* @author Michael Beck
* @version $Id$
*
- * Note that we use the terminology from Click's work here, which is different
+ * This is a slightly enhanced version of Cliff Clicks combo algorithm
+ * - support for commutative nodes is added, Add(a,b) and Add(b,a) ARE congruent
+ * - supports all Firm direct (by a data edge) identities except Mux
+ * (Mux can be a 2-input or 1-input identity, only 2-input is implemented yet)
+ * - supports Confirm nodes (handle them like Copies but do NOT remove them)
+ * - let Cmp nodes calculate Top like all othe data nodes: this would let
+ * Mux nodes to calculate Unknown instead of taking the true result
+ * - let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated
+ * IFF the predecessor changed its type. Because nodes are initialized with Top
+ * this never happens, let all Proj(Cond) be unreachable.
+ * We avoid this condition by the same way we work around Phi: whenever a Block
+ * node is placed on the list, place its Cond nodes (and because they are Tuple
+ * all its Proj-nodes either on the cprop list)
+ * Especially, this changes the meaning of Click's example:
+ *
+ * int main() {
+ * int x;
+ *
+ * if (x == 2)
+ * printf("x == 2\n");
+ * if (x == 3)
+ * printf("x == 3\n");
+ * }
+ *
+ * Would print:
+ * x == 2
+ * x == 3
+ *
+ * using Click's version while is silent with our.
+ * - support for global congruences is implemented but not tested yet
+ *
+ * Note further that we use the terminology from Click's work here, which is different
* in some cases from Firm terminology. Especially, Click's type is a
* Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
*/
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
#include <assert.h>
#include "iroptimize.h"
+#include "archop.h"
#include "irflag.h"
#include "ircons.h"
#include "list.h"
-#include "array.h"
#include "set.h"
#include "pmap.h"
#include "obstack.h"
#include "irop.h"
#include "irouts.h"
#include "irgmod.h"
+#include "iropt_dbg.h"
#include "debug.h"
+#include "array_t.h"
#include "error.h"
#include "tv_t.h"
#include "irdump.h"
/* define this to check that all type translations are monotone */
-#define VERIFY_MONOTONE
+#undef VERIFY_MONOTONE
+
+/* define this to check the consistency of partitions */
+#define CHECK_PARTITIONS
+
+
+/* allow optimization of non-strict programs */
+#define WITH_UNKNOWN
typedef struct node_t node_t;
typedef struct partition_t partition_t;
struct opcode_key_t {
ir_opcode code; /**< The Firm opcode. */
ir_mode *mode; /**< The mode of all nodes in the partition. */
+ int arity; /**< The arity of this opcode (needed for Phi etc. */
union {
long proj; /**< For Proj nodes, its proj number */
ir_entity *ent; /**< For Sel Nodes, its entity */
*/
struct node_t {
ir_node *node; /**< The IR-node itself. */
- list_head node_list; /**< Double-linked list of entries. */
+ list_head node_list; /**< Double-linked list of leader/follower entries. */
list_head cprop_list; /**< Double-linked partition.cprop list. */
partition_t *part; /**< points to the partition this node belongs to */
node_t *next; /**< Next node on local list (partition.touched, fallen). */
+ node_t *race_next; /**< Next node on race list. */
lattice_elem_t type; /**< The associated lattice element "type". */
int max_user_input; /**< Maximum input number of Def-Use edges. */
int next_edge; /**< Index of the next Def-Use edge to use. */
+ int n_followers; /**< Number of Follower in the outs set. */
unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
+ unsigned is_follower:1; /**< Set, if this node is a follower. */
+ unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */
+ unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
+ node_t *cond; /**< if this is a Block node, points to its Cond if any */
};
/**
* A partition containing congruent nodes.
*/
struct partition_t {
- list_head entries; /**< The head of partition node list. */
+ list_head Leader; /**< The head of partition Leader node list. */
+ list_head Follower; /**< The head of partition Follower node list. */
list_head cprop; /**< The head of partition.cprop list. */
partition_t *wl_next; /**< Next entry in the work list if any. */
partition_t *touched_next; /**< Points to the next partition in the touched set. */
partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
+ partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
node_t *touched; /**< The partition.touched set of this partition. */
- unsigned n_nodes; /**< Number of entries in this partition. */
+ unsigned n_leader; /**< Number of entries in this partition.Leader. */
unsigned n_touched; /**< Number of entries in the partition.touched. */
- int max_arity; /**< Maximum arity of all entries. */
int max_user_inputs; /**< Maximum number of user inputs of all entries. */
unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
+ unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
#ifdef DEBUG_libfirm
partition_t *dbg_next; /**< Link all partitions for debugging */
unsigned nr; /**< A unique number for (what-)mapping, >0. */
pmap *type2id_map; /**< The type->id map. */
int end_idx; /**< -1 for local and 0 for global congruences. */
int lambda_input; /**< Captured argument for lambda_partition(). */
+ char modified; /**< Set, if the graph was modified. */
+ char commutative; /**< Set, if commutation nodes should be handled specially. */
#ifdef DEBUG_libfirm
partition_t *dbg_list; /**< List of all partitions. */
#endif
/** Type of the what function. */
typedef void *(*what_func)(const node_t *node, environment_t *env);
-#define get_irn_node(irn) ((node_t *)get_irn_link(irn))
-#define set_irn_node(irn, node) set_irn_link(irn, node)
+#define get_irn_node(follower) ((node_t *)get_irn_link(follower))
+#define set_irn_node(follower, node) set_irn_link(follower, node)
/* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
#undef tarval_unreachable
/** The debug module handle. */
DEBUG_ONLY(static firm_dbg_module_t *dbg;)
+/** The what reason. */
+DEBUG_ONLY(static const char *what_reason;)
+
/** Next partition number. */
DEBUG_ONLY(static unsigned part_nr = 0);
+/** The tarval returned by Unknown nodes. */
+static tarval *tarval_UNKNOWN;
+
+/* forward */
+static node_t *identity(node_t *node);
+
+#ifdef CHECK_PARTITIONS
+/**
+ * Check a partition.
+ */
+static void check_partition(const partition_t *T) {
+ node_t *node;
+ unsigned n = 0;
+
+ list_for_each_entry(node_t, node, &T->Leader, node_list) {
+ assert(node->is_follower == 0);
+ assert(node->flagged == 0);
+ assert(node->part == T);
+ ++n;
+ }
+ assert(n == T->n_leader);
+
+ list_for_each_entry(node_t, node, &T->Follower, node_list) {
+ assert(node->is_follower == 1);
+ assert(node->flagged == 0);
+ assert(node->part == T);
+ }
+} /* check_partition */
+
+/**
+ * check that all leader nodes in the partition have the same opcode.
+ */
+static void check_opcode(const partition_t *Z) {
+ node_t *node;
+ opcode_key_t key;
+ int first = 1;
+
+ list_for_each_entry(node_t, node, &Z->Leader, node_list) {
+ ir_node *irn = node->node;
+
+ if (first) {
+ key.code = get_irn_opcode(irn);
+ key.mode = get_irn_mode(irn);
+ key.arity = get_irn_arity(irn);
+ key.u.proj = 0;
+ key.u.ent = NULL;
+
+ switch (get_irn_opcode(irn)) {
+ case iro_Proj:
+ key.u.proj = get_Proj_proj(irn);
+ break;
+ case iro_Sel:
+ key.u.ent = get_Sel_entity(irn);
+ break;
+ default:
+ break;
+ }
+ first = 0;
+ } else {
+ assert(key.code == get_irn_opcode(irn));
+ assert(key.mode == get_irn_mode(irn));
+ assert(key.arity == get_irn_arity(irn));
+
+ switch (get_irn_opcode(irn)) {
+ case iro_Proj:
+ assert(key.u.proj == get_Proj_proj(irn));
+ break;
+ case iro_Sel:
+ assert(key.u.ent == get_Sel_entity(irn));
+ break;
+ default:
+ break;
+ }
+ }
+ }
+} /* check_opcode */
+
+static void check_all_partitions(environment_t *env) {
+#ifdef DEBUG_libfirm
+ partition_t *P;
+ node_t *node;
+
+ for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
+ check_partition(P);
+ if (! P->type_is_T_or_C)
+ check_opcode(P);
+ list_for_each_entry(node_t, node, &P->Follower, node_list) {
+ node_t *leader = identity(node);
+
+ assert(leader != node && leader->part == node->part);
+ }
+ }
+#endif
+}
+
+/**
+ * Check list.
+ */
+static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
+ const node_t *e;
+
+#define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
+ for (e = list; e != NULL; e = NEXT(e)) {
+ assert(e->part == Z);
+ }
+#undef NEXT
+} /* ido_check_list */
+
+/**
+ * Check a local list.
+ */
+static void check_list(const node_t *list, const partition_t *Z) {
+ do_check_list(list, offsetof(node_t, next), Z);
+} /* check_list */
+
+#else
+#define check_partition(T)
+#define check_list(list, Z)
+#define check_all_partitions(env)
+#endif /* CHECK_PARTITIONS */
+
#ifdef DEBUG_libfirm
-static INLINE lattice_elem_t get_partition_type(const partition_t *X);
+static inline lattice_elem_t get_partition_type(const partition_t *X);
/**
* Dump partition to output.
int first = 1;
lattice_elem_t type = get_partition_type(part);
- DB((dbg, LEVEL_2, "%s part%u (%u, %+F) {\n ", msg, part->nr, part->n_nodes, type));
- list_for_each_entry(node_t, node, &part->entries, node_list) {
+ DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
+ msg, part->nr, part->type_is_T_or_C ? "*" : "",
+ part->n_leader, type));
+ list_for_each_entry(node_t, node, &part->Leader, node_list) {
DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
first = 0;
}
+ if (! list_empty(&part->Follower)) {
+ DB((dbg, LEVEL_2, "\n---\n "));
+ first = 1;
+ list_for_each_entry(node_t, node, &part->Follower, node_list) {
+ DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
+ first = 0;
+ }
+ }
DB((dbg, LEVEL_2, "\n}\n"));
-}
+} /* dump_partition */
+
+/**
+ * Dumps a list.
+ */
+static void do_dump_list(const char *msg, const node_t *node, int ofs) {
+ const node_t *p;
+ int first = 1;
+
+#define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
+
+ DB((dbg, LEVEL_3, "%s = {\n ", msg));
+ for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
+ DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
+ first = 0;
+ }
+ DB((dbg, LEVEL_3, "\n}\n"));
+
+#undef GET_LINK
+} /* do_dump_list */
+
+/**
+ * Dumps a race list.
+ */
+static void dump_race_list(const char *msg, const node_t *list) {
+ do_dump_list(msg, list, offsetof(node_t, race_next));
+} /* dump_race_list */
+
+/**
+ * Dumps a local list.
+ */
+static void dump_list(const char *msg, const node_t *list) {
+ do_dump_list(msg, list, offsetof(node_t, next));
+} /* dump_list */
/**
* Dump all partitions.
DB((dbg, LEVEL_2, "All partitions\n===============\n"));
for (P = env->dbg_list; P != NULL; P = P->dbg_next)
dump_partition("", P);
-}
+} /* dump_all_partitions */
+
+/**
+ * Sump a split list.
+ */
+static void dump_split_list(const partition_t *list) {
+ const partition_t *p;
+
+ DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
+ for (p = list; p != NULL; p = p->split_next)
+ DB((dbg, LEVEL_2, "part%u, ", p->nr));
+ DB((dbg, LEVEL_2, "\n}\n"));
+} /* dump_split_list */
+
+/**
+ * Dump partition and type for a node.
