X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fcfopt.c;h=7ca15daa4f252f72d955a3238e06d6449de4a5e1;hb=0ada12ba1e11d8ff2e39e3183eb43c4783df2d91;hp=a68d286746bc4c36d43217628df20a4f613428fd;hpb=8ea5207a45f6de59e29ab2cff0968e4139f79e4e;p=libfirm diff --git a/ir/opt/cfopt.c b/ir/opt/cfopt.c index a68d28674..7ca15daa4 100644 --- a/ir/opt/cfopt.c +++ b/ir/opt/cfopt.c @@ -21,7 +21,6 @@ * @file * @brief Control flow optimizations. * @author Goetz Lindenmaier, Michael Beck, Sebastian Hack - * @version $Id$ * * Removes Bad control flow predecessors and empty blocks. A block is empty * if it contains only a Jmp node. Blocks can only be removed if they are not @@ -56,6 +55,8 @@ #include "irflag_t.h" #include "firmstat.h" #include "irpass.h" +#include "irnodehashmap.h" +#include "irtools.h" #include "iropt_dbg.h" @@ -63,25 +64,38 @@ typedef struct merge_env { bool changed; /**< Set if the graph was changed. */ bool phis_moved; /**< Set if Phi nodes were moved. */ - ir_node **switch_conds; /**< Helper list for all found Switch Conds. */ } merge_env; +/** set or reset the removable property of a block. */ static void set_Block_removable(ir_node *block, bool removable) { set_Block_mark(block, removable); } -static bool is_Block_removable(ir_node *block) +/** check if a block has the removable property set. */ +static bool is_Block_removable(const ir_node *block) { return get_Block_mark(block); } -static void clear_link(ir_node *node, void *ctx) +/** checks if a given Cond node is a switch Cond. */ +static bool is_switch_Cond(const ir_node *cond) +{ + ir_node *sel = get_Cond_selector(cond); + return get_irn_mode(sel) != mode_b; +} + +/** Walker: clear link fields and mark all blocks as removable. */ +static void clear_link_and_mark_blocks_removable(ir_node *node, void *ctx) { (void) ctx; set_irn_link(node, NULL); - if (is_Block(node)) + if (is_Block(node)) { set_Block_removable(node, true); + set_Block_phis(node, NULL); + } else if (is_Phi(node)) { + set_Phi_next(node, NULL); + } } /** @@ -93,19 +107,23 @@ static void clear_link(ir_node *node, void *ctx) */ static void collect_nodes(ir_node *n, void *ctx) { - merge_env *env = (merge_env*)ctx; - + (void) ctx; if (is_Phi(n)) { /* Collect Phi nodes to compact ins along with block's ins. */ ir_node *block = get_nodes_block(n); - set_irn_link(n, get_irn_link(block)); - set_irn_link(block, n); + add_Block_phi(block, n); } else if (is_Block(n)) { - if (has_Block_entity(n)) + if (get_Block_entity(n) != NULL) { + /* block with a jump label attached cannot be removed. */ set_Block_removable(n, false); + } + } else if (is_Bad(n) || is_Jmp(n)) { + /* ignore these */ return; - } else if (!is_Jmp(n)) { /* Check for non-empty block. */ + } else { + /* Check for non-empty block. */ ir_node *block = get_nodes_block(n); + set_Block_removable(block, false); if (is_Proj(n)) { @@ -113,18 +131,12 @@ static void collect_nodes(ir_node *n, void *ctx) ir_node *pred = get_Proj_pred(n); set_irn_link(n, get_irn_link(pred)); set_irn_link(pred, n); - } else if (is_Cond(n)) { - ir_node *sel = get_Cond_selector(n); - if (get_irn_mode(sel) != mode_b) { - /* found a switch-Cond, collect */ - ARR_APP1(ir_node*, env->switch_conds, n); - } } } } -/** Returns true if pred is predecessor of block. */ -static bool is_pred_of(ir_node *pred, ir_node *b) +/** Returns true if pred is predecessor of block b. */ +static bool is_pred_of(const ir_node *pred, const ir_node *b) { int i; @@ -136,20 +148,6 @@ static bool is_pred_of(ir_node *pred, ir_node *b) return false; } -static unsigned count_non_bad_inputs(const ir_node *node) -{ - int arity = get_irn_arity(node); - unsigned result = 