/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
* Mupen64plus - new_dynarec.c *
- * Copyright (C) 2009-2010 Ari64 *
+ * Copyright (C) 2009-2011 Ari64 *
* *
* This program is free software; you can redistribute it and/or modify *
* it under the terms of the GNU General Public License as published by *
#else
static const u_int using_tlb=0;
#endif
+ static u_int sp_in_mirror;
u_int stop_after_jal;
extern u_char restore_candidate[512];
extern int cycle_count;
int is_const(struct regstat *cur,signed char reg)
{
int hr;
+ if(reg<0) return 0;
if(!reg) return 1;
for (hr=0;hr<HOST_REGS;hr++) {
if((cur->regmap[hr]&63)==reg) {
int j;
int b=-1;
int rn=10;
- int hr;
- u_char hsn[MAXREG+1];
- int preferred_reg;
-
- memset(hsn,10,sizeof(hsn));
- lsn(hsn,i,&preferred_reg);
if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
{
}
}
}*/
- for(hr=0;hr<HOST_REGS;hr++) {
- if(hr!=EXCLUDE_REG) {
- if(rn<hsn[hr]) return 1;
- }
- }
+ if(rn<10) return 1;
return 0;
}
//if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
- if(rt1[i]) {
+ if(rt1[i]&&!((current->u>>rt1[i])&1)) {
alloc_reg(current,i,rt1[i]);
- if(get_reg(current->regmap,rt1[i])<0) {
- // dummy load, but we still need a register to calculate the address
- alloc_reg_temp(current,i,-1);
- minimum_free_regs[i]=1;
- }
+ assert(get_reg(current->regmap,rt1[i])>=0);
if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
{
current->is32&=~(1LL<<rt1[i]);
}
else
{
- // Load to r0 (dummy load)
+ // Load to r0 or unneeded register (dummy load)
// but we still need a register to calculate the address
if(opcode[i]==0x22||opcode[i]==0x26)
{
alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
}
+ // If using TLB, need a register for pointer to the mapping table
+ if(using_tlb) alloc_reg(current,i,TLREG);
alloc_reg_temp(current,i,-1);
minimum_free_regs[i]=1;
if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
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);
+ }
+ else
+ #endif
emit_cmpimm(addr,RAM_SIZE);
jaddr=(int)out;
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if (opcode[i]==0x21||opcode[i]==0x25) x=2; // LH/LHU
map=get_reg(i_regs->regmap,TLREG);
assert(map>=0);
+ reglist&=~(1<<map);
map=do_tlb_r(addr,tl,map,x,-1,-1,c,constmap[i][s]+offset);
do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr);
}
else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
#else
if(!c) a=addr;
+#endif
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
#endif
emit_movsbl_indexed_tlb(x,a,map,tl);
}
else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
#else
if(!c) a=addr;
+#endif
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
#endif
//#ifdef
//emit_movswl_indexed_tlb(x,tl,map,tl);
if (opcode[i]==0x23) { // LW
if(!c||memtarget) {
if(!dummy) {
+ int a=addr;
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
//emit_readword_indexed((int)rdram-0x80000000,addr,tl);
#ifdef HOST_IMM_ADDR32
if(c)
emit_readword_tlb(constmap[i][s]+offset,map,tl);
else
#endif
- emit_readword_indexed_tlb(0,addr,map,tl);
+ emit_readword_indexed_tlb(0,a,map,tl);
}
if(jaddr)
add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
#else
if(!c) a=addr;
+#endif
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
#endif
emit_movzbl_indexed_tlb(x,a,map,tl);
}
else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
#else
if(!c) a=addr;
+#endif
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
#endif
//#ifdef
//emit_movzwl_indexed_tlb(x,tl,map,tl);
assert(th>=0);
if(!c||memtarget) {
if(!dummy) {
+ int a=addr;
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
//emit_readword_indexed((int)rdram-0x80000000,addr,tl);
#ifdef HOST_IMM_ADDR32
if(c)
emit_readword_tlb(constmap[i][s]+offset,map,tl);
else
#endif
- emit_readword_indexed_tlb(0,addr,map,tl);
+ emit_readword_indexed_tlb(0,a,map,tl);
}
if(jaddr)
add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
if (opcode[i]==0x37) { // LD
if(!c||memtarget) {
if(!dummy) {
+ int a=addr;
