X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=c9e168847d3fc8f89f17b1f98841e915e300a15f;hb=af700b411e17806c3afb9e5f607317adc00dd546;hp=d369b815a311fb908605498232102c948ce2b40f;hpb=b7ad2f2c5261e3e94b6d85ab186b4aac5fb79172;p=pcsx_rearmed.git

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index d369b815..c9e16884 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -29,6 +29,10 @@
 #ifdef _3DS
 #include <3ds_utils.h>
 #endif
+#ifdef HAVE_LIBNX
+#include <switch.h>
+static Jit g_jit;
+#endif
 
 #include "new_dynarec_config.h"
 #include "../psxhle.h"
@@ -107,6 +111,17 @@ static struct ndrc_mem *ndrc;
 static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
 static struct ndrc_mem *ndrc = &ndrc_;
 #endif
+#ifdef NDRC_WRITE_OFFSET
+# ifdef __GLIBC__
+# include <sys/types.h>
+# include <sys/stat.h>
+# include <fcntl.h>
+# include <unistd.h>
+# endif
+static long ndrc_write_ofs;
+#else
+#define ndrc_write_ofs 0
+#endif
 
 // stubs
 enum stub_type {
@@ -179,7 +194,8 @@ struct block_info
   u_int tc_offs;
   //u_int tc_len;
   u_int reg_sv_flags;
-  u_short is_dirty;
+  u_char is_dirty;
+  u_char inv_near_misses;
   u_short jump_in_cnt;
   struct {
     u_int vaddr;
@@ -423,6 +439,16 @@ static void mprotect_w_x(void *start, void *end, int is_x)
     sceKernelCloseVMDomain();
   else
     sceKernelOpenVMDomain();
+  #elif defined(HAVE_LIBNX)
+  Result rc;
+  if (is_x)
+    rc = jitTransitionToExecutable(&g_jit);
+  else
+    rc = jitTransitionToWritable(&g_jit);
+  if (R_FAILED(rc))
+    SysPrintf("jitTransition %d %08x\n", is_x, rc);
+  #elif defined(NDRC_WRITE_OFFSET)
+  // separated rx and rw areas are always available
   #else
   u_long mstart = (u_long)start & ~4095ul;
   u_long mend = (u_long)end;
@@ -433,9 +459,10 @@ static void mprotect_w_x(void *start, void *end, int is_x)
 #endif
 }
 
-static void start_tcache_write(void *start, void *end)
+static void *start_tcache_write(void *start, void *end)
 {
   mprotect_w_x(start, end, 0);
+  return (char *)start + ndrc_write_ofs;
 }
 
 static void end_tcache_write(void *start, void *end)
@@ -450,6 +477,8 @@ static void end_tcache_write(void *start, void *end)
   sceKernelSyncVMDomain(sceBlock, start, len);
   #elif defined(_3DS)
   ctr_flush_invalidate_cache();
+  #elif defined(HAVE_LIBNX)
+  // handled in mprotect_w_x()
   #elif defined(__aarch64__)
   // as of 2021, __clear_cache() is still broken on arm64
   // so here is a custom one :(
@@ -477,6 +506,28 @@ static void end_block(void *start)
   end_tcache_write(start, out);
 }
 
+#ifdef NDRC_CACHE_FLUSH_ALL
+
+static int needs_clear_cache;
+
+static void mark_clear_cache(void *target)
+{
+  if (!needs_clear_cache) {
+    start_tcache_write(ndrc, ndrc + 1);
+    needs_clear_cache = 1;
+  }
+}
+
+static void do_clear_cache(void)
+{
+  if (needs_clear_cache) {
+    end_tcache_write(ndrc, ndrc + 1);
+    needs_clear_cache = 0;
+  }
+}
+
+#else
+
 // also takes care of w^x mappings when patching code
 static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
 
@@ -520,7 +571,7 @@ static void do_clear_cache(void)
   }
 }
 
-//#define DEBUG_CYCLE_COUNT 1
+#endif // NDRC_CACHE_FLUSH_ALL
 
 #define NO_CYCLE_PENALTY_THR 12
 
@@ -602,9 +653,9 @@ static void hash_table_remove(int vaddr)
 
 static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
 {
+  u_int vaddr_m = vaddr & 0x1fffffff;
   u_int i, j;
-  vaddr &= 0x1fffffff;
-  for (i = vaddr & ~0xfff; i < vaddr + len; i += 0x1000) {
+  for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
     // ram mirrors, but should not hurt bios
     for (j = 0; j < 0x800000; j += 0x200000) {
       invalid_code[(i|j) >> 12] =
@@ -645,7 +696,7 @@ static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
       if (memcmp(block->source, block->copy, block->len))
         continue;
 
-      block->is_dirty = 0;
+      block->is_dirty = block->inv_near_misses = 0;
       found_clean = block->jump_in[i].addr;
       hash_table_add(vaddr, found_clean);
       mark_invalid_code(block->start, block->len, 0);
@@ -1153,8 +1204,8 @@ static void *get_trampoline(const void *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;
+    const void **d = start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
+    *d = f;
     end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
   }
   return &ndrc->tramp.ops[i];
@@ -1347,10 +1398,11 @@ static void invalidate_block(struct block_info *block)
 static int invalidate_range(u_int start, u_int end,
   u32 *inv_start_ret, u32 *inv_end_ret)
 {
+  struct block_info *last_block = NULL;
   u_int start_page = get_page_prev(start);
   u_int end_page = get_page(end - 1);
   u_int start_m = pmmask(start);
-  u_int end_m = pmmask(end);
+  u_int end_m = pmmask(end - 1);
   u_int inv_start, inv_end;
   u_int blk_start_m, blk_end_m;
   u_int page;
@@ -1366,6 +1418,7 @@ static int invalidate_range(u_int start, u_int end,
     for (block = blocks[page]; block != NULL; block = block->next) {
       if (block->is_dirty)
         continue;
+      last_block = block;
       blk_end_m = pmmask(block->start + block->len);
       if (blk_end_m <= start_m) {
         inv_start = max(inv_start, blk_end_m);
@@ -1385,12 +1438,22 @@ static int invalidate_range(u_int start, u_int end,
     }
   }
 