0; - int i; - - for (i = 0; i < arity; ++i) { - ir_node *in = get_irn_n(node, i); - if (!is_Bad(in)) - ++result; - } - return result; -} - /** Test whether we can optimize away pred block pos of b. * * @param b A block node. @@ -177,16 +175,14 @@ static unsigned count_non_bad_inputs(const ir_node *node) * To perform the test for pos, we must regard predecessors before pos * as already removed. **/ -static unsigned test_whether_dispensable(ir_node *b, int pos) +static unsigned test_whether_dispensable(const ir_node *b, int pos) { ir_node *pred = get_Block_cfgpred(b, pos); ir_node *predb = get_nodes_block(pred); - /* Bad blocks will be optimized away, so we don't need space for them */ - if (is_Bad(pred)) - return 0; - if (!is_Block_removable(predb)) + if (is_Bad(pred) || !is_Block_removable(predb)) return 1; + /* can't remove self-loops */ if (predb == b) goto non_dispensable; @@ -194,7 +190,7 @@ static unsigned test_whether_dispensable(ir_node *b, int pos) goto non_dispensable; /* Seems to be empty. At least we detected this in collect_nodes. */ - if (get_irn_link(b) != NULL) { + if (get_Block_phis(b) != NULL) { int n_cfgpreds = get_Block_n_cfgpreds(b); int i; /* there are Phi nodes */ @@ -230,7 +226,7 @@ static unsigned test_whether_dispensable(ir_node *b, int pos) if (Block_block_visited(predb)) return 1; /* if we get here, the block is dispensable, count useful preds */ - return count_non_bad_inputs(predb); + return get_irn_arity(predb); non_dispensable: set_Block_removable(predb, false); @@ -238,8 +234,28 @@ non_dispensable: } /** - * This method removed Bad cf predecessors from Blocks and Phis, and removes - * empty blocks. A block is empty if it only contains Phi and Jmp nodes. + * This method merges blocks. A block is applicable to be merged, if it + * has only one predecessor with an unconditional jump to this block; + * and if this block does not contain any phis. + */ +static void merge_blocks(ir_node *b, void *env) +{ + (void) env; + + if (get_Block_n_cfgpreds(b) == 1) { + ir_node* pred = get_Block_cfgpred(b, 0); + if (is_Jmp(pred)) { + ir_node* pred_block = get_nodes_block(pred); + if (get_Block_phis(b) == NULL) { + exchange(b, pred_block); + } + } + } +} + +/** + * This method removes empty blocks. A block is empty if it only contains Phi + * and Jmp nodes. * * We first adapt Phi nodes, then Block nodes, as we need the old ins * of the Block to adapt the Phi nodes. We do this by computing new @@ -287,10 +303,16 @@ non_dispensable: 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, *next; + ir_node *phi; + ir_node *next; ir_node **in; merge_env *env = (merge_env*)ctx; + if (get_Block_dom_depth(b) < 0) { + /* ignore unreachable blocks */ + return; + } + /* Count the number of predecessor if this block is merged with pred blocks that are empty. */ max_preds = 0; @@ -300,27 +322,36 @@ static void optimize_blocks(ir_node *b, void *ctx) in = XMALLOCN(ir_node*, max_preds); /*- Fix the Phi nodes of the current block -*/ - for (phi = (ir_node*)get_irn_link(b); phi != NULL; phi = (ir_node*)next) { - assert(is_Phi(phi)); - next = (ir_node*)get_irn_link(phi); + for (phi = get_Block_phis(b); phi != NULL; phi = next) { + next = get_Phi_next(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); + ir_graph *irg = get_irn_irg(b); + ir_node *predx = get_Block_cfgpred(b, i); + ir_node *pred; + + /* case Phi 1: maintain Bads, as somebody else is responsible to + * remove them */ + if (is_Bad(predx)) { + in[p_preds++] = new_r_Bad(irg, get_irn_mode(phi)); + continue; + } - if (is_Bad(pred)) { - /* case Phi 1: Do nothing */ - } else if (is_Block_removable(pred) && !Block_block_visited(pred)) { - /* case