#include "irloop_t.h"
#include "irbackedge_t.h"
#include "cgana.h"
+#include "trouts.h"
#include "irflag_t.h"
#include "irhooks.h"
+#include "iredges_t.h"
/* Defined in iropt.c */
pset *new_identities (void);
}
#if 0 /* Old version. Avoids Ids.
- This is not necessary: we do a postwalk, and get_irn_n
+ This is not necessary: we do a post walk, and get_irn_n
removes ids anyways. So it's much cheaper to call the
optimization less often and use the exchange() algorithm. */
static void
set_irg_loopinfo_inconsistent(current_ir_graph);
- /* Clean the value_table in irg for the cse. */
+ /* Clean the value_table in irg for the CSE. */
del_identities(current_ir_graph->value_table);
current_ir_graph->value_table = new_identities();
do_local_optimize(n);
current_ir_graph = rem;
-
}
void
}
}
-/**
- * Copies the node to the new obstack. The Ins of the new node point to
- * the predecessors on the old obstack. For block/phi nodes not all
- * predecessors might be copied. n->link points to the new node.
+/*
+ * Copies a node to the current_ir_graph. The Ins of the new node point to
+ * the predecessors on the graph of the old node. For block/phi nodes not all
+ * predecessors might be copied.
* For Phi and Block nodes the function allocates in-arrays with an arity
* only for useful predecessors. The arity is determined by counting
* the non-bad predecessors of the block.
- *
- * @param n The node to be copied
- * @param env if non-NULL, the node number attribute will be copied to the new node
- *
- * Note: Also used for loop unrolling.
*/
-static void
-firm_copy_node (ir_node *n, void *env) {
+ir_node *copy_irn(ir_node *n, int copy_node_nr) {
ir_node *nn, *block;
int new_arity;
- opcode op = get_irn_opcode(n);
- int copy_node_nr = env != NULL;
+ ir_op *op = get_irn_op(n);
/* The end node looses it's flexible in array. This doesn't matter,
as dead node elimination builds End by hand, inlineing doesn't use
the End node. */
- /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
+ /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
- if (op == iro_Bad) {
+ if (op == op_Bad) {
/* node copied already */
- return;
- } else if (op == iro_Block) {
+ return NULL;
+ } else if (op == op_Block) {
block = NULL;
new_arity = compute_new_arity(n);
n->attr.block.graph_arr = NULL;
} else {
block = get_nodes_block(n);
- if (get_irn_opcode(n) == iro_Phi) {
+ if (op == op_Phi) {
new_arity = compute_new_arity(block);
} else {
new_arity = get_irn_arity(n);
nn = new_ir_node(get_irn_dbg_info(n),
current_ir_graph,
block,
- get_irn_op(n),
+ op,
get_irn_mode(n),
new_arity,
- get_irn_in(n));
+ get_irn_in(n) + 1);
/* Copy the attributes. These might point to additional data. If this
was allocated on the old obstack the pointers now are dangling. This
frees e.g. the memory of the graph_arr allocated in new_immBlock. */
copy_node_attr(n, nn);
new_backedge_info(nn);
- set_new_node(n, nn);
#if DEBUG_libfirm
if (copy_node_nr) {
}
#endif
- /* printf("\n old node: "); DDMSG2(n);
- printf(" new node: "); DDMSG2(nn); */
+ return nn;
}
+/**
+ * Copies the node to the new obstack. The Ins of the new node point to
+ * the predecessors on the old obstack. For block/phi nodes not all
+ * predecessors might be copied. n->link points to the new node.
+ * For Phi and Block nodes the function allocates in-arrays with an arity
+ * only for useful predecessors. The arity is determined by counting
+ * the non-bad predecessors of the block.
+ *
+ * @param n The node to be copied
+ * @param env if non-NULL, the node number attribute will be copied to the new node
+ *
+ * Note: Also used for loop unrolling.
+ */
+static void firm_copy_node (ir_node *n, void *env) {
+ ir_node *nn = copy_irn(n, env != NULL);
+
+ if (nn)
+ set_new_node(n, nn);
+}
+
+
/**
* Copies new predecessors of old node to new node remembered in link.
* Spare the Bad predecessors of Phi and Block nodes.
*/
-static void
+void
copy_preds (ir_node *n, void *env) {
ir_node *nn, *block;
int i, j, irn_arity;
printf(" new node: "); DDMSG2(nn);
printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
- if (get_irn_opcode(n) == iro_Block) {
+ if (is_Block(n)) {
/* Don't copy Bad nodes. */
j = 0;
irn_arity = get_irn_arity(n);
for (i = 0; i < irn_arity; i++)
- if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
+ if (! is_Bad(get_irn_n(n, i))) {
set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
/*if (is_backedge(n, i)) set_backedge(nn, j);*/
j++;
exchange(nn, old);
}
}
- } else if (get_irn_opcode(n) == iro_Phi) {
+ } else if (get_irn_op(n) == op_Phi) {
/* Don't copy node if corresponding predecessor in block is Bad.
