2 * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Implementation of the RA-Interface for an external, (non-SSA) register allocator.
23 * @author Daniel Grund
27 * The external register allocator is a program:
28 * PROG -i INPUTFILE -o OUTPUTFILE
30 * 1) Input file defines the interference graph
31 * 2) Output file contains the instructions to perform
36 ----------------------
38 inputfile ::= regs nodes interf affinities .
40 regs ::= 'regs' regcount . // Anzahl der register (0..regcount-1), die zur Verfuegung stehen
42 nodes ::= 'nodes' '{' node* '}' . // All nodes in the graph
45 | node-info '<' reg-nr '>' . // Reg-nr is present in case of constraints
47 node-info ::= node-nr spill-costs .
49 interf ::= 'interferences' '{' i-edge* '}' . // Interference edges of the graph
51 i-edge ::= '(' node-nr ',' node-nr ')' .
53 affinities ::= 'affinities' '{' a-edge* '}' . // Affinity edges of the graph
55 a-edge ::= '(' node-nr ',' node-nr ',' weight ')' .
58 weight, regcount, node-nr ::= int32 .
59 spill-costs ::= int32 . // negative spill costs indicate unspillable
61 The output file format
62 -----------------------
64 outputfile ::= spills | allocs .
66 spills ::= 'spills' node-nr+ .
68 allocs ::= 'allocs' alloc* .
70 alloc ::= node-nr reg-nr .
81 #include <libcore/lc_opts.h>
82 #include <libcore/lc_opts_enum.h>
88 #include "raw_bitset.h"
91 #include "irprintf_t.h"
93 #include "irgraph_t.h"
95 #include "iredges_t.h"
100 #include "beraextern.h"
102 #include "bearch_t.h"
103 #include "benode_t.h"
104 #include "beirgmod.h"
105 #include "besched_t.h"
107 #include "belive_t.h"
108 #include "beinsn_t.h"
110 #include "bessadestrsimple.h"
115 * Environment with all the needed stuff
117 typedef struct _be_raext_env_t {
119 const arch_register_class_t *cls;
123 FILE *f; /**< file handle used for out- and input file */
124 set *vars; /**< contains all be_var_info_t */
125 int n_cls_vars; /**< length of the array cls_vars */
126 be_var_info_t **cls_vars; /**< only the var_infos for current cls. needed for double iterating */
127 DEBUG_ONLY(firm_dbg_module_t *dbg;)
132 /******************************************************************************
135 | |__| | ___| |_ __ ___ _ __ ___
136 | __ |/ _ \ | '_ \ / _ \ '__/ __|
137 | | | | __/ | |_) | __/ | \__ \
138 |_| |_|\___|_| .__/ \___|_| |___/
141 *****************************************************************************/
144 #define pset_foreach(pset, irn) for(irn=pset_first(pset); irn; irn=pset_next(pset))
145 #define set_foreach(set, e) for(e=set_first(set); e; e=set_next(set))
148 * Checks if _the_ result of the irn belongs to the
149 * current register class (raenv->cls)
150 * NOTE: Only the first result is checked.
