/*
- * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* @file
* @brief Dead node elimination and Procedure Inlining.
* @author Michael Beck, Goetz Lindenmaier
- * @version $Id$
*/
#include "config.h"
#include <limits.h>
+#include <stdbool.h>
#include <assert.h>
#include "irnode_t.h"
#include "irouts.h"
#include "irloop_t.h"
#include "irbackedge_t.h"
-#include "opt_inline_t.h"
+#include "opt_init.h"
#include "cgana.h"
#include "trouts.h"
#include "error.h"
#include "irhooks.h"
#include "irtools.h"
#include "iropt_dbg.h"
+#include "irpass_t.h"
+#include "irnodemap.h"
DEBUG_ONLY(static firm_dbg_module_t *dbg;)
/**
* Remember the new node in the old node by using a field all nodes have.
*/
-#define set_new_node(oldn, newn) set_irn_link(oldn, newn)
-
-/**
- * Get this new node, before the old node is forgotten.
- */
-#define get_new_node(oldn) get_irn_link(oldn)
-
-/**
- * Check if a new node was set.
- */
-#define has_new_node(n) (get_new_node(n) != NULL)
-
-/**
- * We use the block_visited flag to mark that we have computed the
- * number of useful predecessors for this block.
- * Further we encode the new arity in this flag in the old blocks.
- * Remembering the arity is useful, as it saves a lot of pointer
- * accesses. This function is called for all Phi and Block nodes
- * in a Block.
- */
-static inline int
-compute_new_arity(ir_node *b) {
- int i, res, irn_arity;
- int irg_v, block_v;
-
- irg_v = get_irg_block_visited(current_ir_graph);
- block_v = get_Block_block_visited(b);
- if (block_v >= irg_v) {
- /* we computed the number of preds for this block and saved it in the
- block_v flag */
- return block_v - irg_v;
- } else {
- /* compute the number of good predecessors */
- res = irn_arity = get_irn_arity(b);
- for (i = 0; i < irn_arity; i++)
- if (is_Bad(get_irn_n(b, i))) res--;
- /* save it in the flag. */
- set_Block_block_visited(b, irg_v + res);
- return res;
- }
-}
-
-/**
- * Copies the node to the new obstack. The Ins of the new node point to
- * the predecessors on the old obstack. For block/phi nodes not all
- * predecessors might be copied. n->link points to the new node.
- * For Phi and Block nodes the function allocates in-arrays with an arity
- * only for useful predecessors. The arity is determined by counting
- * the non-bad predecessors of the block.
- *
- * @param n The node to be copied
- * @param env if non-NULL, the node number attribute will be copied to the new node
- *
- * Note: Also used for loop unrolling.
- */
-static void copy_node(ir_node *n, void *env) {
- ir_node *nn, *block;
- int new_arity;
- ir_op *op = get_irn_op(n);
- (void) env;
-
- if (op == op_Bad) {
- /* node copied already */
- return;
- } else if (op == op_Block) {
- block = NULL;
- new_arity = compute_new_arity(n);
- n->attr.block.graph_arr = NULL;
- } else {
- block = get_nodes_block(n);
- if (op == op_Phi) {
- new_arity = compute_new_arity(block);
- } else {
- new_arity = get_irn_arity(n);
- }
- }
- nn = new_ir_node(get_irn_dbg_info(n),
- current_ir_graph,
- block,
- op,
- get_irn_mode(n),
- new_arity,
- get_irn_in(n) + 1);
- /* Copy the attributes. These might point to additional data. If this
- was allocated on the old obstack the pointers now are dangling. This
- frees e.g. the memory of the graph_arr allocated in new_immBlock. */
- if (op == op_Block) {
- /* we cannot allow blocks WITHOUT macroblock input */
- set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
- }
- copy_node_attr(n, nn);
-
- if (env != NULL) {
- /* for easier debugging, we want to copy the node numbers too */
- nn->node_nr = n->node_nr;
- }
-
- set_new_node(n, nn);
- hook_dead_node_elim_subst(current_ir_graph, n, nn);
-}
-
-/**
- * Copies new predecessors of old node to new node remembered in link.
- * Spare the Bad predecessors of Phi and Block nodes.
- */
-static void copy_preds(ir_node *n, void *env) {
- ir_node *nn, *block;
- int i, j, irn_arity;
- (void) env;
-
- nn = get_new_node(n);
-
- if (is_Block(n)) {
- /* copy the macro block header */
- ir_node *mbh = get_Block_MacroBlock(n);
-
- if (mbh == n) {
- /* this block is a macroblock header */
- set_Block_MacroBlock(nn, nn);
- } else {
- /* get the macro block header */
- ir_node *nmbh = get_new_node(mbh);
- assert(nmbh != NULL);
- set_Block_MacroBlock(nn, nmbh);
- }
-
- /* Don't copy Bad nodes. */
- j = 0;
- irn_arity = get_irn_arity(n);
- for (i = 0; i < irn_arity; i++) {
- if (! is_Bad(get_irn_n(n, i))) {
- ir_node *pred = get_irn_n(n, i);
- set_irn_n(nn, j, get_new_node(pred));
- j++;
- }
- }
- /* repair the block visited flag from above misuse. Repair it in both
- graphs so that the old one can still be used. */
- set_Block_block_visited(nn, 0);
- set_Block_block_visited(n, 0);
- /* Local optimization could not merge two subsequent blocks if
- in array contained Bads. Now it's possible.
- We don't call optimize_in_place as it requires
- that the fields in ir_graph are set properly. */
- if (!has_Block_label(nn) &&
- get_opt_control_flow_straightening() &&
- get_Block_n_cfgpreds(nn) == 1 &&
- is_Jmp(get_Block_cfgpred(nn, 0))) {
- ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
- if (nn == old) {
- /* Jmp jumps into the block it is in -- deal self cycle. */
- assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
- exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
- } else {
- exchange(nn, old);
- }
- }
- } else if (is_Phi(n) && get_irn_arity(n) > 0) {
- /* Don't copy node if corresponding predecessor in block is Bad.
- The Block itself should not be Bad. */
- block = get_nodes_block(n);
- set_nodes_block(nn, get_new_node(block));
- j = 0;
- irn_arity = get_irn_arity(n);
- for (i = 0; i < irn_arity; i++) {
- if (! is_Bad(get_irn_n(block, i))) {
- ir_node *pred = get_irn_n(n, i);
- set_irn_n(nn, j, get_new_node(pred));
- /*if (is_backedge(n, i)) set_backedge(nn, j);*/
- j++;
- }
- }
- /* If the pre walker reached this Phi after the post walker visited the
- block block_visited is > 0. */
- set_Block_block_visited(get_nodes_block(n), 0);
- /* Compacting the Phi's ins might generate Phis with only one
- predecessor. */
- if (get_irn_arity(nn) == 1)
- exchange(nn, get_irn_n(nn, 0));
- } else {
- irn_arity = get_irn_arity(n);
- for (i = -1; i < irn_arity; i++)
- set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
- }
- /* Now the new node is complete. We can add it to the hash table for CSE.