+  if (!hit && last_block && last_block->source) {
+    // could be some leftover unused block, uselessly trapping writes
+    last_block->inv_near_misses++;
+    if (last_block->inv_near_misses > 128) {
+      invalidate_block(last_block);
+      stat_inc(stat_inv_hits);
+      hit++;
+    }
+  }
   if (hit) {
     do_clear_cache();
 #ifdef USE_MINI_HT
     memset(mini_ht, -1, sizeof(mini_ht));
 #endif
   }
+
   if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
     // the whole page is empty now
     mark_invalid_code(start, 1, 1);
@@ -2061,7 +2124,7 @@ static void cop0_alloc(struct regstat *current,int i)
   }
   else
   {
-    // TLBR/TLBWI/TLBWR/TLBP/ERET
+    // RFE
     assert(dops[i].opcode2==0x10);
     alloc_all(current,i);
   }
@@ -5311,7 +5374,7 @@ static void rjump_assemble(int i, const struct regstat *i_regs)
   //assert(adj==0);
   emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
   add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
-  if(dops[i+1].itype==COP0&&(source[i+1]&0x3f)==0x10)
+  if(dops[i+1].itype==COP0 && dops[i+1].opcode2==0x10)
     // special case for RFE
     emit_jmp(0);
   else
@@ -6044,7 +6107,7 @@ static void new_dynarec_test(void)
   }
 
   SysPrintf("testing if we can run recompiled code @%p...\n", out);
-  ((volatile u_int *)out)[0]++; // make cache dirty
+  ((volatile u_int *)(out + ndrc_write_ofs))[0]++; // make the cache dirty
 
   for (i = 0; i < ARRAY_SIZE(ret); i++) {
     out = ndrc->translation_cache;
@@ -6115,19 +6178,39 @@ void new_dynarec_init(void)
   #elif defined(_MSC_VER)
   ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
     PAGE_EXECUTE_READWRITE);
+  #elif defined(HAVE_LIBNX)
+  Result rc = jitCreate(&g_jit, sizeof(*ndrc));
+  if (R_FAILED(rc))
+    SysPrintf("jitCreate failed: %08x\n", rc);
+  SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
+  ndrc = g_jit.rx_addr;
+  ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
   #else
   uintptr_t desired_addr = 0;
+  int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
+  int flags = MAP_PRIVATE | MAP_ANONYMOUS;
+  int fd = -1;
   #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);
+  #ifdef NDRC_WRITE_OFFSET
+  // mostly for testing
+  fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
+  ftruncate(fd, sizeof(*ndrc));
+  void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
+                  (flags = MAP_SHARED), fd, 0);
+  assert(mw != MAP_FAILED);
+  prot = PROT_READ | PROT_EXEC;
+  #endif
+  ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
   if (ndrc == MAP_FAILED) {
     SysPrintf("mmap() failed: %s\n", strerror(errno));
     abort();
   }
+  #ifdef NDRC_WRITE_OFFSET
+  ndrc_write_ofs = (char *)mw - (char *)ndrc;
+  #endif
   #endif
 #else
   #ifndef NO_WRITE_EXEC
@@ -6162,9 +6245,13 @@ void new_dynarec_cleanup(void)
   // sceBlock is managed by retroarch's bootstrap code
   //sceKernelFreeMemBlock(sceBlock);
   //sceBlock = -1;
+  #elif defined(HAVE_LIBNX)
+  jitClose(&g_jit);
+  ndrc = NULL;
   #else
   if (munmap(ndrc, sizeof(*ndrc)) < 0)
     SysPrintf("munmap() failed\n");
+  ndrc = NULL;
   #endif
 #endif
   for (n = 0; n < ARRAY_SIZE(blocks); n++)
@@ -7013,9 +7100,9 @@ static noinline void pass2_unneeded_regs(int istart,int iend,int r)
       // SYSCALL instruction (software interrupt)
       u=1;
     }
-    else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
+    else if(dops[i].itype==COP0 && dops[i].opcode2==0x10)
     {
-      // ERET instruction (return from interrupt)
+      // RFE
       u=1;
     }
     //u=1; // DEBUG
@@ -8758,7 +8845,7 @@ static noinline void pass10_expire_blocks(void)
     u_int phase = (expirep >> (base_shift - 1)) & 1u;
     if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
       inv_debug("EXP: base_offs %x/%x phase %u\n", base_offs,
-        out - ndrc->translation_cache phase);
+        out - ndrc->translation_cache, phase);
     }
 
     if (!phase) {
@@ -8793,6 +8880,7 @@ static struct block_info *new_block_info(u_int start, u_int len,
   block->tc_offs = beginning - ndrc->translation_cache;
   //block->tc_len = out - beginning;
   block->is_dirty = 0;
+  block->inv_near_misses = 0;
   block->jump_in_cnt = jump_in_count;
 
   // insert sorted by start mirror-unmasked vaddr