloc(u_int addr)
         abort();
       }
     }
-    dops[i].is_ds=ds;
+    assert(dops[i].is_ds == ds);
     if(ds) {
       ds=0; // Skip delay slot, already allocated as part of branch
       // ...but we need to alloc it in case something jumps here
@@ -7764,7 +7776,6 @@ static noinline void pass3_register_alloc(u_int addr)
         }
       }
     }
-    if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
     //regs[i].waswritten=current.waswritten;
   }
 }
@@ -8971,6 +8982,8 @@ static int new_recompile_block(u_int addr)
 
   pass2_unneeded_regs(0,slen-1,0);
 
+  pass2a_unneeded_other();
+
   /* Pass 3 - Register allocation */
 
   pass3_register_alloc(addr);
@@ -8987,7 +9000,7 @@ static int new_recompile_block(u_int addr)
   /* Pass 6 - Optimize clean/dirty state */
   pass6_clean_registers(0, slen-1, 1);
 
-  /* Pass 7 - Identify 32-bit registers */
+  /* Pass 7 */
   for (i=slen-1;i>=0;i--)
   {
     if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)