#include <stdlib.h>
#include <stdint.h> //include for uint64_t
#include <assert.h>
+#include <sys/mman.h>
#include "emu_if.h" //emulator interface
-#include <sys/mman.h>
+//#define DISASM
+//#define assem_debug printf
+//#define inv_debug printf
+#define assem_debug(...)
+#define inv_debug(...)
#ifdef __i386__
#include "assem_x86.h"
#define MAXBLOCK 4096
#define MAX_OUTPUT_BLOCK_SIZE 262144
-#define CLOCK_DIVIDER 2
+
+int cycle_multiplier; // 100 for 1.0
+#define CLOCK_ADJUST(x) (((x) * cycle_multiplier + 50) / 100)
struct regstat
{
uint64_t uu;
u_int wasconst;
u_int isconst;
+ u_int loadedconst; // host regs that have constants loaded
+ u_int waswritten; // MIPS regs that were used as store base before
uint64_t constmap[HOST_REGS];
};
u_char dep1[MAXBLOCK];
u_char dep2[MAXBLOCK];
u_char lt1[MAXBLOCK];
+ static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
+ static uint64_t gte_rt[MAXBLOCK];
+ static uint64_t gte_unneeded[MAXBLOCK];
+ static u_int smrv[32]; // speculated MIPS register values
+ static u_int smrv_strong; // mask or regs that are likely to have correct values
+ static u_int smrv_weak; // same, but somewhat less likely
+ static u_int smrv_strong_next; // same, but after current insn executes
+ static u_int smrv_weak_next;
int imm[MAXBLOCK];
u_int ba[MAXBLOCK];
char likely[MAXBLOCK];
#else
static const u_int using_tlb=0;
#endif
- static u_int sp_in_mirror;
+ int new_dynarec_did_compile;
+ int new_dynarec_hacks;
u_int stop_after_jal;
extern u_char restore_candidate[512];
extern int cycle_count;
//#define DEBUG_CYCLE_COUNT 1
-void nullf() {}
-//#define assem_debug printf
-//#define inv_debug printf
-#define assem_debug nullf
-#define inv_debug nullf
-
static void tlb_hacks()
{
#ifndef DISABLE_TLB
//printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
invalid_code[vaddr>>12]=0;
inv_code_start=inv_code_end=~0;
+#ifndef DISABLE_TLB
memory_map[vaddr>>12]|=0x40000000;
+#endif
if(vpage<2048) {
#ifndef DISABLE_TLB
if(tlb_LUT_r[vaddr>>12]) {
}
}
-
+#ifndef FORCE32
void div64(int64_t dividend,int64_t divisor)
{
lo=dividend/divisor;
else original=loaded;
return original;
}
+#endif
#ifdef __i386__
#include "assem_x86.c"
inv|=invalid_code[i];
}
}
+#ifndef DISABLE_TLB
if((signed int)head->vaddr>=(signed int)0xC0000000) {
u_int addr = (head->vaddr+(memory_map[head->vaddr>>12]<<2));
//printf("addr=%x start=%x end=%x\n",addr,start,end);
if(addr<start||addr>=end) inv=1;
}
+#endif
else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
inv=1;
}
void shiftimm_alloc(struct regstat *current,int i)
{
- clear_const(current,rs1[i]);
- clear_const(current,rt1[i]);
if(opcode2[i]<=0x3) // SLL/SRL/SRA
{
if(rt1[i]) {
alloc_reg(current,i,rt1[i]);
current->is32|=1LL<<rt1[i];
dirty_reg(current,rt1[i]);
+ if(is_const(current,rs1[i])) {
+ int v=get_const(current,rs1[i]);
+ if(opcode2[i]==0x00) set_const(current,rt1[i],v<<imm[i]);
+ if(opcode2[i]==0x02) set_const(current,rt1[i],(u_int)v>>imm[i]);
+ if(opcode2[i]==0x03) set_const(current,rt1[i],v>>imm[i]);
+ }
