-/********************************************************************/
-/* Apply inlineing to small methods. */
-/********************************************************************/
-
-/** Represents a possible inlinable call in a graph. */
-typedef struct _call_entry call_entry;
-struct _call_entry {
- ir_node *call; /**< the Call */
- ir_graph *callee; /**< the callee called here */
- call_entry *next; /**< for linking the next one */
-};
-
-/**
- * environment for inlining small irgs
- */
-typedef struct _inline_env_t {
- struct obstack obst; /**< an obstack where call_entries are allocated on. */
- call_entry *head; /**< the head of the call entry list */
- call_entry *tail; /**< the tail of the call entry list */
-} inline_env_t;
-
-/**
- * Returns the irg called from a Call node. If the irg is not
- * known, NULL is returned.
- */
-static ir_graph *get_call_called_irg(ir_node *call) {
- ir_node *addr;
- ir_graph *called_irg = NULL;
-
- addr = get_Call_ptr(call);
- if (is_SymConst(addr) && get_SymConst_kind(addr) == symconst_addr_ent) {
- called_irg = get_entity_irg(get_SymConst_entity(addr));
- }
-
- return called_irg;
-}
-
-/**
- * Walker: Collect all calls to known graphs inside a graph.
- */
-static void collect_calls(ir_node *call, void *env) {
- if (is_Call(call)) {
- ir_graph *called_irg = get_call_called_irg(call);
- if (called_irg) {
- /* The Call node calls a locally defined method. Remember to inline. */
- inline_env_t *ienv = env;
- call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
- entry->call = call;
- entry->callee = called_irg;
- entry->next = NULL;
-
- if (ienv->tail == NULL)
- ienv->head = entry;
- else
- ienv->tail->next = entry;
- ienv->tail = entry;
- }
- }
-}
-
-/**
- * Inlines all small methods at call sites where the called address comes
- * from a Const node that references the entity representing the called
- * method.
- * The size argument is a rough measure for the code size of the method:
- * Methods where the obstack containing the firm graph is smaller than
- * size are inlined.
- */
-void inline_small_irgs(ir_graph *irg, int size) {
- ir_graph *rem = current_ir_graph;
- inline_env_t env;
- call_entry *entry;
- DEBUG_ONLY(firm_dbg_module_t *dbg;)
-
- if (!(get_opt_optimize() && get_opt_inline())) return;
-
- FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
-
- current_ir_graph = irg;
- /* Handle graph state */
- assert(get_irg_phase_state(irg) != phase_building);
- free_callee_info(irg);
-
- /* Find Call nodes to inline.
- (We can not inline during a walk of the graph, as inlineing the same
- method several times changes the visited flag of the walked graph:
- after the first inlineing visited of the callee equals visited of
- the caller. With the next inlineing both are increased.) */
- obstack_init(&env.obst);
- env.head = env.tail = NULL;
- irg_walk_graph(irg, NULL, collect_calls, &env);
-
- if (env.head != NULL) {
- /* There are calls to inline */
- collect_phiprojs(irg);
- for (entry = env.head; entry != NULL; entry = entry->next) {
- ir_graph *callee = entry->callee;
- if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
- (get_irg_inline_property(callee) >= irg_inline_forced)) {
- inline_method(entry->call, callee);
- }
- }
- }
- obstack_free(&env.obst, NULL);
- current_ir_graph = rem;
-}
-
-/**
- * Environment for inlining irgs.
- */
-typedef struct {
- int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
- int n_nodes_orig; /**< for statistics */
- call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
- call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
- int n_call_nodes; /**< Number of Call nodes in the graph. */
- int n_call_nodes_orig; /**< for statistics */
- int n_callers; /**< Number of known graphs that call this graphs. */
- int n_callers_orig; /**< for statistics */
- int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
-} inline_irg_env;
-
-/**
- * Allocate a new environment for inlining.
- */
-static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
- inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
- env->n_nodes = -2; /* do not count count Start, End */
- env->n_nodes_orig = -2; /* do not count Start, End */
- env->call_head = NULL;
- env->call_tail = NULL;
- env->n_call_nodes = 0;
- env->n_call_nodes_orig = 0;
- env->n_callers = 0;
- env->n_callers_orig = 0;
- env->got_inline = 0;
- return env;
-}
-
-typedef struct walker_env {
- struct obstack *obst; /**< the obstack for allocations. */
- inline_irg_env *x; /**< the inline environment */
- int ignore_runtime; /**< the ignore runtime flag */
-} wenv_t;
-
-/**
- * post-walker: collect all calls in the inline-environment
- * of a graph and sum some statistics.
