#include "config.h"
#endif
+#include <stdlib.h>
+
+#include "ircons.h"
+#include "debug.h"
+#include "irhooks.h"
+
#include "bearch.h"
#include "belower.h"
#include "benode_t.h"
#include "bechordal_t.h"
-#include "besched.h"
+#include "besched_t.h"
+#include "bestat.h"
+#include "benodesets.h"
#include "irgmod.h"
-#include "iredges.h"
+#include "iredges_t.h"
#include "irgwalk.h"
+#ifdef HAVE_MALLOC_H
+ #include <malloc.h>
+#endif
+#ifdef HAVE_ALLOCA_H
+ #include <alloca.h>
+#endif
+
+#undef KEEP_ALIVE_COPYKEEP_HACK
+
#undef is_Perm
-#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
+#define is_Perm(arch_env, irn) (arch_irn_class_is(arch_env, irn, perm))
+
+/* associates op with it's copy and CopyKeep */
+typedef struct {
+ ir_node *op; /* an irn which must be different */
+ pset *copies; /* all non-spillable copies of this irn */
+ const arch_register_class_t *cls;
+} op_copy_assoc_t;
+
+/* environment for constraints */
+typedef struct {
+ be_irg_t *birg;
+ pset *op_set;
+ struct obstack obst;
+ DEBUG_ONLY(firm_dbg_module_t *dbg;)
+} constraint_env_t;
/* lowering walker environment */
typedef struct _lower_env_t {
- be_chordal_env_t *chord_env;
- int do_copy;
+ be_chordal_env_t *chord_env;
+ unsigned do_copy:1;
+ DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
} lower_env_t;
/* holds a perm register pair */
PERM_COPY
} perm_type_t;
+/* structure to represent cycles or chains in a perm */
typedef struct _perm_cycle_t {
const arch_register_t **elems; /**< the registers in the cycle */
int n_elems; /**< number of elements in the cycle */
perm_type_t type; /**< type (CHAIN or CYCLE) */
} perm_cycle_t;
+static void kill_node(ir_node *irn, int kill_block) {
+ int i, arity = get_irn_arity(irn);
+ ir_graph *irg = get_irn_irg(irn);
+
+ for (i = -1; i < arity; ++i) {
+ set_irn_n(irn, i, new_r_Bad(irg));
+ }
+}
+
+/* Compare the two operands */
+static int cmp_op_copy_assoc(const void *a, const void *b) {
+ const op_copy_assoc_t *op1 = a;
+ const op_copy_assoc_t *op2 = b;
+
+ return op1->op != op2->op;
+}
+
/* Compare the in registers of two register pairs */
static int compare_reg_pair(const void *a, const void *b) {
const reg_pair_t *pair_a = a;
return -1;
}
+/* returns the number register pairs marked as checked */
static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
int i, n_checked = 0;
static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
int i;
- for (i = 0; i < n; i++) {
- if (in_out) {
+ if (in_out) {
+ for (i = 0; i < n; i++) {
/* out register matches */
if (pairs[i].out_reg->index == reg->index)
return pairs[i].out_node;
}
- else {
+ }
+ else {
+ for (i = 0; i < n; i++) {
/* in register matches */
if (pairs[i].in_reg->index == reg->index)
return pairs[i].in_node;
return NULL;
}
+/**
+ * Gets the index in the register pair array where the in/out register
+ * corresponds to reg_idx.
