/* Allocation for dataflow support routines.
Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007,
- 2008 Free Software Foundation, Inc.
+ 2008, 2009 Free Software Foundation, Inc.
Originally contributed by Michael P. Hayes
(m.hayes@elec.canterbury.ac.nz, mhayes@redhat.com)
Major rewrite contributed by Danny Berlin (dberlin@dberlin.org)
d) If the pass modifies all of the insns, as does register
allocation, it is simply better to rescan the entire function.
- e) If the pass uses either non-standard or ancient techniques to
- modify insns, automatic detection of the insns that need to be
- rescanned may be impractical. Cse and regrename fall into this
- category.
-
2) Deferred rescanning - Calls to df_insn_rescan, df_notes_rescan, and
df_insn_delete do not immediately change the insn but instead make
a note that the insn needs to be rescanned. The next call to
cause all of the pending rescans to be processed.
This is the technique of choice if either 1a, 1b, or 1c are issues
- in the pass. In the case of 1a or 1b, a call to df_remove_problem
- (df_chain) should be made before the next call to df_analyze or
- df_process_deferred_rescans.
+ in the pass. In the case of 1a or 1b, a call to df_finish_pass
+ (either manually or via TODO_df_finish) should be made before the
+ next call to df_analyze or df_process_deferred_rescans.
+
+ This mode is also used by a few passes that still rely on note_uses,
+ note_stores and for_each_rtx instead of using the DF data. This
+ can be said to fall under case 1c.
To enable this mode, call df_set_flags (DF_DEFER_INSN_RESCAN).
(This mode can be cleared by calling df_clear_flags
(DF_DEFER_INSN_RESCAN) but this does not cause the deferred insns to
be rescanned.
- 3) Total rescanning - In this mode the rescanning is disabled.
- However, the df information associated with deleted insn is delete
- at the time the insn is deleted. At the end of the pass, a call
- must be made to df_insn_rescan_all. This method is used by the
- register allocator since it generally changes each insn multiple
- times (once for each ref) and does not need to make use of the
- updated scanning information.
-
- It is also currently used by two older passes (cse, and regrename)
- which change insns in hard to track ways. It is hoped that this
- will be fixed soon since this it is expensive to rescan all of the
- insns when only a small number of them have really changed.
+3) Total rescanning - In this mode the rescanning is disabled.
+ Only when insns are deleted is the df information associated with
+ it also deleted. At the end of the pass, a call must be made to
+ df_insn_rescan_all. This method is used by the register allocator
+ since it generally changes each insn multiple times (once for each ref)
+ and does not need to make use of the updated scanning information.
4) Do it yourself - In this mechanism, the pass updates the insns
itself using the low level df primitives. Currently no pass does
/* Set the MASK flags in the DFLOW problem. The old flags are
returned. If a flag is not allowed to be changed this will fail if
checking is enabled. */
-enum df_changeable_flags
-df_set_flags (enum df_changeable_flags changeable_flags)
+int
+df_set_flags (int changeable_flags)
{
- enum df_changeable_flags old_flags = df->changeable_flags;
+ int old_flags = df->changeable_flags;
df->changeable_flags |= changeable_flags;
return old_flags;
}
/* Clear the MASK flags in the DFLOW problem. The old flags are
returned. If a flag is not allowed to be changed this will fail if
checking is enabled. */
-enum df_changeable_flags
-df_clear_flags (enum df_changeable_flags changeable_flags)
+int
+df_clear_flags (int changeable_flags)
{
- enum df_changeable_flags old_flags = df->changeable_flags;
+ int old_flags = df->changeable_flags;
df->changeable_flags &= ~changeable_flags;
return old_flags;
}
NULL, /* sub */
NULL, /* next */
0, /* static_pass_number */
- 0, /* tv_id */
+ TV_NONE, /* tv_id */
0, /* properties_required */
0, /* properties_provided */
0, /* properties_destroyed */
NULL, /* sub */
NULL, /* next */
