X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbeilpsched.c;h=f09298881792d58b865768fd4e4d69bba5753cd4;hb=2589efb23f5c24a341529d4f75d6f5f7eddaba00;hp=0c3da1d2cb80ffe55f8260bde85effbb23b02a62;hpb=70469bd0bc61de41ecd872acfae34f01d91f2056;p=libfirm diff --git a/ir/be/beilpsched.c b/ir/be/beilpsched.c index 0c3da1d2c..f09298881 100644 --- a/ir/be/beilpsched.c +++ b/ir/be/beilpsched.c @@ -1,16 +1,35 @@ +/* + * 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. + */ + /** - * Scheduling algorithms. + * @file + * @brief ILP based instruction scheduling. + * @author Christian Wuerdig + * @date 22.10.2006 + * @version $Id$ + * * An ILP scheduler based on * "ILP-based Instruction Scheduling for IA-64" * by Daniel Kaestner and Sebastian Winkel - * - * @date 22.10.2005 - * @author Christian Wuerdig - * @cvs-id $Id$ + * extended with register pressure constraints by Christian Wuerdig */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #ifdef WITH_ILP @@ -30,26 +49,30 @@ #include "debug.h" #include "irtools.h" #include "irdump.h" +#include "irprintf.h" #include "plist.h" +#include "irprintf.h" +#include "timing.h" #include #include -#ifdef WITH_LIBCORE -#include -#include -#include -#endif /* WITH_LIBCORE */ +#include "lc_opts.h" +#include "lc_opts_enum.h" #include "be.h" -#include "benode_t.h" -#include "besched_t.h" +#include "benode.h" +#include "besched.h" #include "beilpsched.h" #include "beutil.h" +#include "bestat.h" +#include "beirg.h" +#include "bemachine.h" +#include "belistsched.h" +#include "bemodule.h" typedef struct _ilpsched_options_t { unsigned regpress; - unsigned limit_dead; unsigned time_limit; char log_file[1024]; } ilpsched_options_t; @@ -65,10 +88,18 @@ typedef struct _unit_type_info_t { typedef struct _ilp_var_types_t { int *x; /* x_{nt}^k variables */ int *a; /* a_{nt}^k variables */ - int *d; /* d_{nt}^k variables */ int *y; /* y_{nt}^k variables */ } ilp_var_types_t; +/** + * Holds alive variables for a node live-in to a block. + */ +typedef struct _ilp_livein_node_t { + ir_node *irn; + unsigned max_alive_steps; + int *a; +} ilp_livein_node_t; + /* attributes for a node */ typedef struct _ilpsched_node_attr_t { unsigned asap; /**< The ASAP scheduling control step */ @@ -85,7 +116,7 @@ typedef struct _ilpsched_node_attr_t { unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */ bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors for ASAP, successors for ALAP */ - unsigned n_unit_types; /**< number of allowed execution unit types */ + unsigned n_unit_types; /**< number of allowed execution unit types */ unit_type_info_t *type_info; /**< list of allowed execution unit types */ ilp_var_types_t ilp_vars; /**< the different ILP variables */ } ilpsched_node_attr_t; @@ -97,6 +128,7 @@ typedef struct _ilpsched_block_attr_t { unsigned max_steps; /**< Upper bound for block execution */ plist_t *root_nodes; /**< A list of nodes having no user in current block */ ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */ + pset *livein_nodes; /**< A set of nodes which are live-in to this block */ } ilpsched_block_attr_t; typedef union _ilpsched_attr_ { @@ -106,20 +138,18 @@ typedef union _ilpsched_attr_ { /* A irn for the phase and it's attributes (either node or block) */ typedef struct { - ir_node *irn; + const ir_node *irn; ilpsched_attr_t attr; } be_ilpsched_irn_t; /* The ILP scheduling environment */ typedef struct { - phase_t ph; /**< The phase */ + ir_phase ph; /**< The phase */ ir_graph *irg; /**< The current irg */ heights_t *height; /**< The heights object of the irg */ void *irg_env; /**< An environment for the irg scheduling, provided by the backend */ void *block_env; /**< An environment for scheduling a block, provided by the backend */ const arch_env_t *arch_env; - const arch_isa_t *isa; /**< The ISA */ - const be_main_env_t *main_env; const be_machine_t *cpu; /**< the current abstract machine */ ilpsched_options_t *opts; /**< the ilp options for current irg */ const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */ @@ -133,9 +163,9 @@ typedef struct { #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr) /* check if node is considered for ILP scheduling */ -#define consider_for_sched(isa, irn) \ +#define consider_for_sched(env, irn) \ (! (is_Block(irn) || \ - is_normal_Proj(isa, irn) || \ + is_normal_Proj(env, irn) || \ is_Phi(irn) || \ is_NoMem(irn) || \ is_Unknown(irn) || \ @@ -159,34 +189,20 @@ typedef struct { /* check if a double value is within an epsilon environment of 0 */ #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10) -#ifdef WITH_LIBCORE - #define ilp_timer_push(t) lc_timer_push((t)) - #define ilp_timer_pop() lc_timer_pop() - #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t)) -#else /* WITH_LIBCORE */ - #define ilp_timer_push(t) - #define ilp_timer_pop() - #define ilp_timer_elapsed_usec(t) 0.0 -#endif /* WITH_LIBCORE */ - /* option variable */ static ilpsched_options_t ilp_opts = { 1, /* default is with register pressure constraints */ - 120, /* if we have more than 70 nodes: use alive nodes constraint */ 300, /* 300 sec per block time limit */ "" /* no log file */ }; -#ifdef WITH_LIBCORE /* ILP options */ static const lc_opt_table_entry_t ilpsched_option_table[] = { LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress), - LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead), LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit), LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)), - { NULL } + LC_OPT_LAST }; -#endif /* WITH_LIBCORE */ /* We need this global variable as we compare nodes dependent on heights, @@ -198,40 +214,51 @@ static heights_t *glob_heights; * Check if irn is a Proj, which has no execution units assigned. * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise */ -static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) { - return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL); +static inline int is_normal_Proj(const arch_env_t *env, const ir_node *irn) +{ + return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL); } /** * Skips normal Projs. * @return predecessor if irn is a normal Proj, otherwise irn. */ -static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) { - if (is_normal_Proj(isa, irn)) +static inline ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn) +{ + if (is_normal_Proj(env, irn)) return get_Proj_pred(irn); return irn; } -static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) { +static inline int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) +{ unsigned lat = be_ilp_sched_latency(sel, irn, env); if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn)) lat = 1; return lat; } +static int cmp_live_in_nodes(const void *a, const void *b) +{ + const ilp_livein_node_t *n1 = a; + const ilp_livein_node_t *n2 = b; + + return n1->irn != n2->irn; +} /** * Compare scheduling time steps of two be_ilpsched_irn's. */ -static int cmp_ilpsched_irn(const void *a, const void *b) { +static int cmp_ilpsched_irn(const void *a, const void *b) +{ be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a; be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b; ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1); ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2); if (n1_a->sched_point == n2_a->sched_point) { - ir_node *irn_a = n1->irn; - ir_node *irn_b = n2->irn; + const ir_node *irn_a = n1->irn; + const ir_node *irn_b = n2->irn; if (heights_reachable_in_block(glob_heights, irn_a, irn_b)) return 1; @@ -251,7 +278,8 @@ static int cmp_ilpsched_irn(const void *a, const void *b) { /** * In case there is no phase information for irn, initialize it. */ -static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) { +static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old) +{ be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0])); if (res == old) { @@ -288,6 +316,7 @@ static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) { ba->block_last_idx = 0; ba->root_nodes = plist_new(); ba->head_ilp_nodes = NULL; + ba->livein_nodes = new_pset(cmp_live_in_nodes, 16); ba->max_steps = 0; } else { @@ -301,13 +330,15 @@ static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) { /** * Assign a per block unique number to each node. */ -static void build_block_idx(ir_node *irn, void *walk_env) { +static void build_block_idx(ir_node *irn, void *walk_env) +{ be_ilpsched_env_t *env = walk_env; be_ilpsched_irn_t *node, *block_node; ilpsched_node_attr_t *na; ilpsched_block_attr_t *ba; - if (! consider_for_sched(env->arch_env->isa, irn)) + set_irn_link(irn, NULL); + if (! consider_for_sched(env->arch_env, irn)) return; node = get_ilpsched_irn(env, irn); @@ -332,7 +363,8 @@ static void build_block_idx(ir_node *irn, void *walk_env) { /** * Add all nodes having no user in current block to last_nodes list. */ -static void collect_alap_root_nodes(ir_node *irn, void *walk_env) { +static void collect_alap_root_nodes(ir_node *irn, void *walk_env) +{ ir_node *block; const ir_edge_t *edge; be_ilpsched_irn_t *block_node, *node; @@ -344,9 +376,9 @@ static void collect_alap_root_nodes(ir_node *irn, void *walk_env) { unsigned n_consumer = 0; ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP }; ir_node **consumer; - int idx; + unsigned idx; - if (! consider_for_sched(env->arch_env->isa, irn)) + if (! consider_for_sched(env->arch_env, irn)) return; block = get_nodes_block(irn); @@ -360,7 +392,7 @@ static void collect_alap_root_nodes(ir_node *irn, void *walk_env) { foreach_out_edge_kind(irn, edge, ekind[i]) { ir_node *user = get_edge_src_irn(edge); - if (is_normal_Proj(env->arch_env->isa, user)) { + if (is_normal_Proj(env->arch_env, user)) { const ir_edge_t *user_edge; if (get_irn_mode(user) == mode_X) @@ -421,6 +453,9 @@ static void collect_alap_root_nodes(ir_node *irn, void *walk_env) { /* mark user visited by this node */ ua->consumer_idx = idx; } + else if (get_nodes_block(user) != block) { + n_consumer++; + } } } @@ -450,7 +485,8 @@ static void collect_alap_root_nodes(ir_node *irn, void *walk_env) { /** * Calculate the ASAP scheduling step for current irn. */ -static void calculate_irn_asap(ir_node *irn, void *walk_env) { +static void calculate_irn_asap(ir_node *irn, void *walk_env) +{ be_ilpsched_env_t *env = walk_env; int i; ir_node *block; @@ -459,7 +495,7 @@ static void calculate_irn_asap(ir_node *irn, void *walk_env) { ilpsched_block_attr_t *ba; /* These nodes are handled separate */ - if (! consider_for_sched(env->arch_env->isa, irn)) + if (! consider_for_sched(env->arch_env, irn)) return; DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn)); @@ -470,7 +506,7 @@ static void calculate_irn_asap(ir_node *irn, void *walk_env) { na->asap = 1; for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) { - ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i)); + ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i)); /* check for greatest distance to top */ if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) { @@ -499,7 +535,8 @@ static void calculate_irn_asap(ir_node *irn, void *walk_env) { * Calculate the ALAP scheduling step of all irns in current block. * Depends on max_steps being calculated. */ -static void calculate_block_alap(ir_node *block, void *walk_env) { +static void calculate_block_alap(ir_node *block, void *walk_env) +{ be_ilpsched_env_t *env = walk_env; be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block); ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); @@ -543,7 +580,7 @@ static void calculate_block_alap(ir_node *block, void *walk_env) { /* set the alap's of all predecessors */ for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) { - ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i)); + ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i)); /* check for greatest distance to bottom */ if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) { @@ -590,28 +627,32 @@ static void calculate_block_alap(ir_node *block, void *walk_env) { } /** - * We can free the list of root nodes here. + * Free list of root nodes and the set of live-in nodes. */ -static void clear_unwanted_data(ir_node *block, void *walk_env) { +static void clear_unwanted_data(ir_node *block, void *walk_env) +{ be_ilpsched_env_t *env = walk_env; be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block); ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); plist_free(ba->root_nodes); ba->root_nodes = NULL; + del_pset(ba->livein_nodes); + ba->livein_nodes = NULL; } /** * Refine the {ASAP(n), ALAP(n)} interval for the nodes. * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones. */ -static void refine_asap_alap_times(ir_node *irn, void *walk_env) { +static void refine_asap_alap_times(ir_node *irn, void *walk_env) +{ be_ilpsched_env_t *env = walk_env; ir_node *pred = irn; be_ilpsched_irn_t *node, *pred_node; ilpsched_node_attr_t *na, *pna; - if (! consider_for_sched(env->arch_env->isa, irn)) + if (! consider_for_sched(env->arch_env, irn)) return; if (! is_Proj(irn) && ! be_is_Keep(irn)) @@ -648,8 +689,10 @@ static void refine_asap_alap_times(ir_node *irn, void *walk_env) { * *******************************************/ -static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) { +static inline void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn) +{ const ir_edge_t *edge; + (void) block; foreach_out_edge(irn, edge) { ir_node *user = get_edge_src_irn(edge); @@ -664,18 +707,19 @@ static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) { /** * Inserts @p irn before @p before into schedule and notifies backend. */ -static INLINE void notified_sched_add_before(be_ilpsched_env_t *env, - ir_node *before, ir_node *irn, unsigned cycle) +static inline void notified_sched_add_before(be_ilpsched_env_t *env, + const ir_node *before, const ir_node *irn, unsigned cycle) { be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env); - sched_add_before(before, irn); + sched_add_before((ir_node*) before, (ir_node*) irn); } /** * Adds a node, it's Projs (in case of mode_T nodes) and * it's Keeps to schedule. */ -static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) { +static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle) +{ const ir_edge_t *edge; waitq *keeps = new_waitq(); @@ -691,8 +735,11 @@ static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, u foreach_out_edge(irn, edge) { ir_node *user = get_edge_src_irn(edge); - if (to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) + if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) && + get_irn_n_edges(user) > 0) + { notified_sched_add_before(env, block, user, cycle); + } check_for_keeps(keeps, block, user); } @@ -714,18 +761,17 @@ static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, u /** * Schedule all nodes in the given block, according to the ILP solution. */ -static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) { +static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) +{ be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block); ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); - sched_info_t *info = get_irn_sched_info(block); be_ilpsched_irn_t **sched_nodes; unsigned i, l; ir_node *cfop, *irn; const ir_edge_t *edge; /* init block schedule list */ - INIT_LIST_HEAD(&info->list); - info->scheduled = 1; + sched_init_block(block); /* collect nodes and their scheduling time step */ sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0); @@ -751,6 +797,20 @@ static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) { cur_var = 0; found = 0; + if (! na->is_dummy_node) { + for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) { + for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) { + double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]); + + if (! LPP_VALUE_IS_0(cost)) { + DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost)); + found = 1; + } + } + } + } + + found = 0; /* go over all variables of a node until the non-zero one is found */ for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) { for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) { @@ -782,7 +842,7 @@ static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) { case iro_Phi: add_to_sched(env, block, irn, 0); break; - case iro_Start: + case beo_Start: case iro_End: case iro_Proj: case iro_Bad: @@ -824,7 +884,8 @@ static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) { /** * Check if node can be executed on given unit type. */ -static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) { +static inline int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) +{ int i; ilpsched_node_attr_t *na = get_ilpsched_node_attr(node); @@ -846,63 +907,101 @@ static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp * ************************************************/ +static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) +{ + const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn); + unsigned n_unit_types = 0; + be_ilpsched_irn_t *node; + ilpsched_node_attr_t *na; + unsigned unit_idx, tp_idx; + + /* count number of available unit types for this node */ + for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types) + /* just count */ ; + + node = get_ilpsched_irn(env, irn); + na = get_ilpsched_node_attr(node); + + if (! na->type_info) { + na->n_unit_types = n_unit_types; + na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types); + + /* fill the type info array */ + for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) { + for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) { + /* beware: we also count number of available units here */ + if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx])) + na->is_dummy_node = 1; + } + + na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp; + na->type_info[tp_idx].n_units = unit_idx; + } + } + + return n_unit_types; +} + +/** + * Returns the largest alap time of a user of @p irn. + * The user must be in block @p block. + */ +static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block) +{ + const ir_edge_t *edge; + unsigned max_alap = 0; + + foreach_out_edge(irn, edge) { + ir_node *user = get_edge_src_irn(edge); + + if (get_nodes_block(user) == block) { + be_ilpsched_irn_t *node = get_ilpsched_irn(env, user); + ilpsched_node_attr_t *na = get_ilpsched_node_attr(node); + + max_alap = MAX(max_alap, na->alap); + } + } + + assert(max_alap > 0); + return max_alap; +} + /** * Create the following variables: * - x_{nt}^k binary weigthed with: t * node n is scheduled at time step t to unit type k * ==>> These variables represent the schedule * - * - d_{nt}^k binary weighted with: t - * node n dies at time step t on unit type k * - a_{nt}^k binary weighted with num_nodes * node n is alive at time step t on unit type k * - * - y_{nt}^k binary weighted with: num_nodes^2 - * node n is scheduled at time step t to unit type k - * although all units of this type are occupied + * - y_{nt}^k continuous weighted with: num_nodes^2 + * register pressure over limit for unit type k * ==>> These variables represent the register pressure * */ -static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) { +static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) +{ char buf[1024]; ir_node *irn; unsigned num_block_var, num_nodes; + ilp_livein_node_t *livein; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes; -#ifdef WITH_LIBCORE - lc_timer_t *t_var = lc_timer_register("beilpsched_var", "create ilp variables"); -#endif /* WITH_LIBCORE */ - ilp_timer_push(t_var); + stat_ev_tim_push(); + num_block_var = num_nodes = 0; foreach_linked_irns(ba->head_ilp_nodes, irn) { - const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn); - be_ilpsched_irn_t *node; - ilpsched_node_attr_t *na; - unsigned