X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fassem_arm.c;h=f3f89e1e3764720acf1a9d89996854c6f7596f94;hp=3abef7f5050b5c6760e39e813c98bc7666fcb7ce;hb=cfbd3c6ee21bde6e848eafb3b7994db626b70b72;hpb=514ed0d98e058596720f94af4af347b609980de9

diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c
index 3abef7f5..f3f89e1e 100644
--- a/libpcsxcore/new_dynarec/assem_arm.c
+++ b/libpcsxcore/new_dynarec/assem_arm.c
@@ -142,7 +142,7 @@ add_literal(int addr,int val)
   literalcount++; 
 } 
 
-void kill_pointer(void *stub)
+void *kill_pointer(void *stub)
 {
   int *ptr=(int *)(stub+4);
   assert((*ptr&0x0ff00000)==0x05900000);
@@ -150,6 +150,7 @@ void kill_pointer(void *stub)
   int **l_ptr=(void *)ptr+offset+8;
   int *i_ptr=*l_ptr;
   set_jump_target((int)i_ptr,(int)stub);
+  return i_ptr;
 }
 
 int get_pointer(void *stub)
@@ -823,6 +824,11 @@ u_int genimm(u_int imm,u_int *encoded)
   }
   return 0;
 }
+void genimm_checked(u_int imm,u_int *encoded)
+{
+  u_int ret=genimm(imm,encoded);
+  assert(ret);
+}
 u_int genjmp(u_int addr)
 {
   int offset=addr-(int)out-8;
@@ -961,7 +967,7 @@ void emit_testimm(int rs,int imm)
 {
   u_int armval;
   assem_debug("tst %s,$%d\n",regname[rs],imm);
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   output_w32(0xe3100000|rd_rn_rm(0,rs,0)|armval);
 }
 
@@ -969,7 +975,7 @@ void emit_testeqimm(int rs,int imm)
 {
   u_int armval;
   assem_debug("tsteq %s,$%d\n",regname[rs],imm);
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   output_w32(0x03100000|rd_rn_rm(0,rs,0)|armval);
 }
 
@@ -1002,6 +1008,15 @@ void emit_or_and_set_flags(int rs1,int rs2,int rt)
   output_w32(0xe1900000|rd_rn_rm(rt,rs1,rs2));
 }
 