- */
-static void collect_calls2(ir_node *call, void *ctx) {
- wenv_t *env = ctx;
- inline_irg_env *x = env->x;
- ir_op *op = get_irn_op(call);
- ir_graph *callee;
- call_entry *entry;
-
- /* count meaningful nodes in irg */
- if (op != op_Proj && op != op_Tuple && op != op_Sync) {
- ++x->n_nodes;
- ++x->n_nodes_orig;
- }
-
- if (op != op_Call) return;
-
- /* check, if it's a runtime call */
- if (env->ignore_runtime) {
- ir_node *symc = get_Call_ptr(call);
-
- if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
- ir_entity *ent = get_SymConst_entity(symc);
-
- if (get_entity_additional_properties(ent) & mtp_property_runtime)
- return;
- }
- }
-
- /* collect all call nodes */
- ++x->n_call_nodes;
- ++x->n_call_nodes_orig;
-
- callee = get_call_called_irg(call);
- if (callee) {
- inline_irg_env *callee_env = get_irg_link(callee);
- /* count all static callers */
- ++callee_env->n_callers;
- ++callee_env->n_callers_orig;
-
- /* link it in the list of possible inlinable entries */
- entry = obstack_alloc(env->obst, sizeof(*entry));
- entry->call = call;
- entry->callee = callee;
- entry->next = NULL;
- if (x->call_tail == NULL)
- x->call_head = entry;
- else
- x->call_tail->next = entry;
- x->call_tail = entry;
- }
-}
-
-/**
- * Returns TRUE if the number of callers is 0 in the irg's environment,
- * hence this irg is a leave.
- */
-INLINE static int is_leave(ir_graph *irg) {
- inline_irg_env *env = get_irg_link(irg);
- return env->n_call_nodes == 0;
-}
-
-/**
- * Returns TRUE if the number of nodes in the callee is
- * smaller then size in the irg's environment.
- */
-INLINE static int is_smaller(ir_graph *callee, int size) {
- inline_irg_env *env = get_irg_link(callee);
- return env->n_nodes < size;
-}
-
-/**
- * Append the nodes of the list src to the nodes of the list in environment dst.
- */
-static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
- call_entry *entry, *nentry;
-
- /* Note that the src list points to Call nodes in the inlined graph, but
- we need Call nodes in our graph. Luckily the inliner leaves this information
- in the link field. */
- for (entry = src; entry != NULL; entry = entry->next) {
- nentry = obstack_alloc(obst, sizeof(*nentry));
- nentry->call = get_irn_link(entry->call);
- nentry->callee = entry->callee;
- nentry->next = NULL;
- dst->call_tail->next = nentry;
- dst->call_tail = nentry;
- }
-}
-
-/*
- * Inlines small leave methods at call sites where the called address comes
- * from a Const node that references the entity representing the called
- * method.
- * The size argument is a rough measure for the code size of the method:
- * Methods where the obstack containing the firm graph is smaller than
- * size are inlined.
- */
-void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
- inline_irg_env *env;
- ir_graph *irg;
- int i, n_irgs;
- ir_graph *rem;
- int did_inline;
- wenv_t wenv;
- call_entry *entry, *tail;
- const call_entry *centry;
- struct obstack obst;
- DEBUG_ONLY(firm_dbg_module_t *dbg;)
-
- if (!(get_opt_optimize() && get_opt_inline())) return;
-
- FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
- rem = current_ir_graph;
- obstack_init(&obst);
-
- /* extend all irgs by a temporary data structure for inlining. */
- n_irgs = get_irp_n_irgs();
- for (i = 0; i < n_irgs; ++i)
- set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
-
- /* Precompute information in temporary data structure. */
- wenv.obst = &obst;
- wenv.ignore_runtime = ignore_runtime;
- for (i = 0; i < n_irgs; ++i) {
- ir_graph *irg = get_irp_irg(i);
-
- assert(get_irg_phase_state(irg) != phase_building);
- free_callee_info(irg);
-
- wenv.x = get_irg_link(irg);
- irg_walk_graph(irg, NULL, collect_calls2, &wenv);
- }
-
- /* -- and now inline. -- */
-
- /* Inline leaves recursively -- we might construct new leaves. */
- do {
- did_inline = 0;
-
- for (i = 0; i < n_irgs; ++i) {
- ir_node *call;
- int phiproj_computed = 0;
-
- current_ir_graph = get_irp_irg(i);
- env = (inline_irg_env *)get_irg_link(current_ir_graph);
-
- tail = NULL;
