Fix GC heap corruption on ARM. (#1389) accepted/tizen/5.5/unified/20200121.031421 submit/tizen_5.5/20200120.050447
authorAnton Lapounov <antonl@microsoft.com>
Wed, 8 Jan 2020 10:12:25 +0000 (02:12 -0800)
committerAlexander Soldatov/AI Compiler Lab /SRR/Staff Engineer/Samsung Electronics <soldatov.a@samsung.com>
Thu, 9 Jan 2020 13:48:46 +0000 (16:48 +0300)
The allocate_in_free code path in allocate_in_expanded_heap incorrectly calculated the large (double) alignment padding size when limiting the plug size (SHORT_PLUGS) if set_padding_on_saved_p was true:

    set_padding_in_expand (old_loc, set_padding_on_saved_p, pinned_plug_entry); // Sets the padding flag on the saved plug
    ...
    pad += switch_alignment_size (is_plug_padded (old_loc)); // Reads the padding flag from the old (different!) plug

That caused access violation during a later heap walk since the g_gc_pFreeObjectMethodTable pointer marking the gap was not placed at the right address.

Change-Id: I10a5fa443b3f5a614909983cd3c06d5420d22858

src/gc/gc.cpp

index f493664..f13cf17 100644 (file)
@@ -14961,7 +14961,7 @@ allocate_in_free:
 #else // FEATURE_STRUCTALIGN
         if (!((old_loc == 0) || same_large_alignment_p (old_loc, result+pad)))
         {
-            pad += switch_alignment_size (is_plug_padded (old_loc));
+            pad += switch_alignment_size (pad != 0);
             set_node_realigned (old_loc);
             dprintf (3, ("Allocation realignment old_loc: %Ix, new_loc:%Ix",
                          (size_t)old_loc, (size_t)(result+pad)));