+ */
+static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
+ ir_node *irn = local != NULL ? local : n;
+ node_t *node = get_irn_node(irn);
+
+ ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
+ return 1;
+} /* dump_partition_hook */
#else
#define dump_partition(msg, part)
+#define dump_race_list(msg, list)
+#define dump_list(msg, list)
#define dump_all_partitions(env)
+#define dump_split_list(list)
#endif
#if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
return;
}
panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
-}
+} /* verify_type */
#else
#define verify_type(old_type, new_type)
#endif
* @param map the listmap
* @param id the id to search for
*
- * @return the asociated listmap entry for the given id
+ * @return the associated listmap entry for the given id
*/
static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
listmap_entry_t key, *entry;
* @return a hash value for the given opcode map entry
*/
static unsigned opcode_hash(const opcode_key_t *entry) {
- return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent);
+ return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent) + entry->arity;
} /* opcode_hash */
/**
(void) size;
return o1->code != o2->code || o1->mode != o2->mode ||
+ o1->arity != o2->arity ||
o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
} /* cmp_opcode */
*
* @return the associated type of this node
*/
-static INLINE lattice_elem_t get_node_type(const ir_node *irn) {
+static inline lattice_elem_t get_node_type(const ir_node *irn) {
return get_irn_node(irn)->type;
} /* get_node_type */
*
* @return the associated type of this node
*/
-static INLINE tarval *get_node_tarval(const ir_node *irn) {
+static inline tarval *get_node_tarval(const ir_node *irn) {
lattice_elem_t type = get_node_type(irn);
if (is_tarval(type.tv))
/**
* Add a partition to the worklist.
*/
-static INLINE void add_to_worklist(partition_t *X, environment_t *env) {
+static inline void add_to_worklist(partition_t *X, environment_t *env) {
assert(X->on_worklist == 0);
+ DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
X->wl_next = env->worklist;
X->on_worklist = 1;
env->worklist = X;
-}
+} /* add_to_worklist */
/**
* Create a new empty partition.
*
* @return a newly allocated partition
*/
-static INLINE partition_t *new_partition(environment_t *env) {
+static inline partition_t *new_partition(environment_t *env) {
partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
- INIT_LIST_HEAD(&part->entries);
+ INIT_LIST_HEAD(&part->Leader);
+ INIT_LIST_HEAD(&part->Follower);
INIT_LIST_HEAD(&part->cprop);
part->wl_next = NULL;
part->touched_next = NULL;
part->cprop_next = NULL;
+ part->split_next = NULL;
part->touched = NULL;
- part->n_nodes = 0;
+ part->n_leader = 0;
part->n_touched = 0;
- part->max_arity = 0;
part->max_user_inputs = 0;
part->on_worklist = 0;
part->on_touched = 0;
part->on_cprop = 0;
+ part->type_is_T_or_C = 0;
#ifdef DEBUG_libfirm
part->dbg_next = env->dbg_list;
env->dbg_list = part;
/**
* Get the first node from a partition.
*/
-static INLINE node_t *get_first_node(const partition_t *X) {
- return list_entry(X->entries.next, node_t, node_list);
-}
+static inline node_t *get_first_node(const partition_t *X) {
+ return list_entry(X->Leader.next, node_t, node_list);
+} /* get_first_node */
/**
* Return the type of a partition (assuming partition is non-empty and
*
* @return the type of the first element of the partition
*/
-static INLINE lattice_elem_t get_partition_type(const partition_t *X) {
+static inline lattice_elem_t get_partition_type(const partition_t *X) {
const node_t *first = get_first_node(X);
return first->type;
} /* get_partition_type */
node->node = irn;
node->part = part;
node->next = NULL;
+ node->race_next = NULL;
node->type.tv = tarval_top;
node->max_user_input = 0;
node->next_edge = 0;
+ node->n_followers = 0;
node->on_touched = 0;
node->on_cprop = 0;
node->on_fallen = 0;
+ node->is_follower = 0;
+ node->by_all_const = 0;
+ node->flagged = 0;
+ node->cond = NULL;
set_irn_node(irn, node);
- list_add_tail(&node->node_list, &part->entries);
- ++part->n_nodes;
+ list_add_tail(&node->node_list, &part->Leader);
+ ++part->n_leader;
return node;
} /* create_partition_node */
if (is_Block(irn)) {
set_Block_phis(irn, NULL);
}
-}
+} /* init_block_phis */
/**
* Post-Walker, initialize all Nodes' type to U or top and place
environment_t *env = ctx;
partition_t *part = env->initial;
node_t *node;
- int arity;
node = create_partition_node(irn, part, env);
sort_irn_outs(node);
- arity = get_irn_arity(irn);
- if (arity > part->max_arity)
- part->max_arity = arity;
if (node->max_user_input > part->max_user_inputs)
part->max_user_inputs = node->max_user_input;
if (is_Phi(irn)) {
add_Block_phi(get_nodes_block(irn), irn);
- }
-} /* create_initial_partitions */
+ } else if (is_Cond(irn)) {
+ node_t *block = get_irn_node(get_nodes_block(irn));
-/**
- * Add a partition to the touched set if not already there.
- *
- * @param part the partition
- * @param env the environment
- */
-static INLINE void add_to_touched(partition_t *part, environment_t *env) {
- if (part->on_touched == 0) {
- part->touched_next = env->touched;
- env->touched = part;
- part->on_touched = 1;
+ /* link every block with its Cond node if any */
+ block->cond = node;
}
-} /* add_to_touched */
+} /* create_initial_partitions */
/**
- * Add a node to the entry.partition.touched set if not already there.
+ * Add a node to the entry.partition.touched set and
+ * node->partition to the touched set if not already there.
*
- * @param y a node
+ * @param y a node
+ * @param env the environment
*/
-static INLINE void add_to_partition_touched(node_t *y) {
+static inline void add_to_touched(node_t *y, environment_t *env) {
if (y->on_touched == 0) {
partition_t *part = y->part;
part->touched = y;
y->on_touched = 1;
++part->n_touched;
+
+ if (part->on_touched == 0) {
+ part->touched_next = env->touched;
+ env->touched = part;
+ part->on_touched = 1;
+ }
+
+ check_list(part->touched, part);
+ }
+} /* add_to_touched */
+
+/**
+ * Place a node on the cprop list.
+ *
+ * @param y the node
+ * @param env the environment
+ */
+static void add_to_cprop(node_t *y, environment_t *env) {
+ /* Add y to y.partition.cprop. */
+ if (y->on_cprop == 0) {
+ partition_t *Y = y->part;
+
+ list_add_tail(&y->cprop_list, &Y->cprop);
+ y->on_cprop = 1;
+
+ DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
+
+ /* place its partition on the cprop list */
+ if (Y->on_cprop == 0) {
+ Y->cprop_next = env->cprop;
+ env->cprop = Y;
+ Y->on_cprop = 1;
+ }
+ }
+ if (get_irn_mode(y->node) == mode_T) {
+ /* mode_T nodes always produce tarval_bottom, so we must explicitly
+ add it's Proj's to get constant evaluation to work */
+ int i;
+
+ for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) {
+ node_t *proj = get_irn_node(get_irn_out(y->node, i));
+
+ add_to_cprop(proj, env);
+ }
+ } else if (is_Block(y->node)) {
+ /* Due to the way we handle Phi's, we must place all Phis of a block on the list
+ * if someone placed the block. The Block is only placed if the reachability
+ * changes, and this must be re-evaluated in compute_Phi(). */
+ ir_node *phi;
+ for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) {
+ node_t *p = get_irn_node(phi);
+ add_to_cprop(p, env);
+ }
+ /* same for Conds: they must be re-evaluated due to the way we handle Top */
+ if (y->cond != NULL)
+ add_to_cprop(y->cond, env);
}
-} /* add_to_partition_touched */
+} /* add_to_cprop */
/**
* Update the worklist: If Z is on worklist then add Z' to worklist.
* @param env the environment
*/
static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
- if (Z->on_worklist || Z_prime->n_nodes < Z->n_nodes) {
+ if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
add_to_worklist(Z_prime, env);
} else {
add_to_worklist(Z, env);
} /* update_worklist */
/**
- * Split a partition by a local list.
+ * Make all inputs to x no longer be F.def_use edges.
+ *
+ * @param x the node
+ */
+static void move_edges_to_leader(node_t *x) {
+ ir_node *irn = x->node;
+ int i, j, k;
+
+ for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
+ node_t *pred = get_irn_node(get_irn_n(irn, i));
+ ir_node *p;
+ int n;
+
+ p = pred->node;
+ n = get_irn_n_outs(p);
+ for (j = 1; j <= pred->n_followers; ++j) {
+ if (p->out[j].pos == i && p->out[j].use == irn) {
+ /* found a follower edge to x, move it to the Leader */
+ ir_def_use_edge edge = p->out[j];
+
+ /* remove this edge from the Follower set */
+ p->out[j] = p->out[pred->n_followers];
+ --pred->n_followers;
+
+ /* sort it into the leader set */
+ for (k = pred->n_followers + 2; k <= n; ++k) {
+ if (p->out[k].pos >= edge.pos)
+ break;
+ p->out[k - 1] = p->out[k];
+ }
+ /* place the new edge here */
+ p->out[k - 1] = edge;
+
+ /* edge found and moved */
+ break;
+ }
+ }
+ }
+} /* move_edges_to_leader */
+
+/**
+ * Split a partition that has NO followers by a local list.
*
- * @param Z the Z partition to split
+ * @param Z partition to split
* @param g a (non-empty) node list
* @param env the environment
*
* @return a new partition containing the nodes of g
*/
-static partition_t *split(partition_t *Z, node_t *g, environment_t *env) {
+static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
partition_t *Z_prime;
node_t *node;
unsigned n = 0;
- int max_input, max_arity, arity;
+ int max_input;
dump_partition("Splitting ", Z);
+ dump_list("by list ", g);
assert(g != NULL);
/* Remove g from Z. */
for (node = g; node != NULL; node = node->next) {
+ assert(node->part == Z);
list_del(&node->node_list);
++n;
}
- assert(n < Z->n_nodes);
- Z->n_nodes -= n;
+ assert(n < Z->n_leader);
+ Z->n_leader -= n;
- /* Move g to a new partition, Z\92. */
+ /* Move g to a new partition, Z'. */
Z_prime = new_partition(env);
- max_arity = max_input = 0;
+ max_input = 0;
for (node = g; node != NULL; node = node->next) {
- list_add(&node->node_list, &Z_prime->entries);
+ list_add_tail(&node->node_list, &Z_prime->Leader);
node->part = Z_prime;
- arity = get_irn_arity(node->node);
- if (arity > max_arity)
- max_arity = arity;
if (node->max_user_input > max_input)
max_input = node->max_user_input;
}
- Z_prime->max_arity = max_arity;
Z_prime->max_user_inputs = max_input;
- Z_prime->n_nodes = n;
+ Z_prime->n_leader = n;
+
+ check_partition(Z);
+ check_partition(Z_prime);
+
+ /* for now, copy the type info tag, it will be adjusted in split_by(). */
+ Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
update_worklist(Z, Z_prime, env);
dump_partition("Now ", Z);
dump_partition("Created new ", Z_prime);
return Z_prime;
-} /* split */
+} /* split_no_followers */
/**
- * Returns non-zero if the i'th input of a Phi node is live.
+ * Make the Follower -> Leader transition for a node.
