Use symbolic names instead of magic values for the position parameter of get_irn_n().
[libfirm] / ir / be / belower.c
index a02a73c..3b2b47a 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
 #include "besched_t.h"
 #include "bestat.h"
 #include "bessaconstr.h"
-#include "benodesets.h"
+#include "beintlive_t.h"
 
 #undef KEEP_ALIVE_COPYKEEP_HACK
 
-/* associates op with it's copy and CopyKeep */
+/** Associates an op with it's copy and CopyKeep. */
 typedef struct {
-       ir_node *op;         /* an irn which must be different */
-       ir_nodeset_t copies; /* all non-spillable copies of this irn */
+       ir_node *op;         /**< an irn which must be different */
+       ir_nodeset_t copies; /**< all non-spillable copies of this irn */
        const arch_register_class_t *cls;
 } op_copy_assoc_t;
 
-/* environment for constraints */
+/** Environment for constraints. */
 typedef struct {
        be_irg_t       *birg;
        pset           *op_set;
@@ -64,7 +64,7 @@ typedef struct {
        DEBUG_ONLY(firm_dbg_module_t *dbg;)
 } constraint_env_t;
 
-/* lowering walker environment */
+/** Lowering walker environment. */
 typedef struct _lower_env_t {
        be_irg_t         *birg;
        const arch_env_t *arch_env;
@@ -72,7 +72,7 @@ typedef struct _lower_env_t {
        DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
 } lower_env_t;
 
-/* holds a perm register pair */
+/** Holds a Perm register pair. */
 typedef struct _reg_pair_t {
        const arch_register_t *in_reg;    /**< a perm IN register */
        ir_node               *in_node;   /**< the in node to which the register belongs */
@@ -90,15 +90,14 @@ typedef enum _perm_type_t {
        PERM_COPY
 } perm_type_t;
 
-/* structure to represent cycles or chains in a perm */
+/** 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;
 
-//
-/* Compare the two operands */
+/** 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;
@@ -106,7 +105,7 @@ static int cmp_op_copy_assoc(const void *a, const void *b) {
        return op1->op != op2->op;
 }
 
-/* Compare the in registers of two register pairs */
+/** 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;
        const reg_pair_t *pair_b = b;
@@ -117,7 +116,7 @@ static int compare_reg_pair(const void *a, const void *b) {
                return -1;
 }
 
-/* returns the number register pairs marked as checked */
+/** 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;
 
@@ -177,14 +176,14 @@ static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_
        if (in_out) {
                for (i = 0; i < n; i++) {
                        /* out register matches */
-                       if (pairs[i].out_reg->index == reg_idx)
+                       if ((int) 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)
+                       if ((int) pairs[i].in_reg->index == reg_idx)
                                return i;
                }
        }
@@ -308,6 +307,7 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                        should be ok.
        */
        sched_point = sched_prev(irn);
+       DBG((mod, LEVEL_1, "perm: %+F\n", irn));
        DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
        assert(sched_point && "Perm is not scheduled or has no predecessor");
 
@@ -348,9 +348,6 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
                        }
 
-                       /* remove the proj from the schedule */
-                       sched_remove(pairs[i].out_node);
-
                        /* reroute the edges from the proj to the argument */
                        exchange(pairs[i].out_node, pairs[i].in_node);
                        //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
@@ -462,20 +459,12 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                        /* 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]);
 
@@ -495,9 +484,6 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                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);
 
@@ -518,23 +504,19 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
        /* remove the perm from schedule */
        if (! keep_perm) {
                sched_remove(irn);
-               be_kill_node(irn);
+               kill_node(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 int has_irn_users(const ir_node *irn) {
+       return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
 }
 
+/**
+ * Skip all Proj nodes.
+ */
 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
        while(is_Proj(irn))
                irn = get_Proj_pred(irn);
@@ -578,7 +560,7 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
        /* 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 */
+       /* check if already exists such a copy in the schedule immediately before */
        cpy = find_copy(env, belower_skip_proj(irn), other_different);
        if (! cpy) {
                cpy = be_new_Copy(cls, irg, block, other_different);
@@ -594,13 +576,13 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
 
        /* 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, irg, block, 2, in);
-       }
-       else {
+       if (has_irn_users(other_different)) {
                keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
                be_node_set_reg_class(keep, 1, cls);
        }
+       else {
+               keep = be_new_Keep(cls, irg, block, 2, in);
+       }
 
        DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
 
@@ -611,8 +593,8 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
        sched_add_after(irn, keep);
 
        /* insert the other different and it's copies into the set */
-       key.op         = other_different;
-       entry          = pset_find(op_set, &key, nodeset_hash(other_different));
+       key.op = other_different;
+       entry  = pset_find(op_set, &key, hash_irn(other_different));
 
        if (! entry) {
                entry         = obstack_alloc(&env->obst, sizeof(*entry));
@@ -629,12 +611,12 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
                ir_nodeset_insert(&entry->copies, keep);
        }
 
