/* SSA Dominator optimizations for trees
- Copyright (C) 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007
+ Free Software Foundation, Inc.
Contributed by Diego Novillo <dnovillo@redhat.com>
This file is part of GCC.
and its maximum index rather than use a varray. */
tree *cond_equivalences;
unsigned int max_cond_equivalences;
-
- /* If we can thread this edge this field records the new target. */
- edge redirection_target;
};
expressions are removed from AVAIL_EXPRS. Else we may change the
hash code for an expression and be unable to find/remove it from
AVAIL_EXPRS. */
-static VEC(tree,heap) *stmts_to_rescan;
+typedef tree *tree_p;
+DEF_VEC_P(tree_p);
+DEF_VEC_ALLOC_P(tree_p,heap);
+
+static VEC(tree_p,heap) *stmts_to_rescan;
/* Structure for entries in the expression hash table.
restored during finalization of this block. */
static VEC(tree,heap) *const_and_copies_stack;
-/* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not
- know their exact value. */
-static bitmap nonzero_vars;
-
-/* Bitmap of blocks that are scheduled to be threaded through. This
- is used to communicate with thread_through_blocks. */
-static bitmap threaded_blocks;
-
-/* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
- when the current block is finalized.
-
- A NULL entry is used to mark the end of names needing their
- entry in NONZERO_VARS cleared during finalization of this block. */
-static VEC(tree,heap) *nonzero_vars_stack;
-
/* Track whether or not we have changed the control flow graph. */
static bool cfg_altered;
long num_re;
long num_const_prop;
long num_copy_prop;
- long num_iterations;
};
static struct opt_stats_d opt_stats;
-/* Value range propagation record. Each time we encounter a conditional
- of the form SSA_NAME COND CONST we create a new vrp_element to record
- how the condition affects the possible values SSA_NAME may have.
-
- Each record contains the condition tested (COND), and the range of
- values the variable may legitimately have if COND is true. Note the
- range of values may be a smaller range than COND specifies if we have
- recorded other ranges for this variable. Each record also contains the
- block in which the range was recorded for invalidation purposes.
-
- Note that the current known range is computed lazily. This allows us
- to avoid the overhead of computing ranges which are never queried.
-
- When we encounter a conditional, we look for records which constrain
- the SSA_NAME used in the condition. In some cases those records allow
- us to determine the condition's result at compile time. In other cases
- they may allow us to simplify the condition.
-
- We also use value ranges to do things like transform signed div/mod
- operations into unsigned div/mod or to simplify ABS_EXPRs.
-
- Simple experiments have shown these optimizations to not be all that
- useful on switch statements (much to my surprise). So switch statement
- optimizations are not performed.
-
- Note carefully we do not propagate information through each statement
- in the block. i.e., if we know variable X has a value defined of
- [0, 25] and we encounter Y = X + 1, we do not track a value range
- for Y (which would be [1, 26] if we cared). Similarly we do not
- constrain values as we encounter narrowing typecasts, etc. */
-
-struct vrp_element
-{
- /* The highest and lowest values the variable in COND may contain when
- COND is true. Note this may not necessarily be the same values
- tested by COND if the same variable was used in earlier conditionals.
-
- Note this is computed lazily and thus can be NULL indicating that
- the values have not been computed yet. */
- tree low;
- tree high;
-
- /* The actual conditional we recorded. This is needed since we compute
- ranges lazily. */
- tree cond;
-
- /* The basic block where this record was created. We use this to determine
- when to remove records. */
- basic_block bb;
-};
-
-/* A hash table holding value range records (VRP_ELEMENTs) for a given
- SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
- that gets awful wasteful, particularly since the density objects
- with useful information is very low. */
-static htab_t vrp_data;
-
-typedef struct vrp_element *vrp_element_p;
-
-DEF_VEC_P(vrp_element_p);
-DEF_VEC_ALLOC_P(vrp_element_p,heap);
-
-/* An entry in the VRP_DATA hash table. We record the variable and a
- varray of VRP_ELEMENT records associated with that variable. */
-struct vrp_hash_elt
-{
- tree var;
- VEC(vrp_element_p,heap) *records;
-};
-
-/* Array of variables which have their values constrained by operations
- in this basic block. We use this during finalization to know
- which variables need their VRP data updated. */
-
-/* Stack of SSA_NAMEs which had their values constrained by operations
- in this basic block. During finalization of this block we use this
- list to determine which variables need their VRP data updated.
-
- A NULL entry marks the end of the SSA_NAMEs associated with this block. */
-static VEC(tree,heap) *vrp_variables_stack;
-
struct eq_expr_value
{
tree src;
basic_block bb,
block_stmt_iterator);
static tree lookup_avail_expr (tree, bool);
-static hashval_t vrp_hash (const void *);
-static int vrp_eq (const void *, const void *);
static hashval_t avail_expr_hash (const void *);
static hashval_t real_avail_expr_hash (const void *);
static int avail_expr_eq (const void *, const void *);
static void record_cond (tree, tree);
static void record_const_or_copy (tree, tree);
static void record_equality (tree, tree);
-static tree update_rhs_and_lookup_avail_expr (tree, tree, bool);
-static tree simplify_rhs_and_lookup_avail_expr (tree, int);
-static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
-static tree simplify_switch_and_lookup_avail_expr (tree, int);
-static tree find_equivalent_equality_comparison (tree);
-static void record_range (tree, basic_block);
-static bool extract_range_from_cond (tree, tree *, tree *, int *);
static void record_equivalences_from_phis (basic_block);
static void record_equivalences_from_incoming_edge (basic_block);
-static bool eliminate_redundant_computations (tree, stmt_ann_t);
+static bool eliminate_redundant_computations (tree);
static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
-static void thread_across_edge (struct dom_walk_data *, edge);
+static void dom_thread_across_edge (struct dom_walk_data *, edge);
static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
static void remove_local_expressions_from_table (void);
static void restore_vars_to_original_value (void);
static edge single_incoming_edge_ignoring_loop_edges (basic_block);
-static void restore_nonzero_vars_to_original_value (void);
-static inline bool unsafe_associative_fp_binop (tree);
-
-
-/* Local version of fold that doesn't introduce cruft. */
-
-static tree
-local_fold (tree t)
-{
- t = fold (t);
- /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
- may have been added by fold, and "useless" type conversions that might
- now be apparent due to propagation. */
- STRIP_USELESS_TYPE_CONVERSION (t);
-
- return t;
-}
/* Allocate an EDGE_INFO for edge E and attach it to E.
Return the new EDGE_INFO structure. */
{
struct edge_info *edge_info;
- edge_info = xcalloc (1, sizeof (struct edge_info));
+ edge_info = XCNEW (struct edge_info);
e->aux = edge_info;
return edge_info;
{
FOR_EACH_EDGE (e, ei, bb->preds)
{
- struct edge_info *edge_info = e->aux;
+ struct edge_info *edge_info = (struct edge_info *) e->aux;
if (edge_info)
{
- e->aux = edge_info->redirection_target;
if (edge_info->cond_equivalences)
free (edge_info->cond_equivalences);
free (edge_info);
+ e->aux = NULL;
}
}
}
}
-/* Free an instance of vrp_hash_elt. */
-
-static void
-vrp_free (void *data)
-{
- struct vrp_hash_elt *elt = data;
- struct VEC(vrp_element_p,heap) **vrp_elt = &elt->records;
-
- VEC_free (vrp_element_p, heap, *vrp_elt);
- free (elt);
-}
-
/* Jump threading, redundancy elimination and const/copy propagation.
This pass may expose new symbols that need to be renamed into SSA. For
every new symbol exposed, its corresponding bit will be set in
VARS_TO_RENAME. */
-static void
+static unsigned int
tree_ssa_dominator_optimize (void)
{
struct dom_walk_data walk_data;
unsigned int i;
- struct loops loops_info;
memset (&opt_stats, 0, sizeof (opt_stats));
/* Create our hash tables. */
avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
- vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq,
- vrp_free);
avail_exprs_stack = VEC_alloc (tree, heap, 20);
const_and_copies_stack = VEC_alloc (tree, heap, 20);
- nonzero_vars_stack = VEC_alloc (tree, heap, 20);
- vrp_variables_stack = VEC_alloc (tree, heap, 20);
- stmts_to_rescan = VEC_alloc (tree, heap, 20);
- nonzero_vars = BITMAP_ALLOC (NULL);
- threaded_blocks = BITMAP_ALLOC (NULL);
+ stmts_to_rescan = VEC_alloc (tree_p, heap, 20);
need_eh_cleanup = BITMAP_ALLOC (NULL);
/* Setup callbacks for the generic dominator tree walker. */
init_walk_dominator_tree (&walk_data);
calculate_dominance_info (CDI_DOMINATORS);
+ cfg_altered = false;
+
+ /* We need to know loop structures in order to avoid destroying them
+ in jump threading. Note that we still can e.g. thread through loop
+ headers to an exit edge, or through loop header to the loop body, assuming
+ that we update the loop info. */
+ loop_optimizer_init (LOOPS_HAVE_SIMPLE_LATCHES);
+
+ /* We need accurate information regarding back edges in the CFG
+ for jump threading; this may include back edes that are not part of
+ a single loop. */
+ mark_dfs_back_edges ();
+
+ /* Recursively walk the dominator tree optimizing statements. */
+ walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
- /* We need to know which edges exit loops so that we can
- aggressively thread through loop headers to an exit
- edge. */
- flow_loops_find (&loops_info);
- mark_loop_exit_edges (&loops_info);
- flow_loops_free (&loops_info);
-
- /* Clean up the CFG so that any forwarder blocks created by loop
- canonicalization are removed. */
- cleanup_tree_cfg ();
- calculate_dominance_info (CDI_DOMINATORS);
-
- /* If we prove certain blocks are unreachable, then we want to
- repeat the dominator optimization process as PHI nodes may
- have turned into copies which allows better propagation of
- values. So we repeat until we do not identify any new unreachable
- blocks. */
- do
- {
- /* Optimize the dominator tree. */
- cfg_altered = false;
-
- /* We need accurate information regarding back edges in the CFG
- for jump threading. */
- mark_dfs_back_edges ();
-
- /* Recursively walk the dominator tree optimizing statements. */
- walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
-
+ {
+ block_stmt_iterator bsi;
+ basic_block bb;
+ FOR_EACH_BB (bb)
{
- block_stmt_iterator bsi;
- basic_block bb;
- FOR_EACH_BB (bb)
- {
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
- {
- update_stmt_if_modified (bsi_stmt (bsi));
- }
- }
+ for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ update_stmt_if_modified (bsi_stmt (bsi));
}
+ }
- /* If we exposed any new variables, go ahead and put them into
- SSA form now, before we handle jump threading. This simplifies
- interactions between rewriting of _DECL nodes into SSA form
- and rewriting SSA_NAME nodes into SSA form after block
- duplication and CFG manipulation. */
- update_ssa (TODO_update_ssa);
-
- free_all_edge_infos ();
-
- /* Thread jumps, creating duplicate blocks as needed. */
- cfg_altered |= thread_through_all_blocks (threaded_blocks);
-
- /* Removal of statements may make some EH edges dead. Purge
- such edges from the CFG as needed. */
- if (!bitmap_empty_p (need_eh_cleanup))
- {
- cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
- bitmap_zero (need_eh_cleanup);
- }
-
- if (cfg_altered)
- free_dominance_info (CDI_DOMINATORS);
-
- /* Only iterate if we threaded jumps AND the CFG cleanup did
- something interesting. Other cases generate far fewer
- optimization opportunities and thus are not worth another
- full DOM iteration. */
- cfg_altered &= cleanup_tree_cfg ();
-
- if (rediscover_loops_after_threading)
- {
- /* Rerun basic loop analysis to discover any newly
- created loops and update the set of exit edges. */
- rediscover_loops_after_threading = false;
- flow_loops_find (&loops_info);
- mark_loop_exit_edges (&loops_info);
- flow_loops_free (&loops_info);
-
- /* Remove any forwarder blocks inserted by loop
- header canonicalization. */
- cleanup_tree_cfg ();
- }
-
- calculate_dominance_info (CDI_DOMINATORS);
+ /* If we exposed any new variables, go ahead and put them into
+ SSA form now, before we handle jump threading. This simplifies
+ interactions between rewriting of _DECL nodes into SSA form
+ and rewriting SSA_NAME nodes into SSA form after block
+ duplication and CFG manipulation. */
+ update_ssa (TODO_update_ssa);
- update_ssa (TODO_update_ssa);
+ free_all_edge_infos ();
- /* Reinitialize the various tables. */
- bitmap_clear (nonzero_vars);
- bitmap_clear (threaded_blocks);
- htab_empty (avail_exprs);
- htab_empty (vrp_data);
+ /* Thread jumps, creating duplicate blocks as needed. */
+ cfg_altered |= thread_through_all_blocks (first_pass_instance);
- /* Finally, remove everything except invariants in SSA_NAME_VALUE.
