Set dump consts local to false for schedule dumping
[libfirm] / ir / be / bespillilp.c
index 07e2a4f..0a33479 100644 (file)
@@ -39,7 +39,7 @@
 
 #define MAX(a,b) ((a) > (b) ? (a) : (b))
 
-#define DBG_LEVEL SET_LEVEL_4
+#define DBG_LEVEL SET_LEVEL_3
 
 #define DUMP_SOLUTION
 #define DUMP_ILP
@@ -98,13 +98,13 @@ typedef struct _irn_use_head_t {
 } irn_use_head_t;
 
 struct _live_range_t {
-  struct list_head list;
+       struct list_head list;
        irn_use_head_t *use_head;
-  ir_node *user;
-  ir_node *irn;
+       ir_node *user;
+       ir_node *irn;
        int pos;
-  int in_mem_var;
-  int is_remat_var;
+       int in_mem_var;
+       int is_remat_var;
 };
 
 /*
@@ -119,7 +119,7 @@ typedef struct _first_use_t {
 
 
 
-static int has_reg_class(const spill_ilp_t *si, const ir_node *irn)
+static INLINE int has_reg_class(const spill_ilp_t *si, const ir_node *irn)
 {
   return arch_irn_has_reg_class(si->session->main_env->arch_env,
       irn, arch_pos_make_out(0), si->cls);
@@ -141,6 +141,13 @@ static int cmp_irn_use_head(const void *a, const void *b, size_t n)
        return !(p->irn == q->irn);
 }
 
+static irn_use_head_t *get_use_head(spill_ilp_t *si, const ir_node *irn)
+{
+       irn_use_head_t templ;
+       templ.irn = (ir_node *) irn;
+       return set_find(si->irn_use_heads, &templ, sizeof(templ), HASH_PTR(irn));
+}
+
 static int cmp_first_use(const void *a, const void *b, size_t n)
 {
   const first_use_t *p = a;
@@ -197,26 +204,26 @@ static INLINE int can_remat(const spill_ilp_t *si, const ir_node *irn, pset *liv
 
 static live_range_t *get_live_range(spill_ilp_t *si, ir_node *irn, ir_node *user, int pos)
 {
-  live_range_t lr, *res;
+       live_range_t lr, *res;
        irn_use_head_t iuh, *head;
        int is_new;
-  unsigned hash = HASH_COMBINE(HASH_PTR(irn), HASH_PTR(user));
+       unsigned hash = HASH_COMBINE(HASH_PTR(irn), HASH_PTR(user));
 
-  lr.user    = user;
-  lr.irn     = irn;
-  lr.pos     = pos;
-  lr.in_mem_var = -1;
+       lr.user         = user;
+       lr.irn          = irn;
+       lr.pos          = pos;
+       lr.in_mem_var   = -1;
        lr.is_remat_var = -1;
 
-  res = set_insert(si->live_ranges, &lr, sizeof(lr), hash);
+       res = set_insert(si->live_ranges, &lr, sizeof(lr), hash);
        is_new = res->in_mem_var == -1;
 
-  if(is_new) {
-    char buf[128];
-    ir_snprintf(buf, sizeof(buf), "m_%s%N_%N_%d",
-                               is_Phi(irn) ? "phi_" : "", irn, user, MAX(pos, 0));
-    res->in_mem_var = lpp_add_var(si->lpp, buf, lpp_binary, pos >= 0 ? COST_LOAD : 0.0);
-  }
+       if(is_new) {
+               char buf[128];
+               ir_snprintf(buf, sizeof(buf), "m_%s%N_%N_%d",
+                                       is_Phi(irn) ? "phi_" : "", irn, user, MAX(pos, 0));
+               res->in_mem_var = lpp_add_var(si->lpp, buf, lpp_binary, pos >= 0 ? COST_LOAD : 0.0);
+       }
 
        memset(&iuh, 0, sizeof(iuh));
        iuh.irn = irn;
@@ -234,121 +241,136 @@ static live_range_t *get_live_range(spill_ilp_t *si, ir_node *irn, ir_node *user
 
        res->use_head = head;
 
