PR6028 / BZ468143: ia64 fetch-register speedup

4 files changed, 36 insertions, 6 deletions

diff --git a/ChangeLog b/ChangeLog
index 6e9cd7f4..f7ecbb90 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,12 @@
+2008-11-04  Frank Ch. Eigler  <fche@elastic.org>
+
+	From Masami Hiramatsu <mhiramat@redhat.com>, PR 6028.
+	* tapsets.cxx (dwarf_derived_probe): New field access_vars, for
+	tracking result from ..
+	(dwarf_var_expanding_copy_visitor): Collect a flag about whether
+	any $expressions appear within.
+	(emit_probe_local_init): ... and run ia64 bspcache() on if true.
+
 2008-11-03  Jim Keniston  <jkenisto@us.ibm.com>
 
 	Add a version of uprobes that works with the 2.6.27 version of utrace.
diff --git a/runtime/ChangeLog b/runtime/ChangeLog
index a59af1b7..e4ea3330 100644
--- a/runtime/ChangeLog
+++ b/runtime/ChangeLog
@@ -1,3 +1,8 @@
+2008-11-04  Frank Ch. Eigler  <fche@elastic.org>
+
+	From Masami Hiramatsu <mhiramat@redhat.com>, PR 6028.
+	* regs-ia64.c (bspcache): Expand documentation.
+
 2008-10-28  Frank Ch. Eigler  <fche@elastic.org>
 
 	From Jim Keniston:
diff --git a/runtime/regs-ia64.c b/runtime/regs-ia64.c
index f884c5f8..c78a757e 100644
--- a/runtime/regs-ia64.c
+++ b/runtime/regs-ia64.c
@@ -36,8 +36,17 @@ static void ia64_stap_get_arbsp(struct unw_frame_info *info, void *arg)
 }
 
 /*
- * bspcache: get cached unwound address and
- * 	     set a probe local cache of the offset of unwound address.
+ * bspcache: get cached unwound address from current BSP and
+ * 	     set a static local cache of the offset of unwound address
+ * 	     if the static local cache(variable) is not set.
+ * 	     The unwound address means the BSP value when the kprobe
+ * 	     was hit. This macro stores the difference of BSP between
+ * 	     when the kprobe was hit and when this macro was called.
+ * 	     Since the difference depends on how much stack is consumed
+ * 	     from when the kprobe was hit, this macro *MUST NOT* be put
+ * 	     on the path which several functions execute (in that case,
+ * 	     each time the difference is changed and this macro can't
+ * 	     return correct unwound address).
  */
 #define bspcache(cache, regs)\
 	if(regs) {\
diff --git a/tapsets.cxx b/tapsets.cxx
index 0df55d51..5acf50cb 100644
--- a/tapsets.cxx
+++ b/tapsets.cxx
@@ -2397,6 +2397,7 @@ struct dwarf_derived_probe: public derived_probe
   bool has_return;
   bool has_maxactive;
   long maxactive_val;
+  bool access_vars;
 
   void printsig (std::ostream &o) const;
   void join_group (systemtap_session& s);
@@ -4104,9 +4105,10 @@ struct dwarf_var_expanding_copy_visitor: public var_expanding_copy_visitor
   block *add_block;
   probe *add_probe;
   std::map<std::string, symbol *> return_ts_map;
+  bool visited;
 
   dwarf_var_expanding_copy_visitor(dwarf_query & q, Dwarf_Die *sd, Dwarf_Addr a):
-    q(q), scope_die(sd), addr(a), add_block(NULL), add_probe(NULL) {}
+    q(q), scope_die(sd), addr(a), add_block(NULL), add_probe(NULL), visited(false) {}
   void visit_target_symbol (target_symbol* e);
 };
 
@@ -4172,6 +4174,7 @@ void
 dwarf_var_expanding_copy_visitor::visit_target_symbol (target_symbol *e)
 {
   assert(e->base_name.size() > 0 && e->base_name[0] == '$');
+  visited = true;
 
   bool lvalue = is_active_lvalue(e);
   if (lvalue && !q.sess.guru_mode)
@@ -4709,6 +4712,7 @@ dwarf_derived_probe::dwarf_derived_probe(const string& funcname,
     throw semantic_error ("inconsistent relocation address", q.base_loc->tok);
 
   this->tok = q.base_probe->tok;
+  this->access_vars = false;
 
   // XXX: hack for strange g++/gcc's
 #ifndef USHRT_MAX
@@ -4726,12 +4730,12 @@ dwarf_derived_probe::dwarf_derived_probe(const string& funcname,
     {
       dwarf_var_expanding_copy_visitor v (q, scope_die, dwfl_addr);
       require <statement*> (&v, &(this->body), this->body);
+      this->access_vars = v.visited;
 
       // If during target-variable-expanding the probe, we added a new block
       // of code, add it to the start of the probe.
       if (v.add_block)
         this->body = new block(v.add_block, this->body);
-
       // If when target-variable-expanding the probe, we added a new
       // probe, add it in a new file to the list of files to be processed.
       if (v.add_probe)
@@ -4857,8 +4861,11 @@ dwarf_derived_probe::register_patterns(match_node * root)
 void
 dwarf_derived_probe::emit_probe_local_init(translator_output * o)
 {
-  // emit bsp cache setup
-  o->newline() << "bspcache(c->unwaddr, c->regs);";
+  if (access_vars)
+    {
+      // if accessing $variables, emit bsp cache setup for speeding up
+      o->newline() << "bspcache(c->unwaddr, c->regs);";
+    }
 }
 
 // ------------------------------------------------------------------------