#endif
#include "new_dynarec_config.h"
-#include "../psxhle.h" //emulator interface
+#include "../psxhle.h"
+#include "../psxinterpreter.h"
#include "emu_if.h" //emulator interface
#define noinline __attribute__((noinline,noclone))
#define MAXBLOCK 4096
#define MAX_OUTPUT_BLOCK_SIZE 262144
+struct ndrc_mem
+{
+ u_char translation_cache[1 << TARGET_SIZE_2];
+ struct
+ {
+ struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
+ const void *f[2048 / sizeof(void *)];
+ } tramp;
+};
+
+#ifdef BASE_ADDR_DYNAMIC
+static struct ndrc_mem *ndrc;
+#else
+static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
+static struct ndrc_mem *ndrc = &ndrc_;
+#endif
+
// stubs
enum stub_type {
CC_STUB = 1,
static char ooo[MAXBLOCK];
static uint64_t unneeded_reg[MAXBLOCK];
static uint64_t branch_unneeded_reg[MAXBLOCK];
- static signed char regmap_pre[MAXBLOCK][HOST_REGS];
- static uint64_t current_constmap[HOST_REGS];
- static uint64_t constmap[MAXBLOCK][HOST_REGS];
+ static signed char regmap_pre[MAXBLOCK][HOST_REGS]; // pre-instruction i?
+ // contains 'real' consts at [i] insn, but may differ from what's actually
+ // loaded in host reg as 'final' value is always loaded, see get_final_value()
+ static uint32_t current_constmap[HOST_REGS];
+ static uint32_t constmap[MAXBLOCK][HOST_REGS];
static struct regstat regs[MAXBLOCK];
static struct regstat branch_regs[MAXBLOCK];
static signed char minimum_free_regs[MAXBLOCK];
#endif
int new_dynarec_hacks;
+ int new_dynarec_hacks_pergame;
int new_dynarec_did_compile;
+ #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
+
extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
extern int last_count; // last absolute target, often = next_interupt
extern int pcaddr;
#define DJT_2 (void *)2l
// asm linkage
-int new_recompile_block(int addr);
+int new_recompile_block(u_int addr);
void *get_addr_ht(u_int vaddr);
void invalidate_block(u_int block);
void invalidate_addr(u_int addr);
void cc_interrupt();
void fp_exception();
void fp_exception_ds();
-void jump_syscall_hle();
-void jump_hlecall();
-void jump_intcall();
+void jump_to_new_pc();
void new_dyna_leave();
// Needed by assembler
static void load_regs_entry(int t);
static void load_all_consts(signed char regmap[],u_int dirty,int i);
-static int verify_dirty(u_int *ptr);
+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 *get_direct_memhandler(void *table, u_int addr,
enum stub_type type, uintptr_t *addr_host);
static void pass_args(int a0, int a1);
+static void emit_far_jump(const void *f);
+static void emit_far_call(const void *f);
static void mprotect_w_x(void *start, void *end, int is_x)
{
static void end_tcache_write(void *start, void *end)
{
-#ifdef __arm__
+#if defined(__arm__) || defined(__aarch64__)
size_t len = (char *)end - (char *)start;
#if defined(__BLACKBERRY_QNX__)
msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
sceKernelSyncVMDomain(sceBlock, start, len);
#elif defined(_3DS)
ctr_flush_invalidate_cache();
+ #elif defined(__aarch64__)
+ // as of 2021, __clear_cache() is still broken on arm64
+ // so here is a custom one :(
+ clear_cache_arm64(start, end);
#else
__clear_cache(start, end);
#endif
(void)len;
-#else
- __clear_cache(start, end);
#endif
mprotect_w_x(start, end, 1);
static void *start_block(void)
{
u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
- if (end > translation_cache + (1<<TARGET_SIZE_2))
- end = translation_cache + (1<<TARGET_SIZE_2);
+ if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
+ end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
start_tcache_write(out, end);
return out;
}
end_tcache_write(start, out);
}
+// also takes care of w^x mappings when patching code
+static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
+
+static void mark_clear_cache(void *target)
+{
+ uintptr_t offset = (u_char *)target - ndrc->translation_cache;
+ u_int mask = 1u << ((offset >> 12) & 31);
+ if (!(needs_clear_cache[offset >> 17] & mask)) {
+ char *start = (char *)((uintptr_t)target & ~4095l);
+ start_tcache_write(start, start + 4095);
+ needs_clear_cache[offset >> 17] |= mask;
+ }
+}
+
+// Clearing the cache is rather slow on ARM Linux, so mark the areas
+// that need to be cleared, and then only clear these areas once.
