There are two problems I am fixing here (see
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59233 for full analysis).
The first is the original ICE in crossjump optimization, which was
exposed by enabling -freorder-blocks-and-partition which was
conservatively preventing some upstream block combining optimizations.
The issue is that the block ended in a NOTE_INSN_DELETED, which
old_insns_match_p could not handle. I am fixing this by passing
old_insns_match_p the last instruction in the block that it knows how
to handle.
The second issue this exposed was that we were unnecessarily marking
landing pad branches EDGE_PRESERVE since
flag_reorder_blocks_and_partition was on, even though this was -Os and
we will gate partitioning off. Added a method for checking whether we
will really invoke partitioning.
Fixes the original problem (confirmed via a cross-compiler for
x86_64-apple-darwin10) and was bootstrapped and tested on
x86-64-unknown-linux-gnu. Ok for trunk?
Thanks,
Teresa
2013-11-21 Teresa Johnson <tejohn...@google.com>
* bb-reorder.c (do_partition_blocks): New function.
(gate_handle_partition_blocks): Call do_partition_blocks.
* bb-reorder.h (do_partition_blocks): Declare.
* cfgcleanup.c (outgoing_edges_match): Walk up past note instructions
not understood by old_insns_match_p.
* except.c (dw2_build_landing_pads): Call do_partition_blocks.
Index: bb-reorder.c
===================================================================
--- bb-reorder.c (revision 205063)
+++ bb-reorder.c (working copy)
@@ -2532,8 +2532,13 @@ make_pass_duplicate_computed_gotos (gcc::context *
return new pass_duplicate_computed_gotos (ctxt);
}
-static bool
-gate_handle_partition_blocks (void)
+/* This function will return true if hot/cold partitioning will
+ be performed, and is declared non-static so that other passes
+ (namely landing pad expansion in except.c) can more accurately
+ determine whether they need to assume that partitioning will occur. */
+
+bool
+do_partition_blocks (void)
{
/* The optimization to partition hot/cold basic blocks into separate
sections of the .o file does not work well with linkonce or with
@@ -2548,6 +2553,12 @@ make_pass_duplicate_computed_gotos (gcc::context *
&& !user_defined_section_attribute);
}
+static bool
+gate_handle_partition_blocks (void)
+{
+ return do_partition_blocks ();
+}
+
/* This function is the main 'entrance' for the optimization that
partitions hot and cold basic blocks into separate sections of the
.o file (to improve performance and cache locality). Ideally it
Index: bb-reorder.h
===================================================================
--- bb-reorder.h (revision 205063)
+++ bb-reorder.h (working copy)
@@ -37,4 +37,5 @@ extern int get_uncond_jump_length (void);
extern void insert_section_boundary_note (void);
+extern bool do_partition_blocks (void);
#endif
Index: cfgcleanup.c
===================================================================
--- cfgcleanup.c (revision 205063)
+++ cfgcleanup.c (working copy)
@@ -1743,12 +1743,20 @@ outgoing_edges_match (int mode, basic_block bb1, b
}
}
+ /* Find the last non-debug non-note instruction in each bb, except
+ stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
+ handles that case specially. old_insns_match_p does not handle
+ other types of instruction notes. */
rtx last1 = BB_END (bb1);
rtx last2 = BB_END (bb2);
- if (DEBUG_INSN_P (last1))
- last1 = prev_nondebug_insn (last1);
- if (DEBUG_INSN_P (last2))
- last2 = prev_nondebug_insn (last2);
+ while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
+ (DEBUG_INSN_P (last1) || NOTE_P (last1)))
+ last1 = PREV_INSN (last1);
+ while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
+ (DEBUG_INSN_P (last2) || NOTE_P (last2)))
+ last2 = PREV_INSN (last2);
+ gcc_assert (last1 && last2);
+
/* First ensure that the instructions match. There may be many outgoing
edges so this test is generally cheaper. */
if (old_insns_match_p (mode, last1, last2) != dir_both)
Index: except.c
===================================================================
--- except.c (revision 205063)
+++ except.c (working copy)
@@ -125,6 +125,7 @@ along with GCC; see the file COPYING3. If not see
#include "except.h"
#include "hard-reg-set.h"
#include "basic-block.h"
+#include "bb-reorder.h"
#include "output.h"
#include "dwarf2asm.h"
#include "dwarf2out.h"
@@ -1014,7 +1015,7 @@ dw2_build_landing_pads (void)
new landing pads later, which means that we need to hold on to
the post-landing-pad block. Prevent it from being merged away.
We'll remove this bit after partitioning. */
- if (flag_reorder_blocks_and_partition)
+ if (do_partition_blocks ())
e_flags |= EDGE_PRESERVE;
for (i = 1; vec_safe_iterate (cfun->eh->lp_array, i, &lp); ++i)
--
Teresa Johnson | Software Engineer | tejohn...@google.com | 408-460-2413
