X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbelower.c;h=45ea686bc1df9e65207d2d6a2b0fbd3be0199c11;hb=17e47394fa72025d14172a2acef2d258a067aa42;hp=9299ab2a9ad905ffc958f4e1c86c6afc80f77711;hpb=99ddeda82f88b645f244b84a8034165b967a2204;p=libfirm diff --git a/ir/be/belower.c b/ir/be/belower.c index 9299ab2a9..45ea686bc 100644 --- a/ir/be/belower.c +++ b/ir/be/belower.c @@ -14,17 +14,24 @@ #include #include "ircons.h" +#include "debug.h" #include "bearch.h" #include "belower.h" #include "benode_t.h" #include "bechordal_t.h" -#include "besched.h" +#include "besched_t.h" #include "irgmod.h" #include "iredges_t.h" #include "irgwalk.h" +#ifdef _WIN32 +#include +#else +#include +#endif + #undef is_Perm #define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm) @@ -162,22 +169,36 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int 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) + 1; + int n_pairs_done = get_n_checked_pairs(pairs, n); int idx; + perm_type_t cycle_tp = PERM_CYCLE; + + /* 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); + + if (cur_pair_idx < 0) { + cycle_tp = PERM_CHAIN; + break; + } + 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 = calloc(n - n_pairs_done, sizeof(cycle->elems[0])); + cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0])); cycle->n_elems = 2; /* initial number of elements is 2 */ cycle->elems[0] = pairs[start].in_reg; cycle->elems[1] = pairs[start].out_reg; - cycle->type = PERM_CHAIN; /* default is CHAIN, only changed when we found a cycle */ - - /* mark the first pair as checked */ - pairs[start].checked = 1; + cycle->type = cycle_tp; + cur_idx = pairs[start].out_reg->index; idx = 2; /* check for cycle or end of a chain */ - while (cur_idx != head && n_pairs_done < n) { + while (cur_idx != head) { /* goto next register in cycle or chain */ cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0); @@ -190,16 +211,24 @@ 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++; - - n_pairs_done++; } else { /* we are there where we started -> CYCLE */ cycle->type = PERM_CYCLE; } + } - /* mark the pair as checked */ - pairs[cur_pair_idx].checked = 1; + /* 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); + + 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); + + if (cur_pair_idx >= 0) + pairs[cur_pair_idx].checked = 1; } return cycle; @@ -223,12 +252,12 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { reg_pair_t *pairs; const ir_edge_t *edge; perm_cycle_t *cycle; - int n, i, pn, do_copy; + int n, i, pn, do_copy, j; ir_node *sched_point, *block, *in[2]; ir_node *arg1, *arg2, *res1, *res2; ir_node *cpyxchg = NULL; - arch_env = env->chord_env->main_env->arch_env; + arch_env = env->chord_env->birg->main_env->arch_env; do_copy = env->do_copy; mod = env->dbg_module; block = get_nodes_block(irn); @@ -247,7 +276,7 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { 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 = calloc(n, sizeof(pairs[0])); + pairs = alloca(n * sizeof(pairs[0])); /* build the list of register pairs (in, out) */ i = 0; @@ -301,9 +330,15 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { /* go to the first not-checked pair */ while (pairs[i].checked) i++; - cycle = calloc(1, sizeof(*cycle)); + cycle = xcalloc(1, sizeof(*cycle)); cycle = 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((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 */ @@ -335,6 +370,33 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { 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", @@ -342,8 +404,21 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in); + if (i > 0) { + /* cycle is not done yet */ + int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0); + + /* create intermediate proj */ + res2 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0); + + /* set as in for next Perm */ + pairs[pidx].in_node = res2; + } + else { + sched_remove(res2); + } + sched_remove(res1); - sched_remove(res2); set_Proj_pred(res2, cpyxchg); set_Proj_proj(res2, 0); @@ -372,11 +447,13 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { /* 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; DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point)); } -// free(cycle->elems); + free((void *) cycle->elems); free(cycle); } @@ -386,87 +463,6 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { -/** - * Adds Projs to keep nodes for each register class, which eats the - * caller saved registers. - * Note: The caller has to make sure, that call is a Call - * - * @param call The Call node - * @param walk_env The walker environment - */ -static void lower_call_node(ir_node *call, void *walk_env) { - lower_env_t *env = walk_env; - const arch_env_t *arch_env = env->chord_env->main_env->arch_env; - firm_dbg_module_t *mod = env->dbg_module; - const arch_register_class_t *reg_class; - int i, j, set_size = 0, pn, keep_arity; - arch_isa_t *isa = arch_env_get_isa(arch_env); - const ir_node *proj_T = NULL; - ir_node **in_keep, *block = get_nodes_block(call); - bitset_t *proj_set; - const ir_edge_t *edge; - const arch_register_t *reg; - - /* Prepare the bitset where we store the projnums which are already in use*/ - for (i = 0; i < arch_isa_get_n_reg_class(isa); i++) { - reg_class = arch_isa_get_reg_class(isa, i); - set_size += arch_register_class_n_regs(reg_class); - } - - in_keep = malloc(set_size * sizeof(ir_node *)); - - proj_set = bitset_malloc(set_size); - bitset_clear_all(proj_set); - - /* check if there is a ProjT node and which arguments are used */ - foreach_out_edge(call, edge) { - if (get_irn_mode(get_edge_src_irn(edge)) == mode_T) - proj_T = get_edge_src_irn(edge); - } - - /* set all used arguments */ - if (proj_T) { - foreach_out_edge(proj_T, edge) { - bitset_set(proj_set, get_Proj_proj(get_edge_src_irn(edge))); - } - } - else { - proj_T = new_r_Proj(current_ir_graph, block, call, mode_T, pn_Call_T_result); - } - - /* Create for each caller save register a proj (keep node arguement) */ - /* if this proj is not already present */ - for (i = 0; i < arch_isa_get_n_reg_class(isa); i++) { - - /* reset the keep input, as we need one keep for each register class */ - memset(in_keep, 0, set_size * sizeof(ir_node *)); - keep_arity = 0; - reg_class = arch_isa_get_reg_class(isa, i); - - for (j = 0; j < arch_register_class_n_regs(reg_class); j++) { - reg = arch_register_for_index(reg_class, j); - - /* only check caller save registers */ - if (arch_register_type_is(reg, caller_saved)) { - pn = isa->impl->get_projnum_for_register(isa, reg); - if (!bitset_is_set(proj_set, pn)) { - in_keep[keep_arity++] = new_r_Proj(current_ir_graph, block, (ir_node *)proj_T, mode_Is, pn); - } - } - } - - /* ok, we found some caller save register which are not in use but must be saved */ - if (keep_arity) { - be_new_Keep(reg_class, current_ir_graph, block, keep_arity, in_keep); - } - } - - bitset_free(proj_set); - return; -} - - - /** * Calls the backend code generator functions to lower Spill and * Reload nodes into Store and Load. The backend is fully responsible @@ -479,8 +475,8 @@ static void lower_call_node(ir_node *call, void *walk_env) { */ static void lower_spill_reload(ir_node *irn, void *walk_env) { lower_env_t *env = walk_env; - arch_code_generator_t *cg = env->chord_env->main_env->cg; - const arch_env_t *aenv = env->chord_env->main_env->arch_env; + arch_code_generator_t *cg = env->chord_env->birg->cg; + const arch_env_t *aenv = env->chord_env->birg->main_env->arch_env; ir_node *res = NULL; ir_node *sched_point; @@ -516,41 +512,39 @@ static void lower_spill_reload(ir_node *irn, void *walk_env) { * @param irn The node to be checked for lowering * @param walk_env The walker environment */ -static void lower_nodes_walker(ir_node *irn, void *walk_env) { +static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) { lower_env_t *env = walk_env; - const arch_env_t *arch_env = env->chord_env->main_env->arch_env; + const arch_env_t *arch_env = env->chord_env->birg->main_env->arch_env; - if (!is_Block(irn)) { - if (!is_Proj(irn)) { - if (is_Perm(arch_env, irn)) { - lower_perm_node(irn, walk_env); - } - else if (is_Call(arch_env, irn)) { - lower_call_node(irn, walk_env); - } - else if (be_is_Spill(irn) || be_is_Reload(irn)) { - lower_spill_reload(irn, walk_env); - } + if (!is_Block(irn) && !is_Proj(irn)) { + if (is_Perm(arch_env, irn)) { + lower_perm_node(irn, walk_env); + } + else if (be_is_Spill(irn) || be_is_Reload(irn)) { + lower_spill_reload(irn, walk_env); } } return; } + + /** - * Walks over all blocks in an irg and performs some lowering. + * Walks over all blocks in an irg and performs lowering need to be + * done after register allocation (e.g. perm and spill/reload lowering). * * @param chord_env The chordal environment containing the irg * @param do_copy 1 == resolve cycles with a free reg if available */ -void lower_nodes(be_chordal_env_t *chord_env, int do_copy) { +void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy) { lower_env_t env; env.chord_env = chord_env; env.do_copy = do_copy; - env.dbg_module = firm_dbg_register("ir.be.lower"); + env.dbg_module = firm_dbg_register("firm.be.lower"); - irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_walker, &env); + irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env); } #undef is_Perm