drc: do gte flag liveness detection

[pcsx_rearmed.git] / libpcsxcore / new_dynarec / assem_arm.c

diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c
index af45ce3..adbde59 100644 (file)
--- a/libpcsxcore/new_dynarec/assem_arm.c
+++ b/libpcsxcore/new_dynarec/assem_arm.c
@@ -170,6 +170,7 @@ void set_jump_target_fillslot(int addr,u_int target,int copy)
 /* Literal pool */
 add_literal(int addr,int val)
 {
+  assert(literalcount<sizeof(literals)/sizeof(literals[0]));
   literals[literalcount][0]=addr;
   literals[literalcount][1]=val;
   literalcount++; 
@@ -186,11 +187,15 @@ void *kill_pointer(void *stub)
   return i_ptr;
 }
 
+// find where external branch is liked to using addr of it's stub:
+// get address that insn one after stub loads (dyna_linker arg1),
+// treat it as a pointer to branch insn,
+// return addr where that branch jumps to
 int get_pointer(void *stub)
 {
   //printf("get_pointer(%x)\n",(int)stub);
   int *ptr=(int *)(stub+4);
-  assert((*ptr&0x0ff00000)==0x05900000);
+  assert((*ptr&0x0fff0000)==0x059f0000);
   u_int offset=*ptr&0xfff;
   int **l_ptr=(void *)ptr+offset+8;
   int *i_ptr=*l_ptr;
@@ -222,7 +227,7 @@ int verify_dirty(int addr)
   u_int *ptr=(u_int *)addr;
   #ifdef ARMv5_ONLY
   // get from literal pool
-  assert((*ptr&0xFFF00000)==0xe5900000);
+  assert((*ptr&0xFFFF0000)==0xe59f0000);
   u_int offset=*ptr&0xfff;
   u_int *l_ptr=(void *)ptr+offset+8;
   u_int source=l_ptr[0];
@@ -275,7 +280,7 @@ void get_bounds(int addr,u_int *start,u_int *end)
   u_int *ptr=(u_int *)addr;
   #ifdef ARMv5_ONLY
   // get from literal pool
-  assert((*ptr&0xFFF00000)==0xe5900000);
+  assert((*ptr&0xFFFF0000)==0xe59f0000);
   u_int offset=*ptr&0xfff;
   u_int *l_ptr=(void *)ptr+offset+8;
   u_int source=l_ptr[0];
@@ -1057,7 +1062,7 @@ void emit_test(int rs, int rt)
 void emit_testimm(int rs,int imm)
 {
   u_int armval;
-  assem_debug("tst %s,$%d\n",regname[rs],imm);
+  assem_debug("tst %s,#%d\n",regname[rs],imm);
   genimm_checked(imm,&armval);
   output_w32(0xe3100000|rd_rn_rm(0,rs,0)|armval);
 }
@@ -1450,10 +1455,10 @@ void emit_cmpimm(int rs,int imm)
 {
   u_int armval;
   if(genimm(imm,&armval)) {
-    assem_debug("cmp %s,$%d\n",regname[rs],imm);
+    assem_debug("cmp %s,#%d\n",regname[rs],imm);
     output_w32(0xe3500000|rd_rn_rm(0,rs,0)|armval);
   }else if(genimm(-imm,&armval)) {
-    assem_debug("cmn %s,$%d\n",regname[rs],imm);
+    assem_debug("cmn %s,#%d\n",regname[rs],imm);
     output_w32(0xe3700000|rd_rn_rm(0,rs,0)|armval);
   }else if(imm>0) {
     assert(imm<65536);
@@ -3791,6 +3796,7 @@ void c2op_assemble(int i,struct regstat *i_regs)
   signed char temp=get_reg(i_regs->regmap,-1);
   u_int c2op=source[i]&0x3f;
   u_int hr,reglist=0;
+  int need_flags;
   for(hr=0;hr<HOST_REGS;hr++) {
     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
   }
@@ -3804,7 +3810,13 @@ void c2op_assemble(int i,struct regstat *i_regs)
       emit_addimm(cc,gte_cycletab[c2op]/2,cc); // XXX: could just adjust ccadj?
     emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0); // cop2 regs
     emit_writeword(1,(int)&psxRegs.code);
-    emit_call((int)gte_handlers[c2op]);
+    need_flags=!(gte_unneeded[i+1]>>63); // +1 because of how liveness detection works
+    assem_debug("gte unneeded %016llx, need_flags %d\n",gte_unneeded[i+1],need_flags);
+#ifdef ARMv5_ONLY
+    // let's take more risk here
+    need_flags=need_flags&&gte_reads_flags;
+#endif
+    emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
   }
 
   if(i>=slen-1||itype[i+1]!=C2OP)
@@ -4493,7 +4505,9 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
         assert(quotient>=0);
         assert(remainder>=0);
         emit_movs(d1,remainder);
-        emit_negmi(remainder,remainder);
+        emit_movimm(0xffffffff,quotient);
+        emit_negmi(quotient,quotient); // .. quotient and ..
+        emit_negmi(remainder,remainder); // .. remainder for div0 case (will be negated back after jump)
         emit_movs(d2,HOST_TEMPREG);
         emit_jeq((int)out+52); // Division by zero
         emit_negmi(HOST_TEMPREG,HOST_TEMPREG);
@@ -4521,12 +4535,13 @@ void multdiv_assemble_arm(int i,struct regstat *i_regs)
         signed char remainder=get_reg(i_regs->regmap,HIREG);
         assert(quotient>=0);
         assert(remainder>=0);
+        emit_mov(d1,remainder);
+        emit_movimm(0xffffffff,quotient); // div0 case
         emit_test(d2,d2);
-        emit_jeq((int)out+44); // Division by zero
+        emit_jeq((int)out+40); // Division by zero
         emit_clz(d2,HOST_TEMPREG);
         emit_movimm(1<<31,quotient);
         emit_shl(d2,HOST_TEMPREG,d2);
-        emit_mov(d1,remainder);
         emit_shr(quotient,HOST_TEMPREG,quotient);
         emit_cmp(remainder,d2);
         emit_subcs(remainder,d2,remainder);