+static void do_clear_cache(void)
+{
+ int i, j;
+ for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
+ {
+ u_int bitmap = needs_clear_cache[i];
+ if (!bitmap)
+ continue;
+ for (j = 0; j < 32; j++)
+ {
+ u_char *start, *end;
+ if (!(bitmap & (1<<j)))
+ continue;
+
+ start = ndrc->translation_cache + i*131072 + j*4096;
+ end = start + 4095;
+ for (j++; j < 32; j++) {
+ if (!(bitmap & (1<<j)))
+ break;
+ end += 4096;
+ }
+ end_tcache_write(start, end);
+ }
+ needs_clear_cache[i] = 0;
+ }
+}
+
//#define DEBUG_CYCLE_COUNT 1
#define NO_CYCLE_PENALTY_THR 12
int cycle_multiplier; // 100 for 1.0
+int cycle_multiplier_override;
static int CLOCK_ADJUST(int x)
{
+ int m = cycle_multiplier_override
+ ? cycle_multiplier_override : cycle_multiplier;
int s=(x>>31)|1;
- return (x * cycle_multiplier + s * 50) / 100;
+ return (x * m + s * 50) / 100;
}
static u_int get_page(u_int vaddr)
}
}
-void set_const(struct regstat *cur,signed char reg,uint64_t value)
+static void set_const(struct regstat *cur, signed char reg, uint32_t value)
{
int hr;
if(!reg) return;
}
}
-void clear_const(struct regstat *cur,signed char reg)
+static void clear_const(struct regstat *cur, signed char reg)
{
int hr;
if(!reg) return;
}
}
-int is_const(struct regstat *cur,signed char reg)
+static int is_const(struct regstat *cur, signed char reg)
{
int hr;
if(reg<0) return 0;
}
return 0;
}
-uint64_t get_const(struct regstat *cur,signed char reg)
+
+static uint32_t get_const(struct regstat *cur, signed char reg)
{
int hr;
if(!reg) return 0;
FUNCNAME(jump_handler_write16),
FUNCNAME(jump_handler_write32),
FUNCNAME(invalidate_addr),
- FUNCNAME(verify_code),
- FUNCNAME(jump_hlecall),
- FUNCNAME(jump_syscall_hle),
+ FUNCNAME(jump_to_new_pc),
FUNCNAME(new_dyna_leave),
FUNCNAME(pcsx_mtc0),
FUNCNAME(pcsx_mtc0_ds),
FUNCNAME(do_insn_cmp),
+#ifdef __arm__
+ FUNCNAME(verify_code),
+#endif
};
static const char *func_name(const void *a)
#include "assem_arm64.c"
#endif
+static void *get_trampoline(const void *f)
+{
+ size_t i;
+
+ for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
+ if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
+ break;
+ }
+ if (i == ARRAY_SIZE(ndrc->tramp.f)) {
+ SysPrintf("trampoline table is full, last func %p\n", f);
+ abort();
+ }
+ if (ndrc->tramp.f[i] == NULL) {
+ start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
+ ndrc->tramp.f[i] = f;
+ end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
+ }
+ return &ndrc->tramp.ops[i];
+}
+
+static void emit_far_jump(const void *f)
+{
+ if (can_jump_or_call(f)) {
+ emit_jmp(f);
+ return;
+ }
+
+ f = get_trampoline(f);
+ emit_jmp(f);
+}
+
+static void emit_far_call(const void *f)
+{
+ if (can_jump_or_call(f)) {
+ emit_call(f);
+ return;
+ }
+
+ f = get_trampoline(f);
+ emit_call(f);
+}
+
// Add virtual address mapping to linked list
void ll_add(struct ll_entry **head,int vaddr,void *addr)
{
{
inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
void *host_addr=find_extjump_insn(head->addr);
- #if defined(__arm__) || defined(__aarch64__)
- mark_clear_cache(host_addr);
- #endif
+ mark_clear_cache(host_addr);
set_jump_target(host_addr, head->addr);
}
head=head->next;
while(head!=NULL) {
inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
void *host_addr=find_extjump_insn(head->addr);
- #if defined(__arm__) || defined(__aarch64__)
- mark_clear_cache(host_addr);
- #endif
+ mark_clear_cache(host_addr);
set_jump_target(host_addr, head->addr);
next=head->next;
free(head);
for(first=page+1;first<last;first++) {
invalidate_page(first);
}
- #if defined(__arm__) || defined(__aarch64__)
- do_clear_cache();
- #endif
+ do_clear_cache();
// Don't trap writes
invalid_code[block]=1;
// This is called when loading a save state.
// Anything could have changed, so invalidate everything.