-       pset_insert(op_set, entry, nodeset_hash(other_different));
+       pset_insert(op_set, entry, hash_irn(other_different));
 }
 
 /**
- * Checks if node has a should_be_different constraint in output
- * and adds a Keep then to assure the constraint.
+ * Checks if node has a must_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;
@@ -642,19 +624,35 @@ static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
 
        req = arch_get_register_req(arch_env, irn, -1);
 
-       if (arch_register_req_is(req, should_be_different)) {
-               ir_node *different_from = get_irn_n(belower_skip_proj(irn), req->other_different);
-               gen_assure_different_pattern(irn, different_from, 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);
+       if (arch_register_req_is(req, must_be_different)) {
+               const unsigned other = req->other_different;
+               int i;
+
+               if (arch_register_req_is(req, should_be_same)) {
+                       const unsigned same = req->other_same;
+
+                       if (is_po2(other) && is_po2(same)) {
+                               int idx_other = ntz(other);
+                               int idx_same  = ntz(same);
+
+                               /*
+                                * We can safely ignore a should_be_same x must_be_different y
+                                * IFF both inputs are equal!
+                                */
+                               if (get_irn_n(irn, idx_other) == get_irn_n(irn, idx_same)) {
+                                       return;
+                               }
+                       }
+               }
+               for (i = 0; 1U << i <= other; ++i) {
+                       if (other & (1U << i)) {
+                               ir_node *different_from = get_irn_n(belower_skip_proj(irn), i);
+                               gen_assure_different_pattern(irn, different_from, env);
+                       }
                }
        }
 }
 
-
-
 /**
  * Calls the functions to assure register constraints.
  *
@@ -667,8 +665,6 @@ static void assure_constraints_walker(ir_node *irn, void *walk_env) {
 
        if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
                assure_different_constraints(irn, walk_env);
-
-       return;
 }
 
 /**
@@ -756,7 +752,7 @@ static void melt_copykeeps(constraint_env_t *cenv) {
                                        /* now, we can kill the melted keep, except the */
                                        /* ref one, we still need some information      */
                                        if (melt_arr[j] != ref)
-                                               be_kill_node(melt_arr[j]);
+                                               kill_node(melt_arr[j]);
                                }
 
 #ifdef KEEP_ALIVE_COPYKEEP_HACK
@@ -766,27 +762,24 @@ static void melt_copykeeps(constraint_env_t *cenv) {
                                new_ck = be_new_CopyKeep(entry->cls, 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 */
+                               /* set register class for all kept inputs */
                                for (j = 1; j <= n_melt; ++j)
                                        be_node_set_reg_class(new_ck, j, entry->cls);
 
                                ir_nodeset_insert(&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_pt = ref_mode_T;
+                               do {
+                                       /* just walk along the schedule until a non-Keep/CopyKeep node is found */
+                                       sched_pt = sched_next(sched_pt);
+                               } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
 
                                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 */
-                               be_kill_node(ref);
+                               kill_node(ref);
                        }
                }
 
@@ -858,12 +851,12 @@ void assure_constraints(be_irg_t *birg) {
                                int     n = get_irn_arity(cp);
 
                                keep = be_new_Keep(arch_get_irn_reg_class(arch_env, cp, -1),
-                                       irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
+                                       irg, get_nodes_block(cp), n, 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. */
                                sched_remove(cp);
-                               be_kill_node(cp);
+                               kill_node(cp);
                        }
                }
 
@@ -872,10 +865,169 @@ void assure_constraints(be_irg_t *birg) {
 
        del_pset(cenv.op_set);
        obstack_free(&cenv.obst, NULL);
-       be_invalidate_liveness(birg);
+       be_liveness_invalidate(be_get_birg_liveness(birg));
 }
 
