#ifdef _3DS
#include <3ds_utils.h>
#endif
-#ifdef VITA
-#include <psp2/kernel/sysmem.h>
-static int sceBlock;
-#endif
#include "new_dynarec_config.h"
#include "../psxhle.h"
//#define DISASM
//#define ASSEM_PRINT
+//#define REG_ALLOC_PRINT
#ifdef ASSEM_PRINT
#define assem_debug printf
#define MAXBLOCK 4096
#define MAX_OUTPUT_BLOCK_SIZE 262144
+#ifdef VITA
+// apparently Vita has a 16MB limit, so either we cut tc in half,
+// or use this hack (it's a hack because tc size was designed to be power-of-2)
+#define TC_REDUCE_BYTES 4096
+#else
+#define TC_REDUCE_BYTES 0
+#endif
+
struct ndrc_mem
{
- u_char translation_cache[1 << TARGET_SIZE_2];
+ u_char translation_cache[(1 << TARGET_SIZE_2) - TC_REDUCE_BYTES];
struct
{
struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
INVCODE_STUB = 14,
};
+// regmap_pre[i] - regs before [i] insn starts; dirty things here that
+// don't match .regmap will be written back
+// [i].regmap_entry - regs that must be set up if someone jumps here
+// [i].regmap - regs [i] insn will read/(over)write
+// branch_regs[i].* - same as above but for branches, takes delay slot into account
struct regstat
{
- signed char regmap_entry[HOST_REGS]; // pre-insn + loop preloaded regs?
+ signed char regmap_entry[HOST_REGS];
signed char regmap[HOST_REGS];
uint64_t wasdirty;
uint64_t dirty;
uint64_t u;
- u_int wasconst;
- u_int isconst;
+ u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
+ u_int isconst; // ... but isconst is false when r2 is known
u_int loadedconst; // host regs that have constants loaded
u_int waswritten; // MIPS regs that were used as store base before
};
static u_int ba[MAXBLOCK];
static uint64_t unneeded_reg[MAXBLOCK];
static uint64_t branch_unneeded_reg[MAXBLOCK];
- // pre-instruction [i], excluding loop-preload regs?
+ // see 'struct regstat' for a description
static signed char regmap_pre[MAXBLOCK][HOST_REGS];
// 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 void *copy;
static int expirep;
static u_int stop_after_jal;
- static u_int f1_hack; // 0 - off, ~0 - capture address, else addr
+ static u_int f1_hack;
int new_dynarec_hacks;
int new_dynarec_hacks_pergame;
//#define FLOAT 19 // Floating point unit
//#define FCONV 20 // Convert integer to float
//#define FCOMP 21 // Floating point compare (sets FSREG)
-#define SYSCALL 22// SYSCALL
+#define SYSCALL 22// SYSCALL,BREAK
#define OTHER 23 // Other
#define SPAN 24 // Branch/delay slot spans 2 pages
#define NI 25 // Not implemented
void cc_interrupt();
void fp_exception();
void fp_exception_ds();
+void jump_syscall (u_int u0, u_int u1, u_int pc);
+void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
+void jump_break (u_int u0, u_int u1, u_int pc);
+void jump_break_ds(u_int u0, u_int u1, u_int pc);
void jump_to_new_pc();
void call_gteStall();
void new_dyna_leave();
static void emit_far_jump(const void *f);
static void emit_far_call(const void *f);
+#ifdef VITA
+#include <psp2/kernel/sysmem.h>
+static int sceBlock;
+// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
+extern int getVMBlock();
+int _newlib_vm_size_user = sizeof(*ndrc);
+#endif
+
static void mprotect_w_x(void *start, void *end, int is_x)
{
#ifdef NO_WRITE_EXEC
int cycle_multiplier = CYCLE_MULT_DEFAULT; // 100 for 1.0
int cycle_multiplier_override;
int cycle_multiplier_old;
+static int cycle_multiplier_active;
static int CLOCK_ADJUST(int x)
{
- int m = cycle_multiplier_override && cycle_multiplier == CYCLE_MULT_DEFAULT
- ? cycle_multiplier_override : cycle_multiplier;
- int s=(x>>31)|1;
+ int m = cycle_multiplier_active;
+ int s = (x >> 31) | 1;
return (x * m + s * 50) / 100;
}
return get_addr(vaddr);
}
-void clear_all_regs(signed char regmap[])
+static void clear_all_regs(signed char regmap[])
{
- int hr;
- for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
+ memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
}
static signed char get_reg(const signed char regmap[],int r)
FUNCNAME(jump_handler_write32),
FUNCNAME(invalidate_addr),
FUNCNAME(jump_to_new_pc),
+ FUNCNAME(jump_break),
+ FUNCNAME(jump_break_ds),
+ FUNCNAME(jump_syscall),
+ FUNCNAME(jump_syscall_ds),
FUNCNAME(call_gteStall),
FUNCNAME(new_dyna_leave),
FUNCNAME(pcsx_mtc0),
static void syscall_assemble(int i, const struct regstat *i_regs, int ccadj_)
{