Phi 2: It's an empty block and not yet visited. */ + pred = get_nodes_block(predx); + + /* case Phi 2: It's an empty block and not yet visited. */ + if (is_Block_removable(pred) && !Block_block_visited(pred)) { ir_node *phi_pred = get_Phi_pred(phi, i); for (j = 0, k = get_Block_n_cfgpreds(pred); j < k; j++) { ir_node *pred_pred = 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_pred)) + + if (is_Bad(pred_pred)) { + in[p_preds++] = new_r_Bad(irg, get_irn_mode(phi)); continue; + } if (get_nodes_block(phi_pred) == pred) { /* case Phi 2a: */ @@ -340,11 +371,11 @@ static void optimize_blocks(ir_node *b, void *ctx) assert(p_preds == max_preds); /* Fix the node */ - if (p_preds == 1) - /* By removal of Bad ins the Phi might be degenerated. */ + if (p_preds == 1) { exchange(phi, in[0]); - else + } else { set_irn_in(phi, p_preds, in); + } env->changed = true; } @@ -361,35 +392,49 @@ static void optimize_blocks(ir_node *b, void *ctx) ir_node *next_phi; /* we found a predecessor block at position k that will be removed */ - for (phi = (ir_node*)get_irn_link(predb); phi; phi = next_phi) { + for (phi = get_Block_phis(predb); phi != NULL; phi = next_phi) { int q_preds = 0; - next_phi = (ir_node*)get_irn_link(phi); - - assert(is_Phi(phi)); + next_phi = get_Phi_next(phi); if (get_Block_idom(b) != predb) { - /* predb is not the dominator. There can't be uses of pred's Phi nodes, kill them .*/ - ir_graph *irg = get_irn_irg(b); - exchange(phi, get_irg_bad(irg)); + /* predb is not the dominator. There can't be uses of + * pred's Phi nodes, kill them .*/ + ir_graph *irg = get_irn_irg(b); + ir_mode *mode = get_irn_mode(phi); + exchange(phi, new_r_Bad(irg, mode)); } else { - /* predb is the direct dominator of b. There might be uses of the Phi nodes from - predb in further block, so move this phi from the predecessor into the block b */ + /* predb is the direct dominator of b. There might be uses + * of the Phi nodes from predb in further block, so move + * this phi from the predecessor into the block b */ set_nodes_block(phi, b); - set_irn_link(phi, get_irn_link(b)); - set_irn_link(b, phi); + set_Phi_next(phi, get_Block_phis(b)); + set_Block_phis(b, phi); env->phis_moved = true; /* first, copy all 0..k-1 predecessors */ for (i = 0; i < k; i++) { - pred = get_Block_cfgpred_block(b, i); - - if (is_Bad(pred)) { - /* Do nothing */ - } else if (is_Block_removable(pred) && !Block_block_visited(pred)) { + ir_node *predx = get_Block_cfgpred(b, i); + ir_node *pred_block; + + if (is_Bad(predx)) { + ir_graph *irg = get_irn_irg(b); + ir_mode *mode = get_irn_mode(phi); + in[q_preds++] = new_r_Bad(irg, mode); + continue; + } + pred_block = get_nodes_block(predx); + if (is_Block_removable(pred_block) + && !Block_block_visited(pred_block)) { + int n_cfgpreds = get_Block_n_cfgpreds(pred_block); /* 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))) + for (j = 0; j < n_cfgpreds; j++) { + if (!is_Bad(get_Block_cfgpred(pred_block, j))) { in[q_preds++] = phi; + } else { + ir_graph *irg = get_irn_irg(b); + ir_mode *mode = get_irn_mode(phi); + in[q_preds++] = new_r_Bad(irg, mode); + } } } else { in[q_preds++] = phi; @@ -399,8 +444,7 @@ static void optimize_blocks(ir_node *b, void *ctx) /* now we are at k, copy the phi predecessors */ pred = get_nodes_block(get_Block_cfgpred(b, k)); for (i = 0; i < get_Phi_n_preds(phi); i++) { - if (! is_Bad(get_Block_cfgpred(pred, i))) - in[q_preds++] = get_Phi_pred(phi, i); + in[q_preds++] = get_Phi_pred(phi, i); } /* and now all the rest */ @@ -408,12 +452,19 @@ static void optimize_blocks(ir_node *b, void *ctx) pred = get_Block_cfgpred_block(b, i); if (is_Bad(pred)) { - /* Do nothing */ + ir_graph *irg = get_irn_irg(b); + ir_mode *mode = get_irn_mode(phi); + in[q_preds++] = new_r_Bad(irg, mode); } 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))) + if (! is_Bad(get_Block_cfgpred(pred, j))) { in[q_preds++] = phi; + } else { + ir_graph *irg = get_irn_irg(b); + ir_mode *mode = get_irn_mode(phi); + in[q_preds++] = new_r_Bad(irg, mode); + } } } else { in[q_preds++] = phi; @@ -439,24 +490,28 @@ static void optimize_blocks(ir_node *b, void *ctx) for (i = 0; i < get_Block_n_cfgpreds(b); i++) { ir_node *pred = get_Block_cfgpred(b, i); ir_node *predb = get_nodes_block(pred); + ir_graph *irg = get_irn_irg(pred); - /* case 1: Do nothing */ - if (is_Bad(pred)) + /* case 1: Bad predecessor */ + if (is_Bad(pred)) { + in[n_preds++] = new_r_Bad(irg, mode_X); continue; + } if (is_Block_removable(predb) && !Block_block_visited(predb)) { /* case 2: It's an empty block and not yet visited. */ for (j = 0; j < get_Block_n_cfgpreds(predb); j++) { ir_node *predpred = get_Block_cfgpred(predb, j); - /* because of breaking loops, not all predecessors are - * Bad-clean, so we must check this here again */ - if (is_Bad(predpred)) + if (is_Bad(predpred)) { + in[n_preds++] = new_r_Bad(irg, mode_X); continue; + } + in[n_preds++] = predpred; } /* Remove block+jump as it might be kept alive. */ - exchange(pred, get_irg_bad(get_irn_irg(b))); - exchange(predb, get_irg_bad(get_irn_irg(b))); + exchange(pred, new_r_Bad(get_irn_irg(b), mode_X)); + exchange(predb, new_r_Bad(get_irn_irg(b), mode_BB)); } else { /* case 3: */ in[n_preds++] = pred; @@ -468,111 +523,306 @@ static void optimize_blocks(ir_node *b, void *ctx) env->changed = true; /* see if phi-fix was correct */ - assert(get_irn_link(b) == NULL || p_preds == -1 || (n_preds == p_preds)); + assert(get_Block_phis(b) == NULL || p_preds == -1 || (n_preds == p_preds)); xfree(in); } /** - * Block walker: optimize all blocks using the default optimizations. - * This removes Blocks with only a Jmp predecessor. + * Optimize boolean Conds, where true and false jump to the same block into a Jmp + * Block must contain no Phi nodes. + * + * Cond + * / \ + * projA projB => Jmp Bad + * \ / \ / + * block block */ -static void remove_simple_blocks(ir_node *block, void *ctx) +static bool optimize_pred_cond(ir_node *block, int i, int j) { - merge_env *env = (merge_env*)ctx; - ir_node *new_blk = equivalent_node(block); + ir_node *projA, *projB, *cond, *pred_block, *jmp, *bad; + assert(i != j); + + projA = get_Block_cfgpred(block, i); + if (!is_Proj(projA)) return false; + projB = get_Block_cfgpred(block, j); + if (!is_Proj(projB)) return false; + cond = get_Proj_pred(projA); + if (!is_Cond(cond)) return false; + + if (cond != get_Proj_pred(projB)) return false; + if (is_switch_Cond(cond)) return false; + + /* cond should actually be a Jmp */ + pred_block = get_nodes_block(cond); + jmp = new_r_Jmp(pred_block); + bad = new_r_Bad(get_irn_irg(block), mode_X); + + assert(projA != projB); + exchange(projA, jmp); + exchange(projB, bad); + return true; +} - if (new_blk != block) { - exchange(block, new_blk); - env->changed = true; +typedef enum block_flags_t { + BF_HAS_OPERATIONS = 1 << 0, + BF_HAS_PHIS = 1 << 1, + BF_IS_UNKNOWN_JUMP_TARGET = 1 << 2, +} block_flags_t; + +static bool get_block_flag(const ir_nodehashmap_t *infos, const ir_node *block, + int flag) +{ + return PTR_TO_INT(ir_nodehashmap_get(void, infos, block)) & flag; +} + +static void set_block_flag(ir_nodehashmap_t *infos, ir_node *block, + block_flags_t flag) +{ + int data = PTR_TO_INT(ir_nodehashmap_get(void, infos, block)); + data |= flag; + ir_nodehashmap_insert(infos, block, INT_TO_PTR(data)); +} + +static void clear_block_flag(ir_nodehashmap_t *infos, const ir_node *block) +{ + ir_nodehashmap_remove(infos, block); +} + +static bool has_operations(ir_nodehashmap_t *infos, const ir_node *block) +{ + return get_block_flag(infos, block, BF_HAS_OPERATIONS); +} + +static void set_has_operations(ir_nodehashmap_t *infos, ir_node *block) +{ + set_block_flag(infos, block, BF_HAS_OPERATIONS); +} + +static bool has_phis(ir_nodehashmap_t *infos, const ir_node *block) +{ + return get_block_flag(infos, block, BF_HAS_PHIS); +} + +static void set_has_phis(ir_nodehashmap_t *infos, ir_node *block) +{ + set_block_flag(infos, block, BF_HAS_PHIS); +} + +static bool is_unknown_jump_target(ir_nodehashmap_t *infos, const ir_node *block) +{ + return get_block_flag(infos, block, BF_IS_UNKNOWN_JUMP_TARGET); +} + +static void set_is_unknown_jump_target(ir_nodehashmap_t *infos, ir_node *block) +{ + set_block_flag(infos, block, BF_IS_UNKNOWN_JUMP_TARGET); +} + +/** + * Pre-Walker: fill block info information. + */ +static void compute_block_info(ir_node *n, void *x) +{ + ir_nodehashmap_t *infos = (ir_nodehashmap_t*)x; + + if (is_Block(n)) { + int i, max = get_Block_n_cfgpreds(n); + for (i=0; iblock_infos, block)) + return; + + /* optimize Cond predecessors (might produce Bad predecessors) */ + for (i = 0; i < n_preds; ++i) { + for (j = i+1; j < n_preds; ++j) { + optimize_pred_cond(block, i, j); + } } +} - /* handle Cond nodes with constant argument. In this case the localopt rules - * should have killed all obviously impossible cases. - * So the only case left to handle here is 1 defaultProj + 1case - * (this one case should be the one taken) */ - if (get_irn_link(proj2) == NULL) { - ir_tarval *tv = value_of(sel); - - if (tv != tarval_bad) { - /* we have a constant switch */ - long num = get_tarval_long(tv); - long def_num = get_Cond_default_proj(cond); - ir_graph *irg = get_irn_irg(cond); - ir_node *bad = get_irg_bad(irg); - - if (def_num == get_Proj_proj(proj1)) { - /* first one is the defProj */ - if (num == get_Proj_proj(proj2)) { - ir_node *jmp = new_r_Jmp(blk); - exchange(proj2, jmp); - exchange(proj1, bad); - return true; - } - } else if (def_num == get_Proj_proj(proj2)) { - /* second one is the defProj */ - if (num == get_Proj_proj(proj1)) { - ir_node *jmp = new_r_Jmp(blk); - exchange(proj1, jmp); - exchange(proj2, bad); - return true; - } - } else { - /* neither: strange, Cond was not optimized so far */ - if (num == get_Proj_proj(proj1)) { - ir_node *jmp = new_r_Jmp(blk); - exchange(proj1, jmp); - exchange(proj2, bad); - return true; - } else if (num == get_Proj_proj(proj2)) { - ir_node *jmp = new_r_Jmp(blk); - exchange(proj2, jmp); - exchange(proj1, bad); - return true; - } +/** + * Pre-Block walker: remove empty blocks (only contain a Jmp) + * that are control flow predecessors of the current block. + */ +static void remove_empty_blocks(ir_node *block, void *x) +{ + skip_env *env = (skip_env*)x; + int i; + int n_preds = get_Block_n_cfgpreds(block); + + for (i = 0; i < n_preds; ++i) { + ir_node *jmp, *jmp_block; + int n_jpreds = 0; + + jmp = get_Block_cfgpred(block, i); + if (!is_Jmp(jmp)) + continue; + jmp_block = get_nodes_block(jmp); + if (jmp_block == block) + continue; /* this infinite loop cannot be optimized any further */ + if (is_unknown_jump_target(&env->block_infos, jmp_block)) + continue; /* unknown jump target must not be optimized */ + if (has_phis(&env->block_infos,jmp_block)) + continue; /* this block contains Phis and is not skipped */ + if (Block_block_visited(jmp_block)) { + continue; + /* otherwise we could break the walker, + * if block was