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
//gen_tlb_addr_r(tl,map);
//if(th>=0) emit_readword_indexed((int)rdram-0x80000000,addr,th);
//emit_readword_indexed((int)rdram-0x7FFFFFFC,addr,tl);
emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
else
#endif
- emit_readdword_indexed_tlb(0,addr,map,th,tl);
+ emit_readdword_indexed_tlb(0,a,map,th,tl);
}
if(jaddr)
add_stub(LOADD_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
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);
+ }
+ else
+ #endif
#ifdef R29_HACK
// Strmnnrmn's speed hack
- memtarget=1;
if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
#endif
emit_cmpimm(addr,RAM_SIZE);
if(s==addr) emit_mov(s,temp);
#endif
#ifdef R29_HACK
+ memtarget=1;
if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
#endif
{
if (opcode[i]==0x29) x=2; // SH
map=get_reg(i_regs->regmap,TLREG);
assert(map>=0);
+ reglist&=~(1<<map);
map=do_tlb_w(addr,temp,map,x,c,constmap[i][s]+offset);
do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
}
else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
#else
if(!c) a=addr;
+#endif
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
#endif
//gen_tlb_addr_w(temp,map);
//emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
#else
if(!c) a=addr;
+#endif
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
#endif
//#ifdef
//emit_writehword_indexed_tlb(tl,x,temp,map,temp);
type=STOREH_STUB;
}
if (opcode[i]==0x2B) { // SW
- if(!c||memtarget)
+ if(!c||memtarget) {
+ int a=addr;
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
//emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
- emit_writeword_indexed_tlb(tl,0,addr,map,temp);
+ emit_writeword_indexed_tlb(tl,0,a,map,temp);
+ }
type=STOREW_STUB;
}
if (opcode[i]==0x3F) { // SD
if(!c||memtarget) {
+ int a=addr;
+#ifdef PCSX
+ if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
if(rs2[i]) {
assert(th>=0);
//emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
//emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,addr);
- emit_writedword_indexed_tlb(th,tl,0,addr,map,temp);
+ emit_writedword_indexed_tlb(th,tl,0,a,map,temp);
}else{
// Store zero
//emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
//emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
- emit_writedword_indexed_tlb(tl,tl,0,addr,map,temp);
+ emit_writedword_indexed_tlb(tl,tl,0,a,map,temp);
}
}
type=STORED_STUB;
//if(opcode[i]==0x2B)
/*if(opcode[i]==0x2B || opcode[i]==0x28 || opcode[i]==0x29 || opcode[i]==0x3F)
{
- //emit_pusha();
+ #ifdef __i386__
+ emit_pusha();
+ #endif
+ #ifdef __arm__
save_regs(0x100f);
+ #endif
emit_readword((int)&last_count,ECX);
#ifdef __i386__
if(get_reg(i_regs->regmap,CCREG)<0)
emit_writeword(0,(int)&Count);
#endif
emit_call((int)memdebug);
- //emit_popa();
+ #ifdef __i386__
+ emit_popa();
+ #endif
+ #ifdef __arm__
restore_regs(0x100f);
+ #endif
}/**/
}
}else{ // using tlb
int map=get_reg(i_regs->regmap,TLREG);
assert(map>=0);
+ reglist&=~(1<<map);
map=do_tlb_w(c||s<0||offset?temp:s,temp,map,0,c,constmap[i][s]+offset);
if(!c&&!offset&&s>=0) emit_mov(s,temp);
do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
{
map=get_reg(i_regs->regmap,TLREG);
assert(map>=0);
+ reglist&=~(1<<map);
if (opcode[i]==0x31||opcode[i]==0x35) { // LWC1/LDC1
map=do_tlb_r(offset||c||s<0?ar:s,ar,map,0,-1,-1,c,constmap[i][s]+offset);
}
int ar;
int offset;
int memtarget=0,c=0;
- int jaddr,jaddr2=0,jaddr3,type;
+ int jaddr2=0,jaddr3,type;
int agr=AGEN1+(i&1);
u_int hr,reglist=0;
u_int copr=(source[i]>>16)&0x1f;
void address_generation(int i,struct regstat *i_regs,signed char entry[])
{
if(itype[i]==LOAD||itype[i]==LOADLR||itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS||itype[i]==C2LS) {
- int ra;
+ int ra=-1;
int agr=AGEN1+(i&1);
int mgr=MGEN1+(i&1);
if(itype[i]==LOAD) {
emit_zeroreg(hr);
}
else
- if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG)
+ if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
{
emit_loadreg(i_regmap[hr],hr);
}