-  return res;
+       return res;
 }
 
-static ir_node *process_irn(spill_ilp_t *si, pset *live, ir_node *irn, int *demand)
-{
-       int i, n;
-       int relevant_args = 0, results = 0;
-
-       DBG((si->dbg, LEVEL_1, "at %+F\n", irn));
-
-       while(is_Proj(irn)) {
-               if(has_reg_class(si, irn)) {
-                       assert(pset_find_ptr(live, irn) && "node must be live");
-                       pset_remove_ptr(live, irn);
-                       results++;
-               }
-
-               DBG((si->dbg, LEVEL_1, "skipped proj %+F\n", irn));
-               irn = sched_prev(irn);
-       }
-
-       DBG((si->dbg, LEVEL_1, "\tlanded at irn %+F\n", irn));
-
-       if(results > 0)
-               assert(get_irn_mode(irn) == mode_T && "node before projs must be tuple");
-
-       if(has_reg_class(si, irn)) {
-               assert( pset_find_ptr(live, irn) && "node must be live");
-               pset_remove_ptr(live, irn);
-               results = 1;
-       }
-
-       for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
-               ir_node *op = get_irn_n(irn, i);
-               if(has_reg_class(si, op) && !pset_find_ptr(live, op)) {
-                       relevant_args++;
-                       DBG((si->dbg, LEVEL_1, "\trelevant arg %+F\n", op));
-               }
-       }
-
-       *demand = MAX(results, relevant_args);
-       DBG((si->dbg, LEVEL_1, "\tdemand: %d\n", *demand));
-       return irn;
+static void print_live_set(spill_ilp_t *si, pset *s) {
+       ir_node *n;
+       for(n=pset_first(s); n; n=pset_next(s))
+               DBG((si->dbg, LEVEL_3, "    %+F\n", n));
 }
 
 static void process_block(ir_node *bl, void *data)
 {
        char buf[128];
-       int i, n;
-  spill_ilp_t *si  = data;
-  int step         = 0;
-  int n_regs       = arch_register_class_n_regs(si->cls);
+       int i, n, skipped=0;
+       spill_ilp_t *si  = data;
+       int step         = 0;
+       int n_regs       = arch_register_class_n_regs(si->cls);
        int n_preds      = get_irn_arity(bl);
-  pset *live       = pset_new_ptr_default();
-  irn_live_t *li;
-  ir_node *irn;
-
-  /* as always, bring the live end nodes to life here */
-  live_foreach(bl, li) {
-    if(live_is_end(li) && has_reg_class(si, li->irn)) {
-      ir_node *irn = (ir_node *) li->irn;
-      pset_insert_ptr(live, irn);
-
-      /*
-       * The "user" of the live range to the end of a block
-       * is the block itself. This is quite arbitrary.
-       */
-      set_irn_link(irn, get_live_range(si, irn, bl, -1));
-    }
-  }
+       pset *live       = pset_new_ptr_default();
+       irn_live_t *li;
+       ir_node *irn;
+
+       DBG((si->dbg, LEVEL_3, "\n"));
+       DBG((si->dbg, LEVEL_3, "Processing %+F\n", bl));
+
+       /*
+        * Get all live-end values of this block
+        */
+       live_foreach(bl, li) {
+               if(live_is_end(li) && has_reg_class(si, li->irn)) {
+                       ir_node *irn = (ir_node *) li->irn;
+                       pset_insert_ptr(live, irn);
+
+                       /*The "user" of the live range to the end of a block
+                        * is the block itself. This is quite arbitrary. */
+                       set_irn_link(irn, get_live_range(si, irn, bl, -1));
+               }
+       }
+       DBG((si->dbg, LEVEL_3, "Live-End:\n"));
+       print_live_set(si, live);
 
