u_int ba[MAXBLOCK];
char likely[MAXBLOCK];
char is_ds[MAXBLOCK];
+ char ooo[MAXBLOCK];
uint64_t unneeded_reg[MAXBLOCK];
uint64_t unneeded_reg_upper[MAXBLOCK];
uint64_t branch_unneeded_reg[MAXBLOCK];
signed char regmap[MAXBLOCK][HOST_REGS];
signed char regmap_entry[MAXBLOCK][HOST_REGS];
uint64_t constmap[MAXBLOCK][HOST_REGS];
- uint64_t known_value[HOST_REGS];
- u_int known_reg;
struct regstat regs[MAXBLOCK];
struct regstat branch_regs[MAXBLOCK];
+ signed char minimum_free_regs[MAXBLOCK];
u_int needed_reg[MAXBLOCK];
uint64_t requires_32bit[MAXBLOCK];
u_int wont_dirty[MAXBLOCK];
#define CSREG 35 // Coprocessor status
#define CCREG 36 // Cycle count
#define INVCP 37 // Pointer to invalid_code
-#define TEMPREG 38
-#define FTEMP 38 // FPU/LDL/LDR temporary register
-#define PTEMP 39 // Prefetch temporary register
-#define TLREG 40 // TLB mapping offset
-#define RHASH 41 // Return address hash
-#define RHTBL 42 // Return address hash table address
-#define RTEMP 43 // JR/JALR address register
-#define MAXREG 43
-#define AGEN1 44 // Address generation temporary register
-#define AGEN2 45 // Address generation temporary register
-#define MGEN1 46 // Maptable address generation temporary register
-#define MGEN2 47 // Maptable address generation temporary register
-#define BTREG 48 // Branch target temporary register
+#define MMREG 38 // Pointer to memory_map
+#define ROREG 39 // ram offset (if rdram!=0x80000000)
+#define TEMPREG 40
+#define FTEMP 40 // FPU temporary register
+#define PTEMP 41 // Prefetch temporary register
+#define TLREG 42 // TLB mapping offset
+#define RHASH 43 // Return address hash
+#define RHTBL 44 // Return address hash table address
+#define RTEMP 45 // JR/JALR address register
+#define MAXREG 45
+#define AGEN1 46 // Address generation temporary register
+#define AGEN2 47 // Address generation temporary register
+#define MGEN1 48 // Maptable address generation temporary register
+#define MGEN2 49 // Maptable address generation temporary register
+#define BTREG 50 // Branch target temporary register
/* instruction types */
#define NOP 0 // No operation
if(rs1[i]) alloc_reg(current,i,rs1[i]);
if(rs2[i]) alloc_reg(current,i,rs2[i]);
alloc_reg(current,i,rt1[i]);
- if(rt1[i]==rs2[i]) alloc_reg_temp(current,i,-1);
+ if(rt1[i]==rs2[i]) {
+ alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
+ }
current->is32|=1LL<<rt1[i];
} else { // DSLLV/DSRLV/DSRAV
if(rs1[i]) alloc_reg64(current,i,rs1[i]);
alloc_reg64(current,i,rt1[i]);
current->is32&=~(1LL<<rt1[i]);
if(opcode2[i]==0x16||opcode2[i]==0x17) // DSRLV and DSRAV need a temporary register
+ {
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
+ }
}
clear_const(current,rs1[i]);
clear_const(current,rs2[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;
}
if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
{
alloc_reg64(current,i,rt1[i]);
alloc_all(current,i);
alloc_reg64(current,i,FTEMP);
+ minimum_free_regs[i]=HOST_REGS;
}
else current->is32|=1LL<<rt1[i];
dirty_reg(current,rt1[i]);
{
alloc_reg(current,i,FTEMP);
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
}
}
else
alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
}
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
{
alloc_all(current,i);
alloc_reg64(current,i,FTEMP);
+ minimum_free_regs[i]=HOST_REGS;
}
}
}
}
// We need a temporary register for address generation
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
}
void c1ls_alloc(struct regstat *current,int i)
#endif
