fixed debug output of unary x87 nodes
[libfirm] / ir / be / bemain.c
index 4c8e887..c476f88 100644 (file)
@@ -28,7 +28,9 @@
 #include "iredges_t.h"
 #include "irloop_t.h"
 #include "irtools.h"
+#include "irvrfy.h"
 #include "return.h"
+#include "firmstat.h"
 
 #include "bearch.h"
 #include "firm/bearch_firm.h"
@@ -45,7 +47,6 @@
 #include "besched_t.h"
 #include "belistsched.h"
 #include "belive_t.h"
-#include "bespillilp.h"
 #include "bespillbelady.h"
 #include "bera.h"
 #include "beraextern.h"
@@ -65,6 +66,8 @@
 static be_options_t be_options = {
        DUMP_NONE,                         /* dump options */
        BE_TIME_OFF,                       /* no timing */
+       BE_SCHED_SELECT_HEUR,              /* mueller heuristic selector */
+       0,                                 /* disable mris */
        "i44pc52.info.uni-karlsruhe.de",   /* ilp server */
        "cplex"                            /* ilp solver */
 };
@@ -73,7 +76,7 @@ static be_options_t be_options = {
 static unsigned dump_flags = 0;
 
 /* verify options */
-static unsigned vrfy_option = BE_VRFY_WARN;
+static int vrfy_option = BE_VRFY_WARN;
 
 /* register allocator to use. */
 static const be_ra_t *ra = &be_ra_chordal_allocator;
@@ -81,9 +84,6 @@ static const be_ra_t *ra = &be_ra_chordal_allocator;
 /* back end instruction set architecture to use */
 static const arch_isa_if_t *isa_if = &ia32_isa_if;
 
-/* mris option */
-static int be_enable_mris = 0;
-
 #ifdef WITH_LIBCORE
 
 static lc_opt_entry_t *be_grp_root = NULL;
@@ -128,6 +128,15 @@ static const lc_opt_enum_int_items_t vrfy_items[] = {
        { NULL,     0 }
 };
 
+/* schedule selector options. */
+static const lc_opt_enum_int_items_t sched_select_items[] = {
+       { "isa",      BE_SCHED_SELECT_ISA      },
+       { "muchnik",  BE_SCHED_SELECT_MUCHNIK  },
+       { "heur",     BE_SCHED_SELECT_HEUR     },
+       { "hmuchnik", BE_SCHED_SELECT_HMUCHNIK },
+       { NULL,     0 }
+};
+
 static lc_opt_enum_mask_var_t dump_var = {
        &dump_flags, dump_items
 };
@@ -144,14 +153,19 @@ static lc_opt_enum_int_var_t vrfy_var = {
        &vrfy_option, vrfy_items
 };
 
+static lc_opt_enum_int_var_t sched_select_var = {
+       &be_options.sched_select, sched_select_items
+};
+
 static const lc_opt_table_entry_t be_main_options[] = {
-       LC_OPT_ENT_ENUM_MASK("dump",     "dump irg on several occasions",     &dump_var),
-       LC_OPT_ENT_ENUM_PTR ("ra",       "register allocator",                &ra_var),
-       LC_OPT_ENT_ENUM_PTR ("isa",      "the instruction set architecture",  &isa_var),
-       LC_OPT_ENT_NEGBOOL  ("noomitfp", "do not omit frame pointer",         &be_omit_fp),
-       LC_OPT_ENT_BOOL     ("mris",     "enable mris schedule preparation",  &be_enable_mris),
-       LC_OPT_ENT_ENUM_PTR ("vrfy",     "verify the backend irg (off, warn, assert)",  &vrfy_var),
-       LC_OPT_ENT_BOOL     ("time",     "get backend timing statistics",     &be_options.timing),
+       LC_OPT_ENT_ENUM_MASK("dump",         "dump irg on several occasions",     &dump_var),
+       LC_OPT_ENT_ENUM_PTR ("ra",           "register allocator",                &ra_var),
+       LC_OPT_ENT_ENUM_PTR ("isa",          "the instruction set architecture",  &isa_var),
+       LC_OPT_ENT_NEGBOOL  ("noomitfp",     "do not omit frame pointer",         &be_omit_fp),
+       LC_OPT_ENT_ENUM_PTR ("vrfy",         "verify the backend irg (off, warn, assert)",  &vrfy_var),
+       LC_OPT_ENT_BOOL     ("time",         "get backend timing statistics",     &be_options.timing),
+       LC_OPT_ENT_BOOL     ("sched.mris",   "enable mris schedule preparation",  &be_options.mris),
+       LC_OPT_ENT_ENUM_PTR ("sched.select", "schedule node selector (isa, muchnik, heur, hmuchnik)",&sched_select_var),
 