*
- * @param phi a Phi-node
- * @param i an input number
- *
- * @return non-zero if the i'th input of the given Phi node is live
+ * @param n the node
*/
-static int is_live_input(ir_node *phi, int i) {
- if (i >= 0) {
- ir_node *block = get_nodes_block(phi);
- ir_node *pred = get_Block_cfgpred(block, i);
- lattice_elem_t type = get_node_type(pred);
+static void follower_to_leader(node_t *n) {
+ assert(n->is_follower == 1);
- return type.tv != tarval_unreachable;
- }
- /* else it's the control input, always live */
- return 1;
-} /* is_live_input */
+ DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
+ n->is_follower = 0;
+ move_edges_to_leader(n);
+ list_del(&n->node_list);
+ list_add_tail(&n->node_list, &n->part->Leader);
+ ++n->part->n_leader;
+} /* follower_to_leader */
/**
- * Return non-zero if a type is a constant.
+ * The environment for one race step.
*/
-static int is_constant_type(lattice_elem_t type) {
- if (type.tv != tarval_bottom && type.tv != tarval_top)
- return 1;
- return 0;
-} /* is_constant_type */
+typedef struct step_env {
+ node_t *initial; /**< The initial node list. */
+ node_t *unwalked; /**< The unwalked node list. */
+ node_t *walked; /**< The walked node list. */
+ int index; /**< Next index of Follower use_def edge. */
+ unsigned side; /**< side number. */
+} step_env;
/**
- * Place a node on the cprop list.
+ * Return non-zero, if a input is a real follower
*
- * @param y the node
- * @param env the environment
+ * @param irn the node to check
+ * @param input number of the input
*/
-static void add_node_to_cprop(node_t *y, environment_t *env) {
- /* Add y to y.partition.cprop. */
- if (y->on_cprop == 0) {
- partition_t *Y = y->part;
-
- list_add_tail(&y->cprop_list, &Y->cprop);
- y->on_cprop = 1;
-
- DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
+static int is_real_follower(const ir_node *irn, int input) {
+ node_t *pred;
- /* place its partition on the cprop list */
- if (Y->on_cprop == 0) {
- Y->cprop_next = env->cprop;
- env->cprop = Y;
- Y->on_cprop = 1;
+ switch (get_irn_opcode(irn)) {
+ case iro_Confirm:
+ if (input == 1) {
+ /* ignore the Confirm bound input */
+ return 0;
}
- }
- if (get_irn_mode(y->node) == mode_T) {
- /* mode_T nodes always produce tarval_bottom, so we must explicitly
- add it's Proj's to get constant evaluation to work */
- int i;
-
- for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) {
- node_t *proj = get_irn_node(get_irn_out(y->node, i));
-
- add_node_to_cprop(proj, env);
+ break;
+ case iro_Mux:
+ if (input == 0) {
+ /* ignore the Mux sel input */
+ return 0;
}
- }
+ break;
+ case iro_Phi: {
+ /* dead inputs are not follower edges */
+ ir_node *block = get_nodes_block(irn);
+ node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
- if (is_Block(y->node)) {
- /* Due to the way we handle Phi's, we must place all Phis of a block on the list
- * if someone placed the block. The Block is only placed if the reachability
- * changes, and this must be re-evaluated in compute_Phi(). */
- ir_node *phi;
- for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) {
- node_t *p = get_irn_node(phi);
- add_node_to_cprop(p, env);
- }
+ if (pred->type.tv == tarval_unreachable)
+ return 0;
+ break;
}
-} /* add_node_to_cprop */
-
-/**
- * Check whether a type is neither Top or a constant.
- * Note: U is handled like Top here, R is a constant.
- *
- * @param type the type to check
- */
-static int type_is_neither_top_nor_const(const lattice_elem_t type) {
- if (is_tarval(type.tv)) {
- if (type.tv == tarval_top)
+ case iro_Sub:
+ case iro_Shr:
+ case iro_Shl:
+ case iro_Shrs:
+ case iro_Rotl:
+ if (input == 1) {
+ /* only a Sub x,0 / Shift x,0 might be a follower */
return 0;
- if (tarval_is_constant(type.tv))
+ }
+ break;
+ case iro_Add:
+ case iro_Or:
+ case iro_Eor:
+ pred = get_irn_node(get_irn_n(irn, input));
+ if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
return 0;
- } else {
- /* is a symconst */
- return 0;
+ break;
+ case iro_Mul:
+ pred = get_irn_node(get_irn_n(irn, input));
+ if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
+ return 0;
+ break;
+ case iro_And:
+ pred = get_irn_node(get_irn_n(irn, input));
+ if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
+ return 0;
+ break;
+ case iro_Min:
+ case iro_Max:
+ /* all inputs are followers */
+ return 1;
+ default:
+ assert(!"opcode not implemented yet");
+ break;
}
return 1;
-}
+} /* is_real_follower */
/**
- * Split the partitions if caused by the first entry on the worklist.
- *
- * @param env the environment
+ * Do one step in the race.
*/
-static void cause_splits(environment_t *env) {
- partition_t *X, *Y, *Z;
- node_t *x, *y, *e;
- int i, end_idx;
- ir_opcode code;
- ir_node *succ;
+static int step(step_env *env) {
+ node_t *n;
- /* remove the first partition from the worklist */
- X = env->worklist;
- env->worklist = X->wl_next;
- X->on_worklist = 0;
+ if (env->initial != NULL) {
+ /* Move node from initial to unwalked */
+ n = env->initial;
+ env->initial = n->race_next;
- dump_partition("Cause_split: ", X);
- end_idx = env->end_idx;
- for (i = -1; i <= X->max_user_inputs; ++i) {
- /* empty the touched set: already done, just clear the list */
- env->touched = NULL;
+ n->race_next = env->unwalked;
+ env->unwalked = n;
- list_for_each_entry(node_t, x, &X->entries, node_list) {
- int num_edges;
+ return 0;
+ }
- if (i == -1) {
- x->next_edge = 1;
- }
- num_edges = get_irn_n_outs(x->node);
+ while (env->unwalked != NULL) {
+ /* let n be the first node in unwalked */
+ n = env->unwalked;
+ while (env->index < n->n_followers) {
+ const ir_def_use_edge *edge = &n->node->out[1 + env->index];
- while (x->next_edge <= num_edges) {
- ir_def_use_edge *edge = &x->node->out[x->next_edge];
+ /* let m be n.F.def_use[index] */
+ node_t *m = get_irn_node(edge->use);
- /* check if we have necessary edges */
- if (edge->pos > i)
- break;
+ assert(m->is_follower);
+ /*
+ * Some inputs, like the get_Confirm_bound are NOT
+ * real followers, sort them out.
+ */
+ if (! is_real_follower(m->node, edge->pos)) {
+ ++env->index;
+ continue;
+ }
+ ++env->index;
- ++x->next_edge;
+ /* only followers from our partition */
+ if (m->part != n->part)
+ continue;
- succ = edge->use;
+ if ((m->flagged & env->side) == 0) {
+ m->flagged |= env->side;
- /* ignore the "control input" for non-pinned nodes
- if we are running in GCSE mode */
- if (i < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
- continue;
-
- y = get_irn_node(succ);
- if (is_constant_type(y->type)) {
- code = get_irn_opcode(succ);
- if (code == iro_Sub || (code == iro_Proj && is_Cmp(get_Proj_pred(succ))))
- add_node_to_cprop(y, env);
+ if (m->flagged != 3) {
+ /* visited the first time */
+ /* add m to unwalked not as first node (we might still need to
+ check for more follower node */
+ m->race_next = n->race_next;
+ n->race_next = m;
+ return 0;
}
+ /* else already visited by the other side and on the other list */
+ }
+ }
+ /* move n to walked */
+ env->unwalked = n->race_next;
+ n->race_next = env->walked;
+ env->walked = n;
+ env->index = 0;
+ }
+ return 1;
+} /* step */
- /* Partitions of constants should not be split simply because their Nodes have unequal
- functions or incongruent inputs. */
- if (type_is_neither_top_nor_const(y->type) &&
- (! is_Phi(y->node) || is_live_input(y->node, i))) {
- Y = y->part;
- add_to_touched(Y, env);
- add_to_partition_touched(y);
- }
+/**
+ * Clear the flags from a list and check for
+ * nodes that where touched from both sides.
+ *
+ * @param list the list
+ */
+static int clear_flags(node_t *list) {
+ int res = 0;
+ node_t *n;
+
+ for (n = list; n != NULL; n = n->race_next) {
+ if (n->flagged == 3) {
+ /* we reach a follower from both sides, this will split congruent
+ * inputs and make it a leader. */
+ follower_to_leader(n);
+ res = 1;
+ }
+ n->flagged = 0;
+ }
+ return res;
+} /* clear_flags */
+
+/**
+ * Split a partition by a local list using the race.
+ *
+ * @param pX pointer to the partition to split, might be changed!
+ * @param gg a (non-empty) node list
+ * @param env the environment
+ *
+ * @return a new partition containing the nodes of gg
+ */
+static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
+ partition_t *X = *pX;
+ partition_t *X_prime;
+ list_head tmp;
+ step_env senv[2];
+ node_t *g, *h, *node, *t;
+ int max_input, transitions, winner, shf;
+ unsigned n;
+ DEBUG_ONLY(static int run = 0;)
+
+ DB((dbg, LEVEL_2, "Run %d ", run++));
+ if (list_empty(&X->Follower)) {
+ /* if the partition has NO follower, we can use the fast
+ splitting algorithm. */
+ return split_no_followers(X, gg, env);
+ }
+ /* else do the race */
+
+ dump_partition("Splitting ", X);
+ dump_list("by list ", gg);
+
+ INIT_LIST_HEAD(&tmp);
+
+ /* Remove gg from X.Leader and put into g */
+ g = NULL;
+ for (node = gg; node != NULL; node = node->next) {
+ assert(node->part == X);
+ assert(node->is_follower == 0);
+
+ list_del(&node->node_list);
+ list_add_tail(&node->node_list, &tmp);
+ node->race_next = g;
+ g = node;
+ }
+ /* produce h */
+ h = NULL;
+ list_for_each_entry(node_t, node, &X->Leader, node_list) {
+ node->race_next = h;
+ h = node;
+ }
+ /* restore X.Leader */
+ list_splice(&tmp, &X->Leader);
+
+ senv[0].initial = g;
+ senv[0].unwalked = NULL;
+ senv[0].walked = NULL;
+ senv[0].index = 0;
+ senv[0].side = 1;
+
+ senv[1].initial = h;
+ senv[1].unwalked = NULL;
+ senv[1].walked = NULL;
+ senv[1].index = 0;
+ senv[1].side = 2;
+
+ /*
+ * Some informations on the race that are not stated clearly in Click's
+ * thesis.
+ * 1) A follower stays on the side that reach him first.