+ else clear_const(current,rt1[i]);
}
}
+ else
+ {
+ clear_const(current,rs1[i]);
+ clear_const(current,rt1[i]);
+ }
+
if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
{
if(rt1[i]) {
t=get_reg(i_regs->regmap,rt1[i]);
s=get_reg(i_regs->regmap,rs1[i]);
//assert(t>=0);
- if(t>=0){
+ if(t>=0&&!((i_regs->isconst>>t)&1)){
if(rs1[i]==0)
{
emit_zeroreg(t);
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);
- fastload_reg_override=HOST_TEMPREG;
- }
- else
- #endif
- emit_cmpimm(addr,RAM_SIZE);
- jaddr=(int)out;
- #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
- // Hint to branch predictor that the branch is unlikely to be taken
- if(rs1[i]>=28)
- emit_jno_unlikely(0);
- else
- #endif
- emit_jno(0);
+ jaddr=emit_fastpath_cmp_jump(i,addr,&fastload_reg_override);
}
}
}else{ // using tlb
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);
- faststore_reg_override=HOST_TEMPREG;
- }
- else
- #endif
+ #ifndef PCSX
#ifdef R29_HACK
// Strmnnrmn's speed hack
if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
#endif
emit_jno(0);
}
+ #else
+ jaddr=emit_fastpath_cmp_jump(i,addr,&faststore_reg_override);
+ #endif
}
}else{ // using tlb
int x=0;
}
type=STORED_STUB;
}
- if(!using_tlb) {
+#ifdef PCSX
+ if(jaddr) {
+ // PCSX store handlers don't check invcode again
+ reglist|=1<<addr;
+ add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
+ jaddr=0;
+ }
+#endif
+ if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
if(!c||memtarget) {
#ifdef DESTRUCTIVE_SHIFT
// The x86 shift operation is 'destructive'; it overwrites the
}
if(!c||!memtarget)
add_stub(STORELR_STUB,jaddr,(int)out,i,(int)i_regs,temp,ccadj[i],reglist);
- if(!using_tlb) {
+ if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
#ifdef RAM_OFFSET
int map=get_reg(i_regs->regmap,ROREG);
if(map<0) map=HOST_TEMPREG;
emit_writedword_indexed_tlb(th,tl,0,offset||c||s<0?temp:s,map,temp);
type=STORED_STUB;
}
- if(!using_tlb) {
+ if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
#ifndef DESTRUCTIVE_SHIFT
temp=offset||c||s<0?ar:s;
int memtarget=0,c=0;
int jaddr2=0,jaddr3,type;
int agr=AGEN1+(i&1);
+ int fastio_reg_override=0;
u_int hr,reglist=0;
u_int copr=(source[i]>>16)&0x1f;
s=get_reg(i_regs->regmap,rs1[i]);
}
else {
if(!c) {
- emit_cmpimm(offset||c||s<0?ar:s,RAM_SIZE);
- jaddr2=(int)out;
- emit_jno(0);
+ jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
}
if (opcode[i]==0x32) { // LWC2
#ifdef HOST_IMM_ADDR32
if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl);
else
#endif
- emit_readword_indexed(0,ar,tl);
+ int a=ar;
+ if(fastio_reg_override) a=fastio_reg_override;
+ emit_readword_indexed(0,a,tl);
}
if (opcode[i]==0x3a) { // SWC2
#ifdef DESTRUCTIVE_SHIFT
if(!offset&&!c&&s>=0) emit_mov(s,ar);
#endif
- emit_writeword_indexed(tl,0,ar);
+ int a=ar;
+ if(fastio_reg_override) a=fastio_reg_override;
+ emit_writeword_indexed(tl,0,a);
}
}
if(jaddr2)
add_stub(type,jaddr2,(int)out,i,ar,(int)i_regs,ccadj[i],reglist);
- if (opcode[i]==0x3a) { // SWC2
+ if(opcode[i]==0x3a) // SWC2
