improve fix some statistics in bechordal
[libfirm] / ir / be / bechordal_main.c
index 7091153..5e2f38f 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -38,9 +38,8 @@
 #include "iterator.h"
 #include "firm_config.h"
 
-#include <libcore/lc_opts.h>
-#include <libcore/lc_opts_enum.h>
-#include <libcore/lc_timing.h>
+#include "lc_opts.h"
+#include "lc_opts_enum.h"
 
 #include "ircons_t.h"
 #include "irmode_t.h"
@@ -77,8 +76,6 @@
 #include "bera.h"
 #include "beirg_t.h"
 
-#include "bespillbelady.h"
-#include "bespillmorgan.h"
 #include "bespillslots.h"
 #include "bespilloptions.h"
 #include "belower.h"
@@ -109,19 +106,7 @@ typedef struct _post_spill_env_t {
        double                      pre_spill_cost;
 } post_spill_env_t;
 
-static be_ra_timer_t ra_timer = {
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-};
+static be_options_t  *main_opts;
 
 static const lc_opt_enum_int_items_t lower_perm_items[] = {
        { "copy", BE_CH_LOWER_PERM_COPY },
@@ -252,8 +237,6 @@ typedef struct _node_stat_t {
 struct node_stat_walker {
        node_stat_t      *stat;
        const arch_env_t *arch_env;
-       bitset_t         *mem_phis;
-       const arch_register_class_t *cls;
 };
 
 static void node_stat_walker(ir_node *irn, void *data)
@@ -261,57 +244,40 @@ static void node_stat_walker(ir_node *irn, void *data)
        struct node_stat_walker *env  = data;
        const arch_env_t        *aenv = env->arch_env;
 
-       if (arch_irn_consider_in_reg_alloc(aenv, env->cls, irn)) {
-
-               /* if the node is a normal phi */
-               if(is_Phi(irn))
+       /* if the node is a normal phi */
+       if(is_Phi(irn)) {
+               if (get_irn_mode(irn) == mode_M) {
+                       env->stat->n_mem_phis++;
+               } else {
                        env->stat->n_phis++;
+               }
+       } else {
+               arch_irn_class_t classify = arch_irn_classify(aenv, irn);
 
-               else if(arch_irn_classify(aenv, irn) & arch_irn_class_spill)
+               if(classify & arch_irn_class_spill)
                        ++env->stat->n_spills;
-
-               else if(arch_irn_classify(aenv, irn) & arch_irn_class_reload)
+               if(classify & arch_irn_class_reload)
                        ++env->stat->n_reloads;
-
-               else if(arch_irn_classify(aenv, irn) & arch_irn_class_copy)
+               if(classify & arch_irn_class_copy)
                        ++env->stat->n_copies;
-
-               else if(arch_irn_classify(aenv, irn) & arch_irn_class_perm)
+               if(classify & arch_irn_class_perm)
                        ++env->stat->n_perms;
        }
-
-       /* a mem phi is a PhiM with a mem phi operand or a Spill operand */
-       else if(is_Phi(irn) && get_irn_mode(irn) == mode_M) {
-               int i;
-
-               for(i = get_irn_arity(irn) - 1; i >= 0; --i) {
-                       ir_node *op = get_irn_n(irn, i);
-
-                       if((is_Phi(op) && bitset_contains_irn(env->mem_phis, op)) || (arch_irn_classify(aenv, op) & arch_irn_class_spill)) {
-                               bitset_add_irn(env->mem_phis, irn);
-                               env->stat->n_mem_phis++;
-                               break;
-                       }
-               }
-       }
 }
 
-static void node_stats(be_irg_t *birg, const arch_register_class_t *cls, node_stat_t *stat)
+static void node_stats(be_irg_t *birg, node_stat_t *stat)
 {
        struct node_stat_walker env;
 