// We need a temporary register for address generation
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
}
#ifndef multdiv_alloc
current->is32&=~(1LL<<LOREG);
dirty_reg(current,HIREG);
dirty_reg(current,LOREG);
+ minimum_free_regs[i]=HOST_REGS;
}
}
else
assert(opcode2[i]==0x10);
alloc_all(current,i);
}
+ minimum_free_regs[i]=HOST_REGS;
}
void cop1_alloc(struct regstat *current,int i)
alloc_reg_temp(current,i,-1);
}
}
+ minimum_free_regs[i]=1;
}
void fconv_alloc(struct regstat *current,int i)
{
alloc_reg(current,i,CSREG); // Load status
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
}
void float_alloc(struct regstat *current,int i)
{
alloc_reg(current,i,CSREG); // Load status
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
}
void c2op_alloc(struct regstat *current,int i)
{
alloc_reg(current,i,FSREG); // Load flags
dirty_reg(current,FSREG); // Flag will be modified
alloc_reg_temp(current,i,-1);
+ minimum_free_regs[i]=1;
}
void syscall_alloc(struct regstat *current,int i)
alloc_cc(current,i);
dirty_reg(current,CCREG);
alloc_all(current,i);
+ minimum_free_regs[i]=HOST_REGS;
current->isconst=0;
}
current->isconst=0;
current->wasconst=0;
regs[i].wasconst=0;
+ minimum_free_regs[i]=HOST_REGS;
alloc_all(current,i);
alloc_cc(current,i);
dirty_reg(current,CCREG);
if (opcode[i]==0x28) { // SB
if(!c||memtarget) {
- int x=0;
+ int x=0,a=temp;
#ifdef BIG_ENDIAN_MIPS
if(!c) emit_xorimm(addr,3,temp);
else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
#else
- if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
- else if (addr!=temp) emit_mov(addr,temp);
+ if(!c) a=addr;
#endif
//gen_tlb_addr_w(temp,map);
//emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
- emit_writebyte_indexed_tlb(tl,x,temp,map,temp);
+ emit_writebyte_indexed_tlb(tl,x,a,map,a);
}
type=STOREB_STUB;
}
if (opcode[i]==0x29) { // SH
if(!c||memtarget) {
- int x=0;
+ int x=0,a=temp;
#ifdef BIG_ENDIAN_MIPS
if(!c) emit_xorimm(addr,2,temp);
else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
#else
- if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
- else if (addr!=temp) emit_mov(addr,temp);
+ if(!c) a=addr;
#endif
//#ifdef
//emit_writehword_indexed_tlb(tl,x,temp,map,temp);
//#else
if(map>=0) {
- gen_tlb_addr_w(temp,map);
- emit_writehword_indexed(tl,x,temp);
+ gen_tlb_addr_w(a,map);
+ emit_writehword_indexed(tl,x,a);
}else
- emit_writehword_indexed(tl,(int)rdram-0x80000000+x,temp);
+ emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
}
type=STOREH_STUB;
}
}
type=STORED_STUB;
}
- if(!using_tlb&&(!c||memtarget))
- // addr could be a temp, make sure it survives STORE*_STUB
- reglist|=1<<addr;
- if(jaddr) {
- add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
- } else if(!memtarget) {
- inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
- }
if(!using_tlb) {
if(!c||memtarget) {
#ifdef DESTRUCTIVE_SHIFT
#else
emit_cmpmem_indexedsr12_imm((int)invalid_code,addr,1);
#endif
+ #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
+ emit_callne(invalidate_addr_reg[addr]);
+ #else
jaddr2=(int)out;
emit_jne(0);
add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
+ #endif
}
}
+ if(jaddr) {
+ add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
+ } else if(c&&!memtarget) {
+ inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
+ }
//if(opcode[i]==0x2B || opcode[i]==0x3F)
//if(opcode[i]==0x2B || opcode[i]==0x28)
//if(opcode[i]==0x2B || opcode[i]==0x29)
#else
emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