+ if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
#if defined(HOST_IMM8)
int ir=get_reg(i_regs->regmap,INVCP);
assert(ir>=0);
assert(ccreg==HOST_CCREG);
assert(!is_delayslot);
emit_movimm(start+i*4,EAX); // Get PC
- emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
emit_jmp((int)jump_syscall_hle); // XXX
}
assert(!is_delayslot);
emit_movimm(start+i*4+4,0); // Get PC
emit_movimm((int)psxHLEt[source[i]&7],1);
- emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // XXX
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
emit_jmp((int)jump_hlecall);
}
assert(ccreg==HOST_CCREG);
assert(!is_delayslot);
emit_movimm(start+i*4,0); // Get PC
- emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG);
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
emit_jmp((int)jump_intcall);
}
void ds_assemble(int i,struct regstat *i_regs)
{
+ speculate_register_values(i);
is_delayslot=1;
switch(itype[i]) {
case ALU:
// printf("poor load scheduling!\n");
}
else if(c) {
+#ifndef DISABLE_TLB
if(rm>=0) {
if(!entry||entry[rm]!=mgr) {
if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a) {
}
}
}
+#endif
if(rs1[i]!=rt1[i]||itype[i]!=LOAD) {
if(!entry||entry[ra]!=agr) {
if (opcode[i]==0x22||opcode[i]==0x26) {
(using_tlb&&((signed int)constmap[i][rs]+offset)>=(signed int)0xC0000000))
#endif
emit_movimm(constmap[i][rs]+offset,ra);
+ regs[i].loadedconst|=1<<ra;
}
} // else did it in the previous cycle
} // else load_consts already did it
// Preload constants for next instruction
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) {
int agr,ra;
- #ifndef HOST_IMM_ADDR32
+ #if !defined(HOST_IMM_ADDR32) && !defined(DISABLE_TLB)
// Mapper entry
agr=MGEN1+((i+1)&1);
ra=get_reg(i_regs->regmap,agr);
(using_tlb&&((signed int)constmap[i+1][rs]+offset)>=(signed int)0xC0000000))
#endif
emit_movimm(constmap[i+1][rs]+offset,ra);
+ regs[i+1].loadedconst|=1<<ra;
}
}
else if(rs1[i+1]==0) {
// Load registers with known constants
void load_consts(signed char pre[],signed char regmap[],int is32,int i)
{
- int hr;
+ int hr,hr2;
+ // propagate loaded constant flags
+ if(i==0||bt[i])
+ regs[i].loadedconst=0;
+ else {
+ for(hr=0;hr<HOST_REGS;hr++) {
+ if(hr!=EXCLUDE_REG&®map[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
+ &®map[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
+ {
+ regs[i].loadedconst|=1<<hr;
+ }
+ }
+ }
// Load 32-bit regs
for(hr=0;hr<HOST_REGS;hr++) {
if(hr!=EXCLUDE_REG&®map[hr]>=0) {
//if(entry[hr]!=regmap[hr]) {
- if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
+ if(!((regs[i].loadedconst>>hr)&1)) {
if(((regs[i].isconst>>hr)&1)&®map[hr]<64&®map[hr]>0) {
- int value;
+ int value,similar=0;
if(get_final_value(hr,i,&value)) {
- if(value==0) {
+ // see if some other register has similar value
+ for(hr2=0;hr2<HOST_REGS;hr2++) {
+ if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
+ if(is_similar_value(value,constmap[i][hr2])) {
+ similar=1;
+ break;
+ }
+ }
+ }
+ if(similar) {
+ int value2;
+ if(get_final_value(hr2,i,&value2)) // is this needed?
