* - 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
#include <assert.h>
#include "iroptimize.h"
-#include "archop.h"
#include "irflag.h"
#include "ircons.h"
#include "list.h"
#include "debug.h"
#include "array_t.h"
#include "error.h"
+#include "irnodeset.h"
#include "tv_t.h"
#include "irdump.h"
/* define this to check that all type translations are monotone */
-#undef VERIFY_MONOTONE
+#define VERIFY_MONOTONE
/* define this to check the consistency of partitions */
#define CHECK_PARTITIONS
-/* allow optimization of non-strict programs */
-#undef WITH_UNKNOWN
-
typedef struct node_t node_t;
typedef struct partition_t partition_t;
typedef struct opcode_key_t opcode_key_t;
union {
long proj; /**< For Proj nodes, its proj number */
ir_entity *ent; /**< For Sel Nodes, its entity */
+ int intVal; /**< For Conv/Div Nodes: strict/remainderless */
} u;
};
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. */
};
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. */
+ list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) 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 *initial; /**< The initial partition. */
set *opcode2id_map; /**< The opcodeMode->id map. */
pmap *type2id_map; /**< The type->id map. */
+ ir_node **kept_memory; /**< Array of memory nodes that must be kept. */
int end_idx; /**< -1 for local and 0 for global congruences. */
int lambda_input; /**< Captured argument for lambda_partition(). */
- char nonstd_cond; /**< Set, if a Condb note has a non-Cmp predecessor. */
- char modified; /**< Set, if the graph was modified. */
- char commutative; /**< Set, if commutation nodes should be handled specially. */
+ unsigned modified:1; /**< Set, if the graph was modified. */
+ unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
+ /* options driving the optimization */
+ unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
+ unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */
#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(follower) ((node_t *)get_irn_link(follower))
-#define set_irn_node(follower, node) set_irn_link(follower, node)
+#define get_irn_node(irn) ((node_t *)get_irn_link(irn))
+#define set_irn_node(irn, node) set_irn_link(irn, node)
/* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
#undef tarval_unreachable
/** Next partition number. */
DEBUG_ONLY(static unsigned part_nr = 0);
-/** The tarval returned by Unknown nodes. */
+/** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */
static tarval *tarval_UNKNOWN;
/* forward */
case iro_Sel:
key.u.ent = get_Sel_entity(irn);
break;
+ case iro_Conv:
+ key.u.intVal = get_Conv_strict(irn);
+ break;
+ case iro_Div:
+ key.u.intVal = is_Div_remainderless(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));
+ 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:
case iro_Sel:
assert(key.u.ent == get_Sel_entity(irn));
break;
+ case iro_Conv:
+ assert(key.u.intVal == get_Conv_strict(irn));
+ break;
+ case iro_Div:
+ assert(key.u.intVal == is_Div_remainderless(irn));
+ break;
default:
break;
}
} /* check_opcode */
static void check_all_partitions(environment_t *env) {
+#ifdef DEBUG_libfirm
partition_t *P;
node_t *node;
-#ifdef DEBUG_libfirm
for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
check_partition(P);
if (! P->type_is_T_or_C)
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)
/**
* Verify that a type transition is monotone
*/
-static void verify_type(const lattice_elem_t old_type, const lattice_elem_t new_type) {
- if (old_type.tv == new_type.tv) {
+static void verify_type(const lattice_elem_t old_type, node_t *node) {
+ if (old_type.tv == node->type.tv) {
/* no change */
return;
}
/* from Top down-to is always allowed */
return;
}
- if (old_type.tv == tarval_reachable) {
- panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
- }
- if (new_type.tv == tarval_bottom || new_type.tv == tarval_reachable) {
+ if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
/* bottom reached */
return;
}
- panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
+ panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
} /* verify_type */
+
#else
-#define verify_type(old_type, new_type)
+#define verify_type(old_type, node)
#endif
/**
(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;
+ o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent ||
+ o1->u.intVal != o2->u.intVal;
} /* cmp_opcode */
/**
INIT_LIST_HEAD(&part->Leader);
INIT_LIST_HEAD(&part->Follower);
INIT_LIST_HEAD(&part->cprop);
+ INIT_LIST_HEAD(&part->cprop_X);
part->wl_next = NULL;
part->touched_next = NULL;
part->cprop_next = NULL;
node->on_cprop = 0;
node->on_fallen = 0;
node->is_follower = 0;
- node->by_all_const = 0;
node->flagged = 0;
set_irn_node(irn, node);
} /* create_partition_node */
/**
- * Pre-Walker, init all Block-Phi lists.