emit_zeroreg(hr);
}
else
- if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG)
+ if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
{
emit_loadreg(i_regmap[hr],hr);
}
}
// Load 32-bit regs
for(hr=0;hr<HOST_REGS;hr++) {
- if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<64) {
+ if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
if(regs[t].regmap_entry[hr]==0) {
emit_zeroreg(hr);
}
}
// Load 64-bit regs
for(hr=0;hr<HOST_REGS;hr++) {
- if(regs[t].regmap_entry[hr]>=64) {
+ if(regs[t].regmap_entry[hr]>=64&®s[t].regmap_entry[hr]<TEMPREG+64) {
assert(regs[t].regmap_entry[hr]!=64);
if((regs[t].was32>>(regs[t].regmap_entry[hr]&63))&1) {
int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
}
// Load 32-bit regs
for(hr=0;hr<HOST_REGS;hr++) {
- if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<64) {
+ if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
#ifdef DESTRUCTIVE_WRITEBACK
if(i_regmap[hr]!=regs[t].regmap_entry[hr] || ( !((regs[t].dirty>>hr)&1) && ((i_dirty>>hr)&1) && (((i_is32&~unneeded_reg_upper[t])>>i_regmap[hr])&1) ) || (((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
#else
}
//Load 64-bit regs
for(hr=0;hr<HOST_REGS;hr++) {
- if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=64) {
+ if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=64&®s[t].regmap_entry[hr]<TEMPREG+64) {
if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
assert(regs[t].regmap_entry[hr]!=64);
if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
{
if(i_regmap[hr]!=regs[t].regmap_entry[hr])
{
- if(regs[t].regmap_entry[hr]!=-1)
+ if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
{
return 0;
}
else
if((i_dirty>>hr)&1)
{
- if(i_regmap[hr]<64)
+ if(i_regmap[hr]<TEMPREG)
{
if(!((unneeded_reg[t]>>i_regmap[hr])&1))
return 0;
}
- else
+ else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
{
if(!((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1))
return 0;
}
else
{
- emit_cmpimm(HOST_CCREG,-2*(count+2));
+ emit_cmpimm(HOST_CCREG,-CLOCK_DIVIDER*(count+2));
jaddr=(int)out;
emit_jns(0);
}
emit_loadreg(rs2[i],s2l);
#endif
int hr=0;
- int addr,alt,ntaddr;
+ int addr=-1,alt=-1,ntaddr=-1;
while(hr<HOST_REGS)
{
if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
}
+ if(branch_regs[i].regmap[r]>=0) {
+ will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
+ wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
+ }
}
}
//}
//if(ba[i]>start+i*4) { // Disable recursion (for debugging)
for(r=0;r<HOST_REGS;r++) {
if(r!=EXCLUDE_REG) {
- if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
+ signed char target_reg=branch_regs[i].regmap[r];
+ if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
}
- else
- {
- will_dirty_i&=~(1<<r);
+ else if(target_reg>=0) {
+ will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
+ wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
}
// Treat delay slot as part of branch too
/*if(regs[i+1].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
}
}
}
- // Merge in delay slot
+ // Merge in delay slot (won't dirty)
for(r=0;r<HOST_REGS;r++) {
if(r!=EXCLUDE_REG) {
if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
regs[i].wasdirty|=will_dirty_i&(1<<r);
}
}
- else if((nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
+ else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
// Register moved to a different register
will_dirty_i&=~(1<<r);
wont_dirty_i&=~(1<<r);
void new_dynarec_clear_full()
{
int n;
- for(n=0x80000;n<0x80800;n++)
- invalid_code[n]=1;
- for(n=0;n<65536;n++)
- hash_table[n][0]=hash_table[n][2]=-1;
+ out=(u_char *)BASE_ADDR;
+ memset(invalid_code,1,sizeof(invalid_code));
+ memset(hash_table,0xff,sizeof(hash_table));
memset(mini_ht,-1,sizeof(mini_ht));
memset(restore_candidate,0,sizeof(restore_candidate));
memset(shadow,0,sizeof(shadow));
#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
start = (u_int)addr&~3;
//assert(((u_int)addr&1)==0);
#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?