 #ifdef WITH_ILP
        LC_OPT_ENT_STR ("ilp.server", "the ilp server name", be_options.ilp_server, sizeof(be_options.ilp_server)),
@@ -228,7 +242,6 @@ const backend_params *be_init(void)
        be_opt_register();
 
        be_sched_init();
-       be_liveness_init();
        be_numbering_init();
        be_copy_opt_init();
        copystat_init();
@@ -340,15 +353,7 @@ static void be_main_loop(FILE *file_handle)
        unsigned num_nodes_b = 0;
        unsigned num_nodes_a = 0;
        unsigned num_nodes_r = 0;
-       unsigned ra_prolog   = 0;
-       unsigned ra_epilog   = 0;
-       unsigned ra_live     = 0;
-       unsigned ra_spill    = 0;
-       unsigned ra_color    = 0;
-       unsigned ra_ifg      = 0;
-       unsigned ra_copymin  = 0;
-       unsigned ra_ssa      = 0;
-       lc_timer_t *t_prolog, *t_abi, *t_codegen, *t_sched, *t_constr, *t_regalloc, *t_finish, *t_emit, *t_other;
+       lc_timer_t *t_prolog, *t_abi, *t_codegen, *t_sched, *t_constr, *t_regalloc, *t_finish, *t_emit, *t_other, *t_verify;
        be_ra_timer_t *ra_timer;
 
        if (be_options.timing == BE_TIME_ON) {
@@ -360,6 +365,7 @@ static void be_main_loop(FILE *file_handle)
                t_regalloc = lc_timer_register("regalloc", "register allocation");
                t_finish   = lc_timer_register("finish",   "graph finish");
                t_emit     = lc_timer_register("emiter",   "code emiter");
+               t_verify   = lc_timer_register("verify",   "graph verification");
                t_other    = lc_timer_register("other",    "other");
        }
 
@@ -370,8 +376,26 @@ static void be_main_loop(FILE *file_handle)
        /* for debugging, anchors helps */
        // dump_all_anchors(1);
 
-#define BE_TIMER_PUSH(timer)  if (be_options.timing == BE_TIME_ON) lc_timer_push(timer)
-#define BE_TIMER_POP()        if (be_options.timing == BE_TIME_ON) lc_timer_pop()
+#define BE_TIMER_PUSH(timer)                                                        \
+       if (be_options.timing == BE_TIME_ON) {                                          \
+               int res = lc_timer_push(timer);                                             \
+               if (vrfy_option == BE_VRFY_ASSERT)                                          \
+                       assert(res && "Timer already on stack, cannot be pushed twice.");       \
+               else if (vrfy_option == BE_VRFY_WARN && ! res)                              \
+                       fprintf(stderr, "Timer %s already on stack, cannot be pushed twice.\n", \
+                               lc_timer_get_name(timer));                                          \
+       }
+#define BE_TIMER_POP(timer)                                                                    \
+       if (be_options.timing == BE_TIME_ON) {                                                     \
+               lc_timer_t *tmp = lc_timer_pop();                                                      \
+               if (vrfy_option == BE_VRFY_ASSERT)                                                     \
+                       assert(tmp == timer && "Attempt to pop wrong timer.");                             \
+               else if (vrfy_option == BE_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;                                                                           \
+       }
+
 #define BE_TIMER_ONLY(code)   if (be_options.timing == BE_TIME_ON) do { code; } while(0)
 