- emit_movimm(0x20,0); // cause code
- emit_movimm(0,1); // not in delay slot
- call_c_cpu_handler(i, i_regs, ccadj_, start+i*4, psxException);
+ // 'break' tends to be littered around to catch things like
+ // division by 0 and is almost never executed, so don't emit much code here
+ void *func = (dops[i].opcode2 == 0x0C)
+ ? (is_delayslot ? jump_syscall_ds : jump_syscall)
+ : (is_delayslot ? jump_break_ds : jump_break);
+ assert(get_reg(i_regs->regmap, CCREG) == HOST_CCREG);
+ emit_movimm(start + i*4, 2); // pc
+ emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
+ emit_far_jump(func);
}
static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
case SPAN:
pagespan_assemble(i, i_regs);
break;
+ case NOP:
case OTHER:
case NI:
// not handled, just skip
int cc;
int match;
match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
- assem_debug("smatch=%d\n",match);
+ assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
int s1l;
int unconditional=0,nevertaken=0;
int invert=0;
wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
address_generation(i+1,&branch_regs[i],0);
- load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
+ if (ram_offset)
+ load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
+ load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
ds_assemble(i+1,&branch_regs[i]);
cc=get_reg(branch_regs[i].regmap,CCREG);
if (cc == -1) {
load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
}
+static void check_regmap(signed char *regmap)
+{
+#ifndef NDEBUG
+ int i,j;
+ for (i = 0; i < HOST_REGS; i++) {
+ if (regmap[i] < 0)
+ continue;
+ for (j = i + 1; j < HOST_REGS; j++)
+ assert(regmap[i] != regmap[j]);
+ }
+#endif
+}
+
// Basic liveness analysis for MIPS registers
-void unneeded_registers(int istart,int iend,int r)
+static void unneeded_registers(int istart,int iend,int r)
{
int i;
uint64_t u,gte_u,b,gte_b;
}
#ifdef DISASM
+#include <inttypes.h>
+void print_regmap(const char *name, const signed char *regmap)
+{
+ char buf[5];
+ int i, l;
+ fputs(name, stdout);
+ for (i = 0; i < HOST_REGS; i++) {
+ l = 0;
+ if (regmap[i] >= 0)
+ l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
+ for (; l < 3; l++)
+ buf[l] = ' ';
+ buf[l] = 0;
+ printf(" r%d=%s", i, buf);
+ }
+ fputs("\n", stdout);
+}
+
/* disassembly */
void disassemble_inst(int i)
{
//printf (" %s %8x\n",insn[i],source[i]);
printf (" %x: %s\n",start+i*4,insn[i]);
}
+ return;
+ printf("D: %"PRIu64" WD: %"PRIu64" U: %"PRIu64"\n",
+ regs[i].dirty, regs[i].wasdirty, unneeded_reg[i]);
+ print_regmap("pre: ", regmap_pre[i]);
+ print_regmap("entry: ", regs[i].regmap_entry);
+ print_regmap("map: ", regs[i].regmap);
+ if (dops[i].is_jump) {
+ print_regmap("bentry:", branch_regs[i].regmap_entry);
+ print_regmap("bmap: ", branch_regs[i].regmap);
+ }
}
#else
static void disassemble_inst(int i) {}
SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
}
- SysPrintf("testing if we can run recompiled code...\n");
+ SysPrintf("testing if we can run recompiled code @%p...\n", out);
((volatile u_int *)out)[0]++; // make cache dirty
for (i = 0; i < ARRAY_SIZE(ret); i++) {
literalcount=0;
stop_after_jal=0;
inv_code_start=inv_code_end=~0;
+ hack_addr=0;
f1_hack=0;
// TLB
for(n=0;n<4096;n++) ll_clear(jump_in+n);
void new_dynarec_init(void)
{
- SysPrintf("Init new dynarec\n");
+ SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
+#ifdef _3DS
+ check_rosalina();
+#endif
#ifdef BASE_ADDR_DYNAMIC
#ifdef VITA
- sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
- if (sceBlock < 0)
- SysPrintf("sceKernelAllocMemBlockForVM failed\n");
+ sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
+ if (sceBlock <= 0)
+ SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
if (ret < 0)
- SysPrintf("sceKernelGetMemBlockBase failed\n");
+ SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
+ sceKernelOpenVMDomain();
+ sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
+ #elif defined(_MSC_VER)
+ ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
+ PAGE_EXECUTE_READWRITE);
#else
uintptr_t desired_addr = 0;
#ifdef __ELF__
#else
#ifndef NO_WRITE_EXEC
// not all systems allow execute in data segment by default
- if (mprotect(ndrc, sizeof(ndrc->translation_cache) + sizeof(ndrc->tramp.ops),
+ // size must be 4K aligned for 3DS?
