#include "new_dynarec_config.h"
#include "../psxhle.h"
#include "../psxinterpreter.h"
-#include "emu_if.h" //emulator interface
+#include "../gte.h"
+#include "emu_if.h" // emulator interface
#define noinline __attribute__((noinline,noclone))
#ifndef ARRAY_SIZE
#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
#endif
+#ifndef min
+#define min(a, b) ((b) < (a) ? (b) : (a))
+#endif
+#ifndef max
+#define max(a, b) ((b) > (a) ? (b) : (a))
+#endif
//#define DISASM
-//#define assem_debug printf
-//#define inv_debug printf
+//#define ASSEM_PRINT
+
+#ifdef ASSEM_PRINT
+#define assem_debug printf
+#else
#define assem_debug(...)
+#endif
+//#define inv_debug printf
#define inv_debug(...)
#ifdef __i386__
#include "assem_arm64.h"
#endif
+#define RAM_SIZE 0x200000
#define MAXBLOCK 4096
#define MAX_OUTPUT_BLOCK_SIZE 262144
int new_dynarec_hacks;
int new_dynarec_hacks_pergame;
+ int new_dynarec_hacks_old;
int new_dynarec_did_compile;
#define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
void fp_exception();
void fp_exception_ds();
void jump_to_new_pc();
+void call_gteStall();
void new_dyna_leave();
// Needed by assembler
static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
static void load_regs_entry(int t);
static void load_all_consts(signed char regmap[],u_int dirty,int i);
+static u_int get_host_reglist(const signed char *regmap);
static int verify_dirty(const u_int *ptr);
static int get_final_value(int hr, int i, int *value);
static void add_stub(enum stub_type type, void *addr, void *retaddr,
u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
- int i, int addr_reg, struct regstat *i_regs, int ccadj, u_int reglist);
+ int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
static void add_to_linker(void *addr, u_int target, int ext);
static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override);
static void *get_direct_memhandler(void *table, u_int addr,
enum stub_type type, uintptr_t *addr_host);
+static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
static void pass_args(int a0, int a1);
static void emit_far_jump(const void *f);
static void emit_far_call(const void *f);
int cycle_multiplier; // 100 for 1.0
int cycle_multiplier_override;
+int cycle_multiplier_old;
static int CLOCK_ADJUST(int x)
{
static void host_tempreg_release(void) {}
#endif
-#ifdef DRC_DBG
+#ifdef ASSEM_PRINT
extern void gen_interupt();
extern void do_insn_cmp();
#define FUNCNAME(f) { f, " " #f }
FUNCNAME(jump_handler_write32),
FUNCNAME(invalidate_addr),
FUNCNAME(jump_to_new_pc),
+ FUNCNAME(call_gteStall),
FUNCNAME(new_dyna_leave),
FUNCNAME(pcsx_mtc0),
FUNCNAME(pcsx_mtc0_ds),
+#ifdef DRC_DBG
FUNCNAME(do_insn_cmp),
+#endif
#ifdef __arm__
FUNCNAME(verify_code),
#endif
}
}
-void ll_remove_matching_addrs(struct ll_entry **head,uintptr_t addr,int shift)
+static void ll_remove_matching_addrs(struct ll_entry **head,
+ uintptr_t base_offs_s, int shift)
{
struct ll_entry *next;
while(*head) {
- if(((uintptr_t)((*head)->addr)>>shift)==(addr>>shift) ||
- ((uintptr_t)((*head)->addr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift))
+ uintptr_t o1 = (u_char *)(*head)->addr - ndrc->translation_cache;
+ uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
+ if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
{
inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
remove_hash((*head)->vaddr);
