X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=6d7069d9c8d2e837294cd5484bc3706fff789c3b;hp=bfe3961e9a9ed7ca7b352d17b5816c1be828c3b2;hb=a151a8d8331cf743eabeab23ce52e9b7726239e5;hpb=1fedf1ea555e4a6be68dd0ba384909ac21da65d0

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index bfe3961e..6d7069d9 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -23,7 +23,19 @@
 #include <assert.h>
 #include <errno.h>
 #include <sys/mman.h>
+#ifdef __MACH__
+#include <libkern/OSCacheControl.h>
+#endif
+#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" //emulator interface
 #include "emu_if.h" //emulator interface
 
 //#define DISASM
@@ -42,19 +54,6 @@
 #include "assem_arm.h"
 #endif
 
-#ifdef __BLACKBERRY_QNX__
-#undef __clear_cache
-#define __clear_cache(start,end) msync(start, (size_t)((void*)end - (void*)start), MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
-#elif defined(__MACH__)
-#include <libkern/OSCacheControl.h>
-#define __clear_cache mach_clear_cache
-static void __clear_cache(void *start, void *end) {
-  size_t len = (char *)end - (char *)start;
-  sys_dcache_flush(start, len);
-  sys_icache_invalidate(start, len);
-}
-#endif
-
 #define MAXBLOCK 4096
 #define MAX_OUTPUT_BLOCK_SIZE 262144
 
@@ -271,6 +270,66 @@ static void add_to_linker(int addr,int target,int ext);
 
 static int tracedebug=0;
 
+static void mprotect_w_x(void *start, void *end, int is_x)
+{
+#ifdef NO_WRITE_EXEC
+  #if defined(VITA)
+  // *Open* enables write on all memory that was
+  // allocated by sceKernelAllocMemBlockForVM()?
+  if (is_x)
+    sceKernelCloseVMDomain();
+  else
+    sceKernelOpenVMDomain();
+  #else
+  u_long mstart = (u_long)start & ~4095ul;
+  u_long mend = (u_long)end;
+  if (mprotect((void *)mstart, mend - mstart,
+               PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
+    SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
+  #endif
+#endif
+}
+
+static void start_tcache_write(void *start, void *end)
+{
+  mprotect_w_x(start, end, 0);
+}
+
+static void end_tcache_write(void *start, void *end)
+{
+#ifdef __arm__
+  size_t len = (char *)end - (char *)start;
+  #if   defined(__BLACKBERRY_QNX__)
+  msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
+  #elif defined(__MACH__)
+  sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
+  #elif defined(VITA)
+  sceKernelSyncVMDomain(sceBlock, start, len);
+  #elif defined(_3DS)
+  ctr_flush_invalidate_cache();
+  #else
+  __clear_cache(start, end);
+  #endif
+  (void)len;
+#endif
+
+  mprotect_w_x(start, end, 1);
+}
+
+static void *start_block(void)
+{
+  u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
+  if (end > (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2))
+    end = (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2);
+  start_tcache_write(out, end);
+  return out;
+}
+
+static void end_block(void *start)
+{
+  end_tcache_write(start, out);
+}
+
 //#define DEBUG_CYCLE_COUNT 1
 
 #define NO_CYCLE_PENALTY_THR 12
@@ -829,7 +888,7 @@ void ll_clear(struct ll_entry **head)
 }
 
 // Dereference the pointers and remove if it matches
-void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
+static void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
 {
   while(head) {
     int ptr=get_pointer(head->addr);
@@ -838,10 +897,11 @@ void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
        (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
     {
       inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
-      u_int host_addr=(u_int)kill_pointer(head->addr);
+      void *host_addr=find_extjump_insn(head->addr);
       #ifdef __arm__
-        needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
+        mark_clear_cache(host_addr);
       #endif
+      set_jump_target((int)host_addr,(int)head->addr);
     }
     head=head->next;
   }
@@ -865,10 +925,11 @@ void invalidate_page(u_int page)
   jump_out[page]=0;
   while(head!=NULL) {
     inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
-    u_int host_addr=(u_int)kill_pointer(head->addr);
+    void *host_addr=find_extjump_insn(head->addr);
     #ifdef __arm__
-      needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
+      mark_clear_cache(host_addr);
     #endif
+    set_jump_target((int)host_addr,(int)head->addr);
     next=head->next;
     free(head);
     head=next;
@@ -3380,12 +3441,17 @@ void syscall_assemble(int i,struct regstat *i_regs)
 
 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
-  emit_movimm((int)psxHLEt[source[i]&7],1);
+  uint32_t hleCode = source[i] & 0x03ffffff;
+  if (hleCode >= (sizeof(psxHLEt) / sizeof(psxHLEt[0])))
+    emit_movimm((int)psxNULL,1);
+  else
+    emit_movimm((int)psxHLEt[hleCode],1);
   emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
   emit_jmp((int)jump_hlecall);
 }
@@ -4197,6 +4263,28 @@ int match_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
   return 1;
 }
 