+ *
+ * @param pairs The array of register pairs
+ * @param n The number of pairs
+ * @param reg The register index to look for
+ * @param in_out 0 == look for IN register, 1 == look for OUT register
+ * @return The corresponding index in pairs or -1 if not found
+ */
+static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
+ int i;
+
+ if (in_out) {
+ for (i = 0; i < n; i++) {
+ /* out register matches */
+ if (pairs[i].out_reg->index == reg_idx)
+ return i;
+ }
+ }
+ else {
+ for (i = 0; i < n; i++) {
+ /* in register matches */
+ if (pairs[i].in_reg->index == reg_idx)
+ return i;
+ }
+ }
+
+ return -1;
+}
+
/**
* Gets an array of register pairs and tries to identify a cycle or chain starting
* at position start.
static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
int head = pairs[start].in_reg->index;
int cur_idx = pairs[start].out_reg->index;
- int n_pairs_done = get_n_checked_pairs(pairs, n) + 1;
+ int cur_pair_idx = start;
+ int n_pairs_done = get_n_checked_pairs(pairs, n);
int idx;
+ perm_type_t cycle_tp = PERM_CYCLE;
+
+ /* We could be right in the middle of a chain, so we need to find the start */
+ while (head != cur_idx) {
+ /* goto previous register in cycle or chain */
+ cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
+
+ if (cur_pair_idx < 0) {
+ cycle_tp = PERM_CHAIN;
+ break;
+ }
+ else {
+ head = pairs[cur_pair_idx].in_reg->index;
+ start = cur_pair_idx;
+ }
+ }
/* assume worst case: all remaining pairs build a cycle or chain */
- cycle->elems = calloc(n - n_pairs_done, sizeof(cycle->elems[0]));
+ cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
cycle->n_elems = 2; /* initial number of elements is 2 */
cycle->elems[0] = pairs[start].in_reg;
cycle->elems[1] = pairs[start].out_reg;
- cycle->type = PERM_CHAIN; /* default is CHAIN, only changed when we found a cycle */
-
- /* mark the first pair as checked */
- pairs[start].checked = 1;
+ cycle->type = cycle_tp;
+ cur_idx = pairs[start].out_reg->index;
idx = 2;
/* check for cycle or end of a chain */
- while (cur_idx != head && n_pairs_done < n) {
+ while (cur_idx != head) {
/* goto next register in cycle or chain */
- cur_idx = pairs[cur_idx].out_reg->index;
+ cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
+
+ if (cur_pair_idx < 0)
+ break;
+
+ cur_idx = pairs[cur_pair_idx].out_reg->index;
/* it's not the first element: insert it */
if (cur_idx != head) {
- cycle->elems[idx++] = pairs[cur_idx].out_reg;
+ cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
cycle->n_elems++;
-
- pairs[cur_idx].checked = 1;
- n_pairs_done++;
}
else {
/* we are there where we started -> CYCLE */
}
}
+ /* mark all pairs having one in/out register with cycle in common as checked */
+ for (idx = 0; idx < cycle->n_elems; idx++) {
+ cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
+
+ if (cur_pair_idx >= 0)
+ pairs[cur_pair_idx].checked = 1;
+
+ cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
+
+ if (cur_pair_idx >= 0)
+ pairs[cur_pair_idx].checked = 1;
+ }
+
return cycle;
}
/**
* Lowers a perm node. Resolves cycles and creates a bunch of
* copy and swap operations to permute registers.
+ * Note: The caller of this function has to make sure, that irn
+ * is a Perm node.
*
* @param irn The perm node
+ * @param block The block the perm node belongs to
* @param walk_env The environment
*/
-static void lower_perms_walker(ir_node *irn, void *walk_env) {
- const be_node_factory_t *fact;
+static void lower_perm_node(ir_node *irn, void *walk_env) {
const arch_register_class_t *reg_class;
const arch_env_t *arch_env;
- lower_env_t *env = walk_env;
+ lower_env_t *env = walk_env;
+ int real_size = 0;
+ int keep_perm = 0;
+ int n, i, pn, do_copy, j, n_ops;
reg_pair_t *pairs;
const ir_edge_t *edge;
perm_cycle_t *cycle;
- int n, i, pn, do_copy;
- ir_node *block, *arg, *res, *sched_point, *in[2];
+ ir_node *sched_point, *block, *in[2];
+ ir_node *arg1, *arg2, *res1, *res2;
ir_node *cpyxchg = NULL;
+ DEBUG_ONLY(firm_dbg_module_t *mod;)
- fact = env->chord_env->main_env->node_factory;
- arch_env = env->chord_env->main_env->arch_env;
+ arch_env = env->chord_env->birg->main_env->arch_env;
do_copy = env->do_copy;
-
- /* check if perm */
- if (! is_Perm(arch_env, irn))
- return;
-
- block = get_nodes_block(irn);
+ DEBUG_ONLY(mod = env->dbg_module;)
+ block = get_nodes_block(irn);
/*
Get the schedule predecessor node to the perm
should be ok.
