X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbechordal.c;h=1fa973b35fe5bf273f0ddaa6e803d0dccadf9643;hb=6e998f91ca26b38b9022cfe5f3316bd3074bf39f;hp=dc6ea1ac5adff195ed8c8f39bbdbb5b2fab8b922;hpb=664c5fbedff1ecee98ca8c8ccbb0fb0e11e7059a;p=libfirm diff --git a/ir/be/bechordal.c b/ir/be/bechordal.c index dc6ea1ac5..1fa973b35 100644 --- a/ir/be/bechordal.c +++ b/ir/be/bechordal.c @@ -1,33 +1,43 @@ -/** - * Chordal register allocation. - * @author Sebastian Hack - * @date 8.12.2004 - * @cvs-id $Id$ +/* + * Copyright (C) 1995-2007 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. * - * Copyright (C) Universitaet Karlsruhe - * Released under the GPL + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. */ +/** + * @file + * @brief Chordal register allocation. + * @author Sebastian Hack + * @date 08.12.2004 + * @version $Id$ + */ #ifdef HAVE_CONFIG_H #include "config.h" #endif -#ifdef HAVE_MALLOC_H -#include -#endif - -#ifdef HAVE_ALLOCA_H -#include -#endif - #include #include "obst.h" #include "pset.h" #include "list.h" #include "bitset.h" +#include "raw_bitset.h" #include "iterator.h" #include "bipartite.h" +#include "hungarian.h" #include "irmode_t.h" #include "irgraph_t.h" @@ -36,30 +46,37 @@ #include "irdump.h" #include "irdom.h" #include "irtools.h" +#include "irbitset.h" #include "debug.h" #include "xmalloc.h" +#include "iredges.h" #include "beutil.h" #include "besched.h" -#include "benumb_t.h" #include "besched_t.h" #include "belive_t.h" #include "benode_t.h" -#include "bearch.h" +#include "bearch_t.h" #include "beirgmod.h" #include "beifg.h" #include "beinsn_t.h" - +#include "bestatevent.h" +#include "beirg_t.h" +#include "beintlive_t.h" +#include "bera.h" #include "bechordal_t.h" #include "bechordal_draw.h" +#include "bemodule.h" -#define DBG_LEVEL SET_LEVEL_0 -#define DBG_LEVEL_CHECK SET_LEVEL_0 +DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) #define NO_COLOR (-1) #define DUMP_INTERVALS +/* new style assign routine without borders. */ +#undef NEW_STYLE_ASSIGN + typedef struct _be_chordal_alloc_env_t { be_chordal_env_t *chordal_env; @@ -69,7 +86,6 @@ typedef struct _be_chordal_alloc_env_t { bitset_t *colors; /**< The color mask. */ bitset_t *in_colors; /**< Colors used by live in values. */ int colors_n; /**< The number of colors. */ - DEBUG_ONLY(firm_dbg_module_t *constr_dbg;) /**< Debug output for the constraint handler. */ } be_chordal_alloc_env_t; #include "fourcc.h" @@ -117,10 +133,10 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head if(!is_def) { border_t *def; - b = obstack_alloc(&env->obst, sizeof(*b)); + b = obstack_alloc(env->obst, sizeof(*b)); /* also allocate the def and tie it to the use. */ - def = obstack_alloc(&env->obst, sizeof(*def)); + def = obstack_alloc(env->obst, sizeof(*def)); memset(def, 0, sizeof(*def)); b->other_end = def; def->other_end = b; @@ -132,8 +148,8 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head */ set_irn_link(irn, def); - b->magic = BORDER_FOURCC; - def->magic = BORDER_FOURCC; + DEBUG_ONLY(b->magic = BORDER_FOURCC); + DEBUG_ONLY(def->magic = BORDER_FOURCC); } /* @@ -153,7 +169,7 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head b->irn = irn; b->step = step; list_add_tail(&b->list, head); - DBG((env->dbg, LEVEL_5, "\t\t%s adding %+F, step: %d\n", is_def ? "def" : "use", irn, step)); + DBG((dbg, LEVEL_5, "\t\t%s adding %+F, step: %d\n", is_def ? "def" : "use", irn, step)); return b; @@ -167,8 +183,7 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head */ static INLINE int has_reg_class(const be_chordal_env_t *env, const ir_node *irn) { - return arch_irn_has_reg_class(env->birg->main_env->arch_env, irn, -1, env->cls); - // return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn); + return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn); } #define has_limited_constr(req, irn) \ @@ -196,7 +211,7 @@ static bitset_t *get_decisive_partner_regs(bitset_t *bs, const be_operand_t *o1, return