152 #define is_res_in_reg_class(irn) arch_irn_has_reg_class(raenv->aenv, irn, -1, raenv->cls)
154 static INLINE ir_node *get_first_non_phi(pset *s) {
163 assert(0 && "There must be a non-phi-irn in this");
167 static INLINE ir_node *get_first_phi(pset *s) {
176 assert(0 && "There must be a phi in this");
180 static int get_loop_weight(ir_node *irn) {
182 ir_loop *loop = get_irn_loop(get_nodes_block(irn));
185 int d = get_loop_depth(loop);
191 #define get_const_weight(irn) (1)
193 #define get_spill_weight(irn) get_loop_weight(irn)
194 #define get_reload_weight(irn) get_loop_weight(irn)
195 #define get_affinity_weight(irn) get_loop_weight(irn)
197 /******************************************************************************
199 / ____| | | / ____| (_)
200 | | ___ _ __ ___| |_ _ __ | | ___ _ __ _ ___ ___
201 | | / _ \| '_ \/ __| __| '__| | | / _ \| '_ \| |/ _ \/ __|
202 | |___| (_) | | | \__ \ |_| | | |___| (_) | |_) | | __/\__ \
203 \_____\___/|_| |_|___/\__|_| \_____\___/| .__/|_|\___||___/
206 *****************************************************************************/
208 static void handle_constraints_insn(be_raext_env_t *env, be_insn_t *insn)
210 ir_node *bl = get_nodes_block(insn->irn);
213 for(i = 0; i < insn->use_start; ++i) {
214 be_operand_t *op = &insn->ops[i];
216 if(op->has_constraints) {
217 ir_node *cpy = be_new_Copy(op->req->cls, env->irg, bl, op->carrier);
218 sched_add_before(insn->next_insn, cpy);
219 edges_reroute(op->carrier, cpy, env->irg);
223 for(i = insn->use_start; i < insn->n_ops; ++i) {
224 be_operand_t *op = &insn->ops[i];
226 if(op->has_constraints) {
227 ir_node *cpy = be_new_Copy(op->req->cls, env->irg, bl, op->carrier);
228 sched_add_before(insn->irn, cpy);
229 set_irn_n(insn->irn, op->pos, cpy);
230 be_set_constr_limited(cpy, BE_OUT_POS(0), op->req);
235 static void handle_constraints_block(ir_node *bl, void *data)
237 be_raext_env_t *raenv = data;
238 int active = bl != get_irg_start_block(raenv->irg);
245 ie.aenv = raenv->aenv;
247 ie.ignore_colors = NULL;
250 irn = sched_first(bl);
251 while(!sched_is_end(irn)) {
252 be_insn_t *insn = be_scan_insn(&ie, irn);
254 if(insn->has_constraints)
255 handle_constraints_insn(raenv, insn);
257 if(be_is_Barrier(irn))
260 irn = insn->next_insn;
261 obstack_free(&obst, insn);
265 static void handle_constraints(be_raext_env_t *raenv) {
266 irg_block_walk_graph(raenv->irg, NULL, handle_constraints_block, raenv);
272 /******************************************************************************
275 | | | |_ _ _ __ ___ _ __ ___ _ __
276 | | | | | | | '_ ` _ \| '_ \ / _ \ '__|
277 | |__| | |_| | | | | | | |_) | __/ |
278 |_____/ \__,_|_| |_| |_| .__/ \___|_|
281 *****************************************************************************/
284 static void extract_vars_of_cls(be_raext_env_t *raenv) {
288 raenv->cls_vars = xmalloc(set_count(raenv->vars) * sizeof(*raenv->cls_vars));
289 assert(raenv->cls_vars);
291 set_foreach(raenv->vars, vi)
292 if (is_res_in_reg_class(get_first_non_phi(vi->values)))
293 raenv->cls_vars[count++] = vi;
295 raenv->cls_vars = realloc(raenv->cls_vars, count * sizeof(*raenv->cls_vars));
296 assert(raenv->cls_vars);
298 raenv->n_cls_vars = count;
303 * Check if node irn has a limited-constraint at position pos.