- */
-static void init_block_phis(ir_node *irn, void *env) {
- (void) env;
-
- if (is_Block(irn)) {
- set_Block_phis(irn, NULL);
- }
-} /* init_block_phis */
-
-/**
- * Post-Walker, initialize all Nodes' type to U or top and place
+ * Pre-Walker, initialize all Nodes' type to U or top and place
* all nodes into the TOP partition.
*/
static void create_initial_partitions(ir_node *irn, void *ctx) {
if (node->max_user_input > part->max_user_inputs)
part->max_user_inputs = node->max_user_input;
+ if (is_Block(irn)) {
+ set_Block_phis(irn, NULL);
+ }
+} /* create_initial_partitions */
+
+/**
+ * Post-Walker, collect all Block-Phi lists, set Cond.
+ */
+static void init_block_phis(ir_node *irn, void *ctx) {
+ (void) ctx;
+
if (is_Phi(irn)) {
add_Block_phi(get_nodes_block(irn), irn);
- } else if (is_Cond(irn)) {
- /* check if all Cond's have a Cmp predecessor. */
- if (get_irn_mode(irn) == mode_b && !is_Cmp(skip_Proj(get_Cond_selector(irn))))
- env->nonstd_cond = 1;
}
-} /* create_initial_partitions */
+} /* init_block_phis */
/**
* Add a node to the entry.partition.touched set and
* @param env the environment
*/
static void add_to_cprop(node_t *y, environment_t *env) {
+ ir_node *irn;
+
/* Add y to y.partition.cprop. */
if (y->on_cprop == 0) {
partition_t *Y = y->part;
+ ir_node *irn = y->node;
- list_add_tail(&y->cprop_list, &Y->cprop);
+ /* place Conds and all its Projs on the cprop_X list */
+ if (is_Cond(skip_Proj(irn)))
+ list_add_tail(&y->cprop_list, &Y->cprop_X);
+ else
+ 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));
Y->on_cprop = 1;
}
}
- if (get_irn_mode(y->node) == mode_T) {
+ irn = y->node;
+ if (get_irn_mode(irn) == 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));
+ for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
+ node_t *proj = get_irn_node(get_irn_out(irn, i));
add_to_cprop(proj, env);
}
- } else if (is_Block(y->node)) {
+ } else if (is_Block(irn)) {
/* 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)) {
+ for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
node_t *p = get_irn_node(phi);
add_to_cprop(p, env);
}
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;
/**
* Collect commutative nodes to the touched list.
*
+ * @param X the partition of the 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) {
+static void collect_commutative_touched(partition_t *X, list_head *list, environment_t *env) {
+ int first = 1;
+ int both_input = 0;
node_t *x, *y;
list_for_each_entry(node_t, x, list, node_list) {
/* 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);
+ int other_idx = edge->pos ^ 1;
+ node_t *other = get_irn_node(get_irn_n(succ, other_idx));
+ int equal = X == other->part;
+
+ /*
+ * Note: op(a, a) is NOT congruent to op(a, b).
+ * So, either all touch nodes must have both inputs congruent,
+ * or not. We decide this by the first occurred node.
+ */
+ if (first) {
+ first = 0;
+ both_input = equal;
+ }
+ if (both_input == equal)
+ add_to_touched(y, env);
}
}
}
/* 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);
+ collect_commutative_touched(X, &X->Leader, env);
+ collect_commutative_touched(X, &X->Follower, env);
for (Z = env->touched; Z != NULL; Z = N) {
node_t *e;
case iro_Sel:
key.u.ent = get_Sel_entity(irn);
break;
+ case iro_Conv:
+ key.u.intVal = get_Conv_strict(irn);
+ break;
+ case iro_Div:
+ key.u.intVal = is_Div_remainderless(irn);
+ break;
default:
break;
}
static void default_compute(node_t *node) {
int i;
ir_node *irn = node->node;
- node_t *block = get_irn_node(get_nodes_block(irn));
-
- if (block->type.tv == tarval_unreachable) {
- node->type.tv = tarval_top;
- return;
- }
/* if any of the data inputs have type top, the result is type top */
for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
int i;
ir_node *block = node->node;
- if (block == get_irg_start_block(current_ir_graph)) {
- /* start block is always reachable */
+ if (block == get_irg_start_block(current_ir_graph) || has_Block_label(block)) {
+ /* start block and labelled blocks are always reachable */
node->type.tv = tarval_reachable;
return;
}
node->type.tv = tarval_reachable;
} /* compute_End */
+/**
+ * (Re-)compute the type for a Call.