}
// Dereference the pointers and remove if it matches
-static void ll_kill_pointers(struct ll_entry *head,uintptr_t addr,int shift)
+static void ll_kill_pointers(struct ll_entry *head,
+ uintptr_t base_offs_s, int shift)
{
while(head) {
- uintptr_t ptr = (uintptr_t)get_pointer(head->addr);
- inv_debug("EXP: Lookup pointer to %lx at %p (%x)\n",(long)ptr,head->addr,head->vaddr);
- if(((ptr>>shift)==(addr>>shift)) ||
- (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
+ u_char *ptr = get_pointer(head->addr);
+ uintptr_t o1 = ptr - ndrc->translation_cache;
+ uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
+ inv_debug("EXP: Lookup pointer to %p at %p (%x)\n",ptr,head->addr,head->vaddr);
+ if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
{
inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
void *host_addr=find_extjump_insn(head->addr);
static void mov_alloc(struct regstat *current,int i)
{
+ if (rs1[i] == HIREG || rs1[i] == LOREG) {
+ // logically this is needed but just won't work, no idea why
+ //alloc_cc(current,i); // for stalls
+ //dirty_reg(current,CCREG);
+ }
+
// Note: Don't need to actually alloc the source registers
//alloc_reg(current,i,rs1[i]);
alloc_reg(current,i,rt1[i]);
// case 0x1F: DDIVU
clear_const(current,rs1[i]);
clear_const(current,rs2[i]);
+ alloc_cc(current,i); // for stalls
if(rs1[i]&&rs2[i])
{
if((opcode2[i]&4)==0) // 32-bit
minimum_free_regs[i]=HOST_REGS;
}
-static void cop12_alloc(struct regstat *current,int i)
+static void cop2_alloc(struct regstat *current,int i)
{
- alloc_reg(current,i,CSREG); // Load status
- if(opcode2[i]<3) // MFC1/CFC1
+ if (opcode2[i] < 3) // MFC2/CFC2
{
+ alloc_cc(current,i); // for stalls
+ dirty_reg(current,CCREG);
if(rt1[i]){
clear_const(current,rt1[i]);
alloc_reg(current,i,rt1[i]);
dirty_reg(current,rt1[i]);
}
- alloc_reg_temp(current,i,-1);
}
- else if(opcode2[i]>3) // MTC1/CTC1
+ else if (opcode2[i] > 3) // MTC2/CTC2
{
if(rs1[i]){
clear_const(current,rs1[i]);
current->u&=~1LL;
alloc_reg(current,i,0);
}
- alloc_reg_temp(current,i,-1);
}
+ alloc_reg_temp(current,i,-1);
minimum_free_regs[i]=1;
}
void c2op_alloc(struct regstat *current,int i)
{
+ alloc_cc(current,i); // for stalls
+ dirty_reg(current,CCREG);
alloc_reg_temp(current,i,-1);
}
cop0_alloc(current,i);
break;
case COP1:
+ break;
case COP2:
- cop12_alloc(current,i);
+ cop2_alloc(current,i);
break;
case C1LS:
c1ls_alloc(current,i);
}
static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
- int i, int addr_reg, struct regstat *i_regs, int ccadj, u_int reglist)
+ int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
{
add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
}
}
}
-static void load_assemble(int i,struct regstat *i_regs)
+static u_int get_host_reglist(const signed char *regmap)
+{
+ u_int reglist = 0, hr;
+ for (hr = 0; hr < HOST_REGS; hr++) {
+ if (hr != EXCLUDE_REG && regmap[hr] >= 0)
+ reglist |= 1 << hr;
+ }
+ return reglist;
+}
+
+static u_int reglist_exclude(u_int reglist, int r1, int r2)
+{
+ if (r1 >= 0)
+ reglist &= ~(1u << r1);
+ if (r2 >= 0)
+ reglist &= ~(1u << r2);
+ return reglist;
+}
+
+// find a temp caller-saved register not in reglist (so assumed to be free)
+static int reglist_find_free(u_int reglist)
+{
+ u_int free_regs = ~reglist & CALLER_SAVE_REGS;
+ if (free_regs == 0)
+ return -1;
+ return __builtin_ctz(free_regs);
