X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbelower.c;h=2c18953bff3e5f341df24bc72c9ee304f4bafa4b;hb=cc8920adef35bcde828590f537e0147f784555b7;hp=2ca04d5dcac77029dff1216fa23f93c51fa04e12;hpb=590df14827a8ff5a5b53c506ac654876107bd8c9;p=libfirm diff --git a/ir/be/belower.c b/ir/be/belower.c index 2ca04d5dc..2c18953bf 100644 --- a/ir/be/belower.c +++ b/ir/be/belower.c @@ -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. * @@ -24,9 +24,7 @@ * @date 14.12.2005 * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include @@ -35,9 +33,11 @@ #include "irhooks.h" #include "xmalloc.h" #include "irnodeset.h" +#include "irnodemap.h" #include "irgmod.h" #include "iredges_t.h" #include "irgwalk.h" +#include "array_t.h" #include "bearch_t.h" #include "belower.h" @@ -45,35 +45,34 @@ #include "besched_t.h" #include "bestat.h" #include "bessaconstr.h" -#include "benodesets.h" #include "beintlive_t.h" #undef KEEP_ALIVE_COPYKEEP_HACK -/* associates op with it's copy and CopyKeep */ +DEBUG_ONLY(static firm_dbg_module_t *dbg;) +DEBUG_ONLY(static firm_dbg_module_t *dbg_constr;) +DEBUG_ONLY(static firm_dbg_module_t *dbg_permmove;) + +/** Associates an ir_node with it's copy and CopyKeep. */ typedef struct { - ir_node *op; /* an irn which must be different */ - ir_nodeset_t copies; /* all non-spillable copies of this irn */ + ir_nodeset_t copies; /**< all non-spillable copies of this irn */ const arch_register_class_t *cls; } op_copy_assoc_t; -/* environment for constraints */ +/** Environment for constraints. */ typedef struct { be_irg_t *birg; - pset *op_set; + ir_nodemap_t op_set; struct obstack obst; - DEBUG_ONLY(firm_dbg_module_t *dbg;) } constraint_env_t; -/* lowering walker environment */ +/** Lowering walker environment. */ typedef struct _lower_env_t { be_irg_t *birg; - const arch_env_t *arch_env; unsigned do_copy : 1; - DEBUG_ONLY(firm_dbg_module_t *dbg_module;) } lower_env_t; -/* holds a perm register pair */ +/** Holds a Perm register pair. */ typedef struct _reg_pair_t { const arch_register_t *in_reg; /**< a perm IN register */ ir_node *in_node; /**< the in node to which the register belongs */ @@ -91,34 +90,14 @@ typedef enum _perm_type_t { PERM_COPY } perm_type_t; -/* structure to represent cycles or chains in a perm */ +/** Structure to represent cycles or chains in a Perm. */ typedef struct _perm_cycle_t { const arch_register_t **elems; /**< the registers in the cycle */ int n_elems; /**< number of elements in the cycle */ perm_type_t type; /**< type (CHAIN or CYCLE) */ } perm_cycle_t; -// -/* Compare the two operands */ -static int cmp_op_copy_assoc(const void *a, const void *b) { - const op_copy_assoc_t *op1 = a; - const op_copy_assoc_t *op2 = b; - - return op1->op != op2->op; -} - -/* Compare the in registers of two register pairs */ -static int compare_reg_pair(const void *a, const void *b) { - const reg_pair_t *pair_a = a; - const reg_pair_t *pair_b = b; - - if (pair_a->in_reg->index > pair_b->in_reg->index) - return 1; - else - return -1; -} - -/* returns the number register pairs marked as checked */ +/** returns the number register pairs marked as checked. */ static int get_n_checked_pairs(reg_pair_t *pairs, int n) { int i, n_checked = 0; @@ -131,102 +110,129 @@ static int get_n_checked_pairs(reg_pair_t *pairs, int n) { } /** - * Gets the node corresponding to a register from an array of register pairs. + * Gets the node corresponding to an IN register from an array of register pairs. * NOTE: The given registers pairs and the register to look for must belong * to the same register class. * * @param pairs The array of register pairs * @param n The number of pairs * @param reg The register to look for - * @param in_out 0 == look for IN register, 1 == look for OUT register * @return The corresponding node or NULL if not found */ -static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) { +static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg) { int i; - if (in_out) { - for (i = 0; i < n; i++) { - /* out register matches */ - if (pairs[i].out_reg->index == reg->index) - return pairs[i].out_node; - } + for (i = 0; i < n; i++) { + /* in register matches */ + if (pairs[i].in_reg->index == reg->index) + return pairs[i].in_node; } - else { - for (i = 0; i < n; i++) { - /* in register matches */ - if (pairs[i].in_reg->index == reg->index) - return pairs[i].in_node; - } + + return NULL; +} + +/** + * Gets the node corresponding to an OUT register from an array of register pairs. + * NOTE: The given registers pairs and the register to look for must belong + * to the same register class. + * + * @param pairs The array of register pairs + * @param n The number of pairs + * @param reg The register to look for + * @return The corresponding node or NULL if not found + */ +static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg) { + int i; + + for (i = 0; i < n; i++) { + /* out register matches */ + if (pairs[i].out_reg->index == reg->index) + return pairs[i].out_node; } return NULL; } /** - * Gets the index in the register pair array where the in/out register + * Gets the index in the register pair array where the