#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_t.h"
+#include "bestat.h"
#include "irgmod.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 is_Perm
#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
-#undef is_Call
-#define is_Call(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_call)
+/* 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 */
+} op_copy_assoc_t;
+
+/* environment for constraints */
+typedef struct {
+ be_irg_t *birg;
+ pset *op_set;
+ struct obstack obst;
+} constraint_env_t;
/* lowering walker environment */
typedef struct _lower_env_t {
be_chordal_env_t *chord_env;
- int do_copy;
- firm_dbg_module_t *dbg_module;
+ unsigned do_copy:1;
+ DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
} lower_env_t;
/* holds a perm register pair */
perm_type_t type; /**< type (CHAIN or CYCLE) */
} perm_cycle_t;
+/* 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;
int cur_idx = pairs[start].out_reg->index;
int cur_pair_idx = start;
int n_pairs_done = get_n_checked_pairs(pairs, n);
- int idx, done = 0;
+ 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 && !done) {
+ 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;
- done = 1;
+ break;
}
else {
head = pairs[cur_pair_idx].in_reg->index;
}
/* assume worst case: all remaining pairs build a cycle or chain */
- cycle->elems = calloc((n - n_pairs_done) * 2, 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 = cycle_tp;
- n_pairs_done++;
+ 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_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
if (cur_idx != head) {
cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
cycle->n_elems++;
-
- n_pairs_done++;
}
else {
/* we are there where we started -> CYCLE */
static void lower_perm_node(ir_node *irn, void *walk_env) {
const arch_register_class_t *reg_class;
const arch_env_t *arch_env;
- firm_dbg_module_t *mod;
- lower_env_t *env = walk_env;
+ lower_env_t *env = walk_env;
+ int real_size = 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, j;
ir_node *sched_point, *block, *in[2];
ir_node *arg1, *arg2, *res1, *res2;
ir_node *cpyxchg = NULL;
+ DEBUG_ONLY(firm_dbg_module_t *mod;)
- arch_env = env->chord_env->main_env->arch_env;
+ arch_env = env->chord_env->birg->main_env->arch_env;
do_copy = env->do_copy;
- mod = env->dbg_module;
+ DEBUG_ONLY(mod = env->dbg_module;)
block = get_nodes_block(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;
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);
DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
//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 */
- for (i = 0; i < cycle->n_elems - 1; i++) {
+ /* build copy/swap nodes from back to front */
+ 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
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(res1);
sched_remove(res2);
set_Proj_pred(res2, cpyxchg);
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 {
DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
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(res2);
/* exchange copy node and proj */
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);
- 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 = 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);
}
-/**
- * Adds Projs to keep nodes for each register class, which eats the
- * caller saved registers.
- * Note: The caller has to make sure, that call is a Call
- *
- * @param call The Call node
- * @param walk_env The walker environment
- */
-static void lower_call_node(ir_node *call, const void *walk_env) {
- const arch_env_t *arch_env = walk_env;
- const arch_register_class_t *reg_class;
- int i, j, set_size = 0, pn, keep_arity;
- arch_isa_t *isa = arch_env_get_isa(arch_env);
- const ir_node *proj_T = NULL;
- ir_node **in_keep, *block = get_nodes_block(call);
- bitset_t *proj_set;
- const ir_edge_t *edge;
- const arch_register_t *reg;
-
- /* Prepare the bitset where we store the projnums which are already in use*/
- for (i = 0; i < arch_isa_get_n_reg_class(isa); i++) {
- reg_class = arch_isa_get_reg_class(isa, i);
- set_size += arch_register_class_n_regs(reg_class);
+static int get_n_out_edges(const ir_node *irn) {
+ const ir_edge_t *edge;
+ int cnt = 0;
+
+ foreach_out_edge(irn, edge) {
+ cnt++;
}
- in_keep = malloc(set_size * sizeof(ir_node *));
+ return cnt;
+}
- proj_set = bitset_malloc(set_size);
- bitset_clear_all(proj_set);
+static ir_node *belower_skip_proj(ir_node *irn) {
+ while(is_Proj(irn))
+ irn = get_Proj_pred(irn);
+ return irn;
+}
- /* check if there is a ProjT node and which arguments are used */
- foreach_out_edge(call, edge) {
- if (get_irn_mode(get_edge_src_irn(edge)) == mode_T)
- proj_T = get_edge_src_irn(edge);
+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;
+ FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower");
+
+ 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;
}
- /* set all used arguments */
- if (proj_T) {
- foreach_out_edge(proj_T, edge) {
- ir_node *proj = get_edge_src_irn(edge);
+ /* 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 */
- assert(is_Proj(proj));
- bitset_set(proj_set, get_Proj_proj(proj));
- }
+ cpy = be_new_Copy(cls, birg->irg, block, other_different);
+ be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
+
+ 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 {
- proj_T = new_r_Proj(current_ir_graph, block, call, mode_T, pn_Call_T_result);
+ keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different));
+ be_node_set_reg_class(keep, 1, cls);
}
- /* Create for each caller save register a proj (keep node argument) */