+ *
+ * @param node the node
+ */
+static void compute_Call(node_t *node) {
+ /*
+ * A Call computes always bottom, even if it has Unknown
+ * predecessors.
+ */
+ node->type.tv = tarval_bottom;
+} /* compute_Call */
+
/**
* (Re-)compute the type for a SymConst node.
*
} 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)))) {
/*
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)
+ if (node->type.tv != tv)
tv = tarval_bottom;
node->type.tv = tv;
} else {
} 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);
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)
+ if (node->type.tv != tv)
tv = tarval_bottom;
node->type.tv = tv;
} else {
lattice_elem_t b = r->type;
if (a.tv == tarval_top || b.tv == tarval_top) {
- /*
- * Top is congruent to any other value, we can
- * calculate the compare result.
- */
+ 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)) {
/* both nodes are constants, we can probably do something */
node->type.tv = tarval_b_true;
- } else if (r->part == l->part) {
- /* both nodes congruent, we can probably do something */
- node->type.tv = tarval_b_true;
} else {
node->type.tv = tarval_bottom;
}
tarval *tv;
if (a.tv == tarval_top || b.tv == tarval_top) {
- /*
- * Top is congruent to any other value, we can
- * calculate the compare result.
- */
- goto congruent;
+ 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)) {
/*
* BEWARE: a == a is NOT always True for floating Point values, as
* NaN != NaN is defined, so we must check this here.
*/
-congruent:
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)
+ if (node->type.tv != tv)
tv = tarval_bottom;
node->type.tv = tv;
} else {
ir_node *sel = get_Cond_selector(cond);
node_t *selector = get_irn_node(sel);
+ /*
+ * Note: it is crucial for the monotony that the Proj(Cond)
+ * are evaluates after all predecessors of the Cond selector are
+ * processed.
+ * Example
+ *
+ * if (x != 0)
+ *
+ * Due to the fact that 0 is a const, the Cmp gets immediately
+ * on the cprop list. It will be evaluated before x is evaluated,
+ * might leaving x as Top. When later x is evaluated, the Cmp
+ * might change its value.
+ * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
+ * gets R, and later changed to F if Cmp is evaluated to True!
+ *
+ * We prevent this by putting Conds in an extra cprop_X queue, which
+ * gets evaluated after the cprop queue is empty.
+ *
+ * Note that this even happens with Click's original algorithm, if
+ * Cmp(x, 0) is evaluated to True first and later changed to False
+ * if x was Top first and later changed to a Const ...
+ * It is unclear how Click solved that problem ...
+ *
+ * However, in rare cases even this does not help, if a Top reaches
+ * a compare through a Phi, than Proj(Cond) is evaluated changing
+ * the type of the Phi to something other.
+ * So, we take the last resort and bind the type to R once
+ * it is calculated.
+ *
+ * (This might be even the way Click works around the whole problem).
+ *
+ * Finally, we may miss some optimization possibilities due to this:
+ *
+ * x = phi(Top, y)
+ * if (x == 0)
+ *
+ * If Top reaches the if first, than we decide for != here.
+ * If y later is evaluated to 0, we cannot revert this decision
+ * and must live with both outputs enabled. If this happens,
+ * we get an unresolved if (true) in the code ...
+ *
+ * In Click's version where this decision is done at the Cmp,
+ * the Cmp is NOT optimized away than (if y evaluated to 1
+ * for instance) and we get a if (1 == 0) here ...
+ *
+ * Both solutions are suboptimal.