- @@@ inlining aborts if we identify End. Why? */
- if (!is_End(nn))
- add_identities(current_ir_graph->value_table, nn);
-}
-
-/**
- * Copies the graph recursively, compacts the keep-alives of the end node.
- *
- * @param irg the graph to be copied
- * @param copy_node_nr If non-zero, the node number will be copied
- */
-static void copy_graph(ir_graph *irg, int copy_node_nr) {
- ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
- ir_node *ka; /* keep alive */
- int i, irn_arity;
- unsigned long vfl;
-
- /* Some nodes must be copied by hand, sigh */
- vfl = get_irg_visited(irg);
- set_irg_visited(irg, vfl + 1);
-
- oe = get_irg_end(irg);
- mark_irn_visited(oe);
- /* copy the end node by hand, allocate dynamic in array! */
- ne = new_ir_node(get_irn_dbg_info(oe),
- irg,
- NULL,
- op_End,
- mode_X,
- -1,
- NULL);
- /* Copy the attributes. Well, there might be some in the future... */
- copy_node_attr(oe, ne);
- set_new_node(oe, ne);
-
- /* copy the Bad node */
- ob = get_irg_bad(irg);
- mark_irn_visited(ob);
- nb = new_ir_node(get_irn_dbg_info(ob),
- irg,
- NULL,
- op_Bad,
- mode_T,
- 0,
- NULL);
- copy_node_attr(ob, nb);
- set_new_node(ob, nb);
-
- /* copy the NoMem node */
- om = get_irg_no_mem(irg);
- mark_irn_visited(om);
- nm = new_ir_node(get_irn_dbg_info(om),
- irg,
- NULL,
- op_NoMem,
- mode_M,
- 0,
- NULL);
- copy_node_attr(om, nm);
- set_new_node(om, nm);
-
- /* copy the live nodes */
- set_irg_visited(irg, vfl);
- irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
-
- /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
-
- /* visit the anchors as well */
- for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
- ir_node *n = get_irg_anchor(irg, i);
-
- if (n && (get_irn_visited(n) <= vfl)) {
- set_irg_visited(irg, vfl);
- irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
- }
- }
-
- /* copy_preds for the end node ... */
- set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
-
- /*- ... and now the keep alives. -*/
- /* First pick the not marked block nodes and walk them. We must pick these
- first as else we will oversee blocks reachable from Phis. */
- irn_arity = get_End_n_keepalives(oe);
- for (i = 0; i < irn_arity; i++) {
- ka = get_End_keepalive(oe, i);
- if (is_Block(ka)) {
- if (get_irn_visited(ka) <= vfl) {
- /* We must keep the block alive and copy everything reachable */
- set_irg_visited(irg, vfl);
- irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
- }
- add_End_keepalive(ne, get_new_node(ka));
- }
- }
-
- /* Now pick other nodes. Here we will keep all! */
- irn_arity = get_End_n_keepalives(oe);
- for (i = 0; i < irn_arity; i++) {
- ka = get_End_keepalive(oe, i);
- if (!is_Block(ka)) {
- if (get_irn_visited(ka) <= vfl) {
- /* We didn't copy the node yet. */
- set_irg_visited(irg, vfl);
- irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
- }
- add_End_keepalive(ne, get_new_node(ka));
- }
- }
-
- /* start block sometimes only reached after keep alives */
- set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
- set_nodes_block(nm, get_new_node(get_nodes_block(om)));
-}
-
-/**
- * Copies the graph reachable from current_ir_graph->end to the obstack
- * in current_ir_graph and fixes the environment.
- * Then fixes the fields in current_ir_graph containing nodes of the
- * graph.
- *
- * @param copy_node_nr If non-zero, the node number will be copied
- */
-static void
-copy_graph_env(int copy_node_nr) {
- ir_graph *irg = current_ir_graph;
- ir_node *old_end, *new_anchor;
- int i;
-
- /* remove end_except and end_reg nodes */
- old_end = get_irg_end(irg);
- set_irg_end_except (irg, old_end);
- set_irg_end_reg (irg, old_end);
-
- /* Not all nodes remembered in irg might be reachable
- from the end node. Assure their link is set to NULL, so that
- we can test whether new nodes have been computed. */
- for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
- ir_node *n = get_irg_anchor(irg, i);
- if (n != NULL)
- set_new_node(n, NULL);
- }
- /* we use the block walk flag for removing Bads from Blocks ins. */
- inc_irg_block_visited(irg);
-
- /* copy the graph */
- copy_graph(irg, copy_node_nr);
-
- /* fix the anchor */
- old_end = get_irg_end(irg);
- new_anchor = new_Anchor(irg);
-
- for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
- ir_node *n = get_irg_anchor(irg, i);
- if (n)
- set_irn_n(new_anchor, i, get_new_node(n));
- }
- free_End(old_end);
- irg->anchor = new_anchor;
-
- /* ensure the new anchor is placed in the endblock */
- set_nodes_block(new_anchor, get_irg_end_block(irg));
-}
-
-/**
- * Copies all reachable nodes to a new obstack. Removes bad inputs
- * from block nodes and the corresponding inputs from Phi nodes.
- * Merges single exit blocks with single entry blocks and removes
- * 1-input Phis.
- * Adds all new nodes to a new hash table for CSE. Does not
- * perform CSE, so the hash table might contain common subexpressions.