+ emit_movimm_from(value2,hr2,value,hr);
+ else
+ emit_movimm(value,hr);
+ }
+ else if(value==0) {
emit_zeroreg(hr);
}
else {
emit_movimm(value,hr);
}
}
+ regs[i].loadedconst|=1<<hr;
}
}
}
void load_regs_entry(int t)
{
int hr;
- if(is_ds[t]) emit_addimm(HOST_CCREG,CLOCK_DIVIDER,HOST_CCREG);
- else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t]*CLOCK_DIVIDER,HOST_CCREG);
+ if(is_ds[t]) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
+ else if(ccadj[t]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[t]),HOST_CCREG);
if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
emit_storereg(CCREG,HOST_CCREG);
}
emit_jmp(0);
}
else if(*adj==0||invert) {
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(count+2),HOST_CCREG);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(count+2),HOST_CCREG);
jaddr=(int)out;
emit_jns(0);
}
else
{
- emit_cmpimm(HOST_CCREG,-CLOCK_DIVIDER*(count+2));
+ emit_cmpimm(HOST_CCREG,-CLOCK_ADJUST(count+2));
jaddr=(int)out;
emit_jns(0);
}
}
// Update cycle count
assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
- if(stubs[n][3]) emit_addimm(HOST_CCREG,CLOCK_DIVIDER*stubs[n][3],HOST_CCREG);
+ if(stubs[n][3]) emit_addimm(HOST_CCREG,CLOCK_ADJUST((int)stubs[n][3]),HOST_CCREG);
emit_call((int)cc_interrupt);
- if(stubs[n][3]) emit_addimm(HOST_CCREG,-CLOCK_DIVIDER*stubs[n][3],HOST_CCREG);
+ if(stubs[n][3]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((int)stubs[n][3]),HOST_CCREG);
if(stubs[n][6]==TAKEN) {
if(internal_branch(branch_regs[i].is32,ba[i]))
load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
#endif
do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
- if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal_branch(branch_regs[i].is32,ba[i]))
assem_debug("branch: internal\n");
//do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
//if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
//assert(adj==0);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
add_stub(CC_STUB,(int)out,jump_vaddr_reg[rs],0,i,-1,TAKEN,0);
+#ifdef PCSX
+ if(itype[i+1]==COP0&&(source[i+1]&0x3f)==0x10)
+ // special case for RFE
+ emit_jmp(0);
+ else
+#endif
emit_jns(0);
//load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
#ifdef USE_MINI_HT
if(unconditional) {
do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
- if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
assem_debug("branch: internal\n");
}
}
else if(nop) {
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
else {
int taken=0,nottaken=0,nottaken1=0;
do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
- if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
if(!only32)
{
assert(s1h>=0);
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
if(adj) {
- emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
+ emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
add_to_linker((int)out,ba[i],internal);
}else{
emit_addnop(13);
}else
#endif
{
- if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
+ if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
if(nottaken1) set_jump_target(nottaken1,(int)out);
if(adj) {
- if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
+ if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
}
} // (!unconditional)
} // if(ooo)
store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
assem_debug("cycle count (adj)\n");
- if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
assem_debug("branch: internal\n");
if(cc==-1&&!likely[i]) {
// Cycle count isn't in a register, temporarily load it then write it out
emit_loadreg(CCREG,HOST_CCREG);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
else{
cc=get_reg(i_regmap,CCREG);
assert(cc==HOST_CCREG);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
if(unconditional) {
do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
- if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
assem_debug("branch: internal\n");
}
}
else if(nevertaken) {
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
else {
int nottaken=0;
do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
- if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
if(!only32)
{
assert(s1h>=0);
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
if(adj) {
- emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
+ emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
add_to_linker((int)out,ba[i],internal);
}else{
emit_addnop(13);
