X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fcfopt.c;h=e0cd1ac0d2af073bf36fef445f6955b921ea71b0;hb=c0bef4f5aded93790e8d181ec3fcdcd6429a3cd3;hp=96015fc36ecd261740cfd856fa65fee8f6510cb1;hpb=1a29d29759ede67f77585faf1d27ef27647804ed;p=libfirm diff --git a/ir/opt/cfopt.c b/ir/opt/cfopt.c index 96015fc36..e0cd1ac0d 100644 --- a/ir/opt/cfopt.c +++ b/ir/opt/cfopt.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. * @@ -27,6 +27,8 @@ # include "config.h" #endif +#include "iroptimize.h" + #include #include "plist.h" @@ -51,7 +53,6 @@ #include "irflag_t.h" #include "firmstat.h" -#include "cfopt.h" #include "iropt_dbg.h" /*------------------------------------------------------------------*/ @@ -61,10 +62,16 @@ /* is empty if it contains only a Jmp node. */ /* Blocks can only be removed if they are not needed for the */ /* semantics of Phi nodes. */ +/* Further, we NEVER remove labeled blocks (even if we could move */ +/* the label. */ /*------------------------------------------------------------------*/ +#define set_Block_removable(block) set_Block_mark(block, 1) +#define set_Block_non_removable(block) set_Block_mark(block, 0) +#define is_Block_removable(block) (get_Block_mark(block) != 0) + /** - * Block walker, replacing binary Conds that jumps twice into the same block + * Replace binary Conds that jumps twice into the same block * by a simple Jmp. * E.g. * @verbatim @@ -80,12 +87,10 @@ * Note that the simple case that Block has only these two * predecessors are already handled in equivalent_node_Block(). */ -static void remove_senseless_conds(ir_node *bl, void *env) { +static int remove_senseless_conds(ir_node *bl) { int i, j; int n = get_Block_n_cfgpreds(bl); - int *changed = env; - - assert(is_Block(bl)); + int changed = 0; for (i = 0; i < n; ++i) { ir_node *pred_i = get_Block_cfgpred(bl, i); @@ -104,18 +109,20 @@ static void remove_senseless_conds(ir_node *bl, void *env) { set_irn_n(bl, j, new_Bad()); DBG_OPT_IFSIM2(cond_i, jmp); - *changed = 1; + changed = 1; break; } } } } + return changed; } /** An environment for merge_blocks and collect nodes. */ typedef struct _merge_env { - int changed; - plist_t *list; + int changed; /**< Set if the graph was changed. */ + int phis_moved; /**< Set if Phi nodes were moved. */ + plist_t *list; /**< Helper list for all found Switch Conds. */ } merge_env; /** @@ -183,63 +190,77 @@ static void merge_blocks(ir_node *node, void *ctx) { } } - /** * Block walker removing control flow from dead block by * inspecting dominance info. * Do not replace blocks by Bad. This optimization shall * ensure, that all Bad control flow predecessors are * removed, and no new other Bads are introduced. + * Further removed useless Conds and clear the mark of all blocks. * * Must be run in the post walker. */ -static void remove_dead_block_cf(ir_node *block, void *env) { +static void remove_unreachable_blocks_and_conds(ir_node *block, void *env) { int i; int *changed = env; - /* check block predecessors and turn control flow into bad */ + /* Check block predecessors and turn control flow into bad. + Beware of Tuple, kill them. */ for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) { - ir_node *pred_X = get_Block_cfgpred(block, i); + ir_node *pred_X = get_Block_cfgpred(block, i); + ir_node *skipped = skip_Tuple(pred_X); - if (! is_Bad(pred_X)) { - ir_node *pred_bl = get_nodes_block(skip_Proj(pred_X)); + if (! is_Bad(skipped)) { + ir_node *pred_bl = get_nodes_block(skip_Proj(skipped)); if (is_Block_dead(pred_bl) || (get_Block_dom_depth(pred_bl) < 0)) { set_Block_dead(pred_bl); exchange(pred_X, new_Bad()); *changed = 1; + } else if (skipped != pred_X) { + set_Block_cfgpred(block, i, skipped); + *changed = 1; } } } + + *changed |= remove_senseless_conds(block); + + /* clear the block mark of all blocks */ + set_Block_removable(block); } /** * Collects all Phi nodes in link list of Block. - * Marks all blocks "block_visited" if they contain a node other + * Marks all blocks "non_removable" if they contain a node other * than Jmp (and Proj). * Links all Proj nodes to their predecessors. * Collects all switch-Conds in a list. */ static void collect_nodes(ir_node *n, void *ctx) { - ir_op *op = get_irn_op(n); + ir_opcode code = get_irn_opcode(n); merge_env *env = ctx; - if (op != op_Block) { - ir_node *b = get_nodes_block(n); + if (code == iro_Block) { + /* mark the block as non-removable if it is labeled */ + if (has_Block_label(n)) + set_Block_non_removable(n); + } else { + ir_node *b = get_nodes_block(n); - if (op == op_Phi) { + if (code == iro_Phi && get_irn_arity(n) > 0) { /* Collect Phi nodes to compact ins along with block's ins. */ set_irn_link(n, get_irn_link(b)); set_irn_link(b, n); - } else if (op != op_Jmp && !is_Bad(b)) { /* Check for non empty block. */ - mark_Block_block_visited(b); + } else if (code != iro_Jmp && !is_Bad(b)) { /* Check for non-empty block. */ + set_Block_non_removable(b); - if (op == op_Proj) { /* link Proj nodes */ + if (code == iro_Proj) { /* link Proj nodes */ ir_node *pred = get_Proj_pred(n); set_irn_link(n, get_irn_link(pred)); set_irn_link(pred, n); - } else if (op == op_Cond) { + } else if (code == iro_Cond) { ir_node *sel = get_Cond_selector(n); if (mode_is_int(get_irn_mode(sel))) { /* found a switch-Cond, collect */ @@ -252,11 +273,12 @@ static void collect_nodes(ir_node *n, void *ctx) { /** Returns true if pred is predecessor of block. */ static int is_pred_of(ir_node *pred, ir_node *b) { - int i, n; + int i; - for (i = 0, n = get_Block_n_cfgpreds(b); i < n; ++i) { + for (i = get_Block_n_cfgpreds(b) - 1; i >= 0; --i) { ir_node *b_pred = get_Block_cfgpred_block(b, i); - if (b_pred == pred) return 1; + if (b_pred == pred) + return 1; } return 0; } @@ -296,17 +318,15 @@ static int test_whether_dispensable(ir_node *b, int pos) { if (is_Block_dead(pred)) return 0; - if (get_Block_block_visited(pred) + 1 - < get_irg_block_visited(current_ir_graph)) { - + if (is_Block_removable(pred)) { if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) { /* Mark block so that is will not be removed: optimization is turned off. */ - set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1); + set_Block_non_removable(pred); return 1; } /* Seems to be empty. At least we detected this in collect_nodes. */ - if (!get_irn_link(b)) { + if (get_irn_link(b) == NULL) { /* There are no Phi nodes ==> all predecessors are dispensable. */ n_preds = get_Block_n_cfgpreds(pred); } else { @@ -314,10 +334,8 @@ static int test_whether_dispensable(ir_node *b, int pos) { Handle all pred blocks with preds < pos as if they were already removed. */ for (i = 0; i < pos; i++) { ir_node *b_pred = get_Block_cfgpred_block(b, i); - if (! is_Block_dead(b_pred) && - get_Block_block_visited(b_pred) + 1 - < get_irg_block_visited(current_ir_graph)) { - for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) { + if (! is_Block_dead(b_pred) && is_Block_removable(b_pred)) { + for (j = get_Block_n_cfgpreds(b_pred) - 1; j >= 0; --j) { ir_node *b_pred_pred = get_Block_cfgpred_block(b_pred, j); if (is_pred_of(b_pred_pred, pred)) goto non_dispensable; @@ -340,7 +358,7 @@ static int test_whether_dispensable(ir_node *b, int pos) { return n_preds; non_dispensable: - set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1); + set_Block_non_removable(pred); return 1; } @@ -391,11 +409,11 @@ non_dispensable: * @@@ It is negotiable whether we should do this ... there might end up a copy * from the Phi in the