-void invalidate_all_pages()
+void invalidate_all_pages(void)
{
u_int page;
for(page=0;page<4096;page++)
#ifdef USE_MINI_HT
memset(mini_ht,-1,sizeof(mini_ht));
#endif
+ do_clear_cache();
}
static void do_invstub(int n)
set_jump_target(stubs[n].addr, out);
save_regs(reglist);
if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
- emit_call(invalidate_addr);
+ emit_far_call(invalidate_addr);
restore_regs(reglist);
emit_jmp(stubs[n].retaddr); // return address
}
else clear_const(current,rt1[i]);
}
else {
- set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
+ set_const(current,rt1[i],imm[i]<<16); // LUI
}
dirty_reg(current,rt1[i]);
}
s2l=get_reg(i_regs->regmap,rs2[i]);
if(rs2[i]==0) // rx<r0
{
- assert(s1l>=0);
- if(opcode2[i]==0x2a) // SLT
+ if(opcode2[i]==0x2a&&rs1[i]!=0) { // SLT
+ assert(s1l>=0);
emit_shrimm(s1l,31,t);
- else // SLTU (unsigned can not be less than zero)
+ }
+ else // SLTU (unsigned can not be less than zero, 0<0)
emit_zeroreg(t);
}
else if(rs1[i]==0) // r0<rx
}
#ifndef shift_assemble
-void shift_assemble(int i,struct regstat *i_regs)
+static void shift_assemble(int i,struct regstat *i_regs)
{
- printf("Need shift_assemble for this architecture.\n");
- abort();
+ signed char s,t,shift;
+ if (rt1[i] == 0)
+ return;
+ assert(opcode2[i]<=0x07); // SLLV/SRLV/SRAV
+ t = get_reg(i_regs->regmap, rt1[i]);
+ s = get_reg(i_regs->regmap, rs1[i]);
+ shift = get_reg(i_regs->regmap, rs2[i]);
+ if (t < 0)
+ return;
+
+ if(rs1[i]==0)
+ emit_zeroreg(t);
+ else if(rs2[i]==0) {
+ assert(s>=0);
+ if(s!=t) emit_mov(s,t);
+ }
+ else {
+ host_tempreg_acquire();
+ emit_andimm(shift,31,HOST_TEMPREG);
+ switch(opcode2[i]) {
+ case 4: // SLLV
+ emit_shl(s,HOST_TEMPREG,t);
+ break;
+ case 6: // SRLV
+ emit_shr(s,HOST_TEMPREG,t);
+ break;
+ case 7: // SRAV
+ emit_sar(s,HOST_TEMPREG,t);
+ break;
+ default:
+ assert(0);
+ }
+ host_tempreg_release();
+ }
}
+
#endif
enum {
else if(type==MTYPE_1F80) { // scratchpad
if (psxH == (void *)0x1f800000) {
host_tempreg_acquire();
- emit_addimm(addr,-0x1f800000,HOST_TEMPREG);
+ emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
emit_cmpimm(HOST_TEMPREG,0x1000);
host_tempreg_release();
jaddr=out;
}
#ifndef loadlr_assemble
-void loadlr_assemble(int i,struct regstat *i_regs)
+static void loadlr_assemble(int i,struct regstat *i_regs)
{
- printf("Need loadlr_assemble for this architecture.\n");
- abort();
+ 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;
+ tl=get_reg(i_regs->regmap,rt1[i]);
+ s=get_reg(i_regs->regmap,rs1[i]);
+ temp=get_reg(i_regs->regmap,-1);
+ temp2=get_reg(i_regs->regmap,FTEMP);
+ 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;
+ if(s>=0) {
+ c=(i_regs->wasconst>>s)&1;
+ if(c) {
+ memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
+ }
+ }
+ if(!c) {
+ emit_shlimm(addr,3,temp);
+ if (opcode[i]==0x22||opcode[i]==0x26) {
+ emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
+ }else{
+ emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
+ }
+ jaddr=emit_fastpath_cmp_jump(i,temp2,&fastio_reg_override);
+ }
+ else {
+ if(ram_offset&&memtarget) {
+ host_tempreg_acquire();
+ emit_addimm(temp2,ram_offset,HOST_TEMPREG);
+ fastio_reg_override=HOST_TEMPREG;
+ }
+ if (opcode[i]==0x22||opcode[i]==0x26) {
+ emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
+ }else{
+ emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
+ }
+ }
+ if (opcode[i]==0x22||opcode[i]==0x26) { // LWL/LWR
+ if(!c||memtarget) {
+ int a=temp2;
+ if(fastio_reg_override>=0) a=fastio_reg_override;
+ emit_readword_indexed(0,a,temp2);
+ if(fastio_reg_override==HOST_TEMPREG) host_tempreg_release();
+ if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj[i],reglist);
+ }
+ else
+ inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj[i],reglist);
+ if(rt1[i]) {
+ assert(tl>=0);
+ emit_andimm(temp,24,temp);
+ if (opcode[i]==0x22) // LWL
+ emit_xorimm(temp,24,temp);
+ host_tempreg_acquire();
+ emit_movimm(-1,HOST_TEMPREG);
+ if (opcode[i]==0x26) {
+ emit_shr(temp2,temp,temp2);
+ emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
+ }else{
+ emit_shl(temp2,temp,temp2);
+ emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
+ }
+ host_tempreg_release();
+ emit_or(temp2,tl,tl);
+ }
+ //emit_storereg(rt1[i],tl); // DEBUG
+ }
+ if (opcode[i]==0x1A||opcode[i]==0x1B) { // LDL/LDR
+ assert(0);
+ }
}
#endif
add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
jaddr=0;
}
- if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
+ if(!(i_regs->waswritten&(1<<rs1[i])) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
if(!c||memtarget) {
#ifdef DESTRUCTIVE_SHIFT
// The x86 shift operation is 'destructive'; it overwrites the
}
// basic current block modification detection..