*/
sched_point = sched_prev(irn);
+ DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
assert(sched_point && "Perm is not scheduled or has no predecessor");
n = get_irn_arity(irn);
assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
- pairs = calloc(n, sizeof(pairs[0]));
+ pairs = alloca(n * sizeof(pairs[0]));
/* build the list of register pairs (in, out) */
i = 0;
/* Mark all equal pairs as checked, and exchange the OUT proj with
the IN node. */
- i = 0;
- while (pairs[i].in_reg->index == pairs[i].out_reg->index) {
- /* remove the proj from the schedule */
- sched_remove(pairs[i].out_node);
-
- /* remove the argument from schedule */
- sched_remove(pairs[i].in_node);
+ for (i = 0; i < n; i++) {
+ if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
+ DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
+ irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
+
+ /* We have to check for a special case:
+ The in-node could be a Proj from a Perm. In this case,
+ we need to correct the projnum */
+ if (is_Perm(arch_env, pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
+ set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
+ }
- /* exchange the proj with the argument */
- exchange(pairs[i].out_node, pairs[i].in_node);
+ /* remove the proj from the schedule */
+ sched_remove(pairs[i].out_node);
- /* add the argument after the magic scheduling point */
- sched_add_after(sched_point, pairs[i].in_node);
+ /* reroute the edges from the proj to the argument */
+ edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg);
+ set_irn_n(pairs[i].out_node, 0, new_Bad());
- pairs[i++].checked = 1;
+ pairs[i].checked = 1;
+ }
}
/* Set do_copy to 0 if it's on but we have no free register */
do_copy = 0;
}
+ real_size = n - get_n_checked_pairs(pairs, n);
+
+ be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
+
/* check for cycles and chains */
while (get_n_checked_pairs(pairs, n) < n) {
- i = 0;
+ i = n_ops = 0;
/* go to the first not-checked pair */
while (pairs[i].checked) i++;
- cycle = calloc(1, sizeof(*cycle));
+ cycle = xcalloc(1, sizeof(*cycle));
cycle = get_perm_cycle(cycle, pairs, n, i);
-//todo: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
+ DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
+ for (j = 0; j < cycle->n_elems; j++) {
+ DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
+ }
+ DB((mod, LEVEL_1, "\n"));
+
+ /*
+ We don't need to do anything if we have a Perm with two
+ elements which represents a cycle, because those nodes
+ already represent exchange nodes
+ */
+ if (n == 2 && cycle->type == PERM_CYCLE) {
+ free(cycle);
+ keep_perm = 1;
+ continue;
+ }
+
+//TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
// the copy cascade (this reduces the cycle into a chain)
/* build copy/swap nodes from back to front */
- for (i = cycle->n_elems - 2; i >= 0; i++) {
- arg = get_node_for_register(pairs, n, cycle->elems[i], 0);
- res = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
+ for (i = cycle->n_elems - 2; i >= 0; i--) {
+ arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
+ arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
+ res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
+ res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
/*
If we have a cycle and don't copy: we need to create exchange nodes
NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
OUT_2 = out node with register i
*/
if (cycle->type == PERM_CYCLE && !do_copy) {
- in[0] = arg;
- in[1] = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
-
- cpyxchg = new_Perm(fact, reg_class, env->chord_env->irg, block, 2, in);
- set_Proj_pred(res, cpyxchg);
- set_Proj_proj(res, 0);
- set_Proj_pred(get_node_for_register(pairs, n, cycle->elems[i], 1), cpyxchg);