bs; } - assert(o1->req.cls == o2->req.cls); + assert(o1->req->cls == o2->req->cls || ! o1->req->cls || ! o2->req->cls); if(bitset_contains(o1->regs, o2->regs)) bitset_copy(bs, o1->regs); @@ -214,21 +229,85 @@ static be_insn_t *chordal_scan_insn(be_chordal_env_t *env, ir_node *irn) ie.ignore_colors = env->ignore_colors; ie.aenv = env->birg->main_env->arch_env; - ie.obst = &env->obst; + ie.obst = env->obst; ie.cls = env->cls; return be_scan_insn(&ie, irn); } static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn) { - be_insn_t *insn = chordal_scan_insn(env, irn); - bitset_t *def_constr = bitset_alloca(env->cls->n_regs); - bitset_t *tmp = bitset_alloca(env->cls->n_regs); - int i; + const be_irg_t *birg = env->birg; + const arch_env_t *aenv = birg->main_env->arch_env; + bitset_t *tmp = bitset_alloca(env->cls->n_regs); + bitset_t *def_constr = bitset_alloca(env->cls->n_regs); + ir_node *bl = get_nodes_block(irn); + be_lv_t *lv = env->birg->lv; + + be_insn_t *insn; + int i, j; + + for (i = get_irn_arity(irn) - 1; i >= 0; --i) { + ir_node *op = get_irn_n(irn, i); + ir_node *copy; + const arch_register_t *reg; + const arch_register_req_t *req; + + if (arch_get_irn_reg_class(aenv, irn, i) != env->cls) + continue; + + reg = arch_get_irn_register(aenv, op); + + if (reg == NULL || !arch_register_type_is(reg, ignore)) + continue; + if(arch_register_type_is(reg, joker)) + continue; + + req = arch_get_register_req(aenv, irn, i); + if (!arch_register_req_is(req, limited)) + continue; + + if (rbitset_is_set(req->limited, reg->index)) + continue; + + copy = be_new_Copy(env->cls, env->irg, bl, op); + be_stat_ev("constr_copy", 1); + + sched_add_before(irn, copy); + set_irn_n(irn, i, copy); + DBG((dbg, LEVEL_3, "inserting ignore arg copy %+F for %+F pos %d\n", copy, irn, i)); + } + + insn = chordal_scan_insn(env, irn); if(!insn->has_constraints) goto end; + /* insert copies for nodes that occur constrained more than once. */ + for(i = insn->use_start; i < insn->n_ops; ++i) { + be_operand_t *op = &insn->ops[i]; + + if(!op->has_constraints) + continue; + + for(j = i + 1; j < insn->n_ops; ++j) { + ir_node *copy; + be_operand_t *a_op = &insn->ops[j]; + + if(a_op->carrier != op->carrier || !a_op->has_constraints) + continue; + + if (be_is_Copy(get_irn_n(insn->irn, a_op->pos))) + continue; + + copy = be_new_Copy(env->cls, env->irg, bl, op->carrier); + be_stat_ev("constr_copy", 1); + + sched_add_before(insn->irn, copy); + set_irn_n(insn->irn, a_op->pos, copy); + DBG((dbg, LEVEL_3, "inserting multiple constr copy %+F for %+F pos %d\n", copy, insn->irn, a_op->pos)); + } + } + /* collect all registers occuring in out constraints. */ for(i = 0; i < insn->use_start; ++i) { be_operand_t *op = &insn->ops[i]; @@ -241,6 +320,7 @@ static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn) and being constrained to a register which also occurs in out constraints. */ for(i = insn->use_start; i < insn->n_ops; ++i) { + ir_node *copy; be_operand_t *op = &insn->ops[i]; bitset_copy(tmp, op->regs); @@ -252,28 +332,30 @@ static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn) 2) lives through the node. 3) is constrained to a register occuring in out constraints. */ - if(op->has_constraints && values_interfere(env->lv, insn->irn, op->carrier) && bitset_popcnt(tmp) > 0) { - ir_node *bl = get_nodes_block(insn->irn); + if(!op->has_constraints || + !values_interfere(birg, insn->irn, op->carrier) || + bitset_popcnt(tmp) == 0) + continue; - /* - only create the copy if the operand is no copy. - this is necessary since the assure constraints phase inserts - Copies and Keeps for operands which must be different from the results. - Additional copies here would destroy this. - */ - if(!be_is_Copy(op->carrier)) { - ir_node *copy = be_new_Copy(env->cls, env->irg, bl, op->carrier); - - sched_add_before(insn->irn, copy); - set_irn_n(insn->irn, op->pos, copy); - DBG((env->dbg, LEVEL_3, "inserting constr copy %+F for %+F pos %d\n", copy, insn->irn, op->pos)); - be_liveness_update(env->lv, op->carrier); - } - } + /* + only create the copy if the operand is no copy. + this is necessary since the assure constraints phase inserts + Copies and Keeps for operands which must be different from the results. + Additional copies here would destroy this. + */ + if (be_is_Copy(get_irn_n(insn->irn, op->pos))) + continue; + + copy = be_new_Copy(env->cls, env->irg, bl, op->carrier); + + sched_add_before(insn->irn, copy); + set_irn_n(insn->irn, op->pos, copy); + DBG((dbg, LEVEL_3, "inserting constr copy %+F for %+F pos %d\n", copy, insn->irn, op->pos)); + be_liveness_update(lv, op->carrier); } end: - obstack_free(&env->obst, insn); + obstack_free(env->obst, insn); return insn->next_insn; } @@ -315,16 +397,19 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t int n_total; const be_operand_t *op = &insn->ops[i]; - if (! values_interfere(env->lv, op->irn, op->carrier) && ! op->partner) { - bitset_clear_all(bs); - bitset_copy(bs, op->regs); - bitset_and(bs, out_op->regs); - n_total = bitset_popcnt(op->regs) + bitset_popcnt(out_op->regs); + if (op->partner != NULL) + continue; + if (values_interfere(env->birg, op->irn, op->carrier)) + continue; - if (bitset_popcnt(bs) > 0 && n_total < smallest_n_regs) { - smallest = i; - smallest_n_regs = n_total; - } + bitset_clear_all(bs); + bitset_copy(bs, op->regs); + bitset_and(bs, out_op->regs); + n_total = bitset_popcnt(op->regs) + bitset_popcnt(out_op->regs); + + if (bitset_popcnt(bs) > 0 && n_total < smallest_n_regs) { + smallest = i; + smallest_n_regs = n_total; } } @@ -337,18 +422,15 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t } -static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, be_insn_t **the_insn) +static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, + be_insn_t **the_insn) { be_chordal_env_t *env = alloc_env->chordal_env; const arch_env_t *aenv = env->birg->main_env->arch_env; be_insn_t *insn = *the_insn; - ir_node *bl = get_nodes_block(insn->irn); - ir_node *copy = NULL; ir_node *perm = NULL; bitset_t *out_constr = bitset_alloca(env->cls->n_regs); - bitset_t *bs = bitset_alloca(env->cls->n_regs); - DEBUG_ONLY(firm_dbg_module_t *dbg = alloc_env->constr_dbg;) - + const ir_edge_t *edge; int i; assert(insn->has_constraints && "only do this for constrained nodes"); @@ -365,75 +447,44 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, be_in bitset_or(out_constr, op->regs); } - /* - Now, figure out which input operand must be copied since it has input - constraints which are also output constraints. - */ - (void) bl; - (void) copy; - (void) bs; - (void) dbg; -#if 0 - for(i = insn->use_start; i < insn->n_ops; ++i) { - be_operand_t *op = &insn->ops[i]; - if(op->has_constraints && (values_interfere(env->lv, op->carrier, insn->irn) || arch_irn_is(aenv, op->carrier, ignore))) { - bitset_copy(bs, op->regs); - bitset_and(bs, out_constr); - - /* - The operand (interfering with the node) has input constraints - which also occur as output constraints, so insert a copy. - */ - if(bitset_popcnt(bs) > 0) { - copy = be_new_Copy(op->req.cls, env->irg, bl, op->carrier); - op->carrier = copy; - sched_add_before(insn->irn, copy); - set_irn_n(insn->irn, op->pos, op->carrier); - - DBG((dbg, LEVEL_2, "adding copy for interfering and constrained op %+F\n", op->carrier)); - } - } - } -#endif - /* Make the Perm, recompute liveness and re-scan the insn since the in operands are now the Projs of the Perm. */ - perm = insert_Perm_after(aenv, env->lv, env->cls, env->dom_front, sched_prev(insn->irn)); + perm = insert_Perm_after(env->birg, env->cls, sched_prev(insn->irn)); /* Registers are propagated by insert_Perm_after(). Clean them here! */ - if(perm) { - const ir_edge_t *edge; + if(perm == NULL) + return NULL; - foreach_out_edge(perm, edge) { - ir_node *proj = get_edge_src_irn(edge); - arch_set_irn_register(aenv, proj, NULL); - } + be_stat_ev("constr_perm", get_irn_arity(perm)); + foreach_out_edge(perm, edge) { + ir_node *proj = get_edge_src_irn(edge); + arch_set_irn_register(aenv, proj, NULL); + } - /* - We also have to re-build the insn since the input operands are now the Projs of - the Perm. Recomputing liveness is also a good idea if a Perm is inserted, since - the live sets may change. - */ - // be_liveness_recompute(env->lv); - obstack_free(&env->obst, insn); - *the_insn = insn = chordal_scan_insn(env, insn->irn); + /* + We also have to re-build the insn since the input operands are now the Projs of + the Perm. Recomputing liveness is also a good idea if a Perm is inserted, since + the live sets may change. + */ + // be_liveness_recompute(lv); + obstack_free(env->obst, insn); + *the_insn = insn = chordal_scan_insn(env, insn->irn); + /* + Copy the input constraints of the insn to the Perm as output + constraints. Succeeding phases (coalescing) will need that. + */ + for(i = insn->use_start; i < insn->n_ops; ++i) { + be_operand_t *op = &insn->ops[i]; + ir_node *proj = op->carrier; /* - Copy the input constraints of the insn to the Perm as output - constraints. Succeeding phases (coalescing will need that). + Note that the predecessor must not be a Proj of the Perm, + since ignore-nodes are not Perm'ed. */ - for(i = insn->use_start; i < insn->n_ops; ++i) { - be_operand_t *op = &insn->ops[i]; - ir_node *proj = op->carrier; - /* - Note that the predecessor must not be a Proj of the Perm, - since ignore-nodes are not Perm'ed. - */ - if(op->has_constraints && is_Proj(proj) && get_Proj_pred(proj) == perm) { - be_set_constr_limited(perm, BE_OUT_POS(get_Proj_proj(proj)), &op->req); - } + if(op->has_constraints && is_Proj(proj) && get_Proj_pred(proj) == perm) { + be_set_constr_limited(perm, BE_OUT_POS(get_Proj_proj(proj)), op->req); } } @@ -442,11 +493,24 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, be_in static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *irn, int *silent) { + const arch_env_t *aenv; + int n_regs; + bitset_t *bs; + ir_node **alloc_nodes; + hungarian_problem_t *bp; + int *assignment; + pmap *partners; + int i, n_alloc; + long col; + const ir_edge_t *edge; + ir_node *perm = NULL; + int match_res, cost; be_chordal_env_t *env = alloc_env->chordal_env; - void *base = obstack_base(&env->obst); + void *base = obstack_base(env->obst); be_insn_t *insn = chordal_scan_insn(env, irn); ir_node *res = insn->next_insn; int be_silent = *silent; + be_irg_t *birg = env->birg; if(insn->pre_colored) { int i; @@ -471,152 +535,156 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i Perms inserted before the constraint handling phase are considered to be correctly precolored. These Perms arise during the ABI handling phase. */ - if(insn->has_constraints) { - const arch_env_t *aenv = env->birg->main_env->arch_env; - int n_regs = env->cls->n_regs; - bitset_t *bs = bitset_alloca(n_regs); - ir_node **alloc_nodes = alloca(n_regs * sizeof(alloc_nodes[0])); - bipartite_t *bp = bipartite_new(n_regs, n_regs); - int *assignment = alloca(n_regs * sizeof(assignment[0])); - pmap *partners = pmap_create(); - DEBUG_ONLY(firm_dbg_module_t *dbg = alloc_env->constr_dbg;) - - int i, n_alloc; - long col; - const ir_edge_t *edge; - ir_node *perm = NULL; + if(!insn->has_constraints) + goto end; - /* - prepare the constraint handling of this node. - Perms are constructed and Copies are created for constrained values - interfering with the instruction. - */ - perm = pre_process_constraints(alloc_env, &insn); + aenv = env->birg->main_env->arch_env; + n_regs = env->cls->n_regs; + bs = bitset_alloca(n_regs); + alloc_nodes = alloca(n_regs * sizeof(alloc_nodes[0])); + bp = hungarian_new(n_regs, n_regs, 2, HUNGARIAN_MATCH_PERFECT); + // bipartite_t *bp = bipartite_new(n_regs, n_regs); + assignment = alloca(n_regs * sizeof(assignment[0])); + partners = pmap_create(); - /* find suitable in operands to the out operands of the node. */ - pair_up_operands(alloc_env, insn); + /* + prepare the constraint handling of this