in register * corresponds to reg_idx. * * @param pairs The array of register pairs * @param n The number of pairs * @param reg The register index to look for - * @param in_out 0 == look for IN register, 1 == look for OUT register + * * @return The corresponding index in pairs or -1 if not found */ -static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) { +static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) { int i; - if (in_out) { - for (i = 0; i < n; i++) { - /* out register matches */ - if ((int) pairs[i].out_reg->index == reg_idx) - return i; - } - } - else { - for (i = 0; i < n; i++) { - /* in register matches */ - if ((int) pairs[i].in_reg->index == reg_idx) - return i; - } + for (i = 0; i < n; i++) { + /* in register matches */ + if (pairs[i].in_reg->index == reg_idx) + return i; } + return -1; +} + +/** + * Gets the index in the register pair array where the out register + * corresponds to reg_idx. + * + * @param pairs The array of register pairs + * @param n The number of pairs + * @param reg The register index to look for + * + * @return The corresponding index in pairs or -1 if not found + */ +static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) { + int i; + for (i = 0; i < n; i++) { + /* out register matches */ + if (pairs[i].out_reg->index == reg_idx) + return i; + } return -1; } /** - * Gets an array of register pairs and tries to identify a cycle or chain starting - * at position start. + * Gets an array of register pairs and tries to identify a cycle or chain + * starting at position start. + * + * @param cycle Variable to hold the cycle + * @param pairs Array of register pairs + * @param n length of the pairs array + * @param start Index to start * - * @param cycle Variable to hold the cycle - * @param pairs Array of register pairs - * @param start Index to start * @return The cycle or chain */ -static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) { - int head = pairs[start].in_reg->index; - int cur_idx = pairs[start].out_reg->index; - int cur_pair_idx = start; - int n_pairs_done = get_n_checked_pairs(pairs, n); - int idx; - perm_type_t cycle_tp = PERM_CYCLE; +static void get_perm_cycle(perm_cycle_t *const cycle, + reg_pair_t *const pairs, + int const n, + int start) +{ + int head = pairs[start].in_reg->index; + int cur_idx = pairs[start].out_reg->index; + int const n_pairs_done = get_n_checked_pairs(pairs, n); + perm_type_t cycle_tp = PERM_CYCLE; + int idx; /* We could be right in the middle of a chain, so we need to find the start */ while (head != cur_idx) { /* goto previous register in cycle or chain */ - cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1); + int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head); if (cur_pair_idx < 0) { cycle_tp = PERM_CHAIN; break; - } - else { + } else { head = pairs[cur_pair_idx].in_reg->index; start = cur_pair_idx; } } /* assume worst case: all remaining pairs build a cycle or chain */ - cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0])); + cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2); cycle->n_elems = 2; /* initial number of elements is 2 */ cycle->elems[0] = pairs[start].in_reg; cycle->elems[1] = pairs[start].out_reg; @@ -237,7 +243,7 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int /* check for cycle or end of a chain */ while (cur_idx != head) { /* goto next register in cycle or chain */ - cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0); + int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx); if (cur_pair_idx < 0) break; @@ -248,8 +254,7 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int if (cur_idx != head) { cycle->elems[idx++] = pairs[cur_pair_idx].out_reg; cycle->n_elems++; - } - else { + } else { /* we are there where we started -> CYCLE */ cycle->type = PERM_CYCLE; } @@ -257,18 +262,16 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int /* mark all pairs having one in/out register with cycle in common as checked */ for (idx = 0; idx < cycle->n_elems; idx++) { - cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0); + int cur_pair_idx; + cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index); if (cur_pair_idx >= 0) pairs[cur_pair_idx].checked = 1; - cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1); - + cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index); if (cur_pair_idx >= 0) pairs[cur_pair_idx].checked = 1; } - - return cycle; } /** @@ -279,386 +282,361 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int * * @param irn The perm node * @param block The block the perm node belongs to - * @param walk_env The environment + * @param env The lowerer environment */ -static void lower_perm_node(ir_node *irn, void *walk_env) { - ir_graph *irg = get_irn_irg(irn); - const arch_register_class_t *reg_class; - const arch_env_t *arch_env; - lower_env_t *env = walk_env; - int real_size = 0; - int keep_perm = 0; - int n, i, pn, do_copy, j, n_ops; - reg_pair_t *pairs; - const ir_edge_t *edge; - perm_cycle_t *cycle; - ir_node *sched_point, *block, *in[2]; - ir_node *arg1, *arg2, *res1, *res2; - ir_node *cpyxchg = NULL; - DEBUG_ONLY(firm_dbg_module_t *mod;) - - arch_env = env->arch_env; - do_copy = env->do_copy; - DEBUG_ONLY(mod = env->dbg_module;) - block = get_nodes_block(irn); - - /* - Get the schedule predecessor node to the perm - NOTE: This works with auto-magic. If we insert the - new copy/exchange nodes after this node, everything - should be ok. - */ - sched_point = sched_prev(irn); - DBG((mod, LEVEL_1, "perm: %+F\n", irn)); - DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point)); +static void lower_perm_node(ir_node *irn, lower_env_t *env) +{ + const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class; + ir_graph *const irg = get_irn_irg(irn); + ir_node *const block = get_nodes_block(irn); + int const n = get_irn_arity(irn); + reg_pair_t *const pairs = alloca(n * sizeof(pairs[0])); + int keep_perm = 0; + int do_copy = env->do_copy; + /* Get the schedule predecessor node to the perm. + * NOTE: This works with auto-magic. If we insert the new copy/exchange + * nodes after this node, everything should be ok. */ + ir_node * sched_point = sched_prev(irn); + const ir_edge_t * edge; + int i; + + DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point)); assert(sched_point && "Perm is not scheduled or has no predecessor"); - n = get_irn_arity(irn); assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different"); - reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class; - pairs = alloca(n * sizeof(pairs[0])); - /* build the list of register pairs (in, out) */ i = 0; foreach_out_edge(irn, edge) { - pairs[i].out_node = get_edge_src_irn(edge); - pn = get_Proj_proj(pairs[i].out_node); - pairs[i].in_node = get_irn_n(irn, pn); - - pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node); - pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node); - - pairs[i].checked = 0; - i++; + reg_pair_t *const pair = &pairs[i++]; + ir_node *const out = get_edge_src_irn(edge); + long const pn = get_Proj_proj(out); + ir_node *const in = get_irn_n(irn, pn); + + pair->in_node = in; + pair->in_reg = arch_get_irn_register(in); + pair->out_node = out; + pair->out_reg = arch_get_irn_register(out); + pair->checked = 0; } - /* sort the register pairs by the indices of the in registers */ - qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair); - - /* Mark all equal pairs as checked, and exchange the OUT proj with - the IN node. */ + /* Mark all equal pairs as checked, and exchange the OUT proj with the IN + * node. */ for (i = 0; i < n; i++) { - if (pairs[i].in_reg->index == pairs[i].out_reg->index) { - DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n", - irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name)); - - /* We have to check for a special case: - The in-node could be a Proj from a Perm. In this case, - we need to correct the projnum */ - if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) { - set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node)); - } + reg_pair_t *const pair = &pairs[i]; + + if (pair->in_reg->index != pair->out_reg->index) + continue; - /* reroute the edges from the proj to the argument */ - exchange(pairs[i].out_node, pairs[i].in_node); - //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg); - //set_irn_n(pairs[i].out_node, 0, new_Bad()); + DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n", + irn, pair->in_node, pair->out_node, pair->out_reg->name)); - pairs[i].checked = 1; - } + /* reroute the edges from the proj to the argument */ + exchange(pair->out_node, pair->in_node); + + pair->checked = 1; } /* Set do_copy to 0 if it's on but we have no free register */ + /* TODO check for free register */ if (do_copy) { do_copy = 0; } - real_size = n - get_n_checked_pairs(pairs, n); - - be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size); - /* check for cycles and chains */ while (get_n_checked_pairs(pairs, n) < n) { - i = n_ops = 0; + perm_cycle_t cycle; + int j; /* go to the first not-checked pair */ - while (pairs[i].checked) i++; - cycle = xcalloc(1, sizeof(*cycle)); - cycle = get_perm_cycle(cycle, pairs, n, i); + for (i = 0; pairs[i].checked; ++i) {} + get_perm_cycle(&cycle, pairs, n, i); - DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle")); - for (j = 0; j < cycle->n_elems; j++) { - DB((mod, LEVEL_1, " %s", cycle->elems[j]->name)); + DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle")); + for (j = 0; j < cycle.n_elems; j++) { + DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name)); } - DB((mod, LEVEL_1, "\n")); - - /* - We don't need to do anything if we have a Perm with two - elements which represents a cycle, because those nodes - already represent exchange nodes - */ - if (n == 2 && cycle->type == PERM_CYCLE) { - free(cycle); + DB((dbg, LEVEL_1, "\n")); + + if (n == 2 && cycle.type == PERM_CYCLE) { + /* We don't need to do anything if we have a Perm with two elements + * which represents a cycle, because those nodes already represent + * exchange nodes */ keep_perm = 1; - continue; - } + } else { + /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and + * insert copy to/from it before/after the copy cascade (this + * reduces the cycle into a chain) */ + + /* build