-       memset(stat, 0, sizeof(stat[0]));
+       memset(stat, 0, sizeof(*stat));
        env.arch_env = birg->main_env->arch_env;
-       env.mem_phis = bitset_irg_malloc(birg->irg);
        env.stat     = stat;
-       env.cls      = cls;
        irg_walk_graph(birg->irg, NULL, node_stat_walker, &env);
-       bitset_free(env.mem_phis);
 }
 
 static void insn_count_walker(ir_node *irn, void *data)
 {
-       int *cnt = data;
+       unsigned long *cnt = data;
 
        switch(get_irn_opcode(irn)) {
        case iro_Proj:
@@ -324,99 +290,58 @@ static void insn_count_walker(ir_node *irn, void *data)
        }
 }
 
-static unsigned int count_insns(ir_graph *irg)
+static unsigned long count_insns(ir_graph *irg)
 {
-       int cnt = 0;
+       unsigned long cnt = 0;
        irg_walk_graph(irg, insn_count_walker, NULL, &cnt);
        return cnt;
 }
 
-/**
- * Initialize all timers.
- */
-static void be_init_timer(be_options_t *main_opts)
+static void block_count_walker(ir_node *node, void *data)
 {
-       if (main_opts->timing == BE_TIME_ON) {
-               ra_timer.t_prolog     = lc_timer_register("ra_prolog",     "regalloc prolog");
-               ra_timer.t_epilog     = lc_timer_register("ra_epilog",     "regalloc epilog");
-               ra_timer.t_live       = lc_timer_register("ra_liveness",   "be liveness");
-               ra_timer.t_spill      = lc_timer_register("ra_spill",      "spiller");
-               ra_timer.t_spillslots = lc_timer_register("ra_spillslots", "spillslots");
-               ra_timer.t_color      = lc_timer_register("ra_color",      "graph coloring");
-               ra_timer.t_ifg        = lc_timer_register("ra_ifg",        "interference graph");
-               ra_timer.t_copymin    = lc_timer_register("ra_copymin",    "copy minimization");
-               ra_timer.t_ssa        = lc_timer_register("ra_ssadestr",   "ssa destruction");
-               ra_timer.t_verify     = lc_timer_register("ra_verify",     "graph verification");
-               ra_timer.t_other      = lc_timer_register("ra_other",      "other time");
-
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_prolog);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_epilog);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_live);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_spill);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_spillslots);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_color);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_ifg);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_copymin);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_ssa);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_verify);
-               LC_STOP_AND_RESET_TIMER(ra_timer.t_other);
-
-               global_ra_timer = &ra_timer;
-       }
+       unsigned long *cnt = data;
+       if (node == get_irg_end_block(current_ir_graph))
+               return;
+       (*cnt)++;
 }
 
-#define BE_TIMER_INIT(main_opts)       be_init_timer(main_opts)
-
-#define BE_TIMER_PUSH(timer)                                                            \
-       if (main_opts->timing == BE_TIME_ON) {                                              \
-               if (! lc_timer_push(timer)) {                                                   \
-                       if (options.vrfy_option == BE_CH_VRFY_ASSERT)                               \
-                               assert(!"Timer already on stack, cannot be pushed twice.");             \
-                       else if (options.vrfy_option == BE_CH_VRFY_WARN)                            \
-                               fprintf(stderr, "Timer %s already on stack, cannot be pushed twice.\n", \
-                                       lc_timer_get_name(timer));                                          \
-               }                                                                               \
-       }
-#define BE_TIMER_POP(timer)                                                                    \
-       if (main_opts->timing == BE_TIME_ON) {                                                     \
-               lc_timer_t *tmp = lc_timer_pop();                                                      \
-               if (options.vrfy_option == BE_CH_VRFY_ASSERT)                                          \
-                       assert(tmp == timer && "Attempt to pop wrong timer.");                             \
-               else if (options.vrfy_option == BE_CH_VRFY_WARN && tmp != timer)                       \
-                       fprintf(stderr, "Attempt to pop wrong timer. %s is on stack, trying to pop %s.\n", \
-                               lc_timer_get_name(tmp), lc_timer_get_name(timer));                             \
-               timer = tmp;                                                                           \
-       }
+static unsigned long count_blocks(ir_graph *irg)
+{
+       unsigned long cnt = 0;
+       irg_block_walk_graph(irg, block_count_walker, NULL, &cnt);
+       return cnt;
+}
+
+static node_stat_t last_node_stat;
 