+ if (cfg_altered)
+ free_dominance_info (CDI_DOMINATORS);
- This must be done before we iterate as we might have a
- reference to an SSA_NAME which was removed by the call to
- update_ssa.
+ /* Removal of statements may make some EH edges dead. Purge
+ such edges from the CFG as needed. */
+ if (!bitmap_empty_p (need_eh_cleanup))
+ {
+ tree_purge_all_dead_eh_edges (need_eh_cleanup);
+ bitmap_zero (need_eh_cleanup);
+ }
- Long term we will be able to let everything in SSA_NAME_VALUE
- persist. However, for now, we know this is the safe thing to do. */
- for (i = 0; i < num_ssa_names; i++)
- {
- tree name = ssa_name (i);
- tree value;
+ /* Finally, remove everything except invariants in SSA_NAME_VALUE.
- if (!name)
- continue;
+ Long term we will be able to let everything in SSA_NAME_VALUE
+ persist. However, for now, we know this is the safe thing to do. */
+ for (i = 0; i < num_ssa_names; i++)
+ {
+ tree name = ssa_name (i);
+ tree value;
- value = SSA_NAME_VALUE (name);
- if (value && !is_gimple_min_invariant (value))
- SSA_NAME_VALUE (name) = NULL;
- }
+ if (!name)
+ continue;
- opt_stats.num_iterations++;
+ value = SSA_NAME_VALUE (name);
+ if (value && !is_gimple_min_invariant (value))
+ SSA_NAME_VALUE (name) = NULL;
}
- while (optimize > 1 && cfg_altered);
/* Debugging dumps. */
if (dump_file && (dump_flags & TDF_STATS))
dump_dominator_optimization_stats (dump_file);
- /* We emptied the hash table earlier, now delete it completely. */
- htab_delete (avail_exprs);
- htab_delete (vrp_data);
+ loop_optimizer_finalize ();
- /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
- CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
- of the do-while loop above. */
+ /* Delete our main hashtable. */
+ htab_delete (avail_exprs);
/* And finalize the dominator walker. */
fini_walk_dominator_tree (&walk_data);
- /* Free nonzero_vars. */
- BITMAP_FREE (nonzero_vars);
- BITMAP_FREE (threaded_blocks);
+ /* Free asserted bitmaps and stacks. */
BITMAP_FREE (need_eh_cleanup);
VEC_free (tree, heap, avail_exprs_stack);
VEC_free (tree, heap, const_and_copies_stack);
- VEC_free (tree, heap, nonzero_vars_stack);
- VEC_free (tree, heap, vrp_variables_stack);
- VEC_free (tree, heap, stmts_to_rescan);
+ VEC_free (tree_p, heap, stmts_to_rescan);
+ return 0;
}
static bool
0, /* todo_flags_start */
TODO_dump_func
| TODO_update_ssa
+ | TODO_cleanup_cfg
| TODO_verify_ssa, /* todo_flags_finish */
0 /* letter */
};
}
}
}
-/* We are exiting E->src, see if E->dest ends with a conditional
- jump which has a known value when reached via E.
-
- Special care is necessary if E is a back edge in the CFG as we
- will have already recorded equivalences for E->dest into our
- various tables, including the result of the conditional at
- the end of E->dest. Threading opportunities are severely
- limited in that case to avoid short-circuiting the loop
- incorrectly.
-
- Note it is quite common for the first block inside a loop to
- end with a conditional which is either always true or always
- false when reached via the loop backedge. Thus we do not want
- to blindly disable threading across a loop backedge. */
-
-static void
-thread_across_edge (struct dom_walk_data *walk_data, edge e)
-{
- block_stmt_iterator bsi;
- tree stmt = NULL;
- tree phi;
- int stmt_count = 0;
- int max_stmt_count;
-
-
- /* If E->dest does not end with a conditional, then there is
- nothing to do. */
- bsi = bsi_last (e->dest);
- if (bsi_end_p (bsi)
- || ! bsi_stmt (bsi)
- || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
- && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
- && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
- return;
-
- /* The basic idea here is to use whatever knowledge we have
- from our dominator walk to simplify statements in E->dest,
- with the ultimate goal being to simplify the conditional
- at the end of E->dest.
-
- Note that we must undo any changes we make to the underlying
- statements as the simplifications we are making are control
- flow sensitive (ie, the simplifications are valid when we
- traverse E, but may not be valid on other paths to E->dest. */
-
- /* Each PHI creates a temporary equivalence, record them. Again
- these are context sensitive equivalences and will be removed
- by our caller. */
- for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
- {
- tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
- tree dst = PHI_RESULT (phi);
-
- /* Do not include virtual PHIs in our statement count as
- they never generate code. */
- if (is_gimple_reg (dst))
- stmt_count++;
-
- /* If the desired argument is not the same as this PHI's result
- and it is set by a PHI in E->dest, then we can not thread
- through E->dest. */
- if (src != dst
- && TREE_CODE (src) == SSA_NAME
- && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
- && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
- return;
-
- record_const_or_copy (dst, src);
- }
-
- /* Try to simplify each statement in E->dest, ultimately leading to
- a simplification of the COND_EXPR at the end of E->dest.
-
- We might consider marking just those statements which ultimately
- feed the COND_EXPR. It's not clear if the overhead of bookkeeping
- would be recovered by trying to simplify fewer statements.
-
- If we are able to simplify a statement into the form
- SSA_NAME = (SSA_NAME | gimple invariant), then we can record
- a context sensitive equivalency which may help us simplify
- later statements in E->dest.
-
- Failure to simplify into the form above merely means that the
- statement provides no equivalences to help simplify later
- statements. This does not prevent threading through E->dest. */
- max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS);
- for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
- {
- tree cached_lhs = NULL;
-
- stmt = bsi_stmt (bsi);
-
- /* Ignore empty statements and labels. */
- if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
- continue;
-
- /* If duplicating this block is going to cause too much code
- expansion, then do not thread through this block. */
- stmt_count++;
- if (stmt_count > max_stmt_count)
- return;
-
- /* Safely handle threading across loop backedges. This is
- over conservative, but still allows us to capture the
- majority of the cases where we can thread across a loop
- backedge. */
- if ((e->flags & EDGE_DFS_BACK) != 0
- && TREE_CODE (stmt) != COND_EXPR
- && TREE_CODE (stmt) != SWITCH_EXPR)
- return;
-
- /* If the statement has volatile operands, then we assume we
- can not thread through this block. This is overly
- conservative in some ways. */
- if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
- return;
-
- /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
- value, then do not try to simplify this statement as it will
- not simplify in any way that is helpful for jump threading. */
- if (TREE_CODE (stmt) != MODIFY_EXPR
- || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
- continue;
-
- /* At this point we have a statement which assigns an RHS to an
- SSA_VAR on the LHS. We want to try and simplify this statement
- to expose more context sensitive equivalences which in turn may
- allow us to simplify the condition at the end of the loop. */
- if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
- cached_lhs = TREE_OPERAND (stmt, 1);
- else
- {
- /* Copy the operands. */
- tree *copy, pre_fold_expr;
- ssa_op_iter iter;
- use_operand_p use_p;
- unsigned int num, i = 0;
-
- num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
- copy = xcalloc (num, sizeof (tree));
-
- /* Make a copy of the uses & vuses into USES_COPY, then cprop into
- the operands. */
- FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
- {
- tree tmp = NULL;
- tree use = USE_FROM_PTR (use_p);
-
- copy[i++] = use;
- if (TREE_CODE (use) == SSA_NAME)
- tmp = SSA_NAME_VALUE (use);
- if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
- SET_USE (use_p, tmp);
- }
-
- /* Try to fold/lookup the new expression. Inserting the
- expression into the hash table is unlikely to help
- Sadly, we have to handle conditional assignments specially
- here, because fold expects all the operands of an expression
- to be folded before the expression itself is folded, but we
- can't just substitute the folded condition here. */
- if (TREE_CODE (TREE_OPERAND (stmt, 1)) == COND_EXPR)
- {
- tree cond = COND_EXPR_COND (TREE_OPERAND (stmt, 1));
- cond = fold (cond);
- if (cond == boolean_true_node)
- pre_fold_expr = COND_EXPR_THEN (TREE_OPERAND (stmt, 1));
- else if (cond == boolean_false_node)
- pre_fold_expr = COND_EXPR_ELSE (TREE_OPERAND (stmt, 1));
- else
- pre_fold_expr = TREE_OPERAND (stmt, 1);
- }
- else
- pre_fold_expr = TREE_OPERAND (stmt, 1);
-
- if (pre_fold_expr)
- {
- cached_lhs = fold (pre_fold_expr);
- if (TREE_CODE (cached_lhs) != SSA_NAME
- && !is_gimple_min_invariant (cached_lhs))
- cached_lhs = lookup_avail_expr (stmt, false);
- }
-
- /* Restore the statement's original uses/defs. */
- i = 0;
- FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
- SET_USE (use_p, copy[i++]);
-
- free (copy);
- }
-
- /* Record the context sensitive equivalence if we were able
- to simplify this statement. */
- if (cached_lhs
- && (TREE_CODE (cached_lhs) == SSA_NAME
- || is_gimple_min_invariant (cached_lhs)))
- record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
- }
-
- /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
- will be taken. */
- if (stmt
- && (TREE_CODE (stmt) == COND_EXPR
- || TREE_CODE (stmt) == GOTO_EXPR
- || TREE_CODE (stmt) == SWITCH_EXPR))
- {
- tree cond, cached_lhs;
-
