Hi!
When tree-prof/bb-reorg.c is compiled with 32-bit host cc1 with
-fprofile-use, it is miscompiled. The problem is
that partition_hot_cold_basic_blocks leaves garbage (other bb
pointers) in bb->aux of some basic blocks and df_analyze
assumes (like other passes) that it is cleared and uses
->aux cast to ptrdiff_t as age value.
If the bb pointer is in the second half of address space (likely
on 32-bit, unlikely on 64-bit hosts), they are treated as negative age.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux,
ok for trunk?
We really should have some
#ifdef ENABLE_CHECKING
FOR_ALL_BB (bb)
gcc_assert (bb->aux == NULL);
#endif
somewhere in the pass manager and early in df_analyze.
2012-01-06 Jakub Jelinek <ja...@redhat.com>
PR gcov-profile/50127
* bb-reorder.c (partition_hot_cold_basic_blocks): Clear
bb->aux before running df_analyze.
--- gcc/bb-reorder.c.jj 2011-11-10 18:09:12.000000000 +0100
+++ gcc/bb-reorder.c 2012-01-06 13:47:39.619401395 +0100
@@ -2219,6 +2219,7 @@ static unsigned
partition_hot_cold_basic_blocks (void)
{
VEC(edge, heap) *crossing_edges;
+ basic_block bb;
if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1)
return 0;
@@ -2254,6 +2255,10 @@ partition_hot_cold_basic_blocks (void)
add_reg_crossing_jump_notes ();
+ /* Clear bb->aux fields that the above routines were using. */
+ FOR_EACH_BB (bb)
+ bb->aux = NULL;
+
VEC_free (edge, heap, crossing_edges);
/* ??? FIXME: DF generates the bb info for a block immediately.
Jakub