}else
#endif
{
- if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
+ if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
}
if(adj) {
- if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
+ if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
}
} // (!unconditional)
} // if(ooo)
store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
assem_debug("cycle count (adj)\n");
- if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
assem_debug("branch: internal\n");
if(cc==-1&&!likely[i]) {
// Cycle count isn't in a register, temporarily load it then write it out
emit_loadreg(CCREG,HOST_CCREG);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
else{
cc=get_reg(i_regmap,CCREG);
assert(cc==HOST_CCREG);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
assem_debug("cycle count (adj)\n");
if(1) {
int nottaken=0;
- if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
if(1) {
assert(fs>=0);
emit_testimm(fs,0x800000);
} // if(!only32)
if(invert) {
- if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
+ if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
else if(match) emit_addnop(13);
#endif
}
if(adj) {
- if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
+ if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
}
} // (!unconditional)
} // if(ooo)
store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
assem_debug("cycle count (adj)\n");
- if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
+ if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
if(internal)
assem_debug("branch: internal\n");
if(cc==-1&&!likely[i]) {
// Cycle count isn't in a register, temporarily load it then write it out
emit_loadreg(CCREG,HOST_CCREG);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
else{
cc=get_reg(i_regmap,CCREG);
assert(cc==HOST_CCREG);
- emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
+ emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
int jaddr=(int)out;
emit_jns(0);
add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
if((opcode[i]&0x2e)==4||opcode[i]==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
}
- emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
if(opcode[i]==2) // J
{
unconditional=1;
void unneeded_registers(int istart,int iend,int r)
{
int i;
- uint64_t u,uu,b,bu;
- uint64_t temp_u,temp_uu;
+ uint64_t u,uu,gte_u,b,bu,gte_bu;
+ uint64_t temp_u,temp_uu,temp_gte_u=0;
uint64_t tdep;
+ uint64_t gte_u_unknown=0;
+ if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
+ gte_u_unknown=~0ll;
if(iend==slen-1) {
u=1;uu=1;
+ gte_u=gte_u_unknown;
}else{
u=unneeded_reg[iend+1];
uu=unneeded_reg_upper[iend+1];
u=1;uu=1;
+ gte_u=gte_unneeded[iend+1];
}
+
for (i=iend;i>=istart;i--)
{
//printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
// Branch out of this block, flush all regs
u=1;
uu=1;
+ gte_u=gte_u_unknown;
/* Hexagon hack
if(itype[i]==UJUMP&&rt1[i]==31)
{
uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
u|=1;uu|=1;
+ gte_u|=gte_rt[i+1];
+ gte_u&=~gte_rs[i+1];
// If branch is "likely" (and conditional)
// then we skip the delay slot on the fall-thru path
if(likely[i]) {
if(i<slen-1) {
u&=unneeded_reg[i+2];
uu&=unneeded_reg_upper[i+2];
+ gte_u&=gte_unneeded[i+2];
}
else
{
u=1;
uu=1;
+ gte_u=gte_u_unknown;
}
}
}
{
// Unconditional branch
temp_u=1;temp_uu=1;
+ temp_gte_u=0;
} else {
// Conditional branch (not taken case)
temp_u=unneeded_reg[i+2];
temp_uu=unneeded_reg_upper[i+2];
+ temp_gte_u&=gte_unneeded[i+2];
}
// Merge in delay slot
tdep=(~temp_uu>>rt1[i+1])&1;
temp_uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
temp_uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
temp_u|=1;temp_uu|=1;
+ temp_gte_u|=gte_rt[i+1];
+ temp_gte_u&=~gte_rs[i+1];
// If branch is "likely" (and conditional)
// then we skip the delay slot on the fall-thru path
if(likely[i]) {
if(i<slen-1) {
temp_u&=unneeded_reg[i+2];
temp_uu&=unneeded_reg_upper[i+2];
+ temp_gte_u&=gte_unneeded[i+2];
}
else
{
temp_u=1;
temp_uu=1;
+ temp_gte_u=gte_u_unknown;
}
}
tdep=(~temp_uu>>rt1[i])&1;
temp_uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
temp_uu&=~((tdep<<dep1[i])|(tdep<<dep2[i]));
temp_u|=1;temp_uu|=1;
+ temp_gte_u|=gte_rt[i];
+ temp_gte_u&=~gte_rs[i];
unneeded_reg[i]=temp_u;
unneeded_reg_upper[i]=temp_uu;
+ gte_unneeded[i]=temp_gte_u;
// Only go three levels deep. This recursion can take an
// excessive amount of time if there are a lot of nested loops.