 /**
  * Perform things which need to be done per register class before spilling.
  */
-static void pre_spill(const arch_isa_t *isa, int cls_idx, post_spill_env_t *pse) {
+static void pre_spill(post_spill_env_t *pse, const arch_register_class_t *cls)
+{
        be_chordal_env_t    *chordal_env = &pse->cenv;
        be_irg_t            *birg        = pse->birg;
        ir_graph            *irg         = be_get_birg_irg(birg);
        const be_main_env_t *main_env    = birg->main_env;
-       node_stat_t          node_stat;
 
-       pse->cls                   = arch_isa_get_reg_class(isa, cls_idx);
-       chordal_env->cls           = pse->cls;
+       pse->cls                   = cls;
+       chordal_env->cls           = cls;
        chordal_env->border_heads  = pmap_create();
        chordal_env->ignore_colors = bitset_malloc(chordal_env->cls->n_regs);
 
        be_assure_liveness(birg);
        be_liveness_assure_chk(be_get_birg_liveness(birg));
-       stat_ev_ctx_push("cls", pse->cls->name);
-       stat_ev_do(node_stats(birg, pse->cls, &node_stat));
+
        stat_ev_do(pse->pre_spill_cost = be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq));
-       stat_ev_dbl("phis_before_spill", node_stat.n_phis);
 
        /* put all ignore registers into the ignore register set. */
        be_put_ignore_regs(birg, pse->cls, chordal_env->ignore_colors);
 
+       BE_TIMER_PUSH(t_ra_constr);
        be_pre_spill_prepare_constr(chordal_env);
-       dump(BE_CH_DUMP_CONSTR, birg->irg, pse->cls, "-constr-pre", dump_ir_block_graph_sched);
+       BE_TIMER_POP(t_ra_constr);
 
-       stat_ev_ctx_pop();
+       dump(BE_CH_DUMP_CONSTR, birg->irg, pse->cls, "-constr-pre", dump_ir_block_graph_sched);
 }
 
 /**
@@ -427,97 +352,86 @@ static void post_spill(post_spill_env_t *pse, int iteration) {
        be_irg_t            *birg        = pse->birg;
        ir_graph            *irg         = birg->irg;
        const be_main_env_t *main_env    = birg->main_env;
-       be_options_t        *main_opts   = main_env->options;
-       node_stat_t         node_stat;
-       int                 colors_n     = arch_register_class_n_regs(chordal_env->cls);
+       int                  colors_n     = arch_register_class_n_regs(chordal_env->cls);
        int             allocatable_regs = colors_n - be_put_ignore_regs(birg, chordal_env->cls, NULL);
 
        /* some special classes contain only ignore regs, no work to be done */
        if (allocatable_regs > 0) {
-
-               stat_ev_ctx_push("cls", pse->cls->name);
-               stat_ev_do(node_stats(birg, pse->cls, &node_stat));
-               stat_ev_dbl("phis_after_spill", node_stat.n_phis);
-               stat_ev_dbl("mem_phis", node_stat.n_mem_phis);
-               stat_ev_dbl("reloads", node_stat.n_reloads);
-               stat_ev_dbl("spills", node_stat.n_spills);
-               stat_ev_dbl("spillcosts", be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq) - pse->pre_spill_cost);
+               stat_ev_dbl("bechordal_spillcosts", be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq) - pse->pre_spill_cost);
 
