Merge pull request snabbco#186 from lukego/fix-jit-traces-alloc

lj_trace.c: Fix and simplify allocation of J->trace

src/lj_jit.h

@@ -206,9 +206,11 @@ typedef struct GCtrace {
   uint8_t unused1;
 } GCtrace;
 
+#define TRACE_MAX 65535
+
 #define gco2trace(o)	check_exp((o)->gch.gct == ~LJ_TTRACE, (GCtrace *)(o))
 #define traceref(J, n) \
-  check_exp((n)>0 && (MSize)(n)<J->sizetrace, (GCtrace *)gcref(J->trace[(n)]))
+  check_exp((n)>0 && (MSize)(n)<TRACE_MAX, (GCtrace *)gcref(J->trace[(n)]))
 
 LJ_STATIC_ASSERT(offsetof(GChead, gclist) == offsetof(GCtrace, gclist));
 
@@ -359,7 +361,7 @@ typedef struct jit_State {
 
   GCRef *trace;		/* Array of traces. */
   TraceNo freetrace;	/* Start of scan for next free trace. */
-  MSize sizetrace;	/* Size of trace array. */
+  uint16_t ntraces;	/* Number of traces created since last flush. */
   IRRef1 ktrace;	/* Reference to KGC with GCtrace. */
 
   IRRef1 chain[IR__MAX];  /* IR instruction skip-list chain anchors. */

src/lj_state.c

@@ -172,6 +172,7 @@ static void close_state(lua_State *L)
   lj_mem_free(g, J->snapmapbuf, sizeof(SnapEntry)*65536);
   lj_mem_free(g, J->snapbuf, sizeof(SnapShot)*65536);
   lj_mem_free(g, J->irbuf, 65536*sizeof(IRIns));
+  lj_mem_free(g, J->trace, TRACE_MAX * sizeof(GCRef *));
 #if 0
   /* XXX Fix deallocation so that this assertion succeeds. */
   lua_assert(g->gc.total == sizeof(GG_State));
@@ -224,9 +225,10 @@ LUA_API lua_State *lua_newstate(lua_Alloc f, void *ud)
   J->bclog = (BCRecLog *)lj_mem_new(L, sizeof(BCRecLog)*J->maxbclog);
   J->nbclog = 0;
   J->irbuf = (IRIns *)lj_mem_new(L, sizeof(IRIns)*65536);
-  if (J->irbuf == NULL || J->snapbuf == NULL ||
-      J->bclog == NULL || J->snapmapbuf == NULL)
+  J->trace = (GCRef *)lj_mem_new(L, TRACE_MAX * sizeof(GCRef *));
+  if (!(J->irbuf && J->snapbuf && J->bclog && J->snapmapbuf && J->trace))
     return NULL;
+  memset(J->trace, 0, TRACE_MAX * sizeof(GCRef *));
   lj_dispatch_init((GG_State *)L);
   L->status = LUA_ERRERR+1;  /* Avoid touching the stack upon memory error. */
   if (lj_vm_cpcall(L, NULL, NULL, cpluaopen) != 0) {

src/lj_trace.c

@@ -58,22 +58,14 @@ void lj_trace_err_info(jit_State *J, TraceError e)
 /* Find a free trace number. */
 static TraceNo trace_findfree(jit_State *J)
 {
-  MSize osz, lim;
   if (J->freetrace == 0)
     J->freetrace = 1;
-  for (; J->freetrace < J->sizetrace; J->freetrace++)
+  /* Search for a free slot. */
+  for (; J->freetrace < TRACE_MAX; J->freetrace++)
     if (traceref(J, J->freetrace) == NULL)
       return J->freetrace++;
-  /* Need to grow trace array. */
-  lim = (MSize)J->param[JIT_P_maxtrace] + 1;
-  if (lim < 2) lim = 2; else if (lim > 65535) lim = 65535;
-  osz = J->sizetrace;
-  if (osz >= lim)
-    return 0;  /* Too many traces. */
-  lj_mem_growvec(J->L, J->trace, J->sizetrace, lim, GCRef);
-  for (; osz < J->sizetrace; osz++)
-    setgcrefnull(J->trace[osz]);
-  return J->freetrace;
+  /* No free slot in trace array. */
+  return 0;
 }
 
 #define TRACE_APPENDVEC(field, szfield, tp) \
@@ -223,7 +215,7 @@ static void trace_flushroot(jit_State *J, GCtrace *T)
 /* Flush a trace. Only root traces are considered. */
 void lj_trace_flush(jit_State *J, TraceNo traceno)
 {
-  if (traceno > 0 && traceno < J->sizetrace) {
+  if (traceno > 0 && traceno < TRACE_MAX) {
     GCtrace *T = traceref(J, traceno);
     if (T && T->root == 0)
       trace_flushroot(J, T);
@@ -246,17 +238,20 @@ int lj_trace_flushall(lua_State *L)
   if ((J2G(J)->hookmask & HOOK_GC))
     return 1;
   lj_auditlog_trace_flushall(J);
-  for (i = (ptrdiff_t)J->sizetrace-1; i > 0; i--) {
-    GCtrace *T = traceref(J, i);
-    if (T) {
-      if (T->root == 0)
-	trace_flushroot(J, T);
-      lj_gdbjit_deltrace(J, T);
-      T->traceno = T->link = 0;  /* Blacklist the link for cont_stitch. */
-      setgcrefnull(J->trace[i]);
+  if (J->trace) {
+    for (i = (ptrdiff_t)TRACE_MAX-1; i > 0; i--) {
+      GCtrace *T = traceref(J, i);
+      if (T) {
+        if (T->root == 0)
+          trace_flushroot(J, T);
+        lj_gdbjit_deltrace(J, T);
+        T->traceno = T->link = 0;  /* Blacklist the link for cont_stitch. */
+        setgcrefnull(J->trace[i]);
+      }
     }
   }
   J->cur.traceno = 0;
+  J->ntraces = 0;
   J->freetrace = 0;
   g->lasttrace = 0;
   /* Unpatch blacklisted byte codes. */
@@ -306,12 +301,11 @@ void lj_trace_freestate(global_State *g)
 #ifdef LUA_USE_ASSERT
   {  /* This assumes all traces have already been freed. */
     ptrdiff_t i;
-    for (i = 1; i < (ptrdiff_t)J->sizetrace; i++)
+    for (i = 1; i < (ptrdiff_t)TRACE_MAX-1; i++)
       lua_assert(i == (ptrdiff_t)J->cur.traceno || traceref(J, i) == NULL);
   }
 #endif
   lj_mcode_free(J);
-  lj_mem_freevec(g, J->trace, J->sizetrace, GCRef);
 }
 
 /* -- Penalties and blacklisting ------------------------------------------ */
@@ -381,7 +375,7 @@ static void trace_start(jit_State *J)
 
   /* Get a new trace number. */
   traceno = trace_findfree(J);
-  if (LJ_UNLIKELY(traceno == 0)) {  /* No free trace? */
+  if (traceno == 0 || J->ntraces >= J->param[JIT_P_maxtrace]) {  /* No free trace? */
     lua_assert((J2G(J)->hookmask & HOOK_GC) == 0);
     lj_trace_flushall(J->L);
     J->state = LJ_TRACE_IDLE;  /* Silently ignored. */
@@ -469,6 +463,7 @@ static void trace_stop(jit_State *J)
   lj_mcode_commit(J, J->cur.mcode);
   J->postproc = LJ_POST_NONE;
   trace_save(J, T);
+  J->ntraces++;
 
   /* Clear any penalty after successful recording. */
   for (i = 0; i < PENALTY_SLOTS; i++)