+ * At least, we could easily detect this problem and run
+ * cf_opt() (or even combo) again :-(
+ */
+ if (node->type.tv == tarval_reachable)
+ return;
+
if (get_irn_mode(sel) == mode_b) {
/* an IF */
if (pnc == pn_Cond_true) {
node->type.tv = tarval_reachable;
} else {
assert(selector->type.tv == tarval_top);
- node->type.tv = tarval_unreachable;
+ if (tarval_UNKNOWN == tarval_top) {
+ /* any condition based on Top is "!=" */
+ node->type.tv = tarval_unreachable;
+ } else {
+ node->type.tv = tarval_unreachable;
+ }
}
} else {
assert(pnc == pn_Cond_false);
node->type.tv = tarval_reachable;
} else {
assert(selector->type.tv == tarval_top);
- node->type.tv = tarval_unreachable;
+ if (tarval_UNKNOWN == tarval_top) {
+ /* any condition based on Top is "!=" */
+ node->type.tv = tarval_reachable;
+ } else {
+ node->type.tv = tarval_unreachable;
+ }
}
}
} 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 (tarval_UNKNOWN == tarval_top &&
+ 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)) {
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;
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.
*
ir_node *irn = node->node;
compute_func func;
+#ifndef VERIFY_MONOTONE
+ /*
+ * Once a node reaches bottom, the type cannot fall further
+ * in the lattice and we can stop computation.
+ * Do not take this exit if the monotony verifier is
+ * enabled to catch errors.
+ */
+ if (node->type.tv == tarval_bottom)
+ return;
+#endif
+
if (is_no_Block(irn)) {
/* for pinned nodes, check its control input */
- if (get_irn_pinned(irn) == op_pin_state_pinned) {
+ 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) {
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.
*/
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;
}
DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
fallen = NULL;
n_fallen = 0;
- while (! list_empty(&X->cprop)) {
+ for (;;) {
+ int cprop_empty = list_empty(&X->cprop);
+ int cprop_X_empty = list_empty(&X->cprop_X);
+
+ if (cprop_empty && cprop_X_empty) {
+ /* both cprop lists are empty */
+ break;
+ }
+
/* remove the first Node x from X.cprop */
- x = list_entry(X->cprop.next, node_t, cprop_list);
+ if (cprop_empty) {
+ /* Get a node from the cprop_X list only if
+ * all data nodes are processed.
+ * This ensures, that all inputs of the Cond
+ * predecessor are processed if its type is still Top.
+ */
+ x = list_entry(X->cprop_X.next, node_t, cprop_list);
+ } else {
+ x = list_entry(X->cprop.next, node_t, cprop_list);
+ }
+
//assert(x->part == X);
list_del(&x->cprop_list);
x->on_cprop = 0;
DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
compute(x);
if (x->type.tv != old_type.tv) {
- verify_type(old_type, x->type);
DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
+ verify_type(old_type, x);
if (x->on_fallen == 0) {
/* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
return node->node;
} /* get_leader */
+/**
+ * Returns non-zero if a mode_T node has only one reachable output.
+ */
+static int only_one_reachable_proj(ir_node *n) {
+ int i, k = 0;
+
+ for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
+ ir_node *proj = get_irn_out(n, i);
+ node_t *node;
+
+ /* skip non-control flow Proj's */
+ if (get_irn_mode(proj) != mode_X)
+ continue;
+
+ node = get_irn_node(proj);
+ if (node->type.tv == tarval_reachable) {
+ if (++k > 1)
+ return 0;
+ }
+ }
+ return 1;
+} /* only_one_reachable_proj */
+
/**
* Return non-zero if the control flow predecessor node pred
* is the only reachable control flow exit of its block.