- /* Now temporarily cprop the operands and try to find the resulting
- expression in the hash tables. */
- if (TREE_CODE (stmt) == COND_EXPR)
- {
- canonicalize_comparison (stmt);
- cond = COND_EXPR_COND (stmt);
- }
- else if (TREE_CODE (stmt) == GOTO_EXPR)
- cond = GOTO_DESTINATION (stmt);
- else
- cond = SWITCH_COND (stmt);
-
- if (COMPARISON_CLASS_P (cond))
- {
- tree dummy_cond, op0, op1;
- enum tree_code cond_code;
-
- op0 = TREE_OPERAND (cond, 0);
- op1 = TREE_OPERAND (cond, 1);
- cond_code = TREE_CODE (cond);
-
- /* Get the current value of both operands. */
- if (TREE_CODE (op0) == SSA_NAME)
- {
- tree tmp = SSA_NAME_VALUE (op0);
- if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
- op0 = tmp;
- }
-
- if (TREE_CODE (op1) == SSA_NAME)
- {
- tree tmp = SSA_NAME_VALUE (op1);
- if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
- op1 = tmp;
- }
-
- /* Stuff the operator and operands into our dummy conditional
- expression, creating the dummy conditional if necessary. */
- dummy_cond = walk_data->global_data;
- if (! dummy_cond)
- {
- dummy_cond = build2 (cond_code, boolean_type_node, op0, op1);
- dummy_cond = build3 (COND_EXPR, void_type_node,
- dummy_cond, NULL_TREE, NULL_TREE);
- walk_data->global_data = dummy_cond;
- }
- else
- {
- TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
- TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
- TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
- }
-
- /* If the conditional folds to an invariant, then we are done,
- otherwise look it up in the hash tables. */
- cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
- if (! is_gimple_min_invariant (cached_lhs))
- {
- cached_lhs = lookup_avail_expr (dummy_cond, false);
- if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
- cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
- NULL,
- false);
- }
- }
- /* We can have conditionals which just test the state of a
- variable rather than use a relational operator. These are
- simpler to handle. */
- else if (TREE_CODE (cond) == SSA_NAME)
- {
- cached_lhs = cond;
- cached_lhs = SSA_NAME_VALUE (cached_lhs);
- if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
- cached_lhs = NULL;
- }
- else
- cached_lhs = lookup_avail_expr (stmt, false);
-
- if (cached_lhs)
- {
- edge taken_edge = find_taken_edge (e->dest, cached_lhs);
- basic_block dest = (taken_edge ? taken_edge->dest : NULL);
-
- if (dest == e->dest)
- return;
-
- /* If we have a known destination for the conditional, then
- we can perform this optimization, which saves at least one
- conditional jump each time it applies since we get to
- bypass the conditional at our original destination. */
- if (dest)
- {
- struct edge_info *edge_info;
-
- if (e->aux)
- edge_info = e->aux;
- else
- edge_info = allocate_edge_info (e);
- edge_info->redirection_target = taken_edge;
- bitmap_set_bit (threaded_blocks, e->dest->index);
- }
- }
- }
-}
-
/* Initialize local stacks for this optimizer and record equivalences
upon entry to BB. Equivalences can come from the edge traversed to
far to unwind when we finalize this block. */
VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
- VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
- VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
record_equivalences_from_incoming_edge (bb);
else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
{
element->stmt = expr;
- element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
+ element->rhs = GIMPLE_STMT_OPERAND (TREE_OPERAND (expr, 0), 1);
}
else if (TREE_CODE (expr) == GOTO_EXPR)
{
else
{
element->stmt = expr;
- element->rhs = TREE_OPERAND (expr, 1);
+ element->rhs = GENERIC_TREE_OPERAND (expr, 1);
}
element->lhs = lhs;
}
}
-/* Use the SSA_NAMES in LOCALS to restore TABLE to its original
- state, stopping when there are LIMIT entries left in LOCALs. */
-
-static void
-restore_nonzero_vars_to_original_value (void)
-{
- while (VEC_length (tree, nonzero_vars_stack) > 0)
- {
- tree name = VEC_pop (tree, nonzero_vars_stack);
-
- if (name == NULL)
- break;
-
- bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
- }
-}
-
/* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
CONST_AND_COPIES to its original state, stopping when we hit a
NULL marker. */
}
}
+/* A trivial wrapper so that we can present the generic jump
+ threading code with a simple API for simplifying statements. */
+static tree
+simplify_stmt_for_jump_threading (tree stmt, tree within_stmt ATTRIBUTE_UNUSED)
+{
+ return lookup_avail_expr (stmt, false);
+}
+
+/* Wrapper for common code to attempt to thread an edge. For example,
+ it handles lazily building the dummy condition and the bookkeeping
+ when jump threading is successful. */
+
+static void
+dom_thread_across_edge (struct dom_walk_data *walk_data, edge e)
+{
+ /* If we don't already have a dummy condition, build it now. */
+ if (! walk_data->global_data)
+ {
+ tree dummy_cond = build2 (NE_EXPR, boolean_type_node,
+ integer_zero_node, integer_zero_node);
+ dummy_cond = build3 (COND_EXPR, void_type_node, dummy_cond, NULL, NULL);
+ walk_data->global_data = dummy_cond;
+ }
+
+ thread_across_edge ((tree) walk_data->global_data, e, false,
+ &const_and_copies_stack,
+ simplify_stmt_for_jump_threading);
+}
+
/* We have finished processing the dominator children of BB, perform
any finalization actions in preparation for leaving this node in
the dominator tree. */
{
tree last;
+
/* If we have an outgoing edge to a block with multiple incoming and
outgoing edges, then we may be able to thread the edge. ie, we
may be able to statically determine which of the outgoing edges
will be traversed when the incoming edge from BB is traversed. */
if (single_succ_p (bb)
&& (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
- && !single_pred_p (single_succ (bb))
- && !single_succ_p (single_succ (bb)))
-
+ && potentially_threadable_block (single_succ (bb)))
{
- thread_across_edge (walk_data, single_succ_edge (bb));
+ dom_thread_across_edge (walk_data, single_succ_edge (bb));
}
else if ((last = last_stmt (bb))
&& TREE_CODE (last) == COND_EXPR
/* Only try to thread the edge if it reaches a target block with
more than one predecessor and more than one successor. */
- if (!single_pred_p (true_edge->dest) && !single_succ_p (true_edge->dest))
+ if (potentially_threadable_block (true_edge->dest))
{
struct edge_info *edge_info;
unsigned int i;
VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
- edge_info = true_edge->aux;
+ edge_info = (struct edge_info *) true_edge->aux;
/* If we have info associated with this edge, record it into
our equivalency tables. */
}
}
- /* Now thread the edge. */
- thread_across_edge (walk_data, true_edge);
+ dom_thread_across_edge (walk_data, true_edge);
/* And restore the various tables to their state before
we threaded this edge. */
remove_local_expressions_from_table ();
- restore_vars_to_original_value ();
}
/* Similarly for the ELSE arm. */
- if (!single_pred_p (false_edge->dest) && !single_succ_p (false_edge->dest))
+ if (potentially_threadable_block (false_edge->dest))
{
struct edge_info *edge_info;
unsigned int i;
- edge_info = false_edge->aux;
+ VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
+ edge_info = (struct edge_info *) false_edge->aux;
/* If we have info associated with this edge, record it into
our equivalency tables. */
}
}
- thread_across_edge (walk_data, false_edge);
+ /* Now thread the edge. */
+ dom_thread_across_edge (walk_data, false_edge);
/* No need to remove local expressions from our tables
or restore vars to their original value as that will
}
remove_local_expressions_from_table ();
- restore_nonzero_vars_to_original_value ();
restore_vars_to_original_value ();
- /* Remove VRP records associated with this basic block. They are no
- longer valid.
-
- To be efficient, we note which variables have had their values
- constrained in this block. So walk over each variable in the
- VRP_VARIABLEs array. */
- while (VEC_length (tree, vrp_variables_stack) > 0)
- {
- tree var = VEC_pop (tree, vrp_variables_stack);
- struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
- void **slot;
-
- /* Each variable has a stack of value range records. We want to
- invalidate those associated with our basic block. So we walk
- the array backwards popping off records associated with our
- block. Once we hit a record not associated with our block
- we are done. */
- VEC(vrp_element_p,heap) **var_vrp_records;
-
- if (var == NULL)
- break;
-
- vrp_hash_elt.var = var;
- vrp_hash_elt.records = NULL;
-
- slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
-
- vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
- var_vrp_records = &vrp_hash_elt_p->records;
-
- while (VEC_length (vrp_element_p, *var_vrp_records) > 0)
- {
- struct vrp_element *element
- = VEC_last (vrp_element_p, *var_vrp_records);
-
- if (element->bb != bb)
- break;
-
- VEC_pop (vrp_element_p, *var_vrp_records);
- }
- }
-
/* If we queued any statements to rescan in this block, then
go ahead and rescan them now. */
- while (VEC_length (tree, stmts_to_rescan) > 0)
+ while (VEC_length (tree_p, stmts_to_rescan) > 0)
{
- tree stmt = VEC_last (tree, stmts_to_rescan);
+ tree *stmt_p = VEC_last (tree_p, stmts_to_rescan);
+ tree stmt = *stmt_p;
basic_block stmt_bb = bb_for_stmt (stmt);
if (stmt_bb != bb)
break;
- VEC_pop (tree, stmts_to_rescan);
- mark_new_vars_to_rename (stmt);
+ VEC_pop (tree_p, stmts_to_rescan);
+ pop_stmt_changes (stmt_p);
}
}
Ignoring any alternatives which are the same as the result, if
all the alternatives are equal, then the PHI node creates an
- equivalence.