304 * If yes, dump it to FILE raenv->f
306 static INLINE void dump_constraint(be_raext_env_t *raenv, ir_node *irn, int pos) {
307 const arch_register_req_t *req;
309 req = arch_get_register_req(raenv->aenv, irn, pos);
310 if (arch_register_req_is(req, limited)) {
313 reg_nr = rbitset_next(req->limited, 0, 1);
314 fprintf(raenv->f, "<%d>", reg_nr);
315 assert(rbitset_popcnt(req->limited, raenv->cls->n_regs) <= 1
316 && "Constraints with more than 1 possible register are not supported");
320 #define UNSPILLABLE -1
322 static INLINE int get_spill_costs(be_raext_env_t *raenv, be_var_info_t *vi) {
324 int c_spills=0, c_reloads=0;
326 pset_foreach(vi->values, irn) {
327 if (arch_irn_is(raenv->aenv, irn, ignore) || be_is_Reload(irn)) {
328 pset_break(vi->values);
333 /* number of reloads is the number of non-phi uses of all values of this var */
334 const ir_edge_t *edge;
335 foreach_out_edge(irn, edge)
336 if (!is_Phi(edge->src))
337 c_reloads += get_reload_weight(edge->src);
339 /* number of spills is the number of non-phi values for this var */
340 c_spills += get_spill_weight(irn);
344 return c_spills + c_reloads;
347 static void dump_nodes(be_raext_env_t *raenv) {
351 fprintf(f, "\nnodes {\n");
353 for (i=0; i<raenv->n_cls_vars; ++i) {
354 be_var_info_t *vi = raenv->cls_vars[i];
356 if (vi->var_nr == SET_REMOVED)
359 fprintf(f, "%d %d", vi->var_nr, get_spill_costs(raenv, vi));
360 dump_constraint(raenv, get_first_non_phi(vi->values), -1);
369 static void dump_interferences(be_raext_env_t *raenv) {
371 be_var_info_t *vi1, *vi2;
372 ir_node *irn1, *irn2;
374 be_lv_t *lv = raenv->birg->lv;
376 fprintf(f, "\ninterferences {\n");
378 for (i=0; i<raenv->n_cls_vars; ++i) {
379 vi1 = raenv->cls_vars[i];
381 if (vi1->var_nr == SET_REMOVED)
384 for (o=i+1; o<raenv->n_cls_vars; ++o) {
385 vi2 = raenv->cls_vars[o];
387 if (vi2->var_nr == SET_REMOVED)
390 pset_foreach(vi1->values, irn1)
391 pset_foreach(vi2->values, irn2)
392 if (values_interfere(lv, irn1, irn2)) {
393 pset_break(vi1->values);
394 pset_break(vi2->values);
395 fprintf(f, "(%d, %d)\n", vi1->var_nr, vi2->var_nr);
405 static void dump_affinities_walker(ir_node *irn, void *env) {
406 be_raext_env_t *raenv = env;
407 const arch_register_req_t *req;
409 be_var_info_t *vi1, *vi2;
411 if (arch_get_irn_reg_class(raenv->aenv, irn, -1) != raenv->cls || arch_irn_is(raenv->aenv, irn, ignore))
414 vi1 = be_get_var_info(irn);
416 /* copies have affinities */
417 if (arch_irn_class_is(raenv->aenv, irn, copy)) {
418 ir_node *other = be_get_Copy_op(irn);
420 if (! arch_irn_is(raenv->aenv, other, ignore)) {
421 vi2 = be_get_var_info(other);
423 fprintf(raenv->f, "(%d, %d, %d)\n", vi1->var_nr, vi2->var_nr, get_affinity_weight(irn));
428 /* should_be_equal constraints are affinites */
429 for (pos = 0, max = get_irn_arity(irn); pos<max; ++pos) {
430 req = arch_get_register_req(raenv->aenv, irn, pos);
432 if (arch_register_req_is(req, should_be_same)) {
433 ir_node *other = get_irn_n(irn, req->other_same);
434 if(arch_irn_is(raenv->aenv, other, ignore)) {
435 vi2 = be_get_var_info(other);
437 fprintf(raenv->f, "(%d, %d, %d)\n", vi1->var_nr, vi2->var_nr, get_affinity_weight(irn));
444 static void dump_affinities(be_raext_env_t *raenv) {
445 fprintf(raenv->f, "\naffinities {\n");
446 irg_walk_graph(raenv->irg, NULL, dump_affinities_walker, raenv);
447 fprintf(raenv->f, "}\n");
451 * Dump all information needed by the external
452 * register allocator to a single file.