loop when removing the Phis. */ -static void optimize_blocks(ir_node *b, void *env) { +static void optimize_blocks(ir_node *b, void *ctx) { int i, j, k, n, max_preds, n_preds, p_preds = -1; ir_node *pred, *phi; ir_node **in; - int *changed = env; + merge_env *env = ctx; /* Count the number of predecessor if this block is merged with pred blocks that are empty. */ @@ -414,11 +432,9 @@ static void optimize_blocks(ir_node *b, void *env) { for (i = 0, n = get_Block_n_cfgpreds(b); i < n; ++i) { pred = get_Block_cfgpred_block(b, i); - if (is_Bad(get_Block_cfgpred(b, i))) { + if (is_Block_dead(pred)) { /* case Phi 1: Do nothing */ - } - else if (get_Block_block_visited(pred) + 1 - < get_irg_block_visited(current_ir_graph)) { + } else if (is_Block_removable(pred) && !Block_block_visited(pred)) { /* case Phi 2: It's an empty block and not yet visited. */ ir_node *phi_pred = get_Phi_pred(phi, i); @@ -450,7 +466,7 @@ static void optimize_blocks(ir_node *b, void *env) { exchange(phi, in[0]); else set_irn_in(phi, p_preds, in); - *changed = 1; + env->changed = 1; phi = get_irn_link(phi); } @@ -461,7 +477,7 @@ static void optimize_blocks(ir_node *b, void *env) { for (k = 0, n = get_Block_n_cfgpreds(b); k < n; ++k) { ir_node *predb = get_nodes_block(get_Block_cfgpred(b, k)); - if (get_Block_block_visited(predb) + 1 < get_irg_block_visited(current_ir_graph)) { + if (is_Block_removable(predb) && !Block_block_visited(predb)) { ir_node *next_phi; /* we found a predecessor block at position k that will be removed */ @@ -480,15 +496,15 @@ static void optimize_blocks(ir_node *b, void *env) { set_nodes_block(phi, b); set_irn_link(phi, get_irn_link(b)); set_irn_link(b, phi); + env->phis_moved = 1; /* first, copy all 0..k-1 predecessors */ for (i = 0; i < k; i++) { pred = get_Block_cfgpred_block(b, i); - if (is_Bad(pred)) { + if (is_Block_dead(pred)) { /* Do nothing */ - } else if (get_Block_block_visited(pred) + 1 - < get_irg_block_visited(current_ir_graph)) { + } else if (is_Block_removable(pred) && !Block_block_visited(pred)) { /* It's an empty block and not yet visited. */ for (j = 0; j < get_Block_n_cfgpreds(pred); j++) { if (! is_Bad(get_Block_cfgpred(pred, j))) @@ -508,12 +524,11 @@ static void optimize_blocks(ir_node *b, void *env) { /* and now all the rest */ for (i = k+1; i < get_Block_n_cfgpreds(b); i++) { - pred = get_nodes_block(get_Block_cfgpred(b, i)); + pred = get_Block_cfgpred_block(b, i); - if (is_Bad(get_Block_cfgpred(b, i))) { + if (is_Block_dead(pred)) { /* Do nothing */ - } else if (get_Block_block_visited(pred) +1 - < get_irg_block_visited(current_ir_graph)) { + } else if (is_Block_removable(pred) && !Block_block_visited(pred)) { /* It's an empty block and not yet visited. */ for (j = 0; j < get_Block_n_cfgpreds(pred); j++) { if (! is_Bad(get_Block_cfgpred(pred, j))) @@ -529,7 +544,7 @@ static void optimize_blocks(ir_node *b, void *env) { exchange(phi, in[0]); else set_irn_in(phi, q_preds, in); - *changed = 1; + env->changed = 1; assert(q_preds <= max_preds); // assert(p_preds == q_preds && "Wrong Phi Fix"); @@ -543,20 +558,19 @@ static void optimize_blocks(ir_node *b, void *env) { for (i = 0; i < get_Block_n_cfgpreds(b); i++) { pred = get_Block_cfgpred_block(b, i); - if (is_Bad(pred)) { + if (is_Block_dead(pred)) { /* case 1: Do nothing */ - } else if (get_Block_block_visited(pred) +1 - < get_irg_block_visited(current_ir_graph)) { + } else if (is_Block_removable(pred) && !Block_block_visited(pred)) { /* case 2: It's an empty block and not yet visited. */ assert(get_Block_n_cfgpreds(b) > 1); /* Else it should be optimized by equivalent_node. */ for (j = 0; j < get_Block_n_cfgpreds(pred); j++) { - ir_node *pred_block = get_Block_cfgpred(pred, j); + ir_node *pred_X = get_Block_cfgpred(pred, j); /* because of breaking loops, not all predecessors are Bad-clean, * so we must check this here again */ - if (! is_Bad(pred_block)) - in[n_preds++] = pred_block; + if (! is_Bad(pred_X)) + in[n_preds++] = pred_X; } /* Remove block as it might be kept alive. */ exchange(pred, b/*new_Bad()*/); @@ -568,7 +582,7 @@ static void optimize_blocks(ir_node *b, void *env) { assert(n_preds <= max_preds); set_irn_in(b, n_preds, in); - *changed = 1; + env->changed = 1; assert(get_irn_link(b) == NULL || p_preds == -1 || (n_preds == p_preds && "Wrong Phi Fix")); xfree(in); @@ -578,13 +592,13 @@ static void optimize_blocks(ir_node *b, void *env) { * Block walker: optimize all blocks using the default optimizations. * This removes Blocks that with only a Jmp predecessor. */ -static void remove_simple_blocks(ir_node *block, void *env) { +static void remove_simple_blocks(ir_node *block, void *ctx) { ir_node *new_blk = equivalent_node(block); - int *changed = env; + merge_env *env = ctx; if (new_blk != block) { exchange(block, new_blk); - *changed = 1; + env->changed = 1; } } @@ -674,11 +688,8 @@ static int handle_switch_cond(ir_node *cond) { * It walks only over block nodes and adapts these and the Phi nodes in these blocks, * which it finds in a linked list computed by the first pass. * - * We use the block_visited flag to mark empty blocks in the first + * We use the mark flag to mark removable blocks in the first * phase. - * @@@ It would be better to add a struct in the link field - * that keeps the Phi list and the mark. Place it on an obstack, as - * we will lose blocks and thereby generate memory leaks. */ void optimize_cf(ir_graph *irg) { int i, j, n; @@ -696,48 +707,53 @@ void optimize_cf(ir_graph *irg) { current_ir_graph = irg; - /* FIXME: is this still needed? */ + /* FIXME: control flow opt destroys block edges. So edges are deactivated here. Fix the edges! */ edges_deactivate(irg); - env.changed = 0; - if (get_opt_optimize() && get_opt_unreachable_code()) { - ir_node *end; - - /* kill dead blocks using dom info */ - assure_doms(irg); - irg_block_walk_graph(irg, NULL, remove_dead_block_cf, &env.changed); - - /* fix the keep-alives */ - end = get_irg_end(irg); - for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) { - ir_node *ka = get_End_keepalive(end, i); - - if (is_Block(ka)) { - /* do NOT keep dead blocks */ - if (get_Block_dom_depth(ka) < 0) { - set_End_keepalive(end, i, new_Bad()); - env.changed = 1; - } - } else if (is_Block_dead(get_nodes_block(ka)) || - get_Block_dom_depth(get_nodes_block(ka)) < 0) { - /* do NOT keep nodes in dead blocks */ + /* we use the mark flag to mark removable blocks */ + ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK); +restart: + env.changed = 0; + env.phis_moved = 0; + + /* ALWAYS kill unreachable control flow. Backend cannot handle it anyway. + Use dominator info to kill blocks. Also optimize useless Conds. */ + assure_doms(irg); + irg_block_walk_graph(irg, NULL, remove_unreachable_blocks_and_conds, &env.changed); + + /* fix the keep-alives */ + for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) { + ir_node *ka = get_End_keepalive(end, i); + + if (is_Block(ka)) { + /* do NOT keep dead blocks */ + if (is_Block_dead(ka) || get_Block_dom_depth(ka) < 0) { set_End_keepalive(end, i, new_Bad()); env.changed = 1; } + } else if (is_Block_dead(get_nodes_block(ka)) || + get_Block_dom_depth(get_nodes_block(ka)) < 0) { + /* do NOT keep nodes in dead blocks */ + set_End_keepalive(end, i, new_Bad()); + env.changed = 1; } } - irg_block_walk_graph(irg, NULL, remove_senseless_conds, &env.changed); - /* Use block visited flag to mark non-empty blocks. */ - inc_irg_block_visited(irg); - set_using_block_visited(irg); - set_using_irn_link(irg); + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); env.list = plist_new(); irg_walk(end, merge_blocks, collect_nodes, &env); - clear_using_block_visited(irg); - clear_using_irn_link(irg); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); + + if (env.changed) { + /* Handle graph state if was changed. */ + set_irg_outs_inconsistent(irg); + set_irg_doms_inconsistent(irg); + set_irg_extblk_inconsistent(irg); + set_irg_loopinfo_inconsistent(irg); + env.changed = 0; + } /* handle all collected switch-Conds */ foreach_plist(env.list, el) { @@ -747,17 +763,15 @@ void optimize_cf(ir_graph *irg) { plist_free(env.list); if (env.changed) { - /* Handle graph state if was changed. */ - set_irg_outs_inconsistent(irg); - set_irg_doms_inconsistent(irg); - set_irg_extblk_inconsistent(irg); - set_irg_loopinfo_inconsistent(irg); + /* The Cond optimization might generate unreachable code, so restart if + it happens. */ + goto restart; } /* Optimize the standard code. */ env.changed = 0; assure_doms(irg); - irg_block_walk(get_irg_end_block(irg), optimize_blocks, remove_simple_blocks, &env.changed); + irg_block_walk(get_irg_end_block(irg), optimize_blocks, remove_simple_blocks, &env); /* Walk all keep alives, optimize them if block, add to new in-array for end if useful. */ @@ -767,7 +781,7 @@ void optimize_cf(ir_graph *irg) { /* in rare cases a node may be kept alive more than once, use the visited flag to detect this */ inc_irg_visited(irg); - set_using_visited(irg); + ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); /* fix the keep alive */ for (i = j = 0; i < n; i++) { @@ -776,22 +790,18 @@ void optimize_cf(ir_graph *irg) { if (irn_not_visited(ka)) { ir_op *op = get_irn_op(ka); - if ((op == op_Block) && Block_not_block_visited(ka)) { + if ((op == op_Block) && !Block_block_visited(ka)) { /* irg_block_walk() will increase the block visited flag, but we must visit only these blocks that are not visited yet, so decrease it first. */ set_irg_block_visited(irg, get_irg_block_visited(irg) - 1); irg_block_walk(ka, optimize_blocks, remove_simple_blocks, &env.changed); mark_irn_visited(ka); in[j++] = ka; - } else if (op == op_Phi) { + } else { mark_irn_visited(ka); /* don't keep alive dead blocks */ if (! is_Block_dead(get_nodes_block(ka))) in[j++] = ka; - } else if (is_op_keep(op)) { - mark_irn_visited(ka); - if (! is_Block_dead(get_nodes_block(ka))) - in[j++] = ka; } } } @@ -800,7 +810,46 @@ void optimize_cf(ir_graph *irg) { env.changed = 1; } - clear_using_visited(irg); + ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_VISITED); + + if (env.phis_moved) { + /* Bad: when we moved Phi's, we might produce dead Phi nodes + that are kept-alive. + Some other phases cannot copy with this, so will them. + */ + n = get_End_n_keepalives(end); + if (n > 0) { + if (env.changed) { + /* Handle graph state if was changed. */ + set_irg_outs_inconsistent(irg); + } + assure_irg_outs(irg); + + for (i = j = 0; i < n; ++i) { + ir_node *ka = get_End_keepalive(end, i); + + if (is_Phi(ka)) { + int k; + + for (k = get_irn_n_outs(ka) - 1; k >= 0; --k) { + ir_node *user = get_irn_out(ka, k); + + if (user != ka && user != end) { + /* Is it a real user or just a self loop ? */ + break; + } + } + if (k >= 0) + in[j++] = ka; + } else + in[j++] = ka; + } + if (j != n) { + set_End_keepalives(end, j, in); + env.changed = 1; + } + } + } if (env.changed) { /* Handle graph state if was changed. */ @@ -810,6 +859,7 @@ void optimize_cf(ir_graph *irg) { set_irg_loopinfo_inconsistent(irg); } + /* the verifier doesn't work yet with floating nodes */ if (get_irg_pinned(irg) == op_pin_state_pinned) { /* after optimize_cf(), only Bad data flow may remain. */