+void emit_orrshr_imm(u_int rs,u_int imm,u_int rt)
+{
+  assert(rs<16);
+  assert(rt<16);
+  assert(imm<32);
+  assem_debug("orr %s,%s,%s,lsr #%d\n",regname[rt],regname[rt],regname[rs],imm);
+  output_w32(0xe1800020|rd_rn_rm(rt,rt,rs)|(imm<<7));
+}
+
 void emit_xor(u_int rs1,u_int rs2,u_int rt)
 {
   assem_debug("eor %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]);
@@ -1123,14 +1138,14 @@ void emit_addnop(u_int r)
 void emit_adcimm(u_int rs,int imm,u_int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("adc %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0xe2a00000|rd_rn_rm(rt,rs,0)|armval);
 }
 /*void emit_sbcimm(int imm,u_int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("sbc %s,%s,#%d\n",regname[rt],regname[rt],imm);
   output_w32(0xe2c00000|rd_rn_rm(rt,rt,0)|armval);
 }*/
@@ -1154,7 +1169,7 @@ void emit_rscimm(int rs,int imm,u_int rt)
 {
   assert(0);
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("rsc %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0xe2e00000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -1424,28 +1439,28 @@ void emit_cmovne_imm(int imm,int rt)
 {
   assem_debug("movne %s,#%d\n",regname[rt],imm);
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   output_w32(0x13a00000|rd_rn_rm(rt,0,0)|armval);
 }
 void emit_cmovl_imm(int imm,int rt)
 {
   assem_debug("movlt %s,#%d\n",regname[rt],imm);
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   output_w32(0xb3a00000|rd_rn_rm(rt,0,0)|armval);
 }
 void emit_cmovb_imm(int imm,int rt)
 {
   assem_debug("movcc %s,#%d\n",regname[rt],imm);
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   output_w32(0x33a00000|rd_rn_rm(rt,0,0)|armval);
 }
 void emit_cmovs_imm(int imm,int rt)
 {
   assem_debug("movmi %s,#%d\n",regname[rt],imm);
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   output_w32(0x43a00000|rd_rn_rm(rt,0,0)|armval);
 }
 void emit_cmove_reg(int rs,int rt)
@@ -1937,7 +1952,7 @@ void emit_writebyte(int rt, int addr)
 {
   u_int offset = addr-(u_int)&dynarec_local;
   assert(offset<4096);
-  assem_debug("str %s,fp+%d\n",regname[rt],offset);
+  assem_debug("strb %s,fp+%d\n",regname[rt],offset);
   output_w32(0xe5c00000|rd_rn_rm(rt,FP,0)|offset);
 }
 void emit_writeword_imm(int imm, int addr)
@@ -2085,7 +2100,7 @@ void emit_teq(int rs, int rt)
 void emit_rsbimm(int rs, int imm, int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("rsb %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0xe2600000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2351,7 +2366,7 @@ void emit_fmstat()
 void emit_bicne_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("bicne %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x13c00000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2359,7 +2374,7 @@ void emit_bicne_imm(int rs,int imm,int rt)
 void emit_biccs_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("biccs %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x23c00000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2367,7 +2382,7 @@ void emit_biccs_imm(int rs,int imm,int rt)
 void emit_bicvc_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("bicvc %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x73c00000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2375,7 +2390,7 @@ void emit_bicvc_imm(int rs,int imm,int rt)
 void emit_bichi_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("bichi %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x83c00000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2383,7 +2398,7 @@ void emit_bichi_imm(int rs,int imm,int rt)
 void emit_orrvs_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("orrvs %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x63800000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2391,7 +2406,7 @@ void emit_orrvs_imm(int rs,int imm,int rt)
 void emit_orrne_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("orrne %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x13800000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2399,7 +2414,7 @@ void emit_orrne_imm(int rs,int imm,int rt)
 void emit_andne_imm(int rs,int imm,int rt)
 {
   u_int armval;
-  assert(genimm(imm,&armval));
+  genimm_checked(imm,&armval);
   assem_debug("andne %s,%s,#%d\n",regname[rt],regname[rs],imm);
   output_w32(0x12000000|rd_rn_rm(rt,rs,0)|armval);
 }
@@ -2554,8 +2569,10 @@ do_readstub(int n)
     rt=get_reg(i_regmap,rt1[i]);
   }
   assert(rs>=0);
-  assert(rt>=0);
   if(addr<0) addr=rt;
+  if(addr<0)
+    // assume dummy read, no alloced reg
+    addr=get_reg(i_regmap,-1);
   assert(addr>=0);
   int ftable=0;
   if(type==LOADB_STUB||type==LOADBU_STUB)
@@ -2608,19 +2625,22 @@ do_readstub(int n)
   //if((cc=get_reg(regmap,CCREG))>=0) {
   //  emit_loadreg(CCREG,cc);
   //}
-  if(type==LOADB_STUB)
-    emit_movsbl((int)&readmem_dword,rt);
-  if(type==LOADBU_STUB)
-    emit_movzbl((int)&readmem_dword,rt);
-  if(type==LOADH_STUB)
-    emit_movswl((int)&readmem_dword,rt);
-  if(type==LOADHU_STUB)
-    emit_movzwl((int)&readmem_dword,rt);
-  if(type==LOADW_STUB)
-    emit_readword((int)&readmem_dword,rt);
-  if(type==LOADD_STUB) {
-    emit_readword((int)&readmem_dword,rt);
-    if(rth>=0) emit_readword(((int)&readmem_dword)+4,rth);
+  if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) {
+    assert(rt>=0);
+    if(type==LOADB_STUB)
+      emit_movsbl((int)&readmem_dword,rt);
+    if(type==LOADBU_STUB)
+      emit_movzbl((int)&readmem_dword,rt);
+    if(type==LOADH_STUB)
+      emit_movswl((int)&readmem_dword,rt);
+    if(type==LOADHU_STUB)
+      emit_movzwl((int)&readmem_dword,rt);
+    if(type==LOADW_STUB)
+      emit_readword((int)&readmem_dword,rt);
+    if(type==LOADD_STUB) {
+      emit_readword((int)&readmem_dword,rt);
+      if(rth>=0) emit_readword(((int)&readmem_dword)+4,rth);
+    }
   }
   emit_jmp(stubs[n][2]); // return address
 }
@@ -2855,8 +2875,81 @@ inline_writestub(int type, int i, u_int addr, signed char regmap[], int target,
 