+ * 2) If the other side reches a follower, if will be converted to
+ * a leader. /This must be done after the race is over, else the
+ * edges we are iterating on are renumbered./
+ * 3) /New leader might end up on both sides./
+ * 4) /If one side ends up with new Leaders, we must ensure that
+ * they can split out by opcode, hence we have to put _every_
+ * partition with new Leader nodes on the cprop list, as
+ * opcode splitting is done by split_by() at the end of
+ * constant propagation./
+ */
+ for (;;) {
+ if (step(&senv[0])) {
+ winner = 0;
+ break;
+ }
+ if (step(&senv[1])) {
+ winner = 1;
+ break;
+ }
+ }
+ assert(senv[winner].initial == NULL);
+ assert(senv[winner].unwalked == NULL);
+
+ /* clear flags from walked/unwalked */
+ shf = winner;
+ transitions = clear_flags(senv[0].unwalked) << shf;
+ transitions |= clear_flags(senv[0].walked) << shf;
+ shf ^= 1;
+ transitions |= clear_flags(senv[1].unwalked) << shf;
+ transitions |= clear_flags(senv[1].walked) << shf;
+
+ dump_race_list("winner ", senv[winner].walked);
+
+ /* Move walked_{winner} to a new partition, X'. */
+ X_prime = new_partition(env);
+ max_input = 0;
+ n = 0;
+ for (node = senv[winner].walked; node != NULL; node = node->race_next) {
+ list_del(&node->node_list);
+ node->part = X_prime;
+ if (node->is_follower) {
+ list_add_tail(&node->node_list, &X_prime->Follower);
+ } else {
+ list_add_tail(&node->node_list, &X_prime->Leader);
+ ++n;
+ }
+ if (node->max_user_input > max_input)
+ max_input = node->max_user_input;
+ }
+ X_prime->n_leader = n;
+ X_prime->max_user_inputs = max_input;
+ X->n_leader -= X_prime->n_leader;
+
+ /* for now, copy the type info tag, it will be adjusted in split_by(). */
+ X_prime->type_is_T_or_C = X->type_is_T_or_C;
+
+ /*
+ * Even if a follower was not checked by both sides, it might have
+ * loose its congruence, so we need to check this case for all follower.
+ */
+ list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
+ if (identity(node) == node) {
+ follower_to_leader(node);
+ transitions |= 1;
+ }
+ }
+
+ check_partition(X);
+ check_partition(X_prime);
+
+ /* X' is the smaller part */
+ add_to_worklist(X_prime, env);
+
+ /*
+ * If there where follower to leader transitions, ensure that the nodes
+ * can be split out if necessary.
+ */
+ if (transitions & 1) {
+ /* place winner partition on the cprop list */
+ if (X_prime->on_cprop == 0) {
+ X_prime->cprop_next = env->cprop;
+ env->cprop = X_prime;
+ X_prime->on_cprop = 1;
+ }
+ }
+ if (transitions & 2) {
+ /* place other partition on the cprop list */
+ if (X->on_cprop == 0) {
+ X->cprop_next = env->cprop;
+ env->cprop = X;
+ X->on_cprop = 1;
+ }
+ }
+
+ dump_partition("Now ", X);
+ dump_partition("Created new ", X_prime);
+
+ /* we have to ensure that the partition containing g is returned */
+ if (winner != 0) {
+ *pX = X_prime;
+ return X;
+ }
+
+ return X_prime;
+} /* split */
+
+/**
+ * Returns non-zero if the i'th input of a Phi node is live.
+ *
+ * @param phi a Phi-node
+ * @param i an input number
+ *
+ * @return non-zero if the i'th input of the given Phi node is live
+ */
+static int is_live_input(ir_node *phi, int i) {
+ if (i >= 0) {
+ ir_node *block = get_nodes_block(phi);
+ ir_node *pred = get_Block_cfgpred(block, i);
+ lattice_elem_t type = get_node_type(pred);
+
+ return type.tv != tarval_unreachable;
+ }
+ /* else it's the control input, always live */
+ return 1;
+} /* is_live_input */
+
+/**
+ * Return non-zero if a type is a constant.
+ */
+static int is_constant_type(lattice_elem_t type) {
+ if (type.tv != tarval_bottom && type.tv != tarval_top)
+ return 1;
+ return 0;
+} /* is_constant_type */
+
+/**
+ * Check whether a type is neither Top or a constant.
+ * Note: U is handled like Top here, R is a constant.
+ *
+ * @param type the type to check
+ */
+static int type_is_neither_top_nor_const(const lattice_elem_t type) {
+ if (is_tarval(type.tv)) {
+ if (type.tv == tarval_top)
+ return 0;
+ if (tarval_is_constant(type.tv))
+ return 0;
+ } else {
+ /* is a symconst */
+ return 0;
+ }
+ return 1;
+} /* type_is_neither_top_nor_const */
+
+/**
+ * Collect nodes to the touched list.
+ *
+ * @param list the list which contains the nodes that must be evaluated
+ * @param idx the index of the def_use edge to evaluate
+ * @param env the environment
+ */
+static void collect_touched(list_head *list, int idx, environment_t *env) {
+ node_t *x, *y;
+ int end_idx = env->end_idx;
+
+ list_for_each_entry(node_t, x, list, node_list) {
+ int num_edges;
+
+ if (idx == -1) {
+ /* leader edges start AFTER follower edges */
+ x->next_edge = x->n_followers + 1;
+ }
+ num_edges = get_irn_n_outs(x->node);
+
+ /* for all edges in x.L.def_use_{idx} */
+ while (x->next_edge <= num_edges) {
+ const ir_def_use_edge *edge = &x->node->out[x->next_edge];
+ ir_node *succ;
+
+ /* check if we have necessary edges */
+ if (edge->pos > idx)
+ break;
+
+ ++x->next_edge;
+
+ succ = edge->use;
+
+ /* only non-commutative nodes */
+ if (env->commutative &&
+ (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
+ continue;
+
+ /* ignore the "control input" for non-pinned nodes
+ if we are running in GCSE mode */
+ if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
+ continue;
+
+ y = get_irn_node(succ);
+ assert(get_irn_n(succ, idx) == x->node);
+
+ /* ignore block edges touching followers */
+ if (idx == -1 && y->is_follower)
+ continue;
+
+ if (is_constant_type(y->type)) {
+ ir_opcode code = get_irn_opcode(succ);
+ if (code == iro_Sub || code == iro_Cmp)
+ add_to_cprop(y, env);
+ }
+
+ /* Partitions of constants should not be split simply because their Nodes have unequal
+ functions or incongruent inputs. */
+ if (type_is_neither_top_nor_const(y->type) &&
+ (! is_Phi(y->node) || is_live_input(y->node, idx))) {
+ add_to_touched(y, env);
+ }
+ }
+ }
+} /* collect_touched */
+
+/**
+ * Collect commutative nodes to the touched list.
+ *
+ * @param list the list which contains the nodes that must be evaluated
+ * @param env the environment
+ */
+static void collect_commutative_touched(list_head *list, environment_t *env) {
+ node_t *x, *y;
+
+ list_for_each_entry(node_t, x, list, node_list) {
+ int num_edges;
+
+ num_edges = get_irn_n_outs(x->node);
+
+ x->next_edge = x->n_followers + 1;
+
+ /* for all edges in x.L.def_use_{idx} */
+ while (x->next_edge <= num_edges) {
+ const ir_def_use_edge *edge = &x->node->out[x->next_edge];
+ ir_node *succ;
+
+ /* check if we have necessary edges */
+ if (edge->pos > 1)
+ break;
+
+ ++x->next_edge;
+ if (edge->pos < 0)
+ continue;
+
+ succ = edge->use;
+
+ /* only commutative nodes */
+ if (!is_op_commutative(get_irn_op(succ)))
+ continue;
+
+ y = get_irn_node(succ);
+ if (is_constant_type(y->type)) {
+ ir_opcode code = get_irn_opcode(succ);
+ if (code == iro_Eor)
+ add_to_cprop(y, env);
+ }
+
+ /* Partitions of constants should not be split simply because their Nodes have unequal
+ functions or incongruent inputs. */
+ if (type_is_neither_top_nor_const(y->type)) {
+ add_to_touched(y, env);
}
}
+ }
+} /* collect_commutative_touched */
+
+/**
+ * Split the partitions if caused by the first entry on the worklist.
+ *
+ * @param env the environment
+ */
+static void cause_splits(environment_t *env) {
+ partition_t *X, *Z, *N;
+ int idx;
+
+ /* remove the first partition from the worklist */
+ X = env->worklist;
+ env->worklist = X->wl_next;
+ X->on_worklist = 0;
+
+ dump_partition("Cause_split: ", X);
+
+ if (env->commutative) {
+ /* handle commutative nodes first */
+
+ /* empty the touched set: already done, just clear the list */
+ env->touched = NULL;
+
+ collect_commutative_touched(&X->Leader, env);
+ collect_commutative_touched(&X->Follower, env);
+
+ for (Z = env->touched; Z != NULL; Z = N) {
+ node_t *e;
+ node_t *touched = Z->touched;
+ unsigned n_touched = Z->n_touched;
+
+ assert(Z->touched != NULL);
+
+ /* beware, split might change Z */
+ N = Z->touched_next;
- for (Z = env->touched; Z != NULL; Z = Z->touched_next) {
/* remove it from the touched set */
Z->on_touched = 0;
- if (Z->n_nodes != Z->n_touched) {
- DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
- split(Z, Z->touched, env);
+ /* Empty local Z.touched. */
+ for (e = touched; e != NULL; e = e->next) {
+ assert(e->is_follower == 0);
+ e->on_touched = 0;
}
+ Z->touched = NULL;
+ Z->n_touched = 0;
+
+ if (0 < n_touched && n_touched < Z->n_leader) {
+ DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
+ split(&Z, touched, env);
+ } else
+ assert(n_touched <= Z->n_leader);
+ }
+ }
+
+ /* combine temporary leader and follower list */
+ for (idx = -1; idx <= X->max_user_inputs; ++idx) {
+ /* empty the touched set: already done, just clear the list */
+ env->touched = NULL;
+
+ collect_touched(&X->Leader, idx, env);
+ collect_touched(&X->Follower, idx, env);
+
+ for (Z = env->touched; Z != NULL; Z = N) {
+ node_t *e;
+ node_t *touched = Z->touched;
+ unsigned n_touched = Z->n_touched;
+
+ assert(Z->touched != NULL);
+
+ /* beware, split might change Z */
+ N = Z->touched_next;
+
+ /* remove it from the touched set */
+ Z->on_touched = 0;
+
/* Empty local Z.touched. */
- for (e = Z->touched; e != NULL; e = e->next) {
+ for (e = touched; e != NULL; e = e->next) {
+ assert(e->is_follower == 0);
e->on_touched = 0;
}
Z->touched = NULL;
Z->n_touched = 0;
+
+ if (0 < n_touched && n_touched < Z->n_leader) {
+ DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
+ split(&Z, touched, env);
+ } else
+ assert(n_touched <= Z->n_leader);
}
}
} /* cause_splits */
*
* @param X the partition to split
* @param What a function returning an Id for every node of the partition X
- * @param P an flexible array to store the result partitions or NULL
+ * @param P a list to store the result partitions
* @param env the environment
*
- * @return if P != NULL P will be filled with the resulting partitions and returned
+ * @return *P
*/
-static partition_t **split_by_what(partition_t *X, what_func What,
- partition_t **P, environment_t *env) {
+static partition_t *split_by_what(partition_t *X, what_func What,
+ partition_t **P, environment_t *env) {
node_t *x, *S;
listmap_t map;
listmap_entry_t *iter;
/* Let map be an empty mapping from the range of What to (local) list of Nodes. */
listmap_init(&map);
- list_for_each_entry(node_t, x, &X->entries, node_list) {
+ list_for_each_entry(node_t, x, &X->Leader, node_list) {
void *id = What(x, env);
listmap_entry_t *entry;
S = iter->list;
/* Add SPLIT( X, S ) to P. */
- DB((dbg, LEVEL_2, "Split part%d by what\n", X->nr));
- R = split(X, S, env);
- if (P != NULL) {
- ARR_APP1(partition_t *, P, R);
- }
+ DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
+ R = split(&X, S, env);
+ R->split_next = *P;
+ *P = R;
}
/* Add X to P. */
- if (P != NULL) {
- ARR_APP1(partition_t *, P, X);
- }
+ X->split_next = *P;
+ *P = X;
listmap_term(&map);
- return P;
+ return *P;
} /* split_by_what */
/** lambda n.(n.type) */
key.code = get_irn_opcode(irn);
key.mode = get_irn_mode(irn);
+ key.arity = get_irn_arity(irn);
key.u.proj = 0;
key.u.ent = NULL;
int i = env->lambda_input;
if (i >= get_irn_arity(node->node)) {
- /* we are outside the allowed range */
+ /*
+ * We are outside the allowed range: This can happen even
+ * if we have split by opcode first: doing so might move Followers
+ * to Leaders and those will have a different opcode!
