#ifdef WITH_LIBCORE
#include <libcore/lc_opts.h>
#include <libcore/lc_opts_enum.h>
+#include <libcore/lc_timing.h>
#endif /* WITH_LIBCORE */
#include "obst.h"
#include "irloop_t.h"
#include "irtools.h"
#include "return.h"
+#include "firmstat.h"
#include "bearch.h"
#include "firm/bearch_firm.h"
#include "besched_t.h"
#include "belistsched.h"
#include "belive_t.h"
-#include "bespillilp.h"
#include "bespillbelady.h"
#include "bera.h"
#include "beraextern.h"
#include "bestat.h"
#include "beverify.h"
-#define DUMP_INITIAL (1 << 0)
-#define DUMP_ABI (1 << 1)
-#define DUMP_SCHED (1 << 2)
-#define DUMP_PREPARED (1 << 3)
-#define DUMP_RA (1 << 4)
-#define DUMP_FINAL (1 << 5)
-
-enum {
- BE_VRFY_OFF,
- BE_VRFY_WARN,
- BE_VRFY_ASSERT
-};
-
/* options visible for anyone */
static be_options_t be_options = {
- /* ilp server */
- "i44pc52.info.uni-karlsruhe.de",
-
- /* ilp solver */
- "cplex"
+ DUMP_NONE, /* dump options */
+ BE_TIME_OFF, /* no timing */
+ "i44pc52.info.uni-karlsruhe.de", /* ilp server */
+ "cplex" /* ilp solver */
};
/* dump flags */
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;
/* possible dumping options */
static const lc_opt_enum_mask_items_t dump_items[] = {
- { "none", 0 },
+ { "none", DUMP_NONE },
{ "initial", DUMP_INITIAL },
{ "abi", DUMP_ABI },
{ "sched", DUMP_SCHED },
{ "prepared", DUMP_PREPARED },
{ "regalloc", DUMP_RA },
{ "final", DUMP_FINAL },
- { "all", 2 * DUMP_FINAL - 1 },
+ { "be", DUMP_BE },
+ { "all", 2 * DUMP_BE - 1 },
{ NULL, 0 }
};
/* register allocators */
static const lc_opt_enum_const_ptr_items_t ra_items[] = {
- { "chordal", &be_ra_chordal_allocator },
+ { "chordal", &be_ra_chordal_allocator },
{ "external", &be_ra_external_allocator },
{ NULL, NULL }
};
{ "assert", BE_VRFY_ASSERT },
{ NULL, 0 }
};
+
static lc_opt_enum_mask_var_t dump_var = {
&dump_flags, dump_items
};
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),
#ifdef WITH_ILP
LC_OPT_ENT_STR ("ilp.server", "the ilp server name", be_options.ilp_server, sizeof(be_options.ilp_server)),
be_opt_register();
be_sched_init();
- be_liveness_init();
be_numbering_init();
be_copy_opt_init();
copystat_init();
return &be_params;
}
+/**
+ * Initializes the main environment for the backend.
+ *
+ * @param env an empty environment
+ * @param file_handle the file handle where the output will be written to
+ */
static be_main_env_t *be_init_env(be_main_env_t *env, FILE *file_handle)
{
memset(env, 0, sizeof(*env));
obstack_init(&env->obst);
env->arch_env = obstack_alloc(&env->obst, sizeof(env->arch_env[0]));
env->options = &be_options;
+ env->options->dump_flags = dump_flags;
FIRM_DBG_REGISTER(env->dbg, "be.main");
arch_env_init(env->arch_env, isa_if, file_handle);
obstack_free(&env->obst, NULL);
}
+/**
+ * A wrapper around a firm dumper. Dumps only, if
+ * flags are enabled.
+ *
+ * @param mask a bitmask containing the reason what will be dumped
+ * @param irg the IR graph to dump
+ * @param suffix the suffix for the dumper
+ * @param dumper the dumper to be called
+ */
static void dump(int mask, ir_graph *irg, const char *suffix,
void (*dumper)(ir_graph *, const char *))
{
* The Firm backend main loop.
* Do architecture specific lowering for all graphs
* and call the architecture specific code generator.