- for (entry = env->call_head; entry != NULL; entry = entry->next) {
- ir_graph *callee;
-
- if (env->n_nodes > maxsize) break;
-
- call = entry->call;
- callee = entry->callee;
-
- if (is_leave(callee) && is_smaller(callee, leavesize)) {
- if (!phiproj_computed) {
- phiproj_computed = 1;
- collect_phiprojs(current_ir_graph);
- }
- did_inline = inline_method(call, callee);
-
- if (did_inline) {
- /* Do some statistics */
- inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
-
- env->got_inline = 1;
- --env->n_call_nodes;
- env->n_nodes += callee_env->n_nodes;
- --callee_env->n_callers;
-
- /* remove this call from the list */
- if (tail != NULL)
- tail->next = entry->next;
- else
- env->call_head = entry->next;
- continue;
- }
- }
- tail = entry;
- }
- env->call_tail = tail;
- }
- } while (did_inline);
-
- /* inline other small functions. */
- for (i = 0; i < n_irgs; ++i) {
- ir_node *call;
- int phiproj_computed = 0;
-
- current_ir_graph = get_irp_irg(i);
- env = (inline_irg_env *)get_irg_link(current_ir_graph);
-
- /* note that the list of possible calls is updated during the process */
- tail = NULL;
- for (entry = env->call_head; entry != NULL; entry = entry->next) {
- ir_graph *callee;
-
- call = entry->call;
- callee = entry->callee;
-
- if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
- (get_irg_inline_property(callee) >= irg_inline_forced))) {
- if (!phiproj_computed) {
- phiproj_computed = 1;
- collect_phiprojs(current_ir_graph);
- }
- if (inline_method(call, callee)) {
- inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
-
- /* callee was inline. Append it's call list. */
- env->got_inline = 1;
- --env->n_call_nodes;
- append_call_list(&obst, env, callee_env->call_head);
- env->n_call_nodes += callee_env->n_call_nodes;
- env->n_nodes += callee_env->n_nodes;
- --callee_env->n_callers;
-
- /* after we have inlined callee, all called methods inside callee
- are now called once more */
- for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
- inline_irg_env *penv = get_irg_link(centry->callee);
- ++penv->n_callers;
- }
-
- /* remove this call from the list */
- if (tail != NULL)
- tail->next = entry->next;
- else
- env->call_head = entry->next;
- continue;
- }
- }
- tail = entry;
- }
- env->call_tail = tail;
- }
-
- for (i = 0; i < n_irgs; ++i) {
- irg = get_irp_irg(i);
- env = (inline_irg_env *)get_irg_link(irg);
-
- if (env->got_inline) {
- /* this irg got calls inlined */
- set_irg_outs_inconsistent(irg);
- set_irg_doms_inconsistent(irg);
-
- optimize_graph_df(irg);
- optimize_cf(irg);
- }
- if (env->got_inline || (env->n_callers_orig != env->n_callers))
- DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
- env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
- env->n_callers_orig, env->n_callers,
- get_entity_name(get_irg_entity(irg))));
- }
-
- obstack_free(&obst, NULL);
- current_ir_graph = rem;
-}
-
-/*******************************************************************/
-/* Code Placement. Pins all floating nodes to a block where they */
-/* will be executed only if needed. */
-/*******************************************************************/
-
-/**
- * Returns non-zero, is a block is not reachable from Start.
- *
- * @param block the block to test
- */
-static int
-is_Block_unreachable(ir_node *block) {
- return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
-}
-
-/**
- * Find the earliest correct block for node n. --- Place n into the
- * same Block as its dominance-deepest Input.
- *
- * We have to avoid calls to get_nodes_block() here
- * because the graph is floating.
- *
- * move_out_of_loops() expects that place_floats_early() have placed
- * all "living" nodes into a living block. That's why we must
- * move nodes in dead block with "live" successors into a valid
- * block.
- * We move them just into the same block as it's successor (or
- * in case of a Phi into the effective use block). For Phi successors,
- * this may still be a dead block, but then there is no real use, as
- * the control flow will be dead later.