#endif
+ #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
+ emit_callne(invalidate_addr_reg[temp]);
+ #else
jaddr3=(int)out;
emit_jne(0);
add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
+ #endif
}
}
if(jaddr2) add_stub(type,jaddr2,(int)out,i,offset||c||s<0?ar:s,(int)i_regs,ccadj[i],reglist);
#else
emit_cmpmem_indexedsr12_imm((int)invalid_code,ar,1);
#endif
+ #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
+ emit_callne(invalidate_addr_reg[ar]);
+ #else
jaddr3=(int)out;
emit_jne(0);
add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),ar,0,0,0);
+ #endif
}
if (opcode[i]==0x32) { // LWC2
cop2_put_dreg(copr,tl,HOST_TEMPREG);
int prev_cop1_usable=cop1_usable;
int unconditional=0,nop=0;
int only32=0;
- int ooo=1;
int invert=0;
int internal=internal_branch(branch_regs[i].is32,ba[i]);
if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
- if(likely[i]) ooo=0;
if(!match) invert=1;
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(i>(ba[i]-start)>>2) invert=1;
#endif
-
- if(ooo)
- if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
- (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1])))
- {
- // Write-after-read dependency prevents out of order execution
- // First test branch condition, then execute delay slot, then branch
- ooo=0;
- }
-
- if(ooo) {
+
+ if(ooo[i]) {
s1l=get_reg(branch_regs[i].regmap,rs1[i]);
s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
s2l=get_reg(branch_regs[i].regmap,rs2[i]);
only32=(regs[i].was32>>rs1[i])&(regs[i].was32>>rs2[i])&1;
}
- if(ooo) {
+ if(ooo[i]) {
// Out of order execution (delay slot first)
//printf("OOOE\n");
address_generation(i+1,i_regs,regs[i].regmap_entry);
int prev_cop1_usable=cop1_usable;
int unconditional=0,nevertaken=0;
int only32=0;
- int ooo=1;
int invert=0;
int internal=internal_branch(branch_regs[i].is32,ba[i]);
if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
- if(likely[i]) ooo=0;
if(!match) invert=1;
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(i>(ba[i]-start)>>2) invert=1;
//if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
//assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
- if(ooo) {
- if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))
- {
- // Write-after-read dependency prevents out of order execution
- // First test branch condition, then execute delay slot, then branch
- ooo=0;
- }
- if(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))
- // BxxZAL $ra is available to delay insn, so do it in order
- ooo=0;
- }
-
- if(ooo) {
+ if(ooo[i]) {
s1l=get_reg(branch_regs[i].regmap,rs1[i]);
s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
}
only32=(regs[i].was32>>rs1[i])&1;
}
- if(ooo) {
+ if(ooo[i]) {
// Out of order execution (delay slot first)
//printf("OOOE\n");
address_generation(i+1,i_regs,regs[i].regmap_entry);
assem_debug("fmatch=%d\n",match);
int fs,cs;
int eaddr;
- int ooo=1;
int invert=0;
int internal=internal_branch(branch_regs[i].is32,ba[i]);
if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
- if(likely[i]) ooo=0;
if(!match) invert=1;
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(i>(ba[i]-start)>>2) invert=1;
#endif
- if(ooo)
- if(itype[i+1]==FCOMP)
- {
- // Write-after-read dependency prevents out of order execution
- // First test branch condition, then execute delay slot, then branch
- ooo=0;
- }
-
- if(ooo) {
+ if(ooo[i]) {
fs=get_reg(branch_regs[i].regmap,FSREG);
address_generation(i+1,i_regs,regs[i].regmap_entry); // Is this okay?
