[PATCH, Pointer Bounds Checker 40/x] Support in IPA ICF

[PATCH, Pointer Bounds Checker 40/x] Support in IPA ICF

[Ilya Enkovich]

Hi,

After recent merge with trunk I found that new IPA pass ICF requires few modifications for instrumented code:
 - instrumentation thunk existence means we cannot merge function into another one and should generate thunk instead
 - thunk generation should set with_bounds flag for instrumented funnctions
 - when DECL_INITIAL is cleaned for symbol, need_bounds_init should also be reset


Merge also required some minor changes in other patches due to changes includes, etc.  Nothing important to repost.

With these changes I get clean bootstrap and make check on linux-x86_64.  Also have clean runs of instrumented (run in NOP mode) and not isntrumented SPEC2000 and SPEC2006.

I also tried to bootstrap it for Darwin but unfortunately trunk fails to bootstrap there even with patches from PR63534 applied.  Error is the same for mpx branch.

Thanks,
Ilya
--
2014-10-28  Ilya Enkovich  <ilya.enkovich@intel.com>

	* cgraphunit.c (cgraph_node::expand_thunk): Set with_bounds flag
	for created call statement.
	* ipa-icf.c (sem_function::merge): Do not merge when instrumentation
	thunk still exists.
	(sem_variable::merge): Reset need_bounds_init flag.


diff --git a/gcc/cgraphunit.c b/gcc/cgraphunit.c
index f203f6d..c29e9aa 100644
--- a/gcc/cgraphunit.c
+++ b/gcc/cgraphunit.c
@@ -1585,6 +1585,7 @@ cgraph_node::expand_thunk (bool output_asm_thunks, bool force_gimple_thunk)
       call = gimple_build_call_vec (build_fold_addr_expr_loc (0, alias), vargs);
       callees->call_stmt = call;
       gimple_call_set_from_thunk (call, true);
+      gimple_call_set_with_bounds (call, instrumentation_clone);
       if (restmp)
 	{
           gimple_call_set_lhs (call, restmp);
diff --git a/gcc/ipa-icf.c b/gcc/ipa-icf.c
index e8c32c7..0ccda45 100644
--- a/gcc/ipa-icf.c
+++ b/gcc/ipa-icf.c
@@ -562,7 +562,8 @@ sem_function::merge (sem_item *alias_item)
       redirect_callers
 	= (!original_discardable
 	   && alias->get_availability () > AVAIL_INTERPOSABLE
-	   && original->get_availability () > AVAIL_INTERPOSABLE);
+	   && original->get_availability () > AVAIL_INTERPOSABLE
+	   && !alias->instrumented_version);
     }
   else
     {
@@ -1177,6 +1178,7 @@ sem_variable::merge (sem_item *alias_item)
       alias->analyzed = false;
 
       DECL_INITIAL (alias->decl) = NULL;
+      alias->need_bounds_init = false;
       alias->remove_all_references ();
 
       varpool_node::create_alias (alias_var->decl, decl);

Re: [PATCH, Pointer Bounds Checker 40/x] Support in IPA ICF

[Jeff Law]

On 10/28/14 10:38, Ilya Enkovich wrote:
> Hi,
>
> After recent merge with trunk I found that new IPA pass ICF requires few modifications for instrumented code:
>   - instrumentation thunk existence means we cannot merge function into another one and should generate thunk instead
>   - thunk generation should set with_bounds flag for instrumented funnctions
>   - when DECL_INITIAL is cleaned for symbol, need_bounds_init should also be reset
>
>
> Merge also required some minor changes in other patches due to changes includes, etc.  Nothing important to repost.
>
> With these changes I get clean bootstrap and make check on linux-x86_64.  Also have clean runs of instrumented (run in NOP mode) and not isntrumented SPEC2000 and SPEC2006.
>
> I also tried to bootstrap it for Darwin but unfortunately trunk fails to bootstrap there even with patches from PR63534 applied.  Error is the same for mpx branch.
>
> Thanks,
> Ilya
> --
> 2014-10-28  Ilya Enkovich  <ilya.enkovich@intel.com>
>
> 	* cgraphunit.c (cgraph_node::expand_thunk): Set with_bounds flag
> 	for created call statement.
> 	* ipa-icf.c (sem_function::merge): Do not merge when instrumentation
> 	thunk still exists.
> 	(sem_variable::merge): Reset need_bounds_init flag.
OK.

Last few days have been hell and thus hard to find the time to look at 
the state of the Darwin world.  Thanks for trying to test Darwin.

jeff