+#ifdef DRC_DBG
+static void drc_dbg_emit_do_cmp(int i)
+{
+  extern void do_insn_cmp();
+  extern int cycle;
+  u_int hr,reglist=0;
+
+  for(hr=0;hr<HOST_REGS;hr++)
+    if(regs[i].regmap[hr]>=0) reglist|=1<<hr;
+  save_regs(reglist);
+  emit_movimm(start+i*4,0);
+  emit_writeword(0,(int)&pcaddr);
+  emit_call((int)do_insn_cmp);
+  //emit_readword((int)&cycle,0);
+  //emit_addimm(0,2,0);
+  //emit_writeword(0,(int)&cycle);
+  restore_regs(reglist);
+}
+#else
+#define drc_dbg_emit_do_cmp(x)
+#endif
+
 // Used when a branch jumps into the delay slot of another branch
 void ds_assemble_entry(int i)
 {
@@ -4204,6 +4292,7 @@ void ds_assemble_entry(int i)
   if(!instr_addr[t]) instr_addr[t]=(u_int)out;
   assem_debug("Assemble delay slot at %x\n",ba[i]);
   assem_debug("<->\n");
+  drc_dbg_emit_do_cmp(t);
   if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
     wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty,regs[t].was32);
   load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,rs1[t],rs2[t]);
@@ -6936,13 +7025,14 @@ static void disassemble_inst(int i) {}
 static int new_dynarec_test(void)
 {
   int (*testfunc)(void) = (void *)out;
+  void *beginning;
   int ret;
+
+  beginning = start_block();
   emit_movimm(DRC_TEST_VAL,0); // test
   emit_jmpreg(14);
   literal_pool(0);
-#ifdef __arm__
-  __clear_cache((void *)testfunc, out);
-#endif
+  end_block(beginning);
   SysPrintf("testing if we can run recompiled code..\n");
   ret = testfunc();
   if (ret == DRC_TEST_VAL)
@@ -6979,19 +7069,43 @@ void new_dynarec_clear_full()
 void new_dynarec_init()
 {
   SysPrintf("Init new dynarec\n");
-  out=(u_char *)BASE_ADDR;
-#if BASE_ADDR_FIXED
-  if (mmap (out, 1<<TARGET_SIZE_2,
+
+  // 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 VITA
+  sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
+  if (sceBlock < 0)
+    SysPrintf("sceKernelAllocMemBlockForVM failed\n");
+  int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&translation_cache);
+  if (ret < 0)
+    SysPrintf("sceKernelGetMemBlockBase failed\n");
+  #else
+  translation_cache = mmap (NULL, 1 << TARGET_SIZE_2,
             PROT_READ | PROT_WRITE | PROT_EXEC,
-            MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
-            -1, 0) <= 0) {
+            MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+  if (translation_cache == MAP_FAILED) {
     SysPrintf("mmap() failed: %s\n", strerror(errno));
+    abort();
   }
+  #endif
 #else
+  #ifndef NO_WRITE_EXEC
   // not all systems allow execute in data segment by default
-  if (mprotect(out, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
+  if (mprotect((void *)BASE_ADDR, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
     SysPrintf("mprotect() failed: %s\n", strerror(errno));
+  #endif
 #endif
+  out=(u_char *)BASE_ADDR;
   cycle_multiplier=200;
   new_dynarec_clear_full();
 #ifdef HOST_IMM8
@@ -7010,9 +7124,15 @@ void new_dynarec_init()
 void new_dynarec_cleanup()
 {
   int n;
-  #if BASE_ADDR_FIXED
-  if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0) {SysPrintf("munmap() failed\n");}
+#if defined(BASE_ADDR_FIXED) || defined(BASE_ADDR_DYNAMIC)
+  #ifdef VITA
+  sceKernelFreeMemBlock(sceBlock);
+  sceBlock = -1;
+  #else
+  if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0)
+    SysPrintf("munmap() failed\n");
   #endif
+#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);
@@ -7172,16 +7292,15 @@ int new_recompile_block(int addr)
   if (Config.HLE && start == 0x80001000) // hlecall
   {
     // XXX: is this enough? Maybe check hleSoftCall?
-    u_int beginning=(u_int)out;
+    void *beginning=start_block();
     u_int page=get_page(start);
+
     invalid_code[start>>12]=0;
     emit_movimm(start,0);
     emit_writeword(0,(int)&pcaddr);
     emit_jmp((int)new_dyna_leave);
     literal_pool(0);
-#ifdef __arm__
-    __clear_cache((void *)beginning,out);
-#endif
+    end_block(beginning);
     ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
     return 0;
   }
@@ -9883,7 +10002,7 @@ int new_recompile_block(int addr)
   cop1_usable=0;
   uint64_t is32_pre=0;
   u_int dirty_pre=0;
-  u_int beginning=(u_int)out;
+  void *beginning=start_block();
   if((u_int)addr&1) {
     ds=1;
     pagespan_ds();
@@ -9937,6 +10056,8 @@ int new_recompile_block(int addr)
       // branch target entry point
       instr_addr[i]=(u_int)out;
       assem_debug("<->\n");
+      drc_dbg_emit_do_cmp(i);
+
       // load regs
       if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
         wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32);
@@ -10173,14 +10294,12 @@ int new_recompile_block(int addr)
   // Align code
   if(((u_int)out)&7) emit_addnop(13);
   #endif
-  assert((u_int)out-beginning<MAX_OUTPUT_BLOCK_SIZE);
+  assert((u_int)out-(u_int)beginning<MAX_OUTPUT_BLOCK_SIZE);
   //printf("shadow buffer: %x-%x\n",(int)copy,(int)copy+slen*4);
   memcpy(copy,source,slen*4);
   copy+=slen*4;
 
-  #ifdef __arm__
-  __clear_cache((void *)beginning,out);
-  #endif
+  end_block(beginning);
 
   // If we're within 256K of the end of the buffer,
   // start over from the beginning. (Is 256K enough?)