The Block itself should not be Bad. */
block = get_nodes_block(n);
j = 0;
irn_arity = get_irn_arity(n);
for (i = 0; i < irn_arity; i++)
- if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
+ if (! is_Bad(get_irn_n(block, i))) {
set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
/*if (is_backedge(n, i)) set_backedge(nn, j);*/
j++;
set_Block_block_visited(get_nodes_block(n), 0);
/* Compacting the Phi's ins might generate Phis with only one
predecessor. */
- if (get_irn_arity(n) == 1)
- exchange(n, get_irn_n(n, 0));
+ if (get_irn_arity(nn) == 1)
+ exchange(nn, get_irn_n(nn, 0));
} else {
irn_arity = get_irn_arity(n);
for (i = -1; i < irn_arity; i++)
set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
}
- /* Now the new node is complete. We can add it to the hash table for cse.
+ /* Now the new node is complete. We can add it to the hash table for CSE.
@@@ inlinening aborts if we identify End. Why? */
- if(get_irn_op(nn) != op_End)
+ if (get_irn_op(nn) != op_End)
add_identities (current_ir_graph->value_table, nn);
}
* from block nodes and the corresponding inputs from Phi nodes.
* Merges single exit blocks with single entry blocks and removes
* 1-input Phis.
- * Adds all new nodes to a new hash table for cse. Does not
- * perform cse, so the hash table might contain common subexpressions.
+ * Adds all new nodes to a new hash table for CSE. Does not
+ * perform CSE, so the hash table might contain common subexpressions.
*/
void
dead_node_elimination(ir_graph *irg) {
struct obstack *graveyard_obst = NULL;
struct obstack *rebirth_obst = NULL;
+ edges_init_graph(irg);
+
/* inform statistics that we started a dead-node elimination run */
hook_dead_node_elim_start(irg);
/* Handle graph state */
assert(get_irg_phase_state(current_ir_graph) != phase_building);
free_callee_info(current_ir_graph);
- free_outs(current_ir_graph);
+ free_irg_outs(current_ir_graph);
+ free_trouts();
/* @@@ so far we loose loops when copying */
free_loop_information(current_ir_graph);
}
/**
- * Relink bad predeseccors of a block and store the old in array to the
+ * Relink bad predecessors of a block and store the old in array to the
* link field. This function is called by relink_bad_predecessors().
* The array of link field starts with the block operand at position 0.
* If block has bad predecessors, create a new in array without bad preds.
int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
/* if link field of block is NULL, look for bad predecessors otherwise
- this is allready done */
+ this is already done */
if (get_irn_op(n) == op_Block &&
get_irn_link(n) == NULL) {
keep the old one to update Phis. */
new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
- /* set new predeseccors in array */
+ /* set new predecessors in array */
new_in[0] = NULL;
new_irn_n = 1;
for (i = 0; i < old_irn_arity; i++) {
} /* Block is not relinked */
}
-/*
- * Relinks Bad predecesors from Bocks and Phis called by walker
+/**
+ * Relinks Bad predecessors from Blocks and Phis called by walker
* remove_bad_predecesors(). If n is a Block, call
- * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
+ * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
* function of Phi's Block. If this block has bad predecessors, relink preds
- * of the Phinode.
+ * of the Phi-node.
*/
static void relink_bad_predecessors(ir_node *n, void *env) {
ir_node *block, **old_in;
int i, old_irn_arity, new_irn_arity;
- /* relink bad predeseccors of a block */
+ /* relink bad predecessors of a block */
if (get_irn_op(n) == op_Block)
relink_bad_block_predecessors(n, env);
/* If Phi node relink its block and its predecessors */
if (get_irn_op(n) == op_Phi) {
- /* Relink predeseccors of phi's block */
+ /* Relink predecessors of phi's block */
block = get_nodes_block(n);
if (get_irn_link(block) == NULL)
relink_bad_block_predecessors(block, env);
old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
old_irn_arity = ARR_LEN(old_in);
- /* Relink Phi predeseccors if count of predeseccors changed */
+ /* Relink Phi predecessors if count of predecessors changed */
if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
- /* set new predeseccors in array
+ /* set new predecessors in array
n->in[0] remains the same block */
new_irn_arity = 1;
for(i = 1; i < old_irn_arity; i++)
}
/*
- * Removes Bad Bad predecesors from Blocks and the corresponding
+ * Removes Bad Bad predecessors from Blocks and the corresponding
* inputs to Phi nodes as in dead_node_elimination but without
* copying the graph.
