ia32: cleanup handling of 8/16bit operations
[libfirm] / ir / be / bespill.c
index f18206d..7dd59cd 100644 (file)
+/*
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
 /**
- * Author:      Daniel Grund, Sebastian Hack
- * Date:               29.09.2005
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+ * @file
+ * @brief       Spill module selection; Preparation steps
+ * @author      Matthias Braun
+ * @date        29.09.2005
  */
+#include "config.h"
 
-#include "pset.h"
-#include "irnode_t.h"
-#include "ircons_t.h"
+#include "irtools.h"
 #include "debug.h"
+#include "iredges_t.h"
+#include "raw_bitset.h"
+#include "statev.h"
+#include "irgwalk.h"
 
-#include "besched.h"
 #include "bespill.h"
-#include "benode_t.h"
-
-typedef struct _reloader_t reloader_t;
-typedef struct _spill_info_t spill_info_t;
-
-struct _reloader_t {
-       reloader_t *next;
-       ir_node *reloader;
-};
+#include "bemodule.h"
+#include "be.h"
+#include "belive_t.h"
+#include "beirg.h"
+#include "bearch.h"
+#include "benode.h"
+#include "besched.h"
+#include "bera.h"
+#include "beintlive_t.h"
 
-struct _spill_info_t {
-       ir_node *spilled_node;
-       reloader_t *reloaders;
-};
+#include "lc_opts.h"
+#include "lc_opts_enum.h"
 
-typedef struct _spill_ctx_t {
-       ir_node *spilled;  /**< The spilled node. */
-       ir_node *user;     /**< The node this spill is for. */
-       ir_node *spill;    /**< The spill itself. */
-} spill_ctx_t;
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
-struct _spill_env_t {
-       firm_dbg_module_t *dbg;
+typedef struct be_pre_spill_env_t {
+       ir_graph                    *irg;
        const arch_register_class_t *cls;
-       const be_main_session_env_t *session;
-       struct obstack obst;
-       set *spill_ctxs;
-       set *spills;            /**< all spill_info_t's, which must be placed */
-};
-
-static int cmp_spillctx(const void *a, const void *b, size_t n) {
-       const spill_ctx_t *p = a;
-       const spill_ctx_t *q = b;
-       return !(p->user == q->user && p->spilled == q->spilled);
-}
-
-static int cmp_spillinfo(const void *x, const void *y, size_t size) {
-       const spill_info_t *xx = x;
-       const spill_info_t *yy = y;
-       return ! (xx->spilled_node == yy->spilled_node);
-}
-
-spill_env_t *be_new_spill_env(firm_dbg_module_t *dbg,
-               const be_main_session_env_t *session, const arch_register_class_t *cls) {
-       spill_env_t *env = malloc(sizeof(env[0]));
-       env->spill_ctxs = new_set(cmp_spillctx, 1024);
-       env->spills     = new_set(cmp_spillinfo, 1024);
-       env->session    = session;
-       env->cls        = cls;
-       env->dbg        = dbg;
-       obstack_init(&env->obst);
-       return env;
-}
-
-void be_delete_spill_env(spill_env_t *senv) {
-       del_set(senv->spill_ctxs);
-       del_set(senv->spills);
-       obstack_free(&senv->obst, NULL);
-       free(senv);
-}
-
-static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
-       spill_ctx_t templ;
-
-       templ.spilled = to_spill;
-       templ.user    = ctx_irn;
-       templ.spill   = NULL;
+} be_pre_spill_env_t;
+
+static void prepare_constr_insn(be_pre_spill_env_t *env, ir_node *node)
+{
+       const arch_register_class_t *cls = env->cls;
+       ir_node  *block      = get_nodes_block(node);
+       const ir_graph *irg  = env->irg;
+       be_irg_t       *birg = be_birg_from_irg(irg);
+       be_lv_t *lv          = be_get_irg_liveness(irg);
+       unsigned *tmp        = NULL;
+       unsigned *def_constr = NULL;
+       int       arity      = get_irn_arity(node);
+       ir_node  *def;
+
+       int i, i2;
+
+       /* Insert a copy for constraint inputs attached to a value which can't
+        * fulfill the constraint
+        * (typical example: stack pointer as input to copyb)
+        * TODO: This really just checks precolored registers at the moment and
+        *       ignores the general case of not matching in/out constraints
+        */
+       for (i = 0; i < arity; ++i) {
+               ir_node                   *op  = get_irn_n(node, i);
+               const arch_register_req_t *req = arch_get_irn_register_req_in(node, i);
+               const arch_register_t     *reg;
+               ir_node                   *copy;
+
+               if (req->cls != cls)
+                       continue;
+               reg = arch_get_irn_register(op);
+               if (reg == NULL)
+                       continue;
+
+               /* precolored with an ignore register (which is not a joker like
+                  unknown/noreg) */
+               if ((reg->type & arch_register_type_joker) ||
+                   rbitset_is_set(birg->allocatable_regs, reg->global_index))
+                       continue;
+
+               if (! (req->type & arch_register_req_type_limited))
+                       continue;
+               if (rbitset_is_set(req->limited, reg->index))
+                       continue;
+
+               copy = be_new_Copy(block, op);
+               stat_ev_int("constr_copy", 1);
+               sched_add_before(node, copy);
+               set_irn_n(node, i, copy);
+               DBG((dbg, LEVEL_3, "inserting ignore arg copy %+F for %+F pos %d\n",
+                    copy, node, i));
+       }
 
