Handle Cmp(And(x,y) != 0) ==> Test(x,y)
[libfirm] / ir / be / bechordal.c
index 750cfd4..55f2c19 100644 (file)
@@ -62,8 +62,6 @@ typedef struct _be_chordal_alloc_env_t {
        pset *pre_colored;    /**< Set of precolored nodes. */
        bitset_t *live;                         /**< A liveness bitset. */
        bitset_t *colors;                       /**< The color mask. */
-       bitset_t *valid_colors;     /**< A mask of colors which shall be considered during allocation.
-                                                                    Registers with the ignore bit on, must not be considered. */
        bitset_t *in_colors;        /**< Colors used by live in values. */
        int colors_n;               /**< The number of colors. */
 } be_chordal_alloc_env_t;
@@ -162,7 +160,8 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head
  */
 static INLINE int has_reg_class(const be_chordal_env_t *env, const ir_node *irn)
 {
-  return arch_irn_has_reg_class(env->main_env->arch_env, irn, -1, env->cls);
+       // return arch_irn_has_reg_class(env->main_env->arch_env, irn, -1, env->cls);
+       return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn);
 }
 
 #define has_limited_constr(req, irn) \
@@ -183,15 +182,19 @@ typedef struct {
        int n_ops;
        int use_start;
        ir_node *next_insn;
+       unsigned in_constraints  : 1;
+       unsigned out_constraints : 1;
        unsigned has_constraints : 1;
+       unsigned pre_colored     : 1;
 } insn_t;
 
 static insn_t *scan_insn(be_chordal_env_t *env, ir_node *irn, struct obstack *obst)
 {
-       const arch_env_t *arch_env = env->main_env->arch_env;
+       const arch_env_t *arch_env = env->birg->main_env->arch_env;
        operand_t o;
        insn_t *insn;
        int i, n;
+       int pre_colored = 0;
 
        insn = obstack_alloc(obst, sizeof(insn[0]));
        memset(insn, 0, sizeof(insn[0]));
@@ -209,7 +212,8 @@ static insn_t *scan_insn(be_chordal_env_t *env, ir_node *irn, struct obstack *ob
                                arch_get_register_req(arch_env, &o.req, p, -1);
                                obstack_grow(obst, &o, sizeof(o));
                                insn->n_ops++;
-                               insn->has_constraints |= arch_register_req_is(&o.req, limited);
+                               insn->out_constraints |= arch_register_req_is(&o.req, limited);
+                               pre_colored += arch_get_irn_register(arch_env, p) != NULL;
                        }
                }
 
@@ -224,10 +228,12 @@ static insn_t *scan_insn(be_chordal_env_t *env, ir_node *irn, struct obstack *ob
                arch_get_register_req(arch_env, &o.req, irn, -1);
                obstack_grow(obst, &o, sizeof(o));
                insn->n_ops++;
-               insn->has_constraints |= arch_register_req_is(&o.req, limited);
+               insn->out_constraints |= arch_register_req_is(&o.req, limited);
+               pre_colored += arch_get_irn_register(arch_env, irn) != NULL;
        }
 
-       insn->use_start = insn->n_ops;
+       insn->pre_colored = pre_colored == insn->n_ops;
+       insn->use_start   = insn->n_ops;
 
        for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
                ir_node *op = get_irn_n(irn, i);
@@ -240,14 +246,16 @@ static insn_t *scan_insn(be_chordal_env_t *env, ir_node *irn, struct obstack *ob
                        arch_get_register_req(arch_env, &o.req, irn, i);
                        obstack_grow(obst, &o, sizeof(o));
                        insn->n_ops++;
-                       insn->has_constraints |= arch_register_req_is(&o.req, limited);
+                       insn->in_constraints |= arch_register_req_is(&o.req, limited);
                }
        }
 
+       insn->has_constraints = insn->in_constraints | insn->out_constraints;
        insn->ops = obstack_finish(obst);
        return insn;
 }
 
+#if 0
 static operand_t *find_unpaired_use(insn_t *insn, const operand_t *op, int can_be_constrained)
 {
        int i;
@@ -257,11 +265,14 @@ static operand_t *find_unpaired_use(insn_t *insn, const operand_t *op, int can_b
                operand_t *op = &insn->ops[i];
                int has_constraint = arch_register_req_is(&op->req, limited);
 
