PATCH: optimise call graph for coverage data

[Ben Elliston]

On Fri, 2008-04-04 at 03:04 +0200, Michael Matz wrote:

> > tree-cfg.c creates a fake edge from a call to a builtin to the exit
> > block of the enclosing function because the call is not a noreturn
> > function (ie. it may longjmp).  Thus, the builtin is treated as a
> > regular function call.  Many of the GCC builtins are neither const nor
> > pure, so a new ECF constant seems to be needed.  I chose
> > ECF_RETURNS_ONCE.
> 
> Isn't this the same as ECF_NOTHROW && !ECF_NORETURN && !ECF_RETURNS_TWICE?  
> After all, if a function doesn't throw, does return, but doesn't return 
> twice, it returns exactly once.  Couldn't the predicate for 
> need_fake_edge_p be adjusted to take ECF_NOTHROW into account?

Thanks for the review.  How about this revised patch?  (Testing
underway.)

Ben


2008-04-04  Ben Elliston  <bje@au.ibm.com>

        * tree-cfg.c (need_fake_edge_p): Return false for calls to
        builtins that return once and do not throw.

Index: tree-cfg.c
===================================================================
--- tree-cfg.c  (revision 133889)
+++ tree-cfg.c  (working copy)
@@ -6403,6 +6403,16 @@ need_fake_edge_p (tree t)
      the counter incrementation code from -fprofile-arcs
      leads to different results from -fbranch-probabilities.  */
   call = get_call_expr_in (t);
+
+  if (call
+      && CALL_EXPR_FN (call)
+      && TREE_CODE (TREE_OPERAND (CALL_EXPR_FN (call), 0)) == FUNCTION_DECL
+      && DECL_BUILT_IN (TREE_OPERAND (CALL_EXPR_FN (call), 0))
+      && (call_expr_flags (call) & ECF_NOTHROW)
+      && !(call_expr_flags (call) & ECF_NORETURN)
+      && !(call_expr_flags (call) & ECF_RETURNS_TWICE))
+    return false;
+
   if (call
       && !(call_expr_flags (call) & ECF_NORETURN))
     return true;