-
- Additionally, if all the PHI alternatives are known to have a nonzero
- value, then the result of this PHI is known to have a nonzero value,
- even if we do not know its exact value. */
+ equivalence. */
static void
record_equivalences_from_phis (basic_block bb)
if (i == PHI_NUM_ARGS (phi)
&& may_propagate_copy (lhs, rhs))
SSA_NAME_VALUE (lhs) = rhs;
-
- /* Now see if we know anything about the nonzero property for the
- result of this PHI. */
- for (i = 0; i < PHI_NUM_ARGS (phi); i++)
- {
- if (!PHI_ARG_NONZERO (phi, i))
- break;
- }
-
- if (i == PHI_NUM_ARGS (phi))
- bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
}
}
{
unsigned int i;
- edge_info = e->aux;
+ edge_info = (struct edge_info *) e->aux;
if (edge_info)
{
if (cond_equivalences)
{
- bool recorded_range = false;
for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
{
tree expr = cond_equivalences[i];
tree value = cond_equivalences[i + 1];
record_cond (expr, value);
-
- /* For the first true equivalence, record range
- information. We only do this for the first
- true equivalence as it should dominate any
- later true equivalences. */
- if (! recorded_range
- && COMPARISON_CLASS_P (expr)
- && value == boolean_true_node
- && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
- {
- record_range (expr, bb);
- recorded_range = true;
- }
}
}
}
fprintf (file, " Copies propagated: %6ld\n",
opt_stats.num_copy_prop);
- fprintf (file, "\nTotal number of DOM iterations: %6ld\n",
- opt_stats.num_iterations);
-
fprintf (file, "\nHash table statistics:\n");
fprintf (file, " avail_exprs: ");
htab_collisions (htab));
}
-/* Record the fact that VAR has a nonzero value, though we may not know
- its exact value. Note that if VAR is already known to have a nonzero
- value, then we do nothing. */
-
-static void
-record_var_is_nonzero (tree var)
-{
- int indx = SSA_NAME_VERSION (var);
-
- if (bitmap_bit_p (nonzero_vars, indx))
- return;
-
- /* Mark it in the global table. */
- bitmap_set_bit (nonzero_vars, indx);
-
- /* Record this SSA_NAME so that we can reset the global table
- when we leave this block. */
- VEC_safe_push (tree, heap, nonzero_vars_stack, var);
-}
-
/* Enter a statement into the true/false expression hash table indicating
that the condition COND has the value VALUE. */
static void
record_cond (tree cond, tree value)
{
- struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
+ struct expr_hash_elt *element = XCNEW (struct expr_hash_elt);
void **slot;
initialize_hash_element (cond, value, element);
{
case LT_EXPR:
case GT_EXPR:
- edge_info->max_cond_equivalences = 12;
- edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
+ if (FLOAT_TYPE_P (TREE_TYPE (op0)))
+ {
+ edge_info->max_cond_equivalences = 12;
+ edge_info->cond_equivalences = XNEWVEC (tree, 12);
+ build_and_record_new_cond (ORDERED_EXPR, op0, op1,
+ &edge_info->cond_equivalences[8]);
+ build_and_record_new_cond (LTGT_EXPR, op0, op1,
+ &edge_info->cond_equivalences[10]);
+ }
+ else
+ {
+ edge_info->max_cond_equivalences = 8;
+ edge_info->cond_equivalences = XNEWVEC (tree, 8);
+ }
+
build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
? LE_EXPR : GE_EXPR),
op0, op1, &edge_info->cond_equivalences[4]);
- build_and_record_new_cond (ORDERED_EXPR, op0, op1,
- &edge_info->cond_equivalences[6]);
build_and_record_new_cond (NE_EXPR, op0, op1,
- &edge_info->cond_equivalences[8]);
- build_and_record_new_cond (LTGT_EXPR, op0, op1,
- &edge_info->cond_equivalences[10]);
+ &edge_info->cond_equivalences[6]);
break;
case GE_EXPR:
case LE_EXPR:
- edge_info->max_cond_equivalences = 6;
- edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
- build_and_record_new_cond (ORDERED_EXPR, op0, op1,
- &edge_info->cond_equivalences[4]);
+ if (FLOAT_TYPE_P (TREE_TYPE (op0)))
+ {
+ edge_info->max_cond_equivalences = 6;
+ edge_info->cond_equivalences = XNEWVEC (tree, 6);
+ build_and_record_new_cond (ORDERED_EXPR, op0, op1,
+ &edge_info->cond_equivalences[4]);
+ }
+ else
+ {
+ edge_info->max_cond_equivalences = 4;
+ edge_info->cond_equivalences = XNEWVEC (tree, 4);
+ }
break;
case EQ_EXPR:
- edge_info->max_cond_equivalences = 10;
- edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
- build_and_record_new_cond (ORDERED_EXPR, op0, op1,
- &edge_info->cond_equivalences[4]);
+ if (FLOAT_TYPE_P (TREE_TYPE (op0)))
+ {
+ edge_info->max_cond_equivalences = 10;
+ edge_info->cond_equivalences = XNEWVEC (tree, 10);
+ build_and_record_new_cond (ORDERED_EXPR, op0, op1,
+ &edge_info->cond_equivalences[8]);
+ }
+ else
+ {
+ edge_info->max_cond_equivalences = 8;
+ edge_info->cond_equivalences = XNEWVEC (tree, 8);
+ }
build_and_record_new_cond (LE_EXPR, op0, op1,
- &edge_info->cond_equivalences[6]);
+ &edge_info->cond_equivalences[4]);
build_and_record_new_cond (GE_EXPR, op0, op1,
- &edge_info->cond_equivalences[8]);
+ &edge_info->cond_equivalences[6]);
break;
case UNORDERED_EXPR:
edge_info->max_cond_equivalences = 16;
- edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
+ edge_info->cond_equivalences = XNEWVEC (tree, 16);
build_and_record_new_cond (NE_EXPR, op0, op1,
&edge_info->cond_equivalences[4]);
build_and_record_new_cond (UNLE_EXPR, op0, op1,
case UNLT_EXPR:
case UNGT_EXPR:
edge_info->max_cond_equivalences = 8;
- edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
+ edge_info->cond_equivalences = XNEWVEC (tree, 8);
build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
? UNLE_EXPR : UNGE_EXPR),
op0, op1, &edge_info->cond_equivalences[4]);
case UNEQ_EXPR:
edge_info->max_cond_equivalences = 8;
- edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
+ edge_info->cond_equivalences = XNEWVEC (tree, 8);
build_and_record_new_cond (UNLE_EXPR, op0, op1,
&edge_info->cond_equivalences[4]);
build_and_record_new_cond (UNGE_EXPR, op0, op1,
case LTGT_EXPR:
edge_info->max_cond_equivalences = 8;
- edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
+ edge_info->cond_equivalences = XNEWVEC (tree, 8);
build_and_record_new_cond (NE_EXPR, op0, op1,
&edge_info->cond_equivalences[4]);
build_and_record_new_cond (ORDERED_EXPR, op0, op1,
default:
edge_info->max_cond_equivalences = 4;
- edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
+ edge_info->cond_equivalences = XNEWVEC (tree, 4);
break;
}
record_const_or_copy_1 (x, y, prev_x);
}
-/* Return true, if it is ok to do folding of an associative expression.
- EXP is the tree for the associative expression. */
-
-static inline bool
-unsafe_associative_fp_binop (tree exp)
-{
- enum tree_code code = TREE_CODE (exp);
- return !(!flag_unsafe_math_optimizations
- && (code == MULT_EXPR || code == PLUS_EXPR
- || code == MINUS_EXPR)
- && FLOAT_TYPE_P (TREE_TYPE (exp)));
-}
-
/* Returns true when STMT is a simple iv increment. It detects the
following situation:
tree lhs, rhs, preinc, phi;
unsigned i;
- if (TREE_CODE (stmt) != MODIFY_EXPR)
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- lhs = TREE_OPERAND (stmt, 0);
+ lhs = GIMPLE_STMT_OPERAND (stmt, 0);
if (TREE_CODE (lhs) != SSA_NAME)
return false;
- rhs = TREE_OPERAND (stmt, 1);
+ rhs = GIMPLE_STMT_OPERAND (stmt, 1);
if (TREE_CODE (rhs) != PLUS_EXPR
&& TREE_CODE (rhs) != MINUS_EXPR)
return false;
}
-/* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
- hash tables. Try to simplify the RHS using whatever equivalences
- we may have recorded.
+/* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
+ known value for that SSA_NAME (or NULL if no value is known).
- If we are able to simplify the RHS, then lookup the simplified form in
- the hash table and return the result. Otherwise return NULL. */
+ Propagate values from CONST_AND_COPIES into the PHI nodes of the
+ successors of BB. */
-static tree
-simplify_rhs_and_lookup_avail_expr (tree stmt, int insert)
+static void
+cprop_into_successor_phis (basic_block bb)
{
- tree rhs = TREE_OPERAND (stmt, 1);
- enum tree_code rhs_code = TREE_CODE (rhs);
- tree result = NULL;
-
- /* If we have lhs = ~x, look and see if we earlier had x = ~y.
- In which case we can change this statement to be lhs = y.
- Which can then be copy propagated.
+ edge e;
+ edge_iterator ei;
- Similarly for negation. */
- if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
- && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
+ FOR_EACH_EDGE (e, ei, bb->succs)
{
- /* Get the definition statement for our RHS. */
- tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
-
- /* See if the RHS_DEF_STMT has the same form as our statement. */
- if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
- && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
- {
- tree rhs_def_operand;
-
- rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
-
- /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
- if (TREE_CODE (rhs_def_operand) == SSA_NAME
- && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
- result = update_rhs_and_lookup_avail_expr (stmt,
- rhs_def_operand,
- insert);
- }
- }
-
- /* Optimize *"foo" into 'f'. This is done here rather than
- in fold to avoid problems with stuff like &*"foo". */
- if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
- {
- tree t = fold_read_from_constant_string (rhs);
-
- if (t)
- result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
- }
-
- return result;
-}
-
-/* COND is a condition of the form:
-
- x == const or x != const
-
- Look back to x's defining statement and see if x is defined as
-
- x = (type) y;
-
- If const is unchanged if we convert it to type, then we can build
- the equivalent expression:
-
-
- y == const or y != const
-
- Which may allow further optimizations.
-
- Return the equivalent comparison or NULL if no such equivalent comparison
- was found. */
-
-static tree
-find_equivalent_equality_comparison (tree cond)
-{
- tree op0 = TREE_OPERAND (cond, 0);
- tree op1 = TREE_OPERAND (cond, 1);
- tree def_stmt = SSA_NAME_DEF_STMT (op0);
-
- /* OP0 might have been a parameter, so first make sure it
- was defined by a MODIFY_EXPR. */
- if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
- {
- tree def_rhs = TREE_OPERAND (def_stmt, 1);
-
-
- /* If either operand to the comparison is a pointer to
- a function, then we can not apply this optimization
- as some targets require function pointers to be
- canonicalized and in this case this optimization would
- eliminate a necessary canonicalization. */
- if ((POINTER_TYPE_P (TREE_TYPE (op0))
- && TREE_CODE (TREE_TYPE (TREE_TYPE (op0))) == FUNCTION_TYPE)
- || (POINTER_TYPE_P (TREE_TYPE (op1))
- && TREE_CODE (TREE_TYPE (TREE_TYPE (op1))) == FUNCTION_TYPE))
- return NULL;
-
- /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
- if ((TREE_CODE (def_rhs) == NOP_EXPR
- || TREE_CODE (def_rhs) == CONVERT_EXPR)
- && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
- {
- tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
- tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
- tree new;
-
- if (TYPE_PRECISION (def_rhs_inner_type)
- > TYPE_PRECISION (TREE_TYPE (def_rhs)))
- return NULL;
-
- /* If the inner type of the conversion is a pointer to
- a function, then we can not apply this optimization
- as some targets require function pointers to be
- canonicalized. This optimization would result in
- canonicalization of the pointer when it was not originally
- needed/intended. */
- if (POINTER_TYPE_P (def_rhs_inner_type)
- && TREE_CODE (TREE_TYPE (def_rhs_inner_type)) == FUNCTION_TYPE)
- return NULL;
-
- /* What we want to prove is that if we convert OP1 to
- the type of the object inside the NOP_EXPR that the
- result is still equivalent to SRC.