+ if (mprotect(ndrc, sizeof(*ndrc),
PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
SysPrintf("mprotect() failed: %s\n", strerror(errno));
#endif
int n;
#ifdef BASE_ADDR_DYNAMIC
#ifdef VITA
- sceKernelFreeMemBlock(sceBlock);
- sceBlock = -1;
+ // sceBlock is managed by retroarch's bootstrap code
+ //sceKernelFreeMemBlock(sceBlock);
+ //sceBlock = -1;
#else
if (munmap(ndrc, sizeof(*ndrc)) < 0)
SysPrintf("munmap() failed\n");
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))
{
// 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;
+ cycle_multiplier_active = 200;
*limit = (addr & 0xfff00000) | 0x80000;
return (u_int *)((u_char *)psxR + (addr&0x7ffff));
memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
}
+static int apply_hacks(void)
+{
+ int i;
+ if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
+ return 0;
+ /* special hack(s) */
+ for (i = 0; i < slen - 4; i++)
+ {
+ // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
+ if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
+ && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
+ && imm[i+3] == 0x6e40 && dops[i+3].rs1 == 2)
+ {
+ SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
+ dops[i + 3].itype = NOP;
+ }
+ }
+ i = slen;
+ if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
+ && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
+ && dops[i-7].itype == STORE)
+ {
+ i = i-8;
+ if (dops[i].itype == IMM16)
+ i--;
+ // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
+ if (dops[i].itype == STORELR && dops[i].rs1 == 6
+ && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
+ {
+ SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
+ f1_hack = 1;
+ return 1;
+ }
+ }
+ return 0;
+}
+
int new_recompile_block(u_int addr)
{
u_int pagelimit = 0;
ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
return 0;
}
- else if (f1_hack == ~0u || (f1_hack != 0 && start == f1_hack)) {
+ else if (f1_hack && hack_addr == 0) {
void *beginning = start_block();
u_int page = get_page(start);
+ emit_movimm(start, 0);
+ emit_writeword(0, &hack_addr);
emit_readword(&psxRegs.GPR.n.sp, 0);
emit_readptr(&mem_rtab, 1);
emit_shrimm(0, 12, 2);
ll_add_flags(jump_in + page, start, state_rflags, beginning);
SysPrintf("F1 hack to %08x\n", start);
- f1_hack = start;
return 0;
}
+ cycle_multiplier_active = cycle_multiplier_override && cycle_multiplier == CYCLE_MULT_DEFAULT
+ ? cycle_multiplier_override : cycle_multiplier;
+
source = get_source_start(start, &pagelimit);
if (source == NULL) {
SysPrintf("Compile at bogus memory address: %08x\n", addr);
/* Pass 1 disassembly */
- for(i=0;!done;i++) {
- dops[i].bt=0;
- dops[i].ooo=0;
+ for (i = 0; !done; i++)
+ {
+ memset(&dops[i], 0, sizeof(dops[i]));
op2=0;
minimum_free_regs[i]=0;
dops[i].opcode=op=source[i]>>26;
case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
- case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
+ case 0x0D: strcpy(insn[i],"BREAK"); type=SYSCALL; break;
case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
do_in_intrp=1;
}
}
- if(do_in_intrp) {
- dops[i-1].rs1=CCREG;
- dops[i-1].rs2=dops[i-1].rt1=dops[i-1].rt2=0;
- ba[i-1]=-1;
- dops[i-1].itype=INTCALL;
- done=2;
+ if (do_in_intrp) {
+ memset(&dops[i-1], 0, sizeof(dops[i-1]));
+ dops[i-1].itype = INTCALL;
+ dops[i-1].rs1 = CCREG;
+ ba[i-1] = -1;
+ done = 2;
i--; // don't compile the DS
}
}
// Don't get too close to the limit
if(i>MAXBLOCK/2) done=1;
}
- if(dops[i].itype==SYSCALL&&stop_after_jal) done=1;
- if(dops[i].itype==HLECALL||dops[i].itype==INTCALL) done=2;
- if(done==2) {
+ if (dops[i].itype == SYSCALL || dops[i].itype == HLECALL || dops[i].itype == INTCALL)
+ done = stop_after_jal ? 1 : 2;
+ if (done == 2) {
// Does the block continue due to a branch?