        /* For all graphs */
@@ -391,9 +415,10 @@ static void be_main_loop(FILE *file_handle)
                        LC_STOP_AND_RESET_TIMER(t_regalloc);
                        LC_STOP_AND_RESET_TIMER(t_finish);
                        LC_STOP_AND_RESET_TIMER(t_emit);
+                       LC_STOP_AND_RESET_TIMER(t_verify);
                        LC_STOP_AND_RESET_TIMER(t_other);
                }
-               BE_TIMER_PUSH(t_other);
+               BE_TIMER_PUSH(t_other);   /* t_other */
 
                BE_TIMER_ONLY(num_nodes_b = get_num_reachable_nodes(irg));
 
@@ -420,23 +445,25 @@ static void be_main_loop(FILE *file_handle)
                /* create the code generator and generate code. */
                prepare_graph(&birg);
 
+               BE_TIMER_POP(t_prolog);
+
                /* some transformations need to be done before abi introduce */
+               BE_TIMER_PUSH(t_codegen);
                arch_code_generator_before_abi(birg.cg);
-
-               BE_TIMER_POP();
-               BE_TIMER_PUSH(t_abi);
+               BE_TIMER_POP(t_codegen);
 
                /* implement the ABI conventions. */
+               BE_TIMER_PUSH(t_abi);
                birg.abi = be_abi_introduce(&birg);
-               dump(DUMP_ABI, irg, "-abi", dump_ir_block_graph);
+               BE_TIMER_POP(t_abi);
 
+               dump(DUMP_ABI, irg, "-abi", dump_ir_block_graph);
                be_do_stat_nodes(irg, "02 Abi");
 
-               BE_TIMER_POP();
-               BE_TIMER_PUSH(t_codegen);
-
                /* generate code */
+               BE_TIMER_PUSH(t_codegen);
                arch_code_generator_prepare_graph(birg.cg);
+               BE_TIMER_POP(t_codegen);
 
                be_do_stat_nodes(irg, "03 Prepare");
 
@@ -451,26 +478,30 @@ static void be_main_loop(FILE *file_handle)
 
                /* Compute loop nesting information (for weighting copies) */
                construct_cf_backedges(irg);
-
                dump(DUMP_PREPARED, irg, "-prepared", dump_ir_block_graph);
-
-               BE_TIMER_POP();
                BE_TIMER_ONLY(num_nodes_r = get_num_reachable_nodes(irg));
-               BE_TIMER_PUSH(t_sched);
 
-               /* Schedule the graphs. */
+               /* let backend prepare scheduling */
+               BE_TIMER_PUSH(t_codegen);
                arch_code_generator_before_sched(birg.cg);
-               list_sched(&birg, be_enable_mris);
+               BE_TIMER_POP(t_codegen);
+
+               /* schedule the irg */
+               BE_TIMER_PUSH(t_sched);
+               list_sched(&birg, &be_options);
+               BE_TIMER_POP(t_sched);
+
                dump(DUMP_SCHED, irg, "-sched", dump_ir_block_graph_sched);
 
                /* check schedule */
+               BE_TIMER_PUSH(t_verify);
                be_sched_vrfy(birg.irg, vrfy_option);
+               BE_TIMER_POP(t_verify);
 
                be_do_stat_nodes(irg, "04 Schedule");
 
-               BE_TIMER_POP();
+               /* introduce patterns to assure constraints */
                BE_TIMER_PUSH(t_constr);
-
                /* we switch off optimizations here, because they might cause trouble */
                save_optimization_state(&state);
                set_optimize(0);
@@ -479,73 +510,91 @@ static void be_main_loop(FILE *file_handle)
                /* add Keeps for should_be_different constrained nodes  */
                /* beware: needs schedule due to usage of be_ssa_constr */
                assure_constraints(&birg);
-               dump(DUMP_SCHED, irg, "-assured", dump_ir_block_graph_sched);
+               BE_TIMER_POP(t_constr);
 