}
cop1_usable=1;
}
- if(ooo) {
+ if(ooo[i]) {
// Out of order execution (delay slot first)
//printf("OOOE\n");
ds_assemble(i+1,i_regs);
}
}
-void new_dynarec_init()
+// clear the state completely, instead of just marking
+// things invalid like invalidate_all_pages() does
+void new_dynarec_clear_full()
{
- printf("Init new dynarec\n");
- out=(u_char *)BASE_ADDR;
- 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");}
-#ifdef MUPEN64
- rdword=&readmem_dword;
- fake_pc.f.r.rs=&readmem_dword;
- fake_pc.f.r.rt=&readmem_dword;
- fake_pc.f.r.rd=&readmem_dword;
-#endif
int n;
for(n=0x80000;n<0x80800;n++)
invalid_code[n]=1;
hash_table[n][0]=hash_table[n][2]=-1;
memset(mini_ht,-1,sizeof(mini_ht));
memset(restore_candidate,0,sizeof(restore_candidate));
+ memset(shadow,0,sizeof(shadow));
copy=shadow;
expirep=16384; // Expiry pointer, +2 blocks
pending_exception=0;
literalcount=0;
-#ifdef HOST_IMM8
- // Copy this into local area so we don't have to put it in every literal pool
- invc_ptr=invalid_code;
-#endif
stop_after_jal=0;
// TLB
using_tlb=0;
memory_map[n]=((u_int)rdram-0x80000000)>>2;
for(n=526336;n<1048576;n++) // 0x80800000 .. 0xFFFFFFFF
memory_map[n]=-1;
+ 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);
+}
+
+void new_dynarec_init()
+{
+ printf("Init new dynarec\n");
+ out=(u_char *)BASE_ADDR;
+ 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");}
+#ifdef MUPEN64
+ rdword=&readmem_dword;
+ fake_pc.f.r.rs=&readmem_dword;
+ fake_pc.f.r.rt=&readmem_dword;
+ fake_pc.f.r.rd=&readmem_dword;
+#endif
+ int n;
+ new_dynarec_clear_full();
+#ifdef HOST_IMM8
+ // Copy this into local area so we don't have to put it in every literal pool
+ invc_ptr=invalid_code;
+#endif
#ifdef MUPEN64
for(n=0;n<0x8000;n++) { // 0 .. 0x7FFFFFFF
writemem[n] = write_nomem_new;
/* Pass 1 disassembly */
for(i=0;!done;i++) {
- bt[i]=0;likely[i]=0;op2=0;
+ bt[i]=0;likely[i]=0;ooo[i]=0;op2=0;
+ minimum_free_regs[i]=0;
opcode[i]=op=source[i]>>26;
switch(op)
{
#endif
//current.is32|=1LL<<rt1[i];
}
+ ooo[i]=1;
delayslot_alloc(¤t,i+1);
//current.isconst=0; // DEBUG
ds=1;
alloc_reg(¤t,i,RTEMP);
}
//current.isconst=0; // DEBUG
+ ooo[i]=1;
ds=1;
break;
case CJUMP:
(rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1]))) {
// The delay slot overwrites one of our conditions.
// Allocate the branch condition registers instead.
- // Note that such a sequence of instructions could
- // be considered a bug since the branch can not be
- // re-executed if an exception occurs.
current.isconst=0;
current.wasconst=0;
regs[i].wasconst=0;
if(rs2[i]) alloc_reg64(¤t,i,rs2[i]);
}
}
- else delayslot_alloc(¤t,i+1);
+ else
+ {
+ ooo[i]=1;
+ delayslot_alloc(¤t,i+1);
+ }
}
else
if((opcode[i]&0x3E)==6) // BLEZ/BGTZ
if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
// The delay slot overwrites one of our conditions.
// Allocate the branch condition registers instead.
- // Note that such a sequence of instructions could
- // be considered a bug since the branch can not be
- // re-executed if an exception occurs.
current.isconst=0;
current.wasconst=0;
regs[i].wasconst=0;
if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
}
}
- else delayslot_alloc(¤t,i+1);
+ else
+ {
+ ooo[i]=1;
+ delayslot_alloc(¤t,i+1);
+ }
}
else
// Don't alloc the delay slot yet because we might not execute it
//#endif
//current.is32|=1LL<<rt1[i];
}
- if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
- // The delay slot overwrites the branch condition.
+ if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) // The delay slot overwrites the branch condition.
+ ||(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))) { // DS touches $ra
// Allocate the branch condition registers instead.
- // Note that such a sequence of instructions could
- // be considered a bug since the branch can not be
- // re-executed if an exception occurs.
current.isconst=0;
current.wasconst=0;
regs[i].wasconst=0;
if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
}
}
- else delayslot_alloc(¤t,i+1);
+ else
+ {
+ ooo[i]=1;
+ delayslot_alloc(¤t,i+1);
+ }
}
else
// Don't alloc the delay slot yet because we might not execute it
if(itype[i+1]==FCOMP) {
// The delay slot overwrites the branch condition.
// Allocate the branch condition registers instead.