- *
- * @param n the node to be placed
- * @param worklist a worklist, predecessors of non-floating nodes are placed here
- */
-static void
-place_floats_early(ir_node *n, waitq *worklist) {
- int i, irn_arity;
-
- /* we must not run into an infinite loop */
- assert(irn_not_visited(n));
- mark_irn_visited(n);
-
- /* Place floating nodes. */
- if (get_irn_pinned(n) == op_pin_state_floats) {
- ir_node *curr_block = get_nodes_block(n);
- int in_dead_block = is_Block_unreachable(curr_block);
- int depth = 0;
- ir_node *b = NULL; /* The block to place this node in */
-
- assert(is_no_Block(n));
-
- if (is_irn_start_block_placed(n)) {
- /* These nodes will not be placed by the loop below. */
- b = get_irg_start_block(current_ir_graph);
- depth = 1;
- }
-
- /* find the block for this node. */
- irn_arity = get_irn_arity(n);
- for (i = 0; i < irn_arity; i++) {
- ir_node *pred = get_irn_n(n, i);
- ir_node *pred_block;
-
- if ((irn_not_visited(pred))
- && (get_irn_pinned(pred) == op_pin_state_floats)) {
-
- /*
- * If the current node is NOT in a dead block, but one of its
- * predecessors is, we must move the predecessor to a live block.
- * Such thing can happen, if global CSE chose a node from a dead block.
- * We move it simply to our block.
- * Note that neither Phi nor End nodes are floating, so we don't
- * need to handle them here.
- */
- if (! in_dead_block) {
- if (get_irn_pinned(pred) == op_pin_state_floats &&
- is_Block_unreachable(get_nodes_block(pred)))
- set_nodes_block(pred, curr_block);
- }
- place_floats_early(pred, worklist);
- }
-
- /*
- * A node in the Bad block must stay in the bad block,
- * so don't compute a new block for it.
- */
- if (in_dead_block)
- continue;
-
- /* Because all loops contain at least one op_pin_state_pinned node, now all
- our inputs are either op_pin_state_pinned or place_early() has already
- been finished on them. We do not have any unfinished inputs! */
- pred_block = get_nodes_block(pred);
- if ((!is_Block_dead(pred_block)) &&
- (get_Block_dom_depth(pred_block) > depth)) {
- b = pred_block;
- depth = get_Block_dom_depth(pred_block);
- }
- /* Avoid that the node is placed in the Start block */
- if (depth == 1 &&
- get_Block_dom_depth(get_nodes_block(n)) > 1 &&
- get_irg_phase_state(current_ir_graph) != phase_backend) {
- b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
- assert(b != get_irg_start_block(current_ir_graph));
- depth = 2;
- }
- }
- if (b)
- set_nodes_block(n, b);
- }
-
- /*
- * Add predecessors of non floating nodes and non-floating predecessors
- * of floating nodes to worklist and fix their blocks if the are in dead block.
- */
- irn_arity = get_irn_arity(n);
-
- if (is_End(n)) {
- /*
- * Simplest case: End node. Predecessors are keep-alives,
- * no need to move out of dead block.
- */
- for (i = -1; i < irn_arity; ++i) {
- ir_node *pred = get_irn_n(n, i);
- if (irn_not_visited(pred))
- waitq_put(worklist, pred);
- }
- } else if (is_Block(n)) {
- /*
- * Blocks: Predecessors are control flow, no need to move
- * them out of dead block.
- */
- for (i = irn_arity - 1; i >= 0; --i) {
- ir_node *pred = get_irn_n(n, i);
- if (irn_not_visited(pred))
- waitq_put(worklist, pred);
- }
- } else if (is_Phi(n)) {
- ir_node *pred;
- ir_node *curr_block = get_nodes_block(n);
- int in_dead_block = is_Block_unreachable(curr_block);
-
- /*
- * Phi nodes: move nodes from dead blocks into the effective use
- * of the Phi-input if the Phi is not in a bad block.
- */
- pred = get_nodes_block(n);
- if (irn_not_visited(pred))
- waitq_put(worklist, pred);
-
- for (i = irn_arity - 1; i >= 0; --i) {
- ir_node *pred = get_irn_n(n, i);
-
- if (irn_not_visited(pred)) {
- if (! in_dead_block &&
- get_irn_pinned(pred) == op_pin_state_floats &&
- is_Block_unreachable(get_nodes_block(pred))) {
- set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
- }
- waitq_put(worklist, pred);
- }
- }
- } else {
- ir_node *pred;
- ir_node *curr_block = get_nodes_block(n);
- int in_dead_block = is_Block_unreachable(curr_block);
-
- /*
- * All other nodes: move nodes from dead blocks into the same block.