+ *
+ * @param file_handle the file handle the output will be written to
*/
static void be_main_loop(FILE *file_handle)
{
int i, n;
arch_isa_t *isa;
be_main_env_t env;
+ unsigned num_nodes_b = 0;
+ unsigned num_nodes_a = 0;
+ unsigned num_nodes_r = 0;
+ 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) {
+ t_prolog = lc_timer_register("prolog", "prolog");
+ t_abi = lc_timer_register("beabi", "be abi introduction");
+ t_codegen = lc_timer_register("codegen", "codegeneration");
+ t_sched = lc_timer_register("sched", "scheduling");
+ t_constr = lc_timer_register("constr", "assure constraints");
+ 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");
+ }
be_init_env(&env, file_handle);
isa = arch_env_get_isa(env.arch_env);
- // /* for debugging, anchors helps */
+ /* for debugging, anchors helps */
// dump_all_anchors(1);
+#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 */
for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
ir_graph *irg = get_irp_irg(i);
const arch_code_generator_if_t *cg_if;
be_irg_t birg;
- int save_optimize, save_normalize;
+ optimization_state_t state;
+
+ /* stop and reset timers */
+ if (be_options.timing == BE_TIME_ON) {
+ LC_STOP_AND_RESET_TIMER(t_prolog);
+ LC_STOP_AND_RESET_TIMER(t_abi);
+ LC_STOP_AND_RESET_TIMER(t_codegen);
+ LC_STOP_AND_RESET_TIMER(t_sched);
+ LC_STOP_AND_RESET_TIMER(t_constr);
+ 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); /* t_other */
+
+ BE_TIMER_ONLY(num_nodes_b = get_num_reachable_nodes(irg));
birg.irg = irg;
birg.main_env = &env;
DBG((env.dbg, LEVEL_2, "====> IRG: %F\n", irg));
dump(DUMP_INITIAL, irg, "-begin", dump_ir_block_graph);
+ BE_TIMER_PUSH(t_prolog);
+
be_stat_init_irg(env.arch_env, irg);
be_do_stat_nodes(irg, "01 Begin");
/* 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(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");
/* 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");
/* Compute loop nesting information (for weighting copies) */
construct_cf_backedges(irg);
-
dump(DUMP_PREPARED, irg, "-prepared", dump_ir_block_graph);
+ BE_TIMER_ONLY(num_nodes_r = get_num_reachable_nodes(irg));
- /* Schedule the graphs. */
+ /* let backend prepare scheduling */
+ BE_TIMER_PUSH(t_codegen);
arch_code_generator_before_sched(birg.cg);
+ BE_TIMER_POP(t_codegen);
+
+ /* schedule the irg */
+ BE_TIMER_PUSH(t_sched);
list_sched(&birg, be_enable_mris);
+ 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");
+ /* introduce patterns to assure constraints */
+ BE_TIMER_PUSH(t_constr);
/* we switch off optimizations here, because they might cause trouble */
- save_optimize = get_optimize();
- save_normalize = get_opt_normalize();
+ save_optimization_state(&state);
set_optimize(0);
set_opt_normalize(0);
/* 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);
- /* Do register allocation */
+ /* stuff needs to be done after scheduling but before register allocation */
+ BE_TIMER_PUSH(t_codegen);
arch_code_generator_before_ra(birg.cg);
- ra->allocate(&birg);
- dump(DUMP_RA, irg, "-ra", dump_ir_block_graph_sched);
+ BE_TIMER_POP(t_codegen);
+
+ /* Do register allocation */
+ BE_TIMER_ONLY(lc_timer_start(t_regalloc));
+ ra_timer = ra->allocate(&birg);
+ 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");
+ /* 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_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);
be_sched_vrfy(birg.irg, vrfy_option);
+ be_verify_register_allocation(env.arch_env, birg.irg);
+ BE_TIMER_POP(t_verify);
+ /* 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");
+ restore_optimization_state(&state);
+
+ BE_TIMER_ONLY(num_nodes_a = get_num_reachable_nodes(irg));
+ BE_TIMER_POP(t_other);
+
+#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);
+ printf("# nodes before ra: %u\n", num_nodes_r);
+ printf("# nodes at end: %u\n\n", num_nodes_a);
+ LC_EMIT(t_prolog);
+ LC_EMIT(t_abi);
+ LC_EMIT(t_codegen);
+ LC_EMIT(t_sched);
+ LC_EMIT(t_constr);
+ LC_EMIT(t_regalloc);
+ 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
- /* reset the optimizations */
- set_optimize(save_optimize);
- set_opt_normalize(save_normalize);
+ /* switched off due to statistics (statistic module needs all irgs) */
+ if (! stat_is_active())
+ free_ir_graph(irg);
- /* switched off due to statistics (statistic module needs all irgs) */
- // free_ir_graph(irg);
}
be_done_env(&env);
+
+#undef BE_TIMER_POP
+#undef BE_TIMER_PUSH
+#undef BE_TIMER_ONLY
}
/* Main interface to the frontend. */
void be_main(FILE *file_handle)
{
+#ifdef WITH_LIBCORE
+ lc_timer_t *t;
+
+ if (be_options.timing == BE_TIME_ON) {
+ t = lc_timer_register("bemain", "measure complete bemain loop");
+
+ if (lc_timer_enter_high_priority()) {
+ fprintf(stderr, "Warning: Could not enter high priority mode.\n");
+ }
+
+ lc_timer_reset_and_start(t);
+ }
+#endif /* WITH_LIBCORE */
+
/* never build code for pseudo irgs */
set_visit_pseudo_irgs(0);
be_node_init();
be_main_loop(file_handle);
+
+#ifdef WITH_LIBCORE
+ if (be_options.timing == BE_TIME_ON) {
+ lc_timer_stop(t);
+ lc_timer_leave_high_priority();
+ printf("%-20s: %lu msec\n", "BEMAINLOOP", lc_timer_elapsed_msec(t));
+ }
+#endif /* WITH_LIBCORE */
}
/** The debug info retriever function. */
projects / libfirm / blobdiff