- */
-void dead_node_elimination(ir_graph *irg) {
- ir_graph *rem;
-#ifdef INTERPROCEDURAL_VIEW
- int rem_ipview = get_interprocedural_view();
-#endif
- struct obstack *graveyard_obst = NULL;
- struct obstack *rebirth_obst = NULL;
-
- edges_deactivate(irg);
-
- /* inform statistics that we started a dead-node elimination run */
- hook_dead_node_elim(irg, 1);
-
- /* Remember external state of current_ir_graph. */
- rem = current_ir_graph;
- current_ir_graph = irg;
-#ifdef INTERPROCEDURAL_VIEW
- set_interprocedural_view(0);
-#endif
-
- assert(get_irg_phase_state(irg) != phase_building);
-
- /* Handle graph state */
- free_callee_info(irg);
- free_irg_outs(irg);
- free_trouts();
-
- /* @@@ so far we loose loops when copying */
- free_loop_information(irg);
-
- set_irg_doms_inconsistent(irg);
-
- /* A quiet place, where the old obstack can rest in peace,
- until it will be cremated. */
- graveyard_obst = irg->obst;
-
- /* A new obstack, where the reachable nodes will be copied to. */
- rebirth_obst = XMALLOC(struct obstack);
- irg->obst = rebirth_obst;
- obstack_init(irg->obst);
- irg->last_node_idx = 0;
-
- /* We also need a new value table for CSE */
- del_identities(irg->value_table);
- irg->value_table = new_identities();
-
- /* Copy the graph from the old to the new obstack */
- copy_graph_env(/*copy_node_nr=*/1);
-
- /* Free memory from old unoptimized obstack */
- obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
- xfree(graveyard_obst); /* ... then free it. */
-
- /* inform statistics that the run is over */
- hook_dead_node_elim(irg, 0);
-
- current_ir_graph = rem;
-#ifdef INTERPROCEDURAL_VIEW
- set_interprocedural_view(rem_ipview);
-#endif
-}
-
-/**
- * Relink bad predecessors of a block and store the old in array to the
- * link field. This function is called by relink_bad_predecessors().
- * The array of link field starts with the block operand at position 0.
- * If block has bad predecessors, create a new in array without bad preds.
- * Otherwise let in array untouched.
- */
-static void relink_bad_block_predecessors(ir_node *n, void *env) {
- ir_node **new_in, *irn;
- int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
- (void) env;
-
- /* if link field of block is NULL, look for bad predecessors otherwise
- this is already done */
- if (is_Block(n) && get_irn_link(n) == NULL) {
- /* save old predecessors in link field (position 0 is the block operand)*/
- set_irn_link(n, get_irn_in(n));
-
- /* count predecessors without bad nodes */
- old_irn_arity = get_irn_arity(n);
- for (i = 0; i < old_irn_arity; i++)
- if (!is_Bad(get_irn_n(n, i)))
- ++new_irn_arity;
-
- /* arity changing: set new predecessors without bad nodes */
- if (new_irn_arity < old_irn_arity) {
- /* Get new predecessor array. We do not resize the array, as we must
- keep the old one to update Phis. */
- new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
-
- /* set new predecessors in array */
- new_in[0] = NULL;
- new_irn_n = 1;
- for (i = 0; i < old_irn_arity; i++) {
- irn = get_irn_n(n, i);
- if (!is_Bad(irn)) {
- new_in[new_irn_n] = irn;
- is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
- ++new_irn_n;
- }
- }
- /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
- ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
- n->in = new_in;
- } /* ir node has bad predecessors */
- } /* Block is not relinked */
-}
-
-/**
- * Relinks Bad predecessors from Blocks and Phis called by walker
- * remove_bad_predecesors(). If n is a Block, call
- * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
- * function of Phi's Block. If this block has bad predecessors, relink preds
- * of the Phi-node.
- */
-static void relink_bad_predecessors(ir_node *n, void *env) {
- ir_node *block, **old_in;
- int i, old_irn_arity, new_irn_arity;
-
- /* relink bad predecessors of a block */
- if (is_Block(n))
- relink_bad_block_predecessors(n, env);
-
- /* If Phi node relink its block and its predecessors */
- if (is_Phi(n)) {
- /* Relink predecessors of phi's block */
- block = get_nodes_block(n);
- if (get_irn_link(block) == NULL)
- relink_bad_block_predecessors(block, env);
-
- old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
- old_irn_arity = ARR_LEN(old_in);
-
- /* Relink Phi predecessors if count of predecessors changed */
- if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
- /* set new predecessors in array
- n->in[0] remains the same block */
- new_irn_arity = 1;
- for(i = 1; i < old_irn_arity; i++)
- if (!is_Bad(old_in[i])) {
- n->in[new_irn_arity] = n->in[i];
- is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
- ++new_irn_arity;
- }
-
- ARR_SETLEN(ir_node *, n->in, new_irn_arity);
- ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
- }
- } /* n is a Phi node */
-}
-
-/*
- * Removes Bad Bad predecessors from Blocks and the corresponding
- * inputs to Phi nodes as in dead_node_elimination but without
- * copying the graph.
- * On walking up set the link field to NULL, on walking down call
- * relink_bad_predecessors() (This function stores the old in array
- * to the link field and sets a new in array if arity of predecessors
- * changes).
- */
-void remove_bad_predecessors(ir_graph *irg) {
- panic("Fix backedge handling first");
- irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
-}
-
-
-/*
- __ _ __ __
- (_ __ o _ | \/ |_
- __)|_| | \_/ | \_/(/_ |_/\__|__
-
- The following stuff implements a facility that automatically patches
- registered ir_node pointers to the new node when a dead node elimination occurs.
-*/
-
-struct _survive_dce_t {
- struct obstack obst;
- pmap *places;
- pmap *new_places;
- hook_entry_t dead_node_elim;
- hook_entry_t dead_node_elim_subst;
-};
-
-typedef struct _survive_dce_list_t {
- struct _survive_dce_list_t *next;
- ir_node **place;
-} survive_dce_list_t;
-
-static void dead_node_hook(void *context, ir_graph *irg, int start) {
- survive_dce_t *sd = context;
- (void) irg;
-
- /* Create a new map before the dead node elimination is performed. */
- if (start) {
- sd->new_places = pmap_create_ex(pmap_count(sd->places));
- } else {
- /* Patch back all nodes if dead node elimination is over and something is to be done. */
- pmap_destroy(sd->places);
- sd->places = sd->new_places;
- sd->new_places = NULL;
- }
-}
-
-/**
- * Hook called when dead node elimination replaces old by nw.
- */
-static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
- survive_dce_t *sd = context;
- survive_dce_list_t *list = pmap_get(sd->places, old);
- (void) irg;
-
- /* If the node is to be patched back, write the new address to all registered locations. */
- if (list) {
- survive_dce_list_t *p;
-
- for (p = list; p; p = p->next)
- *(p->place) = nw;
-
- pmap_insert(sd->new_places, nw, list);
- }
-}
-
-/**
- * Make a new Survive DCE environment.
