[PATCH, PR 39259] Set calls_setjmp and calls_alloca if necessary when versioning functions
Martin Jambor
mjambor@suse.cz
Wed Feb 25 18:20:00 GMT 2009
On Tue, Feb 24, 2009 at 02:58:02PM +0100, Richard Guenther wrote:
> On Tue, Feb 24, 2009 at 2:43 PM, Martin Jambor <mjambor@suse.cz> wrote:
> > Hi,
> >
> > the following patch removes the assert that was hit in PR 39259 which
> > was hit when a function calling setjump was being versioned altogether
> > with a similar one for functions calling alloca. It also sets the
> > flags for the new version if it detects such calls when copying
> > function bodies.
> >
> > Honza suggested I do this rather than simply copying the flags because
> > in near future we are likely to do copies of only selected parts of
> > functions.
> >
> > It does fix the reported testcase and I have bootstrapped and
> > regression tested this patch (on x86_64-linux-gnu, rev 144403) with
> > extra asserts to make sure the flags in the old and new version were
> > the same. I will redo this with this exact patch to follow the rules
> > but I do not expect any problems.
> >
> > I do not really know what a testcase for a removed assert like this
> > should look like.
> >
> > Is this ok for trunk if it rebootstraps fine?
>
> Ok if you add the testcase from the PR to g++.dg/torture/
>
> Thanks,
> Richard.
OK, this is what I committed as revision 144428.
Thanks,
Martin
2009-02-25 Martin Jambor <mjambor@suse.cz>
* tree-inline.c (initialize_cfun): Remove asserts for calls_setjmp and
alls_alloca function flags.
(copy_bb): Set calls_setjmp and alls_alloca function flags if such
calls are detected.
Index: gcc/testsuite/g++.dg/torture/pr39259.C
===================================================================
--- gcc/testsuite/g++.dg/torture/pr39259.C (revision 0)
+++ gcc/testsuite/g++.dg/torture/pr39259.C (revision 0)
@@ -0,0 +1,40 @@
+// PR tree-optimization/39259
+// { dg-do compile }
+// { dg-options "-O2" }
+
+
+extern "C" int __mysetjmp () __attribute__ ((__returns_twice__));
+
+class TContStatus {};
+
+class TContEvent
+{
+public:
+ inline void Execute () throw();
+};
+
+class TCont
+{
+public:
+ TContStatus ReadD (void* buf, int deadline)
+ {
+ TContEvent event;
+ event.Execute ();
+ return TContStatus();
+ }
+ TContStatus ReadI (void *buf)
+ {
+ return ReadD (buf, 1);
+ }
+};
+
+void TContEvent::Execute () throw ()
+{
+ __mysetjmp();
+}
+
+void Broken (TCont *mCont)
+{
+ mCont->ReadI(0);
+ mCont->ReadI(0);
+}
Index: gcc/tree-inline.c
===================================================================
--- gcc/tree-inline.c (revision 144406)
+++ gcc/tree-inline.c (working copy)
@@ -1398,6 +1398,7 @@
{
struct cgraph_node *node;
struct cgraph_edge *edge;
+ int flags;
switch (id->transform_call_graph_edges)
{
@@ -1429,6 +1430,13 @@
default:
gcc_unreachable ();
}
+
+ flags = gimple_call_flags (stmt);
+
+ if (flags & ECF_MAY_BE_ALLOCA)
+ cfun->calls_alloca = true;
+ if (flags & ECF_RETURNS_TWICE)
+ cfun->calls_setjmp = true;
}
/* If you think we can abort here, you are wrong.
@@ -1745,10 +1753,6 @@
gcc_assert (cfun->cfg == NULL);
gcc_assert (cfun->decl == new_fndecl);
- /* No need to copy; this is initialized later in compilation. */
- gcc_assert (!src_cfun->calls_setjmp);
- gcc_assert (!src_cfun->calls_alloca);
-
/* Copy items we preserve during clonning. */
cfun->static_chain_decl = src_cfun->static_chain_decl;
cfun->nonlocal_goto_save_area = src_cfun->nonlocal_goto_save_area;
More information about the Gcc-patches
mailing list