-
- If that is true, the build and return new equivalent
- condition which uses the source of the typecast and the
- new constant (which has only changed its type). */
- new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
- new = local_fold (new);
- if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
- return build2 (TREE_CODE (cond), TREE_TYPE (cond),
- def_rhs_inner, new);
- }
- }
- return NULL;
-}
-
-/* STMT is a COND_EXPR for which we could not trivially determine its
- result. This routine attempts to find equivalent forms of the
- condition which we may be able to optimize better. It also
- uses simple value range propagation to optimize conditionals. */
-
-static tree
-simplify_cond_and_lookup_avail_expr (tree stmt,
- stmt_ann_t ann,
- int insert)
-{
- tree cond = COND_EXPR_COND (stmt);
-
- if (COMPARISON_CLASS_P (cond))
- {
- tree op0 = TREE_OPERAND (cond, 0);
- tree op1 = TREE_OPERAND (cond, 1);
-
- if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
- {
- int limit;
- tree low, high, cond_low, cond_high;
- int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
- VEC(vrp_element_p,heap) **vrp_records;
- struct vrp_element *element;
- struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
- void **slot;
-
- /* First see if we have test of an SSA_NAME against a constant
- where the SSA_NAME is defined by an earlier typecast which
- is irrelevant when performing tests against the given
- constant. */
- if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
- {
- tree new_cond = find_equivalent_equality_comparison (cond);
-
- if (new_cond)
- {
- /* Update the statement to use the new equivalent
- condition. */
- COND_EXPR_COND (stmt) = new_cond;
-
- /* If this is not a real stmt, ann will be NULL and we
- avoid processing the operands. */
- if (ann)
- mark_stmt_modified (stmt);
-
- /* Lookup the condition and return its known value if it
- exists. */
- new_cond = lookup_avail_expr (stmt, insert);
- if (new_cond)
- return new_cond;
-
- /* The operands have changed, so update op0 and op1. */
- op0 = TREE_OPERAND (cond, 0);
- op1 = TREE_OPERAND (cond, 1);
- }
- }
-
- /* Consult the value range records for this variable (if they exist)
- to see if we can eliminate or simplify this conditional.
-
- Note two tests are necessary to determine no records exist.
- First we have to see if the virtual array exists, if it
- exists, then we have to check its active size.
-
- Also note the vast majority of conditionals are not testing
- a variable which has had its range constrained by an earlier
- conditional. So this filter avoids a lot of unnecessary work. */
- vrp_hash_elt.var = op0;
- vrp_hash_elt.records = NULL;
- slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
- if (slot == NULL)
- return NULL;
-
- vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
- vrp_records = &vrp_hash_elt_p->records;
-
- limit = VEC_length (vrp_element_p, *vrp_records);
-
- /* If we have no value range records for this variable, or we are
- unable to extract a range for this condition, then there is
- nothing to do. */
- if (limit == 0
- || ! extract_range_from_cond (cond, &cond_high,
- &cond_low, &cond_inverted))
- return NULL;
-
- /* We really want to avoid unnecessary computations of range
- info. So all ranges are computed lazily; this avoids a
- lot of unnecessary work. i.e., we record the conditional,
- but do not process how it constrains the variable's
- potential values until we know that processing the condition
- could be helpful.
-
- However, we do not want to have to walk a potentially long
- list of ranges, nor do we want to compute a variable's
- range more than once for a given path.
-
- Luckily, each time we encounter a conditional that can not
- be otherwise optimized we will end up here and we will
- compute the necessary range information for the variable
- used in this condition.
-
- Thus you can conclude that there will never be more than one
- conditional associated with a variable which has not been
- processed. So we never need to merge more than one new
- conditional into the current range.
-
- These properties also help us avoid unnecessary work. */
- element = VEC_last (vrp_element_p, *vrp_records);
-
- if (element->high && element->low)
- {
- /* The last element has been processed, so there is no range
- merging to do, we can simply use the high/low values
- recorded in the last element. */
- low = element->low;
- high = element->high;
- }
- else
- {
- tree tmp_high, tmp_low;
- int dummy;
-
- /* The last element has not been processed. Process it now.
- record_range should ensure for cond inverted is not set.
- This call can only fail if cond is x < min or x > max,
- which fold should have optimized into false.
- If that doesn't happen, just pretend all values are
- in the range. */
- if (! extract_range_from_cond (element->cond, &tmp_high,
- &tmp_low, &dummy))
- gcc_unreachable ();
- else
- gcc_assert (dummy == 0);
-
- /* If this is the only element, then no merging is necessary,
- the high/low values from extract_range_from_cond are all
- we need. */
- if (limit == 1)
- {
- low = tmp_low;
- high = tmp_high;
- }
- else
- {
- /* Get the high/low value from the previous element. */
- struct vrp_element *prev
- = VEC_index (vrp_element_p, *vrp_records, limit - 2);
- low = prev->low;
- high = prev->high;
-
- /* Merge in this element's range with the range from the
- previous element.
-
- The low value for the merged range is the maximum of
- the previous low value and the low value of this record.
-
- Similarly the high value for the merged range is the
- minimum of the previous high value and the high value of
- this record. */
- low = (low && tree_int_cst_compare (low, tmp_low) == 1
- ? low : tmp_low);
- high = (high && tree_int_cst_compare (high, tmp_high) == -1
- ? high : tmp_high);
- }
-
- /* And record the computed range. */
- element->low = low;
- element->high = high;
-
- }
-
- /* After we have constrained this variable's potential values,
- we try to determine the result of the given conditional.
-
- To simplify later tests, first determine if the current
- low value is the same low value as the conditional.
- Similarly for the current high value and the high value
- for the conditional. */
- lowequal = tree_int_cst_equal (low, cond_low);
- highequal = tree_int_cst_equal (high, cond_high);
-
- if (lowequal && highequal)
- return (cond_inverted ? boolean_false_node : boolean_true_node);
-
- /* To simplify the overlap/subset tests below we may want
- to swap the two ranges so that the larger of the two
- ranges occurs "first". */
- swapped = 0;
- if (tree_int_cst_compare (low, cond_low) == 1
- || (lowequal
- && tree_int_cst_compare (cond_high, high) == 1))
- {
- tree temp;
-
- swapped = 1;
- temp = low;
- low = cond_low;
- cond_low = temp;
- temp = high;
- high = cond_high;
- cond_high = temp;
- }
-
- /* Now determine if there is no overlap in the ranges
- or if the second range is a subset of the first range. */
- no_overlap = tree_int_cst_lt (high, cond_low);
- subset = tree_int_cst_compare (cond_high, high) != 1;
-
- /* If there was no overlap in the ranges, then this conditional
- always has a false value (unless we had to invert this
- conditional, in which case it always has a true value). */
- if (no_overlap)
- return (cond_inverted ? boolean_true_node : boolean_false_node);
-
- /* If the current range is a subset of the condition's range,
- then this conditional always has a true value (unless we
- had to invert this conditional, in which case it always
- has a true value). */
- if (subset && swapped)
- return (cond_inverted ? boolean_false_node : boolean_true_node);
-
- /* We were unable to determine the result of the conditional.
- However, we may be able to simplify the conditional. First
- merge the ranges in the same manner as range merging above. */
- low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
- high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
-
- /* If the range has converged to a single point, then turn this
- into an equality comparison. */
- if (TREE_CODE (cond) != EQ_EXPR
- && TREE_CODE (cond) != NE_EXPR
- && tree_int_cst_equal (low, high))
- {
- TREE_SET_CODE (cond, EQ_EXPR);
- TREE_OPERAND (cond, 1) = high;
- }
- }
- }
- return 0;
-}
-
-/* STMT is a SWITCH_EXPR for which we could not trivially determine its
- result. This routine attempts to find equivalent forms of the
- condition which we may be able to optimize better. */
-
-static tree
-simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
-{
- tree cond = SWITCH_COND (stmt);
- tree def, to, ti;
-
- /* The optimization that we really care about is removing unnecessary
- casts. That will let us do much better in propagating the inferred
- constant at the switch target. */
- if (TREE_CODE (cond) == SSA_NAME)
- {
- def = SSA_NAME_DEF_STMT (cond);
- if (TREE_CODE (def) == MODIFY_EXPR)
- {
- def = TREE_OPERAND (def, 1);
- if (TREE_CODE (def) == NOP_EXPR)
- {
- int need_precision;
- bool fail;
-
- def = TREE_OPERAND (def, 0);
-
-#ifdef ENABLE_CHECKING
- /* ??? Why was Jeff testing this? We are gimple... */
- gcc_assert (is_gimple_val (def));
-#endif
-
- to = TREE_TYPE (cond);
- ti = TREE_TYPE (def);
-
- /* If we have an extension that preserves value, then we
- can copy the source value into the switch. */
-
- need_precision = TYPE_PRECISION (ti);
- fail = false;
- if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
- fail = true;
- else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
- need_precision += 1;
- if (TYPE_PRECISION (to) < need_precision)
- fail = true;
-
- if (!fail)
- {
- SWITCH_COND (stmt) = def;
- mark_stmt_modified (stmt);
-
- return lookup_avail_expr (stmt, insert);
- }
- }
- }
- }
-
- return 0;
-}
-
-
-/* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
- known value for that SSA_NAME (or NULL if no value is known).
-
- NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
- even if we don't know their precise value.
-
- Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
- nodes of the successors of BB. */
-
-static void
-cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
-{
- edge e;
- edge_iterator ei;
-
- FOR_EACH_EDGE (e, ei, bb->succs)
- {
- tree phi;
- int indx;
+ tree phi;
+ int indx;
/* If this is an abnormal edge, then we do not want to copy propagate
into the PHI alternative associated with this edge. */
indx = e->dest_idx;
for ( ; phi; phi = PHI_CHAIN (phi))
{
- tree new;
+ tree new_val;
use_operand_p orig_p;
- tree orig;
+ tree orig_val;
/* The alternative may be associated with a constant, so verify
it is an SSA_NAME before doing anything with it. */
orig_p = PHI_ARG_DEF_PTR (phi, indx);
- orig = USE_FROM_PTR (orig_p);
- if (TREE_CODE (orig) != SSA_NAME)
+ orig_val = USE_FROM_PTR (orig_p);
+ if (TREE_CODE (orig_val) != SSA_NAME)
continue;
- /* If the alternative is known to have a nonzero value, record
- that fact in the PHI node itself for future use. */
- if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
- PHI_ARG_NONZERO (phi, indx) = true;
-
/* If we have *ORIG_P in our constant/copy table, then replace
ORIG_P with its value in our constant/copy table. */
- new = SSA_NAME_VALUE (orig);
- if (new
- && new != orig
- && (TREE_CODE (new) == SSA_NAME
- || is_gimple_min_invariant (new))
- && may_propagate_copy (orig, new))
- propagate_value (orig_p, new);
+ new_val = SSA_NAME_VALUE (orig_val);
+ if (new_val
+ && new_val != orig_val
+ && (TREE_CODE (new_val) == SSA_NAME
+ || is_gimple_min_invariant (new_val))
+ && may_propagate_copy (orig_val, new_val))
+ propagate_value (orig_p, new_val);
}
}
}
{
tree labels = SWITCH_LABELS (stmt);
int i, n_labels = TREE_VEC_LENGTH (labels);
- tree *info = xcalloc (last_basic_block, sizeof (tree));
+ tree *info = XCNEWVEC (tree, last_basic_block);
edge e;
edge_iterator ei;
basic_block bb)
{
record_edge_info (bb);
- cprop_into_successor_phis (bb, nonzero_vars);
+ cprop_into_successor_phis (bb);
}
/* Search for redundant computations in STMT. If any are found, then
table. */
static bool
-eliminate_redundant_computations (tree stmt, stmt_ann_t ann)
+eliminate_redundant_computations (tree stmt)
{
tree *expr_p, def = NULL_TREE;
bool insert = true;
bool retval = false;
bool modify_expr_p = false;
- if (TREE_CODE (stmt) == MODIFY_EXPR)
- def = TREE_OPERAND (stmt, 0);
+ if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
+ def = GIMPLE_STMT_OPERAND (stmt, 0);
/* Certain expressions on the RHS can be optimized away, but can not
themselves be entered into the hash tables. */
- if (ann->makes_aliased_stores
- || ! def
+ if (! def
|| TREE_CODE (def) != SSA_NAME
|| SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
- || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
+ || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VDEF)
/* Do not record equivalences for increments of ivs. This would create
overlapping live ranges for a very questionable gain. */
|| simple_iv_increment_p (stmt))
/* Check if the expression has been computed before. */
cached_lhs = lookup_avail_expr (stmt, insert);
- /* If this is an assignment and the RHS was not in the hash table,
- then try to simplify the RHS and lookup the new RHS in the
- hash table. */
- if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
- cached_lhs = simplify_rhs_and_lookup_avail_expr (stmt, insert);
- /* Similarly if this is a COND_EXPR and we did not find its
- expression in the hash table, simplify the condition and
- try again. */
- else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
- cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
- /* Similarly for a SWITCH_EXPR. */
- else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
- cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
-
opt_stats.num_exprs_considered++;
/* Get a pointer to the expression we are trying to optimize. */
expr_p = &SWITCH_COND (stmt);
else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
{
- expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
+ expr_p = &GIMPLE_STMT_OPERAND (TREE_OPERAND (stmt, 0), 1);
modify_expr_p = true;
}
else
{
- expr_p = &TREE_OPERAND (stmt, 1);
+ expr_p = &GENERIC_TREE_OPERAND (stmt, 1);
modify_expr_p = true;
}
if (cached_lhs
&& ((TREE_CODE (cached_lhs) != SSA_NAME
&& (modify_expr_p
- || tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
- TREE_TYPE (cached_lhs))))
+ || useless_type_conversion_p (TREE_TYPE (*expr_p),
+ TREE_TYPE (cached_lhs))))
|| may_propagate_copy (*expr_p, cached_lhs)))
{
if (dump_file && (dump_flags & TDF_DETAILS))
retval = true;
if (modify_expr_p
- && !tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
- TREE_TYPE (cached_lhs)))
+ && !useless_type_conversion_p (TREE_TYPE (*expr_p),
+ TREE_TYPE (cached_lhs)))
cached_lhs = fold_convert (TREE_TYPE (*expr_p), cached_lhs);
propagate_tree_value (expr_p, cached_lhs);
return retval;
}
-/* STMT, a MODIFY_EXPR, may create certain equivalences, in either
+/* STMT, a GIMPLE_MODIFY_STMT, may create certain equivalences, in either
the available expressions table or the const_and_copies table.