454 static void dump_to_file(be_raext_env_t *raenv, char *filename) {
457 if (!(f = fopen(filename, "wt"))) {
458 fprintf(stderr, "Could not open file %s for writing\n", filename);
464 /* dump register info */
465 fprintf(f, "regs %d\n", arch_register_class_n_regs(raenv->cls));
467 /* dump the interference graph */
469 dump_interferences(raenv);
470 dump_affinities(raenv);
475 /******************************************************************************
478 | |__ __ _____ ___ _ _| |_ ___
479 | __| \ \/ / _ \/ __| | | | __/ _ \
480 | |____ > < __/ (__| |_| | || __/
481 |______/_/\_\___|\___|\__,_|\__\___|
482 *****************************************************************************/
485 * Execute the external register allocator specified in the
486 * firm-option firm.be.ra.ext.callee
488 static void execute(char *prog_to_call, char *out_file, char *result_file) {
492 snprintf(cmd_line, sizeof(cmd_line), "%s -i %s -o %s", prog_to_call, out_file, result_file);
493 cmd_line[sizeof(cmd_line) - 1] = '\0';
495 ret_status = system(cmd_line);
496 assert(ret_status != -1 && "Invokation of external register allocator failed");
497 assert(ret_status == 0 && "External register allocator is unhappy with sth.");
500 /******************************************************************************
503 / \ _ __ _ __ | |_ _ | |__) |___ ___ _ _| | |_
504 / /\ \ | '_ \| '_ \| | | | | | _ // _ \/ __| | | | | __|
505 / ____ \| |_) | |_) | | |_| | | | \ \ __/\__ \ |_| | | |_
506 /_/ \_\ .__/| .__/|_|\__, | |_| \_\___||___/\__,_|_|\__|
509 *****************************************************************************/
512 * Spill a variable and add reloads before all uses.
514 static INLINE void var_add_spills_and_reloads(be_raext_env_t *raenv, int var_nr) {
515 be_var_info_t *vi = be_var_find(raenv->vars, var_nr);
516 ir_node *spill=NULL, *ctx, *irn;
518 const ir_edge_t *edge, *ne;
519 pset *spills = pset_new_ptr(4); /* the spills of this variable */
520 pset *reloads = pset_new_ptr(4); /* the reloads of this variable */
521 be_lv_t *lv = raenv->birg->lv;
522 be_dom_front_info_t *dom_front = raenv->birg->dom_front;
523 int new_size, n_spills, n_reloads;
525 assert(vi && "Variable nr does not exist!");
526 assert(pset_count(vi->values) && "There are no values associated to this variable");
528 /* the spill context is set to an arbitrary node of the phi-class,
529 * or the node itself if it is not member of a phi class
531 if (pset_count(vi->values) == 1)
532 ctx = get_first_non_phi(vi->values);
534 ctx = get_first_phi(vi->values);
536 DBG((raenv->dbg, LEVEL_2, "Spill context: %+F\n", ctx));
538 /* for each value of this variable insert the spills */
539 pset_foreach(vi->values, irn) {
545 /* all ordinary nodes must be spilled */
546 DBG((raenv->dbg, LEVEL_2, " spilling %+F\n", irn));
547 spill = be_spill(raenv->aenv, irn);
549 /* remember the spill */
550 pset_insert_ptr(spills, spill);
553 assert(spill && "There must be at least one non-phi-node");
555 mode = get_irn_mode(get_irn_n(spill, be_pos_Spill_val));
557 /* insert reloads and wire them arbitrary*/
558 pset_foreach(vi->values, irn) {
559 foreach_out_edge_safe(irn, edge, ne) {
560 ir_node *reload, *src = edge->src;
561 if (is_Phi(src) || be_is_Spill(src))
564 /* all real uses must be reloaded */
565 DBG((raenv->dbg, LEVEL_2, " reloading before %+F\n", src));
566 reload = be_reload(raenv->aenv, raenv->cls, edge->src, mode, spill);
567 set_irn_n(edge->src, edge->pos, reload);
569 /* remember the reload */
570 pset_insert_ptr(reloads, reload);
574 /* correct the reload->spill pointers... */