// not looking back as that should be in mips cache already
- // (note: doesn't seem to trigger, migh be broken)
+ // (see Spyro2 title->attract mode)
if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
assert(i_regs->regmap==regs[i].regmap); // not delay slot
emit_movimm(start+i*4+4,0);
emit_writeword(0,&pcaddr);
emit_addimm(HOST_CCREG,2,HOST_CCREG);
- emit_call(get_addr_ht);
+ emit_far_call(get_addr_ht);
emit_jmpreg(0);
}
}
}
-void storelr_assemble(int i,struct regstat *i_regs)
+static void storelr_assemble(int i,struct regstat *i_regs)
{
int s,tl;
int temp;
emit_jmp(0);
}
}
- emit_addimm_no_flags(ram_offset,temp);
+ if(ram_offset)
+ emit_addimm_no_flags(ram_offset,temp);
if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
assert(0);
if (opcode[i]==0x2A) { // SWL
emit_writeword_indexed(tl,0,temp);
}
- if (opcode[i]==0x2E) { // SWR
+ else if (opcode[i]==0x2E) { // SWR
emit_writebyte_indexed(tl,3,temp);
}
- if (opcode[i]==0x2C) { // SDL
- assert(0);
- }
- if (opcode[i]==0x2D) { // SDR
+ else
assert(0);
- }
done0=out;
emit_jmp(0);
// 1
emit_writebyte_indexed(tl,1,temp);
if(rs2[i]) emit_rorimm(tl,8,tl);
}
- if (opcode[i]==0x2E) { // SWR
+ else if (opcode[i]==0x2E) { // SWR
// Write two lsb into two most significant bytes
emit_writehword_indexed(tl,1,temp);
}
- if (opcode[i]==0x2C) { // SDL
- assert(0);
- }
- if (opcode[i]==0x2D) { // SDR
- assert(0);
- }
done1=out;
emit_jmp(0);
// 2
emit_writehword_indexed(tl,-2,temp);
if(rs2[i]) emit_rorimm(tl,16,tl);
}
- if (opcode[i]==0x2E) { // SWR
+ else if (opcode[i]==0x2E) { // SWR
// Write 3 lsb into three most significant bytes
emit_writebyte_indexed(tl,-1,temp);
if(rs2[i]) emit_rorimm(tl,8,tl);
emit_writehword_indexed(tl,0,temp);
if(rs2[i]) emit_rorimm(tl,24,tl);
}
- if (opcode[i]==0x2C) { // SDL
- assert(0);
- }
- if (opcode[i]==0x2D) { // SDR
- assert(0);
- }
done2=out;
emit_jmp(0);
// 3
emit_writebyte_indexed(tl,-3,temp);
if(rs2[i]) emit_rorimm(tl,8,tl);
}
- if (opcode[i]==0x2E) { // SWR
+ else if (opcode[i]==0x2E) { // SWR
// Write entire word
emit_writeword_indexed(tl,-3,temp);
}
- if (opcode[i]==0x2C) { // SDL
- assert(0);
- }
- if (opcode[i]==0x2D) { // SDR
- assert(0);
- }
set_jump_target(done0, out);
set_jump_target(done1, out);
set_jump_target(done2, out);
- if (opcode[i]==0x2C) { // SDL
- assert(0);
- }
- if (opcode[i]==0x2D) { // SDR
- assert(0);
- }
if(!c||!memtarget)
add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj[i],reglist);
- if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
+ if(!(i_regs->waswritten&(1<<rs1[i])) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
emit_addimm_no_flags(-ram_offset,temp);
#if defined(HOST_IMM8)
int ir=get_reg(i_regs->regmap,INVCP);
emit_storereg(CCREG,HOST_CCREG);
emit_loadreg(rs1[i],1);
emit_movimm(copr,0);
- emit_call(pcsx_mtc0_ds);
+ emit_far_call(pcsx_mtc0_ds);
emit_loadreg(rs1[i],s);
return;
}
emit_movimm(0,HOST_TEMPREG);
emit_writeword(HOST_TEMPREG,&pending_exception);
}
- //else if(copr==12&&is_delayslot) emit_call((int)MTC0_R12);
- //else
if(s==HOST_CCREG)
emit_loadreg(rs1[i],1);
else if(s!=1)
emit_mov(s,1);
emit_movimm(copr,0);
- emit_call(pcsx_mtc0);
+ emit_far_call(pcsx_mtc0);
if(copr==9||copr==11||copr==12||copr==13) {
emit_readword(&Count,HOST_CCREG);
emit_readword(&next_interupt,HOST_TEMPREG);
emit_jeq(0);
emit_readword(&pcaddr, 0);
emit_addimm(HOST_CCREG,2,HOST_CCREG);
- emit_call(get_addr_ht);
+ emit_far_call(get_addr_ht);
emit_jmpreg(0);
set_jump_target(jaddr, out);
}
if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
emit_movimm(start+(i-ds)*4,EAX); // Get PC
emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
- emit_jmp(ds?fp_exception_ds:fp_exception);
+ emit_far_jump(ds?fp_exception_ds:fp_exception);
}
static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
break;
case 28:
case 29:
- emit_readword(®_cop2d[9],temp);
- emit_testimm(temp,0x8000); // do we need this?