Detect and record those equivalences. */
static void
-record_equivalences_from_stmt (tree stmt,
- int may_optimize_p,
- stmt_ann_t ann)
+record_equivalences_from_stmt (tree stmt, int may_optimize_p, stmt_ann_t ann)
{
- tree lhs = TREE_OPERAND (stmt, 0);
+ tree lhs = GIMPLE_STMT_OPERAND (stmt, 0);
enum tree_code lhs_code = TREE_CODE (lhs);
- int i;
if (lhs_code == SSA_NAME)
{
- tree rhs = TREE_OPERAND (stmt, 1);
+ tree rhs = GIMPLE_STMT_OPERAND (stmt, 1);
/* Strip away any useless type conversions. */
STRIP_USELESS_TYPE_CONVERSION (rhs);
&& (TREE_CODE (rhs) == SSA_NAME
|| is_gimple_min_invariant (rhs)))
SSA_NAME_VALUE (lhs) = rhs;
-
- if (tree_expr_nonzero_p (rhs))
- record_var_is_nonzero (lhs);
}
- /* Look at both sides for pointer dereferences. If we find one, then
- the pointer must be nonnull and we can enter that equivalence into
- the hash tables. */
- if (flag_delete_null_pointer_checks)
- for (i = 0; i < 2; i++)
- {
- tree t = TREE_OPERAND (stmt, i);
-
- /* Strip away any COMPONENT_REFs. */
- while (TREE_CODE (t) == COMPONENT_REF)
- t = TREE_OPERAND (t, 0);
-
- /* Now see if this is a pointer dereference. */
- if (INDIRECT_REF_P (t))
- {
- tree op = TREE_OPERAND (t, 0);
-
- /* If the pointer is a SSA variable, then enter new
- equivalences into the hash table. */
- while (TREE_CODE (op) == SSA_NAME)
- {
- tree def = SSA_NAME_DEF_STMT (op);
-
- record_var_is_nonzero (op);
-
- /* And walk up the USE-DEF chains noting other SSA_NAMEs
- which are known to have a nonzero value. */
- if (def
- && TREE_CODE (def) == MODIFY_EXPR
- && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
- op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
- else
- break;
- }
- }
- }
-
/* A memory store, even an aliased store, creates a useful
equivalence. By exchanging the LHS and RHS, creating suitable
vops and recording the result in the available expression table,
we may be able to expose more redundant loads. */
if (!ann->has_volatile_ops
- && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
- || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
+ && stmt_references_memory_p (stmt)
+ && (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1)) == SSA_NAME
+ || is_gimple_min_invariant (GIMPLE_STMT_OPERAND (stmt, 1)))
&& !is_gimple_reg (lhs))
{
- tree rhs = TREE_OPERAND (stmt, 1);
- tree new;
+ tree rhs = GIMPLE_STMT_OPERAND (stmt, 1);
+ tree new_stmt;
/* FIXME: If the LHS of the assignment is a bitfield and the RHS
is a constant, we need to adjust the constant to fit into the
if (rhs)
{
/* Build a new statement with the RHS and LHS exchanged. */
- new = build2 (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
+ new_stmt = build_gimple_modify_stmt (rhs, lhs);
- create_ssa_artficial_load_stmt (new, stmt);
+ create_ssa_artificial_load_stmt (new_stmt, stmt);
/* Finally enter the statement into the available expression
table. */
- lookup_avail_expr (new, true);
+ lookup_avail_expr (new_stmt, true);
}
}
}
propagation opportunity. */
if (TREE_CODE (val) != SSA_NAME)
{
- if (!lang_hooks.types_compatible_p (op_type, val_type))
+ if (!useless_type_conversion_p (op_type, val_type))
{
val = fold_convert (TREE_TYPE (op), val);
if (!is_gimple_min_invariant (val))
known value for that SSA_NAME (or NULL if no value is known).
Propagate values from CONST_AND_COPIES into the uses, vuses and
- v_may_def_ops of STMT. */
+ vdef_ops of STMT. */
static bool
cprop_into_stmt (tree stmt)
ann = stmt_ann (stmt);
opt_stats.num_stmts++;
may_have_exposed_new_symbols = false;
+ push_stmt_changes (bsi_stmt_ptr (si));
if (dump_file && (dump_flags & TDF_DETAILS))
{
print_generic_stmt (dump_file, stmt, TDF_SLIM);
}
- /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
+ /* Const/copy propagate into USES, VUSES and the RHS of VDEFs. */
may_have_exposed_new_symbols = cprop_into_stmt (stmt);
/* If the statement has been modified with constant replacements,
may_optimize_p = (!ann->has_volatile_ops
&& ((TREE_CODE (stmt) == RETURN_EXPR
&& TREE_OPERAND (stmt, 0)
- && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
+ && TREE_CODE (TREE_OPERAND (stmt, 0))
+ == GIMPLE_MODIFY_STMT
&& ! (TREE_SIDE_EFFECTS
- (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
- || (TREE_CODE (stmt) == MODIFY_EXPR
- && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
+ (GIMPLE_STMT_OPERAND
+ (TREE_OPERAND (stmt, 0), 1))))
+ || (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
+ && ! TREE_SIDE_EFFECTS (GIMPLE_STMT_OPERAND (stmt,
+ 1)))
|| TREE_CODE (stmt) == COND_EXPR
|| TREE_CODE (stmt) == SWITCH_EXPR));
if (may_optimize_p)
- may_have_exposed_new_symbols
- |= eliminate_redundant_computations (stmt, ann);
+ may_have_exposed_new_symbols |= eliminate_redundant_computations (stmt);
/* Record any additional equivalences created by this statement. */
- if (TREE_CODE (stmt) == MODIFY_EXPR)
- record_equivalences_from_stmt (stmt,
- may_optimize_p,
- ann);
+ if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
+ record_equivalences_from_stmt (stmt, may_optimize_p, ann);
/* If STMT is a COND_EXPR and it was modified, then we may know
where it goes. If that is the case, then mark the CFG as altered.
Ultimately I suspect we're going to need to change the interface
into the SSA_NAME manager. */
-
if (ann->modified)
{
tree val = NULL;
}
if (may_have_exposed_new_symbols)
- VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
-}
-
-/* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
- available expression hashtable, then return the LHS from the hash
- table.
-
- If INSERT is true, then we also update the available expression
- hash table to account for the changes made to STMT. */
-
-static tree
-update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
-{
- tree cached_lhs = NULL;
-
- /* Remove the old entry from the hash table. */
- if (insert)
{
- struct expr_hash_elt element;
-
- initialize_hash_element (stmt, NULL, &element);
- htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
+ /* Queue the statement to be re-scanned after all the
+ AVAIL_EXPRS have been processed. The change buffer stack for
+ all the pushed statements will be processed when this queue
+ is emptied. */
+ VEC_safe_push (tree_p, heap, stmts_to_rescan, bsi_stmt_ptr (si));
+ }
+ else
+ {
+ /* Otherwise, just discard the recently pushed change buffer. If
+ not, the STMTS_TO_RESCAN queue will get out of synch with the
+ change buffer stack. */
+ discard_stmt_changes (bsi_stmt_ptr (si));
}
-
- /* Now update the RHS of the assignment. */
- TREE_OPERAND (stmt, 1) = new_rhs;
-
- /* Now lookup the updated statement in the hash table. */
- cached_lhs = lookup_avail_expr (stmt, insert);
-
- /* We have now called lookup_avail_expr twice with two different
- versions of this same statement, once in optimize_stmt, once here.
-
- We know the call in optimize_stmt did not find an existing entry
- in the hash table, so a new entry was created. At the same time
- this statement was pushed onto the AVAIL_EXPRS_STACK vector.
-
- If this call failed to find an existing entry on the hash table,
- then the new version of this statement was entered into the
- hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
- for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
-
- If this call succeeded, we still have one copy of this statement
- on the BLOCK_AVAIL_EXPRs vector.
-
- For both cases, we need to pop the most recent entry off the
- BLOCK_AVAIL_EXPRs vector. For the case where we never found this
- statement in the hash tables, that will leave precisely one
- copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
- we found a copy of this statement in the second hash table lookup
- we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
- if (insert)
- VEC_pop (tree, avail_exprs_stack);
-
- /* And make sure we record the fact that we modified this
- statement. */
- mark_stmt_modified (stmt);
-
- return cached_lhs;
}
/* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
is also added to the stack pointed to by BLOCK_AVAIL_EXPRS_P, so that they
can be removed when we finish processing this block and its children.