                /*
                        If we have a backend provided spiller, post spill is
                        called in a loop after spilling for each register class.
                        But we only need to fix stack nodes once in this case.
                */
+               BE_TIMER_PUSH(t_ra_spill);
+               check_for_memory_operands(chordal_env);
                if (iteration == 0) {
-                       check_for_memory_operands(chordal_env);
                        be_abi_fix_stack_nodes(birg->abi);
                }
+               BE_TIMER_POP(t_ra_spill);
 
-               BE_TIMER_PUSH(ra_timer.t_verify);
+               BE_TIMER_PUSH(t_verify);
 
                /* verify schedule and register pressure */
                if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
                        be_verify_schedule(birg);
                        be_verify_register_pressure(birg, pse->cls, irg);
-               }
-               else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
+               } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
                        assert(be_verify_schedule(birg) && "Schedule verification failed");
                        assert(be_verify_register_pressure(birg, pse->cls, irg)
                                && "Register pressure verification failed");
                }
-               BE_TIMER_POP(ra_timer.t_verify);
+               BE_TIMER_POP(t_verify);
 
                /* Color the graph. */
-               BE_TIMER_PUSH(ra_timer.t_color);
+               BE_TIMER_PUSH(t_ra_color);
                be_ra_chordal_color(chordal_env);
-               BE_TIMER_POP(ra_timer.t_color);
+               BE_TIMER_POP(t_ra_color);
 
                dump(BE_CH_DUMP_CONSTR, irg, pse->cls, "-color", dump_ir_block_graph_sched);
 
                /* Create the ifg with the selected flavor */
-               BE_TIMER_PUSH(ra_timer.t_ifg);
+               BE_TIMER_PUSH(t_ra_ifg);
                chordal_env->ifg = be_create_ifg(chordal_env);
-               BE_TIMER_POP(ra_timer.t_ifg);
+               BE_TIMER_POP(t_ra_ifg);
 
-               {
+               stat_ev_if {
                        be_ifg_stat_t stat;
-
-                       stat_ev_do(be_ifg_stat(birg, chordal_env->ifg, &stat));
-                       stat_ev_dbl("ifg_nodes", stat.n_nodes);
-                       stat_ev_dbl("ifg_edges", stat.n_edges);
-                       stat_ev_dbl("ifg_comps", stat.n_comps);
-
-                       stat_ev_do(node_stats(birg, pse->cls, &node_stat));
-                       stat_ev_dbl("perms_before_coal", node_stat.n_perms);
-                       stat_ev_dbl("copies_before_coal", node_stat.n_copies);
+                       node_stat_t   node_stat;
+
+                       be_ifg_stat(birg, chordal_env->ifg, &stat);
+                       stat_ev_dbl("bechordal_ifg_nodes", stat.n_nodes);
+                       stat_ev_dbl("bechordal_ifg_edges", stat.n_edges);
+                       stat_ev_dbl("bechordal_ifg_comps", stat.n_comps);
+
+                       node_stats(birg, &node_stat);
+                       stat_ev_dbl("bechordal_perms_before_coal",
+                                       node_stat.n_perms - last_node_stat.n_perms);
+                       stat_ev_dbl("bechordal_copies_before_coal",
+                                       node_stat.n_copies - last_node_stat.n_copies);
                }
 
                /* copy minimization */
-               BE_TIMER_PUSH(ra_timer.t_copymin);
+               BE_TIMER_PUSH(t_ra_copymin);
                co_driver(chordal_env);
-               BE_TIMER_POP(ra_timer.t_copymin);
+               BE_TIMER_POP(t_ra_copymin);
 
                dump(BE_CH_DUMP_COPYMIN, irg, pse->cls, "-copymin", dump_ir_block_graph_sched);
 
-               BE_TIMER_PUSH(ra_timer.t_ssa);
 
                /* ssa destruction */
+               BE_TIMER_PUSH(t_ra_ssa);
                be_ssa_destruction(chordal_env);
-
-               BE_TIMER_POP(ra_timer.t_ssa);
+               BE_TIMER_POP(t_ra_ssa);
 
                dump(BE_CH_DUMP_SSADESTR, irg, pse->cls, "-ssadestr", dump_ir_block_graph_sched);
 