- */
-survive_dce_t *new_survive_dce(void) {
- survive_dce_t *res = XMALLOC(survive_dce_t);
- obstack_init(&res->obst);
- res->places = pmap_create();
- res->new_places = NULL;
-
- res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
- res->dead_node_elim.context = res;
- res->dead_node_elim.next = NULL;
-
- res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
- res->dead_node_elim_subst.context = res;
- res->dead_node_elim_subst.next = NULL;
-
- register_hook(hook_dead_node_elim, &res->dead_node_elim);
- register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
- return res;
-}
-
-/**
- * Free a Survive DCE environment.
- */
-void free_survive_dce(survive_dce_t *sd) {
- obstack_free(&sd->obst, NULL);
- pmap_destroy(sd->places);
- unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
- unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
- xfree(sd);
+static void set_new_node(ir_node *node, ir_node *new_node)
+{
+ set_irn_link(node, new_node);
}
/**
- * Register a node pointer to be patched upon DCE.
- * When DCE occurs, the node pointer specified by @p place will be
- * patched to the new address of the node it is pointing to.
- *
- * @param sd The Survive DCE environment.
- * @param place The address of the node pointer.
+ * Get this new node, before the old node is forgotten.
*/
-void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
- if (*place != NULL) {
- ir_node *irn = *place;
- survive_dce_list_t *curr = pmap_get(sd->places, irn);
- survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
-
- nw->next = curr;
- nw->place = place;
-
- pmap_insert(sd->places, irn, nw);
- }
+static inline ir_node *get_new_node(ir_node *old_node)
+{
+ assert(irn_visited(old_node));
+ return (ir_node*) get_irn_link(old_node);
}
/*--------------------------------------------------------------------*/
* inlined procedure. The new entities must be in the link field of
* the entities.
*/
-static void copy_node_inline(ir_node *n, void *env) {
- ir_node *nn;
- ir_type *frame_tp = (ir_type *)env;
-
- copy_node(n, NULL);
- if (is_Sel(n)) {
- nn = get_new_node (n);
- assert(is_Sel(nn));
- if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
- set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
+static void copy_node_inline(ir_node *node, void *env)
+{
+ ir_graph *new_irg = (ir_graph*) env;
+ ir_node *new_node = irn_copy_into_irg(node, new_irg);
+
+ set_new_node(node, new_node);
+ if (is_Sel(node)) {
+ ir_graph *old_irg = get_irn_irg(node);
+ ir_type *old_frame_type = get_irg_frame_type(old_irg);
+ ir_entity *old_entity = get_Sel_entity(node);
+ assert(is_Sel(new_node));
+ /* use copied entities from the new frame */
+ if (get_entity_owner(old_entity) == old_frame_type) {
+ ir_entity *new_entity = (ir_entity*)get_entity_link(old_entity);
+ assert(new_entity != NULL);
+ set_Sel_entity(new_node, new_entity);
}
- } else if (is_Block(n)) {
- nn = get_new_node (n);
- nn->attr.block.irg = current_ir_graph;
+ } else if (is_Block(new_node)) {
+ new_node->attr.block.irg.irg = new_irg;
}
}
-/**
- * Copies new predecessors of old node and move constants to
- * the Start Block.
- */
-static void copy_preds_inline(ir_node *n, void *env) {
- ir_node *nn;
-
- copy_preds(n, env);
- nn = skip_Id(get_new_node(n));
- if (is_irn_constlike(nn)) {
- /* move Constants into the start block */
- set_nodes_block(nn, get_irg_start_block(current_ir_graph));
-
- n = identify_remember(current_ir_graph->value_table, nn);
- if (nn != n) {
- DBG_OPT_CSE(nn, n);
- exchange(nn, n);
- }
+static void set_preds_inline(ir_node *node, void *env)
+{
+ ir_node *new_node;
+
+ irn_rewire_inputs(node);
+
+ /* move constants into start block */
+ new_node = get_new_node(node);
+ if (is_irn_constlike(new_node)) {
+ ir_graph *new_irg = (ir_graph *) env;
+ ir_node *start_block = get_irg_start_block(new_irg);
+ set_nodes_block(new_node, start_block);
}
}
/**
* Walker: checks if P_value_arg_base is used.
*/
-static void find_addr(ir_node *node, void *env) {
- int *allow_inline = env;
- if (is_Sel(node)) {
+static void find_addr(ir_node *node, void *env)
+{
+ bool *allow_inline = (bool*)env;
+
+ if (is_Block(node) && get_Block_entity(node)) {
+ /**
+ * Currently we can't handle blocks whose address was taken correctly
+ * when inlining
+ */
+ *allow_inline = false;
+ } else if (is_Sel(node)) {
ir_graph *irg = current_ir_graph;
if (get_Sel_ptr(node) == get_irg_frame(irg)) {
/* access to frame */
ir_entity *ent = get_Sel_entity(node);
if (get_entity_owner(ent) != get_irg_frame_type(irg)) {
/* access to value_type */
- *allow_inline = 0;
+ *allow_inline = false;
+ }
+ if (is_parameter_entity(ent)) {
+ *allow_inline = false;
}
}
} else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
* into schedule_block cause it to require 2GB of ram instead of 256MB.
*
* Sorrily this is true with our implementation also.
- * Moreover, we cannot differentiate between alloca() and VLA yet, so this
- * disables inlining of functions using VLA (with are completely save).
+ * Moreover, we cannot differentiate between alloca() and VLA yet, so
+ * this disables inlining of functions using VLA (which are completely
+ * save).