case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
- case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
- case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
- case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
- case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
- case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
- case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
- case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
- case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
- case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
- case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
- case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
+#ifndef FORCE32
+ case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
+ case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
+ case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
+ case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
+ case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
+ case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
+ case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
+ case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
+ case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
+ case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
+ case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
+#endif
}
break;
case 0x01: strcpy(insn[i],"regimm"); type=NI;
// Does the block continue due to a branch?
for(j=i-1;j>=0;j--)
{
+ if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
if(ba[j]==start+i*4+4) done=j=0;
if(ba[j]==start+i*4+8) done=j=0;
}
current.wasconst=0;
int ds=0;
int cc=0;
- int hr;
+ int hr=-1;
#ifndef FORCE32
provisional_32bit();
}
if(temp_is32!=current.is32) {
//printf("dumping 32-bit regs (%x)\n",start+i*4);
- #ifdef DESTRUCTIVE_WRITEBACK
+ #ifndef DESTRUCTIVE_WRITEBACK
+ if(ds)
+ #endif
for(hr=0;hr<HOST_REGS;hr++)
{
int r=current.regmap[hr];
}
}
}
- #endif
current.is32=temp_is32;
}
}
clear_const(¤t,rt1[i]);
alloc_cc(¤t,i);
dirty_reg(¤t,CCREG);
+ ooo[i]=1;
+ delayslot_alloc(¤t,i+1);
if (rt1[i]==31) {
alloc_reg(¤t,i,31);
dirty_reg(¤t,31);
#endif
//current.is32|=1LL<<rt1[i];
}
- ooo[i]=1;
- delayslot_alloc(¤t,i+1);
//current.isconst=0; // DEBUG
ds=1;
//printf("i=%d, isconst=%x\n",i,current.isconst);
}
}
// Don't need stuff which is overwritten
- if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
- if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
+ //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
+ //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
// Merge in delay slot
for(hr=0;hr<HOST_REGS;hr++)
{
if(likely[i]) {
regs[i].regmap[hr]=-1;
regs[i].isconst&=~(1<<hr);
- if(i<slen-2) regmap_pre[i+2][hr]=-1;
+ if(i<slen-2) {
+ regmap_pre[i+2][hr]=-1;
+ regs[i+2].wasconst&=~(1<<hr);
+ }
}
}
}
{
if(!likely[i]&&i<slen-2) {
regmap_pre[i+2][hr]=-1;
+ regs[i+2].wasconst&=~(1<<hr);
}
}
}
}
regmap_pre[i+1][hr]=-1;
if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
+ regs[i+1].wasconst&=~(1<<hr);
}
regs[i].regmap[hr]=-1;
regs[i].isconst&=~(1<<hr);
// If a register is allocated during a loop, try to allocate it for the
// entire loop, if possible. This avoids loading/storing registers
// inside of the loop.
-
+
signed char f_regmap[HOST_REGS];
clear_all_regs(f_regmap);
for(i=0;i<slen-1;i++)
{
int t=(ba[i]-start)>>2;
if(t>0&&(itype[t-1]!=UJUMP&&itype[t-1]!=RJUMP&&itype[t-1]!=CJUMP&&itype[t-1]!=SJUMP&&itype[t-1]!=FJUMP)) // loop_preload can't handle jumps into delay slots
- if(t<2||(itype[t-2]!=UJUMP)) // call/ret assumes no registers allocated
+ if(t<2||(itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||rt1[t-2]!=31) // call/ret assumes no registers allocated
for(hr=0;hr<HOST_REGS;hr++)
{
if(regs[i].regmap[hr]>64) {
// a mov, which is of negligible benefit. So such cases are
// skipped below.
