X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbelower.c;h=c08331ba40368f0a030424d1e0c5fff14dd392fc;hb=d35fdb8eda3010ae7e487d9f191e415224db39cb;hp=ee09f6b537abe1e808cdb87631bd8eedbea66805;hpb=dbc419164cf5e14cdf5757f1e105d02693bacb9f;p=libfirm diff --git a/ir/be/belower.c b/ir/be/belower.c index ee09f6b53..c08331ba4 100644 --- a/ir/be/belower.c +++ b/ir/be/belower.c @@ -30,7 +30,6 @@ #include "ircons.h" #include "debug.h" -#include "irhooks.h" #include "xmalloc.h" #include "irnodeset.h" #include "irnodemap.h" @@ -49,6 +48,10 @@ #undef KEEP_ALIVE_COPYKEEP_HACK +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_nodeset_t copies; /**< all non-spillable copies of this irn */ @@ -60,14 +63,12 @@ typedef struct { be_irg_t *birg; ir_nodemap_t op_set; struct obstack obst; - DEBUG_ONLY(firm_dbg_module_t *dbg;) } constraint_env_t; /** Lowering walker environment. */ typedef struct _lower_env_t { be_irg_t *birg; unsigned do_copy : 1; - DEBUG_ONLY(firm_dbg_module_t *dbg_module;) } lower_env_t; /** Holds a Perm register pair. */ @@ -95,126 +96,144 @@ typedef struct _perm_cycle_t { perm_type_t type; /**< type (CHAIN or CYCLE) */ } perm_cycle_t; -/** 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; +/** returns the number register pairs marked as checked. */ +static int get_n_unchecked_pairs(reg_pair_t const *const pairs, int const n) +{ + int n_unchecked = 0; + int i; - if (pair_a->in_reg->index > pair_b->in_reg->index) - return 1; - else - return -1; + for (i = 0; i < n; i++) { + if (!pairs[i].checked) + n_unchecked++; + } + + return n_unchecked; } -/** 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; +/** + * 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 + * @return The corresponding node or NULL if not found + */ +static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg) { + int i; for (i = 0; i < n; i++) { - if (pairs[i].checked) - n_checked++; + /* in register matches */ + if (pairs[i].in_reg->index == reg->index) + return pairs[i].in_node; } - return n_checked; + return NULL; } /** - * Gets the node corresponding to a register from an array of register pairs. + * 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 - * @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_out_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; - } - } - else { - for (i = 0; i < n; i++) { - /* in register matches */ - if (pairs[i].in_reg->index == reg->index) - return pairs[i].in_node; - } + 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_todo = get_n_unchecked_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 = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2); + cycle->elems = XMALLOCNZ(const arch_register_t*, n_pairs_todo * 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; @@ -225,7 +244,7 @@ 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; @@ -236,8 +255,7 @@ 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; } @@ -245,18 +263,16 @@ 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; } /** @@ -267,227 +283,198 @@ 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; - 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; - ir_node *sched_point, *block, *in[2]; - ir_node *arg1, *arg2, *res1, *res2; - ir_node *cpyxchg = NULL; - DEBUG_ONLY(firm_dbg_module_t *mod;) - - 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(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(pairs[i].in_node); - pairs[i].out_reg = arch_get_irn_register(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]; - /* 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()); + if (pair->in_reg->index != pair->out_reg->index) + continue; - pairs[i].checked = 1; - } + 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)); + + /* 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); - /* check for cycles and chains */ - while (get_n_checked_pairs(pairs, n) < n) { - perm_cycle_t *cycle; - - i = n_ops = 0; + while (get_n_unchecked_pairs(pairs, n) > 0) { + perm_cycle_t cycle; + int j; /* go to the first not-checked pair */ - while (pairs[i].checked) i++; - cycle = XMALLOCZ(perm_cycle_t); - 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(res2, cycle->elems[i + 1]); - arch_set_irn_register(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(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; + } } } - free((void *) cycle->elems); - free(cycle); + free((void*)cycle.elems); } /* remove the perm from schedule */ - if (! keep_perm) { + if (!keep_perm) { sched_remove(irn); kill_node(irn); } @@ -499,82 +486,72 @@ static int has_irn_users(const ir_node *irn) { return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0; } -/** - * Skip all Proj nodes. - */ -static INLINE ir_node *belower_skip_proj(ir_node *irn) { - while(is_Proj(irn)) - irn = get_Proj_pred(irn); - return irn; -} - static ir_node *find_copy(ir_node *irn, ir_node *op) { - ir_node *block = get_nodes_block(irn); 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)) - { + 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); - ir_nodemap_t *op_set = &env->op_set; - ir_node *block = get_nodes_block(irn); - const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1); - ir_node *in[2], *keep, *cpy; + 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; - DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;) if (arch_irn_is(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)); + 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 immediately before */ - cpy = find_copy(belower_skip_proj(irn), other_different); + 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)); - } - else { - DBG((mod, LEVEL_1, "using already existing %+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)); } - 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 (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 { + } 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 map */ @@ -591,18 +568,20 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, 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); - } } /** * 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) { - ir_node *skipped_irn = belower_skip_proj(irn); - const arch_register_req_t *req = arch_get_register_req(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, must_be_different)) { const unsigned other = req->other_different; @@ -636,15 +615,29 @@ static void assure_different_constraints(ir_node *irn, constraint_env_t *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 (mode == mode_T) { + const ir_edge_t *edge; - if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn))) - assure_different_constraints(irn, walk_env); + 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); + } + } } /** @@ -699,7 +692,7 @@ 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) { @@ -708,7 +701,7 @@ 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); @@ -718,7 +711,7 @@ 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; } @@ -758,7 +751,7 @@ 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 */ kill_node(ref); @@ -780,14 +773,14 @@ void assure_constraints(be_irg_t *birg) { ir_node **nodes; ir_nodemap_iterator_t map_iter; ir_nodemap_entry_t map_entry; - FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr"); - DEBUG_ONLY(cenv.dbg = mod;) - cenv.birg = birg; + 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 */ @@ -806,16 +799,16 @@ 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 ", map_entry.node)); + 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); @@ -856,15 +849,15 @@ 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; - ir_graph *irg = get_irn_irg(perm); ir_node *bl = get_nodes_block(perm); ir_node *node; @@ -875,14 +868,13 @@ static int push_through_perm(ir_node *perm, void *data) int n_moved; int new_size; ir_node *frontier = bl; - FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove"); int i, n; const ir_edge_t *edge; 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 that Proj. */ edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL); @@ -898,7 +890,7 @@ static int push_through_perm(ir_node *perm, void *data) * 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) { + 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(cls, op) && @@ -910,11 +902,11 @@ static int push_through_perm(ir_node *perm, void *data) } 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; @@ -931,30 +923,30 @@ found_front: } } /* 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(node, modify_flags)) break; - if(is_Proj(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(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(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); @@ -989,8 +981,8 @@ 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; @@ -1015,19 +1007,16 @@ 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); } /** @@ -1039,14 +1028,17 @@ 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.do_copy = do_copy; - FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower"); /* 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); }