*
* 2 Solutions:
* - add a flag to the Alloc node for "real" alloca() calls
- * - add a new Stack-Restore node at the end of a function using alloca()
+ * - add a new Stack-Restore node at the end of a function using
+ * alloca()
*/
- *allow_inline = 0;
+ *allow_inline = false;
}
}
*
* check these conditions here
*/
-static int can_inline(ir_node *call, ir_graph *called_graph) {
- ir_type *call_type = get_Call_type(call);
- int params, ress, i, res;
- assert(is_Method_type(call_type));
-
- params = get_method_n_params(call_type);
- ress = get_method_n_ress(call_type);
-
- /* check parameters for compound arguments */
- for (i = 0; i < params; ++i) {
- ir_type *p_type = get_method_param_type(call_type, i);
-
- if (is_compound_type(p_type))
- return 0;
- }
-
- /* check results for compound arguments */
- for (i = 0; i < ress; ++i) {
- ir_type *r_type = get_method_res_type(call_type, i);
-
- if (is_compound_type(r_type))
- return 0;
- }
-
- res = 1;
- irg_walk_graph(called_graph, find_addr, NULL, &res);
-
- return res;
-}
-
-enum exc_mode {
- exc_handler = 0, /**< There is a handler. */
- exc_to_end = 1, /**< Branches to End. */
- exc_no_handler = 2 /**< Exception handling not represented. */
-};
-
-/* Inlines a method at the given call site. */
-int inline_method(ir_node *call, ir_graph *called_graph) {
- ir_node *pre_call;
- ir_node *post_call, *post_bl;
- ir_node *in[pn_Start_max];
- ir_node *end, *end_bl, *block;
- ir_node **res_pred;
- ir_node **cf_pred;
- ir_node **args_in;
- ir_node *ret, *phi;
- int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
- enum exc_mode exc_handling;
- ir_type *called_frame, *curr_frame, *mtp, *ctp;
- ir_entity *ent;
- ir_graph *rem, *irg;
- irg_inline_property prop = get_irg_inline_property(called_graph);
- unsigned long visited;
+static bool can_inline(ir_node *call, ir_graph *called_graph)
+{
+ ir_entity *called = get_irg_entity(called_graph);
+ ir_type *called_type = get_entity_type(called);
+ ir_type *call_type = get_Call_type(call);
+ size_t n_params = get_method_n_params(called_type);
+ size_t n_arguments = get_method_n_params(call_type);
+ size_t n_res = get_method_n_ress(called_type);
+ irg_inline_property prop = get_irg_inline_property(called_graph);
+ size_t i;
+ bool res;
if (prop == irg_inline_forbidden)
- return 0;
-
- ent = get_irg_entity(called_graph);
+ return false;
- mtp = get_entity_type(ent);
- ctp = get_Call_type(call);
- if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
+ if (n_arguments != n_params) {
/* this is a bad feature of C: without a prototype, we can
* call a function with less parameters than needed. Currently
* we don't support this, although we could use Unknown than. */
- return 0;
+ return false;
+ }
+ if (n_res != get_method_n_ress(call_type)) {
+ return false;
}
/* Argh, compiling C has some bad consequences:
* It is implementation dependent what happens in that case.
* We support inlining, if the bitsize of the types matches AND
* the same arithmetic is used. */
- n_params = get_method_n_params(mtp);
- for (i = n_params - 1; i >= 0; --i) {
- ir_type *param_tp = get_method_param_type(mtp, i);
- ir_type *arg_tp = get_method_param_type(ctp, i);
+ for (i = 0; i < n_params; ++i) {
+ ir_type *param_tp = get_method_param_type(called_type, i);
+ ir_type *arg_tp = get_method_param_type(call_type, i);
if (param_tp != arg_tp) {
ir_mode *pmode = get_type_mode(param_tp);
ir_mode *amode = get_type_mode(arg_tp);
if (pmode == NULL || amode == NULL)
- return 0;
+ return false;
if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
- return 0;
+ return false;
if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
- return 0;
+ return false;
/* otherwise we can simply "reinterpret" the bits */
}
}
+ for (i = 0; i < n_res; ++i) {
+ ir_type *decl_res_tp = get_method_res_type(called_type, i);
+ ir_type *used_res_tp = get_method_res_type(call_type, i);
+
+ if (decl_res_tp != used_res_tp) {
+ ir_mode *decl_mode = get_type_mode(decl_res_tp);
+ ir_mode *used_mode = get_type_mode(used_res_tp);
+ if (decl_mode == NULL || used_mode == NULL)
+ return false;
+ if (get_mode_size_bits(decl_mode) != get_mode_size_bits(used_mode))
+ return false;
+ if (get_mode_arithmetic(decl_mode) != get_mode_arithmetic(used_mode))
+ return false;
+ /* otherwise we can "reinterpret" the bits */
+ }
+ }
- irg = get_irn_irg(call);
+ /* check parameters for compound arguments */
+ for (i = 0; i < n_params; ++i) {
+ ir_type *p_type = get_method_param_type(call_type, i);
- /*
- * We cannot inline a recursive call. The graph must be copied before
- * the call the inline_method() using create_irg_copy().
- */
- if (called_graph == irg)
- return 0;
+ if (is_compound_type(p_type))
+ return false;
+ }
- /*
- * currently, we cannot inline two cases:
- * - call with compound arguments
- * - graphs that take the address of a parameter
- */
+ /* check results for compound arguments */
+ for (i = 0; i < n_res; ++i) {
+ ir_type *r_type = get_method_res_type(call_type, i);
+
+ if (is_compound_type(r_type))
+ return false;
+ }
+
+ res = true;
+ irg_walk_graph(called_graph, find_addr, NULL, &res);
+
+ return res;
+}
+
+enum exc_mode {
+ exc_handler, /**< There is a handler. */
+ exc_no_handler /**< Exception handling not represented. */
+};
+
+/**
+ * copy all entities on the stack frame on 1 irg to the stackframe of another.
+ * Sets entity links of the old entities to the copies
+ */
+static void copy_frame_entities(ir_graph *from, ir_graph *to)
+{
+ ir_type *from_frame = get_irg_frame_type(from);
+ ir_type *to_frame = get_irg_frame_type(to);
+ size_t n_members = get_class_n_members(from_frame);
+ size_t i;
+ assert(from_frame != to_frame);
+
+ for (i = 0; i < n_members; ++i) {
+ ir_entity *old_ent = get_class_member(from_frame, i);
+ ir_entity *new_ent = copy_entity_own(old_ent, to_frame);
+ set_entity_link(old_ent, new_ent);
+ assert (!is_parameter_entity(old_ent));
+ }
+}
+
+/* Inlines a method at the given call site. */
+int inline_method(ir_node *call, ir_graph *called_graph)
+{
+ ir_node *pre_call;
+ ir_node *post_call, *post_bl;
+ ir_node *in[pn_Start_max+1];
+ ir_node *end, *end_bl, *block;
+ ir_node **res_pred;
+ ir_node **cf_pred;
+ ir_node **args_in;
+ ir_node *ret, *phi;
+ int arity, n_ret, n_exc, n_res, i, j, rem_opt;
+ int irn_arity, n_params;
+ int n_mem_phi;
+ enum exc_mode exc_handling;
+ ir_type *mtp;
+ ir_type *ctp;
+ ir_entity *ent;
+ ir_graph *rem;
+ ir_graph *irg = get_irn_irg(call);
+
+ /* we cannot inline some types of calls */
if (! can_inline(call, called_graph))
return 0;
+ /* We cannot inline a recursive call. The graph must be copied before
+ * the call the inline_method() using create_irg_copy(). */
+ if (called_graph == irg)
+ return 0;
+
+ ent = get_irg_entity(called_graph);
+ mtp = get_entity_type(ent);
+ ctp = get_Call_type(call);
+ n_params = get_method_n_params(mtp);
+ n_res = get_method_n_ress(mtp);
+
rem = current_ir_graph;
current_ir_graph = irg;
DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
- /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
+ /* optimizations can cause problems when allocating new nodes */
rem_opt = get_opt_optimize();
set_optimize(0);
assert(get_irg_phase_state(irg) != phase_building);
assert(get_irg_pinned(irg) == op_pin_state_pinned);
assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
- set_irg_outs_inconsistent(irg);
- set_irg_extblk_inconsistent(irg);
- set_irg_doms_inconsistent(irg);
- set_irg_loopinfo_inconsistent(irg);
+ clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
+ | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
- set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
-
- /* -- Check preconditions -- */
- assert(is_Call(call));
+ clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
+ edges_deactivate(irg);
/* here we know we WILL inline, so inform the statistics */
hook_inline(call, called_graph);
for the Call node, or do we branch directly to End on an exception?
exc_handling:
0 There is a handler.
- 1 Branches to End.