+       /*
+        * Walk through the schedule of this block from end to begin.
+        * Phis are handled togther with live ins after this loop.
+        */
        for(irn = sched_last(bl); !sched_is_begin(irn) && !is_Phi(irn); irn = sched_prev(irn)) {
                ir_node *l;
                int cst;
+               int relevant_args, results;
                int demand;
-               int n_live;
+               int n_cand;
                int must_be_in_mem;
+               pset *cand;
+
+               /*
+                * Determine the number of results
+                */
+               /* Special handling of Projs */
+               if(is_Proj(irn)) {
+                       if(has_reg_class(si, irn)) {
+                               assert(pset_find_ptr(live, irn) && "node must be live");
+                               pset_remove_ptr(live, irn);
+                               skipped++;
+                       }
 
-               /* We handle phi togther with live ins after this loop (see below). */
-               if(is_Phi(irn))
-                       break;
+                       DBG((si->dbg, LEVEL_2, "Skipped %+F\n", irn));
+                       continue;
+               }
 
-#if 0
-               if(has_reg_class(si, irn))
-                       pset_remove_ptr(live, irn);
+               DBG((si->dbg, LEVEL_1, "Irn %+F\n", irn));
+               if(skipped > 0) {
+                       /* ModeT node */
+                       assert(get_irn_mode(irn) == mode_T && "node before projs must be tuple");
+                       results = skipped;
+                       skipped = 0;
+               } else {
+                       /* Normal node */
+                       if(has_reg_class(si, irn)) {
+                               assert(get_irn_mode(irn) != mode_T && "node must not be a tuple");
+                               assert(pset_find_ptr(live, irn) && "node must be live");
+                               pset_remove_ptr(live, irn);
+                               results = 1;
+                       } else {
+                               results = 0;
+                       }
+               }
 
-               demand = register_demand(si, live, irn);
-               n_live = pset_count(live);
-#endif
+               /* cand holds the irns which may be spilled */
+               cand = pset_new_ptr(8);
+               for(l=pset_first(live); l; l=pset_next(live))
+                       pset_insert_ptr(cand, l);
 
-               irn = process_irn(si, live, irn, &demand);
-               n_live = pset_count(live);
+               /*
+                * Determine number of arguments
+                */
+               relevant_args = 0;
+               for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
+                       ir_node *op = get_irn_n(irn, i);
+                       if(has_reg_class(si, op)) {
+                               DBG((si->dbg, LEVEL_2, "  arg %+F\n", op));
+                               relevant_args++;
+                               /* arguments must not be spilled */
+                               if(pset_find_ptr(cand, op))
+                                       pset_remove_ptr(cand, op);
+                       }
+               }
 
                /*
-                * Determine, how many values (which are not used at the label)
-                * must be in memory.
-                * demand means the number of registers, the operation will consume.
-                * So there are n_regs - demand registers available to store values
-                * which are not used at this label. The rest must reside in memory.
+                * Determine, how many values must be in memory.
+                * We have 'n_regs' registers.
+                * The instr. needs 'demand'.
+                * So (R:= n_regs - demand) registers can be used for candidates 'cand'.
+                * The rest (M:= n_cand - R) must reside in memory.
                 */
-               must_be_in_mem = MAX(n_live + demand - n_regs, 0);
+               demand = MAX(results, relevant_args);
+               n_cand = pset_count(cand);
+               must_be_in_mem = n_cand - (n_regs - demand);
 
-               DBG((si->dbg, LEVEL_1, "%+F: demand: %d, live: %d, in mem: %d\n",
-                                       irn, demand, n_live, must_be_in_mem));
+               DBG((si->dbg, LEVEL_1, "  Demand: %d, Cands: %d, InMem: %d\n", demand, n_cand, must_be_in_mem));
+               DBG((si->dbg, LEVEL_3, "  Cand-Set:\n"));
+               print_live_set(si, cand);
 
+               /*
+                * Generate the corresponding constraint spilling
+                * enough candidates at this label.
+                */
                if(must_be_in_mem > 0) {
-
-                       /*
-                        * The constraint limiting the pressure at this label to
-                        * the number of free registers.
-                        */
                        ir_snprintf(buf, sizeof(buf), "cp_%N_%N_%d", bl, irn, step);
                        cst = lpp_add_cst(si->lpp, buf, lpp_greater, must_be_in_mem);
 