575 be_ssa_constr_set_ignore(dom_front, lv, spills, NULL);
578 /****** correct the variable <--> values mapping: ******
580 * - if we had a phi class it gets split into several new variables
581 * - all reloads are new variables
583 n_spills = pset_count(spills);
584 n_reloads = pset_count(reloads);
586 /* first make room for new pointers in the cls_var array */
587 new_size = raenv->n_cls_vars + n_reloads + ((n_spills>1) ? n_spills : 0);
588 raenv->cls_vars = realloc(raenv->cls_vars, (new_size) * sizeof(*raenv->cls_vars));
589 assert(raenv->cls_vars && "Out of mem!?");
591 /* if we had a real phi-class, we must... */
592 if (pset_count(spills) > 1) {
593 /* ...remove the old variable corresponding to the phi class */
594 vi->var_nr = SET_REMOVED;
596 /* ...add new vars for each non-phi-member */
597 pset_foreach(spills, irn) {
598 ir_node *spilled = get_irn_n(irn, be_pos_Spill_val);
599 raenv->cls_vars[raenv->n_cls_vars++] = be_var_add_value(raenv->vars, get_irn_node_nr(spilled), spilled);
603 /* add new variables for all reloads */
604 pset_foreach(reloads, irn) {
605 assert(get_irn_node_nr(irn) != 1089);
606 raenv->cls_vars[raenv->n_cls_vars++] = be_var_add_value(raenv->vars, get_irn_node_nr(irn), irn);
613 #define INVALID_FILE_FORMAT assert(0 && "Invalid file format.")
615 #define BUFCONV " %32s "
618 * Read in the actions performed by the external allocator.
619 * Apply these transformations to the irg.
620 * @return 1 if an allocation was read in. 0 otherwise.
622 static int read_and_apply_results(be_raext_env_t *raenv, char *filename) {
625 int is_allocation = 0;
627 if (!(f = fopen(filename, "rt"))) {
628 fprintf(stderr, "Could not open file %s for reading\n", filename);
634 /* read the action */
635 if (fscanf(f, BUFCONV, buf) != 1)
639 if (!strcmp(buf, "spills")) {
641 while (fscanf(f, " %d ", &var_nr) == 1)
642 var_add_spills_and_reloads(raenv, var_nr);
645 /* or do we allocate */
646 if (!strcmp(buf, "allocs")) {
650 while (fscanf(f, " %d %d ", &var_nr, ®_nr) == 2) {
652 pset *vals = be_get_var_values(raenv->vars, var_nr);
654 assert(vals && "Variable nr does not exist!");
655 pset_foreach(vals, irn)
656 arch_set_irn_register(raenv->aenv, irn, arch_register_for_index(raenv->cls, reg_nr));
666 return is_allocation;
669 static void check_allocation(be_raext_env_t *raenv) {
671 be_lv_t *lv = raenv->birg->lv;
673 for (i=0; i<raenv->n_cls_vars; ++i) {
674 be_var_info_t *vi1 = raenv->cls_vars[i];
676 if (vi1->var_nr == SET_REMOVED)
679 for (o=0; o<i; ++o) {
680 be_var_info_t *vi2 = raenv->cls_vars[o];
681 ir_node *irn1, *irn2;
683 if (vi2->var_nr == SET_REMOVED)
686 pset_foreach(vi1->values, irn1)
687 pset_foreach(vi2->values, irn2)
688 if (values_interfere(lv, irn1, irn2) && arch_get_irn_register(raenv->aenv, irn1) == arch_get_irn_register(raenv->aenv, irn2)) {
689 dump_ir_block_graph_sched(raenv->irg, "ERROR");
690 ir_fprintf(stdout, "SSA values %+F and %+F interfere. They belong to variable %d and %d respectively.\n", irn1, irn2, vi1->var_nr, vi2->var_nr);
691 assert(0 && "ERROR graph dumped");
697 /******************************************************************************
701 | |\/| |/ _` | | '_ \
702 | | | | (_| | | | | |
703 |_| |_|\__,_|_|_| |_|
704 *****************************************************************************/
707 * Default values for options
709 static char callee[128] = "\"E:/user/kimohoff/public/register allocator\"";
710 //static char callee[128] = "/ben/kimohoff/ipd-registerallocator/register_allocator";
714 * Allocate registers with an external program using a text-file interface.