- // Note that such a sequence of instructions could
- // be considered a bug since the branch can not be
- // re-executed if an exception occurs.
alloc_cc(¤t,i);
dirty_reg(¤t,CCREG);
alloc_reg(¤t,i,CSREG);
alloc_reg(¤t,i,FSREG);
}
else {
+ ooo[i]=1;
delayslot_alloc(¤t,i+1);
alloc_reg(¤t,i+1,CSREG);
}
{
cc=0;
}
+#ifdef PCSX
+ else if(/*itype[i]==LOAD||*/itype[i]==STORE||itype[i]==C1LS) // load causes weird timing issues
+ {
+ cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
+ }
+ else if(itype[i]==C2LS)
+ {
+ cc+=4;
+ }
+#endif
else
{
cc++;
f_regmap[hr]=branch_regs[i].regmap[hr];
}
}
- if(itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
- ||itype[i+1]==SHIFT||itype[i+1]==COP1||itype[i+1]==FLOAT
- ||itype[i+1]==FCOMP||itype[i+1]==FCONV
- ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
- {
- // Test both in case the delay slot is ooo,
- // could be done better...
- if(count_free_regs(branch_regs[i].regmap)<2
- ||count_free_regs(regs[i].regmap)<2)
+ if(ooo[i]) {
+ if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
+ f_regmap[hr]=branch_regs[i].regmap[hr];
+ }else{
+ if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
f_regmap[hr]=branch_regs[i].regmap[hr];
}
// Avoid dirty->clean transition
- // #ifdef DESTRUCTIVE_WRITEBACK here?
+ #ifdef DESTRUCTIVE_WRITEBACK
if(t>0) if(get_reg(regmap_pre[t],f_regmap[hr])>=0) if((regs[t].wasdirty>>get_reg(regmap_pre[t],f_regmap[hr]))&1) f_regmap[hr]=-1;
+ #endif
+ // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
+ // case above, however it's always a good idea. We can't hoist the
+ // load if the register was already allocated, so there's no point
+ // wasting time analyzing most of these cases. It only "succeeds"
+ // when the mapping was different and the load can be replaced with
+ // 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) {
+ if(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++)
{
// register is lower numbered than the lower-half
// register. Not sure if it's worth fixing...
if(get_reg(regs[j].regmap,r&63)<0) break;
+ if(get_reg(regs[j].regmap_entry,r&63)<0) break;
if(regs[j].is32&(1LL<<(r&63))) break;
}
if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
}
k=i;
while(k>1&®s[k-1].regmap[hr]==-1) {
- if(itype[k-1]==STORE||itype[k-1]==STORELR
- ||itype[k-1]==C1LS||itype[k-1]==SHIFT||itype[k-1]==COP1
- ||itype[k-1]==FLOAT||itype[k-1]==FCONV||itype[k-1]==FCOMP
- ||itype[k-1]==COP2||itype[k-1]==C2LS||itype[k-1]==C2OP) {
- if(count_free_regs(regs[k-1].regmap)<2) {
- //printf("no free regs for store %x\n",start+(k-1)*4);
- break;
- }
+ if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
+ //printf("no free regs for store %x\n",start+(k-1)*4);
+ break;
}
- else
- if(itype[k-1]!=NOP&&itype[k-1]!=MOV&&itype[k-1]!=ALU&&itype[k-1]!=SHIFTIMM&&itype[k-1]!=IMM16&&itype[k-1]!=LOAD) break;
if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
//printf("no-match due to different register\n");
break;
}
}
for(k=t;k<j;k++) {
+ // Alloc register clean at beginning of loop,
+ // but may dirty it in pass 6
regs[k].regmap_entry[hr]=f_regmap[hr];
regs[k].regmap[hr]=f_regmap[hr];
- regmap_pre[k+1][hr]=f_regmap[hr];
- regs[k+1].wasdirty&=~(1<<hr);
regs[k].dirty&=~(1<<hr);
regs[k].wasconst&=~(1<<hr);
regs[k].isconst&=~(1<<hr);
+ if(itype[k]==UJUMP||itype[k]==RJUMP||itype[k]==CJUMP||itype[k]==SJUMP||itype[k]==FJUMP) {
+ branch_regs[k].regmap_entry[hr]=f_regmap[hr];