+}
+
+static void load_assemble(int i, const struct regstat *i_regs)
{
int s,tl,addr;
int offset;
void *jaddr=0;
int memtarget=0,c=0;
int fastio_reg_override=-1;
- u_int hr,reglist=0;
+ u_int reglist=get_host_reglist(i_regs->regmap);
tl=get_reg(i_regs->regmap,rt1[i]);
s=get_reg(i_regs->regmap,rs1[i]);
offset=imm[i];
- for(hr=0;hr<HOST_REGS;hr++) {
- if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
- }
if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
if(s>=0) {
c=(i_regs->wasconst>>s)&1;
}
#ifndef loadlr_assemble
-static void loadlr_assemble(int i,struct regstat *i_regs)
+static void loadlr_assemble(int i, const struct regstat *i_regs)
{
int s,tl,temp,temp2,addr;
int offset;
void *jaddr=0;
int memtarget=0,c=0;
int fastio_reg_override=-1;
- u_int hr,reglist=0;
+ u_int reglist=get_host_reglist(i_regs->regmap);
tl=get_reg(i_regs->regmap,rt1[i]);
s=get_reg(i_regs->regmap,rs1[i]);
temp=get_reg(i_regs->regmap,-1);
addr=get_reg(i_regs->regmap,AGEN1+(i&1));
assert(addr<0);
offset=imm[i];
- for(hr=0;hr<HOST_REGS;hr++) {
- if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
- }
reglist|=1<<temp;
if(offset||s<0||c) addr=temp2;
else addr=s;
}
#endif
-void store_assemble(int i,struct regstat *i_regs)
+void store_assemble(int i, const struct regstat *i_regs)
{
int s,tl;
int addr,temp;
int memtarget=0,c=0;
int agr=AGEN1+(i&1);
int fastio_reg_override=-1;
- u_int hr,reglist=0;
+ u_int reglist=get_host_reglist(i_regs->regmap);
tl=get_reg(i_regs->regmap,rs2[i]);
s=get_reg(i_regs->regmap,rs1[i]);
temp=get_reg(i_regs->regmap,agr);
}
assert(tl>=0);
assert(temp>=0);
- for(hr=0;hr<HOST_REGS;hr++) {
- if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
- }
if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
if(offset||s<0||c) addr=temp;
else addr=s;
}
}
-static void storelr_assemble(int i,struct regstat *i_regs)
+static void storelr_assemble(int i, const struct regstat *i_regs)
{
int s,tl;
int temp;
void *done0, *done1, *done2;
int memtarget=0,c=0;
int agr=AGEN1+(i&1);
- u_int hr,reglist=0;
+ u_int reglist=get_host_reglist(i_regs->regmap);
tl=get_reg(i_regs->regmap,rs2[i]);
s=get_reg(i_regs->regmap,rs1[i]);
temp=get_reg(i_regs->regmap,agr);
}
}
assert(tl>=0);
- for(hr=0;hr<HOST_REGS;hr++) {
- if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
- }
assert(temp>=0);
if(!c) {
emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
emit_far_jump(ds?fp_exception_ds:fp_exception);
}
+static int cop2_is_stalling_op(int i, int *cycles)
+{
+ if (opcode[i] == 0x3a) { // SWC2
+ *cycles = 0;
+ return 1;
+ }
+ if (itype[i] == COP2 && (opcode2[i] == 0 || opcode2[i] == 2)) { // MFC2/CFC2
+ *cycles = 0;
+ return 1;
+ }
+ if (itype[i] == C2OP) {
+ *cycles = gte_cycletab[source[i] & 0x3f];
+ return 1;
+ }
+ // ... what about MTC2/CTC2/LWC2?
+ return 0;
+}
+
+#if 0
+static void log_gte_stall(int stall, u_int cycle)
+{
+ if ((u_int)stall <= 44)
+ printf("x stall %2d %u\n", stall, cycle + last_count);
+}
+
+static void emit_log_gte_stall(int i, int stall, u_int reglist)
+{
+ save_regs(reglist);
+ if (stall > 0)
+ emit_movimm(stall, 0);
+ else
+ emit_mov(HOST_TEMPREG, 0);
+ emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
+ emit_far_call(log_gte_stall);
+ restore_regs(reglist);
+}
+#endif
+
+static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
+{
+ int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