node. + Perms are constructed and Copies are created for constrained values + interfering with the instruction. + */ + perm = pre_process_constraints(alloc_env, &insn); - /* - look at the in/out operands and add each operand (and its possible partner) - to a bipartite graph (left: nodes with partners, right: admissible colors). - */ - for(i = 0, n_alloc = 0; i < insn->n_ops; ++i) { - be_operand_t *op = &insn->ops[i]; + /* find suitable in operands to the out operands of the node. */ + pair_up_operands(alloc_env, insn); - /* - If the operand has no partner or the partner has not been marked - for allocation, determine the admissible registers and mark it - for allocation by associating the node and its partner with the - set of admissible registers via a bipartite graph. - */ - if(!op->partner || !pmap_contains(partners, op->partner->carrier)) { + /* + look at the in/out operands and add each operand (and its possible partner) + to a bipartite graph (left: nodes with partners, right: admissible colors). + */ + for(i = 0, n_alloc = 0; i < insn->n_ops; ++i) { + be_operand_t *op = &insn->ops[i]; - pmap_insert(partners, op->carrier, op->partner ? op->partner->carrier : NULL); - alloc_nodes[n_alloc] = op->carrier; + /* + If the operand has no partner or the partner has not been marked + for allocation, determine the admissible registers and mark it + for allocation by associating the node and its partner with the + set of admissible registers via a bipartite graph. + */ + if(!op->partner || !pmap_contains(partners, op->partner->carrier)) { - DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, op->partner ? op->partner->carrier : NULL)); + pmap_insert(partners, op->carrier, op->partner ? op->partner->carrier : NULL); + alloc_nodes[n_alloc] = op->carrier; - bitset_clear_all(bs); - get_decisive_partner_regs(bs, op, op->partner); + DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, op->partner ? op->partner->carrier : NULL)); - DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier, bs)); + bitset_clear_all(bs); + get_decisive_partner_regs(bs, op, op->partner); - bitset_foreach(bs, col) - bipartite_add(bp, n_alloc, col); + DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier, bs)); - n_alloc++; + bitset_foreach(bs, col) { + hungarian_add(bp, n_alloc, col, 1); + // bipartite_add(bp, n_alloc, col); } + + n_alloc++; } + } - /* - Put all nodes which live by the constrained instruction also to the - allocation bipartite graph. They are considered unconstrained. - */ - if(perm) { - foreach_out_edge(perm, edge) { - ir_node *proj = get_edge_src_irn(edge); + /* + Put all nodes which live through the constrained instruction also to the + allocation bipartite graph. They are considered unconstrained. + */ + if(perm != NULL) { + foreach_out_edge(perm, edge) { + ir_node *proj = get_edge_src_irn(edge); - assert(is_Proj(proj)); + assert(is_Proj(proj)); - if(values_interfere(env->lv, proj, irn) && !pmap_contains(partners, proj)) { - assert(n_alloc < n_regs); - alloc_nodes[n_alloc] = proj; - pmap_insert(partners, proj, NULL); + if(!values_interfere(birg, proj, irn) || pmap_contains(partners, proj)) + continue; - bitset_clear_all(bs); - arch_put_non_ignore_regs(aenv, env->cls, bs); - bitset_foreach(bs, col) - bipartite_add(bp, n_alloc, col); + assert(n_alloc < n_regs); + alloc_nodes[n_alloc] = proj; + pmap_insert(partners, proj, NULL); - n_alloc++; - } + bitset_clear_all(bs); + arch_put_non_ignore_regs(aenv, env->cls, bs); + bitset_andnot(bs, env->ignore_colors); + bitset_foreach(bs, col) { + hungarian_add(bp, n_alloc, col, 1); + // bipartite_add(bp, n_alloc, col); } + + n_alloc++; } + } - /* Compute a valid register allocation. */ - bipartite_matching(bp, assignment); + /* Compute a valid register allocation. */ + hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL); + match_res = hungarian_solve(bp, assignment, &cost, 1); + assert(match_res == 0 && "matching failed"); + //bipartite_matching(bp, assignment); - /* Assign