#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__
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))
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_call_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
+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(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);
inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
void *host_addr=find_extjump_insn(head->addr);
mark_clear_cache(host_addr);
- set_jump_target(host_addr, head->addr);
+ set_jump_target(host_addr, head->addr); // point back to dyna_linker
next=head->next;
free(head);
head=next;
// Add an entry to jump_out after making a link
// src should point to code by emit_extjump2()
-void add_link(u_int vaddr,void *src)
+void add_jump_out(u_int vaddr,void *src)
{
u_int page=get_page(vaddr);
- inv_debug("add_link: %p -> %x (%d)\n",src,vaddr,page);
+ inv_debug("add_jump_out: %p -> %x (%d)\n",src,vaddr,page);
check_extjump2(src);
ll_add(jump_out+page,vaddr,src);
- //void *ptr=get_pointer(src);
- //inv_debug("add_link: Pointer is to %p\n",ptr);
+ //inv_debug("add_jump_out: to %p\n",get_pointer(src));
}
// If a code block was found to be unmodified (bit was set in
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
{
if ((u_int)stall <= 44)
printf("x stall %2d %u\n", stall, cycle + last_count);
- if (cycle + last_count > 1215348544) exit(1);
}
static void emit_log_gte_stall(int i, int stall, u_int reglist)
}
#endif
-static void cop2_call_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
+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_GTE_NO_STALL))
+ if (HACK_ENABLED(NDHACK_NO_STALLS))
return;
- //assert(get_reg(i_regs->regmap, CCREG) == HOST_CCREG);
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);
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 = 0; // can't stall
if (stall == -MAXBLOCK && rtmp >= 0) {
// unknown stall, do the expensive runtime check
- assem_debug("; cop2_call_stall_check\n");
+ assem_debug("; cop2_do_stall_check\n");
#if 0 // too slow
save_regs(reglist);
emit_movimm(gte_cycletab[op], 0);
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) {
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_call_stall_check(0, i, i_regs, reglist_exclude(reglist, tl, -1));
cop2_get_dreg(copr,tl,-1);
type=STOREW_STUB;
}
u_int copr = (source[i]>>11) & 0x1f;
signed char temp = get_reg(i_regs->regmap, -1);
- if (opcode2[i] == 0 || opcode2[i] == 2) { // MFC2/CFC2
- if (!HACK_ENABLED(NDHACK_GTE_NO_STALL)) {
+ 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]);
- u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), tl, temp);
- cop2_call_stall_check(0, i, i_regs, reglist);
+ 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]);
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:
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:
emit_extjump_ds(branch_addr, target_addr);
if(compiled_target_addr) {
set_jump_target(branch_addr, compiled_target_addr);
- add_link(target_addr,stub);
+ add_jump_out(target_addr,stub);
}
else set_jump_target(branch_addr, stub);
if(likely[i]) {
emit_extjump_ds(branch_addr, target_addr);
if(compiled_target_addr) {
set_jump_target(branch_addr, compiled_target_addr);
- add_link(target_addr,stub);
+ add_jump_out(target_addr,stub);
}
else set_jump_target(branch_addr, stub);
}
u_int page=get_page(vaddr);
u_int vpage=get_vpage(vaddr);
ll_add(jump_dirty+vpage,vaddr,(void *)out);
- do_dirty_stub_ds();
+ do_dirty_stub_ds(slen*4);
ll_add(jump_in+page,vaddr,(void *)out);
assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
if(regs[0].regmap[HOST_CCREG]!=CCREG)
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)
// 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_GTE_NO_STALL) ? gtec/2 : 2;
+ 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)
{
// same as with C2OP
- cc += HACK_ENABLED(NDHACK_GTE_NO_STALL) ? 4 : 2;
+ 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:
literal_pool_jumpover(256);
}
}
- //assert(is_ujump(i-2));
+
+ assert(slen > 0);
+ if (itype[slen-1] == INTCALL) {
+ // no ending needed for this block since INTCALL never returns
+ }
// If the block did not end with an unconditional branch,
// add a jump to the next instruction.
- if(i>1) {
+ else if (i > 1) {
if(!is_ujump(i-2)&&itype[i-1]!=SPAN) {
assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP);
assert(i==slen);
emit_extjump(link_addr[i].addr, link_addr[i].target);
if (addr) {
set_jump_target(link_addr[i].addr, addr);
- add_link(link_addr[i].target,stub);
+ add_jump_out(link_addr[i].target,stub);
}
else
set_jump_target(link_addr[i].addr, stub);
//#endif
}
}
+
+ u_int source_len = slen*4;
+ if (itype[slen-1] == INTCALL && source_len > 4)
+ // no need to treat the last instruction as compiled
+ // as interpreter fully handles it
+ source_len -= 4;
+
+ if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
+ copy = shadow;
+
// External Branch Targets (jump_in)
- if(copy+slen*4>(void *)shadow+sizeof(shadow)) copy=shadow;
for(i=0;i<slen;i++)
{
if(bt[i]||i==0)
assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
assem_debug("jump_in: %x\n",start+i*4);
ll_add(jump_dirty+vpage,vaddr,out);
- void *entry_point = do_dirty_stub(i);
+ void *entry_point = do_dirty_stub(i, source_len);
ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
// If there was an existing entry in the hash table,
// replace it with the new address.
#endif
assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
//printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
- memcpy(copy,source,slen*4);
- copy+=slen*4;
+ memcpy(copy, source, source_len);
+ copy += source_len;
end_block(beginning);
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