- NOTE: This function assumes that STMT is a MODIFY_EXPR node that
+ NOTE: This function assumes that STMT is a GIMPLE_MODIFY_STMT node that
contains no CALL_EXPR on its RHS and makes no volatile nor
aliased references. */
void **slot;
tree lhs;
tree temp;
- struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
+ struct expr_hash_elt *element = XNEW (struct expr_hash_elt);
- lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
+ lhs = TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
+ ? GIMPLE_STMT_OPERAND (stmt, 0) : NULL;
initialize_hash_element (stmt, lhs, element);
return NULL_TREE;
}
- /* If this is an equality test against zero, see if we have recorded a
- nonzero value for the variable in question. */
- if ((TREE_CODE (element->rhs) == EQ_EXPR
- || TREE_CODE (element->rhs) == NE_EXPR)
- && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
- && integer_zerop (TREE_OPERAND (element->rhs, 1)))
- {
- int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
-
- if (bitmap_bit_p (nonzero_vars, indx))
- {
- tree t = element->rhs;
- free (element);
- return constant_boolean_node (TREE_CODE (t) != EQ_EXPR,
- TREE_TYPE (t));
- }
- }
-
/* Finally try to find the expression in the main expression hash table. */
slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
(insert ? INSERT : NO_INSERT));
return lhs;
}
-/* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
- range of values that result in the conditional having a true value.
-
- Return true if we are successful in extracting a range from COND and
- false if we are unsuccessful. */
-
-static bool
-extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
-{
- tree op1 = TREE_OPERAND (cond, 1);
- tree high, low, type;
- int inverted;
-
- type = TREE_TYPE (op1);
-
- /* Experiments have shown that it's rarely, if ever useful to
- record ranges for enumerations. Presumably this is due to
- the fact that they're rarely used directly. They are typically
- cast into an integer type and used that way. */
- if (TREE_CODE (type) != INTEGER_TYPE)
- return 0;
-
- switch (TREE_CODE (cond))
- {
- case EQ_EXPR:
- high = low = op1;
- inverted = 0;
- break;
-
- case NE_EXPR:
- high = low = op1;
- inverted = 1;
- break;
-
- case GE_EXPR:
- low = op1;
-
- /* Get the highest value of the type. If not a constant, use that
- of its base type, if it has one. */
- high = TYPE_MAX_VALUE (type);
- if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type))
- high = TYPE_MAX_VALUE (TREE_TYPE (type));
- inverted = 0;
- break;
-
- case GT_EXPR:
- high = TYPE_MAX_VALUE (type);
- if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type))
- high = TYPE_MAX_VALUE (TREE_TYPE (type));
- if (!tree_int_cst_lt (op1, high))
- return 0;
- low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
- inverted = 0;
- break;
-
- case LE_EXPR:
- high = op1;
- low = TYPE_MIN_VALUE (type);
- if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type))
- low = TYPE_MIN_VALUE (TREE_TYPE (type));
- inverted = 0;
- break;
-
- case LT_EXPR:
- low = TYPE_MIN_VALUE (type);
- if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type))
- low = TYPE_MIN_VALUE (TREE_TYPE (type));
- if (!tree_int_cst_lt (low, op1))
- return 0;
- high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
- inverted = 0;
- break;
-
- default:
- return 0;
- }
-
- *hi_p = high;
- *lo_p = low;
- *inverted_p = inverted;
- return 1;
-}
-
-/* Record a range created by COND for basic block BB. */
-
-static void
-record_range (tree cond, basic_block bb)
-{
- enum tree_code code = TREE_CODE (cond);
-
- /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
- They rarely allow for meaningful range optimizations and significantly
- complicate the implementation. */
- if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
- || code == GE_EXPR || code == EQ_EXPR)
- && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
- {
- struct vrp_hash_elt *vrp_hash_elt;
- struct vrp_element *element;
- VEC(vrp_element_p,heap) **vrp_records_p;
- void **slot;
-
-
- vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
- vrp_hash_elt->var = TREE_OPERAND (cond, 0);
- vrp_hash_elt->records = NULL;
- slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
-
- if (*slot == NULL)
- *slot = (void *) vrp_hash_elt;
- else
- vrp_free (vrp_hash_elt);
-
- vrp_hash_elt = (struct vrp_hash_elt *) *slot;
- vrp_records_p = &vrp_hash_elt->records;
-
- element = ggc_alloc (sizeof (struct vrp_element));
- element->low = NULL;
- element->high = NULL;
- element->cond = cond;
- element->bb = bb;
-
- VEC_safe_push (vrp_element_p, heap, *vrp_records_p, element);
- VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
- }
-}
-
-/* Hashing and equality functions for VRP_DATA.
-
- Since this hash table is addressed by SSA_NAMEs, we can hash on
- their version number and equality can be determined with a
- pointer comparison. */
-
-static hashval_t
-vrp_hash (const void *p)
-{
- tree var = ((struct vrp_hash_elt *)p)->var;
-
- return SSA_NAME_VERSION (var);
-}
-
-static int
-vrp_eq (const void *p1, const void *p2)
-{
- tree var1 = ((struct vrp_hash_elt *)p1)->var;
- tree var2 = ((struct vrp_hash_elt *)p2)->var;
-
- return var1 == var2;
-}
-
/* Hashing and equality functions for AVAIL_EXPRS. The table stores
- MODIFY_EXPR statements. We compute a value number for expressions using
- the code of the expression and the SSA numbers of its operands. */
+ GIMPLE_MODIFY_STMT statements. We compute a value number for expressions
+ using the code of the expression and the SSA numbers of its operands. */
static hashval_t
avail_expr_hash (const void *p)
/* In case of a collision, both RHS have to be identical and have the
same VUSE operands. */
- if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
- || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
+ if (types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2))
&& operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
{
bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
return false;
}
+
+/* PHI-ONLY copy and constant propagation. This pass is meant to clean
+ up degenerate PHIs created by or exposed by jump threading. */
+
+/* Given PHI, return its RHS if the PHI is a degenerate, otherwise return
+ NULL. */
+
+static tree
+degenerate_phi_result (tree phi)
+{
+ tree lhs = PHI_RESULT (phi);
+ tree val = NULL;
+ int i;
+
+ /* Ignoring arguments which are the same as LHS, if all the remaining
+ arguments are the same, then the PHI is a degenerate and has the
+ value of that common argument. */
+ for (i = 0; i < PHI_NUM_ARGS (phi); i++)
+ {
+ tree arg = PHI_ARG_DEF (phi, i);
+
+ if (arg == lhs)
+ continue;
+ else if (!val)
+ val = arg;
+ else if (!operand_equal_p (arg, val, 0))
+ break;
+ }
+ return (i == PHI_NUM_ARGS (phi) ? val : NULL);
+}
+
+/* Given a tree node T, which is either a PHI_NODE or GIMPLE_MODIFY_STMT,
+ remove it from the IL. */
+
+static void
+remove_stmt_or_phi (tree t)
+{
+ if (TREE_CODE (t) == PHI_NODE)
+ remove_phi_node (t, NULL, true);
+ else
+ {
+ block_stmt_iterator bsi = bsi_for_stmt (t);
+ bsi_remove (&bsi, true);
+ release_defs (t);
+ }
+}
+
+/* Given a tree node T, which is either a PHI_NODE or GIMPLE_MODIFY_STMT,
+ return the "rhs" of the node, in the case of a non-degenerate
+ PHI, NULL is returned. */
+
+static tree
+get_rhs_or_phi_arg (tree t)
+{
+ if (TREE_CODE (t) == PHI_NODE)
+ return degenerate_phi_result (t);
+ else if (TREE_CODE (t) == GIMPLE_MODIFY_STMT)
+ return GIMPLE_STMT_OPERAND (t, 1);
+ gcc_unreachable ();
+}
+
+
+/* Given a tree node T, which is either a PHI_NODE or a GIMPLE_MODIFY_STMT,
+ return the "lhs" of the node. */
+
+static tree
+get_lhs_or_phi_result (tree t)
+{
+ if (TREE_CODE (t) == PHI_NODE)
+ return PHI_RESULT (t);
+ else if (TREE_CODE (t) == GIMPLE_MODIFY_STMT)
+ return GIMPLE_STMT_OPERAND (t, 0);
+ gcc_unreachable ();
+}
+
+/* Propagate RHS into all uses of LHS (when possible).
+
+ RHS and LHS are derived from STMT, which is passed in solely so
+ that we can remove it if propagation is successful.
+
+ When propagating into a PHI node or into a statement which turns
+ into a trivial copy or constant initialization, set the
+ appropriate bit in INTERESTING_NAMEs so that we will visit those
+ nodes as well in an effort to pick up secondary optimization
+ opportunities. */
+
+static void
+propagate_rhs_into_lhs (tree stmt, tree lhs, tree rhs, bitmap interesting_names)
+{
+ /* First verify that propagation is valid and isn't going to move a
+ loop variant variable outside its loop. */
+ if (! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)
+ && (TREE_CODE (rhs) != SSA_NAME
+ || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs))
+ && may_propagate_copy (lhs, rhs)
+ && loop_depth_of_name (lhs) >= loop_depth_of_name (rhs))
+ {
+ use_operand_p use_p;
+ imm_use_iterator iter;
+ tree use_stmt;
+ bool all = true;
+
+ /* Dump details. */
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, " Replacing '");
+ print_generic_expr (dump_file, lhs, dump_flags);
+ fprintf (dump_file, "' with %s '",
+ (TREE_CODE (rhs) != SSA_NAME ? "constant" : "variable"));
+ print_generic_expr (dump_file, rhs, dump_flags);
+ fprintf (dump_file, "'\n");
+ }
+
+ /* Walk over every use of LHS and try to replace the use with RHS.
+ At this point the only reason why such a propagation would not
+ be successful would be if the use occurs in an ASM_EXPR. */
+ FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
+ {
+
+ /* It's not always safe to propagate into an ASM_EXPR. */
+ if (TREE_CODE (use_stmt) == ASM_EXPR
+ && ! may_propagate_copy_into_asm (lhs))
+ {
+ all = false;
+ continue;
+ }
+
+ /* Dump details. */
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, " Original statement:");
+ print_generic_expr (dump_file, use_stmt, dump_flags);
+ fprintf (dump_file, "\n");
+ }
+
+ push_stmt_changes (&use_stmt);
+
+ /* Propagate the RHS into this use of the LHS. */
+ FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
+ propagate_value (use_p, rhs);
+
+ /* Special cases to avoid useless calls into the folding
+ routines, operand scanning, etc.
+
+ First, propagation into a PHI may cause the PHI to become
+ a degenerate, so mark the PHI as interesting. No other
+ actions are necessary.