-               if(!values_interfere(op->carrier, op->irn) && !op->partner && (!has_constraint || can_be_constrained)) {
-                       if(arch_register_req_is(&op->req, should_be_same) && op->req.other_same == op->carrier)
-                               return op;
-                       else
-                               res = op;
+               if(!values_interfere(op->carrier, op->irn) && !op->partner) {
+
+                       if(!has_constraint || can_be_constrained) {
+                               if(arch_register_req_is(&op->req, should_be_same) && op->req.other_same == op->carrier)
+                                       return op;
+                               else
+                                       res = op;
+                       }
                }
        }
 
@@ -284,6 +295,77 @@ static void pair_up_operands(insn_t *insn)
                }
        }
 }
+#endif
+
+static void add_possible_partners(insn_t *insn, int curr, const arch_register_t *out_reg, bipartite_t *bp, int can_be_constrained)
+{
+       bitset_t *bs;
+       int i;
+
+       if(out_reg)
+               bs = bitset_alloca(out_reg->reg_class->n_regs);
+
+       for(i = insn->use_start; i < insn->n_ops; ++i) {
+               const operand_t *op = &insn->ops[i];
+
+               /*
+                       The in operand can only be paired with a def, if the node defining the
+                       operand's value does not interfere with the instruction itself. That
+                       would mean, that it is live at the instruction, so no result of the instruction
+                       can have the same register as the operand.
+
+                       Furthermore, if the operand has already been paired (due to previous calls)
+                       to this function, we do not touch this partnership.
+               */
+               if(!values_interfere(op->irn, op->carrier)) {
+                       int has_constraint = arch_register_req_is(&op->req, limited);
+
+                       if(has_constraint && out_reg && out_reg->reg_class == op->req.cls) {
+                               bitset_clear_all(bs);
+                               op->req.limited(op->req.limited_env, bs);
+                               if(bitset_is_set(bs, out_reg->index)) {
+                                       bipartite_add(bp, curr, i - insn->use_start);
+                               }
+                       }
+
+                       if(!has_constraint || can_be_constrained) {
+                               bipartite_add(bp, curr, i - insn->use_start);
+#if 0
+                               if(arch_register_req_is(&op->req, should_be_same) && op->req.other_same == op->carrier)
+                                       return;
+#endif
+                       }
+               }
+       }
+}
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+
+static void pair_up_operands(const arch_env_t *arch_env, insn_t *insn) {
+       int i;
+       int m = MAX(insn->n_ops - insn->use_start, 1);
+       bipartite_t *bp = bipartite_new(MAX(insn->use_start, 1), m);
+       int *match      = alloca(insn->use_start * sizeof(match[0]));
+
+       for(i = 0; i < insn->use_start; ++i) {
+               operand_t *op = &insn->ops[i];
+               const arch_register_t *reg = arch_get_irn_register(arch_env, op->carrier);
+               int has_constraint         = arch_register_req_is(&op->req, limited);
+               add_possible_partners(insn, i, reg, bp, !has_constraint);
+       }
+
+       bipartite_matching(bp, match);
+       for(i = 0; i < insn->use_start; ++i) {
+               int p = match[i] + insn->use_start;
+
+               if(p >= insn->use_start) {
+                       insn->ops[i].partner = &insn->ops[p];
+                       insn->ops[p].partner = &insn->ops[i];
+               }
+       }
+
+       bipartite_free(bp);
+}
 
 static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *irn)
 {
@@ -292,11 +374,26 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
        insn_t *insn           = scan_insn(env, irn, &env->obst);
        ir_node *res           = insn->next_insn;
 
+
+       if(insn->pre_colored) {
+               int i;
+               for(i = 0; i < insn->use_start; ++i)
+                       pset_insert_ptr(alloc_env->pre_colored, insn->ops[i].carrier);
+       }
+
+       if(be_is_Perm(irn) || be_is_RegParams(irn) || (be_is_Barrier(irn) && !insn->in_constraints))
+               goto end;
+
+       /*
+               Perms inserted before the constraint handling phase are considered to be
+               correctly precolored. These Perms arise during the ABI handling phase.
+       */
        if(insn->has_constraints) {
                firm_dbg_module_t *dbg = firm_dbg_register("firm.be.chordal.constr");
-               const arch_env_t *aenv = env->main_env->arch_env;
+               const arch_env_t *aenv = env->birg->main_env->arch_env;
                int n_regs             = env->cls->n_regs;
                bitset_t *bs           = bitset_alloca(n_regs);
+               bitset_t *non_ignore   = bitset_alloca(n_regs);
                ir_node **alloc_nodes  = alloca(n_regs * sizeof(alloc_nodes[0]));
                bipartite_t *bp        = bipartite_new(n_regs, n_regs);
                int *assignment        = alloca(n_regs * sizeof(assignment[0]));
@@ -305,7 +402,12 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                int i, n_alloc;
                long col;
                const ir_edge_t *edge;
-               ir_node *perm = insert_Perm_after(aenv, env->cls, env->dom_front, sched_prev(irn));
+               ir_node *perm = NULL;
+
+//             if(!insn->pre_colored || insn->in_constraints)
+               perm = insert_Perm_after(aenv, env->cls, env->dom_front, sched_prev(irn));
+
+               arch_put_non_ignore_regs(aenv, env->cls, non_ignore);
 