-       return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
-}
+       /* insert copies for nodes that occur constrained more than once. */
+       for (i = 0; i < arity; ++i) {
+               ir_node                   *in;
+               ir_node                   *copy;
+               const arch_register_req_t *req;
+
+               req = arch_get_irn_register_req_in(node, i);
+               if (req->cls != cls)
+                       continue;
+
+               if (! (req->type & arch_register_req_type_limited))
+                       continue;
+
+               in = get_irn_n(node, i);
+               if (!arch_irn_consider_in_reg_alloc(cls, in))
+                       continue;
+
+               for (i2 = i + 1; i2 < arity; ++i2) {
+                       ir_node *in2;
+                       const arch_register_req_t *req2;
+
+                       req2 = arch_get_irn_register_req_in(node, i2);
+                       if (req2->cls != cls)
+                               continue;
+                       if (! (req2->type & arch_register_req_type_limited))
+                               continue;
+
+                       in2 = get_irn_n(node, i2);
+                       if (in2 != in)
+                               continue;
+
+                       /* if the constraint is the same, no copy is necessary
+                        * TODO generalise unequal but overlapping constraints */
+                       if (rbitsets_equal(req->limited, req2->limited, cls->n_regs))
+                               continue;
+
+                       copy = be_new_Copy(block, in);
+                       stat_ev_int("constr_copy", 1);
+
+                       sched_add_before(node, copy);
+                       set_irn_n(node, i2, copy);
+                       DBG((dbg, LEVEL_3,
+                            "inserting multiple constr copy %+F for %+F pos %d\n",
+                            copy, node, i2));
+               }
+       }
 
-static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
-       spill_ctx_t *ctx;
-       DBG((senv->dbg, LEVEL_1, "spill_irn %+F\n", irn));
+       /* collect all registers occurring in out constraints. */
+       be_foreach_definition(node, cls, def,
+               if (! (req_->type & arch_register_req_type_limited))
+                       continue;
+               if (def_constr == NULL) {
+                       rbitset_alloca(def_constr, cls->n_regs);
+               }
+               rbitset_or(def_constr, req_->limited, cls->n_regs);
+       );
 
-       ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
-       if(!ctx->spill) {
-               const be_main_env_t *env = senv->session->main_env;
-               ctx->spill = be_spill(env->node_factory, env->arch_env, irn);
+       /* no output constraints => we're good */
+       if (def_constr == NULL) {
+               return;
        }
 