if(r<2) {
}else{
unneeded_reg[(ba[i]-start)>>2]=1;
unneeded_reg_upper[(ba[i]-start)>>2]=1;
+ gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
}
} /*else*/ if(1) {
if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
// Unconditional branch
u=unneeded_reg[(ba[i]-start)>>2];
uu=unneeded_reg_upper[(ba[i]-start)>>2];
+ gte_u=gte_unneeded[(ba[i]-start)>>2];
branch_unneeded_reg[i]=u;
branch_unneeded_reg_upper[i]=uu;
//u=1;
uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
u|=1;uu|=1;
+ gte_u|=gte_rt[i+1];
+ gte_u&=~gte_rs[i+1];
} else {
// Conditional branch
b=unneeded_reg[(ba[i]-start)>>2];
bu=unneeded_reg_upper[(ba[i]-start)>>2];
+ gte_bu=gte_unneeded[(ba[i]-start)>>2];
branch_unneeded_reg[i]=b;
branch_unneeded_reg_upper[i]=bu;
//b=1;
bu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
bu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
b|=1;bu|=1;
+ gte_bu|=gte_rt[i+1];
+ gte_bu&=~gte_rs[i+1];
// If branch is "likely" then we skip the
// delay slot on the fall-thru path
if(likely[i]) {
u=b;
uu=bu;
+ gte_u=gte_bu;
if(i<slen-1) {
u&=unneeded_reg[i+2];
uu&=unneeded_reg_upper[i+2];
+ gte_u&=gte_unneeded[i+2];
//u=1;
//uu=1;
}
} else {
u&=b;
uu&=bu;
+ gte_u&=gte_bu;
//u=1;
//uu=1;
}
u|=1LL<<rt2[i];
uu|=1LL<<rt1[i];
uu|=1LL<<rt2[i];
+ gte_u|=gte_rt[i];
// Accessed registers are needed
u&=~(1LL<<rs1[i]);
u&=~(1LL<<rs2[i]);
uu&=~(1LL<<us1[i]);
uu&=~(1LL<<us2[i]);
+ gte_u&=~gte_rs[i];
// Source-target dependencies
uu&=~(tdep<<dep1[i]);
uu&=~(tdep<<dep2[i]);
// Save it
unneeded_reg[i]=u;
unneeded_reg_upper[i]=uu;
+ gte_unneeded[i]=gte_u;
/*
printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
printf("U:");
}
}
+#ifdef DISASM
/* disassembly */
void disassemble_inst(int i)
{
printf (" %x: %s\n",start+i*4,insn[i]);
}
}
+#else
+static void disassemble_inst(int i) {}
+#endif // DISASM
// clear the state completely, instead of just marking
// things invalid like invalidate_all_pages() does
// TLB
#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
memory_map[n]=((u_int)rdram-0x80000000)>>2;
for(n=526336;n<1048576;n++) // 0x80800000 .. 0xFFFFFFFF
memory_map[n]=-1;
+#endif
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);
fake_pc.f.r.rd=&readmem_dword;
#endif
int n;
+ cycle_multiplier=200;
new_dynarec_clear_full();
#ifdef HOST_IMM8
// Copy this into local area so we don't have to put it in every literal pool
//rlist();
start = (u_int)addr&~3;
//assert(((u_int)addr&1)==0);
+ new_dynarec_did_compile=1;
#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?
#endif
#ifdef PCSX
case 0x12: strcpy(insn[i],"COP2"); type=NI;
- // note: COP MIPS-1 encoding differs from MIPS32
op2=(source[i]>>21)&0x1f;
- if (source[i]&0x3f) {
+ //if (op2 & 0x10) {
+ if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
if (gte_handlers[source[i]&0x3f]!=NULL) {
- snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
+ if (gte_regnames[source[i]&0x3f]!=NULL)
+ strcpy(insn[i],gte_regnames[source[i]&0x3f]);
+ else
+ snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
type=C2OP;
}
}
us2[i]=0;
dep1[i]=0;
dep2[i]=0;
+ gte_rs[i]=gte_rt[i]=0;
switch(type) {
case LOAD:
rs1[i]=(source[i]>>21)&0x1f;
if(op2==16) if((source[i]&0x3f)==0x18) rs2[i]=CCREG; // ERET
break;
case COP1:
- case COP2:
rs1[i]=0;
rs2[i]=0;
rt1[i]=0;
if(op2==5) us1[i]=rs1[i]; // DMTC1
rs2[i]=CSREG;
break;
+ case COP2:
+ rs1[i]=0;
+ rs2[i]=0;
+ rt1[i]=0;
+ rt2[i]=0;
+ if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC2/CFC2
+ if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC2/CTC2
+ rs2[i]=CSREG;
+ int gr=(source[i]>>11)&0x1F;
+ switch(op2)
+ {
+ case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
+ case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
+ case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