+
+ Second, if we're propagating a virtual operand and the
+ propagation does not change the underlying _DECL node for
+ the virtual operand, then no further actions are necessary. */
+ if (TREE_CODE (use_stmt) == PHI_NODE
+ || (! is_gimple_reg (lhs)
+ && TREE_CODE (rhs) == SSA_NAME
+ && SSA_NAME_VAR (lhs) == SSA_NAME_VAR (rhs)))
+ {
+ /* Dump details. */
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, " Updated statement:");
+ print_generic_expr (dump_file, use_stmt, dump_flags);
+ fprintf (dump_file, "\n");
+ }
+
+ /* Propagation into a PHI may expose new degenerate PHIs,
+ so mark the result of the PHI as interesting. */
+ if (TREE_CODE (use_stmt) == PHI_NODE)
+ {
+ tree result = get_lhs_or_phi_result (use_stmt);
+ bitmap_set_bit (interesting_names, SSA_NAME_VERSION (result));
+ }
+
+ discard_stmt_changes (&use_stmt);
+ continue;
+ }
+
+ /* From this point onward we are propagating into a
+ real statement. Folding may (or may not) be possible,
+ we may expose new operands, expose dead EH edges,
+ etc. */
+ fold_stmt_inplace (use_stmt);
+
+ /* Sometimes propagation can expose new operands to the
+ renamer. Note this will call update_stmt at the
+ appropriate time. */
+ pop_stmt_changes (&use_stmt);
+
+ /* Dump details. */
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, " Updated statement:");
+ print_generic_expr (dump_file, use_stmt, dump_flags);
+ fprintf (dump_file, "\n");
+ }
+
+ /* If we replaced a variable index with a constant, then
+ we would need to update the invariant flag for ADDR_EXPRs. */
+ if (TREE_CODE (use_stmt) == GIMPLE_MODIFY_STMT
+ && TREE_CODE (GIMPLE_STMT_OPERAND (use_stmt, 1)) == ADDR_EXPR)
+ recompute_tree_invariant_for_addr_expr
+ (GIMPLE_STMT_OPERAND (use_stmt, 1));
+
+ /* If we cleaned up EH information from the statement,
+ mark its containing block as needing EH cleanups. */
+ if (maybe_clean_or_replace_eh_stmt (use_stmt, use_stmt))
+ {
+ bitmap_set_bit (need_eh_cleanup, bb_for_stmt (use_stmt)->index);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, " Flagged to clear EH edges.\n");
+ }
+
+ /* Propagation may expose new trivial copy/constant propagation
+ opportunities. */
+ if (TREE_CODE (use_stmt) == GIMPLE_MODIFY_STMT
+ && TREE_CODE (GIMPLE_STMT_OPERAND (use_stmt, 0)) == SSA_NAME
+ && (TREE_CODE (GIMPLE_STMT_OPERAND (use_stmt, 1)) == SSA_NAME
+ || is_gimple_min_invariant (GIMPLE_STMT_OPERAND (use_stmt,
+ 1))))
+ {
+ tree result = get_lhs_or_phi_result (use_stmt);
+ bitmap_set_bit (interesting_names, SSA_NAME_VERSION (result));
+ }
+
+ /* Propagation into these nodes may make certain edges in
+ the CFG unexecutable. We want to identify them as PHI nodes
+ at the destination of those unexecutable edges may become
+ degenerates. */
+ else if (TREE_CODE (use_stmt) == COND_EXPR
+ || TREE_CODE (use_stmt) == SWITCH_EXPR
+ || TREE_CODE (use_stmt) == GOTO_EXPR)
+ {
+ tree val;
+
+ if (TREE_CODE (use_stmt) == COND_EXPR)
+ val = COND_EXPR_COND (use_stmt);
+ else if (TREE_CODE (use_stmt) == SWITCH_EXPR)
+ val = SWITCH_COND (use_stmt);
+ else
+ val = GOTO_DESTINATION (use_stmt);
+
+ if (is_gimple_min_invariant (val))
+ {
+ basic_block bb = bb_for_stmt (use_stmt);
+ edge te = find_taken_edge (bb, val);
+ edge_iterator ei;
+ edge e;
+ block_stmt_iterator bsi;
+
+ /* Remove all outgoing edges except TE. */
+ for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei));)
+ {
+ if (e != te)
+ {
+ tree phi;
+
+ /* Mark all the PHI nodes at the destination of
+ the unexecutable edge as interesting. */
+ for (phi = phi_nodes (e->dest);
+ phi;
+ phi = PHI_CHAIN (phi))
+ {
+ tree result = PHI_RESULT (phi);
+ int version = SSA_NAME_VERSION (result);
+
+ bitmap_set_bit (interesting_names, version);
+ }
+
+ te->probability += e->probability;
+
+ te->count += e->count;
+ remove_edge (e);
+ cfg_altered = true;
+ }
+ else
+ ei_next (&ei);
+ }
+
+ bsi = bsi_last (bb_for_stmt (use_stmt));
+ bsi_remove (&bsi, true);
+
+ /* And fixup the flags on the single remaining edge. */
+ te->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
+ te->flags &= ~EDGE_ABNORMAL;
+ te->flags |= EDGE_FALLTHRU;
+ if (te->probability > REG_BR_PROB_BASE)
+ te->probability = REG_BR_PROB_BASE;
+ }
+ }
+ }
+
+ /* Ensure there is nothing else to do. */
+ gcc_assert (!all || has_zero_uses (lhs));
+
+ /* If we were able to propagate away all uses of LHS, then
+ we can remove STMT. */
+ if (all)
+ remove_stmt_or_phi (stmt);
+ }
+}
+
+/* T is either a PHI node (potentially a degenerate PHI node) or
+ a statement that is a trivial copy or constant initialization.
+
+ Attempt to eliminate T by propagating its RHS into all uses of
+ its LHS. This may in turn set new bits in INTERESTING_NAMES
+ for nodes we want to revisit later.
+
+ All exit paths should clear INTERESTING_NAMES for the result
+ of T. */
+
+static void
+eliminate_const_or_copy (tree t, bitmap interesting_names)
+{
+ tree lhs = get_lhs_or_phi_result (t);
+ tree rhs;
+ int version = SSA_NAME_VERSION (lhs);
+
+ /* If the LHS of this statement or PHI has no uses, then we can
+ just eliminate it. This can occur if, for example, the PHI
+ was created by block duplication due to threading and its only
+ use was in the conditional at the end of the block which was
+ deleted. */
+ if (has_zero_uses (lhs))
+ {
+ bitmap_clear_bit (interesting_names, version);
+ remove_stmt_or_phi (t);
+ return;
+ }
+
+ /* Get the RHS of the assignment or PHI node if the PHI is a
+ degenerate. */
+ rhs = get_rhs_or_phi_arg (t);
+ if (!rhs)
+ {
+ bitmap_clear_bit (interesting_names, version);
+ return;
+ }
+
+ propagate_rhs_into_lhs (t, lhs, rhs, interesting_names);
+
+ /* Note that T may well have been deleted by now, so do
+ not access it, instead use the saved version # to clear
+ T's entry in the worklist. */
+ bitmap_clear_bit (interesting_names, version);
+}
+
+/* The first phase in degenerate PHI elimination.
+
+ Eliminate the degenerate PHIs in BB, then recurse on the
+ dominator children of BB. */
+
+static void
+eliminate_degenerate_phis_1 (basic_block bb, bitmap interesting_names)
+{
+ tree phi, next;
+ basic_block son;
+
+ for (phi = phi_nodes (bb); phi; phi = next)
+ {
+ next = PHI_CHAIN (phi);
+ eliminate_const_or_copy (phi, interesting_names);
+ }
+
+ /* Recurse into the dominator children of BB. */
+ for (son = first_dom_son (CDI_DOMINATORS, bb);
+ son;
+ son = next_dom_son (CDI_DOMINATORS, son))
+ eliminate_degenerate_phis_1 (son, interesting_names);
+}
+
+
+/* A very simple pass to eliminate degenerate PHI nodes from the
+ IL. This is meant to be fast enough to be able to be run several
+ times in the optimization pipeline.
+
+ Certain optimizations, particularly those which duplicate blocks
+ or remove edges from the CFG can create or expose PHIs which are
+ trivial copies or constant initializations.
+
+ While we could pick up these optimizations in DOM or with the
+ combination of copy-prop and CCP, those solutions are far too
+ heavy-weight for our needs.
+
+ This implementation has two phases so that we can efficiently
+ eliminate the first order degenerate PHIs and second order
+ degenerate PHIs.
+
+ The first phase performs a dominator walk to identify and eliminate
+ the vast majority of the degenerate PHIs. When a degenerate PHI
+ is identified and eliminated any affected statements or PHIs
+ are put on a worklist.
+
+ The second phase eliminates degenerate PHIs and trivial copies
+ or constant initializations using the worklist. This is how we
+ pick up the secondary optimization opportunities with minimal
+ cost. */
+
+static unsigned int
+eliminate_degenerate_phis (void)
+{
+ bitmap interesting_names;
+ bitmap interesting_names1;
+
+ /* Bitmap of blocks which need EH information updated. We can not
+ update it on-the-fly as doing so invalidates the dominator tree. */
+ need_eh_cleanup = BITMAP_ALLOC (NULL);
+
+ /* INTERESTING_NAMES is effectively our worklist, indexed by
+ SSA_NAME_VERSION.
+
+ A set bit indicates that the statement or PHI node which
+ defines the SSA_NAME should be (re)examined to determine if
+ it has become a degenerate PHI or trivial const/copy propagation
+ opportunity.
+
+ Experiments have show we generally get better compilation
+ time behavior with bitmaps rather than sbitmaps. */
+ interesting_names = BITMAP_ALLOC (NULL);
+ interesting_names1 = BITMAP_ALLOC (NULL);
+
+ calculate_dominance_info (CDI_DOMINATORS);
+ cfg_altered = false;
+
+ /* First phase. Eliminate degenerate PHIs via a dominator
+ walk of the CFG.
+
+ Experiments have indicated that we generally get better
+ compile-time behavior by visiting blocks in the first
+ phase in dominator order. Presumably this is because walking
+ in dominator order leaves fewer PHIs for later examination
+ by the worklist phase. */
+ eliminate_degenerate_phis_1 (ENTRY_BLOCK_PTR, interesting_names);
+
+ /* Second phase. Eliminate second order degenerate PHIs as well
+ as trivial copies or constant initializations identified by
+ the first phase or this phase. Basically we keep iterating
+ until our set of INTERESTING_NAMEs is empty. */
+ while (!bitmap_empty_p (interesting_names))
+ {
+ unsigned int i;
+ bitmap_iterator bi;
+
+ /* EXECUTE_IF_SET_IN_BITMAP does not like its bitmap
+ changed during the loop. Copy it to another bitmap and
+ use that. */
+ bitmap_copy (interesting_names1, interesting_names);
+
+ EXECUTE_IF_SET_IN_BITMAP (interesting_names1, 0, i, bi)
+ {
+ tree name = ssa_name (i);
+
+ /* Ignore SSA_NAMEs that have been released because
+ their defining statement was deleted (unreachable). */
+ if (name)
+ eliminate_const_or_copy (SSA_NAME_DEF_STMT (ssa_name (i)),
+ interesting_names);
+ }
+ }
+
+ if (cfg_altered)
+ free_dominance_info (CDI_DOMINATORS);
+
+ /* Propagation of const and copies may make some EH edges dead. Purge
+ such edges from the CFG as needed. */
+ if (!bitmap_empty_p (need_eh_cleanup))
+ {
+ tree_purge_all_dead_eh_edges (need_eh_cleanup);
+ BITMAP_FREE (need_eh_cleanup);
+ }
+
+ BITMAP_FREE (interesting_names);
+ BITMAP_FREE (interesting_names1);
+ return 0;
+}
+
+struct tree_opt_pass pass_phi_only_cprop =
+{
+ "phicprop", /* name */
+ gate_dominator, /* gate */
+ eliminate_degenerate_phis, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_TREE_PHI_CPROP, /* tv_id */
+ PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_cleanup_cfg
+ | TODO_dump_func
+ | TODO_ggc_collect
+ | TODO_verify_ssa
+ | TODO_verify_stmts
+ | TODO_update_ssa, /* todo_flags_finish */
+ 0 /* letter */
+};
OSDN Git Service