for(j=i-1;j>=0;j--)
{
}
assert(slen>0);
- /* spacial hack(s) */
- if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
- && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
- && dops[i-7].itype == STORE)
- {
- i = i-8;
- if (dops[i].itype == IMM16)
- i--;
- // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
- if (dops[i].itype == STORELR && dops[i].rs1 == 6
- && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
- {
- SysPrintf("F1 hack from %08x\n", start);
- if (f1_hack == 0)
- f1_hack = ~0u;
- }
- }
+ int clear_hack_addr = apply_hacks();
/* Pass 2 - Register dependencies and branch targets */
/* Pass 3 - Register allocation */
struct regstat current; // Current register allocations/status
- current.dirty=0;
- current.u=unneeded_reg[0];
+ clear_all_regs(current.regmap_entry);
clear_all_regs(current.regmap);
- alloc_reg(¤t,0,CCREG);
- dirty_reg(¤t,CCREG);
- current.isconst=0;
- current.wasconst=0;
- current.waswritten=0;
+ current.wasdirty = current.dirty = 0;
+ current.u = unneeded_reg[0];
+ alloc_reg(¤t, 0, CCREG);
+ dirty_reg(¤t, CCREG);
+ current.wasconst = 0;
+ current.isconst = 0;
+ current.loadedconst = 0;
+ current.waswritten = 0;
int ds=0;
int cc=0;
int hr=-1;
memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
regs[i].wasconst=current.isconst;
regs[i].wasdirty=current.dirty;
+ regs[i].dirty=0;
+ regs[i].u=0;
+ regs[i].isconst=0;
regs[i].loadedconst=0;
if (!dops[i].is_jump) {
if(i+1<slen) {
current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
current.u|=1;
- } else { SysPrintf("oops, branch at end of block with no delay slot\n");abort(); }
+ } else {
+ SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
+ abort();
+ }
}
dops[i].is_ds=ds;
if(ds) {
{
regs[i].regmap[hr]=-1;
regs[i].isconst&=~(1<<hr);
+ regs[i].dirty&=~(1<<hr);
+ regs[i+1].wasdirty&=~(1<<hr);
if((branch_regs[i].regmap[hr]&63)!=dops[i].rs1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 &&
(branch_regs[i].regmap[hr]&63)!=dops[i].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rt2 &&
(branch_regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
}
regs[i].regmap[hr]=-1;
regs[i].isconst&=~(1<<hr);
+ regs[i].dirty&=~(1<<hr);
+ regs[i+1].wasdirty&=~(1<<hr);
}
}
}
//printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
int k;
if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
+ if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
- if(r>63) {
- if(get_reg(regs[i].regmap,r&63)<0) break;
- if(get_reg(branch_regs[i].regmap,r&63)<0) break;
- }
k=i;
while(k>1&®s[k-1].regmap[hr]==-1) {
if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
break;
}
- assert(r < 64);
k--;
}
if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
- else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
+ else {
+ regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
+ regs[i+1].isconst&=~(1<<hr);
+ }
assert(hr>=0);
if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
{
dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
}
-#ifdef DISASM
+#ifdef REG_ALLOC_PRINT
/* Debug/disassembly */
for(i=0;i<slen;i++)
{
#endif
}
}
-#endif // DISASM
+#endif // REG_ALLOC_PRINT
/* Pass 8 - Assembly */
linkcount=0;stubcount=0;
}
for(i=0;i<slen;i++)
{
+ check_regmap(regmap_pre[i]);
+ check_regmap(regs[i].regmap_entry);
+ check_regmap(regs[i].regmap);
//if(ds) printf("ds: ");
disassemble_inst(i);
if(ds) {
instr_addr[i] = out;
assem_debug("<->\n");
drc_dbg_emit_do_cmp(i, ccadj[i]);
+ if (clear_hack_addr) {
+ emit_movimm(0, 0);
+ emit_writeword(0, &hack_addr);
+ clear_hack_addr = 0;
+ }
// load regs
if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
notaz.gp2x.de / pcsx_rearmed.git / blobdiff