+ case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
+ }
+ break;
case C1LS:
rs1[i]=(source[i]>>21)&0x1F;
rs2[i]=CSREG;
rt1[i]=0;
rt2[i]=0;
imm[i]=(short)source[i];
+ if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
+ else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
+ break;
+ case C2OP:
+ rs1[i]=0;
+ rs2[i]=0;
+ rt1[i]=0;
+ rt2[i]=0;
+ gte_rs[i]=gte_reg_reads[source[i]&0x3f];
+ gte_rt[i]=gte_reg_writes[source[i]&0x3f];
+ gte_rt[i]|=1ll<<63; // every op changes flags
break;
case FLOAT:
case FCONV:
dirty_reg(¤t,CCREG);
current.isconst=0;
current.wasconst=0;
+ current.waswritten=0;
int ds=0;
int cc=0;
int hr=-1;
if(current.regmap[hr]==0) current.regmap[hr]=-1;
}
current.isconst=0;
+ current.waswritten=0;
}
if(i>1)
{
regs[i].wasconst=current.isconst;
regs[i].was32=current.is32;
regs[i].wasdirty=current.dirty;
+ regs[i].loadedconst=0;
#if defined(DESTRUCTIVE_WRITEBACK) && !defined(FORCE32)
// To change a dirty register from 32 to 64 bits, we must write
// it out during the previous cycle (for branches, 2 cycles)
}
memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
}
+
+ if(i>0&&(itype[i-1]==STORE||itype[i-1]==STORELR||(itype[i-1]==C2LS&&opcode[i-1]==0x3a))&&(u_int)imm[i-1]<0x800)
+ current.waswritten|=1<<rs1[i-1];
+ current.waswritten&=~(1<<rt1[i]);
+ current.waswritten&=~(1<<rt2[i]);
+ if((itype[i]==STORE||itype[i]==STORELR||(itype[i]==C2LS&&opcode[i]==0x3a))&&(u_int)imm[i]>=0x800)
+ current.waswritten&=~(1<<rs1[i]);
+
/* Branch post-alloc */
if(i>0)
{
cc=0;
}
#ifdef PCSX
+ else if(itype[i]==C2OP&>e_cycletab[source[i]&0x3f]>2)
+ {
+ // GTE runs in parallel until accessed, divide by 2 for a rough guess
+ cc+=gte_cycletab[source[i]&0x3f]/2;
+ }
else if(/*itype[i]==LOAD||*/itype[i]==STORE||itype[i]==C1LS) // load causes weird timing issues
{
cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
}
}
if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
+ regs[i].waswritten=current.waswritten;
}
/* Pass 4 - Cull unused host registers */
if(itype[slen-1]==SPAN) {
bt[slen-1]=1; // Mark as a branch target so instruction can restart after exception
}
-
+
+#ifdef DISASM
/* Debug/disassembly */
- if((void*)assem_debug==(void*)printf)
for(i=0;i<slen;i++)
{
printf("U:");
#endif
}
}
+#endif // DISASM
/* Pass 8 - Assembly */
linkcount=0;stubcount=0;
for(i=0;i<slen;i++)
{
//if(ds) printf("ds: ");
- if((void*)assem_debug==(void*)printf) disassemble_inst(i);
+ disassemble_inst(i);
if(ds) {
ds=0; // Skip delay slot
if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
instr_addr[i]=0;
} else {
+ speculate_register_values(i);
#ifndef DESTRUCTIVE_WRITEBACK
if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
{
store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
emit_loadreg(CCREG,HOST_CCREG);
- emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i-1]+1),HOST_CCREG);
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
}
else if(!likely[i-2])
{
store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
emit_loadreg(CCREG,HOST_CCREG);
- emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i-1]+1),HOST_CCREG);
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
add_to_linker((int)out,start+i*4,0);
emit_jmp(0);
}
}
inv_code_start=inv_code_end=~0;
#ifdef PCSX
- // PCSX maps all RAM mirror invalid_code tests to 0x80000000..0x80000000+RAM_SIZE
+ // for PCSX we need to mark all mirrors too
if(get_page(start)<(RAM_SIZE>>12))
for(i=start>>12;i<=(start+slen*4)>>12;i++)
- invalid_code[((u_int)0x80000000>>12)|i]=0;
+ invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
+ invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
+ invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
#endif
/* Pass 10 - Free memory by expiring oldest blocks */
notaz.gp2x.de / pcsx_rearmed.git / blobdiff