- emit_andimm(temp,0xf80,temp);
- emit_andne_imm(temp,0,temp);
- emit_shrimm(temp,7,tl);
- emit_readword(®_cop2d[10],temp);
- emit_testimm(temp,0x8000);
- emit_andimm(temp,0xf80,temp);
- emit_andne_imm(temp,0,temp);
- emit_orrshr_imm(temp,2,tl);
- emit_readword(®_cop2d[11],temp);
- emit_testimm(temp,0x8000);
- emit_andimm(temp,0xf80,temp);
- emit_andne_imm(temp,0,temp);
- emit_orrshl_imm(temp,3,tl);
- emit_writeword(tl,®_cop2d[copr]);
+ c2op_mfc2_29_assemble(tl,temp);
break;
default:
emit_readword(®_cop2d[copr],tl);
emit_writeword(sl,®_cop2d[28]);
break;
case 30:
- emit_movs(sl,temp);
- emit_mvnmi(temp,temp);
+ emit_xorsar_imm(sl,sl,31,temp);
#if defined(HAVE_ARMV5) || defined(__aarch64__)
emit_clz(temp,temp);
#else
assert(ar>=0);
if (opcode[i]==0x3a) { // SWC2
- cop2_get_dreg(copr,tl,HOST_TEMPREG);
+ cop2_get_dreg(copr,tl,-1);
type=STOREW_STUB;
}
else
if(jaddr2)
add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj[i],reglist);
if(opcode[i]==0x3a) // SWC2
- if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
+ if(!(i_regs->waswritten&(1<<rs1[i])) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
#if defined(HOST_IMM8)
int ir=get_reg(i_regs->regmap,INVCP);
assert(ir>=0);
emit_signextend16(sl,temp);
break;
case 31:
- //value = value & 0x7ffff000;
- //if (value & 0x7f87e000) value |= 0x80000000;
- emit_shrimm(sl,12,temp);
- emit_shlimm(temp,12,temp);
- emit_testimm(temp,0x7f000000);
- emit_testeqimm(temp,0x00870000);
- emit_testeqimm(temp,0x0000e000);
- emit_orrne_imm(temp,0x80000000,temp);
+ c2op_ctc2_31_assemble(sl,temp);
break;
default:
temp=sl;
}
}
+static void do_unalignedwritestub(int n)
+{
+ assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
+ literal_pool(256);
+ set_jump_target(stubs[n].addr, out);
+
+ int i=stubs[n].a;
+ struct regstat *i_regs=(struct regstat *)stubs[n].c;
+ int addr=stubs[n].b;
+ u_int reglist=stubs[n].e;
+ signed char *i_regmap=i_regs->regmap;
+ int temp2=get_reg(i_regmap,FTEMP);
+ int rt;
+ rt=get_reg(i_regmap,rs2[i]);
+ assert(rt>=0);
+ assert(addr>=0);
+ assert(opcode[i]==0x2a||opcode[i]==0x2e); // SWL/SWR only implemented
+ reglist|=(1<<addr);
+ reglist&=~(1<<temp2);
+
+#if 1
+ // don't bother with it and call write handler
+ save_regs(reglist);
+ pass_args(addr,rt);
+ int cc=get_reg(i_regmap,CCREG);
+ if(cc<0)
+ emit_loadreg(CCREG,2);
+ emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n].d+1),2);
+ emit_far_call((opcode[i]==0x2a?jump_handle_swl:jump_handle_swr));
+ emit_addimm(0,-CLOCK_ADJUST((int)stubs[n].d+1),cc<0?2:cc);
+ if(cc<0)
+ emit_storereg(CCREG,2);
+ restore_regs(reglist);
+ emit_jmp(stubs[n].retaddr); // return address
+#else
+ emit_andimm(addr,0xfffffffc,temp2);
+ emit_writeword(temp2,&address);
+
+ save_regs(reglist);
+ emit_shrimm(addr,16,1);
+ int cc=get_reg(i_regmap,CCREG);
+ if(cc<0) {
+ emit_loadreg(CCREG,2);
+ }
+ emit_movimm((u_int)readmem,0);
+ emit_addimm(cc<0?2:cc,2*stubs[n].d+2,2);
+ emit_call((int)&indirect_jump_indexed);
+ restore_regs(reglist);
+
+ emit_readword(&readmem_dword,temp2);
+ int temp=addr; //hmh
+ emit_shlimm(addr,3,temp);
+ emit_andimm(temp,24,temp);
+ if (opcode[i]==0x2a) // SWL
+ emit_xorimm(temp,24,temp);
+ emit_movimm(-1,HOST_TEMPREG);
+ if (opcode[i]==0x2a) { // SWL
+ emit_bic_lsr(temp2,HOST_TEMPREG,temp,temp2);
+ emit_orrshr(rt,temp,temp2);
+ }else{
+ emit_bic_lsl(temp2,HOST_TEMPREG,temp,temp2);
+ emit_orrshl(rt,temp,temp2);
+ }
+ emit_readword(&address,addr);
+ emit_writeword(temp2,&word);
+ //save_regs(reglist); // don't need to, no state changes
+ emit_shrimm(addr,16,1);
+ emit_movimm((u_int)writemem,0);
+ //emit_call((int)&indirect_jump_indexed);
+ emit_mov(15,14);
+ emit_readword_dualindexedx4(0,1,15);
+ emit_readword(&Count,HOST_TEMPREG);
+ emit_readword(&next_interupt,2);
+ emit_addimm(HOST_TEMPREG,-2*stubs[n].d-2,HOST_TEMPREG);
+ emit_writeword(2,&last_count);
+ emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc);
+ if(cc<0) {
+ emit_storereg(CCREG,HOST_TEMPREG);
+ }
+ restore_regs(reglist);
+ emit_jmp(stubs[n].retaddr); // return address
+#endif
+}
+
#ifndef multdiv_assemble
void multdiv_assemble(int i,struct regstat *i_regs)
{
}
}
-static void syscall_assemble(int i,struct regstat *i_regs)
+// call interpreter, exception handler, things that change pc/regs/cycles ...
+static void call_c_cpu_handler(int i, const struct regstat *i_regs, u_int pc, void *func)
{
signed char ccreg=get_reg(i_regs->regmap,CCREG);
assert(ccreg==HOST_CCREG);
assert(!is_delayslot);
(void)ccreg;
- emit_movimm(start+i*4,EAX); // Get PC
- emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
- emit_jmp(jump_syscall_hle); // XXX
+
+ emit_movimm(pc,3); // Get PC
+ emit_readword(&last_count,2);
+ emit_writeword(3,&psxRegs.pc);
+ emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
+ emit_add(2,HOST_CCREG,2);
+ emit_writeword(2,&psxRegs.cycle);
+ emit_far_call(func);
+ emit_far_jump(jump_to_new_pc);
+}
+
+static void syscall_assemble(int i,struct regstat *i_regs)
+{
+ emit_movimm(0x20,0); // cause code
+ emit_movimm(0,1); // not in delay slot
+ call_c_cpu_handler(i,i_regs,start+i*4,psxException);
}
static void hlecall_assemble(int i,struct regstat *i_regs)
{
- extern void psxNULL();
- signed char ccreg=get_reg(i_regs->regmap,CCREG);
- assert(ccreg==HOST_CCREG);
- assert(!is_delayslot);
- (void)ccreg;
- emit_movimm(start+i*4+4,0); // Get PC
+ void *hlefunc = psxNULL;
uint32_t hleCode = source[i] & 0x03ffffff;
- if (hleCode >= ARRAY_SIZE(psxHLEt))
- emit_movimm((uintptr_t)psxNULL,1);
- else
- emit_movimm((uintptr_t)psxHLEt[hleCode],1);
- emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
- emit_jmp(jump_hlecall);
+ if (hleCode < ARRAY_SIZE(psxHLEt))
+ hlefunc = psxHLEt[hleCode];
+
+ call_c_cpu_handler(i,i_regs,start+i*4+4,hlefunc);
}
static void intcall_assemble(int i,struct regstat *i_regs)
{
- signed char ccreg=get_reg(i_regs->regmap,CCREG);
- assert(ccreg==HOST_CCREG);
- assert(!is_delayslot);
- (void)ccreg;
- emit_movimm(start+i*4,0); // Get PC
- emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
- emit_jmp(jump_intcall);
+ call_c_cpu_handler(i,i_regs,start+i*4,execI);
}
static void speculate_mov(int rs,int rt)
static void drc_dbg_emit_do_cmp(int i)
{
extern void do_insn_cmp();
- extern int cycle;
+ //extern int cycle;
u_int hr,reglist=0;
- for(hr=0;hr<HOST_REGS;hr++)
+ 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]) {
+ for (hr = 0; hr < HOST_REGS; hr++) {
+ int reg = regs[i-1].regmap[hr];
+ if (hr == EXCLUDE_REG || reg < 0)
+ continue;
+ if (!((regs[i-1].isconst >> hr) & 1))
+ continue;
+ if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
+ continue;
+ emit_movimm(constmap[i-1][hr],0);
+ emit_storereg(reg, 0);
+ }
+ }
emit_movimm(start+i*4,0);
emit_writeword(0,&pcaddr);
- emit_call(do_insn_cmp);
+ emit_far_call(do_insn_cmp);
//emit_readword(&cycle,0);
//emit_addimm(0,2,0);
//emit_writeword(0,&cycle);
+ (void)get_reg2;
restore_regs(reglist);
+ assem_debug("\\\\do_insn_cmp\n");
}
#else
#define drc_dbg_emit_do_cmp(x)
else if(*adj==0||invert) {
int cycles=CLOCK_ADJUST(count+2);
// faster loop HACK
+#if 0
if (t&&*adj) {
int rel=t-i;
if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
}
+#endif
emit_addimm_and_set_flags(cycles,HOST_CCREG);
jaddr=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].a) emit_addimm(HOST_CCREG,CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
- emit_call(cc_interrupt);
+ emit_far_call(cc_interrupt);
if(stubs[n].a) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
if(stubs[n].d==TAKEN) {
if(internal_branch(ba[i]))
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(i>(ba[i]-start)>>2) invert=1;
#endif
+ #ifdef __aarch64__
+ invert=1; // because of near cond. branches
+ #endif
if(ooo[i]) {