- set_Proj_proj(get_node_for_register(pairs, n, cycle->elems[i], 1), 1);
+ in[0] = arg1;
+ in[1] = arg2;
+
+ /* At this point we have to handle the following problem: */
+ /* */
+ /* If we have a cycle with more than two elements, then */
+ /* this could correspond to the following Perm node: */
+ /* */
+ /* +----+ +----+ +----+ */
+ /* | r1 | | r2 | | r3 | */
+ /* +-+--+ +-+--+ +--+-+ */
+ /* | | | */
+ /* | | | */
+ /* +-+--------+---------+-+ */
+ /* | Perm | */
+ /* +-+--------+---------+-+ */
+ /* | | | */
+ /* | | | */
+ /* +-+--+ +-+--+ +--+-+ */
+ /* |Proj| |Proj| |Proj| */
+ /* | r2 | | r3 | | r1 | */
+ /* +----+ +----+ +----+ */
+ /* */
+ /* This node is about to be split up into two 2x Perm's */
+ /* for which we need 4 Proj's and the one additional Proj */
+ /* of the first Perm has to be one IN of the second. So in */
+ /* general we need to create one additional Proj for each */
+ /* "middle" Perm and set this to one in node of the successor */
+ /* Perm. */
+
+ DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
+ irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
+ DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
+ irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
+
+ cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
+ n_ops++;
+
+ if (i > 0) {
+ /* cycle is not done yet */
+ int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
+
+ /* create intermediate proj */
+ res1 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0);
+
+ /* set as in for next Perm */
+ pairs[pidx].in_node = res1;
+ }
+ else {
+ sched_remove(res1);
+ }
+
+ sched_remove(res2);
+
+ set_Proj_pred(res2, cpyxchg);
+ set_Proj_proj(res2, 0);
+ set_Proj_pred(res1, cpyxchg);
+ set_Proj_proj(res1, 1);
+
+ sched_add_after(sched_point, res1);
+ sched_add_after(sched_point, res2);
+
+ arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
+ arch_set_irn_register(arch_env, res1, cycle->elems[i]);
+
+ /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
+ sched_add_after(sched_point, cpyxchg);
+
+ DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
+
+ /* set the new scheduling point */
+ sched_point = res1;
}
else {
- cpyxchg = new_Copy(fact, reg_class, env->chord_env->irg, block, arg);
+ DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
+ irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
+
+ cpyxchg = be_new_Copy(reg_class, env->chord_env->irg, block, arg1);
arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
+ n_ops++;
/* remove the proj from the schedule */
- sched_remove(res);
+ sched_remove(res2);
/* exchange copy node and proj */
- exchange(res, cpyxchg);
- }
+ exchange(res2, cpyxchg);
- /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
- sched_add_after(sched_point, cpyxchg);
+ /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
+ sched_add_after(sched_point, cpyxchg);
+
+ /* set the new scheduling point */
+ sched_point = cpyxchg;
+ }
}
- free(cycle->elems);
+ be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
+
+ free((void *) cycle->elems);
free(cycle);
}
/* remove the perm from schedule */
- sched_remove(irn);
+ if (! keep_perm) {
+ kill_node(irn, 0);
+ sched_remove(irn);
+ }
+}
+
+
+
+static int get_n_out_edges(const ir_node *irn) {
+ const ir_edge_t *edge;
+ int cnt = 0;
+
+ foreach_out_edge(irn, edge) {
+ cnt++;
+ }
+
+ return cnt;
+}
+
+static ir_node *belower_skip_proj(ir_node *irn) {
+ while(is_Proj(irn))
+ irn = get_Proj_pred(irn);
+ return irn;
+}
+
+static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) {
+ const arch_env_t *arch_env = env->birg->main_env->arch_env;
+ ir_node *block = get_nodes_block(irn);
+ ir_node *cur_node;
+
+ for (cur_node = sched_prev(irn);
+ ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