-                       for(l = pset_first(live); l; l = pset_next(live)) {
+                       for(l = pset_first(cand); l; l = pset_next(cand)) {
                                live_range_t *lr = get_irn_link(l);
                                lpp_set_factor_fast(si->lpp, cst, lr->in_mem_var, 1.0);
                        }
@@ -359,16 +381,17 @@ static void process_block(ir_node *bl, void *data)
 
                        if(has_reg_class(si, op)) {
                                live_range_t *op_lr = get_live_range(si, op, irn, i);
-
                                set_irn_link(op, op_lr);
 
-                               /*
-                                * The operand is reloaded at its usage, so it must not occur
-                                * in the constraint which determines which values live at the
-                                * instruction must reside in memory.
-                                */
-                               if(must_be_in_mem > 0)
-                                       lpp_set_factor_fast(si->lpp, cst, op_lr->in_mem_var, 0.0);
+//                             /*
+//                              * The operand is reloaded at its usage, so it must not occur
+//                              * in the constraint which determines which values live at the
+//                              * instruction must reside in memory.
+//                              */
+//                             if(must_be_in_mem > 0) {
+//                                     DBG((si->dbg, LEVEL_3, " Resetting %+F to 0:\n", op));
+//                                     lpp_set_factor_fast(si->lpp, cst, op_lr->in_mem_var, 0.0);
+//                             }
 
                                /*
                                 * Check, if the node is a rematerializable node and
@@ -407,12 +430,16 @@ static void process_block(ir_node *bl, void *data)
                        }
                }
 
-                       for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
-                               ir_node *op = get_irn_n(irn, i);
-                               if(has_reg_class(si, op) && !is_Phi(irn))
-                                       pset_insert_ptr(live, op);
-                       }
+               /*
+                * Insert arguments of current instr into the live set
+                */
+               for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
+                       ir_node *op = get_irn_n(irn, i);
+                       if(has_reg_class(si, op))
+                               pset_insert_ptr(live, op);
+               }
 
+               del_pset(cand);
                step++;
        }
 
@@ -420,9 +447,11 @@ static void process_block(ir_node *bl, void *data)
                goto end;
 
        /*
-        * Here, only the phis in the block and the values live in are in the
-        * live set.
+        * Here, the live set contains
+        * - phis of the block
+        * - live-in values of the block
         *
+        * TODO: comment is wrong
         * If a value is live in, it must be in a register in all predecessor
         * blocks or in memory at the end of all predecessor blocks. Also, the
         * closest use in the current block must then be from register or
@@ -433,6 +462,9 @@ static void process_block(ir_node *bl, void *data)
                int is_phi = is_Phi(irn) && get_nodes_block(irn) == bl;
                int cst;
 
+               assert(has_reg_class(si, irn));
+               assert(is_Phi(irn) || is_live_in(bl, irn));
+
                /* Deprecated: Can be done with the first uses map */
                if(is_phi)
                        lr->use_head->closest_use = lr;
@@ -443,25 +475,19 @@ static void process_block(ir_node *bl, void *data)
                 */
                add_first_use(si, bl, irn, lr);
 
-               assert(has_reg_class(si, irn));
-               assert(is_Phi(irn) || is_live_in(bl, irn));
-
                for(i = 0; i < n_preds; ++i) {
                        ir_node *pred_bl     = get_Block_cfgpred_block(bl, i);
                        ir_node *end_node    = is_phi ? get_irn_n(irn, i) : irn;
                        live_range_t *op_lr  = get_live_range(si, end_node, pred_bl, -1);
                        edge_reload_t *edge  = obstack_alloc(si->obst, sizeof(edge[0]));
 
-                       ir_snprintf(buf, sizeof(buf), "edge_%N_%N_%N_%N",
-                                       bl, pred_bl, end_node, op_lr->irn);
+                       ir_snprintf(buf, sizeof(buf), "edge_%N_%N_%N_%N", bl, pred_bl, end_node, op_lr->irn);
                        edge->in_mem_var = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD);
                        edge->bl         = bl;
                        edge->irn        = end_node;
                        edge->pos        = i;
 