 do_unalignedwritestub(int n)
 {
+  assem_debug("do_unalignedwritestub %x\n",start+stubs[n][3]*4);
+  literal_pool(256);
   set_jump_target(stubs[n][1],(int)out);
-  output_w32(0xef000000);
+
+  int i=stubs[n][3];
+  struct regstat *i_regs=(struct regstat *)stubs[n][4];
+  int addr=stubs[n][5];
+  u_int reglist=stubs[n][7];
+  signed char *i_regmap=i_regs->regmap;
+  int temp2=get_reg(i_regmap,FTEMP);
+  int rt;
+  int ds, real_rs;
+  rt=get_reg(i_regmap,rs2[i]);
+  assert(rt>=0);
+  assert(addr>=0);
+  assert(opcode[i]==0x2a||opcode[i]==0x2e); // SWL/SWR only implemented
+  reglist|=(1<<addr);
+  reglist&=~(1<<temp2);
+
+  emit_andimm(addr,0xfffffffc,temp2);
+  emit_writeword(temp2,(int)&address);
+
+  save_regs(reglist);
+  ds=i_regs!=&regs[i];
+  real_rs=get_reg(i_regmap,rs1[i]);
+  u_int cmask=ds?-1:(0x100f|~i_regs->wasconst);
+  if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i);
+  wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs)));
+  if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i);
+  emit_shrimm(addr,16,1);
+  int cc=get_reg(i_regmap,CCREG);
+  if(cc<0) {
+    emit_loadreg(CCREG,2);
+  }
+  emit_movimm((u_int)readmem,0);
+  emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2);
+  emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3); // XXX: can be rm'd?
+  emit_call((int)&indirect_jump_indexed);
+  restore_regs(reglist);
+
+  emit_readword((int)&readmem_dword,temp2);
+  int temp=addr; //hmh
+  emit_shlimm(addr,3,temp);
+  emit_andimm(temp,24,temp);
+#ifdef BIG_ENDIAN_MIPS
+  if (opcode[i]==0x2e) // SWR
+#else
+  if (opcode[i]==0x2a) // SWL
+#endif
+    emit_xorimm(temp,24,temp);
+  emit_movimm(-1,HOST_TEMPREG);
+  if (opcode[i]==0x2a) { // SWL
+    emit_bic_lsr(temp2,HOST_TEMPREG,temp,temp2);
+    emit_orrshr(rt,temp,temp2);
+  }else{
+    emit_bic_lsl(temp2,HOST_TEMPREG,temp,temp2);
+    emit_orrshl(rt,temp,temp2);
+  }
+  emit_readword((int)&address,addr);
+  emit_writeword(temp2,(int)&word);
+  //save_regs(reglist); // don't need to, no state changes
+  emit_shrimm(addr,16,1);
+  emit_movimm((u_int)writemem,0);
+  //emit_call((int)&indirect_jump_indexed);
+  emit_mov(15,14);
+  emit_readword_dualindexedx4(0,1,15);
+  emit_readword((int)&Count,HOST_TEMPREG);
+  emit_readword((int)&next_interupt,2);
+  emit_addimm(HOST_TEMPREG,-2*stubs[n][6]-2,HOST_TEMPREG);
+  emit_writeword(2,(int)&last_count);
+  emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
+  if(cc<0) {
+    emit_storereg(CCREG,HOST_TEMPREG);
+  }
+  restore_regs(reglist);
   emit_jmp(stubs[n][2]); // return address
 }
 