-               BE_TIMER_PUSH(ra_timer.t_verify);
                if (chordal_env->opts->vrfy_option != BE_CH_VRFY_OFF) {
+                       BE_TIMER_PUSH(t_verify);
                        be_ssa_destruction_check(chordal_env);
+                       BE_TIMER_POP(t_verify);
                }
-               BE_TIMER_POP(ra_timer.t_verify);
-
-               stat_ev_do(node_stats(birg, pse->cls, &node_stat));
-               stat_ev_dbl("perms_after_coal", node_stat.n_perms);
-               stat_ev_dbl("copies_after_coal", node_stat.n_copies);
-               stat_ev_ctx_pop();
 
                /* the ifg exists only if there are allocatable regs */
                be_ifg_free(chordal_env->ifg);
@@ -536,19 +450,19 @@ static void post_spill(post_spill_env_t *pse, int iteration) {
  */
 static void be_ra_chordal_main(be_irg_t *birg)
 {
-       const be_main_env_t *main_env  = birg->main_env;
-       const arch_isa_t    *isa       = arch_env_get_isa(main_env->arch_env);
-       ir_graph            *irg       = birg->irg;
-       be_options_t        *main_opts = main_env->options;
+       const be_main_env_t *main_env = birg->main_env;
+       const arch_env_t    *arch_env = main_env->arch_env;
+       ir_graph            *irg      = birg->irg;
        int                 j, m;
        be_chordal_env_t    chordal_env;
        struct obstack      obst;
 
-       BE_TIMER_INIT(main_opts);
-       BE_TIMER_PUSH(ra_timer.t_other);
-       BE_TIMER_PUSH(ra_timer.t_prolog);
+       main_opts = main_env->options;
+
+       BE_TIMER_PUSH(t_ra_other);
+
+       BE_TIMER_PUSH(t_ra_prolog);
 
-       be_assure_dom_front(birg);
        be_assure_liveness(birg);
 
        chordal_env.obst          = &obst;
@@ -561,53 +475,80 @@ static void be_ra_chordal_main(be_irg_t *birg)
 
        obstack_init(&obst);
 
-       BE_TIMER_POP(ra_timer.t_prolog);
-
-       be_stat_ev("insns_before", count_insns(irg));
-
+       BE_TIMER_POP(t_ra_prolog);
 
+       stat_ev_if {
+               be_stat_ev("bechordal_insns_before", count_insns(irg));
+               be_stat_ev("bechordal_blocks_before", count_blocks(irg));
+               node_stats(birg, &last_node_stat);
+       }
 
        if (! arch_code_generator_has_spiller(birg->cg)) {
                /* use one of the generic spiller */
 
                /* Perform the following for each register class. */
-               for (j = 0, m = arch_isa_get_n_reg_class(isa); j < m; ++j) {
+               for (j = 0, m = arch_env_get_n_reg_class(arch_env); j < m; ++j) {
                        post_spill_env_t pse;
+                       const arch_register_class_t *cls
+                               = arch_env_get_reg_class(arch_env, j);
+
+                       if(arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
+                               continue;
+
+
+                       stat_ev_ctx_push_str("bechordal_cls", cls->name);
 
                        memcpy(&pse.cenv, &chordal_env, sizeof(chordal_env));
                        pse.birg = birg;
-                       pre_spill(isa, j, &pse);
-
-                       /* this is a hack, TODO remove me later */
-                       if(j == 2) {
-                               be_do_stat_reg_pressure(birg);
-                       }
+                       pre_spill(&pse, cls);
 