+               dump(DUMP_SCHED, irg, "-assured", dump_ir_block_graph_sched);
                be_do_stat_nodes(irg, "05 Constraints");
 
                /* connect all stack modifying nodes together (see beabi.c) */
-               be_abi_fix_stack_nodes(birg.abi);
+               BE_TIMER_PUSH(t_abi);
+               be_abi_fix_stack_nodes(birg.abi, NULL);
+               BE_TIMER_POP(t_abi);
+
                dump(DUMP_SCHED, irg, "-fix_stack", dump_ir_block_graph_sched);
 
                /* check schedule */
+               BE_TIMER_PUSH(t_verify);
                be_sched_vrfy(birg.irg, vrfy_option);
+               BE_TIMER_POP(t_verify);
 
                /* do some statistics */
                be_do_stat_reg_pressure(&birg);
 
                /* stuff needs to be done after scheduling but before register allocation */
+               BE_TIMER_PUSH(t_codegen);
                arch_code_generator_before_ra(birg.cg);
-
-               BE_TIMER_POP();
-               BE_TIMER_PUSH(t_regalloc);
+               BE_TIMER_POP(t_codegen);
 
                /* Do register allocation */
+               BE_TIMER_ONLY(lc_timer_start(t_regalloc));
                ra_timer = ra->allocate(&birg);
-               dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched);
-
-               if (be_options.timing == BE_TIME_ON && ra_timer) {
-                       ra_prolog  = lc_timer_elapsed_msec(ra_timer->t_prolog);
-                       ra_epilog  = lc_timer_elapsed_msec(ra_timer->t_epilog);
-                       ra_live    = lc_timer_elapsed_msec(ra_timer->t_live);
-                       ra_spill   = lc_timer_elapsed_msec(ra_timer->t_spill);
-                       ra_color   = lc_timer_elapsed_msec(ra_timer->t_color);
-                       ra_copymin = lc_timer_elapsed_msec(ra_timer->t_copymin);
-                       ra_ssa     = lc_timer_elapsed_msec(ra_timer->t_ssa);
-                       ra_ifg     = lc_timer_elapsed_msec(ra_timer->t_ifg);
-               }
+               BE_TIMER_ONLY(lc_timer_stop(t_regalloc));
 
+               dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched);
                be_do_stat_nodes(irg, "06 Register Allocation");
 
-               BE_TIMER_POP();
+               /* let the codegenerator prepare the graph for emitter */
                BE_TIMER_PUSH(t_finish);
-
                arch_code_generator_after_ra(birg.cg);
+               BE_TIMER_POP(t_finish);
+
+               /* fix stack offsets */
+               BE_TIMER_PUSH(t_abi);
                be_abi_fix_stack_bias(birg.abi);
+               BE_TIMER_POP(t_abi);
 
-               /* check schedule */
-               be_sched_vrfy(birg.irg, vrfy_option);
+               BE_TIMER_PUSH(t_finish);
+               arch_code_generator_finish(birg.cg);
+               BE_TIMER_POP(t_finish);
+
+               dump(DUMP_FINAL, irg, "-finish", dump_ir_block_graph_sched);
+
+               /* check schedule and register allocation */
+               BE_TIMER_PUSH(t_verify);
+               if (vrfy_option == BE_VRFY_WARN) {
+                       //irg_verify(birg.irg, VRFY_ENFORCE_SSA);
+                       be_check_dominance(birg.irg);
+                       be_verify_schedule(birg.irg);
+                       be_verify_register_allocation(env.arch_env, birg.irg);
+               }
+               else if (vrfy_option == BE_VRFY_ASSERT) {
+                       //assert(irg_verify(birg.irg, VRFY_ENFORCE_SSA) && "irg verification failed");
+                       assert(be_check_dominance(birg.irg) && "Dominance verification failed");
+                       assert(be_verify_schedule(birg.irg) && "Schedule verification failed");
+                       assert(be_verify_register_allocation(env.arch_env, birg.irg)
+                              && "register allocation verification failed");
+               }
+               BE_TIMER_POP(t_verify);
 