716 * Do some computations (SSA-destruction and mapping of values--vars)
718 * Execute external program
719 * Read in results and apply them
722 static void be_ra_extern_main(be_irg_t *birg) {
723 be_main_env_t *env = birg->main_env;
724 ir_graph *irg = birg->irg;
726 be_raext_env_t raenv;
729 be_assure_dom_front(birg);
730 be_assure_liveness(birg);
735 raenv.aenv = env->arch_env;
736 FIRM_DBG_REGISTER(raenv.dbg, "firm.be.raextern");
738 /* Insert copies for constraints */
739 for(clsnr = 0, clss = arch_isa_get_n_reg_class(raenv.aenv->isa); clsnr < clss; ++clsnr) {
740 raenv.cls = arch_isa_get_reg_class(raenv.aenv->isa, clsnr);
741 handle_constraints(&raenv);
744 be_dump(irg, "-extern-constr", dump_ir_block_graph_sched);
746 /* SSA destruction respectively transformation into "Conventional SSA" */
747 raenv.vars = be_ssa_destr_simple(irg, env->arch_env);
748 be_dump(irg, "-extern-ssadestr", dump_ir_block_graph_sched);
751 /* For all register classes */
752 for(clsnr = 0, clss = arch_isa_get_n_reg_class(raenv.aenv->isa); clsnr < clss; ++clsnr) {
754 char out[256], in[256];
756 raenv.cls = arch_isa_get_reg_class(raenv.aenv->isa, clsnr);
758 extract_vars_of_cls(&raenv);
761 ir_snprintf(out, sizeof(out), "%F-%s-%d.ra", irg, raenv.cls->name, round);
762 ir_snprintf(in, sizeof(in), "%F-%s-%d.ra.res", irg, raenv.cls->name, round);
766 dump_to_file(&raenv, out);
767 execute(callee, out, in);
768 done = read_and_apply_results(&raenv, in);
769 be_abi_fix_stack_nodes(birg->abi);
771 ir_snprintf(in, sizeof(in), "-extern-%s-round-%d", raenv.cls->name, round);
772 be_dump(irg, in, dump_ir_block_graph_sched);
777 check_allocation(&raenv);
779 free(raenv.cls_vars);
782 be_dump(irg, "-extern-alloc", dump_ir_block_graph_sched);
785 free_ssa_destr_simple(raenv.vars);
787 be_invalidate_liveness(birg);
790 /******************************************************************************
793 | | | |_ __ | |_ _ ___ _ __ ___
794 | | | | '_ \| __| |/ _ \| '_ \/ __|
795 | |__| | |_) | |_| | (_) | | | \__ \
796 \____/| .__/ \__|_|\___/|_| |_|___/
799 *****************************************************************************/
801 static const lc_opt_enum_func_ptr_items_t ssa_destr_items[] = {
802 { "simple", (int (*)(void)) be_ssa_destr_simple }, /* TODO make (void*) casts nicer */
806 static set* (*ssa_destr)(ir_graph*,const arch_env_t*) = be_ssa_destr_simple;
808 static lc_opt_enum_func_ptr_var_t ssa_destr_var = {
809 (int (**)(void)) &ssa_destr, ssa_destr_items
812 static const lc_opt_table_entry_t be_ra_extern_options[] = {
813 LC_OPT_ENT_ENUM_FUNC_PTR("ssa_destr", "SSA destruction flavor", &ssa_destr_var),
814 LC_OPT_ENT_STR("callee", "The external program to call", callee, sizeof(callee)),
818 static be_ra_t be_ra_external_allocator = {
822 void be_init_raextern(void) {
823 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
824 lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "ra");
825 lc_opt_entry_t *ext_grp = lc_opt_get_grp(blocksched_grp, "ext");
827 lc_opt_add_table(ext_grp, be_ra_extern_options);
829 be_register_allocator("ext", &be_ra_external_allocator);
831 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_raextern);