+ * Note that in this case the partition is on the cprop list and will be
+ * split again.
+ */
return NULL;
}
return p->part;
} /* lambda_partition */
+/** lambda n.(n[i].partition) for commutative nodes */
+static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
+ ir_node *irn = node->node;
+ ir_node *skipped = skip_Proj(irn);
+ ir_node *pred, *left, *right;
+ node_t *p;
+ partition_t *pl, *pr;
+ int i = env->lambda_input;
+
+ if (i >= get_irn_arity(node->node)) {
+ /*
+ * We are outside the allowed range: This can happen even
+ * if we have split by opcode first: doing so might move Followers
+ * to Leaders and those will have a different opcode!
+ * Note that in this case the partition is on the cprop list and will be
+ * split again.
+ */
+ return NULL;
+ }
+
+ /* ignore the "control input" for non-pinned nodes
+ if we are running in GCSE mode */
+ if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
+ return NULL;
+
+ if (i == -1) {
+ pred = get_irn_n(skipped, i);
+ p = get_irn_node(pred);
+ return p->part;
+ }
+
+ if (is_op_commutative(get_irn_op(irn))) {
+ /* normalize partition order by returning the "smaller" on input 0,
+ the "bigger" on input 1. */
+ left = get_binop_left(irn);
+ pl = get_irn_node(left)->part;
+ right = get_binop_right(irn);
+ pr = get_irn_node(right)->part;
+
+ if (i == 0)
+ return pl < pr ? pl : pr;
+ else
+ return pl > pr ? pl : pr;
+ } else {
+ /* a not split out Follower */
+ pred = get_irn_n(irn, i);
+ p = get_irn_node(pred);
+
+ return p->part;
+ }
+} /* lambda_commutative_partition */
+
/**
- * Checks whether a type is a constant.
+ * Returns true if a type is a constant (and NOT Top
+ * or Bottom).
*/
-static int is_type_constant(lattice_elem_t type) {
+static int is_con(const lattice_elem_t type) {
+ /* be conservative */
if (is_tarval(type.tv))
return tarval_is_constant(type.tv);
- /* else it is a symconst */
- return 1;
-}
+ return is_entity(type.sym.entity_p);
+} /* is_con */
/**
* Implements split_by().
* @param env the environment
*/
static void split_by(partition_t *X, environment_t *env) {
- partition_t **P = NEW_ARR_F(partition_t *, 0);
- int i, j, k;
+ partition_t *I, *P = NULL;
+ int input;
- DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr));
- P = split_by_what(X, lambda_type, P, env);
- for (i = ARR_LEN(P) - 1; i >= 0; --i) {
- partition_t *Y = P[i];
+ dump_partition("split_by", X);
- if (Y->n_nodes > 1) {
- lattice_elem_t type = get_partition_type(Y);
+ if (X->n_leader == 1) {
+ /* we have only one leader, no need to split, just check it's type */
+ node_t *x = get_first_node(X);
+ X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
+ return;
+ }
- /* we do not want split the TOP or constant partitions */
- if (type.tv != tarval_top && !is_type_constant(type)) {
- partition_t **Q = NEW_ARR_F(partition_t *, 0);
+ DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
+ P = split_by_what(X, lambda_type, &P, env);
+ dump_split_list(P);
- DB((dbg, LEVEL_2, "WHAT = lambda n.(n.opcode) on part%d\n", Y->nr));
- Q = split_by_what(Y, lambda_opcode, Q, env);
+ /* adjust the type tags, we have split partitions by type */
+ for (I = P; I != NULL; I = I->split_next) {
+ node_t *x = get_first_node(I);
+ I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
+ }
- for (j = ARR_LEN(Q) - 1; j >= 0; --j) {
- partition_t *Z = Q[j];
+ do {
+ partition_t *Y = P;
- for (k = Z->max_arity - 1; k >= -1; --k) {
- if (Z->n_nodes > 1) {
- env->lambda_input = k;
- DB((dbg, LEVEL_2, "WHAT = lambda n.(n[%d].partition) on part%d\n", k, Z->nr));
- split_by_what(Z, lambda_partition, NULL, env);
+ P = P->split_next;
+ if (Y->n_leader > 1) {
+ /* we do not want split the TOP or constant partitions */
+ if (! Y->type_is_T_or_C) {
+ partition_t *Q = NULL;
+
+ DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
+ Q = split_by_what(Y, lambda_opcode, &Q, env);
+ dump_split_list(Q);
+
+ do {
+ partition_t *Z = Q;
+
+ Q = Q->split_next;
+ if (Z->n_leader > 1) {
+ const node_t *first = get_first_node(Z);
+ int arity = get_irn_arity(first->node);
+ partition_t *R, *S;
+ what_func what = lambda_partition;
+ DEBUG_ONLY(char buf[64];)
+
+ if (env->commutative && is_op_commutative(get_irn_op(first->node)))
+ what = lambda_commutative_partition;
+
+ /*
+ * BEWARE: during splitting by input 2 for instance we might
+ * create new partitions which are different by input 1, so collect
+ * them and split further.
+ */
+ Z->split_next = NULL;
+ R = Z;
+ S = NULL;
+ for (input = arity - 1; input >= -1; --input) {
+ do {
+ partition_t *Z_prime = R;
+
+ R = R->split_next;
+ if (Z_prime->n_leader > 1) {
+ env->lambda_input = input;
+ DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
+ DEBUG_ONLY(what_reason = buf;)
+ S = split_by_what(Z_prime, what, &S, env);
+ dump_split_list(S);
+ } else {
+ Z_prime->split_next = S;
+ S = Z_prime;
+ }
+ } while (R != NULL);
+ R = S;
+ S = NULL;
}
}
- }
- DEL_ARR_F(Q);
+ } while (Q != NULL);
}
}
- }
- DEL_ARR_F(P);
+ } while (P != NULL);
} /* split_by */
/**
int i;
ir_node *block = node->node;
+ if (block == get_irg_start_block(current_ir_graph)) {
+ /* start block is always reachable */
+ node->type.tv = tarval_reachable;
+ return;
+ }
+
for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
* @param node the node
*/
static void compute_Unknown(node_t *node) {
- /* While Unknown nodes compute Top, but this is dangerous:
- * a if (unknown) would lead to BOTH control flows unreachable.
+ /* While Unknown nodes should compute Top this is dangerous:
+ * a Top input to a Cond would lead to BOTH control flows unreachable.
* While this is correct in the given semantics, it would destroy the Firm
* graph.
- * For now, we compute bottom here.
+ *
+ * It would be safe to compute Top IF it can be assured, that only Cmp
+ * nodes are inputs to Conds. We check that first.
+ * This is the way Frontends typically build Firm, but some optimizations
+ * (cond_eval for instance) might replace them by Phib's...
*/
- node->type.tv = tarval_bottom;
+ node->type.tv = tarval_UNKNOWN;
} /* compute_Unknown */
/**
node->type = block->type;
} /* compute_Jmp */
+/**
+ * (Re-)compute the type for the Return node.
+ *
+ * @param node the node
+ */
+static void compute_Return(node_t *node) {
+ /* The Return node is NOT dead if it is in a reachable block.
+ * This is already checked in compute(). so we can return
+ * Reachable here. */
+ node->type.tv = tarval_reachable;
+} /* compute_Return */
+
/**
* (Re-)compute the type for the End node.
*
static void compute_End(node_t *node) {
/* the End node is NOT dead of course */
node->type.tv = tarval_reachable;
-}
+} /* compute_End */
/**
* (Re-)compute the type for a SymConst node.
node_t *r = get_irn_node(get_Add_right(sub));
lattice_elem_t a = l->type;
lattice_elem_t b = r->type;
- node_t *block = get_irn_node(get_nodes_block(sub));
ir_mode *mode;
- if (block->type.tv == tarval_unreachable) {
- node->type.tv = tarval_top;
- return;
- }
-
if (a.tv == tarval_top || b.tv == tarval_top) {
node->type.tv = tarval_top;
} else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
}
} /* compute_Add */
-/**
- * Returns true if a type is a constant.
- */
-static int is_con(const lattice_elem_t type) {
- return is_entity(type.sym.entity_p) || tarval_is_constant(type.tv);
-}
-
/**
* (Re-)compute the type for a Sub. Special case: both nodes are congruent.
*
node_t *r = get_irn_node(get_Sub_right(sub));
lattice_elem_t a = l->type;
lattice_elem_t b = r->type;
- node_t *block = get_irn_node(get_nodes_block(sub));
+ tarval *tv;
+
+ if (a.tv == tarval_top || b.tv == tarval_top) {
+ node->type.tv = tarval_top;
+ } else if (is_con(a) && is_con(b)) {
+ if (is_tarval(a.tv) && is_tarval(b.tv)) {
+ node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
+ } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
+ node->type = b;
+ } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
+ node->type = a;
+ } else {
+ node->type.tv = tarval_bottom;
+ }
+ node->by_all_const = 1;
+ } else if (r->part == l->part &&
+ (!mode_is_float(get_irn_mode(l->node)))) {
+ /*
+ * BEWARE: a - a is NOT always 0 for floating Point values, as
+ * NaN op NaN = NaN, so we must check this here.
+ */
+ ir_mode *mode = get_irn_mode(sub);
+ tv = get_mode_null(mode);
+
+ /* if the node was ONCE evaluated by all constants, but now
+ this breaks AND we get from the argument partitions a different
+ result, switch to bottom.
+ This happens because initially all nodes are in the same partition ... */
+ if (node->by_all_const && node->type.tv != tv)
+ tv = tarval_bottom;
+ node->type.tv = tv;
+ } else {
+ node->type.tv = tarval_bottom;
+ }
+} /* compute_Sub */
+
+/**
+ * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
+ *
+ * @param node the node
+ */
+static void compute_Eor(node_t *node) {
+ ir_node *eor = node->node;
+ node_t *l = get_irn_node(get_Eor_left(eor));
+ node_t *r = get_irn_node(get_Eor_right(eor));
+ lattice_elem_t a = l->type;
+ lattice_elem_t b = r->type;
+ tarval *tv;
+
+ if (a.tv == tarval_top || b.tv == tarval_top) {
+ node->type.tv = tarval_top;
+ } else if (is_con(a) && is_con(b)) {
+ if (is_tarval(a.tv) && is_tarval(b.tv)) {
+ node->type.tv = tarval_eor(a.tv, b.tv);
+ } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
+ node->type = b;
+ } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
+ node->type = a;
+ } else {
+ node->type.tv = tarval_bottom;
+ }
+ node->by_all_const = 1;
+ } else if (r->part == l->part) {
+ ir_mode *mode = get_irn_mode(eor);
+ tv = get_mode_null(mode);
+
+ /* if the node was ONCE evaluated by all constants, but now
+ this breaks AND we get from the argument partitions a different
+ result, switch to bottom.