colors obtained from the matching. */ - for(i = 0; i < n_alloc; ++i) { - const arch_register_t *reg; - ir_node *nodes[2]; - int j; + /* Assign colors obtained from the matching. */ + for(i = 0; i < n_alloc; ++i) { + const arch_register_t *reg; + ir_node *nodes[2]; + int j; - assert(assignment[i] >= 0 && "there must have been a register assigned"); - reg = arch_register_for_index(env->cls, assignment[i]); + assert(assignment[i] >= 0 && "there must have been a register assigned"); + reg = arch_register_for_index(env->cls, assignment[i]); - nodes[0] = alloc_nodes[i]; - nodes[1] = pmap_get(partners, alloc_nodes[i]); + nodes[0] = alloc_nodes[i]; + nodes[1] = pmap_get(partners, alloc_nodes[i]); - for(j = 0; j < 2; ++j) { - if(!nodes[j]) - continue; + for(j = 0; j < 2; ++j) { + if(!nodes[j]) + continue; - arch_set_irn_register(aenv, nodes[j], reg); - pset_hinsert_ptr(alloc_env->pre_colored, nodes[j]); - DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", nodes[j], reg->name)); - } + arch_set_irn_register(aenv, nodes[j], reg); + (void) pset_hinsert_ptr(alloc_env->pre_colored, nodes[j]); + DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", nodes[j], reg->name)); } + } + /* Allocate the non-constrained Projs of the Perm. */ + if(perm != NULL) { + bitset_clear_all(bs); - /* Allocate the non-constrained Projs of the Perm. */ - if(perm) { - - bitset_clear_all(bs); - - /* Put the colors of all Projs in a bitset. */ - foreach_out_edge(perm, edge) { - ir_node *proj = get_edge_src_irn(edge); - const arch_register_t *reg = arch_get_irn_register(aenv, proj); + /* Put the colors of all Projs in a bitset. */ + foreach_out_edge(perm, edge) { + ir_node *proj = get_edge_src_irn(edge); + const arch_register_t *reg = arch_get_irn_register(aenv, proj); - if(reg != NULL) - bitset_set(bs, reg->index); - } + if(reg != NULL) + bitset_set(bs, reg->index); + } - /* Assign the not yet assigned Projs of the Perm a suitable color. */ - foreach_out_edge(perm, edge) { - ir_node *proj = get_edge_src_irn(edge); - const arch_register_t *reg = arch_get_irn_register(aenv, proj); + /* Assign the not yet assigned Projs of the Perm a suitable color. */ + foreach_out_edge(perm, edge) { + ir_node *proj = get_edge_src_irn(edge); + const arch_register_t *reg = arch_get_irn_register(aenv, proj); - DBG((dbg, LEVEL_2, "\tchecking reg of %+F: %s\n", proj, reg ? reg->name : "")); + DBG((dbg, LEVEL_2, "\tchecking reg of %+F: %s\n", proj, reg ? reg->name : "")); - if(reg == NULL) { - col = get_next_free_reg(alloc_env, bs); - reg = arch_register_for_index(env->cls, col); - bitset_set(bs, reg->index); - arch_set_irn_register(aenv, proj, reg); - pset_insert_ptr(alloc_env->pre_colored, proj); - DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", proj, reg->name)); - } + if(reg == NULL) { + col = get_next_free_reg(alloc_env, bs); + reg = arch_register_for_index(env->cls, col); + bitset_set(bs, reg->index); + arch_set_irn_register(aenv, proj, reg); + pset_insert_ptr(alloc_env->pre_colored, proj); + DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", proj, reg->name)); } } - - bipartite_free(bp); - pmap_destroy(partners); } + //bipartite_free(bp); + hungarian_free(bp); + pmap_destroy(partners); + end: - obstack_free(&env->obst, base); + obstack_free(env->obst, base); return res; } @@ -669,20 +737,20 @@ static void pressure(ir_node *block, void *env_ptr) be_chordal_env_t *env = alloc_env->chordal_env; bitset_t *live = alloc_env->live; ir_node *irn; - DEBUG_ONLY(firm_dbg_module_t *dbg = env->dbg;) + be_lv_t *lv = env->birg->lv; int i, n; unsigned step = 0; unsigned pressure = 0; struct list_head *head; - pset *live_in = be_lv_pset_put_in(env->lv, block, pset_new_ptr_default()); - pset *live_end = be_lv_pset_put_end(env->lv, block, pset_new_ptr_default()); + pset *live_in = be_lv_pset_put_in(lv, block, pset_new_ptr_default()); + pset *live_end = be_lv_pset_put_end(lv, block, pset_new_ptr_default()); DBG((dbg, LEVEL_1, "Computing