copy/swap nodes from back to front */ + for (i = cycle.n_elems - 2; i >= 0; i--) { + ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]); + ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]); + + ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]); + ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]); + /* If we have a cycle and don't copy: we need to create exchange + * nodes + * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs + * IN_1 = in node with register i + * IN_2 = in node with register i + 1 + * OUT_1 = out node with register i + 1 + * OUT_2 = out node with register i */ + if (cycle.type == PERM_CYCLE && !do_copy) { + ir_node *in[2]; + ir_node *cpyxchg; + + in[0] = arg1; + in[1] = arg2; + + /* At this point we have to handle the following problem: + * + * If we have a cycle with more than two elements, then this + * could correspond to the following Perm node: + * + * +----+ +----+ +----+ + * | r1 | | r2 | | r3 | + * +-+--+ +-+--+ +--+-+ + * | | | + * | | | + * +-+--------+---------+-+ + * | Perm | + * +-+--------+---------+-+ + * | | | + * | | | + * +-+--+ +-+--+ +--+-+ + * |Proj| |Proj| |Proj| + * | r2 | | r3 | | r1 | + * +----+ +----+ +----+ + * + * This node is about to be split up into two 2x Perm's for + * which we need 4 Proj's and the one additional Proj of the + * first Perm has to be one IN of the second. So in general + * we need to create one additional Proj for each "middle" + * Perm and set this to one in node of the successor Perm. */ + + DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n", + irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name)); + DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n", + irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name)); + + cpyxchg = be_new_Perm(reg_class, irg, block, 2, in); + + if (i > 0) { + /* cycle is not done yet */ + int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index); + + /* create intermediate proj */ + res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0); + + /* set as in for next Perm */ + pairs[pidx].in_node = res1; + } -//TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after -// the copy cascade (this reduces the cycle into a chain) - - /* build copy/swap nodes from back to front */ - for (i = cycle->n_elems - 2; i >= 0; i--) { - arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0); - arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0); - - res1 = get_node_for_register(pairs, n, cycle->elems[i], 1); - res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1); - /* - If we have a cycle and don't copy: we need to create exchange nodes - NOTE: An exchange node is a perm node with 2 INs and 2 OUTs - IN_1 = in node with register i - IN_2 = in node with register i + 1 - OUT_1 = out node with register i + 1 - OUT_2 = out node with register i - */ - if (cycle->type == PERM_CYCLE && !do_copy) { - in[0] = arg1; - in[1] = arg2; - - /* At this point we have to handle the following problem: */ - /* */ - /* If we have a cycle with more than two elements, then */ - /* this could correspond to the following Perm node: */ - /* */ - /* +----+ +----+ +----+ */ - /* | r1 | | r2 | | r3 | */ - /* +-+--+ +-+--+ +--+-+ */ - /* | | | */ - /* | | | */ - /* +-+--------+---------+-+ */ - /* | Perm | */ - /* +-+--------+---------+-+ */ - /* | | | */ - /* | | | */ - /* +-+--+ +-+--+ +--+-+ */ - /* |Proj| |Proj| |Proj| */ - /* | r2 | | r3 | | r1 | */ - /* +----+ +----+ +----+ */ - /* */ - /* This node is about to be split up into two 2x Perm's */ - /* for which we need 4 Proj's and the one additional Proj */ - /* of the first Perm has to be one IN of the second. So in */ - /* general we need to create one additional Proj for each */ - /* "middle" Perm and set this to one in node of the successor */ - /* Perm. */ - - DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n", - irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name)); - DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n", - irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name)); - - cpyxchg = be_new_Perm(reg_class, irg, block, 2, in); - n_ops++; - - if (i > 0) { - /* cycle is not done yet */ - int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0); - - /* create intermediate proj */ - res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0); - - /* set as in for next Perm */ - pairs[pidx].in_node = res1; - } + set_Proj_pred(res2, cpyxchg); + set_Proj_proj(res2, 0); + set_Proj_pred(res1, cpyxchg); + set_Proj_proj(res1, 1); - set_Proj_pred(res2, cpyxchg); - set_Proj_proj(res2, 0); - set_Proj_pred(res1, cpyxchg); - set_Proj_proj(res1, 1); + arch_set_irn_register(res2, cycle.elems[i + 1]); + arch_set_irn_register(res1, cycle.elems[i]); - arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]); - arch_set_irn_register(arch_env, res1, cycle->elems[i]); + /* insert the copy/exchange node in schedule after the magic schedule node (see above) */ + sched_add_after(sched_point, cpyxchg); - /* insert the copy/exchange node in schedule after the magic schedule node (see above) */ - sched_add_after(sched_point, cpyxchg); + DBG((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point)); - DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point)); + /* set the new scheduling point */ + sched_point = res1; + } else { + ir_node *cpyxchg; - /* set the new scheduling point */ - sched_point = res1; - } - else { - DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n", - irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name)); + DBG((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n", + irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name)); - cpyxchg = be_new_Copy(reg_class, irg, block, arg1); - arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]); - n_ops++; + cpyxchg = be_new_Copy(reg_class, irg, block, arg1); + arch_set_irn_register(cpyxchg, cycle.elems[i + 1]); - /* exchange copy node and proj */ - exchange(res2, cpyxchg); + /* exchange copy node and proj */ + exchange(res2, cpyxchg); - /* insert the copy/exchange node in schedule after the magic schedule node (see above) */ - sched_add_after(sched_point, cpyxchg); + /* insert the copy/exchange node in schedule after the magic schedule node (see above) */ + sched_add_after(sched_point, cpyxchg); - /* set the new scheduling point */ - sched_point = cpyxchg; + /* set the new scheduling point */ + sched_point = cpyxchg; + } } } - be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops); - - free((void *) cycle->elems); - free(cycle); + free((void*)cycle.elems); } /* remove the perm from schedule */ - if (! keep_perm) { + if (!keep_perm) { sched_remove(irn); - be_kill_node(irn); + kill_node(irn); } } -static int get_n_out_edges(const ir_node *irn) { - const ir_edge_t *edge; - int cnt = 0; - - foreach_out_edge(irn, edge) { - cnt++; - } - - return cnt; -} - -static INLINE ir_node *belower_skip_proj(ir_node *irn) { - while(is_Proj(irn)) - irn = get_Proj_pred(irn); - return irn; +static int has_irn_users(const ir_node *irn) { + return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0; } -static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) { - const arch_env_t *arch_env = be_get_birg_arch_env(env->birg); - ir_node *block = get_nodes_block(irn); - ir_node *cur_node; +static ir_node *find_copy(ir_node *irn, ir_node *op) +{ + ir_node *cur_node; - for (cur_node = sched_prev(irn); - ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block; - cur_node = sched_prev(cur_node)) - { - if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill)) + for (cur_node = irn;;) { + cur_node = sched_prev(cur_node); + if (! be_is_Copy(cur_node)) + return NULL; + if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill)) return cur_node; } - - return NULL; } static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) { - be_irg_t *birg = env->birg; - ir_graph *irg = be_get_birg_irg(birg); - pset *op_set = env->op_set; - const arch_env_t *arch_env = be_get_birg_arch_env(birg); - ir_node *block = get_nodes_block(irn); - const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1); - ir_node *in[2], *keep, *cpy; - op_copy_assoc_t key, *entry; - DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;) - - if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) { - DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn)); + ir_graph *irg; + ir_nodemap_t *op_set; + ir_node *block; + const arch_register_class_t *cls; + ir_node *keep, *cpy; + op_copy_assoc_t *entry; + + if (arch_irn_is(other_different, ignore) || + !mode_is_datab(get_irn_mode(other_different))) { + DBG((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn)); return; } + irg = be_get_birg_irg(env->birg); + op_set = &env->op_set; + block = get_nodes_block(irn); + cls = arch_get_irn_reg_class(other_different, -1); + /* Make a not spillable copy of the different node */ /* this is needed because the different irn could be */ /* in block far far away */ /* The copy is optimized later if not needed */ - /* check if already exists such a copy in the schedule immediatly before */ - cpy = find_copy(env, belower_skip_proj(irn), other_different); + /* check if already exists such a copy in the schedule immediately before */ + cpy = find_copy(skip_Proj(irn), other_different); if (! cpy) { cpy = be_new_Copy(cls, irg, block, other_different); be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill); - DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different)); + DBG((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different)); + } else { + DBG((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different)); } - else { - DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different)); - } - - in[0] = irn; - in[1] = cpy; /* Add the Keep resp. CopyKeep and reroute the users */ /* of the other_different irn in case of CopyKeep. */ - if (get_n_out_edges(other_different) == 0) { - keep = be_new_Keep(cls, irg, block, 2, in); - } - else { + if (has_irn_users(other_different)) { keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different)); be_node_set_reg_class(keep, 1, cls); + } else { + ir_node *in[2]; + + in[0] = irn; + in[1] = cpy; + keep = be_new_Keep(cls, irg, block, 