-       return ctx->spill;
+       /*
+        * insert copies for all constrained arguments living through the node
+        * and being constrained to a register which also occurs in out constraints.
+        */
+       rbitset_alloca(tmp, cls->n_regs);
+       for (i = 0; i < arity; ++i) {
+               const arch_register_req_t *req;
+               ir_node                   *in;
+               ir_node                   *copy;
+
+               /*
+                * Check, if
+                * 1) the operand is constrained.
+                * 2) lives through the node.
+                * 3) is constrained to a register occurring in out constraints.
+                */
+               req = arch_get_irn_register_req_in(node, i);
+               if (req->cls != cls)
+                       continue;
+               if (!(req->type & arch_register_req_type_limited))
+                       continue;
+
+               in = get_irn_n(node, i);
+               if (!arch_irn_consider_in_reg_alloc(cls, in))
+                       continue;
+               if (!be_values_interfere(lv, node, in))
+                       continue;
+
+               rbitset_copy(tmp, req->limited, cls->n_regs);
+               rbitset_and(tmp, def_constr, cls->n_regs);
+
+               if (rbitset_is_empty(tmp, cls->n_regs))
+                       continue;
+
+               /*
+                * only create the copy if the operand is no copy.
+                * this is necessary since the assure constraints phase inserts
+                * Copies and Keeps for operands which must be different from the
+                * results. Additional copies here would destroy this.
+                */
+               if (be_is_Copy(in))
+                       continue;
+
+               copy = be_new_Copy(block, in);
+               sched_add_before(node, copy);
+               set_irn_n(node, i, copy);
+               DBG((dbg, LEVEL_3, "inserting constr copy %+F for %+F pos %d\n",
+                    copy, node, i));
+               be_liveness_update(lv, in);
+       }
 }
 
-/**
- * If the first usage of a phi result would be out of memory
- * there is no sense in allocating a register for it.
- * Thus we spill it and all its operands to the same spill slot.
- * Therefore the phi/dataB becomes a phi/Memory
- */
-static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn,
-               pset *mem_phis) {
-       int i, n = get_irn_arity(phi);
-       ir_node **ins, *bl = get_nodes_block(phi);
-       ir_graph *irg = senv->session->irg;
-       spill_ctx_t *ctx;
-
-       assert(is_Phi(phi));
-       DBG((senv->dbg, LEVEL_1, "spill_phi %+F\n", phi));
-
-       /* search an existing spill for this context */
-       ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
-
-       /* if not found spill the phi */
-       if(!ctx->spill) {
-               /* build a new PhiM with dummy in-array */
-               ins  = malloc(n * sizeof(ins[0]));
-               for(i=0; i<n; ++i)
-                       ins[i] = new_r_Unknown(irg, mode_M);
-               ctx->spill = new_r_Phi(senv->session->irg, bl, n, ins, mode_M);
-               free(ins);
-
-               /* re-wire the phiM */
-               for(i=0; i<n; ++i) {
-                       ir_node *arg = get_irn_n(phi, i);
-                       ir_node *sub_res;
-
-                       if(is_Phi(arg) && pset_find_ptr(mem_phis, arg))
-                               sub_res = be_spill_phi(senv, arg, ctx_irn, mem_phis);
-                       else
-                               sub_res = be_spill_irn(senv, arg, ctx_irn);
-
-                       set_irn_n(ctx->spill, i, sub_res);
-               }
+static void pre_spill_prepare_constr_walker(ir_node *block, void *data)
+{
+       be_pre_spill_env_t *env = (be_pre_spill_env_t*)data;
+       sched_foreach(block, node) {
+               prepare_constr_insn(env, node);
        }
-       return ctx->spill;
 }
 
-static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill, pset *mem_phis) {
-       ir_node *res;
-       if (pset_find_ptr(mem_phis, to_spill))
-               res = be_spill_phi(senv, to_spill, to_spill, mem_phis);
-       else
-               res = be_spill_irn(senv, to_spill, to_spill);
+void be_pre_spill_prepare_constr(ir_graph *irg,
+                                 const arch_register_class_t *cls)
+{
+       be_pre_spill_env_t env;
+       memset(&env, 0, sizeof(env));
+       env.irg = irg;
+       env.cls = cls;
 
-       return res;
+       be_assure_live_sets(irg);
+
+       irg_block_walk_graph(irg, pre_spill_prepare_constr_walker, NULL, &env);
 }
 
-void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set, decide_irn_t is_mem_phi, void *data) {
-       ir_graph *irg = senv->session->irg;
-       ir_node *irn;
-       spill_info_t *si;
-       struct obstack ob;
-       pset *mem_phis = pset_new_ptr_default();
 