- */
- pred = get_nodes_block(n);
- if (irn_not_visited(pred))
- waitq_put(worklist, pred);
-
- for (i = irn_arity - 1; i >= 0; --i) {
- ir_node *pred = get_irn_n(n, i);
-
- if (irn_not_visited(pred)) {
- if (! in_dead_block &&
- get_irn_pinned(pred) == op_pin_state_floats &&
- is_Block_unreachable(get_nodes_block(pred))) {
- set_nodes_block(pred, curr_block);
- }
- waitq_put(worklist, pred);
- }
- }
- }
-}
-
-/**
- * Floating nodes form subgraphs that begin at nodes as Const, Load,
- * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
- * places all floating nodes reachable from its argument through floating
- * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
- *
- * @param worklist a worklist, used for the algorithm, empty on in/output
- */
-static void place_early(waitq *worklist) {
- assert(worklist);
- inc_irg_visited(current_ir_graph);
-
- /* this inits the worklist */
- place_floats_early(get_irg_end(current_ir_graph), worklist);
-
- /* Work the content of the worklist. */
- while (!waitq_empty(worklist)) {
- ir_node *n = waitq_get(worklist);
- if (irn_not_visited(n))
- place_floats_early(n, worklist);
- }
-
- set_irg_outs_inconsistent(current_ir_graph);
- set_irg_pinned(current_ir_graph, op_pin_state_pinned);
-}
-
-/**
- * Compute the deepest common ancestor of block and dca.
- */
-static ir_node *calc_dca(ir_node *dca, ir_node *block) {
- assert(block);
-
- /* we do not want to place nodes in dead blocks */
- if (is_Block_dead(block))
- return dca;
-
- /* We found a first legal placement. */
- if (!dca) return block;
-
- /* Find a placement that is dominates both, dca and block. */
- while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
- block = get_Block_idom(block);
-
- while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
- dca = get_Block_idom(dca);
- }
-
- while (block != dca) {
- block = get_Block_idom(block); dca = get_Block_idom(dca);
- }
-
- return dca;
-}
-
-/** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
- * I.e., DCA is the block where we might place PRODUCER.
- * A data flow edge points from producer to consumer.
- */
-static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
-{
- /* Compute the last block into which we can place a node so that it is
- before consumer. */
- if (is_Phi(consumer)) {
- /* our consumer is a Phi-node, the effective use is in all those
- blocks through which the Phi-node reaches producer */
- ir_node *phi_block = get_nodes_block(consumer);
- int arity = get_irn_arity(consumer);
- int i;
-
- for (i = 0; i < arity; i++) {
- if (get_Phi_pred(consumer, i) == producer) {
- ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
-
- if (!is_Block_unreachable(new_block))
- dca = calc_dca(dca, new_block);
- }
- }
- } else {
- dca = calc_dca(dca, get_nodes_block(consumer));
- }
-
- return dca;
-}
-
-/* FIXME: the name clashes here with the function from ana/field_temperature.c
- * please rename. */
-static INLINE int get_irn_loop_depth(ir_node *n) {
- return get_loop_depth(get_irn_loop(n));
-}
-
-/**
- * Move n to a block with less loop depth than it's current block. The
- * new block must be dominated by early.
- *
- * @param n the node that should be moved
- * @param early the earliest block we can n move to
- */
-static void move_out_of_loops(ir_node *n, ir_node *early) {
- ir_node *best, *dca;
- assert(n && early);
-
-
- /* Find the region deepest in the dominator tree dominating
- dca with the least loop nesting depth, but still dominated
- by our early placement. */
- dca = get_nodes_block(n);
-
- best = dca;
- while (dca != early) {
- dca = get_Block_idom(dca);
- if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
- if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
- best = dca;
- }
- }
- if (best != get_nodes_block(n)) {
- /* debug output
- printf("Moving out of loop: "); DDMN(n);
- printf(" Outermost block: "); DDMN(early);
- printf(" Best block: "); DDMN(best);
- printf(" Innermost block: "); DDMN(get_nodes_block(n));
- */
- set_nodes_block(n, best);
- }
-}
-
-/* deepest common ancestor in the dominator tree of all nodes'
- blocks depending on us; our final placement has to dominate DCA. */
-static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
-{
- int i;
-
- for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
- ir_node *succ = get_irn_out(node, i);
-
- if (is_End(succ)) {
- /*
- * This consumer is the End node, a keep alive edge.
- * This is not a real consumer, so we ignore it
- */
- continue;
- }
-
- if (is_Proj(succ)) {
- dca = get_deepest_common_ancestor(succ, dca);
- } else {
- /* ignore if succ is in dead code */
- ir_node *succ_blk = get_nodes_block(succ);
- if (is_Block_unreachable(succ_blk))
- continue;
- dca = consumer_dom_dca(dca, succ, node);
- }
- }
-
- return dca;
-}
-
-static void set_projs_block(ir_node *node, ir_node *block)