pressure in block %+F\n", block)); bitset_clear_all(live); /* Set up the border list in the block info */ - head = obstack_alloc(&env->obst, sizeof(*head)); + head = obstack_alloc(env->obst, sizeof(*head)); INIT_LIST_HEAD(head); assert(pmap_get(env->border_heads, block) == NULL); pmap_insert(env->border_heads, block, head); @@ -770,11 +838,11 @@ static void assign(ir_node *block, void *env_ptr) bitset_t *in_colors = alloc_env->in_colors; const arch_env_t *arch_env = env->birg->main_env->arch_env; struct list_head *head = get_block_border_head(env, block); - pset *live_in = be_lv_pset_put_in(env->lv, block, pset_new_ptr_default()); + be_lv_t *lv = env->birg->lv; + pset *live_in = be_lv_pset_put_in(lv, block, pset_new_ptr_default()); const ir_node *irn; border_t *b; - DEBUG_ONLY(firm_dbg_module_t *dbg = env->dbg;) bitset_clear_all(colors); bitset_clear_all(live); @@ -812,8 +880,7 @@ static void assign(ir_node *block, void *env_ptr) } /* - * Mind that the sequence - * of defs from back to front defines a perfect + * Mind that the sequence of defs from back to front defines a perfect * elimination order. So, coloring the definitions from first to last * will work. */ @@ -826,17 +893,15 @@ static void assign(ir_node *block, void *env_ptr) * Assign a color, if it is a local def. Global defs already have a * color. */ - if(b->is_def && !be_is_live_in(env->lv, block, irn)) { + if(b->is_def && !be_is_live_in(lv, block, irn)) { const arch_register_t *reg; int col = NO_COLOR; - if(pset_find_ptr(alloc_env->pre_colored, irn) || ignore) { + if(ignore || pset_find_ptr(alloc_env->pre_colored, irn)) { reg = arch_get_irn_register(arch_env, irn); col = reg->index; assert(!bitset_is_set(colors, col) && "pre-colored register must be free"); - } - - else { + } else { col = get_next_free_reg(alloc_env, colors); reg = arch_register_for_index(env->cls, col); assert(arch_get_irn_register(arch_env, irn) == NULL && "This node must not have been assigned a register yet"); @@ -860,7 +925,11 @@ static void assign(ir_node *block, void *env_ptr) assert(reg && "Register must have been assigned"); col = arch_register_get_index(reg); - assert(bitset_is_set(live, nr) && "Cannot have a non live use"); +#ifndef NDEBUG + if(!arch_register_type_is(reg, ignore)) { + assert(bitset_is_set(live, nr) && "Cannot have a non live use"); + } +#endif bitset_clear(colors, col); bitset_clear(live, nr); @@ -870,15 +939,146 @@ static void assign(ir_node *block, void *env_ptr) del_pset(live_in); } +/** + * A new assign... + */ +static void assign_new(ir_node *block, be_chordal_alloc_env_t *alloc_env, bitset_t *live_end_dom) +{ + be_chordal_env_t *env = alloc_env->chordal_env; + bitset_t *colors = alloc_env->colors; + bitset_t *in_colors = alloc_env->in_colors; + bitset_t *live = bitset_irg_malloc(env->irg); + const arch_env_t *arch_env = env->birg->main_env->arch_env; + be_irg_t *birg = env->birg; + lv_chk_t *lv = be_get_birg_liveness_chk(birg); + + bitset_pos_t elm; + ir_node *irn; + + bitset_clear_all(colors); + bitset_clear_all(in_colors); + + /* + * All variables which are live in to this block are live out + * of the immediate dominator thanks to SSA properties. As we + * have already visited the immediate dominator, we know these + * variables. The only tjing left is to check wheather they are live + * in here (they also could be phi arguments to some ohi not + * in this block, hence we have to check). + */ + bitset_foreach (live_end_dom, elm) { + ir_node *irn = get_idx_irn(env->irg, elm); + if (lv_chk_bl_in(lv, block, irn)) { + const arch_register_t *reg = arch_get_irn_register(arch_env, irn); + int col; + + assert(be_is_live_in(env->birg->lv, block, irn)); + assert(reg && "Node must have been assigned a register"); + col = arch_register_get_index(reg); + + DBG((dbg, LEVEL_4, "%+F has reg %s\n", irn, reg->name)); + + /* Mark the color of the live in value as used. */ + bitset_set(colors, col); + bitset_set(in_colors, col); + + /* Mark the