n_unit_types, tp_idx, unit_idx, n_var, cur_unit; - unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad; - - /* count number of available unit types for this node */ - for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types) - /* just count */ ; - - node = get_ilpsched_irn(env, irn); - na = get_ilpsched_node_attr(node); - - na->n_unit_types = n_unit_types; - na->type_info = NEW_ARR_D(unit_type_info_t, var_obst, n_unit_types); - - /* fill the type info array */ - for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) { - for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) { - /* beware: we also count number of available units here */ - if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx])) - na->is_dummy_node = 1; - } + be_ilpsched_irn_t *node; + ilpsched_node_attr_t *na; + unsigned n_unit_types, tp_idx, n_var, cur_unit; + unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad; + int i; - na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp; - na->type_info[tp_idx].n_units = unit_idx; - } + node = get_ilpsched_irn(env, irn); + na = get_ilpsched_node_attr(node); + n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst); /* allocate space for ilp variables */ na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na)); @@ -913,16 +1012,9 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na)); memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0])); - num_ad = ba->max_steps - na->asap + 1; - - if (ba->n_interesting_nodes > env->opts->limit_dead) { - na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad); - memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0])); - } - else { - na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad); - memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0])); - } + num_ad = ba->max_steps - na->asap + 1; + na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad); + memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0])); } DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n", @@ -947,7 +1039,7 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn /* y_{nt}^k variables */ snprintf(buf, sizeof(buf), "y_n%u_%s_%u", get_irn_idx(irn), na->type_info[tp_idx].tp->name, t); - na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_binary, (double)(weigth_y)); + na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y)); DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf)); /* variable counter */ @@ -960,18 +1052,10 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn if (! na->is_dummy_node) { for (t = na->asap - 1; t <= ba->max_steps; ++t) { - if (ba->n_interesting_nodes > env->opts->limit_dead) { - /* a_{nt}^k variables */ - snprintf(buf, sizeof(buf), "a_n%u_%s_%u", - get_irn_idx(irn), na->type_info[tp_idx].tp->name, t); - na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes)); - } - else { - /* d_{nt}^k variables */ - snprintf(buf, sizeof(buf), "d_n%u_%s_%u", - get_irn_idx(irn), na->type_info[tp_idx].tp->name, t); - na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1)); - } + /* a_{nt}^k variables */ + snprintf(buf, sizeof(buf), "a_n%u_%s_%u", + get_irn_idx(irn), na->type_info[tp_idx].tp->name, t); + na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes)); DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf)); /* variable counter */ @@ -979,14 +1063,59 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn num_block_var++; } } + + /* collect live-in nodes */ + for (i = get_irn_arity(irn) - 1; i >= 0; --i) { + ir_node *pred = get_irn_n(irn, i); + + if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) { + be_ilpsched_set_type_info(env, pred, var_obst); + if (! na->is_dummy_node) { + ilp_livein_node_t *entry = OALLOC(var_obst, ilp_livein_node_t); + entry->irn = pred; + entry->a = NULL; + pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred)); + } + } + } } DB((env->dbg, LEVEL_3, "%u variables created\n", n_var)); num_nodes++; } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n", - num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0)); + + /* create alive variables a_{nt}^k for live-ins */ + foreach_pset(ba->livein_nodes, livein) { + be_ilpsched_irn_t *node; + ilpsched_node_attr_t *na; + unsigned tp_idx, var_idx; + ir_node *irn; + + irn = livein->irn; + node = get_ilpsched_irn(env, irn); + na = get_ilpsched_node_attr(node); + + livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn); + + livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps); + var_idx = 0; + + /* create variables */ + for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) { + unsigned t; + + for (t = 0; t < livein->max_alive_steps; ++t) { + /* a_{nt}^k variables */ + snprintf(buf, sizeof(buf), "al_n%u_%s_%u", + get_irn_idx(irn), na->type_info[tp_idx].tp->name, t); + livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes)); + DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf)); + num_block_var++; + } + } + } + + stat_ev_tim_pop("beilpsched_var"); } /******************************************************* @@ -999,6 +1128,32 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn * *******************************************************/ +/** + * Collect all operands and nodes @p irn depends on. + * If there is a Proj within the dependencies, all other Projs of the parent node are added as well. + */ +static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps) +{ + int i; + + for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) { + ir_node *p = get_irn_in_or_dep(irn, i); + + if (is_Proj(p)) { + const ir_edge_t *edge; + + p = get_Proj_pred(p); + foreach_out_edge(p, edge) { + ir_node *src = get_edge_src_irn(edge); + ir_nodeset_insert(deps, src); + } + } + else { + ir_nodeset_insert(deps, p); + } + } +} + /** * Create following ILP constraints: * - the assignment constraints: @@ -1008,49 +1163,39 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn * - the precedence constraints: * assure that no data dependencies are violated */ -static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ unsigned num_cst_assign, num_cst_prec, num_cst_dead; char buf[1024]; ir_node *irn; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst_assign = lc_timer_register("beilpsched_cst_assign", "create assignment constraints"); - lc_timer_t *t_cst_dead = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints"); - lc_timer_t *t_cst_prec = lc_timer_register("beilpsched_cst_prec", "create precedence constraints"); -#endif /* WITH_LIBCORE */ num_cst_assign = num_cst_prec = num_cst_dead = 0; foreach_linked_irns(ba->head_ilp_nodes, irn) { - int cst, tp_idx, i; + int