if(f_regmap[hr]>0) {
- if(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0) {
+ if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
int r=f_regmap[hr];
for(j=t;j<=i;j++)
{
break;
}
// call/ret fast path assumes no registers allocated
- if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)) {
+ if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)&&rt1[k-3]==31) {
break;
}
if(r>63) {
}
}
}else{
- int count=0;
+ // Non branch or undetermined branch target
for(hr=0;hr<HOST_REGS;hr++)
{
if(hr!=EXCLUDE_REG) {
f_regmap[hr]=regs[i].regmap[hr];
}
}
- else if(regs[i].regmap[hr]<0) count++;
}
}
// Try to restore cycle count at branch targets
}
}
+ // Cache memory offset or tlb map pointer if a register is available
+ #ifndef HOST_IMM_ADDR32
+ #ifndef RAM_OFFSET
+ if(using_tlb)
+ #endif
+ {
+ int earliest_available[HOST_REGS];
+ int loop_start[HOST_REGS];
+ int score[HOST_REGS];
+ int end[HOST_REGS];
+ int reg=using_tlb?MMREG:ROREG;
+
+ // Init
+ for(hr=0;hr<HOST_REGS;hr++) {
+ score[hr]=0;earliest_available[hr]=0;
+ loop_start[hr]=MAXBLOCK;
+ }
+ for(i=0;i<slen-1;i++)
+ {
+ // Can't do anything if no registers are available
+ if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i]) {
+ for(hr=0;hr<HOST_REGS;hr++) {
+ score[hr]=0;earliest_available[hr]=i+1;
+ loop_start[hr]=MAXBLOCK;
+ }
+ }
+ if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
+ if(!ooo[i]) {
+ if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1]) {
+ for(hr=0;hr<HOST_REGS;hr++) {
+ score[hr]=0;earliest_available[hr]=i+1;
+ loop_start[hr]=MAXBLOCK;
+ }
+ }
+ }else{
+ if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1]) {
+ for(hr=0;hr<HOST_REGS;hr++) {
+ score[hr]=0;earliest_available[hr]=i+1;
+ loop_start[hr]=MAXBLOCK;
+ }
+ }
+ }
+ }
+ // Mark unavailable registers
+ for(hr=0;hr<HOST_REGS;hr++) {
+ if(regs[i].regmap[hr]>=0) {
+ score[hr]=0;earliest_available[hr]=i+1;
+ loop_start[hr]=MAXBLOCK;
+ }
+ if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
+ if(branch_regs[i].regmap[hr]>=0) {
+ score[hr]=0;earliest_available[hr]=i+2;
+ loop_start[hr]=MAXBLOCK;
+ }
+ }
+ }
+ // No register allocations after unconditional jumps
+ if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
+ {
+ for(hr=0;hr<HOST_REGS;hr++) {
+ score[hr]=0;earliest_available[hr]=i+2;
+ loop_start[hr]=MAXBLOCK;
+ }
+ i++; // Skip delay slot too
+ //printf("skip delay slot: %x\n",start+i*4);
+ }
+ else
+ // Possible match
+ if(itype[i]==LOAD||itype[i]==LOADLR||
+ itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS) {
+ for(hr=0;hr<HOST_REGS;hr++) {
+ if(hr!=EXCLUDE_REG) {
+ end[hr]=i-1;
+ for(j=i;j<slen-1;j++) {
+ if(regs[j].regmap[hr]>=0) break;
+ if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
+ if(branch_regs[j].regmap[hr]>=0) break;
+ if(ooo[j]) {
+ if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1]) break;
+ }else{
+ if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1]) break;
+ }
+ }
+ else if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) break;
+ if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
+ int t=(ba[j]-start)>>2;
+ if(t<j&&t>=earliest_available[hr]) {
+ if(t==1||(t>1&&itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||(t>1&&rt1[t-2]!=31)) { // call/ret assumes no registers allocated
+ // Score a point for hoisting loop invariant
+ if(t<loop_start[hr]) loop_start[hr]=t;
+ //printf("set loop_start: i=%x j=%x (%x)\n",start+i*4,start+j*4,start+t*4);
+ score[hr]++;
+ end[hr]=j;
+ }
+ }
+ else if(t<j) {
+ if(regs[t].regmap[hr]==reg) {
+ // Score a point if the branch target matches this register
+ score[hr]++;
+ end[hr]=j;
+ }
+ }
+ if(itype[j+1]==LOAD||itype[j+1]==LOADLR||
+ itype[j+1]==STORE||itype[j+1]==STORELR||itype[j+1]==C1LS) {
+ score[hr]++;
+ end[hr]=j;
+ }
+ }