+ cur_node = sched_prev(cur_node))
+ {
+ if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill))
+ return cur_node;
+ }
+
+ return NULL;
}
+static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
+ be_irg_t *birg = env->birg;
+ pset *op_set = env->op_set;
+ const arch_env_t *arch_env = birg->main_env->arch_env;
+ ir_node *block = get_nodes_block(irn);
+ const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
+ ir_node *in[2], *keep, *cpy;
+ op_copy_assoc_t key, *entry;
+ DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
+
+ if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
+ DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
+ return;
+ }
+
+ /* Make a not spillable copy of the different node */
+ /* this is needed because the different irn could be */
+ /* in block far far away */
+ /* The copy is optimized later if not needed */
+
+ /* check if already exists such a copy in the schedule immediatly before */
+ cpy = find_copy(env, belower_skip_proj(irn), other_different);
+ if (! cpy) {
+ cpy = be_new_Copy(cls, birg->irg, block, other_different);
+ be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
+ DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
+ }
+ else {
+ DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
+ }
+
+ in[0] = irn;
+ in[1] = cpy;
+
+ /* Add the Keep resp. CopyKeep and reroute the users */
+ /* of the other_different irn in case of CopyKeep. */
+ if (get_n_out_edges(other_different) == 0) {
+ keep = be_new_Keep(cls, birg->irg, block, 2, in);
+ }
+ else {
+ keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different));
+ be_node_set_reg_class(keep, 1, cls);
+ }
+
+ DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
+
+ /* insert copy and keep into schedule */
+ assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
+ if (! sched_is_scheduled(cpy))
+ sched_add_before(belower_skip_proj(irn), cpy);
+ sched_add_after(irn, keep);
+
+ /* insert the other different and it's copies into the set */
+ key.op = other_different;
+ key.copies = NULL;
+ entry = pset_find(op_set, &key, nodeset_hash(other_different));
+
+ if (! entry) {
+ entry = obstack_alloc(&env->obst, sizeof(*entry));
+ entry->copies = pset_new_ptr_default();
+ entry->op = other_different;
+ entry->cls = cls;
+ }
+
+ /* insert copy */
+ pset_insert_ptr(entry->copies, cpy);
+
+ /* insert keep in case of CopyKeep */
+ if (be_is_CopyKeep(keep))
+ pset_insert_ptr(entry->copies, keep);
+
+ pset_insert(op_set, entry, nodeset_hash(other_different));
+}
+
+/**
+ * Checks if node has a should_be_different constraint in output
+ * and adds a Keep then to assure the constraint.
+ */
+static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
+ const arch_register_req_t *req;
+ arch_register_req_t req_temp;
+
+ req = arch_get_register_req(env->birg->main_env->arch_env, &req_temp, irn, -1);
+
+ if (req) {
+ if (arch_register_req_is(req, should_be_different)) {
+ gen_assure_different_pattern(irn, req->other_different, env);
+ }
+ else if (arch_register_req_is(req, should_be_different_from_all)) {
+ int i, n = get_irn_arity(belower_skip_proj(irn));
+ for (i = 0; i < n; i++) {
+ gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env);
+ }
+ }
+ }
+}
+
+
+
/**
- * Walks over all nodes in an irg and performs perm lowering.
+ * Calls the functions to assure register constraints.
+ *
+ * @param irn The node to be checked for lowering
+ * @param walk_env The walker environment
+ */
+static void assure_constraints_walker(ir_node *irn, void *walk_env) {
+ if (is_Block(irn))
+ return;
+
+ if (mode_is_datab(get_irn_mode(irn)))
+ assure_different_constraints(irn, walk_env);
+
+ return;
+}
+
+/**
+ * Melt all copykeeps pointing to the same node
+ * (or Projs of the same node), copying the same operand.