reached via + * KeepAlive edge -> jmp_block -> A ---> block, + * because the walker cannot handle Id nodes. + * + * A B + * \ / + * jmp_block + * / \ + * block End + */ + } + + /* jmp_block is an empty block and can be optimized! */ + + n_jpreds = get_Block_n_cfgpreds(jmp_block); + /** + * If the jmp block has only one predecessor this is straightforward. + * However, if there are more predecessors, we only handle this, + * if block has no Phis. + */ + if (n_jpreds == 1) { + ir_node *pred = get_Block_cfgpred(jmp_block, 0); + ir_node *pred_block = get_nodes_block(pred); + if (has_operations(&env->block_infos,jmp_block)) { + if (get_irg_start_block(get_irn_irg(pred_block)) == pred_block) + continue; /* must not merge operations into start block */ + if (!is_Jmp(pred)) + continue; /* must not create partially dead code, especially when it is mode_M */ + } + + /* skip jmp block by rerouting its predecessor to block + * + * A A + * | | + * jmp_block => | + * | | + * block block + */ + exchange(jmp, pred); + + /* cleanup: jmp_block might have a Keep edge! */ + exchange(jmp_block, pred_block); + env->changed = true; + } else if ( !has_phis(&env->block_infos, block) && + !has_operations(&env->block_infos,jmp_block)) + { + /* all predecessors can skip the jmp block, so block gets some new + * predecessors + * + * A B A B + * \ / | | + * jmp_block C => Bad C | | + * \ / \ | | / + * block block + */ + ir_node **ins = ALLOCAN(ir_node*, n_preds+n_jpreds); + int j; + /* first copy the old predecessors, because the outer loop (i) + * still walks over them */ + for (j = 0; j < n_preds; ++j) { + ins[j] = get_Block_cfgpred(block, j); } + /* now append the new predecessors */ + for (j = 0; j < n_jpreds; ++j) { + ir_node *pred = get_Block_cfgpred(jmp_block, j); + ins[n_preds+j] = pred; + } + set_irn_in(block, n_preds+n_jpreds, ins); + /* convert the jmp_block to Bad */ + ir_graph *irg = get_irn_irg(block); + exchange(jmp_block, new_r_Bad(irg, mode_BB)); + exchange(jmp, new_r_Bad(irg, mode_X)); + /* let the outer loop walk over the new predecessors as well */ + n_preds += n_jpreds; + env->changed = true; + // TODO What if jmp_block had a KeepAlive edge? + } else { + /* This would involve Phis ... */ } } - return false; } -/* Optimizations of the control flow that also require changes of Phi nodes. - * - * This optimization performs two passes over the graph. - * - * The first pass collects all Phi nodes in a link list in the block - * nodes. Further it performs simple control flow optimizations. - * Finally it marks all blocks that do not contain useful - * computations, i.e., these blocks might be removed. - * - * The second pass performs the optimizations intended by this algorithm. - * 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. +/* + * All cfg optimizations, which do not touch Phi nodes. * - * We use the mark flag to mark removable blocks in the first phase. + * Note that this might create critical edges. */ +static void cfgopt_ignoring_phis(ir_graph *irg) +{ + skip_env env; + + env.changed = true; + ir_nodehashmap_init(&env.block_infos); + + while (env.changed) { + irg_walk_graph(irg, compute_block_info, NULL, &env.block_infos); + env.changed = false; + + /* Remove blocks, which only consist of a Jmp */ + irg_block_walk_graph(irg, remove_empty_blocks, NULL, &env); + + /* Optimize Cond->Jmp, where then- and else-block are the same. */ + irg_block_walk_graph(irg, NULL, optimize_ifs, &env); + + if (env.changed) { + confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE); + /* clear block info, because it must be recomputed */ + irg_block_walk_graph(irg, clear_block_info, NULL, &env.block_infos); + /* Removing