-       obstack_init(&ob);
 
-       /* get all special spilled phis */
-       for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
-               irn = si->spilled_node;
-               if (is_Phi(irn) && is_mem_phi(irn, data))
-                       pset_insert_ptr(mem_phis, irn);
-       }
+int be_coalesce_spill_slots = 1;
+int be_do_remats = 1;
 
-       /* process each spilled node */
-       for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
-               reloader_t *rld;
-               ir_node **reloads;
-               int n_reloads = 0;
-               ir_mode *mode = get_irn_mode(si->spilled_node);
-
-               /* go through all reloads for this spill */
-               for(rld = si->reloaders; rld; rld = rld->next) {
-                       /* the spill for this reloader */
-                       ir_node *spill   = be_spill_node(senv, si->spilled_node, mem_phis);
-
-                       /* the reload */
-                       ir_node *bl      = is_Block(rld->reloader) ? rld->reloader : get_nodes_block(rld->reloader);
-                       ir_node *reload  = new_Reload(senv->session->main_env->node_factory,
-                                                                                                               senv->cls, irg, bl, mode, spill);
-
-                       DBG((senv->dbg, LEVEL_2, " RELOADER %+F   Reload %+F of %+F\n",
-                                               rld->reloader, reload, si->spilled_node));
-                       if(reload_set)
-                               pset_insert_ptr(reload_set, reload);
-
-                       /* remember the reaload */
-                       obstack_ptr_grow(&ob, reload);
-                       sched_add_before(rld->reloader, reload);
-                       n_reloads++;
-               }
+static const lc_opt_table_entry_t be_spill_options[] = {
+       LC_OPT_ENT_BOOL ("coalesce_slots", "coalesce the spill slots", &be_coalesce_spill_slots),
+       LC_OPT_ENT_BOOL ("remat", "try to rematerialize values instead of reloading", &be_do_remats),
+       LC_OPT_LAST
+};
 
-               assert(n_reloads > 0);
-               reloads = obstack_finish(&ob);
-               be_introduce_copies_ignore(senv->session->dom_front, si->spilled_node,
-                               n_reloads, reloads, mem_phis);
-               obstack_free(&ob, reloads);
-       }
+static be_module_list_entry_t *spillers = NULL;
+static const be_spiller_t *selected_spiller = NULL;
 
-       obstack_free(&ob, NULL);
+void be_register_spiller(const char *name, be_spiller_t *spiller)
+{
+       if (selected_spiller == NULL)
+               selected_spiller = spiller;
+       be_add_module_to_list(&spillers, name, spiller);
+}
 
-       for(irn = pset_first(mem_phis); irn; irn = pset_next(mem_phis)) {
-               int i, n;
-               for(i = 0, n = get_irn_arity(irn); i < n; ++i)
-                       set_irn_n(irn, i, new_r_Bad(senv->session->irg));
-               sched_remove(irn);
-       }
+void be_do_spill(ir_graph *irg, const arch_register_class_t *cls)
+{
+       assert(selected_spiller != NULL);
 
-       del_pset(mem_phis);
+       selected_spiller->spill(irg, cls);
 }
 
-void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
-       spill_info_t templ, *res;
-       reloader_t *rel;
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spilloptions)
+void be_init_spilloptions(void)
+{
+       lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
+       lc_opt_entry_t *spill_grp = lc_opt_get_grp(be_grp, "spill");
 
-       templ.spilled_node = to_spill;
-       templ.reloaders    = NULL;
-       res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
-
-       rel           = obstack_alloc(&senv->obst, sizeof(rel[0]));
-       rel->reloader = before;
-       rel->next     = res->reloaders;
-       res->reloaders = rel;
-}
+       lc_opt_add_table(spill_grp, be_spill_options);
+       be_add_module_list_opt(be_grp, "spiller", "spill algorithm",
+                              &spillers, (void**) &selected_spiller);
 
-void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
-       ir_node *insert_bl = get_irn_arity(bl) == 1
-               ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
-       be_add_reload(senv, to_spill, insert_bl);
+       FIRM_DBG_REGISTER(dbg, "firm.be.spillprepare");
 }