+ */
+static void melt_copykeeps(constraint_env_t *cenv) {
+ op_copy_assoc_t *entry;
+
+ /* for all */
+ foreach_pset(cenv->op_set, entry) {
+ int idx, num_ck;
+ ir_node *cp;
+ struct obstack obst;
+ ir_node **ck_arr, **melt_arr;
+
+ obstack_init(&obst);
+
+ /* collect all copykeeps */
+ num_ck = idx = 0;
+ foreach_pset(entry->copies, cp) {
+ if (be_is_CopyKeep(cp)) {
+ obstack_grow(&obst, &cp, sizeof(cp));
+ ++num_ck;
+ }
+#ifdef KEEP_ALIVE_COPYKEEP_HACK
+ else {
+ set_irn_mode(cp, mode_ANY);
+ keep_alive(cp);
+ }
+#endif /* KEEP_ALIVE_COPYKEEP_HACK */
+ }
+
+ /* compare each copykeep with all other copykeeps */
+ ck_arr = (ir_node **)obstack_finish(&obst);
+ for (idx = 0; idx < num_ck; ++idx) {
+ ir_node *ref, *ref_mode_T;
+
+ if (ck_arr[idx]) {
+ int j, n_melt;
+ ir_node **new_ck_in;
+ ir_node *new_ck;
+ ir_node *sched_pt = NULL;
+
+ n_melt = 1;
+ ref = ck_arr[idx];
+ ref_mode_T = skip_Proj(get_irn_n(ref, 1));
+ obstack_grow(&obst, &ref, sizeof(ref));
+
+ DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
+
+ /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
+ for (j = 0; j < num_ck; ++j) {
+ ir_node *cur_ck = ck_arr[j];
+
+ if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
+ obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
+ pset_remove_ptr(entry->copies, cur_ck);
+ DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
+ ck_arr[j] = NULL;
+ ++n_melt;
+ sched_remove(cur_ck);
+ }
+ }
+ ck_arr[idx] = NULL;
+
+ /* check, if we found some candidates for melting */
+ if (n_melt == 1) {
+ DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
+ continue;
+ }
+
+ pset_remove_ptr(entry->copies, ref);
+ sched_remove(ref);
+
+ melt_arr = (ir_node **)obstack_finish(&obst);
+ /* melt all found copykeeps */
+ NEW_ARR_A(ir_node *, new_ck_in, n_melt);
+ for (j = 0; j < n_melt; ++j) {
+ new_ck_in[j] = get_irn_n(melt_arr[j], 1);
+
+ /* now, we can kill the melted keep, except the */
+ /* ref one, we still need some information */
+ if (melt_arr[j] != ref)
+ kill_node(melt_arr[j], 0);
+ }
+
+#ifdef KEEP_ALIVE_COPYKEEP_HACK
+ new_ck = be_new_CopyKeep(entry->cls, cenv->birg->irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
+ keep_alive(new_ck);
+#else
+ new_ck = be_new_CopyKeep(entry->cls, cenv->birg->irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, get_irn_mode(ref));
+#endif /* KEEP_ALIVE_COPYKEEP_HACK */
+
+ /* set register class for all keeped inputs */
+ for (j = 1; j <= n_melt; ++j)
+ be_node_set_reg_class(new_ck, j, entry->cls);
+
+ pset_insert_ptr(entry->copies, new_ck);
+
+ /* find scheduling point */
+ if (get_irn_mode(ref_mode_T) == mode_T) {
+ /* walk along the Projs */
+ for (sched_pt = sched_next(ref_mode_T); is_Proj(sched_pt) || be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt); sched_pt = sched_next(sched_pt))
+ /* just walk along the schedule until a non-Proj/Keep/CopyKeep node is found*/ ;
+ }
+ else {
+ sched_pt = ref_mode_T;
+ }
+
+ sched_add_before(sched_pt, new_ck);
+ DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
+
+ /* finally: kill the reference copykeep */
+ kill_node(ref, 0);
+ }
+ }
+
+ obstack_free(&obst, NULL);
+ }
+}
+
+/**
+ * Walks over all nodes to assure register constraints.