s1l=get_reg(branch_regs[i].regmap,rs1[i]);
#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
if(i>(ba[i]-start)>>2) invert=1;
#endif
+ #ifdef __aarch64__
+ invert=1; // because of near cond. branches
+ #endif
//if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
//assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
uint64_t u,gte_u,b,gte_b;
uint64_t temp_u,temp_gte_u=0;
uint64_t gte_u_unknown=0;
- if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
+ if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
gte_u_unknown=~0ll;
if(iend==slen-1) {
u=1;
((volatile u_int *)out)[0]++; // make cache dirty
for (i = 0; i < ARRAY_SIZE(ret); i++) {
- out = translation_cache;
+ out = ndrc->translation_cache;
beginning = start_block();
emit_movimm(DRC_TEST_VAL + i, 0); // test
emit_ret();
SysPrintf("test passed.\n");
else
SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
- out = translation_cache;
+ out = ndrc->translation_cache;
}
// clear the state completely, instead of just marking
// things invalid like invalidate_all_pages() does
-void new_dynarec_clear_full()
+void new_dynarec_clear_full(void)
{
int n;
- out = translation_cache;
+ out = ndrc->translation_cache;
memset(invalid_code,1,sizeof(invalid_code));
memset(hash_table,0xff,sizeof(hash_table));
memset(mini_ht,-1,sizeof(mini_ht));
for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
}
-void new_dynarec_init()
+void new_dynarec_init(void)
{
SysPrintf("Init new dynarec\n");
- // allocate/prepare a buffer for translation cache
- // see assem_arm.h for some explanation
-#if defined(BASE_ADDR_FIXED)
- if (mmap(translation_cache, 1 << TARGET_SIZE_2,
- PROT_READ | PROT_WRITE | PROT_EXEC,
- MAP_PRIVATE | MAP_ANONYMOUS,
- -1, 0) != translation_cache) {
- SysPrintf("mmap() failed: %s\n", strerror(errno));
- SysPrintf("disable BASE_ADDR_FIXED and recompile\n");
- abort();
- }
-#elif defined(BASE_ADDR_DYNAMIC)
+#ifdef BASE_ADDR_DYNAMIC
#ifdef VITA
sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
if (sceBlock < 0)
SysPrintf("sceKernelAllocMemBlockForVM failed\n");
- int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&translation_cache);
+ int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
if (ret < 0)
SysPrintf("sceKernelGetMemBlockBase failed\n");
#else
- translation_cache = mmap (NULL, 1 << TARGET_SIZE_2,
+ uintptr_t desired_addr = 0;
+ #ifdef __ELF__
+ extern char _end;
+ desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
+ #endif
+ ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
- if (translation_cache == MAP_FAILED) {
+ if (ndrc == MAP_FAILED) {
SysPrintf("mmap() failed: %s\n", strerror(errno));
abort();
}
#else
#ifndef NO_WRITE_EXEC
// not all systems allow execute in data segment by default
- if (mprotect(translation_cache, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
+ if (mprotect(ndrc, sizeof(ndrc->translation_cache) + sizeof(ndrc->tramp.ops),
+ PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
SysPrintf("mprotect() failed: %s\n", strerror(errno));
#endif
#endif
- out = translation_cache;
+ out = ndrc->translation_cache;
cycle_multiplier=200;
new_dynarec_clear_full();
#ifdef HOST_IMM8
SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
}
-void new_dynarec_cleanup()
+void new_dynarec_cleanup(void)
{
int n;
-#if defined(BASE_ADDR_FIXED) || defined(BASE_ADDR_DYNAMIC)
+#ifdef BASE_ADDR_DYNAMIC
#ifdef VITA
sceKernelFreeMemBlock(sceBlock);
sceBlock = -1;
#else
- if (munmap(translation_cache, 1<<TARGET_SIZE_2) < 0)
+ if (munmap(ndrc, sizeof(*ndrc)) < 0)
SysPrintf("munmap() failed\n");
#endif
#endif
static u_int *get_source_start(u_int addr, u_int *limit)
{
+ if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
+ cycle_multiplier_override = 0;
+
if (addr < 0x00200000 ||
- (0xa0000000 <= addr && addr < 0xa0200000)) {
+ (0xa0000000 <= addr && addr < 0xa0200000))
+ {
// used for BIOS calls mostly?