* On walking up set the link field to NULL, on walking down call
}
#endif
- /* -- Turn cse back on. -- */
+ /* -- Turn CSE back on. -- */
set_optimize(rem_opt);
return 1;
int n_callers_orig; /**< for statistics */
} inline_irg_env;
+/**
+ * Allocate a new nvironment for inlining.
+ */
static inline_irg_env *new_inline_irg_env(void) {
- inline_irg_env *env = xmalloc(sizeof(*env));
- env->n_nodes = -2; /* uncount Start, End */
- env->n_nodes_orig = -2; /* uncount Start, End */
- env->call_nodes = eset_create();
- env->n_call_nodes = 0;
+ inline_irg_env *env = xmalloc(sizeof(*env));
+ env->n_nodes = -2; /* do not count count Start, End */
+ env->n_nodes_orig = -2; /* do not count Start, End */
+ env->call_nodes = eset_create();
+ env->n_call_nodes = 0;
env->n_call_nodes_orig = 0;
- env->n_callers = 0;
- env->n_callers_orig = 0;
+ env->n_callers = 0;
+ env->n_callers_orig = 0;
return env;
}
+/**
+ * destroy an environment for inlining.
+ */
static void free_inline_irg_env(inline_irg_env *env) {
eset_destroy(env->call_nodes);
free(env);
}
+/**
+ * post-walker: collect all calls in the inline-environment
+ * of a graph and sum some statistics.
+ */
static void collect_calls2(ir_node *call, void *env) {
inline_irg_env *x = (inline_irg_env *)env;
ir_op *op = get_irn_op(call);
ir_graph *callee;
- /* count nodes in irg */
+ /* count meaningful nodes in irg */
if (op != op_Proj && op != op_Tuple && op != op_Sync) {
x->n_nodes++;
x->n_nodes_orig++;
/* count all static callers */
callee = get_call_called_irg(call);
if (callee) {
- ((inline_irg_env *)get_irg_link(callee))->n_callers++;
- ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
+ inline_irg_env *callee_env = get_irg_link(callee);
+ callee_env->n_callers++;
+ callee_env->n_callers_orig++;
}
}
+/**
+ * Returns TRUE if the number of callers in 0 in the irg's environment,
+ * hence this irg is a leave.
+ */
INLINE static int is_leave(ir_graph *irg) {
return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
}
+/**
+ * Returns TRUE if the number of callers is smaller size in the irg's environment.
+ */
INLINE static int is_smaller(ir_graph *callee, int size) {
return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
}
for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
ir_graph *callee;
- if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
+ if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
callee = get_call_called_irg(call);
if (env->n_nodes > maxsize) continue; // break;
/* will be executed only if needed. */
/*******************************************************************/
+/**
+ * Returns non-zero, is a block is not reachable from Start.
+ */
+static int
+is_Block_unreachable(ir_node *block) {
+ return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
+}
+
/**
* Find the earliest correct block for N. --- Place N into the
* same Block as its dominance-deepest Input.
+ *
+ * We have to avoid calls to get_nodes_block() here
+ * because the graph is floating.
*/
static void
place_floats_early(ir_node *n, pdeq *worklist)
/* Place floating nodes. */
if (get_irn_pinned(n) == op_pin_state_floats) {
int depth = 0;
- ir_node *b = new_Bad(); /* The block to place this node in */
- int bad_recursion = is_Bad(get_nodes_block(n));
+ ir_node *b = NULL; /* The block to place this node in */
+ int bad_recursion = is_Block_unreachable(get_irn_n(n, -1));
assert(get_irn_op(n) != op_Block);
/* Because all loops contain at least one op_pin_state_pinned node, now all
our inputs are either op_pin_state_pinned or place_early has already
been finished on them. We do not have any unfinished inputs! */
- dep_block = get_nodes_block(dep);
- if ((!is_Bad(dep_block)) &&
+ dep_block = get_irn_n(dep, -1);
+ if ((!is_Block_dead(dep_block)) &&
(get_Block_dom_depth(dep_block) > depth)) {
b = dep_block;
depth = get_Block_dom_depth(dep_block);
}
/* Avoid that the node is placed in the Start block */
- if ((depth == 1) && (get_Block_dom_depth(get_nodes_block(n)) > 1)) {
+ if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
assert(b != get_irg_start_block(current_ir_graph));
depth = 2;
}
}
- set_nodes_block(n, b);
+ if (b)
+ set_nodes_block(n, b);
}
/* Add predecessors of non floating nodes on worklist. */
current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
}
-/** Compute the deepest common ancestor of block and dca. */
+/**
+ * Compute the deepest common ancestor of block and dca.