-                       BE_TIMER_PUSH(ra_timer.t_spill);
-                       be_do_spill(birg, pse.cls);
-                       BE_TIMER_POP(ra_timer.t_spill);
+                       BE_TIMER_PUSH(t_ra_spill);
+                       be_do_spill(birg, cls);
+                       BE_TIMER_POP(t_ra_spill);
 
-                       dump(BE_CH_DUMP_SPILL, irg, pse.cls, "-spill", dump_ir_block_graph_sched);
+                       dump(BE_CH_DUMP_SPILL, irg, pse.cls, "-spill",
+                            dump_ir_block_graph_sched);
 
                        post_spill(&pse, 0);
+
+                       stat_ev_if {
+                               node_stat_t node_stat;
+
+                               node_stats(birg, &node_stat);
+                               stat_ev_dbl("bechordal_phis",
+                                               node_stat.n_phis - last_node_stat.n_phis);
+                               stat_ev_dbl("bechordal_mem_phis",
+                                               node_stat.n_mem_phis - last_node_stat.n_mem_phis);
+                               stat_ev_dbl("bechordal_reloads",
+                                               node_stat.n_reloads - last_node_stat.n_reloads);
+                               stat_ev_dbl("bechordal_spills",
+                                               node_stat.n_spills - last_node_stat.n_spills);
+                               stat_ev_dbl("bechordal_perms_after_coal",
+                                               node_stat.n_perms - last_node_stat.n_perms);
+                               stat_ev_dbl("bechordal_copies_after_coal",
+                                               node_stat.n_copies - last_node_stat.n_copies);
+
+                               last_node_stat = node_stat;
+                               stat_ev_ctx_pop("bechordal_cls");
+                       }
                }
        } else {
                post_spill_env_t *pse;
 
-               /* the backend has it's own spiller */
-               m = arch_isa_get_n_reg_class(isa);
+               /* the backend has its own spiller */
+               m = arch_env_get_n_reg_class(arch_env);
 
                pse = alloca(m * sizeof(pse[0]));
 
                for (j = 0; j < m; ++j) {
                        memcpy(&pse[j].cenv, &chordal_env, sizeof(chordal_env));
                        pse[j].birg = birg;
-                       pre_spill(isa, j, &pse[j]);
+                       pre_spill(&pse[j], pse[j].cls);
                }
 
-               BE_TIMER_PUSH(ra_timer.t_spill);
+               BE_TIMER_PUSH(t_ra_spill);
                arch_code_generator_spill(birg->cg, birg);
-               BE_TIMER_POP(ra_timer.t_spill);
+               BE_TIMER_POP(t_ra_spill);
                dump(BE_CH_DUMP_SPILL, irg, NULL, "-spill", dump_ir_block_graph_sched);
 
                for (j = 0; j < m; ++j) {
@@ -615,22 +556,23 @@ static void be_ra_chordal_main(be_irg_t *birg)
                }
        }
 
-
+       BE_TIMER_PUSH(t_verify);
        be_verify_register_allocation(birg);
+       BE_TIMER_POP(t_verify);
 
-       BE_TIMER_PUSH(ra_timer.t_epilog);
+       BE_TIMER_PUSH(t_ra_epilog);
        lower_nodes_after_ra(birg, options.lower_perm_opt & BE_CH_LOWER_PERM_COPY ? 1 : 0);
        dump(BE_CH_DUMP_LOWER, irg, NULL, "-belower-after-ra", dump_ir_block_graph_sched);
 
        obstack_free(&obst, NULL);
        be_liveness_invalidate(be_get_birg_liveness(birg));
-       BE_TIMER_POP(ra_timer.t_epilog);
-
-       BE_TIMER_POP(ra_timer.t_other);
+       BE_TIMER_POP(t_ra_epilog);
 
-       be_stat_ev("insns_after", count_insns(irg));
+       BE_TIMER_POP(t_ra_other);
 
-       return;
+       stat_ev_if {
+               be_stat_ev("bechordal_insns_after", count_insns(irg));
+       }
 }
 
 static be_ra_t be_ra_chordal_allocator = {