+ This happens because initially all nodes are in the same partition ... */
+ if (node->by_all_const && node->type.tv != tv)
+ tv = tarval_bottom;
+ node->type.tv = tv;
+ } else {
+ node->type.tv = tarval_bottom;
+ }
+} /* compute_Eor */
+
+/**
+ * (Re-)compute the type for Cmp.
+ *
+ * @param node the node
+ */
+static void compute_Cmp(node_t *node) {
+ ir_node *cmp = node->node;
+ node_t *l = get_irn_node(get_Cmp_left(cmp));
+ node_t *r = get_irn_node(get_Cmp_right(cmp));
+ lattice_elem_t a = l->type;
+ lattice_elem_t b = r->type;
- if (block->type.tv == tarval_unreachable) {
- node->type.tv = tarval_top;
- return;
- }
if (a.tv == tarval_top || b.tv == tarval_top) {
node->type.tv = tarval_top;
+ } else if (r->part == l->part) {
+ /* both nodes congruent, we can probably do something */
+ node->type.tv = tarval_b_true;
} else if (is_con(a) && is_con(b)) {
- if (is_tarval(a.tv) && is_tarval(b.tv)) {
- node->type.tv = tarval_sub(a.tv, b.tv);
- } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
- node->type = b;
- } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
- node->type = a;
- } else {
- node->type.tv = tarval_bottom;
- }
- } else if (r->part == l->part &&
- (!mode_is_float(get_irn_mode(l->node)))) {
- if (node->type.tv == tarval_top) {
- /*
- * BEWARE: a - a is NOT always 0 for floating Point values, as
- * NaN op NaN = NaN, so we must check this here.
- */
- ir_mode *mode = get_irn_mode(sub);
- node->type.tv = get_mode_null(mode);
- } else {
- node->type.tv = tarval_bottom;
- }
+ /* both nodes are constants, we can probably do something */
+ node->type.tv = tarval_b_true;
} else {
node->type.tv = tarval_bottom;
}
-} /* compute_Sub */
+} /* compute_Cmp */
/**
* (Re-)compute the type for a Proj(Cmp).
lattice_elem_t a = l->type;
lattice_elem_t b = r->type;
pn_Cmp pnc = get_Proj_proj(proj);
+ tarval *tv;
if (a.tv == tarval_top || b.tv == tarval_top) {
- node->type.tv = tarval_top;
+ node->type.tv = tarval_undefined;
} else if (is_con(a) && is_con(b)) {
default_compute(node);
+ node->by_all_const = 1;
} else if (r->part == l->part &&
(!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
- if (node->type.tv == tarval_top) {
- /*
- * BEWARE: a == a is NOT always True for floating Point values, as
- * NaN != NaN is defined, so we must check this here.
- */
- node->type.tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b);
- } else {
- node->type.tv = tarval_bottom;
- }
+ /*
+ * BEWARE: a == a is NOT always True for floating Point values, as
+ * NaN != NaN is defined, so we must check this here.
+ */
+ tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
+
+ /* if the node was ONCE evaluated by all constants, but now
+ this breaks AND we get from the argument partitions a different
+ result, switch to bottom.
+ This happens because initially all nodes are in the same partition ... */
+ if (node->by_all_const && node->type.tv != tv)
+ tv = tarval_bottom;
+ node->type.tv = tv;
} else {
node->type.tv = tarval_bottom;
}
node->type.tv = tarval_reachable;
} else {
assert(selector->type.tv == tarval_top);
+ /* any condition based on Top is "!=" */
node->type.tv = tarval_unreachable;
}
} else {
node->type.tv = tarval_reachable;
} else {
assert(selector->type.tv == tarval_top);
- node->type.tv = tarval_unreachable;
+ /* any condition based on Top is "!=" */
+ node->type.tv = tarval_reachable;
}
}
} else {
if (selector->type.tv == tarval_bottom) {
node->type.tv = tarval_reachable;
} else if (selector->type.tv == tarval_top) {
- node->type.tv = tarval_unreachable;
+ if (pnc == get_Cond_defaultProj(cond)) {
+ /* a switch based of Top is always "default" */
+ node->type.tv = tarval_reachable;
+ } else
+ node->type.tv = tarval_unreachable;
} else {
long value = get_tarval_long(selector->type.tv);
if (pnc == get_Cond_defaultProj(cond)) {
} /* compute_Proj_Cond */
/**
- * (Re-)compute the type for a Proj-Nodes.
+ * (Re-)compute the type for a Proj-Node.
*
* @param node the node
*/
node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
ir_node *pred = get_Proj_pred(proj);
- if (get_Proj_proj(proj) == pn_Start_X_initial_exec && is_Start(pred)) {
- /* The initial_exec node is ALWAYS reachable. */
- node->type.tv = tarval_reachable;
- return;
- }
-
if (block->type.tv == tarval_unreachable) {
/* a Proj in a unreachable Block stay Top */
node->type.tv = tarval_top;
return;
}
- if (get_irn_node(pred)->type.tv == tarval_top) {
+ if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
/* if the predecessor is Top, its Proj follow */
node->type.tv = tarval_top;
return;
} /* compute_Proj */
/**
- * (Re-)compute the type for a Confirm-Nodes.
+ * (Re-)compute the type for a Confirm.
*
* @param node the node
*/
node->type = pred->type;
} /* compute_Confirm */
+/**
+ * (Re-)compute the type for a Max.
+ *
+ * @param node the node
+ */
+static void compute_Max(node_t *node) {
+ ir_node *op = node->node;
+ node_t *l = get_irn_node(get_binop_left(op));
+ node_t *r = get_irn_node(get_binop_right(op));
+ lattice_elem_t a = l->type;
+ lattice_elem_t b = r->type;
+
+ if (a.tv == tarval_top || b.tv == tarval_top) {
+ node->type.tv = tarval_top;
+ } else if (is_con(a) && is_con(b)) {
+ /* both nodes are constants, we can probably do something */
+ if (a.tv == b.tv) {
+ /* this case handles SymConsts as well */
+ node->type = a;
+ } else {
+ ir_mode *mode = get_irn_mode(op);
+ tarval *tv_min = get_mode_min(mode);
+
+ if (a.tv == tv_min)
+ node->type = b;
+ else if (b.tv == tv_min)
+ node->type = a;
+ else if (is_tarval(a.tv) && is_tarval(b.tv)) {
+ if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
+ node->type.tv = a.tv;
+ else
+ node->type.tv = b.tv;
+ } else {
+ node->type.tv = tarval_bad;
+ }
+ }
+ } else if (r->part == l->part) {
+ /* both nodes congruent, we can probably do something */
+ node->type = a;
+ } else {
+ node->type.tv = tarval_bottom;
+ }
+} /* compute_Max */
+
+/**
+ * (Re-)compute the type for a Min.
+ *
+ * @param node the node
+ */
+static void compute_Min(node_t *node) {
+ ir_node *op = node->node;
+ node_t *l = get_irn_node(get_binop_left(op));
+ node_t *r = get_irn_node(get_binop_right(op));
+ lattice_elem_t a = l->type;
+ lattice_elem_t b = r->type;
+
+ if (a.tv == tarval_top || b.tv == tarval_top) {
+ node->type.tv = tarval_top;
+ } else if (is_con(a) && is_con(b)) {
+ /* both nodes are constants, we can probably do something */
+ if (a.tv == b.tv) {
+ /* this case handles SymConsts as well */
+ node->type = a;
+ } else {
+ ir_mode *mode = get_irn_mode(op);
+ tarval *tv_max = get_mode_max(mode);
+
+ if (a.tv == tv_max)
+ node->type = b;
+ else if (b.tv == tv_max)
+ node->type = a;
+ else if (is_tarval(a.tv) && is_tarval(b.tv)) {
+ if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
+ node->type.tv = a.tv;
+ else
+ node->type.tv = b.tv;
+ } else {
+ node->type.tv = tarval_bad;
+ }
+ }
+ } else if (r->part == l->part) {
+ /* both nodes congruent, we can probably do something */
+ node->type = a;
+ } else {
+ node->type.tv = tarval_bottom;
+ }
+} /* compute_Min */
+
/**
* (Re-)compute the type for a given node.
*
* @param node the node
*/
static void compute(node_t *node) {
- compute_func func = (compute_func)node->node->op->ops.generic;
+ ir_node *irn = node->node;
+ compute_func func;
+
+ if (is_no_Block(irn)) {
+ /* for pinned nodes, check its control input */
+ if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
+ node_t *block = get_irn_node(get_nodes_block(irn));
+ if (block->type.tv == tarval_unreachable) {
+ node->type.tv = tarval_top;
+ return;
+ }
+ }
+ }
+
+ func = (compute_func)node->node->op->ops.generic;
if (func != NULL)
func(node);
} /* compute */
+/*
+ * Identity functions: Note that one might thing that identity() is just a
+ * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
+ * here, because it expects that the identity node is one of the inputs, which is NOT
+ * always true for equivalent_node() which can handle (and does sometimes) DAGs.
+ * So, we have our own implementation, which copies some parts of equivalent_node()
+ */
+
+/**
+ * Calculates the Identity for Phi nodes
+ */
+static node_t *identity_Phi(node_t *node) {
+ ir_node *phi = node->node;
+ ir_node *block = get_nodes_block(phi);
+ node_t *n_part = NULL;
+ int i;
+
+ for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
+ node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
+
+ if (pred_X->type.tv == tarval_reachable) {
+ node_t *pred = get_irn_node(get_Phi_pred(phi, i));
+
+ if (n_part == NULL)
+ n_part = pred;
+ else if (n_part->part != pred->part) {
+ /* incongruent inputs, not a follower */
+ return node;
+ }
+ }
+ }
+ /* if n_part is NULL here, all inputs path are dead, the Phi computes
+ * tarval_top, is in the TOP partition and should NOT being split! */
+ assert(n_part != NULL);
+ return n_part;
+} /* identity_Phi */
+
+/**
+ * Calculates the Identity for commutative 0 neutral nodes.
+ */
+static node_t *identity_comm_zero_binop(node_t *node) {
+ ir_node *op = node->node;
+ node_t *a = get_irn_node(get_binop_left(op));
+ node_t *b = get_irn_node(get_binop_right(op));
+ ir_mode *mode = get_irn_mode(op);
+ tarval *zero;
+
+ /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
+ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
+ return node;
+
+ /* node: no input should be tarval_top, else the binop would be also
+ * Top and not being split. */
+ zero = get_mode_null(mode);
+ if (a->type.tv == zero)
+ return b;
+ if (b->type.tv == zero)
+ return a;
+ return node;
+} /* identity_comm_zero_binop */
+
+/**
+ * Calculates the Identity for Shift nodes.
+ */
+static node_t *identity_shift(node_t *node) {
+ ir_node *op = node->node;
+ node_t *b = get_irn_node(get_binop_right(op));
+ ir_mode *mode = get_irn_mode(b->node);
+ tarval *zero;
+
+ /* node: no input should be tarval_top, else the binop would be also
+ * Top and not being split. */
+ zero = get_mode_null(mode);
+ if (b->type.tv == zero)
+ return get_irn_node(get_binop_left(op));
+ return node;
+} /* identity_shift */
+
+/**
+ * Calculates the Identity for Mul nodes.