-               BE_TIMER_POP();
+               /* emit assembler code */
                BE_TIMER_PUSH(t_emit);
-
                arch_code_generator_done(birg.cg);
+               BE_TIMER_POP(t_emit);
+
                dump(DUMP_FINAL, irg, "-end", dump_ir_extblock_graph_sched);
+
+               BE_TIMER_PUSH(t_abi);
                be_abi_free(birg.abi);
+               BE_TIMER_POP(t_abi);
 
                be_do_stat_nodes(irg, "07 Final");
-
-               BE_TIMER_POP();
-
                restore_optimization_state(&state);
 
                BE_TIMER_ONLY(num_nodes_a = get_num_reachable_nodes(irg));
+               BE_TIMER_POP(t_other);
 
-               /* switched off due to statistics (statistic module needs all irgs) */
-               //              free_ir_graph(irg);
-               BE_TIMER_POP();
-
-#define LC_EMIT(timer)     printf("%-20s: %u msec\n", lc_timer_get_description(timer), lc_timer_elapsed_msec(timer))
-#define EMIT_RA_TIME(n, t) printf("%-20s: %u msec\n", n, t)
+#define LC_EMIT(timer)    printf("%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0)
+#define LC_EMIT_RA(timer) printf("\t%-20s: %.3lf msec\n", lc_timer_get_description(timer), (double)lc_timer_elapsed_usec(timer) / 1000.0)
                if (be_options.timing == BE_TIME_ON) {
                        printf("==>> IRG %s <<==\n", get_entity_name(get_irg_entity(irg)));
                        printf("# nodes at begin:  %u\n", num_nodes_b);
@@ -557,25 +606,34 @@ static void be_main_loop(FILE *file_handle)
                        LC_EMIT(t_sched);
                        LC_EMIT(t_constr);
                        LC_EMIT(t_regalloc);
-                       EMIT_RA_TIME("prolog",    ra_prolog);
-                       EMIT_RA_TIME("liveness",  ra_live);
-                       EMIT_RA_TIME("spilling",  ra_spill);
-                       EMIT_RA_TIME("coloring",  ra_color);
-                       EMIT_RA_TIME("ifg build", ra_ifg);
-                       EMIT_RA_TIME("copymin",   ra_copymin);
-                       EMIT_RA_TIME("ssa destr", ra_ssa);
-                       EMIT_RA_TIME("epilog",    ra_epilog);
+                       LC_EMIT_RA(ra_timer->t_prolog);
+                       LC_EMIT_RA(ra_timer->t_live);
+                       LC_EMIT_RA(ra_timer->t_spill);
+                        LC_EMIT_RA(ra_timer->t_spillslots);
+                       LC_EMIT_RA(ra_timer->t_color);
+                       LC_EMIT_RA(ra_timer->t_ifg);
+                       LC_EMIT_RA(ra_timer->t_copymin);
+                       LC_EMIT_RA(ra_timer->t_ssa);
+                       LC_EMIT_RA(ra_timer->t_epilog);
+                       LC_EMIT_RA(ra_timer->t_verify);
+                       LC_EMIT_RA(ra_timer->t_other);
                        LC_EMIT(t_finish);
                        LC_EMIT(t_emit);
+                       LC_EMIT(t_verify);
                        LC_EMIT(t_other);
                }
 #undef LC_EMIT
+
+        /* switched off due to statistics (statistic module needs all irgs) */
+               if (! stat_is_active())
+                       free_ir_graph(irg);
+
        }
        be_done_env(&env);
 
-#undef BE_TIME_START
-#undef BE_TIME_STOP
-#undef BE_TIME_ONLY
+#undef BE_TIMER_POP
+#undef BE_TIMER_PUSH
+#undef BE_TIMER_ONLY
 }
 
 /* Main interface to the frontend. */