+ */
static ir_node *calc_dca(ir_node *dca, ir_node *block)
{
assert(block);
+
+ /* we do not want to place nodes in dead blocks */
+ if (is_Block_dead(block))
+ return dca;
+
+ /* We found a first legal placement. */
if (!dca) return block;
+
+ /* Find a placement that is dominates both, dca and block. */
while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
block = get_Block_idom(block);
+
while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
dca = get_Block_idom(dca);
}
+
while (block != dca)
{ block = get_Block_idom(block); dca = get_Block_idom(dca); }
if (get_irn_n(consumer, i) == producer) {
ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
- block = calc_dca(block, new_block);
+ if (! is_Block_unreachable(new_block))
+ block = calc_dca(block, new_block);
}
}
+
+ if (! block)
+ block = get_irn_n(producer, -1);
+
} else {
assert(is_no_Block(consumer));
block = get_nodes_block(consumer);
return calc_dca(dca, block);
}
+/* FIXME: the name clashes here with the function from ana/field_temperature.c
+ * please rename. */
static INLINE int get_irn_loop_depth(ir_node *n) {
return get_loop_depth(get_irn_loop(n));
}
/**
* Move n to a block with less loop depth than it's current block. The
* new block must be dominated by early.
+ *
+ * @param n the node that should be moved
+ * @param early the earliest block we can n move to
*/
static void
move_out_of_loops (ir_node *n, ir_node *early)
dca with the least loop nesting depth, but still dominated
by our early placement. */
dca = get_nodes_block(n);
+
best = dca;
while (dca != early) {
dca = get_Block_idom(dca);
* `optimal' Block between the latest and earliest legal block.
* The `optimal' block is the dominance-deepest block of those
* with the least loop-nesting-depth. This places N out of as many
- * loops as possible and then makes it as control dependant as
+ * loops as possible and then makes it as control dependent as
* possible.
*/
static void
(get_irn_mode(n) != mode_X)) {
/* Remember the early placement of this block to move it
out of loop no further than the early placement. */
- early = get_nodes_block(n);
+ early = get_irn_n(n, -1);
- /* Do not move code not reachable from Start. For
- * these we could not compute dominator information. */
- if (is_Bad(early) || get_Block_dom_depth(early) == -1)
- return;
+ /*
+ * BEWARE: Here we also get code, that is live, but
+ * was in a dead block. If the node is life, but because
+ * of CSE in a dead block, we still might need it.
+ */
/* Assure that our users are all placed, except the Phi-nodes.
--- Each data flow cycle contains at least one Phi-node. We
final region of our users, which is OK with Phi-nodes, as they
are op_pin_state_pinned, and they never have to be placed after a
producer of one of their inputs in the same block anyway. */
- for (i = 0; i < get_irn_n_outs(n); i++) {
+ for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
ir_node *succ = get_irn_out(n, i);
if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
place_floats_late(succ, worklist);
dominator tree of all nodes'
blocks depending on us; our final
placement has to dominate DCA. */
- for (i = 0; i < get_irn_n_outs(n); i++) {
+ for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
ir_node *out = get_irn_out(n, i);
+ ir_node *outbl;
+
+ if (get_irn_op(out) == op_End) {
+ /*
+ * This consumer is the End node, a keep alive edge.
+ * This is not a real consumer, so we ignore it
+ */
+ continue;
+ }
+
/* ignore if out is in dead code */
- ir_node *outbl = get_nodes_block(out);
- if (is_Bad(outbl) || get_Block_dom_depth(outbl) == -1)
+ outbl = get_irn_n(out, -1);
+ if (is_Block_unreachable(outbl))
continue;
- dca = consumer_dom_dca (dca, out, n);
+ dca = consumer_dom_dca(dca, out, n);
}
if (dca) {
set_nodes_block(n, dca);
-
move_out_of_loops (n, early);
}
/* else all outs are in dead code */
}
/* Add predecessors of all non-floating nodes on list. (Those of floating
- nodes are placeded already and therefore are marked.) */
+ nodes are placed already and therefore are marked.) */
for (i = 0; i < get_irn_n_outs(n); i++) {
ir_node *succ = get_irn_out(n, i);
if (irn_not_visited(get_irn_out(n, i))) {
place_early(worklist);
/* place_early invalidates the outs, place_late needs them. */
- compute_outs(irg);
+ compute_irg_outs(irg);
/* Now move the nodes down in the dominator tree. This reduces the
unnecessary executions of the node. */
place_late(worklist);