+ */
+static node_t *identity_Mul(node_t *node) {
+ ir_node *op = node->node;
+ node_t *a = get_irn_node(get_Mul_left(op));
+ node_t *b = get_irn_node(get_Mul_right(op));
+ ir_mode *mode = get_irn_mode(op);
+ tarval *one;
+
+ /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
+ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
+ return node;
+
+ /* node: no input should be tarval_top, else the binop would be also
+ * Top and not being split. */
+ one = get_mode_one(mode);
+ if (a->type.tv == one)
+ return b;
+ if (b->type.tv == one)
+ return a;
+ return node;
+} /* identity_Mul */
+
+/**
+ * Calculates the Identity for Sub nodes.
+ */
+static node_t *identity_Sub(node_t *node) {
+ ir_node *sub = node->node;
+ node_t *b = get_irn_node(get_Sub_right(sub));
+ ir_mode *mode = get_irn_mode(sub);
+
+ /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
+ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
+ return node;
+
+ /* node: no input should be tarval_top, else the binop would be also
+ * Top and not being split. */
+ if (b->type.tv == get_mode_null(mode))
+ return get_irn_node(get_Sub_left(sub));
+ return node;
+} /* identity_Sub */
+
+/**
+ * Calculates the Identity for And nodes.
+ */
+static node_t *identity_And(node_t *node) {
+ ir_node *and = node->node;
+ node_t *a = get_irn_node(get_And_left(and));
+ node_t *b = get_irn_node(get_And_right(and));
+ tarval *neutral = get_mode_all_one(get_irn_mode(and));
+
+ /* node: no input should be tarval_top, else the And would be also
+ * Top and not being split. */
+ if (a->type.tv == neutral)
+ return b;
+ if (b->type.tv == neutral)
+ return a;
+ return node;
+} /* identity_And */
+
+/**
+ * Calculates the Identity for Confirm nodes.
+ */
+static node_t *identity_Confirm(node_t *node) {
+ ir_node *confirm = node->node;
+
+ /* a Confirm is always a Copy */
+ return get_irn_node(get_Confirm_value(confirm));
+} /* identity_Confirm */
+
+/**
+ * Calculates the Identity for Mux nodes.
+ */
+static node_t *identity_Mux(node_t *node) {
+ ir_node *mux = node->node;
+ node_t *t = get_irn_node(get_Mux_true(mux));
+ node_t *f = get_irn_node(get_Mux_false(mux));
+ /*node_t *sel; */
+
+ if (t->part == f->part)
+ return t;
+
+ /* for now, the 1-input identity is not supported */
+#if 0
+ sel = get_irn_node(get_Mux_sel(mux));
+
+ /* Mux sel input is mode_b, so it is always a tarval */
+ if (sel->type.tv == tarval_b_true)
+ return t;
+ if (sel->type.tv == tarval_b_false)
+ return f;
+#endif
+ return node;
+} /* identity_Mux */
+
+/**
+ * Calculates the Identity for Min nodes.
+ */
+static node_t *identity_Min(node_t *node) {
+ ir_node *op = node->node;
+ node_t *a = get_irn_node(get_binop_left(op));
+ node_t *b = get_irn_node(get_binop_right(op));
+ ir_mode *mode = get_irn_mode(op);
+ tarval *tv_max;
+
+ if (a->part == b->part) {
+ /* leader of multiple predecessors */
+ return a;
+ }
+
+ /* works even with NaN */
+ tv_max = get_mode_max(mode);
+ if (a->type.tv == tv_max)
+ return b;
+ if (b->type.tv == tv_max)
+ return a;
+ return node;
+} /* identity_Min */
+
+/**
+ * Calculates the Identity for Max nodes.
+ */
+static node_t *identity_Max(node_t *node) {
+ ir_node *op = node->node;
+ node_t *a = get_irn_node(get_binop_left(op));
+ node_t *b = get_irn_node(get_binop_right(op));
+ ir_mode *mode = get_irn_mode(op);
+ tarval *tv_min;
+
+ if (a->part == b->part) {
+ /* leader of multiple predecessors */
+ return a;
+ }
+
+ /* works even with NaN */
+ tv_min = get_mode_min(mode);
+ if (a->type.tv == tv_min)
+ return b;
+ if (b->type.tv == tv_min)
+ return a;
+ return node;
+} /* identity_Max */
+
+/**
+ * Calculates the Identity for nodes.
+ */
+static node_t *identity(node_t *node) {
+ ir_node *irn = node->node;
+
+ switch (get_irn_opcode(irn)) {
+ case iro_Phi:
+ return identity_Phi(node);
+ case iro_Mul:
+ return identity_Mul(node);
+ case iro_Add:
+ case iro_Or:
+ case iro_Eor:
+ return identity_comm_zero_binop(node);
+ case iro_Shr:
+ case iro_Shl:
+ case iro_Shrs:
+ case iro_Rotl:
+ return identity_shift(node);
+ case iro_And:
+ return identity_And(node);
+ case iro_Sub:
+ return identity_Sub(node);
+ case iro_Confirm:
+ return identity_Confirm(node);
+ case iro_Mux:
+ return identity_Mux(node);
+ case iro_Min:
+ return identity_Min(node);
+ case iro_Max:
+ return identity_Max(node);
+ default:
+ return node;
+ }
+} /* identity */
+
+/**
+ * Node follower is a (new) follower of leader, segregate Leader
+ * out edges.
+ */
+static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
+ ir_node *l = leader->node;
+ int j, i, n = get_irn_n_outs(l);
+
+ DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
+ /* The leader edges must remain sorted, but follower edges can
+ be unsorted. */
+ for (i = leader->n_followers + 1; i <= n; ++i) {
+ if (l->out[i].use == follower) {
+ ir_def_use_edge t = l->out[i];
+
+ for (j = i - 1; j >= leader->n_followers + 1; --j)
+ l->out[j + 1] = l->out[j];
+ ++leader->n_followers;
+ l->out[leader->n_followers] = t;
+ break;
+ }
+ }
+} /* segregate_def_use_chain_1 */
+
+/**
+ * Node follower is a (new) follower segregate its Leader
+ * out edges.
+ *
+ * @param follower the follower IR node
+ */
+static void segregate_def_use_chain(const ir_node *follower) {
+ int i;
+
+ for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
+ node_t *pred = get_irn_node(get_irn_n(follower, i));
+
+ segregate_def_use_chain_1(follower, pred);
+ }
+} /* segregate_def_use_chain */
+
/**
* Propagate constant evaluation.
*
node_t *x;
lattice_elem_t old_type;
node_t *fallen;
- unsigned n_fallen;
+ unsigned n_fallen, old_type_was_T_or_C;
int i;
while (env->cprop != NULL) {
+ void *oldopcode = NULL;
+
/* remove the first partition X from cprop */
- X = env->cprop;
+ X = env->cprop;
X->on_cprop = 0;
- env->cprop = X->cprop_next;
+ env->cprop = X->cprop_next;
+
+ old_type_was_T_or_C = X->type_is_T_or_C;
DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
fallen = NULL;
while (! list_empty(&X->cprop)) {
/* remove the first Node x from X.cprop */
x = list_entry(X->cprop.next, node_t, cprop_list);
+ //assert(x->part == X);
list_del(&x->cprop_list);
x->on_cprop = 0;
+ if (x->is_follower && identity(x) == x) {
+ /* check the opcode first */
+ if (oldopcode == NULL) {
+ oldopcode = lambda_opcode(get_first_node(X), env);
+ }
+ if (oldopcode != lambda_opcode(x, env)) {
+ if (x->on_fallen == 0) {
+ /* different opcode -> x falls out of this partition */
+ x->next = fallen;
+ x->on_fallen = 1;
+ fallen = x;
+ ++n_fallen;
+ DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
+ }
+ }
+
+ /* x will make the follower -> leader transition */
+ follower_to_leader(x);
+ }
+
/* compute a new type for x */
old_type = x->type;
DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
node_t *y = get_irn_node(succ);
/* Add y to y.partition.cprop. */
- add_node_to_cprop(y, env);
+ add_to_cprop(y, env);
}
}
}
- if (n_fallen > 0 && n_fallen != X->n_nodes) {
+ if (n_fallen > 0 && n_fallen != X->n_leader) {
DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
- Y = split(X, fallen, env);
+ Y = split(&X, fallen, env);
+ /*
+ * We have split out fallen node. The type of the result
+ * partition is NOT set yet.
+ */
+ Y->type_is_T_or_C = 0;
} else {
Y = X;
}
- /* remove the nodes from the fallen list */
+ /* remove the flags from the fallen list */
for (x = fallen; x != NULL; x = x->next)
x->on_fallen = 0;
- if (Y->n_nodes > 1)
- split_by(Y, env);
+ if (old_type_was_T_or_C) {
+ node_t *y, *tmp;
+
+ /* check if some nodes will make the leader -> follower transition */
+ list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
+ if (y->type.tv != tarval_top && ! is_con(y->type)) {
+ node_t *eq_node = identity(y);
+
+ if (eq_node != y && eq_node->part == y->part) {
+ DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
+ /* move to Follower */
+ y->is_follower = 1;
+ list_del(&y->node_list);
+ list_add_tail(&y->node_list, &Y->Follower);
+ --Y->n_leader;
+
+ segregate_def_use_chain(y->node);
+ }
+ }
+ }
+ }
+ split_by(Y, env);
}
} /* propagate */
static ir_node *get_leader(node_t *node) {
partition_t *part = node->part;
- if (part->n_nodes > 1) {
- DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
+ if (part->n_leader > 1 || node->is_follower) {
+ if (node->is_follower) {
+ DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
+ }
+ else
+ DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
return get_first_node(part)->node;
}
* @param pred the control flow exit
*/
static int can_exchange(ir_node *pred) {
- if (is_Jmp(pred))
+ if (is_Start(pred))
+ return 0;
+ else if (is_Jmp(pred))
return 1;
else if (get_irn_mode(pred) == mode_T) {
int i, k;
/* if the predecessor block has more than one
- reachable outputs we cannot remove the block */
+ reachable outputs we cannot remove the block */
k = 0;
for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
ir_node *proj = get_irn_out(pred, i);
return 1;
}
return 0;
-}
+} /* can_exchange */
/**
* Block Post-Walker, apply the analysis results on control flow by
* shortening Phi's and Block inputs.
*/
static void apply_cf(ir_node *block, void *ctx) {
- node_t *node = get_irn_node(block);
- int i, j, k, n;
- ir_node **ins, **in_X;
- ir_node *phi, *next;
-
- (void) ctx;
- if (block == get_irg_end_block(current_ir_graph) ||
- block == get_irg_start_block(current_ir_graph)) {
+ environment_t *env = ctx;
+ node_t *node = get_irn_node(block);
+ int i, j, k, n;
+ ir_node **ins, **in_X;
+ ir_node *phi, *next;
+
+ n = get_Block_n_cfgpreds(block);
+
+ if (node->type.tv == tarval_unreachable) {
+ env->modified = 1;
+
+ for (i = n - 1; i >= 0; --i) {
+ ir_node *pred = get_Block_cfgpred(block, i);
+
+ if (! is_Bad(pred)) {
+ node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
+
+ if (pred_bl->flagged == 0) {
+ pred_bl->flagged = 3;
+
+ if (pred_bl->type.tv == tarval_reachable) {
+ /*
+ * We will remove an edge from block to its pred.