*
- * @param pred the control flow exit
+ * @param pred the control flow exit
+ * @param block the destination block
*/
-static int can_exchange(ir_node *pred) {
- if (is_Start(pred))
+static int can_exchange(ir_node *pred, ir_node *block) {
+ if (is_Start(pred) || has_Block_label(block))
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 */
- k = 0;
- for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
- ir_node *proj = get_irn_out(pred, i);
- node_t *node;
-
- /* skip non-control flow Proj's */
- if (get_irn_mode(proj) != mode_X)
- continue;
-
- node = get_irn_node(proj);
- if (node->type.tv == tarval_reachable) {
- if (++k > 1)
- return 0;
- }
- }
- return 1;
+ return only_one_reachable_proj(pred);
}
return 0;
} /* can_exchange */
/* only one predecessor combine */
ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
- if (can_exchange(pred)) {
+ if (can_exchange(pred, block)) {
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);
if (k >= n)
return;
+ /* fix Phi's */
NEW_ARR_A(ir_node *, ins, n);
for (phi = get_Block_phis(block); phi != NULL; phi = next) {
node_t *node = get_irn_node(phi);
if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
/* this Phi is replaced by a constant */
tarval *tv = node->type.tv;
- ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
+ ir_node *c = new_Const(tv);
set_irn_node(c, node);
node->node = c;
}
}
+ /* fix block */
if (k == 1) {
/* this Block has only one live predecessor */
ir_node *pred = skip_Proj(in_X[0]);
- if (can_exchange(pred)) {
+ if (can_exchange(pred, block)) {
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;
+ return;
}
- } else {
- set_irn_in(block, k, in_X);
- env->modified = 1;
}
+ set_irn_in(block, k, in_X);
+ env->modified = 1;
} /* apply_cf */
/**
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
+ /* The conv is a no-op, so we are free to place it
* 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. */
leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
}
exchange(irn, leader);
-}
+} /* exchange_leader */
+
+/**
+ * Check, if all users of a mode_M node are dead. Use
+ * the Def-Use edges for this purpose, as they still
+ * reflect the situation.
+ */
+static int all_users_are_dead(const ir_node *irn) {
+ int i, n = get_irn_n_outs(irn);
+
+ for (i = 1; i <= n; ++i) {
+ const ir_node *succ = irn->out[i].use;
+ const node_t *block = get_irn_node(get_nodes_block(succ));
+ const node_t *node;
+
+ if (block->type.tv == tarval_unreachable) {
+ /* block is unreachable */
+ continue;
+ }
+ node = get_irn_node(succ);
+ if (node->type.tv != tarval_top) {
+ /* found a reachable user */
+ return 0;
+ }
+ }
+ /* all users are unreachable */
+ return 1;
+} /* all_user_are_dead */
+
+/**
+ * Walker: Find reachable mode_M nodes that have only
+ * unreachable users. These nodes must be kept later.
+ */
+static void find_kept_memory(ir_node *irn, void *ctx) {
+ environment_t *env = ctx;
+ node_t *node, *block;
+
+ if (get_irn_mode(irn) != mode_M)
+ return;
+
+ block = get_irn_node(get_nodes_block(irn));
+ if (block->type.tv == tarval_unreachable)
+ return;
+
+ node = get_irn_node(irn);
+ if (node->type.tv == tarval_top)
+ return;
+
+ /* ok, we found a live memory node. */
+ if (all_users_are_dead(irn)) {
+ DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
+ ARR_APP1(ir_node *, env->kept_memory, irn);
+ }
+} /* find_kept_memory */
/**
* Post-Walker, apply the analysis results;
DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
exchange(irn, bad);
env->modified = 1;
- }
- 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);
+ } else if (node->type.tv == tarval_top) {
+ ir_mode *mode = get_irn_mode(irn);
+
+ if (mode == mode_M) {
+ /* never kill a mode_M node */
+ if (is_Proj(irn)) {
+ ir_node *pred = get_Proj_pred(irn);
+ node_t *pnode = get_irn_node(pred);
+
+ if (pnode->type.tv == tarval_top) {
+ /* skip the predecessor */
+ ir_node *mem = get_memop_mem(pred);
+ node->node = mem;
+ DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
+ exchange(irn, mem);
+ env->modified = 1;
+ }
+ }
+ /* leave other nodes, especially PhiM */
+ } else if (mode == mode_T) {
+ /* Do not kill mode_T nodes, kill their Projs */
+ } else if (! is_Unknown(irn)) {
+ /* don't kick away Unknown's, they might be still needed */
+ ir_node *unk = new_r_Unknown(current_ir_graph, mode);
+
+ /* control flow should already be handled at apply_cf() */
+ assert(mode != mode_X);
/* see comment above */
- set_irn_node(bad, node);
- node->node = bad;
- DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
- exchange(irn, bad);
+ set_irn_node(unk, node);
+ node->node = unk;
+ DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
+ exchange(irn, unk);
env->modified = 1;
}
}
ir_node *cond = get_Proj_pred(irn);
if (is_Cond(cond)) {
- 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 and the Proj is reachable, make a Jmp */
- ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
+ if (only_one_reachable_proj(cond)) {
+ 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;
+ } else {
+ node_t *sel = get_irn_node(get_Cond_selector(cond));
+ tarval *tv = sel->type.tv;
+
+ if (is_tarval(tv) && tarval_is_constant(tv)) {
+ /* The selector is a constant, but more
+ * than one output is active: An unoptimized
+ * case found. */
+ env->unopt_cf = 1;
+ }
}
}
}
*/
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);
+ ir_node *c = new_Const(tv);
set_irn_node(c, node);
node->node = c;
DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
SET(Confirm);
SET(Return);
SET(End);
+ SET(Call);
+} /* set_compute_functions */
- if (op_Max != NULL)
- SET(Max);
- if (op_Min != NULL)
- SET(Min);
+/**
+ * Add memory keeps.