cst, tp_idx; unsigned cur_var; be_ilpsched_irn_t *node; ilpsched_node_attr_t *na; + ir_node *pred; + ir_nodeset_t deps; + ir_nodeset_iterator_t iter; + + ir_nodeset_init(&deps); node = get_ilpsched_irn(env, irn); na = get_ilpsched_node_attr(node); cur_var = 0; /* the assignment constraint */ - ilp_timer_push(t_cst_assign); + stat_ev_tim_push(); snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn)); cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0); DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf)); num_cst_assign++; lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0); - ilp_timer_pop(); - - /* the dead node assignment constraint */ - if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) { - ilp_timer_push(t_cst_dead); - snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn)); - cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0); - DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf)); - - lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0); - ilp_timer_pop(); - } + stat_ev_tim_pop("beilpsched_cst_assign"); /* We have separate constraints for Projs and Keeps */ // ILP becomes infeasible ?!? @@ -1058,15 +1203,17 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, // continue; /* the precedence constraints */ - ilp_timer_push(t_cst_prec); - bs_block_irns = bitset_clear_all(bs_block_irns); - for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) { - ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i)); + stat_ev_tim_push(); + bitset_clear_all(bs_block_irns); + + sta_collect_in_deps(irn, &deps); + foreach_ir_nodeset(&deps, pred, iter) { unsigned t_low, t_high, t; be_ilpsched_irn_t *pred_node; ilpsched_node_attr_t *pna; unsigned delay; + pred = skip_normal_Proj(env->arch_env, pred); if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred)) continue; @@ -1133,12 +1280,9 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, DEL_ARR_F(tmp_var_idx); } } - ilp_timer_pop(); + ir_nodeset_destroy(&deps); + stat_ev_tim_pop("beilpsched_cst_prec"); } - DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n", - num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0)); - DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n", - num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0)); } /** @@ -1146,16 +1290,14 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, * - assure that for each time step not more instructions are scheduled * to the same unit types as units of this type are available */ -static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ int glob_type_idx; char buf[1024]; unsigned num_cst_resrc = 0; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst_rsrc = lc_timer_register("beilpsched_cst_rsrc", "create resource constraints"); -#endif /* WITH_LIBCORE */ - ilp_timer_push(t_cst_rsrc); + stat_ev_tim_push(); for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) { unsigned t; be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx]; @@ -1193,9 +1335,7 @@ static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ DEL_ARR_F(tmp_var_idx); } } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n", - num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0)); + stat_ev_tim_pop("beilpsched_cst_rsrc"); } /** @@ -1203,17 +1343,15 @@ static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ * - assure, at most bundle_size * bundles_per_cycle instructions * can be started at a certain point. */ -static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ char buf[1024]; unsigned t; unsigned num_cst_bundle = 0; unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst_bundle = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints"); -#endif /* WITH_LIBCORE */ - ilp_timer_push(t_cst_bundle); + stat_ev_tim_push(); for (t = 0; t < ba->max_steps; ++t) { ir_node *irn; int cst; @@ -1253,109 +1391,21 @@ static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilp DEL_ARR_F(tmp_var_idx); } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n", - num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0)); + stat_ev_tim_pop("beilpsched_cst_bundle"); } /** - * Create ILP dying nodes constraints: - * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k + * Create ILP alive nodes constraints: + * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k */ -static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { - char buf[1024]; - unsigned t; - unsigned num_cst = 0; - ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints"); -#endif /* WITH_LIBCORE */ - - ilp_timer_push(t_cst); - /* check all time_steps */ - for (t = 0; t < ba->max_steps; ++t) { - ir_node *irn; - - /* for all nodes */ - foreach_linked_irns(ba->head_ilp_nodes, irn) { - be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn); - ilpsched_node_attr_t *na = get_ilpsched_node_attr(node); - - /* if node has no consumer within current block, it cannot die here */ - /* we also ignore nodes assigned to dummy unit */ - if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node) - continue; - - /* node can only die here if t at least asap(n) */ - if (t >= na->asap - 1) { - int node_tp_idx; - - /* for all unit types */ - for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) { - int tp_idx, i, cst; - int *tmp_var_idx = NEW_ARR_F(int, 0); - - snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn)); - cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1)); - DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf)); - num_cst++; - - /* number of consumer scheduled till t */ - for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) { - be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]); - ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons); - - for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) { - unsigned tm; - - for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) { - int idx = ILPVAR_IDX(ca, tp_idx, tm); - ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]); - } - } - } - - /* could be that no consumer can be scheduled at this point */ - if (ARR_LEN(tmp_var_idx)) { - int idx; - unsigned tn; - - /* subtract possible prior kill points */ - for (tn = na->asap - 1; tn < t; ++tn) { - idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn); - lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0); - } - - idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t); - lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer)); - lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0); - } - - DEL_ARR_F(tmp_var_idx); - } - } - - } - } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n", - num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0)); -} - -/** -* Create