+ branch_regs[k].regmap[hr]=f_regmap[hr];
+ branch_regs[k].dirty&=~(1<<hr);
+ branch_regs[k].wasconst&=~(1<<hr);
+ branch_regs[k].isconst&=~(1<<hr);
+ if(itype[k]!=RJUMP&&itype[k]!=UJUMP&&(source[k]>>16)!=0x1000) {
+ regmap_pre[k+2][hr]=f_regmap[hr];
+ regs[k+2].wasdirty&=~(1<<hr);
+ assert((branch_regs[k].is32&(1LL<<f_regmap[hr]))==
+ (regs[k+2].was32&(1LL<<f_regmap[hr])));
+ }
+ }
+ else
+ {
+ regmap_pre[k+1][hr]=f_regmap[hr];
+ regs[k+1].wasdirty&=~(1<<hr);
+ }
}
if(regs[j].regmap[hr]==f_regmap[hr])
regs[j].regmap_entry[hr]=f_regmap[hr];
//printf("32/64 mismatch %x %d\n",start+j*4,hr);
break;
}
- if(itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS
- ||itype[j]==SHIFT||itype[j]==COP1||itype[j]==FLOAT
- ||itype[j]==FCOMP||itype[j]==FCONV
- ||itype[j]==COP2||itype[j]==C2LS||itype[j]==C2OP) {
- if(count_free_regs(regs[j].regmap)<2) {
- //printf("No free regs for store %x\n",start+j*4);
+ if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
+ {
+ // Stop on unconditional branch
+ break;
+ }
+ if(itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP)
+ {
+ 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;
+ }
+ if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
+ //printf("no-match due to different register (branch)\n");
break;
}
}
- else if(itype[j]!=NOP&&itype[j]!=MOV&&itype[j]!=ALU&&itype[j]!=SHIFTIMM&&itype[j]!=IMM16&&itype[j]!=LOAD) break;
+ if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
+ //printf("No free regs for store %x\n",start+j*4);
+ break;
+ }
if(f_regmap[hr]>=64) {
if(regs[j].is32&(1LL<<(f_regmap[hr]&63))) {
break;
if(bt[i]) {
for(j=i;j<slen-1;j++) {
if(regs[j].regmap[HOST_CCREG]!=-1) break;
- if(itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS
- ||itype[j]==SHIFT||itype[j]==COP1||itype[j]==FLOAT
- ||itype[j]==FCOMP||itype[j]==FCONV
- ||itype[j]==COP2||itype[j]==C2LS||itype[j]==C2OP) {
- if(count_free_regs(regs[j].regmap)<2) {
- //printf("no free regs for store %x\n",start+j*4);
- break;
- }
+ if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
+ //printf("no free regs for store %x\n",start+j*4);
+ break;
}
- else
- if(itype[j]!=NOP&&itype[j]!=MOV&&itype[j]!=ALU&&itype[j]!=SHIFTIMM&&itype[j]!=IMM16&&itype[j]!=LOAD) break;
}
if(regs[j].regmap[HOST_CCREG]==CCREG) {
int k=i;
int k;
k=i;
while(regs[k-1].regmap[HOST_CCREG]==-1) {
- if(itype[k-1]==STORE||itype[k-1]==STORELR||itype[k-1]==C1LS
- ||itype[k-1]==SHIFT||itype[k-1]==COP1||itype[k-1]==FLOAT
- ||itype[k-1]==FCONV||itype[k-1]==FCOMP
- ||itype[k-1]==COP2||itype[k-1]==C2LS||itype[k-1]==C2OP) {
- if(count_free_regs(regs[k-1].regmap)<2) {
- //printf("no free regs for store %x\n",start+(k-1)*4);
- break;
- }
+ if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
+ //printf("no free regs for store %x\n",start+(k-1)*4);
+ break;
}
- else
- if(itype[k-1]!=NOP&&itype[k-1]!=MOV&&itype[k-1]!=ALU&&itype[k-1]!=SHIFTIMM&&itype[k-1]!=IMM16&&itype[k-1]!=LOAD) break;
k--;
}
if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
if(itype[i]!=STORE&&itype[i]!=STORELR&&itype[i]!=C1LS&&itype[i]!=SHIFT&&
itype[i]!=NOP&&itype[i]!=MOV&&itype[i]!=ALU&&itype[i]!=SHIFTIMM&&
itype[i]!=IMM16&&itype[i]!=LOAD&&itype[i]!=COP1&&itype[i]!=FLOAT&&
- itype[i]!=FCONV&&itype[i]!=FCOMP&&
- itype[i]!=COP2&&itype[i]!=C2LS&&itype[i]!=C2OP)
+ itype[i]!=FCONV&&itype[i]!=FCOMP)
{
memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
}
notaz.gp2x.de / pcsx_rearmed.git / blobdiff