+ int rtmp = reglist_find_free(reglist);
+
+ if (HACK_ENABLED(NDHACK_NO_STALLS))
+ return;
+ if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
+ // happens occasionally... cc evicted? Don't bother then
+ //printf("no cc %08x\n", start + i*4);
+ return;
+ }
+ if (!bt[i]) {
+ for (j = i - 1; j >= 0; j--) {
+ //if (is_ds[j]) break;
+ if (cop2_is_stalling_op(j, &other_gte_op_cycles) || bt[j])
+ break;
+ }
+ j = max(j, 0);
+ }
+ cycles_passed = CLOCK_ADJUST(ccadj[i] - ccadj[j]);
+ if (other_gte_op_cycles >= 0)
+ stall = other_gte_op_cycles - cycles_passed;
+ else if (cycles_passed >= 44)
+ stall = 0; // can't stall
+ if (stall == -MAXBLOCK && rtmp >= 0) {
+ // unknown stall, do the expensive runtime check
+ assem_debug("; cop2_do_stall_check\n");
+#if 0 // too slow
+ save_regs(reglist);
+ emit_movimm(gte_cycletab[op], 0);
+ emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
+ emit_far_call(call_gteStall);
+ restore_regs(reglist);
+#else
+ host_tempreg_acquire();
+ emit_readword(&psxRegs.gteBusyCycle, rtmp);
+ emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
+ emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
+ emit_cmpimm(HOST_TEMPREG, 44);
+ emit_cmovb_reg(rtmp, HOST_CCREG);
+ //emit_log_gte_stall(i, 0, reglist);
+ host_tempreg_release();
+#endif
+ }
+ else if (stall > 0) {
+ //emit_log_gte_stall(i, stall, reglist);
+ emit_addimm(HOST_CCREG, stall, HOST_CCREG);
+ }
+
+ // save gteBusyCycle, if needed
+ if (gte_cycletab[op] == 0)
+ return;
+ other_gte_op_cycles = -1;
+ for (j = i + 1; j < slen; j++) {
+ if (cop2_is_stalling_op(j, &other_gte_op_cycles))
+ break;
+ if (is_jump(j)) {
+ // check ds
+ if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
+ j++;
+ break;
+ }
+ }
+ if (other_gte_op_cycles >= 0)
+ // will handle stall when assembling that op
+ return;
+ cycles_passed = CLOCK_ADJUST(ccadj[min(j, slen -1)] - ccadj[i]);
+ if (cycles_passed >= 44)
+ return;
+ assem_debug("; save gteBusyCycle\n");
+ host_tempreg_acquire();
+#if 0
+ emit_readword(&last_count, HOST_TEMPREG);
+ emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
+ emit_addimm(HOST_TEMPREG, CLOCK_ADJUST(ccadj[i]), HOST_TEMPREG);
+ emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
+ emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
+#else
+ emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + gte_cycletab[op], HOST_TEMPREG);
+ emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
+#endif
+ host_tempreg_release();
+}
+
+static int is_mflohi(int i)
+{
+ return (itype[i] == MOV && (rs1[i] == HIREG || rs1[i] == LOREG));
+}
+
+static int check_multdiv(int i, int *cycles)
+{
+ if (itype[i] != MULTDIV)
+ return 0;
+ if (opcode2[i] == 0x18 || opcode2[i] == 0x19) // MULT(U)
+ *cycles = 11; // approx from 7 11 14
+ else
+ *cycles = 37;
+ return 1;
+}
+
+static void multdiv_prepare_stall(int i, const struct regstat *i_regs)
+{
+ int j, found = 0, c = 0;
+ if (HACK_ENABLED(NDHACK_NO_STALLS))
+ return;
+ if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
+ // happens occasionally... cc evicted? Don't bother then
+ return;
+ }
+ for (j = i + 1; j < slen; j++) {
+ if (bt[j])
+ break;
+ if ((found = is_mflohi(j)))
+ break;
+ if (is_jump(j)) {
+ // check ds
+ if (j + 1 < slen && (found = is_mflohi(j + 1)))
+ j++;
+ break;
+ }
+ }
+ if (found)
+ // handle all in multdiv_do_stall()