ILP alive nodes constraints: -* - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k -*/ -static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ char buf[1024]; ir_node *irn; unsigned num_cst = 0; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints"); -#endif /* WITH_LIBCORE */ - ilp_timer_push(t_cst); + stat_ev_tim_push(); /* for each node */ foreach_linked_irns(ba->head_ilp_nodes, irn) { be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn); @@ -1421,110 +1471,83 @@ static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be } } } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n", - num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0)); + stat_ev_tim_pop("beilpsched_cst_alive_nodes"); } /** - * Create ILP pressure constraints, based on dead nodes: - * - add additional costs to objective function if a node is scheduled - * on a unit although all units of this type are currently occupied + * Create ILP alive nodes constraints for live-in nodes: + * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k */ -static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ char buf[1024]; - ir_node *cur_irn; + ilp_livein_node_t *livein; unsigned num_cst = 0; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints"); -#endif /* WITH_LIBCORE */ - - ilp_timer_push(t_cst); - /* y_{nt}^k is set for each node and timestep and unit type */ - foreach_linked_irns(ba->head_ilp_nodes, cur_irn) { - unsigned cur_idx = get_irn_idx(cur_irn); - be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn); - ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node); - int glob_type_idx; - - /* we ignore nodes assigned to DUMMY unit here */ - if (cur_na->is_dummy_node) - continue; - /* for all types */ - for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) { - be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx]; - int cur_tp_idx; - unsigned t; - - /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */ - - /* check if node can be executed on this unit type */ - cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node); - if (cur_tp_idx < 0) - continue; + stat_ev_tim_push(); + /* for each node */ + foreach_pset(ba->livein_nodes, livein) { + ir_node *irn = livein->irn; + be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn); + ilpsched_node_attr_t *na = get_ilpsched_node_attr(node); + unsigned t; - /* check all time_steps */ - for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) { - int cst, y_idx; - ir_node *irn; - int *tmp_idx_1 = NEW_ARR_F(int, 0); - int *tmp_idx_m1 = NEW_ARR_F(int, 0); + /* check check all time steps: 0 <= t < max_alive_steps */ + for (t = 0; t < livein->max_alive_steps; ++t) { + int node_tp_idx; - snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name); - cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1)); - DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf)); - num_cst++; + /* for all unit types available for this node */ + for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) { + const ir_edge_t *edge; + unsigned idx; + int cst, num_block_user; + int *tmp_var_idx_m = NEW_ARR_F(int, 0); - /* - - accumulate all nodes scheduled on unit type k till t - - subtract all nodes died on unit type k till t - */ - foreach_linked_irns(ba->head_ilp_nodes, irn) { - be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn); - ilpsched_node_attr_t *na = get_ilpsched_node_attr(node); - unsigned tn, tmax; + /* check the number of consumer scheduled so far */ + num_block_user = 0; + foreach_out_edge(irn, edge) { + ir_node *user = get_edge_src_irn(edge); + be_ilpsched_irn_t *cons; + ilpsched_node_attr_t *ca; int tp_idx; + unsigned tm, tm_max; - tmax = MIN(t, na->alap - 1); - tp_idx = is_valid_unit_type_for_node(cur_tp, node); - - /* current unit type is not suitable for current node */ - if (tp_idx < 0) + /* check only users within current block */ + if (get_nodes_block(user) != block_node->irn) continue; - for (tn = na->asap - 1; tn <= tmax; ++tn) { - int idx; - - /* node scheduled */ - idx = ILPVAR_IDX(na, tp_idx, tn); - ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]); + num_block_user++; + cons = get_ilpsched_irn(env, user); + ca = get_ilpsched_node_attr(cons); - /* node dead */ - idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn); - ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]); + tm_max = MIN(ca->alap - 1, t); + for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) { + for (tm = ca->asap - 1; tm <= tm_max; ++tm) { + int idx = ILPVAR_IDX(ca, tp_idx, tm); + ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]); + } } } - if (ARR_LEN(tmp_idx_1) > 0) - lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0); - - if (ARR_LEN(tmp_idx_m1) > 0) - lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0); + snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s", + t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name); + cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user); + DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf)); + num_cst++; - /* BEWARE: t is unsigned, so (double)(-t) won't work */ - y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t); - lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(t)); + /* sum(scheduled users) */ + if (ARR_LEN(tmp_var_idx_m) > 0) + lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0); + DEL_ARR_F(tmp_var_idx_m); - DEL_ARR_F(tmp_idx_1); - DEL_ARR_F(tmp_idx_m1); + /* + c * a_{nt}^k */ + idx = node_tp_idx * livein->max_alive_steps + t; + lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user)); } } } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n", - num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0)); + stat_ev_tim_pop("beilpsched_cst_alive_livein_nodes"); } /** @@ -1532,16 +1555,14 @@ static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, * - add additional costs to objective function if a node is scheduled * on a unit although all units of this type are currently occupied */ -static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ char buf[1024]; ir_node *cur_irn; unsigned num_cst = 0; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints"); -#endif /* WITH_LIBCORE */ - ilp_timer_push(t_cst); + stat_ev_tim_push(); /* y_{nt}^k is set for each node and timestep and unit type */ foreach_linked_irns(ba->head_ilp_nodes, cur_irn) { unsigned cur_idx = get_irn_idx(cur_irn); @@ -1571,6 +1592,7 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, int cst, y_idx; ir_node *irn; int *tmp_var_idx = NEW_ARR_F(int, 0); + ilp_livein_node_t *livein; snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name); cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1)); @@ -1596,6 +1618,25 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t); ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]); } + /* do the same for livein nodes */ + foreach_pset(ba->livein_nodes, livein) { + ir_node *irn = livein->irn; + be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn); + int a_idx, tp_idx; + + /* check if node can be alive here */ + if (t >= livein->max_alive_steps) + continue; + + tp_idx = is_valid_unit_type_for_node(cur_tp, node); + + /* current type is not suitable */ + if (tp_idx < 0) + continue; + + a_idx = tp_idx * livein->max_alive_steps + t; + ARR_APP1(int, tmp_var_idx, livein->a[a_idx]); + } if (ARR_LEN(tmp_var_idx) > 0) lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0); @@ -1603,26 +1644,116 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, /* - num_nodes * y_{nt}^k */ y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t); - lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(ba->n_interesting_nodes)); + lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0); + } + } + } + stat_ev_tim_pop("beilpsched_cst_pressure"); +} + +/** + * Create ILP branch constraints: + * Assure, alle nodes are scheduled prior to cfg op. + */ +static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ + char buf[1024]; + ir_node *cur_irn, *cfop; + unsigned num_cst = 0; + unsigned num_non_branches = 0; + ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); + + stat_ev_tim_push(); + cfop = NULL; + /* determine number of non-branch nodes and the one and only branch node */ + foreach_linked_irns(ba->head_ilp_nodes, cur_irn) { + switch (get_irn_opcode(cur_irn)) { + case iro_Phi: + case beo_Start: + case iro_End: + case iro_Proj: + case iro_Bad: + case iro_Unknown: + num_non_branches++; + break; + default: + if (is_cfop(cur_irn)) { + assert(cfop == NULL && "Highlander - there can be only one to be constrained"); + cfop = cur_irn; + } + else { + num_non_branches++; + } + break; + } + } + + if (cfop) { + be_ilpsched_irn_t *cf_node = get_ilpsched_irn(env, cfop); + ilpsched_node_attr_t *cf_na = get_ilpsched_node_attr(cf_node); + unsigned t; + + /* for each time step */ + for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) { + int *non_branch_vars, *branch_vars; + int cst; + + snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop)); + cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0); + DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf)); + num_cst++; + + /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */ + non_branch_vars = NEW_ARR_F(int, 0); + branch_vars = NEW_ARR_F(int, 0); + foreach_linked_irns(ba->head_ilp_nodes, cur_irn) { + be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn); + ilpsched_node_attr_t *na = get_ilpsched_node_attr(node); + int tp_idx; + + if (cur_irn == cfop) { + /* for all unit types available for this node */ + for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) { + unsigned idx = ILPVAR_IDX(na, tp_idx, t); + ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]); + } + } + else { + /* sum up all possible schedule points for this node upto current timestep */ + for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) { + unsigned tn; + unsigned tmax = MIN(t, na->alap - 1); + + for (tn = na->asap - 1; tn <= tmax; ++tn) { + unsigned idx = ILPVAR_IDX(na, tp_idx, tn); + ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]); + } + } + } + } + + if (ARR_LEN(non_branch_vars) > 0) + lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0); + if (ARR_LEN(branch_vars) > 0) + lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches); + + DEL_ARR_F(branch_vars); + DEL_ARR_F(non_branch_vars); } } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n", - num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0)); + stat_ev_tim_pop("beilpsched_cst_branch"); } #if 0 -static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) { +static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) +{ char buf[1024]; ir_node *irn; unsigned num_cst = 0; ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); -#ifdef WITH_LIBCORE - lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints"); -#endif /* WITH_LIBCORE */ - ilp_timer_push(t_cst); + stat_ev_tim_push(); /* check all nodes */ foreach_linked_irns(ba->head_ilp_nodes, irn) { be_ilpsched_irn_t *node; @@ -1682,11 +1813,9 @@ static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ } } } - ilp_timer_pop(); - DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n", - num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0)); + stat_ev_tim_pop("beilpsched_cst_projkeep"); } -#endif +#endif /* if 0 */ /*************************************************** * _____ _ _____ _ @@ -1701,12 +1830,14 @@ static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ /** * Create the ilp (add variables, build constraints, solve, build schedule from solution). */ -static void create_ilp(ir_node *block, void *walk_env) { +static void create_ilp(ir_node *block, void *walk_env) +{ be_ilpsched_env_t *env = walk_env; be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block); ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node); FILE *logfile = NULL; lpp_t *lpp = NULL; + int need_heur = 0; struct obstack var_obst; char name[1024]; @@ -1727,9 +1858,11 @@ static void create_ilp(ir_node *block, void *walk_env) { int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes; int estimated_n_var = (int)((double)base_num * fact_var); int estimated_n_cst = (int)((double)base_num * fact_cst); + be_options_t *options = be_get_irg_options(env->irg); DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n", estimated_n_var, estimated_n_cst)); + (void) estimated_n_var; /* set up the LPP object */ snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg))); @@ -1750,22 +1883,19 @@ static void create_ilp(ir_node *block, void *walk_env) { create_assignment_and_precedence_constraints(env, lpp, block_node); create_ressource_constraints(env, lpp, block_node); create_bundle_constraints(env, lpp, block_node); + create_branch_constraint(env, lpp, block_node); //create_proj_keep_constraints(env, lpp, block_node); if (env->opts->regpress) { - if (ba->n_interesting_nodes > env->opts->limit_dead) { - create_alive_nodes_constraint(env, lpp, block_node); - create_pressure_alive_constraint(env, lpp, block_node); - } else { - create_dying_nodes_constraint(env, lpp, block_node); - create_pressure_dead_constraint(env, lpp, block_node); - } + create_alive_nodes_constraint(env, lpp, block_node); + create_alive_livein_nodes_constraint(env, lpp, block_node); + create_pressure_alive_constraint(env, lpp, block_node); } - DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next)); + DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp_get_var_count(lpp), lpp_get_cst_count(lpp))); /* debug stuff, dump lpp when debugging is on */ - DEBUG_ONLY( + DEBUG_ONLY({ if (firm_dbg_get_mask(env->dbg) > 1) { char buf[1024]; FILE *f; @@ -1777,7 +1907,7 @@ static void create_ilp(ir_node *block, void *walk_env) { snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block)); lpp_dump(lpp, buf); } - ); + }) /* set solve time limit */ lpp_set_time_limit(lpp, env->opts->time_limit); @@ -1798,41 +1928,65 @@ static void create_ilp(ir_node *block, void *walk_env) { } /* solve the ILP */ - lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver); + lpp_solve_net(lpp, options->ilp_server, options->ilp_solver); if (logfile) fclose(logfile); /* check for valid solution */ if (! lpp_is_sol_valid(lpp)) { - char buf[1024]; - FILE *f; - - snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block)); - f = fopen(buf, "w"); - lpp_dump_plain(lpp, f); - fclose(f); - snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block)); - lpp_dump(lpp, buf); - dump_ir_block_graph(env->irg, "-assert"); - - assert(0 && "ILP solution is not feasible!"); + DEBUG_ONLY({ + char buf[1024]; + FILE *f; + + if (firm_dbg_get_mask(env->dbg) >= 2) { + snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block)); + f = fopen(buf, "w"); + lpp_dump_plain(lpp, f); + fclose(f); + snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block)); + lpp_dump(lpp, buf); + dump_ir_graph(env->irg, "infeasible"); + } + }) + + ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg); + need_heur = 1; } DBG((env->dbg, LEVEL_1, "\nSolution:\n")); - DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0)); - DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0)); - DBG((env->dbg, LEVEL_1, "\tmatrix: %u elements, density %.2f%%, size %.2fMB\n", lpp->n_elems, lpp->density, (double)lpp->matrix_mem / 1024.0 / 1024.0)); - DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations)); - DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time)); - DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval)); - DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound)); - - DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst))); + //DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0)); + //DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0)); + //DBG((env->dbg, LEVEL_1, "\tmatrix: %u elements, density %.2f%%, size %.2fMB\n", lpp->n_elems, lpp->density, (double)lpp->matrix_mem / 1024.0 / 1024.0)); + DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp_get_iter_cnt(lpp))); + DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp_get_sol_time(lpp))); + //DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval)); + //DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound)); + + //DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst))); } /* apply solution */ - apply_solution(env, lpp, block); + be_stat_ev("nodes", ba->block_last_idx); + be_stat_ev("vars", lpp ? lpp_get_var_count(lpp) : 0); + be_stat_ev("csts", lpp ? lpp_get_cst_count(lpp) : 0); + if (need_heur) { + be_stat_ev("time", -1); + be_stat_ev_dbl("opt", 0.0); + list_sched_single_block(env->irg, block); + } else { +#if 0 + if (lpp) { + double opt = lpp_get_sol_state(lpp) == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval; + be_stat_ev_dbl("time", lpp->sol_time); + be_stat_ev_dbl("opt", opt); + } else { + be_stat_ev_dbl("time", 0.0); + be_stat_ev_dbl("opt", 100.0); + } +#endif + apply_solution(env, lpp, block); + } if (lpp) free_lpp(lpp); @@ -1844,25 +1998,24 @@ static void create_ilp(ir_node *block, void *walk_env) { /** * Perform ILP scheduling on the given irg. */ -void be_ilp_sched(const be_irg_t *birg) { +void be_ilp_sched(ir_graph *irg) +{ be_ilpsched_env_t env; - const char *name = "be ilp scheduling"; - arch_isa_t *isa = birg->main_env->arch_env->isa; - const ilp_sched_selector_t *sel = isa->impl->get_ilp_sched_selector(isa); + const arch_env_t *arch_env = be_get_irg_arch_env(irg); + const ilp_sched_selector_t *sel = arch_env->impl->get_ilp_sched_selector(arch_env); FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp"); -// firm_dbg_set_mask(env.dbg, 1); + stat_ev_ctx_push("ilpsched"); - env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg); + env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, irg); env.sel = sel; - env.irg = birg->irg; - env.height = heights_new(birg->irg); - env.main_env = birg->main_env; - env.arch_env = birg->main_env->arch_env; - env.cpu = arch_isa_get_machine(birg->main_env->arch_env->isa); + env.irg = irg; + env.height = heights_new(irg); + env.arch_env = arch_env; + env.cpu = arch_env_get_machine(arch_env); env.opts = &ilp_opts; - phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn); + phase_init(&env.ph, env.irg, init_ilpsched_irn); /* assign a unique per block number to all interesting nodes */ irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env); @@ -1884,7 +2037,7 @@ void be_ilp_sched(const be_irg_t *birg) { irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env); /* perform ILP scheduling */ - irg_block_walk_graph(env.irg, clear_unwanted_data, create_ilp, &env); + irg_block_walk_graph(env.irg, NULL, create_ilp, &env); DEBUG_ONLY( if (firm_dbg_get_mask(env.dbg)) { @@ -1895,18 +2048,23 @@ void be_ilp_sched(const be_irg_t *birg) { } ); + /* free data allocated dynamically */ + irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env); + /* free all allocated object */ - phase_free(&env.ph); + phase_deinit(&env.ph); heights_free(env.height); /* notify backend */ - be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env); + be_ilp_sched_finish_irg_ilp_schedule(sel, irg, env.irg_env); + + stat_ev_ctx_pop("ilpsched"); } -#ifdef WITH_LIBCORE /** * Register ILP scheduler options. */ +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_ilpsched); void be_init_ilpsched(void) { lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); @@ -1914,11 +2072,10 @@ void be_init_ilpsched(void) lc_opt_add_table(sched_grp, ilpsched_option_table); } -#endif /* WITH_LIBCORE */ #else /* WITH_ILP */ -static INLINE void some_picky_compiler_do_not_allow_empty_files(void) +static inline void some_picky_compiler_do_not_allow_empty_files(void) {} #endif /* WITH_ILP */