> Missing test cases?
Good point. The testcases I have needs rest of the patches from the series to
hit the mainline. For now I have added the following to test the basic scenario
* g++.dg/ipa/devirt-15.C: New testcase.
Index: g++.dg/ipa/devirt-15.C
===================================================================
--- g++.dg/ipa/devirt-15.C (revision 0)
+++ g++.dg/ipa/devirt-15.C (working copy)
@@ -0,0 +1,40 @@
+/* Check that we speculatively devirutalize call to FOO to A::foo becuase
+ B is noreturn. */
+/* { dg-do run } */
+/* { dg-options "-O2 -fdump-ipa-devirt" } */
+class A {
+public:
+ virtual int foo(void)
+ {
+ throw (1);
+ return 0;
+ }
+};
+
+
+class B : public A {
+public:
+ virtual int foo(void);
+};
+
+int
+B::foo(void)
+{
+ return 1;
+}
+class A a, *b=&a;
+void
+m(void)
+{
+ b->foo();
+}
+main()
+{
+ m();
+}
+
+/* { dg-final { scan-ipa-dump "Speculatively devirtualizing call" "devirt"} }
*/
+/* { dg-final { cleanup-ipa-dump "devirt" } } */
+/* Match if (PROF_6 == foo) to verify that the speculation survived. */
+/* { dg-final { scan-tree-dump "== foo" "optimized"} } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
>
> Have you tested the optimization with SPEC2k and SPEC06? There are a
> couple of benchmarks benefit greatly from devirtualization, such as
> eon, povray etc. I believe astar will probably improve with this
> optimization at O2 (it has hot virtual functions that are not
> overridden at all). For eon, the assumption at O2 for speculative
> devirt may not work well.
I trusted Martin (Jambor) that only interesting testcase for devirtualization
in SPEC in xalancbmk. We will definitely see tomorrow.
Concerning eon, I think it includes all the classes for individual objects so
it should not get devirtualized the wrong way. I will double check. Wrong
guess in general should not be terribly expensive. Many speculative calls are
eliminated by inlining and other will just end up with the extra
conditinal+jump (to jump). For firefox the code growth for all those 30000
devirtualization is about 1.8%. Here about 70% of them survive till
final executable since they use virtual functions for very cheap accessors
that gets inlined well.
There is an accounting bug in ipa-cp that makes us to mistakely remove
references to some virutal functions as dead after devirtualization, so I
disabled tracking of described references for virtual calls. Once this is
fixed, the code size may get better.
Honza