 
+/**
+ * Push nodes that do not need to be permed through the Perm.
+ * This is commonly a reload cascade at block ends.
+ * @note This routine needs interference.
+ * @note Probably, we can implement it a little more efficient.
+ *       Especially searching the frontier lazily might be better.
+ * @param perm The perm.
+ * @param data The walker data (lower_env_t).
+ * @return     1, if there is something left to perm over.
+ *             0, if removed the complete perm.
+ */
+static int push_through_perm(ir_node *perm, void *data)
+{
+       lower_env_t *env       = data;
+       const arch_env_t *aenv = env->arch_env;
+
+       ir_graph *irg     = get_irn_irg(perm);
+       ir_node *bl       = get_nodes_block(perm);
+       ir_node *node;
+       int  arity        = get_irn_arity(perm);
+       int *map;
+       int *proj_map;
+       bitset_t *moved   = bitset_alloca(arity);
+       int n_moved;
+       int new_size;
+       ir_node *frontier = bl;
+       FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
+
+       int i, n;
+       const ir_edge_t *edge;
+       ir_node *one_proj = NULL, *irn;
+       const arch_register_class_t *cls = NULL;
+
+       DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
+
+       /* get some Proj and find out the register class of that Proj. */
+       edge     = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
+       one_proj = get_edge_src_irn(edge);
+       assert(is_Proj(one_proj));
+       cls      = arch_get_irn_reg_class(aenv, one_proj, -1);
+
+       /* Find the point in the schedule after which the
+        * potentially movable nodes must be defined.
+        * A Perm will only be pushed up to first instruction
+        * which lets an operand of itself die.
+        * If we would allow to move the Perm above this instruction,
+        * the former dead operand would be live now at the point of
+        * the Perm, increasing the register pressure by one.
+        */
+       sched_foreach_reverse_from (sched_prev(perm), irn) {
+               for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
+                       ir_node *op = get_irn_n(irn, i);
+                       if (arch_irn_consider_in_reg_alloc(aenv, cls, op) &&
+                           !values_interfere(env->birg, op, one_proj)) {
+                               frontier = irn;
+                               goto found_front;
+                       }
+               }
+       }
+found_front:
+
+       DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
+
+       node = sched_prev(perm);
+       n_moved = 0;
+       while(!sched_is_begin(node)) {
+               const arch_register_req_t *req;
+               int                        input = -1;
+               ir_node                   *proj;
+
+               /* search if node is a INPUT of Perm */
+               foreach_out_edge(perm, edge) {
+                       ir_node *out = get_edge_src_irn(edge);
+                       int      pn  = get_Proj_proj(out);
+                       ir_node *in  = get_irn_n(perm, pn);
+                       if (node == in) {
+                               proj  = out;
+                               input = pn;
+                               break;
+                       }
+               }
+               /* it wasn't an input to the perm, we can't do anything more */
+               if(input < 0)
+                       break;
+               if(!sched_comes_after(frontier, node))
+                       break;
+               if(arch_irn_is(aenv, node, modify_flags))
+                       break;
+               if(is_Proj(node)) {
+                       req = arch_get_register_req(aenv, get_Proj_pred(node),
+                                                   -1 - get_Proj_proj(node));
+               } else {
+                       req = arch_get_register_req(aenv, node, -1);
+               }
+               if(req->type != arch_register_req_type_normal)
+                       break;
+               for(i = get_irn_arity(node) - 1; i >= 0; --i) {
+                       ir_node *opop = get_irn_n(node, i);
+                       if (arch_irn_consider_in_reg_alloc(aenv, cls, opop)) {
+                               break;
+                       }
+               }
+               if(i >= 0)
+                       break;
+
+               DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
+
+               /* move the movable node in front of the Perm */
+               sched_remove(node);
+               sched_add_after(perm, node);
+
+               /* give it the proj's register */
+               arch_set_irn_register(aenv, node, arch_get_irn_register(aenv, proj));
+
+               /* reroute all users of the proj to the moved node. */
+               edges_reroute(proj, node, irg);
+
+               /* and kill it */
+               set_Proj_pred(proj, new_Bad());
+               kill_node(proj);
+
+               bitset_set(moved, input);
+               n_moved++;
+
+               node = sched_prev(node);
+       }
+
+       /* well, we could not push anything through the perm */
+       if(n_moved == 0)
+               return 1;
+
+       new_size = arity - n_moved;
+       if(new_size == 0) {
+               return 0;
+       }
+
+       map      = alloca(new_size * sizeof(map[0]));
+       proj_map = alloca(arity * sizeof(proj_map[0]));
+       memset(proj_map, -1, sizeof(proj_map[0]));
+       n   = 0;
+       for(i = 0; i < arity; ++i) {
+               if(bitset_is_set(moved, i))
+                       continue;
+               map[n]      = i;
+               proj_map[i] = n;
+               n++;
+       }
+       assert(n == new_size);
+       foreach_out_edge(perm, edge) {
+               ir_node *proj = get_edge_src_irn(edge);
+               int      pn   = get_Proj_proj(proj);
+               pn = proj_map[pn];
+               assert(pn >= 0);
+               set_Proj_proj(proj, pn);
+       }
+
+       be_Perm_reduce(perm, new_size, map);
+       return 1;
+}
 
 /**
  * Calls the corresponding lowering function for the node.
@@ -884,13 +1036,18 @@ void assure_constraints(be_irg_t *birg) {
  * @param walk_env The walker environment
  */
 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
-       if (! is_Block(irn) && ! is_Proj(irn)) {
-               if (be_is_Perm(irn)) {
-                       lower_perm_node(irn, walk_env);
-               }
-       }
+       int perm_stayed;
 
-       return;
+       if (is_Block(irn) || is_Proj(irn))
+               return;
+       if (!be_is_Perm(irn))
+               return;
+
+       perm_stayed = push_through_perm(irn, walk_env);
+       if (!perm_stayed)
+               return;
+
+       lower_perm_node(irn, walk_env);
 }
 
 /**
@@ -909,5 +1066,8 @@ void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
        env.do_copy  = do_copy;
        FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
 
+       /* we will need interference */
+       be_liveness_assure_chk(be_get_birg_liveness(birg));
+
        irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);
 }