0, /* static_pass_number */
- 0, /* tv_id */
+ TV_NONE, /* tv_id */
0, /* properties_required */
0, /* properties_provided */
0, /* properties_destroyed */
NULL, /* sub */
NULL, /* next */
0, /* static_pass_number */
- 0, /* tv_id */
+ TV_NONE, /* tv_id */
0, /* properties_required */
0, /* properties_provided */
0, /* properties_destroyed */
/* This will free "pending". */
-static void
-df_worklist_dataflow_overeager (struct dataflow *dataflow,
- bitmap pending,
- sbitmap considered,
- int *blocks_in_postorder,
- unsigned *bbindex_to_postorder)
-{
- enum df_flow_dir dir = dataflow->problem->dir;
- int count = 0;
-
- while (!bitmap_empty_p (pending))
- {
- unsigned bb_index;
- int index;
- count++;
-
- index = bitmap_first_set_bit (pending);
- bitmap_clear_bit (pending, index);
-
- bb_index = blocks_in_postorder[index];
-
- if (dir == DF_FORWARD)
- df_worklist_propagate_forward (dataflow, bb_index,
- bbindex_to_postorder,
- pending, considered);
- else
- df_worklist_propagate_backward (dataflow, bb_index,
- bbindex_to_postorder,
- pending, considered);
- }
-
- BITMAP_FREE (pending);
-
- /* Dump statistics. */
- if (dump_file)
- fprintf (dump_file, "df_worklist_dataflow_overeager:"
- "n_basic_blocks %d n_edges %d"
- " count %d (%5.2g)\n",
- n_basic_blocks, n_edges,
- count, count / (float)n_basic_blocks);
-}
static void
df_worklist_dataflow_doublequeue (struct dataflow *dataflow,
/* Worklist-based dataflow solver. It uses sbitmap as a worklist,
with "n"-th bit representing the n-th block in the reverse-postorder order.
- This is so-called over-eager algorithm where it propagates
- changes on demand. This algorithm may visit blocks more than
- iterative method if there are deeply nested loops.
- Worklist algorithm works better than iterative algorithm
- for CFGs with no nested loops.
- In practice, the measurement shows worklist algorithm beats
- iterative algorithm by some margin overall.
- Note that this is slightly different from the traditional textbook worklist solver,
- in that the worklist is effectively sorted by the reverse postorder.
- For CFGs with no nested loops, this is optimal.
-
- The overeager algorithm while works well for typical inputs,
- it could degenerate into excessive iterations given CFGs with high loop nests
- and unstructured loops. To cap the excessive iteration on such case,
- we switch to double-queueing when the original algorithm seems to
- get into such.
- */
+ The solver is a double-queue algorithm similar to the "double stack" solver
+ from Cooper, Harvey and Kennedy, "Iterative data-flow analysis, Revisited".
+ The only significant difference is that the worklist in this implementation
+ is always sorted in RPO of the CFG visiting direction. */
void
df_worklist_dataflow (struct dataflow *dataflow,
if (dataflow->problem->init_fun)
dataflow->problem->init_fun (blocks_to_consider);
- /* Solve it. Determine the solving algorithm
- based on a simple heuristic. */
- if (n_edges > PARAM_VALUE (PARAM_DF_DOUBLE_QUEUE_THRESHOLD_FACTOR)
- * n_basic_blocks)
- {
- /* High average connectivity, meaning dense graph
- with more likely deep nested loops
- or unstructured loops. */
- df_worklist_dataflow_doublequeue (dataflow, pending, considered,
- blocks_in_postorder,
- bbindex_to_postorder);
- }
- else
- {
- /* Most inputs fall into this case
- with relatively flat or structured CFG. */
- df_worklist_dataflow_overeager (dataflow, pending, considered,
- blocks_in_postorder,
- bbindex_to_postorder);
- }
+ /* Solve it. */
+ df_worklist_dataflow_doublequeue (dataflow, pending, considered,
+ blocks_in_postorder,
+ bbindex_to_postorder);
sbitmap_free (considered);
free (bbindex_to_postorder);
DF_REF_FLAGS (ref),
DF_REF_TYPE (ref));
if (DF_REF_LOC (ref))
- fprintf (file, "loc %p(%p) chain ", (void *)DF_REF_LOC (ref), (void *)*DF_REF_LOC (ref));
+ {
+ if (flag_dump_noaddr)
+ fprintf (file, "loc #(#) chain ");
+ else
+ fprintf (file, "loc %p(%p) chain ", (void *)DF_REF_LOC (ref),
+ (void *)*DF_REF_LOC (ref));
+ }
else
fprintf (file, "chain ");
df_chain_dump (DF_REF_CHAIN (ref), file);
OSDN Git Service