2, in); } - DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy)); + DBG((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy)); /* insert copy and keep into schedule */ assert(sched_is_scheduled(irn) && "need schedule to assure constraints"); if (! sched_is_scheduled(cpy)) - sched_add_before(belower_skip_proj(irn), cpy); + sched_add_before(skip_Proj(irn), cpy); sched_add_after(irn, keep); - /* insert the other different and it's copies into the set */ - key.op = other_different; - entry = pset_find(op_set, &key, nodeset_hash(other_different)); - + /* insert the other different and it's copies into the map */ + entry = ir_nodemap_get(op_set, other_different); if (! entry) { - entry = obstack_alloc(&env->obst, sizeof(*entry)); + entry = obstack_alloc(&env->obst, sizeof(*entry)); + entry->cls = cls; ir_nodeset_init(&entry->copies); - entry->op = other_different; - entry->cls = cls; + + ir_nodemap_insert(op_set, other_different, entry); } /* insert copy */ ir_nodeset_insert(&entry->copies, cpy); /* insert keep in case of CopyKeep */ - if (be_is_CopyKeep(keep)) { + if (be_is_CopyKeep(keep)) ir_nodeset_insert(&entry->copies, keep); - } - - pset_insert(op_set, entry, nodeset_hash(other_different)); } /** - * Checks if node has a should_be_different constraint in output - * and adds a Keep then to assure the constraint. + * Checks if node has a must_be_different constraint in output and adds a Keep + * then to assure the constraint. + * + * @param irn the node to check + * @param skipped_irn if irn is a Proj node, its predecessor, else irn + * @param env the constraint environment */ -static void assure_different_constraints(ir_node *irn, constraint_env_t *env) { - const arch_register_req_t *req; - const arch_env_t *arch_env = be_get_birg_arch_env(env->birg); - - req = arch_get_register_req(arch_env, irn, -1); +static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) { + const arch_register_req_t *req = arch_get_register_req(irn, -1); - if (arch_register_req_is(req, should_be_different)) { + if (arch_register_req_is(req, must_be_different)) { const unsigned other = req->other_different; int i; + if (arch_register_req_is(req, should_be_same)) { + const unsigned same = req->other_same; + + if (is_po2(other) && is_po2(same)) { + int idx_other = ntz(other); + int idx_same = ntz(same); + + /* + * We can safely ignore a should_be_same x must_be_different y + * IFF both inputs are equal! + */ + if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) { + return; + } + } + } for (i = 0; 1U << i <= other; ++i) { if (other & (1U << i)) { - ir_node *different_from = get_irn_n(belower_skip_proj(irn), i); + ir_node *different_from = get_irn_n(skipped_irn, i); gen_assure_different_pattern(irn, different_from, env); } } } } - - /** * Calls the functions to assure register constraints. * - * @param irn The node to be checked for lowering + * @param block The block to be checked * @param walk_env The walker environment */ -static void assure_constraints_walker(ir_node *irn, void *walk_env) { - if (is_Block(irn)) - return; +static void assure_constraints_walker(ir_node *block, void *walk_env) { + ir_node *irn; + + sched_foreach_reverse(block, irn) { + ir_mode *mode = get_irn_mode(irn); - if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn))) - assure_different_constraints(irn, walk_env); + if (mode == mode_T) { + const ir_edge_t *edge; - return; + foreach_out_edge(irn, edge) { + ir_node *proj = get_edge_src_irn(edge); + + mode = get_irn_mode(proj); + if (mode_is_datab(mode)) + assure_different_constraints(proj, irn, walk_env); + } + } else if (mode_is_datab(mode)) { + assure_different_constraints(irn, irn, walk_env); + } + } } /** @@ -668,10 +646,12 @@ static void assure_constraints_walker(ir_node *irn, void *walk_env) { static void melt_copykeeps(constraint_env_t *cenv) { be_irg_t *birg = cenv->birg; ir_graph *irg = be_get_birg_irg(birg); - op_copy_assoc_t *entry; + ir_nodemap_iterator_t map_iter; + ir_nodemap_entry_t map_entry; /* for all */ - foreach_pset(cenv->op_set, entry) { + foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) { + op_copy_assoc_t *entry = map_entry.data; int idx, num_ck; ir_node *cp; struct obstack obst; @@ -711,7 +691,7 @@ static void melt_copykeeps(constraint_env_t *cenv) { ref_mode_T = skip_Proj(get_irn_n(ref, 1)); obstack_grow(&obst, &ref, sizeof(ref)); - DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref)); + DBG((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref)); /* check for copykeeps pointing to the same mode_T node as the reference copykeep */ for (j = 0; j < num_ck; ++j) { @@ -720,7 +700,7 @@ static void melt_copykeeps(constraint_env_t *cenv) { if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) { obstack_grow(&obst, &cur_ck, sizeof(cur_ck)); ir_nodeset_remove(&entry->copies, cur_ck); - DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck)); + DBG((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck)); ck_arr[j] = NULL; ++n_melt; sched_remove(cur_ck); @@ -730,7 +710,7 @@ static void melt_copykeeps(constraint_env_t *cenv) { /* check, if we found some candidates for melting */ if (n_melt == 1) { - DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n")); + DBG((dbg_constr, LEVEL_1, "\tno candidate found\n")); continue; } @@ -746,7 +726,7 @@ static void melt_copykeeps(constraint_env_t *cenv) { /* now, we can kill the melted keep, except the */ /* ref one, we still need some information */ if (melt_arr[j] != ref) - be_kill_node(melt_arr[j]); + kill_node(melt_arr[j]); } #ifdef KEEP_ALIVE_COPYKEEP_HACK @@ -770,10 +750,10 @@ static void melt_copykeeps(constraint_env_t *cenv) { } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt)); sched_add_before(sched_pt, new_ck); - DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt)); + DBG((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt)); /* finally: kill the reference copykeep */ - be_kill_node(ref); + kill_node(ref); } } @@ -787,21 +767,19 @@ static void melt_copykeeps(constraint_env_t *cenv) { * @param birg The birg structure containing the irg */ void assure_constraints(be_irg_t *birg) { - ir_graph *irg = be_get_birg_irg(birg); - const arch_env_t *arch_env = be_get_birg_arch_env(birg); - constraint_env_t cenv; - op_copy_assoc_t *entry; - ir_node **nodes; - FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr"); - - be_assure_dom_front(birg); - - DEBUG_ONLY(cenv.dbg = mod;) - cenv.birg = birg; - cenv.op_set = new_pset(cmp_op_copy_assoc, 16); + ir_graph *irg = be_get_birg_irg(birg); + constraint_env_t cenv; + ir_node **nodes; + ir_nodemap_iterator_t map_iter; + ir_nodemap_entry_t map_entry; + + FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr"); + + cenv.birg = birg; + ir_nodemap_init(&cenv.op_set); obstack_init(&cenv.obst); - irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv); + irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv); /* melt copykeeps, pointing to projs of */ /* the same mode_T node and keeping the */ @@ -809,7 +787,8 @@ void assure_constraints(be_irg_t *birg) { melt_copykeeps(&cenv); /* for all */ - foreach_pset(cenv.op_set, entry) { + foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) { + op_copy_assoc_t *entry = map_entry.data; int n; ir_node *cp; ir_nodeset_iterator_t iter; @@ -819,22 +798,22 @@ void assure_constraints(be_irg_t *birg) { nodes = alloca(n * sizeof(nodes[0])); /* put the node in an array */ - DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op)); + DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node)); /* collect all copies */ n = 0; foreach_ir_nodeset(&entry->copies, cp, iter) { nodes[n++] = cp; - DB((mod, LEVEL_1, ", %+F ", cp)); + DB((dbg_constr, LEVEL_1, ", %+F ", cp)); } - DB((mod, LEVEL_1, "\n")); + DB((dbg_constr, LEVEL_1, "\n")); /* introduce the copies for the operand and it's copies */ be_ssa_construction_init(&senv, birg); - be_ssa_construction_add_copy(&senv, entry->op); + be_ssa_construction_add_copy(&senv, map_entry.node); be_ssa_construction_add_copies(&senv, nodes, n); - be_ssa_construction_fix_users(&senv, entry->op); + be_ssa_construction_fix_users(&senv, map_entry.node); be_ssa_construction_destroy(&senv); /* Could be that not all CopyKeeps are really needed, */ @@ -844,20 +823,20 @@ void assure_constraints(be_irg_t *birg) { ir_node *keep; int n = get_irn_arity(cp); - keep = be_new_Keep(arch_get_irn_reg_class(arch_env, cp, -1), - irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]); + keep = be_new_Keep(arch_get_irn_reg_class(cp, -1), + irg, get_nodes_block(cp), n, get_irn_in(cp) + 1); sched_add_before(cp, keep); /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */ sched_remove(cp); - be_kill_node(cp); + kill_node(cp); } } ir_nodeset_destroy(&entry->copies); } - del_pset(cenv.op_set); + ir_nodemap_destroy(&cenv.op_set); obstack_free(&cenv.obst, NULL); be_liveness_invalidate(be_get_birg_liveness(birg)); } @@ -869,16 +848,15 @@ void assure_constraints(be_irg_t *birg) { * @note This routine needs interference. * @note Probably, we can implement it a little more efficient. * Especially searching the frontier lazily might be better. - * @param perm The perm. - * @param data The walker data (lower_env_t). + * + * @param perm The perm + * @param env The lowerer environment + * * @return 1, if there is something left to perm over. * 0, if removed the complete perm. */ -static int push_through_perm(ir_node *perm, void *data) +static int push_through_perm(ir_node *perm, lower_env_t *env) { - lower_env_t *env = data; - const arch_env_t *aenv = env->arch_env; - ir_graph *irg = get_irn_irg(perm); ir_node *bl = get_nodes_block(perm); ir_node *node; @@ -888,99 +866,100 @@ static int push_through_perm(ir_node *perm, void *data) bitset_t *moved = bitset_alloca(arity); int n_moved; int new_size; - ir_node *frontier = sched_first(bl); - FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove"); + ir_node *frontier = bl; int i, n; const ir_edge_t *edge; - ir_node *last_proj = NULL, *irn; + ir_node *one_proj = NULL, *irn; const arch_register_class_t *cls = NULL; - DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg)); + DBG((dbg_permmove, LEVEL_1, "perm move %+F irg %+F\n", perm, irg)); - /* get some proj and find out the register class of the proj. */ - foreach_out_edge (perm, edge) { - last_proj = get_edge_src_irn(edge); - cls = arch_get_irn_reg_class(aenv, last_proj, -1); - assert(is_Proj(last_proj)); - break; - } + /* get some Proj and find out the register class of that Proj. */ + edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL); + one_proj = get_edge_src_irn(edge); + assert(is_Proj(one_proj)); + cls = arch_get_irn_reg_class(one_proj, -1); - /* find the point in the schedule after which the + /* Find the point in the schedule after which the * potentially movable nodes must be defined. - * A perm will only be pushed up to first instruction - * which lets an operand of itself die. */ - - sched_foreach_reverse_from (sched_prev(perm), irn) { - for(i = get_irn_arity(irn) - 1; i >= 0; --i) { + * A Perm will only be pushed up to first instruction + * which lets an operand of itself die. + * If we would allow to move the Perm above this instruction, + * the former dead operand would be live now at the point of + * the Perm, increasing the register pressure by one. + */ + sched_foreach_reverse_from(sched_prev(perm), irn) { + for (i = get_irn_arity(irn) - 1; i >= 0; --i) { ir_node *op = get_irn_n(irn, i); - if(arch_irn_consider_in_reg_alloc(aenv, cls, op) - && !values_interfere(env->birg, op, last_proj)) { - frontier = sched_next(irn); + if (arch_irn_consider_in_reg_alloc(cls, op) && + !values_interfere(env->birg, op, one_proj)) { + frontier = irn; goto found_front; } } } found_front: - DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier)); + DBG((dbg_permmove, LEVEL_2, "\tfrontier: %+F\n", frontier)); node = sched_prev(perm); n_moved = 0; - while(!sched_is_begin(node)) { + while (!sched_is_begin(node)) { const arch_register_req_t *req; int input = -1; ir_node *proj; + /* search if node is a INPUT of Perm */ foreach_out_edge(perm, edge) { ir_node *out = get_edge_src_irn(edge); int pn = get_Proj_proj(out); ir_node *in = get_irn_n(perm, pn); - if(node == in) { + if (node == in) { proj = out; input = pn; break; } } /* it wasn't an input to the perm, we can't do anything more */ - if(input < 0) + if (input < 0) break; - if(!sched_comes_after(frontier, node)) + if (!sched_comes_after(frontier, node)) break; - if(arch_irn_is(aenv, node, modify_flags)) + if (arch_irn_is(node, modify_flags)) break; - if(is_Proj(node)) { - req = arch_get_register_req(aenv, get_Proj_pred(node), + if (is_Proj(node)) { + req = arch_get_register_req(get_Proj_pred(node), -1 - get_Proj_proj(node)); } else { - req = arch_get_register_req(aenv, node, -1); + req = arch_get_register_req(node, -1); } - if(req->type != arch_register_req_type_normal) + if (req->type != arch_register_req_type_normal) break; - for(i = get_irn_arity(node) - 1; i >= 0; --i) { + for (i = get_irn_arity(node) - 1; i >= 0; --i) { ir_node *opop = get_irn_n(node, i); - if (arch_irn_consider_in_reg_alloc(aenv, cls, opop)) { + if (arch_irn_consider_in_reg_alloc(cls, opop)) { break; } } - if(i >= 0) + if (i >= 0) break; - DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj)); + DBG((dbg_permmove, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj)); /* move the movable node in front of the Perm */ sched_remove(node); sched_add_after(perm, node); /* give it the proj's register */ - arch_set_irn_register(aenv, node, arch_get_irn_register(aenv, proj)); + arch_set_irn_register(node, arch_get_irn_register(proj)); /* reroute all users of the proj to the moved node. */ edges_reroute(proj, node, irg); /* and kill it */ set_Proj_pred(proj, new_Bad()); - be_kill_node(proj); + kill_node(proj); bitset_set(moved, input); n_moved++; @@ -1001,8 +980,8 @@ found_front: proj_map = alloca(arity * sizeof(proj_map[0])); memset(proj_map, -1, sizeof(proj_map[0])); n = 0; - for(i = 0; i < arity; ++i) { - if(bitset_is_set(moved, i)) + for (i = 0; i < arity; ++i) { + if (bitset_is_set(moved, i)) continue; map[n] = i; proj_map[i] = n; @@ -1027,19 +1006,16 @@ found_front: * @param irn The node to be checked for lowering * @param walk_env The walker environment */ -static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) { +static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) +{ int perm_stayed; - if (is_Block(irn) || is_Proj(irn)) - return; if (!be_is_Perm(irn)) return; perm_stayed = push_through_perm(irn, walk_env); - if (!perm_stayed) - return; - - lower_perm_node(irn, walk_env); + if (perm_stayed) + lower_perm_node(irn, walk_env); } /** @@ -1051,15 +1027,17 @@ static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) { */ void lower_nodes_after_ra(be_irg_t *birg, int do_copy) { lower_env_t env; - ir_graph *irg = be_get_birg_irg(birg); + ir_graph *irg; + + FIRM_DBG_REGISTER(dbg, "firm.be.lower"); + FIRM_DBG_REGISTER(dbg_permmove, "firm.be.lower.permmove"); - env.birg = birg; - env.arch_env = be_get_birg_arch_env(birg); - env.do_copy = do_copy; - FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower"); + env.birg = birg; + env.do_copy = do_copy; /* we will need interference */ be_liveness_assure_chk(be_get_birg_liveness(birg)); - irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env); + irg = be_get_birg_irg(birg); + irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env); }