2 Exception handling not represented in Firm. -- */
{
- ir_node *proj, *Mproj = NULL, *Xproj = NULL;
- for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
+ ir_node *Xproj = NULL;
+ ir_node *proj;
+ for (proj = (ir_node*)get_irn_link(call); proj != NULL;
+ proj = (ir_node*)get_irn_link(proj)) {
long proj_nr = get_Proj_proj(proj);
if (proj_nr == pn_Call_X_except) Xproj = proj;
- if (proj_nr == pn_Call_M_except) Mproj = proj;
}
- if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
- else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
- else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
+ exc_handling = Xproj != NULL ? exc_handler : exc_no_handler;
}
/* create the argument tuple */
- NEW_ARR_A(ir_type *, args_in, n_params);
+ args_in = ALLOCAN(ir_node*, n_params);
block = get_nodes_block(call);
for (i = n_params - 1; i >= 0; --i) {
ir_mode *mode = get_type_mode(param_tp);
if (mode != get_irn_mode(arg)) {
- arg = new_r_Conv(irg, block, arg, mode, 0);
+ arg = new_r_Conv(block, arg, mode);
}
args_in[i] = arg;
}
- /* --
- the procedure and later replaces the Start node of the called graph.
- Post_call is the old Call node and collects the results of the called
- graph. Both will end up being a tuple. -- */
+ /* the procedure and later replaces the Start node of the called graph.
+ * Post_call is the old Call node and collects the results of the called
+ * graph. Both will end up being a tuple. */
post_bl = get_nodes_block(call);
- set_irg_current_block(irg, post_bl);
/* XxMxPxPxPxT of Start + parameter of Call */
- in[pn_Start_X_initial_exec] = new_Jmp();
- in[pn_Start_M] = get_Call_mem(call);
- in[pn_Start_P_frame_base] = get_irg_frame(irg);
- in[pn_Start_P_tls] = get_irg_tls(irg);
- in[pn_Start_T_args] = new_Tuple(n_params, args_in);
- pre_call = new_Tuple(pn_Start_max, in);
+ in[pn_Start_M] = get_Call_mem(call);
+ in[pn_Start_X_initial_exec] = new_r_Jmp(post_bl);
+ in[pn_Start_P_frame_base] = get_irg_frame(irg);
+ in[pn_Start_T_args] = new_r_Tuple(post_bl, n_params, args_in);
+ pre_call = new_r_Tuple(post_bl, pn_Start_max+1, in);
post_call = call;
/* --
predecessors and all Phi nodes. -- */
part_block(pre_call);
- /* -- Prepare state for dead node elimination -- */
- /* Visited flags in calling irg must be >= flag in called irg.
- Else walker and arity computation will not work. */
- if (get_irg_visited(irg) <= get_irg_visited(called_graph))
- set_irg_visited(irg, get_irg_visited(called_graph) + 1);
- if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
- set_irg_block_visited(irg, get_irg_block_visited(called_graph));
- visited = get_irg_visited(irg);
-
- /* Set pre_call as new Start node in link field of the start node of
- calling graph and pre_calls block as new block for the start block
- of calling graph.
- Further mark these nodes so that they are not visited by the
- copying. */
- set_irn_link(get_irg_start(called_graph), pre_call);
- set_irn_visited(get_irg_start(called_graph), visited);
- set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
- set_irn_visited(get_irg_start_block(called_graph), visited);
-
- set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
- set_irn_visited(get_irg_bad(called_graph), visited);
-
- set_irn_link(get_irg_no_mem(called_graph), get_irg_no_mem(current_ir_graph));
- set_irn_visited(get_irg_no_mem(called_graph), visited);
-
- /* Initialize for compaction of in arrays */
- inc_irg_block_visited(irg);
-
- /* -- Replicate local entities of the called_graph -- */
- /* copy the entities. */
- irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
- called_frame = get_irg_frame_type(called_graph);
- curr_frame = get_irg_frame_type(irg);
- for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
- ir_entity *new_ent, *old_ent;
- old_ent = get_class_member(called_frame, i);
- new_ent = copy_entity_own(old_ent, curr_frame);
- set_entity_link(old_ent, new_ent);
- }
+ /* increment visited flag for later walk */
+ inc_irg_visited(called_graph);
+
+ /* link some nodes to nodes in the current graph so instead of copying
+ * the linked nodes will get used.
+ * So the copier will use the created Tuple instead of copying the start
+ * node, similar for singleton nodes like NoMem and Bad.
+ * Note: this will prohibit predecessors to be copied - only do it for
+ * nodes without predecessors */
+ {
+ ir_node *start_block;
+ ir_node *start;
+ ir_node *nomem;
+
+ start_block = get_irg_start_block(called_graph);
+ set_new_node(start_block, get_nodes_block(pre_call));
+ mark_irn_visited(start_block);
- /* visited is > than that of called graph. With this trick visited will
- remain unchanged so that an outer walker, e.g., searching the call nodes
- to inline, calling this inline will not visit the inlined nodes. */
- set_irg_visited(irg, get_irg_visited(irg)-1);
+ start = get_irg_start(called_graph);
+ set_new_node(start, pre_call);
+ mark_irn_visited(start);
- /* -- Performing dead node elimination inlines the graph -- */
- /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
- entities. */
- irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
- get_irg_frame_type(called_graph));
+ nomem = get_irg_no_mem(called_graph);
+ set_new_node(nomem, get_irg_no_mem(irg));
+ mark_irn_visited(nomem);
+ }
+
+ /* entitiy link is used to link entities on old stackframe to the
+ * new stackframe */
+ irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
- irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
+ /* copy entities and nodes */
+ assert(!irn_visited(get_irg_end(called_graph)));
+ copy_frame_entities(called_graph, irg);
+ irg_walk_core(get_irg_end(called_graph), copy_node_inline, set_preds_inline,
+ irg);
- /* Repair called_graph */
- set_irg_visited(called_graph, get_irg_visited(irg));
- set_irg_block_visited(called_graph, get_irg_block_visited(irg));
- set_Block_block_visited(get_irg_start_block(called_graph), 0);
+ irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
/* -- Merge the end of the inlined procedure with the call site -- */
/* We will turn the old Call node into a Tuple with the following
memories.