+ return;
+ check_multdiv(i, &c);
+ assert(c > 0);
+ assem_debug("; muldiv prepare stall %d\n", c);
+ host_tempreg_acquire();
+ emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + c, HOST_TEMPREG);
+ emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
+ host_tempreg_release();
+}
+
+static void multdiv_do_stall(int i, const struct regstat *i_regs)
+{
+ int j, known_cycles = 0;
+ u_int reglist = get_host_reglist(i_regs->regmap);
+ int rtmp = get_reg(i_regs->regmap, -1);
+ if (rtmp < 0)
+ rtmp = reglist_find_free(reglist);
+ if (HACK_ENABLED(NDHACK_NO_STALLS))
+ return;
+ if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
+ // happens occasionally... cc evicted? Don't bother then
+ //printf("no cc/rtmp %08x\n", start + i*4);
+ return;
+ }
+ if (!bt[i]) {
+ for (j = i - 1; j >= 0; j--) {
+ if (is_ds[j]) break;
+ if (check_multdiv(j, &known_cycles) || bt[j])
+ break;
+ if (is_mflohi(j))
+ // already handled by this op
+ return;
+ }
+ j = max(j, 0);
+ }
+ if (known_cycles > 0) {
+ known_cycles -= CLOCK_ADJUST(ccadj[i] - ccadj[j]);
+ assem_debug("; muldiv stall resolved %d\n", known_cycles);
+ if (known_cycles > 0)
+ emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
+ return;
+ }
+ assem_debug("; muldiv stall unresolved\n");
+ host_tempreg_acquire();
+ emit_readword(&psxRegs.muldivBusyCycle, rtmp);
+ emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
+ emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
+ emit_cmpimm(HOST_TEMPREG, 37);
+ emit_cmovb_reg(rtmp, HOST_CCREG);
+ //emit_log_gte_stall(i, 0, reglist);
+ host_tempreg_release();
+}
+
static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
{
switch (copr) {
}
}
-static void c2ls_assemble(int i,struct regstat *i_regs)
+static void c2ls_assemble(int i, const struct regstat *i_regs)
{
int s,tl;
int ar;
enum stub_type type;
int agr=AGEN1+(i&1);
int fastio_reg_override=-1;
- u_int hr,reglist=0;
+ u_int reglist=get_host_reglist(i_regs->regmap);
u_int copr=(source[i]>>16)&0x1f;
s=get_reg(i_regs->regmap,rs1[i]);
tl=get_reg(i_regs->regmap,FTEMP);
assert(rs1[i]>0);
assert(tl>=0);
- for(hr=0;hr<HOST_REGS;hr++) {
- if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
- }
if(i_regs->regmap[HOST_CCREG]==CCREG)
reglist&=~(1<<HOST_CCREG);
if (!offset&&!c&&s>=0) ar=s;
assert(ar>=0);
+ cop2_do_stall_check(0, i, i_regs, reglist);
+
if (opcode[i]==0x3a) { // SWC2
cop2_get_dreg(copr,tl,-1);
type=STOREW_STUB;
}
}
-static void cop2_assemble(int i,struct regstat *i_regs)
+static void cop2_assemble(int i, const struct regstat *i_regs)
{
- u_int copr=(source[i]>>11)&0x1f;
- signed char temp=get_reg(i_regs->regmap,-1);
+ u_int copr = (source[i]>>11) & 0x1f;
+ signed char temp = get_reg(i_regs->regmap, -1);
+
+ if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
+ u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
+ if (opcode2[i] == 0 || opcode2[i] == 2) { // MFC2/CFC2
+ signed char tl = get_reg(i_regs->regmap, rt1[i]);
+ reglist = reglist_exclude(reglist, tl, -1);
+ }
+ cop2_do_stall_check(0, i, i_regs, reglist);
+ }
if (opcode2[i]==0) { // MFC2
signed char tl=get_reg(i_regs->regmap,rt1[i]);
if(tl>=0&&rt1[i]!=0)
else emit_loadreg(rs1[i],tl);
}
}
+ if (rs1[i] == HIREG || rs1[i] == LOREG) // MFHI/MFLO
+ multdiv_do_stall(i, i_regs);
}
// call interpreter, exception handler, things that change pc/regs/cycles ...