+ */
+static void add_memory_keeps(ir_node **kept_memory, int len) {
+ ir_node *end = get_irg_end(current_ir_graph);
+ int i;
+ ir_nodeset_t set;
-} /* set_compute_functions */
+ ir_nodeset_init(&set);
-static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
-#ifdef DEBUG_libfirm
- ir_node *irn = local != NULL ? local : n;
- node_t *node = get_irn_node(irn);
+ /* check, if those nodes are already kept */
+ for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
+ ir_nodeset_insert(&set, get_End_keepalive(end, i));
- ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
- return 1;
-#endif
-}
+ for (i = len - 1; i >= 0; --i) {
+ ir_node *ka = kept_memory[i];
+
+ if (! ir_nodeset_contains(&set, ka)) {
+ add_End_keepalive(end, ka);
+ }
+ }
+ ir_nodeset_destroy(&set);
+} /* add_memory_keeps */
void combo(ir_graph *irg) {
environment_t env;
ir_node *initial_bl;
node_t *start;
ir_graph *rem = current_ir_graph;
+ int len;
current_ir_graph = irg;
#endif
env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
env.type2id_map = pmap_create();
+ env.kept_memory = NEW_ARR_F(ir_node *, 0);
env.end_idx = get_opt_global_cse() ? 0 : -1;
env.lambda_input = 0;
- env.nonstd_cond = 0;
- env.commutative = 1;
env.modified = 0;
+ env.unopt_cf = 0;
+ /* options driving the optimization */
+ env.commutative = 1;
+ env.opt_unknown = 1;
assure_irg_outs(irg);
assure_cf_loop(irg);
set_compute_functions();
DEBUG_ONLY(part_nr = 0);
- ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
+ ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
+
+ if (env.opt_unknown)
+ tarval_UNKNOWN = tarval_top;
+ else
+ tarval_UNKNOWN = tarval_bad;
/* 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);
+ irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
-#ifdef WITH_UNKNOWN
- tarval_UNKNOWN = env.nonstd_cond ? tarval_bad : tarval_top;
-#else
- tarval_UNKNOWN = tarval_bad;
-#endif
+ /* set the hook: from now, every node has a partition and a type */
+ DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
/* all nodes on the initial partition have type Top */
env.initial->type_is_T_or_C = 1;
check_all_partitions(&env);
#if 0
- set_dump_node_vcgattr_hook(dump_partition_hook);
dump_ir_block_graph(irg, "-partition");
- set_dump_node_vcgattr_hook(NULL);
-#else
- (void)dump_partition_hook;
#endif
/* apply the result */
+
+ /* check, which nodes must be kept */
+ irg_walk_graph(irg, NULL, find_kept_memory, &env);
+
+ /* kill unreachable control flow */
irg_block_walk_graph(irg, NULL, apply_cf, &env);
- irg_walk_graph(irg, NULL, apply_result, &env);
+ /* Kill keep-alives of dead blocks: this speeds up apply_result()
+ * and fixes assertion because dead cf to dead blocks is NOT removed by
+ * apply_cf(). */
apply_end(get_irg_end(irg), &env);
+ irg_walk_graph(irg, NULL, apply_result, &env);
+
+ len = ARR_LEN(env.kept_memory);
+ if (len > 0)
+ add_memory_keeps(env.kept_memory, len);
+
+ if (env.unopt_cf) {
+ DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
+ }
if (env.modified) {
/* control flow might changed */
set_irg_loopinfo_inconsistent(irg);
}
- ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+ ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
+
+ /* remove the partition hook */
+ DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
+ DEL_ARR_F(env.kept_memory);
pmap_destroy(env.type2id_map);
del_set(env.opcode2id_map);
obstack_free(&env.obst, NULL);