                /* Registers are propagated by insert_Perm_after(). Clean them here! */
                if(perm) {
@@ -324,22 +426,25 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                /*
                 * If there was no Perm made, nothing was alive in this register class.
                 * This means, that the node has no operands, thus no input constraints.
-                * so it had output constraints. The other results then can be assigned freeliy.
+                * so it had output constraints. The other results then can be assigned freely.
                 */
 
-               pair_up_operands(insn);
+               pair_up_operands(aenv, insn);
 
                for(i = 0, n_alloc = 0; i < insn->n_ops; ++i) {
                        operand_t *op = &insn->ops[i];
-                       if(arch_register_req_is(&op->req, limited)) {
+                       if((op->partner ? !pmap_contains(partners, op->partner->carrier) : 1)
+                               && arch_register_req_is(&op->req, limited)
+                               && arch_irn_consider_in_reg_alloc(aenv, env->cls, op->carrier)) {
+
                                pmap_insert(partners, op->carrier, op->partner ? op->partner->carrier : NULL);
                                alloc_nodes[n_alloc] = op->carrier;
 
                                DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, pmap_get(partners, op->carrier)));
 
-                               bitset_clear_all(bs);
+                               bitset_clear_all(bs);
                                op->req.limited(op->req.limited_env, bs);
-                               bitset_and(bs, alloc_env->valid_colors);
+                               bitset_and(bs, non_ignore);
 
                                DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier, bs));
 
@@ -351,6 +456,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                }
 
                if(perm) {
+
                        foreach_out_edge(perm, edge) {
                                ir_node *proj = get_edge_src_irn(edge);
 
@@ -363,7 +469,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
 
                                        bitset_clear_all(bs);
                                        arch_get_allocatable_regs(aenv, proj, -1, bs);
-                                       bitset_and(bs, alloc_env->valid_colors);
+                                       bitset_and(bs, non_ignore);
                                        bitset_foreach(bs, col)
                                                bipartite_add(bp, n_alloc, col);
 
@@ -374,7 +480,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
 
                bipartite_matching(bp, assignment);
 
-
+               /* Assign colors obtained from the matching. */
                for(i = 0; i < n_alloc; ++i) {
                        int j;
                        ir_node *nodes[2];
@@ -397,8 +503,23 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                }
 
 
+               /* Allocate the non-constrained Projs of the Perm. */
                if(perm) {
+                       /* Make the input constraints of the node to output constraints of the Perm's Projs */
+                       for(i = insn->use_start; i < insn->n_ops; ++i) {
+                               operand_t *op = &insn->ops[i];
+
+                               /*
+                                       If the operand is an "in" operand, constrained and the carrier is a Proj to the Perm,
+                                       then copy the in constraint to the Perm's out constraint
+                               */
+                               if(arch_register_req_is(&op->req, limited) && is_Proj(op->carrier) && perm == get_Proj_pred(op->carrier))
+                                       be_set_constr_limited(perm, BE_OUT_POS(get_Proj_proj(op->carrier)), &op->req);
+                       }
+
                        bitset_clear_all(bs);
+
+                       /* Put the colors of all Projs in a bitset. */
                        foreach_out_edge(perm, edge) {
                                ir_node *proj              = get_edge_src_irn(edge);
                                const arch_register_t *reg = arch_get_irn_register(aenv, proj);
@@ -407,7 +528,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                                        bitset_set(bs, reg->index);
                        }
 
-                       // bitset_or(bs, alloc_env->ignore_colors);
+                       /* Assign the not yet assigned Projs of the Perm a suitable color. */
                        foreach_out_edge(perm, edge) {
                                ir_node *proj              = get_edge_src_irn(edge);
                                const arch_register_t *reg = arch_get_irn_register(aenv, proj);
@@ -428,6 +549,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                pmap_destroy(partners);
        }
 