case C2OP:
c2op_assemble(i,i_regs);break;
case MULTDIV:
- multdiv_assemble(i,i_regs);break;
+ multdiv_assemble(i,i_regs);
+ multdiv_prepare_stall(i,i_regs);
+ break;
case MOV:
mov_assemble(i,i_regs);break;
case SYSCALL:
{
extern void do_insn_cmp();
//extern int cycle;
- u_int hr,reglist=0;
+ u_int hr, reglist = get_host_reglist(regs[i].regmap);
assem_debug("//do_insn_cmp %08x\n", start+i*4);
- for (hr = 0; hr < HOST_REGS; hr++)
- if(regs[i].regmap[hr]>=0) reglist|=1<<hr;
save_regs(reglist);
// write out changed consts to match the interpreter
if (i > 0 && !bt[i]) {
case C2OP:
c2op_assemble(t,®s[t]);break;
case MULTDIV:
- multdiv_assemble(t,®s[t]);break;
+ multdiv_assemble(t,®s[t]);
+ multdiv_prepare_stall(i,®s[t]);
+ break;
case MOV:
mov_assemble(t,®s[t]);break;
case SYSCALL:
case C2OP:
c2op_assemble(0,®s[0]);break;
case MULTDIV:
- multdiv_assemble(0,®s[0]);break;
+ multdiv_assemble(0,®s[0]);
+ multdiv_prepare_stall(0,®s[0]);
+ break;
case MOV:
mov_assemble(0,®s[0]);break;
case SYSCALL:
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);
+
+ cycle_multiplier_old = cycle_multiplier;
+ new_dynarec_hacks_old = new_dynarec_hacks;
}
void new_dynarec_init(void)
cop0_alloc(¤t,i);
break;
case COP1:
+ break;
case COP2:
- cop12_alloc(¤t,i);
+ cop2_alloc(¤t,i);
break;
case C1LS:
c1ls_alloc(¤t,i);
#if !defined(DRC_DBG)
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,store causes weird timing issues
- {
- cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
+ // this should really be removed since the real stalls have been implemented,
+ // but doing so causes sizeable perf regression against the older version
+ u_int gtec = gte_cycletab[source[i] & 0x3f];
+ cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
}
else if(i>1&&itype[i]==STORE&&itype[i-1]==STORE&&itype[i-2]==STORE&&!bt[i])
{
}
else if(itype[i]==C2LS)
{
- cc+=4;
+ // same as with C2OP
+ cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
}
#endif
else
case C2OP:
c2op_assemble(i,®s[i]);break;
case MULTDIV:
- multdiv_assemble(i,®s[i]);break;
+ multdiv_assemble(i,®s[i]);
+ multdiv_prepare_stall(i,®s[i]);
+ break;
case MOV:
mov_assemble(i,®s[i]);break;
case SYSCALL:
while(expirep!=end)
{
int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
- uintptr_t base=(uintptr_t)ndrc->translation_cache+((expirep>>13)<<shift); // Base address of this block
+ uintptr_t base_offs = ((uintptr_t)(expirep >> 13) << shift); // Base offset of this block
+ uintptr_t base_offs_s = base_offs >> shift;
inv_debug("EXP: Phase %d\n",expirep);
switch((expirep>>11)&3)
{
case 0:
// Clear jump_in and jump_dirty
- ll_remove_matching_addrs(jump_in+(expirep&2047),base,shift);
- ll_remove_matching_addrs(jump_dirty+(expirep&2047),base,shift);
- ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base,shift);
- ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base,shift);
+ ll_remove_matching_addrs(jump_in+(expirep&2047),base_offs_s,shift);
+ ll_remove_matching_addrs(jump_dirty+(expirep&2047),base_offs_s,shift);
+ ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base_offs_s,shift);
+ ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base_offs_s,shift);
break;
case 1:
// Clear pointers
- ll_kill_pointers(jump_out[expirep&2047],base,shift);
- ll_kill_pointers(jump_out[(expirep&2047)+2048],base,shift);
+ ll_kill_pointers(jump_out[expirep&2047],base_offs_s,shift);
+ ll_kill_pointers(jump_out[(expirep&2047)+2048],base_offs_s,shift);
break;
case 2:
// Clear hash table
for(i=0;i<32;i++) {
struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
- if (((uintptr_t)ht_bin->tcaddr[1]>>shift) == (base>>shift) ||
- (((uintptr_t)ht_bin->tcaddr[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
+ uintptr_t o1 = (u_char *)ht_bin->tcaddr[1] - ndrc->translation_cache;
+ uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
+ if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
ht_bin->vaddr[1] = -1;
ht_bin->tcaddr[1] = NULL;
}
- if (((uintptr_t)ht_bin->tcaddr[0]>>shift) == (base>>shift) ||
- (((uintptr_t)ht_bin->tcaddr[0]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
+ o1 = (u_char *)ht_bin->tcaddr[0] - ndrc->translation_cache;
+ o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
+ if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
ht_bin->vaddr[0] = ht_bin->vaddr[1];
ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
// Clear jump_out
if((expirep&2047)==0)
do_clear_cache();
- ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
- ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
+ ll_remove_matching_addrs(jump_out+(expirep&2047),base_offs_s,shift);
+ ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base_offs_s,shift);
break;
}
expirep=(expirep+1)&65535;
notaz.gp2x.de / pcsx_rearmed.git / blobdiff