*/
- /* -- Precompute some values -- */
+ /* Precompute some values */
end_bl = get_new_node(get_irg_end_block(called_graph));
- end = get_new_node(get_irg_end(called_graph));
- arity = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret */
- n_res = get_method_n_ress(get_Call_type(call));
+ end = get_new_node(get_irg_end(called_graph));
+ arity = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret */
+ n_res = get_method_n_ress(get_Call_type(call));
res_pred = XMALLOCN(ir_node*, n_res);
cf_pred = XMALLOCN(ir_node*, arity);
- set_irg_current_block(irg, post_bl); /* just to make sure */
-
- /* -- archive keepalives -- */
+ /* archive keepalives */
irn_arity = get_irn_arity(end);
for (i = 0; i < irn_arity; i++) {
ir_node *ka = get_End_keepalive(end, i);
add_End_keepalive(get_irg_end(irg), ka);
}
- /* The new end node will die. We need not free as the in array is on the obstack:
- copy_node() only generated 'D' arrays. */
-
- /* -- Replace Return nodes by Jump nodes. -- */
+ /* replace Return nodes by Jump nodes */
n_ret = 0;
for (i = 0; i < arity; i++) {
ir_node *ret;
ret = get_Block_cfgpred(end_bl, i);
if (is_Return(ret)) {
- cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
+ ir_node *block = get_nodes_block(ret);
+ cf_pred[n_ret] = new_r_Jmp(block);
n_ret++;
}
}
set_irn_in(post_bl, n_ret, cf_pred);
- /* -- Build a Tuple for all results of the method.
- Add Phi node if there was more than one Return. -- */
- turn_into_tuple(post_call, pn_Call_max);
+ /* build a Tuple for all results of the method.
+ * add Phi node if there was more than one Return. */
+ turn_into_tuple(post_call, pn_Call_max+1);
/* First the Memory-Phi */
- n_ret = 0;
+ n_mem_phi = 0;
for (i = 0; i < arity; i++) {
ret = get_Block_cfgpred(end_bl, i);
if (is_Return(ret)) {
- cf_pred[n_ret] = get_Return_mem(ret);
- n_ret++;
+ cf_pred[n_mem_phi++] = get_Return_mem(ret);
+ }
+ /* memory output for some exceptions is directly connected to End */
+ if (is_Call(ret)) {
+ cf_pred[n_mem_phi++] = new_r_Proj(ret, mode_M, 3);
+ } else if (is_fragile_op(ret)) {
+ /* We rely that all cfops have the memory output at the same position. */
+ cf_pred[n_mem_phi++] = new_r_Proj(ret, mode_M, 0);
+ } else if (is_Raise(ret)) {
+ cf_pred[n_mem_phi++] = new_r_Proj(ret, mode_M, 1);
}
}
- phi = new_Phi(n_ret, cf_pred, mode_M);
- set_Tuple_pred(call, pn_Call_M_regular, phi);
+ phi = new_r_Phi(post_bl, n_mem_phi, cf_pred, mode_M);
+ set_Tuple_pred(call, pn_Call_M, phi);
/* Conserve Phi-list for further inlinings -- but might be optimized */
if (get_nodes_block(phi) == post_bl) {
set_irn_link(phi, get_irn_link(post_bl));
}
/* Now the real results */
if (n_res > 0) {
+ ir_node *result_tuple;
for (j = 0; j < n_res; j++) {
+ ir_type *res_type = get_method_res_type(ctp, j);
+ ir_mode *res_mode = get_type_mode(res_type);
n_ret = 0;
for (i = 0; i < arity; i++) {
ret = get_Block_cfgpred(end_bl, i);
if (is_Return(ret)) {
- cf_pred[n_ret] = get_Return_res(ret, j);
+ ir_node *res = get_Return_res(ret, j);
+ if (get_irn_mode(res) != res_mode) {
+ ir_node *block = get_nodes_block(res);
+ res = new_r_Conv(block, res, res_mode);
+ }
+ cf_pred[n_ret] = res;
n_ret++;
}
}
- if (n_ret > 0)
- phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
- else
- phi = new_Bad();
+ if (n_ret > 0) {
+ phi = new_r_Phi(post_bl, n_ret, cf_pred, res_mode);
+ } else {
+ phi = new_r_Bad(irg, res_mode);
+ }
res_pred[j] = phi;
/* Conserve Phi-list for further inlinings -- but might be optimized */
if (get_nodes_block(phi) == post_bl) {
set_Block_phis(post_bl, phi);
}
}
- set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
+ result_tuple = new_r_Tuple(post_bl, n_res, res_pred);
+ set_Tuple_pred(call, pn_Call_T_result, result_tuple);
} else {
- set_Tuple_pred(call, pn_Call_T_result, new_Bad());
+ set_Tuple_pred(call, pn_Call_T_result, new_r_Bad(irg, mode_T));
}
/* handle the regular call */
- set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
-
- /* For now, we cannot inline calls with value_base */
- set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
+ set_Tuple_pred(call, pn_Call_X_regular, new_r_Jmp(post_bl));
/* Finally the exception control flow.
- We have two (three) possible situations:
- First if the Call branches to an exception handler: We need to add a Phi node to
+ We have two possible situations:
+ First if the Call branches to an exception handler:
+ We need to add a Phi node to
collect the memory containing the exception objects. Further we need
to add another block to get a correct representation of this Phi. To
this block we add a Jmp that resolves into the X output of the Call
when the Call is turned into a tuple.
- Second the Call branches to End, the exception is not handled. Just
- add all inlined exception branches to the End node.
- Third: there is no Exception edge at all. Handle as case two. */
+ Second: There is no exception edge. Just add all inlined exception
+ branches to the End node.