-
-                       ir_snprintf(buf, sizeof(buf), "cedge_%N_%N_%N_%N",
-                                       bl, pred_bl, end_node, op_lr->irn);
+                       ir_snprintf(buf, sizeof(buf), "cedge_%N_%N_%N_%N", bl, pred_bl, end_node, op_lr->irn);
                        cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
                        lpp_set_factor_fast(si->lpp, cst, op_lr->in_mem_var, 1.0);
                        lpp_set_factor_fast(si->lpp, cst, lr->in_mem_var, -1.0);
@@ -470,8 +496,7 @@ static void process_block(ir_node *bl, void *data)
        }
 
 end:
-
-  del_pset(live);
+       del_pset(live);
 }
 
 /**
@@ -516,17 +541,16 @@ static int is_spilled(const spill_ilp_t *si, const live_range_t *lr)
        return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
 }
 
+static int is_mem_phi(const ir_node *phi, void *data)
+{
+       spill_ilp_t *si = data;
+       return is_spilled(si, get_use_head(si, phi)->closest_use);
+}
+
 static void writeback_results(spill_ilp_t *si)
 {
        irn_use_head_t *uh;
        edge_reload_t *edge;
-       pset *mem_phis = pset_new_ptr_default();
-
-       /* Put all completely spilled phis into the mem_phis set */
-       for(uh = set_first(si->irn_use_heads); uh; uh = set_next(si->irn_use_heads)) {
-               if(is_Phi(uh->irn) && is_spilled(si, uh->closest_use))
-                       pset_insert_ptr(mem_phis, uh->irn);
-       }
 
        /* Look at each node and examine the usages. */
        for(uh = set_first(si->irn_use_heads); uh; uh = set_next(si->irn_use_heads)) {
@@ -535,17 +559,16 @@ static void writeback_results(spill_ilp_t *si)
                /* Go through all live ranges of the node. */
                list_for_each_entry(live_range_t, lr, &uh->head, list) {
                        if(is_spilled(si, lr) && !is_end_of_block_use(lr))
-                               be_add_spill(si->senv, lr->irn, lr->user);
+                               be_add_reload(si->senv, lr->irn, lr->user);
                }
        }
 
        for(edge = si->edges; edge; edge = edge->next) {
                if(!is_zero(edge->in_mem_var))
-                       be_add_spill_on_edge(si->senv, edge->irn, edge->bl, edge->pos);
+                       be_add_reload_on_edge(si->senv, edge->irn, edge->bl, edge->pos);
        }
 
-       be_insert_spills_reloads(si->senv, mem_phis, NULL);
-       del_pset(mem_phis);
+       be_insert_spills_reloads(si->senv, NULL);
 }
 
 void be_spill_ilp(const be_main_session_env_t *session_env,
@@ -561,7 +584,7 @@ void be_spill_ilp(const be_main_session_env_t *session_env,
        si.session         = session_env;
        si.obst            = &obst;
        si.dbg             = firm_dbg_register("be.ra.spillilp");
-       si.senv            = be_new_spill_env(si.dbg, session_env, cls);
+       si.senv            = be_new_spill_env(si.dbg, session_env, cls, is_mem_phi, &si);
        si.cls             = cls;
        si.lpp             = new_lpp(problem_name, lpp_minimize);
        si.irn_use_heads   = new_set(cmp_irn_use_head, 4096);
@@ -619,7 +642,7 @@ void be_spill_ilp(const be_main_session_env_t *session_env,
        }
 #endif
 
-  // writeback_results(&si);
+       writeback_results(&si);
 
 #ifdef DUMP_STATS
        {
@@ -639,8 +662,8 @@ void be_spill_ilp(const be_main_session_env_t *session_env,
        }
 #endif
 
-  del_set(si.irn_use_heads);
-  del_set(si.live_ranges);
-  free_lpp(si.lpp);
-  obstack_free(&obst, NULL);
+       del_set(si.irn_use_heads);
+       del_set(si.live_ranges);
+       free_lpp(si.lpp);
+       obstack_free(&obst, NULL);
 }