+end:
        obstack_free(&env->obst, base);
        return res;
 }
@@ -443,7 +565,7 @@ static void constraints(ir_node *bl, void *data)
 {
        firm_dbg_module_t *dbg      = firm_dbg_register("firm.be.chordal.constr");
        be_chordal_alloc_env_t *env = data;
-       arch_env_t *arch_env        = env->chordal_env->main_env->arch_env;
+       arch_env_t *arch_env        = env->chordal_env->birg->main_env->arch_env;
        ir_node *irn;
 
        for(irn = sched_first(bl); !sched_is_end(irn);) {
@@ -469,6 +591,7 @@ static void pressure(ir_node *block, void *env_ptr)
 
        be_chordal_alloc_env_t *alloc_env = env_ptr;
        be_chordal_env_t *env             = alloc_env->chordal_env;
+       const arch_env_t *arch_env        = env->birg->main_env->arch_env;
        bitset_t *live                    = alloc_env->live;
        firm_dbg_module_t *dbg            = env->dbg;
        ir_node *irn;
@@ -570,19 +693,17 @@ static void assign(ir_node *block, void *env_ptr)
        bitset_t *live              = alloc_env->live;
        bitset_t *colors            = alloc_env->colors;
        bitset_t *in_colors         = alloc_env->in_colors;
-       const arch_env_t *arch_env  = env->main_env->arch_env;
+       const arch_env_t *arch_env  = env->birg->main_env->arch_env;
 
        const ir_node *irn;
        border_t *b;
        struct list_head *head = get_block_border_head(env, block);
        pset *live_in = put_live_in(block, pset_new_ptr_default());
 
+       bitset_clear_all(colors);
        bitset_clear_all(live);
        bitset_clear_all(in_colors);
 
-       bitset_copy(colors, alloc_env->valid_colors);
-       bitset_flip_all(colors);
-
        DBG((dbg, LEVEL_4, "Assigning colors for block %+F\n", block));
        DBG((dbg, LEVEL_4, "\tusedef chain for block\n"));
        list_for_each_entry(border_t, b, head, list) {
@@ -613,7 +734,8 @@ static void assign(ir_node *block, void *env_ptr)
        }
 
        /*
-        * Mind that the sequence of defs from back to front defines a perfect
+        * Mind that the sequence
+        * of defs from back to front defines a perfect
         * elimination order. So, coloring the definitions from first to last
         * will work.
         */
@@ -642,6 +764,8 @@ static void assign(ir_node *block, void *env_ptr)
                        }
 
                        bitset_set(colors, col);
+
+                       assert(!arch_register_type_is(reg, ignore) && "Must not assign ignore register");
                        arch_set_irn_register(arch_env, irn, reg);
 
                        DBG((dbg, LEVEL_1, "\tassigning register %s(%d) to %+F\n",
@@ -684,18 +808,15 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
        env.chordal_env   = chordal_env;
        env.colors_n      = colors_n;
        env.colors        = bitset_malloc(colors_n);
-       env.valid_colors  = bitset_malloc(colors_n);
        env.in_colors     = bitset_malloc(colors_n);
        env.pre_colored   = pset_new_ptr_default();
 
-       arch_put_non_ignore_regs(chordal_env->main_env->arch_env, chordal_env->cls, env.valid_colors);
-
        /* Handle register targeting constraints */
        dom_tree_walk_irg(irg, constraints, NULL, &env);
 
        if(chordal_env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
                snprintf(buf, sizeof(buf), "-%s-constr", chordal_env->cls->name);
-               dump_ir_block_graph_sched(chordal_env->irg, buf);
+               be_dump(chordal_env->irg, buf, dump_ir_block_graph_sched);
        }
 
        be_numbering(irg);
@@ -711,7 +832,6 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
 
        if(chordal_env->opts->dump_flags & BE_CH_DUMP_TREE_INTV) {
        plotter_t *plotter;
-
                ir_snprintf(buf, sizeof(buf), "ifg_%s_%F.eps", chordal_env->cls->name, irg);
        plotter = new_plotter_ps(buf);
        draw_interval_tree(&draw_chordal_def_opts, chordal_env, plotter);
@@ -721,7 +841,5 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
        free(env.live);
        free(env.colors);
        free(env.in_colors);
-       free(env.valid_colors);
-
        del_pset(env.pre_colored);
 }