+ * This might leave the pred block as an endless loop
+ */
+ if (! is_backedge(block, i))
+ keep_alive(pred_bl->node);
+ }
+ }
+ }
+ }
+
/* the EndBlock is always reachable even if the analysis
finds out the opposite :-) */
+ if (block != get_irg_end_block(current_ir_graph)) {
+ /* mark dead blocks */
+ set_Block_dead(block);
+ DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
+ } else {
+ /* the endblock is unreachable */
+ set_irn_in(block, 0, NULL);
+ }
return;
}
- if (node->type.tv == tarval_unreachable) {
- /* mark dead blocks */
- set_Block_dead(block);
- return;
- }
-
- n = get_Block_n_cfgpreds(block);
if (n == 1) {
/* only one predecessor combine */
ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
- if (can_exchange(pred))
- exchange(block, get_nodes_block(pred));
+ if (can_exchange(pred)) {
+ ir_node *new_block = get_nodes_block(pred);
+ DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
+ DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
+ exchange(block, new_block);
+ node->node = new_block;
+ env->modified = 1;
+ }
return;
}
if (node->type.tv == tarval_reachable) {
in_X[k++] = pred;
+ } else {
+ DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
+ if (! is_Bad(pred)) {
+ node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
+
+ if (pred_bl->flagged == 0) {
+ pred_bl->flagged = 3;
+
+ if (pred_bl->type.tv == tarval_reachable) {
+ /*
+ * We will remove an edge from block to its pred.
+ * This might leave the pred block as an endless loop
+ */
+ if (! is_backedge(block, i))
+ keep_alive(pred_bl->node);
+ }
+ }
+ }
}
}
if (k >= n)
set_irn_node(c, node);
node->node = c;
DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
+ DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
exchange(phi, c);
+ env->modified = 1;
} else {
j = 0;
for (i = 0; i < n; ++i) {
ins[j++] = get_Phi_pred(phi, i);
}
}
- if (j <= 1) {
+ if (j == 1) {
/* this Phi is replaced by a single predecessor */
ir_node *s = ins[0];
+ node_t *phi_node = get_irn_node(phi);
node->node = s;
- DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, s));
+ DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
+ DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
exchange(phi, s);
+ phi_node->node = s;
+ env->modified = 1;
} else {
set_irn_in(phi, j, ins);
+ env->modified = 1;
}
}
}
- if (k <= 1) {
+ if (k == 1) {
/* this Block has only one live predecessor */
ir_node *pred = skip_Proj(in_X[0]);
- if (can_exchange(pred))
- exchange(block, get_nodes_block(pred));
+ if (can_exchange(pred)) {
+ ir_node *new_block = get_nodes_block(pred);
+ DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
+ exchange(block, new_block);
+ node->node = new_block;
+ env->modified = 1;
+ }
} else {
- set_irn_in(block, j, in_X);
+ set_irn_in(block, k, in_X);
+ env->modified = 1;
+ }
+} /* apply_cf */
+
+/**
+ * Exchange a node by its leader.
+ * Beware: in rare cases the mode might be wrong here, for instance
+ * AddP(x, NULL) is a follower of x, but with different mode.
+ * Fix it here.
+ */
+static void exchange_leader(ir_node *irn, ir_node *leader) {
+ ir_mode *mode = get_irn_mode(irn);
+ if (mode != get_irn_mode(leader)) {
+ /* The conv is a no-op, so we are fre to place in
+ * either in the block of the leader OR in irn's block.
+ * Probably placing it into leaders block might reduce
+ * the number of Conv due to CSE. */
+ ir_node *block = get_nodes_block(leader);
+ dbg_info *dbg = get_irn_dbg_info(irn);
+
+ leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
}
+ exchange(irn, leader);
}
/**
* Post-Walker, apply the analysis results;
*/
static void apply_result(ir_node *irn, void *ctx) {
- node_t *node = get_irn_node(irn);
+ environment_t *env = ctx;
+ node_t *node = get_irn_node(irn);
- (void) ctx;
- if (is_Block(irn) || is_End(irn)) {
+ if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
/* blocks already handled, do not touch the End node */
} else {
node_t *block = get_irn_node(get_nodes_block(irn));
if (block->type.tv == tarval_unreachable) {
- if (! is_Bad(irn)) {
- ir_node *bad = get_irg_bad(current_ir_graph);
-
- /* here, bad might already have a node, but this can be safely ignored
- as long as bad has at least ONE valid node */
- set_irn_node(bad, node);
- node->node = bad;
- DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
- exchange(irn, bad);
- }
+ ir_node *bad = get_irg_bad(current_ir_graph);
+
+ /* here, bad might already have a node, but this can be safely ignored
+ as long as bad has at least ONE valid node */
+ set_irn_node(bad, node);
+ node->node = bad;
+ DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
+ exchange(irn, bad);
+ env->modified = 1;
}
- else if (get_irn_mode(irn) == mode_X) {
- if (node->type.tv == tarval_unreachable) {
+ else if (node->type.tv == tarval_unreachable) {
+ /* don't kick away Unknown */
+ if (! is_Unknown(irn)) {
ir_node *bad = get_irg_bad(current_ir_graph);
/* see comment above */
node->node = bad;
DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
exchange(irn, bad);
+ env->modified = 1;
}
- else if (is_Proj(irn)) {
+ }
+ else if (get_irn_mode(irn) == mode_X) {
+ if (is_Proj(irn)) {
/* leave or Jmp */
ir_node *cond = get_Proj_pred(irn);
node_t *sel = get_irn_node(get_Cond_selector(cond));
if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
- /* Cond selector is a constant, make a Jmp */
- ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
+ /* Cond selector is a constant and the Proj is reachable, make a Jmp */
+ ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
set_irn_node(jmp, node);
node->node = jmp;
DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
+ DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
exchange(irn, jmp);
+ env->modified = 1;
}
}
}
if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
tarval *tv = node->type.tv;
- if (! is_Const(irn)) {
+ /*
+ * Beware: never replace mode_T nodes by constants. Currently we must mark
+ * mode_T nodes with constants, but do NOT replace them.
+ */
+ if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
/* can be replaced by a constant */
ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
set_irn_node(c, node);
node->node = c;
DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
- exchange(irn, c);
+ DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
+ exchange_leader(irn, c);
+ env->modified = 1;
}
} else if (is_entity(node->type.sym.entity_p)) {
if (! is_SymConst(irn)) {
- /* can be replaced by a Symconst */
+ /* can be replaced by a SymConst */
ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
set_irn_node(symc, node);
node->node = symc;
DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
- exchange(irn, symc);
+ DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
+ exchange_leader(irn, symc);
+ env->modified = 1;
}
+ } else if (is_Confirm(irn)) {
+ /* Confirms are always follower, but do not kill them here */
} else {
ir_node *leader = get_leader(node);
if (leader != irn) {
- DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
- exchange(irn, leader);
+ int non_strict_phi = 0;
+
+ /*
+ * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
+ * as this might create non-strict programs.
+ */
+ if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
+ int i;
+
+ for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
+ ir_node *pred = get_Phi_pred(irn, i);
+
+ if (is_Unknown(pred)) {
+ non_strict_phi = 1;
+ break;
+ }
+ }
+ }
+ if (! non_strict_phi) {
+ DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
+ if (node->is_follower)
+ DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
+ else
+ DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
+ exchange_leader(irn, leader);
+ env->modified = 1;
+ }
}
}
}
}
} /* apply_result */
+/**
+ * Fix the keep-alives by deleting unreachable ones.
+ */
+static void apply_end(ir_node *end, environment_t *env) {
+ int i, j, n = get_End_n_keepalives(end);
+ ir_node **in;
+
+ if (n > 0)
+ NEW_ARR_A(ir_node *, in, n);
+
+ /* fix the keep alive */
+ for (i = j = 0; i < n; i++) {
+ ir_node *ka = get_End_keepalive(end, i);
+ node_t *node = get_irn_node(ka);
+
+ if (! is_Block(ka))
+ node = get_irn_node(get_nodes_block(ka));
+
+ if (node->type.tv != tarval_unreachable && !is_Bad(ka))
+ in[j++] = ka;
+ }
+ if (j != n) {
+ set_End_keepalives(end, j, in);
+ env->modified = 1;
+ }
+} /* apply_end */
+
#define SET(code) op_##code->ops.generic = (op_func)compute_##code
/**
SET(Phi);
SET(Add);
SET(Sub);
+ SET(Eor);
SET(SymConst);
+ SET(Cmp);
SET(Proj);
SET(Confirm);
+ SET(Return);
SET(End);
-} /* set_compute_functions */
-static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
- ir_node *irn = local != NULL ? local : n;
- node_t *node = get_irn_node(irn);
+ if (op_Max != NULL)
+ SET(Max);
+ if (op_Min != NULL)
+ SET(Min);
- ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
- return 1;
-}
+} /* set_compute_functions */
void combo(ir_graph *irg) {
environment_t env;
- ir_node *initial_X;
+ ir_node *initial_bl;
node_t *start;
ir_graph *rem = current_ir_graph;
/* register a debug mask */
FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
- firm_dbg_set_mask(dbg, SET_LEVEL_1);
- //DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
+ DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
obstack_init(&env.obst);
env.worklist = NULL;
env.type2id_map = pmap_create();
env.end_idx = get_opt_global_cse() ? 0 : -1;
env.lambda_input = 0;
+ env.commutative = 1;
+ env.modified = 0;
assure_irg_outs(irg);
+ assure_cf_loop(irg);
/* we have our own value_of function */
set_value_of_func(get_node_tarval);
set_compute_functions();
DEBUG_ONLY(part_nr = 0);
+ ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
+
/* create the initial partition and place it on the work list */
env.initial = new_partition(&env);
add_to_worklist(env.initial, &env);
irg_walk_graph(irg, init_block_phis, create_initial_partitions, &env);
+ /* set the hook: from now, every node has a partition and a type */
+ DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
+
+#ifdef WITH_UNKNOWN
+ tarval_UNKNOWN = tarval_top;
+#else
+ tarval_UNKNOWN = tarval_bad;
+#endif
+
+ /* all nodes on the initial partition have type Top */
+ env.initial->type_is_T_or_C = 1;
+
/* Place the START Node's partition on cprop.
Place the START Node on its local worklist. */
- initial_X = get_irg_initial_exec(irg);
- start = get_irn_node(initial_X);
- add_node_to_cprop(start, &env);
+ initial_bl = get_irg_start_block(irg);
+ start = get_irn_node(initial_bl);
+ add_to_cprop(start, &env);
do {
propagate(&env);
} while (env.cprop != NULL || env.worklist != NULL);
dump_all_partitions(&env);
+ check_all_partitions(&env);
- set_dump_node_vcgattr_hook(dump_partition_hook);
+#if 0
dump_ir_block_graph(irg, "-partition");
- set_dump_node_vcgattr_hook(NULL);
-
+#endif
/* apply the result */
irg_block_walk_graph(irg, NULL, apply_cf, &env);
irg_walk_graph(irg, NULL, apply_result, &env);
+ apply_end(get_irg_end(irg), &env);
+
+ if (env.modified) {
+ /* control flow might changed */
+ set_irg_outs_inconsistent(irg);
+ set_irg_extblk_inconsistent(irg);
+ set_irg_doms_inconsistent(irg);
+ set_irg_loopinfo_inconsistent(irg);
+ }
+
+ ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+
+ /* remove the partition hook */
+ DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
pmap_destroy(env.type2id_map);
del_set(env.opcode2id_map);
projects / libfirm / blobdiff