X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fcfopt.c;h=e2a658a43d7b4b8538aea4b4524421a6fe80b72a;hb=43332529cf18bf9f2a0ecb384dc5dd0093eb3c8a;hp=96015fc36ecd261740cfd856fa65fee8f6510cb1;hpb=1a29d29759ede67f77585faf1d27ef27647804ed;p=libfirm diff --git a/ir/opt/cfopt.c b/ir/opt/cfopt.c index 96015fc36..e2a658a43 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. * @@ -23,9 +23,9 @@ * @author Goetz Lindenmaier, Michael Beck, Sebastian Hack * @version $Id$ */ -#ifdef HAVE_CONFIG_H -# include "config.h" -#endif +#include "config.h" + +#include "iroptimize.h" #include @@ -43,15 +43,15 @@ #include "irvrfy.h" #include "iredges.h" -#include "array.h" +#include "array_t.h" #include "irouts.h" #include "irbackedge_t.h" #include "irflag_t.h" #include "firmstat.h" +#include "irpass.h" -#include "cfopt.h" #include "iropt_dbg.h" /*------------------------------------------------------------------*/ @@ -61,10 +61,16 @@ /* 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 +86,11 @@ * 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); @@ -99,23 +104,25 @@ static void remove_senseless_conds(ir_node *bl, void *env) { ir_node *cond_j = skip_Proj(pred_j); if (cond_j == cond_i) { - ir_node *jmp = new_r_Jmp(current_ir_graph, get_nodes_block(cond_i)); + ir_node *jmp = new_r_Jmp(get_nodes_block(cond_i)); set_irn_n(bl, i, jmp); 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; /** @@ -125,7 +132,8 @@ typedef struct _merge_env { * Therefore we also optimize at control flow operations, depending * how we first reach the Block. */ -static void merge_blocks(ir_node *node, void *ctx) { +static void merge_blocks(ir_node *node, void *ctx) +{ int i; ir_node *new_block; merge_env *env = ctx; @@ -158,7 +166,7 @@ static void merge_blocks(ir_node *node, void *ctx) { if (!is_Block_dead(b)) { new_block = equivalent_node(b); - while (irn_not_visited(b) && (!is_Block_dead(new_block)) && (new_block != b)) { + while (!irn_visited(b) && !is_Block_dead(new_block) && new_block != b) { /* We would have to run gigo() if new is bad, so we promote it directly below. Nevertheless, we sometimes reach a block the first time through a dataflow node. In this case we optimized the @@ -183,63 +191,82 @@ 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 non labeled blocks */ + if (has_Block_entity(block)) + set_Block_non_removable(block); + else + 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); +static void collect_nodes(ir_node *n, void *ctx) +{ + 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_entity(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 */ @@ -251,12 +278,14 @@ 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; +static int is_pred_of(ir_node *pred, ir_node *b) +{ + 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; } @@ -288,7 +317,8 @@ static int is_pred_of(ir_node *pred, ir_node *b) { * To perform the test for pos, we must regard predecessors before pos * as already removed. **/ -static int test_whether_dispensable(ir_node *b, int pos) { +static int test_whether_dispensable(ir_node *b, int pos) +{ int i, j, n_preds = 1; ir_node *pred = get_Block_cfgpred_block(b, pos); @@ -296,17 +326,15 @@ 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 +342,8 @@ 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 +366,7 @@ 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 +417,12 @@ 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 *pred, *phi, *next; 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. */ @@ -403,22 +430,21 @@ static void optimize_blocks(ir_node *b, void *env) { for (i = 0, k = get_Block_n_cfgpreds(b); i < k; ++i) { max_preds += test_whether_dispensable(b, i); } - in = xmalloc(max_preds * sizeof(*in)); + in = XMALLOCN(ir_node*, max_preds); /*- Fix the Phi nodes of the current block -*/ - for (phi = get_irn_link(b); phi; ) { - assert(get_irn_op(phi) == op_Phi); + for (phi = get_irn_link(b); phi != NULL; phi = next) { + assert(is_Phi(phi)); + next = get_irn_link(phi); /* Find the new predecessors for the Phi */ p_preds = 0; 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); @@ -428,7 +454,7 @@ static void optimize_blocks(ir_node *b, void *env) { if (! is_Bad(get_Block_cfgpred(pred, j))) { if (get_nodes_block(phi_pred) == pred) { /* case Phi 2a: */ - assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */ + assert(is_Phi(phi_pred)); /* Block is empty!! */ in[p_preds++] = get_Phi_pred(phi_pred, j); } else { @@ -450,9 +476,7 @@ static void optimize_blocks(ir_node *b, void *env) { exchange(phi, in[0]); else set_irn_in(phi, p_preds, in); - *changed = 1; - - phi = get_irn_link(phi); + env->changed = 1; } /*- This happens only if merge between loop backedge and single loop entry. @@ -461,7 +485,7 @@ 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 +504,15 @@ 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 +532,11 @@ 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 +552,7 @@ 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 +566,19 @@ 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); + assert(get_Block_n_cfgpreds(b) > 1 || has_Block_entity(b)); /* 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 +590,7 @@ 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 +600,14 @@ 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; } } @@ -600,7 +623,8 @@ static void remove_simple_blocks(ir_node *block, void *env) { * * Expects all Proj's linked to the cond node */ -static int handle_switch_cond(ir_node *cond) { +static int handle_switch_cond(ir_node *cond) +{ ir_node *sel = get_Cond_selector(cond); ir_node *proj1 = get_irn_link(cond); @@ -611,9 +635,9 @@ static int handle_switch_cond(ir_node *cond) { if (proj2 == NULL) { /* this Cond has only one Proj: must be the defProj */ - assert(get_Cond_defaultProj(cond) == get_Proj_proj(proj1)); + assert(get_Cond_default_proj(cond) == get_Proj_proj(proj1)); /* convert it into a Jmp */ - jmp = new_r_Jmp(current_ir_graph, blk); + jmp = new_r_Jmp(blk); exchange(proj1, jmp); return 1; } else if (get_irn_link(proj2) == NULL) { @@ -624,12 +648,12 @@ static int handle_switch_cond(ir_node *cond) { if (tv != tarval_bad) { /* we have a constant switch */ long num = get_tarval_long(tv); - long def_num = get_Cond_defaultProj(cond); + long def_num = get_Cond_default_proj(cond); if (def_num == get_Proj_proj(proj1)) { /* first one is the defProj */ if (num == get_Proj_proj(proj2)) { - jmp = new_r_Jmp(current_ir_graph, blk); + jmp = new_r_Jmp(blk); exchange(proj2, jmp); exchange(proj1, new_Bad()); return 1; @@ -637,7 +661,7 @@ static int handle_switch_cond(ir_node *cond) { } else if (def_num == get_Proj_proj(proj2)) { /* second one is the defProj */ if (num == get_Proj_proj(proj1)) { - jmp = new_r_Jmp(current_ir_graph, blk); + jmp = new_r_Jmp(blk); exchange(proj1, jmp); exchange(proj2, new_Bad()); return 1; @@ -645,12 +669,12 @@ static int handle_switch_cond(ir_node *cond) { } else { /* neither: strange, Cond was not optimized so far */ if (num == get_Proj_proj(proj1)) { - jmp = new_r_Jmp(current_ir_graph, blk); + jmp = new_r_Jmp(blk); exchange(proj1, jmp); exchange(proj2, new_Bad()); return 1; } else if (num == get_Proj_proj(proj2)) { - jmp = new_r_Jmp(current_ir_graph, blk); + jmp = new_r_Jmp(blk); exchange(proj2, jmp); exchange(proj1, new_Bad()); return 1; @@ -674,14 +698,12 @@ 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; +void optimize_cf(ir_graph *irg) +{ + int i, j, n, changed; ir_node **in = NULL; ir_node *cond, *end = get_irg_end(irg); ir_graph *rem = current_ir_graph; @@ -696,48 +718,59 @@ 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) { + /* 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 */ + changed = 0; + 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()); + changed = 1; + } + } else { + ir_node *block = get_nodes_block(ka); + + if (is_Bad(block) || is_Block_dead(block) || get_Block_dom_depth(block) < 0) { /* do NOT keep nodes in dead blocks */ set_End_keepalive(end, i, new_Bad()); - env.changed = 1; + changed = 1; } } } - irg_block_walk_graph(irg, NULL, remove_senseless_conds, &env.changed); + env.changed |= 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); + set_irg_entity_usage_state(irg, ir_entity_usage_not_computed); + env.changed = 0; + } /* handle all collected switch-Conds */ foreach_plist(env.list, el) { @@ -747,60 +780,87 @@ 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); - - /* Walk all keep alives, optimize them if block, add to new in-array - for end if useful. */ - n = get_End_n_keepalives(end); - if (n > 0) - NEW_ARR_A(ir_node *, in, n); + irg_block_walk_graph(irg, optimize_blocks, remove_simple_blocks, &env); /* in rare cases a node may be kept alive more than once, use the visited flag to detect this */ + ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); inc_irg_visited(irg); - set_using_visited(irg); - /* fix the keep alive */ - for (i = j = 0; i < n; i++) { + /* fix the keep-alives again */ + changed = 0; + for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) { ir_node *ka = get_End_keepalive(end, i); - if (irn_not_visited(ka)) { - ir_op *op = get_irn_op(ka); - - if ((op == op_Block) && Block_not_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) { - 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; + 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()); + changed = 1; + } + } else { + ir_node *block = get_nodes_block(ka); + + if (is_Bad(block) || is_Block_dead(block) || get_Block_dom_depth(block) < 0) { + /* do NOT keep nodes in dead blocks */ + set_End_keepalive(end, i, new_Bad()); + changed = 1; } } } - if (j != n) { - set_End_keepalives(end, j, in); - env.changed = 1; - } + env.changed |= changed; + + remove_End_Bads_and_doublets(end); - 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) { + NEW_ARR_A(ir_node *, in, n); + 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. */ @@ -808,8 +868,10 @@ void optimize_cf(ir_graph *irg) { set_irg_doms_inconsistent(irg); set_irg_extblk_inconsistent(irg); set_irg_loopinfo_inconsistent(irg); + set_irg_entity_usage_state(irg, ir_entity_usage_not_computed); } + /* 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. */ @@ -822,3 +884,9 @@ void optimize_cf(ir_graph *irg) { current_ir_graph = rem; } + +/* Creates an ir_graph pass for optimize_cf. */ +ir_graph_pass_t *optimize_cf_pass(const char *name) +{ + return def_graph_pass(name ? name : "optimize_cf", optimize_cf); +} /* optimize_cf_pass */