blocks and Conds might enable more optimizations */ + continue; + } else { + confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL); + break; + } + } + + ir_nodehashmap_destroy(&env.block_infos); +} + +/* Optimizations of the control flow that also require changes of Phi nodes. */ void optimize_cf(ir_graph *irg) { int i, j, n; @@ -581,51 +831,41 @@ void optimize_cf(ir_graph *irg) ir_node *new_end; merge_env env; - assert(get_irg_phase_state(irg) != phase_building); + env.changed = false; + env.phis_moved = false; /* if the graph is not pinned, we cannot determine empty blocks */ assert(get_irg_pinned(irg) != op_pin_state_floats && "Control flow optimization need a pinned graph"); - /* FIXME: control flow opt destroys block edges. So edges are deactivated - * here. Fix the edges! */ - edges_deactivate(irg); - - /* we use the mark flag to mark removable blocks */ - ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK); -restart: - env.changed = false; - env.phis_moved = false; - - assure_doms(irg); + assure_irg_properties(irg, IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE); - env.switch_conds = NEW_ARR_F(ir_node*, 0); - irg_walk(end, clear_link, collect_nodes, &env); + /* First the "simple" optimizations, which do not touch Phis */ + cfgopt_ignoring_phis(irg); - /* handle all collected switch-Conds */ - n = ARR_LEN(env.switch_conds); - for (i = 0; i < n; ++i) { - ir_node *cond = env.switch_conds[i]; - env.changed |= handle_switch_cond(cond); - } - DEL_ARR_F(env.switch_conds); - - 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); - - /* The Cond optimization might generate unreachable code, so restart if - it happens. */ - goto restart; - } - - /* Optimize the standard code. */ - assure_doms(irg); - irg_block_walk_graph(irg, optimize_blocks, remove_simple_blocks, &env); + /* we use the mark flag to mark removable blocks */ + ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK + | IR_RESOURCE_PHI_LIST); + + /* + * This pass collects all Phi nodes in a link list in the block + * nodes. Further it performs simple control flow optimizations. + * Finally it marks all blocks that do not contain useful + * computations, i.e., these blocks might be removed. + */ + irg_walk(end, clear_link_and_mark_blocks_removable, collect_nodes, NULL); + + /* assert due to collect_nodes: + * 1. removable blocks are now marked as such + * 2. phi lists are up to date + */ + + /* Optimize the standard code. + * It walks only over block nodes and adapts these and the Phi nodes in + * these blocks, which it finds in a linked list computed before. + */ + assure_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE); + irg_block_walk_graph(irg, optimize_blocks, merge_blocks, &env); new_end = optimize_in_place(end); if (new_end != end) { @@ -634,20 +874,17 @@ restart: } remove_End_Bads_and_doublets(end); - ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK); + ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK + | IR_RESOURCE_PHI_LIST); 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. + Some other phases cannot copy with this, so kill 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) { @@ -676,23 +913,8 @@ restart: } } - 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); - } - - /* 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. */ - if (irg_verify_bads(irg, BAD_DF | BAD_BLOCK | TUPLE)) { - dump_ir_graph(irg, "-verify-cf"); - fprintf(stderr, "VERIFY_BAD in optimize_cf()\n"); - } - } + confirm_irg_properties(irg, + env.changed ? IR_GRAPH_PROPERTIES_NONE : IR_GRAPH_PROPERTIES_ALL); } /* Creates an ir_graph pass for optimize_cf. */