value live in. */ + bitset_set(live, elm); + } + + else { + assert(!be_is_live_in(env->birg->lv, block, irn)); + } + } + + /* + * Mind that the sequence of defs from back to front defines a perfect + * elimination order. So, coloring the definitions from first to last + * will work. + */ + sched_foreach (block, irn) { + int nr = get_irn_idx(irn); + int ignore = arch_irn_is(arch_env, irn, ignore); + + /* Clear the color upon a last use. */ + if(!is_Phi(irn)) { + int i; + for (i = get_irn_arity(irn) - 1; i >= 0; --i) { + ir_node *op = get_irn_n(irn, i); + + /* + * If the reg class matches and the operand is not live after + * the node, irn is a last use of op and the register can + * be freed. + */ + if (has_reg_class(env, op)) { + if (!be_lv_chk_after_irn(birg, op, irn)) { + const arch_register_t *reg = arch_get_irn_register(arch_env, op); + int col; + + assert(reg && "Register must have been assigned"); + col = arch_register_get_index(reg); + bitset_clear(colors, col); + bitset_clear(live, nr); + } + } + } + } + + if (has_reg_class(env, irn)) { + const arch_register_t *reg; + int col = NO_COLOR; + + /* + * Assign a color, if it is a local def. Global defs already have a + * color. + */ + if(ignore || pset_find_ptr(alloc_env->pre_colored, irn)) { + reg = arch_get_irn_register(arch_env, irn); + col = reg->index; + assert(!bitset_is_set(colors, col) && "pre-colored register must be free"); + } else { + col = get_next_free_reg(alloc_env, colors); + reg = arch_register_for_index(env->cls, col); + assert(arch_get_irn_register(arch_env, irn) == NULL && "This node must not have been assigned a register yet"); + assert(!arch_register_type_is(reg, ignore) && "Must not assign ignore register"); + } + + bitset_set(colors, col); + arch_set_irn_register(arch_env, irn, reg); + + DBG((dbg, LEVEL_1, "\tassigning register %s(%d) to %+F\n", arch_register_get_name(reg), col, irn)); + + assert(!bitset_is_set(live, nr) && "Value's definition must not have been encountered"); + bitset_set(live, nr); + } + + } + + dominates_for_each (block, irn) { + assign_new(irn, alloc_env, live); + } + + bitset_free(live); +} + void be_ra_chordal_color(be_chordal_env_t *chordal_env) { be_chordal_alloc_env_t env; char buf[256]; + be_irg_t *birg = chordal_env->birg; + const arch_register_class_t *cls = chordal_env->cls; - int colors_n = arch_register_class_n_regs(chordal_env->cls); + int colors_n = arch_register_class_n_regs(cls); ir_graph *irg = chordal_env->irg; + int allocatable_regs = colors_n - be_put_ignore_regs(birg, cls, NULL); + /* some special classes contain only ignore regs, no work to be done */ + if(allocatable_regs == 0) + return; + be_assure_dom_front(birg); + be_assure_liveness(birg); assure_doms(irg); env.chordal_env = chordal_env; @@ -887,8 +1087,6 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env) env.tmp_colors = bitset_alloca(colors_n); env.in_colors = bitset_alloca(colors_n); env.pre_colored = pset_new_ptr_default(); - FIRM_DBG_REGISTER(env.constr_dbg, "firm.be.chordal.constr"); - /* Handle register targeting constraints */ dom_tree_walk_irg(irg, constraints, NULL, &env); @@ -898,16 +1096,17 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env) be_dump(chordal_env->irg, buf, dump_ir_block_graph_sched); } - be_numbering(irg); env.live = bitset_malloc(get_irg_last_idx(chordal_env->irg)); /* First, determine the pressure */ dom_tree_walk_irg(irg, pressure, NULL, &env); /* Assign the colors */ +#ifdef NEW_STYLE_ASSIGN + assign_new(get_irg_start_block(irg), &env, env.live); +#else dom_tree_walk_irg(irg, assign, NULL, &env); - - be_numbering_done(irg); +#endif if(chordal_env->opts->dump_flags & BE_CH_DUMP_TREE_INTV) { plotter_t *plotter; @@ -920,3 +1119,10 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env) bitset_free(env.live); del_pset(env.pre_colored); } + +void be_init_chordal(void) +{ + FIRM_DBG_REGISTER(dbg, "firm.be.chordal"); +} + +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_chordal);