@@ -2880,15 +2973,19 @@ do_invstub(int n)
 int do_dirty_stub(int i)
 {
   assem_debug("do_dirty_stub %x\n",start+i*4);
+  u_int addr=(int)start<(int)0xC0000000?(u_int)source:(u_int)start;
+  #ifdef PCSX
+  addr=(u_int)source;
+  #endif
   // Careful about the code output here, verify_dirty needs to parse it.
   #ifdef ARMv5_ONLY
-  emit_loadlp((int)start<(int)0xC0000000?(int)source:(int)start,1);
+  emit_loadlp(addr,1);
   emit_loadlp((int)copy,2);
   emit_loadlp(slen*4,3);
   #else
-  emit_movw(((int)start<(int)0xC0000000?(u_int)source:(u_int)start)&0x0000FFFF,1);
+  emit_movw(addr&0x0000FFFF,1);
   emit_movw(((u_int)copy)&0x0000FFFF,2);
-  emit_movt(((int)start<(int)0xC0000000?(u_int)source:(u_int)start)&0xFFFF0000,1);
+  emit_movt(addr&0xFFFF0000,1);
   emit_movt(((u_int)copy)&0xFFFF0000,2);
   emit_movw(slen*4,3);
   #endif
@@ -3159,7 +3256,7 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs)
   else addr=s;
   if(s>=0) {
     c=(i_regs->wasconst>>s)&1;
-    memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80800000;
+    memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
     if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
   }
   if(tl>=0) {
@@ -3173,7 +3270,7 @@ void loadlr_assemble_arm(int i,struct regstat *i_regs)
         }else{
           emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
         }
-        emit_cmpimm(addr,0x800000);
+        emit_cmpimm(addr,RAM_SIZE);
         jaddr=(int)out;
         emit_jno(0);
       }
@@ -3281,7 +3378,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
     signed char t=get_reg(i_regs->regmap,rt1[i]);
     char copr=(source[i]>>11)&0x1f;
     //assert(t>=0); // Why does this happen?  OOT is weird
-    if(t>=0) {
+    if(t>=0&&rt1[i]!=0) {
 #ifdef MUPEN64
       emit_addimm(FP,(int)&fake_pc-(int)&dynarec_local,0);
       emit_movimm((source[i]>>11)&0x1f,1);
@@ -3380,6 +3477,16 @@ void cop0_assemble(int i,struct regstat *i_regs)
     if((source[i]&0x3f)==0x08) // TLBP
       emit_call((int)TLBP);
 #endif
+#ifdef PCSX
+    if((source[i]&0x3f)==0x10) // RFE
+    {
+      emit_readword((int)&Status,0);
+      emit_andimm(0,0x3c,1);
+      emit_andimm(0,~0xf,0);
+      emit_orrshr_imm(1,2,0);
+      emit_writeword(0,(int)&Status);
+    }
+#else
     if((source[i]&0x3f)==0x18) // ERET
     {
       int count=ccadj[i];
@@ -3387,6 +3494,7 @@ void cop0_assemble(int i,struct regstat *i_regs)
       emit_addimm(HOST_CCREG,CLOCK_DIVIDER*count,HOST_CCREG); // TODO: Should there be an extra cycle here?
       emit_jmp((int)jump_eret);
     }
+#endif
   }
 }
 
@@ -3491,7 +3599,7 @@ void cop2_assemble(int i,struct regstat *i_regs)
   signed char temp=get_reg(i_regs->regmap,-1);
   if (opcode2[i]==0) { // MFC2
     signed char tl=get_reg(i_regs->regmap,rt1[i]);
-    if(tl>=0)
+    if(tl>=0&&rt1[i]!=0)
       cop2_get_dreg(copr,tl,temp);
   }
   else if (opcode2[i]==4) { // MTC2
@@ -3501,7 +3609,7 @@ void cop2_assemble(int i,struct regstat *i_regs)
   else if (opcode2[i]==2) // CFC2
   {
     signed char tl=get_reg(i_regs->regmap,rt1[i]);
-    if(tl>=0)
+    if(tl>=0&&rt1[i]!=0)
       emit_readword((int)&reg_cop2c[copr],tl);
   }
   else if (opcode2[i]==6) // CTC2