+ *
+ * @param birg The birg structure containing the irg
+ */
+void assure_constraints(be_irg_t *birg) {
+ constraint_env_t cenv;
+ op_copy_assoc_t *entry;
+ dom_front_info_t *dom;
+ ir_node **nodes;
+ FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
+
+ DEBUG_ONLY(cenv.dbg = mod;)
+ cenv.birg = birg;
+ cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
+ obstack_init(&cenv.obst);
+
+ irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, &cenv);
+
+ /* melt copykeeps, pointing to projs of */
+ /* the same mode_T node and keeping the */
+ /* same operand */
+ melt_copykeeps(&cenv);
+
+ /* introduce copies needs dominance information */
+ dom = be_compute_dominance_frontiers(birg->irg);
+
+ /* for all */
+ foreach_pset(cenv.op_set, entry) {
+ int n;
+ ir_node *cp;
+
+ n = pset_count(entry->copies);
+ nodes = alloca((n + 1) * sizeof(nodes[0]));
+
+ /* put the node in an array */
+ n = 0;
+ nodes[n++] = entry->op;
+ DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
+
+ /* collect all copies */
+ foreach_pset(entry->copies, cp) {
+ nodes[n++] = cp;
+ DB((mod, LEVEL_1, ", %+F ", cp));
+ }
+
+ DB((mod, LEVEL_1, "\n"));
+
+ /* introduce the copies for the operand and it's copies */
+ be_ssa_constr(dom, NULL, n, nodes);
+
+
+ /* Could be that not all CopyKeeps are really needed, */
+ /* so we transform unnecessary ones into Keeps. */
+ foreach_pset(entry->copies, cp) {
+ if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
+ ir_node *keep;
+ int n = get_irn_arity(cp);
+
+ keep = be_new_Keep(arch_get_irn_reg_class(birg->main_env->arch_env, cp, -1),
+ birg->irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
+ sched_add_before(cp, keep);
+
+ /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
+ kill_node(cp, 1);
+ sched_remove(cp);
+ }
+ }
+
+ del_pset(entry->copies);
+ }
+
+ be_free_dominance_frontiers(dom);
+
+ del_pset(cenv.op_set);
+ obstack_free(&cenv.obst, NULL);
+}
+
+
+
+/**
+ * Calls the corresponding lowering function for the node.
+ *
+ * @param irn The node to be checked for lowering
+ * @param walk_env The walker environment
+ */
+static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
+ lower_env_t *env = walk_env;
+ const arch_env_t *arch_env = env->chord_env->birg->main_env->arch_env;
+
+ if (! is_Block(irn) && ! is_Proj(irn)) {
+ if (is_Perm(arch_env, irn)) {
+ lower_perm_node(irn, walk_env);
+ }
+ }
+
+ return;
+}
+
+/**
+ * Walks over all blocks in an irg and performs lowering need to be
+ * done after register allocation (e.g. perm lowering).
*
* @param chord_env The chordal environment containing the irg
* @param do_copy 1 == resolve cycles with a free reg if available
*/
-void lower_perms(be_chordal_env_t *chord_env, int do_copy) {
+void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy) {
lower_env_t env;
- env.chord_env = chord_env;
- env.do_copy = do_copy;
+ env.chord_env = chord_env;
+ env.do_copy = do_copy;
+ FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
- irg_walk_blkwise_graph(chord_env->irg, NULL, lower_perms_walker, &env);
+ irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);
}
+
+#undef is_Perm
projects / libfirm / blobdiff