*limit = (addr&0xa0000000)|0x00200000;
return (u_int *)(rdram + (addr&0x1fffff));
}
else if (!Config.HLE && (
/* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
- (0xbfc00000 <= addr && addr < 0xbfc80000))) {
- // BIOS
+ (0xbfc00000 <= addr && addr < 0xbfc80000)))
+ {
+ // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
+ // but timings in PCSX are too tied to the interpreter's BIAS
+ if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
+ cycle_multiplier_override = 200;
+
*limit = (addr & 0xfff00000) | 0x80000;
return (u_int *)((u_char *)psxR + (addr&0x7ffff));
}
memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
}
-int new_recompile_block(int addr)
+int new_recompile_block(u_int addr)
{
u_int pagelimit = 0;
u_int state_rflags = 0;
invalid_code[start>>12]=0;
emit_movimm(start,0);
emit_writeword(0,&pcaddr);
- emit_jmp(new_dyna_leave);
+ emit_far_jump(new_dyna_leave);
literal_pool(0);
end_block(beginning);
ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
dirty_reg(&branch_regs[i-1],31);
}
memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
- memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+ memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
break;
case RJUMP:
memcpy(&branch_regs[i-1],¤t,sizeof(current));
}
#endif
memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
- memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+ memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
break;
case CJUMP:
if((opcode[i-1]&0x3E)==4) // BEQ/BNE
branch_regs[i-1].isconst=0;
branch_regs[i-1].wasconst=0;
memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
- memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+ memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
}
else
if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
branch_regs[i-1].isconst=0;
branch_regs[i-1].wasconst=0;
memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
- memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+ memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
}
else
// Alloc the delay slot in case the branch is taken
branch_regs[i-1].isconst=0;
branch_regs[i-1].wasconst=0;
memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
- memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
+ memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
}
else
// Alloc the delay slot in case the branch is taken
if(!is_ds[i]) {
regs[i].dirty=current.dirty;
regs[i].isconst=current.isconst;
- memcpy(constmap[i],current_constmap,sizeof(current_constmap));
+ memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
}
for(hr=0;hr<HOST_REGS;hr++) {
if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
{
if(i<slen-1&&!is_ds[i]) {
assert(regs[i].regmap[hr]<64);
- if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]!=-1)
+ if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
{
SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
}
}
}
- }
- }
+ } // if needed
+ } // for hr
}
/* Pass 5 - Pre-allocate registers */
void *instr_addr0_override = NULL;
if (start == 0x80030000) {
- // nasty hack for fastbios thing
+ // nasty hack for the fastbios thing
// override block entry to this code
instr_addr0_override = out;
emit_movimm(start,0);
emit_writeword(0,&pcaddr);
emit_writeword(0,&address);
emit_cmp(0,1);
+ #ifdef __aarch64__
+ emit_jeq(out + 4*2);
+ emit_far_jump(new_dyna_leave);
+ #else
emit_jne(new_dyna_leave);
+ #endif
}
for(i=0;i<slen;i++)
{
// If we're within 256K of the end of the buffer,
// start over from the beginning. (Is 256K enough?)
- if (out > translation_cache+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE)
- out = translation_cache;
+ if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
+ out = ndrc->translation_cache;
// Trap writes to any of the pages we compiled
for(i=start>>12;i<=(start+slen*4)>>12;i++) {
/* Pass 10 - Free memory by expiring oldest blocks */
- int end=(((out-translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
+ int end=(((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
while(expirep!=end)
{
int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
- uintptr_t base=(uintptr_t)translation_cache+((expirep>>13)<<shift); // Base address of this block
+ uintptr_t base=(uintptr_t)ndrc->translation_cache+((expirep>>13)<<shift); // Base address of this block
inv_debug("EXP: Phase %d\n",expirep);
switch((expirep>>11)&3)
{
break;
case 3:
// Clear jump_out
- #if defined(__arm__) || defined(__aarch64__)
if((expirep&2047)==0)
do_clear_cache();
- #endif
ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
break;