+ if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
+ {
+ // Stop on unconditional branch
+ break;
+ }
+ else
+ if(itype[j]==LOAD||itype[j]==LOADLR||
+ itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS) {
+ score[hr]++;
+ end[hr]=j;
+ }
+ }
+ }
+ }
+ // Find highest score and allocate that register
+ int maxscore=0;
+ for(hr=0;hr<HOST_REGS;hr++) {
+ if(hr!=EXCLUDE_REG) {
+ if(score[hr]>score[maxscore]) {
+ maxscore=hr;
+ //printf("highest score: %d %d (%x->%x)\n",score[hr],hr,start+i*4,start+end[hr]*4);
+ }
+ }
+ }
+ if(score[maxscore]>1)
+ {
+ if(i<loop_start[maxscore]) loop_start[maxscore]=i;
+ for(j=loop_start[maxscore];j<slen&&j<=end[maxscore];j++) {
+ //if(regs[j].regmap[maxscore]>=0) {printf("oops: %x %x was %d=%d\n",loop_start[maxscore]*4+start,j*4+start,maxscore,regs[j].regmap[maxscore]);}
+ assert(regs[j].regmap[maxscore]<0);
+ if(j>loop_start[maxscore]) regs[j].regmap_entry[maxscore]=reg;
+ regs[j].regmap[maxscore]=reg;
+ regs[j].dirty&=~(1<<maxscore);
+ regs[j].wasconst&=~(1<<maxscore);
+ regs[j].isconst&=~(1<<maxscore);
+ if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
+ branch_regs[j].regmap[maxscore]=reg;
+ branch_regs[j].wasdirty&=~(1<<maxscore);
+ branch_regs[j].dirty&=~(1<<maxscore);
+ branch_regs[j].wasconst&=~(1<<maxscore);
+ branch_regs[j].isconst&=~(1<<maxscore);
+ if(itype[j]!=RJUMP&&itype[j]!=UJUMP&&(source[j]>>16)!=0x1000) {
+ regmap_pre[j+2][maxscore]=reg;
+ regs[j+2].wasdirty&=~(1<<maxscore);
+ }
+ // loop optimization (loop_preload)
+ int t=(ba[j]-start)>>2;
+ if(t==loop_start[maxscore]) {
+ if(t==1||(t>1&&itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||(t>1&&rt1[t-2]!=31)) // call/ret assumes no registers allocated
+ regs[t].regmap_entry[maxscore]=reg;
+ }
+ }
+ else
+ {
+ if(j<1||(itype[j-1]!=RJUMP&&itype[j-1]!=UJUMP&&itype[j-1]!=CJUMP&&itype[j-1]!=SJUMP&&itype[j-1]!=FJUMP)) {
+ regmap_pre[j+1][maxscore]=reg;
+ regs[j+1].wasdirty&=~(1<<maxscore);
+ }
+ }
+ }
+ i=j-1;
+ if(itype[j-1]==RJUMP||itype[j-1]==UJUMP||itype[j-1]==CJUMP||itype[j-1]==SJUMP||itype[j-1]==FJUMP) i++; // skip delay slot
+ for(hr=0;hr<HOST_REGS;hr++) {
+ score[hr]=0;earliest_available[hr]=i+i;
+ loop_start[hr]=MAXBLOCK;
+ }
+ }
+ }
+ }
+ }
+ #endif
+
// This allocates registers (if possible) one instruction prior
// to use, which can avoid a load-use penalty on certain CPUs.
for(i=0;i<slen-1;i++)
}
}
}
+ // Preload target address for load instruction (non-constant)
if(itype[i+1]==LOAD&&rs1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
{
}
}
}
+ // Load source into target register
if(lt1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
{
}
}
}
+ // Preload map address
#ifndef HOST_IMM_ADDR32
if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) {
hr=get_reg(regs[i+1].regmap,TLREG);
}
}
#endif
+ // Address for store instruction (non-constant)
if(itype[i+1]==STORE||itype[i+1]==STORELR
||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
}
//requires_32bit[i]=is32[i]&~unneeded_reg_upper[i]; // DEBUG
}
+#else
+ for (i=slen-1;i>=0;i--)
+ {
+ if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
+ {
+ // Conditional branch
+ if((source[i]>>16)!=0x1000&&i<slen-2) {
+ // Mark this address as a branch target since it may be called
+ // upon return from interrupt
+ bt[i+2]=1;
+ }
+ }
+ }
#endif
if(itype[slen-1]==SPAN) {
wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,is32_pre,
unneeded_reg[i],unneeded_reg_upper[i]);
}
- is32_pre=regs[i].is32;
- dirty_pre=regs[i].dirty;
+ if((itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)&&!likely[i]) {
+ is32_pre=branch_regs[i].is32;
+ dirty_pre=branch_regs[i].dirty;
+ }else{
+ is32_pre=regs[i].is32;
+ dirty_pre=regs[i].dirty;
+ }
#endif
// write back
if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
notaz.gp2x.de / pcsx_rearmed.git / blobdiff