- /* if this proj is not already present */
- for (i = 0; i < arch_isa_get_n_reg_class(isa); i++) {
+ /* let the copy point to the other_different irn */
+ set_irn_n(cpy, 0, other_different);
- /* reset the keep input, as we need one keep for each register class */
- memset(in_keep, 0, set_size * sizeof(ir_node *));
- keep_arity = 0;
- reg_class = arch_isa_get_reg_class(isa, i);
+ /* insert copy and keep into schedule */
+ assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
+ sched_add_before(belower_skip_proj(irn), cpy);
+ sched_add_after(irn, keep);
- for (j = 0; j < arch_register_class_n_regs(reg_class); j++) {
- reg = arch_register_for_index(reg_class, j);
+ /* insert the other different and it's copies into the set */
+ key.op = other_different;
+ key.copies = NULL;
+ entry = pset_find(op_set, &key, HASH_PTR(other_different));
- /* only check caller save registers */
- if (arch_register_type_is(reg, caller_saved)) {
- pn = isa->impl->get_projnum_for_register(isa, reg);
- if (!bitset_is_set(proj_set, pn)) {
- ir_node *proj = new_r_Proj(current_ir_graph, block, (ir_node *)proj_T, mode_Is, pn);
+ if (! entry) {
+ entry = obstack_alloc(&env->obst, sizeof(*entry));
+ entry->copies = pset_new_ptr_default();
+ entry->op = other_different;
+ }
- in_keep[keep_arity++] = proj;
- }
- }
- }
+ /* 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);
- /* ok, we found some caller save register which are not in use but must be saved */
- if (keep_arity) {
- be_new_Keep(reg_class, current_ir_graph, block, keep_arity, in_keep);
+ pset_insert(op_set, entry, HASH_PTR(other_different));
+
+ DBG((mod, LEVEL_1, "created %+F for %+F to assure should_be_different\n", keep, irn));
+}
+
+/**
+ * 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);
+ }
}
}
-
- bitset_free(proj_set);
- return;
}
/**
- * Calls the backend code generator functions to lower Spill and
- * Reload nodes into Store and Load. The backend is fully responsible
- * for creating the new nodes and setting their input correct.
- * Note: The caller of this has to make sure that irn is a Spill
- * or Reload!
+ * Calls the functions to assure register constraints.
*
- * @param irn The Spill/Reload node
+ * @param irn The node to be checked for lowering
* @param walk_env The walker environment
*/
-static void lower_spill_reload(ir_node *irn, void *walk_env) {
- lower_env_t *env = walk_env;
- arch_code_generator_t *cg = env->chord_env->main_env->cg;
- const arch_env_t *aenv = env->chord_env->main_env->arch_env;
- ir_node *res = NULL;
- ir_node *sched_point;
-
- if (be_is_Spill(irn) && cg->impl->lower_spill) {
- res = cg->impl->lower_spill(cg, irn);
- }
- else if (be_is_Reload(irn) && cg->impl->lower_reload) {
- res = cg->impl->lower_reload(cg, irn);
- if (res && res != irn) {
- /* copy the result register from the reload to the load */
- arch_set_irn_register(aenv, res, arch_get_irn_register(aenv, irn));
- }
- }
+static void assure_constraints_walker(ir_node *irn, void *walk_env) {
+ if (is_Block(irn))
+ return;
- if (res && res != irn) {
- sched_point = sched_prev(irn);
- sched_remove(irn);
- exchange(irn, res);
- sched_add_after(sched_point, res);
- }
- else {
- DBG((env->dbg_module, LEVEL_1, "node %+F not lowered\n", irn));
- }
+ if (mode_is_datab(get_irn_mode(irn)))
+ assure_different_constraints(irn, walk_env);
return;
}
+
/**
- * Calls the corresponding lowering function for the node.
+ * Walks over all nodes to assure register constraints.
*
- * @param irn The node to be checked for lowering
- * @param walk_env The walker environment
+ * @param birg The birg structure containing the irg
*/
-static void lower_nodes_before_sched_walker(ir_node *irn, const void *walk_env) {
- const arch_env_t *arch_env = walk_env;
+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");
+
+ 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);
+
+ /* 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));
+ }
- if (!is_Block(irn) && !is_Proj(irn)) {
- if (is_Call(arch_env, irn)) {
- lower_call_node(irn, walk_env);
+ DB((mod, LEVEL_1, "\n"));
+
+ /* introduce the copies for the operand and it's copies */
+ be_ssa_constr(dom, 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);
+ sched_remove(cp);
+
+ /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
+ while (--n >= -1)
+ set_irn_n(cp, n, get_irg_bad(birg->irg));
+ }
}
+
+ del_pset(entry->copies);
}
- return;
+ be_free_dominance_frontiers(dom);
+
+ del_pset(cenv.op_set);
+ obstack_free(&cenv.obst, NULL);
}
+
/**
* Calls the corresponding lowering function for the node.
*
*/
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->main_env->arch_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);
}
- else if (be_is_Spill(irn) || be_is_Reload(irn)) {
- lower_spill_reload(irn, walk_env);
- }
}
return;
}
-
-
-/**
- * Walks over all blocks in an irg and performs lowering need
- * to be done before scheduling (e.g. call lowering).
- *
- * @param chord_env The chordal environment containing the irg
- * @param do_copy 1 == resolve cycles with a free reg if available
- */
-void lower_nodes_before_sched(ir_graph *irg, const void *env) {
- irg_walk_blkwise_graph(irg, NULL, lower_nodes_before_sched_walker, env);
-}
-
-
-
/**
* Walks over all blocks in an irg and performs lowering need to be
- * done after register allocation (e.g. perm and spill/reload lowering).
+ * 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
env.chord_env = chord_env;
env.do_copy = do_copy;
- env.dbg_module = firm_dbg_register("ir.be.lower");
+ FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);
}
#undef is_Perm
-#undef is_Call