+ */
if (exc_handling == exc_handler) {
n_exc = 0;
for (i = 0; i < arity; i++) {
}
}
if (n_exc > 0) {
- ir_node *block = new_Block(n_exc, cf_pred);
- set_cur_block(block);
-
- set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
- /* The Phi for the memories with the exception objects */
- n_exc = 0;
- for (i = 0; i < arity; i++) {
- ir_node *ret;
- ret = skip_Proj(get_Block_cfgpred(end_bl, i));
- if (is_Call(ret)) {
- cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
- n_exc++;
- } else if (is_fragile_op(ret)) {
- /* We rely that all cfops have the memory output at the same position. */
- cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
- n_exc++;
- } else if (is_Raise(ret)) {
- cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
- n_exc++;
- }
+ if (n_exc == 1) {
+ /* simple fix */
+ set_Tuple_pred(call, pn_Call_X_except, cf_pred[0]);
+ } else {
+ ir_node *block = new_r_Block(irg, n_exc, cf_pred);
+ set_Tuple_pred(call, pn_Call_X_except, new_r_Jmp(block));
}
- set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
} else {
- set_Tuple_pred(call, pn_Call_X_except, new_Bad());
- set_Tuple_pred(call, pn_Call_M_except, new_Bad());
+ set_Tuple_pred(call, pn_Call_X_except, new_r_Bad(irg, mode_X));
}
} else {
ir_node *main_end_bl;
for (i = 0; i < n_exc; ++i)
end_preds[main_end_bl_arity + i] = cf_pred[i];
set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
- set_Tuple_pred(call, pn_Call_X_except, new_Bad());
- set_Tuple_pred(call, pn_Call_M_except, new_Bad());
+ set_Tuple_pred(call, pn_Call_X_except, new_r_Bad(irg, mode_X));
free(end_preds);
}
free(res_pred);
static struct obstack temp_obst;
/** Represents a possible inlinable call in a graph. */
-typedef struct _call_entry {
+typedef struct call_entry {
ir_node *call; /**< The Call node. */
ir_graph *callee; /**< The callee IR-graph. */
list_head list; /**< List head for linking the next one. */
/**
* environment for inlining small irgs
*/
-typedef struct _inline_env_t {
+typedef struct inline_env_t {
struct obstack obst; /**< An obstack where call_entries are allocated on. */
list_head calls; /**< The call entry list. */
} inline_env_t;
*
* @param call the call node
*/
-static ir_graph *get_call_called_irg(ir_node *call) {
+static ir_graph *get_call_called_irg(ir_node *call)
+{
ir_node *addr;
addr = get_Call_ptr(call);
- if (is_Global(addr)) {
- ir_entity *ent = get_Global_entity(addr);
+ if (is_SymConst_addr_ent(addr)) {
+ ir_entity *ent = get_SymConst_entity(addr);
+ /* we don't know which function gets finally bound to a weak symbol */
+ if (get_entity_linkage(ent) & IR_LINKAGE_WEAK)
+ return NULL;
+
return get_entity_irg(ent);
}
/**
* Walker: Collect all calls to known graphs inside a graph.
*/
-static void collect_calls(ir_node *call, void *env) {
+static void collect_calls(ir_node *call, void *env)
+{
(void) env;
if (is_Call(call)) {
ir_graph *called_irg = get_call_called_irg(call);
if (called_irg != NULL) {
/* 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));
+ inline_env_t *ienv = (inline_env_t*)env;
+ call_entry *entry = OALLOC(&ienv->obst, call_entry);
entry->call = call;
entry->callee = called_irg;
entry->loop_depth = 0;
* Methods where the obstack containing the firm graph is smaller than
* size are inlined.
*/
-void inline_small_irgs(ir_graph *irg, int size) {
+void inline_small_irgs(ir_graph *irg, int size)
+{
ir_graph *rem = current_ir_graph;
inline_env_t env;
- call_entry *entry;
current_ir_graph = irg;
/* Handle graph state */
ir_graph *callee = entry->callee;
irg_inline_property prop = get_irg_inline_property(callee);
- if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
- /* do not inline forbidden / weak graphs */
+ if (prop == irg_inline_forbidden) {
continue;
}
current_ir_graph = rem;
}
+typedef struct inline_small_irgs_pass_t {
+ ir_graph_pass_t pass;
+ int size;
+} inline_small_irgs_pass_t;
+
+/**
+ * Wrapper to run inline_small_irgs() as a pass.
+ */
+static int inline_small_irgs_wrapper(ir_graph *irg, void *context)
+{
+ inline_small_irgs_pass_t *pass = (inline_small_irgs_pass_t*)context;
+
+ inline_small_irgs(irg, pass->size);
+ return 0;
+}
+
+/* create a pass for inline_small_irgs() */
+ir_graph_pass_t *inline_small_irgs_pass(const char *name, int size)
+{
+ inline_small_irgs_pass_t *pass = XMALLOCZ(inline_small_irgs_pass_t);
+
+ pass->size = size;
+ return def_graph_pass_constructor(
+ &pass->pass, name ? name : "inline_small_irgs", inline_small_irgs_wrapper);
+}
+
/**
* Environment for inlining irgs.
*/
unsigned n_callers; /**< Number of known graphs that call this graphs. */
unsigned n_callers_orig; /**< for statistics */
unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
- unsigned local_vars:1; /**< Set, if an inlined function got the address of a local variable. */
unsigned recursive:1; /**< Set, if this function is self recursive. */
} inline_irg_env;
/**
* Allocate a new environment for inlining.
*/
-static inline_irg_env *alloc_inline_irg_env(void) {
- inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
+static inline_irg_env *alloc_inline_irg_env(void)
+{
+ inline_irg_env *env = OALLOC(&temp_obst, inline_irg_env);
INIT_LIST_HEAD(&env->calls);
env->local_weights = NULL;
env->n_nodes = -2; /* do not count count Start, End */
env->n_callers = 0;
env->n_callers_orig = 0;
env->got_inline = 0;
- env->local_vars = 0;
env->recursive = 0;
return env;
}
* 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;
+static void collect_calls2(ir_node *call, void *ctx)
+{
+ wenv_t *env = (wenv_t*)ctx;
inline_irg_env *x = env->x;
- ir_opcode code = get_irn_opcode(call);
+ unsigned code = get_irn_opcode(call);
ir_graph *callee;
call_entry *entry;
if (env->ignore_runtime) {
ir_node *symc = get_Call_ptr(call);
- if (is_Global(symc)) {
- ir_entity *ent = get_Global_entity(symc);
+ if (is_SymConst_addr_ent(symc)) {
+ ir_entity *ent = get_SymConst_entity(symc);
if (get_entity_additional_properties(ent) & mtp_property_runtime)
return;
callee = get_call_called_irg(call);
if (callee != NULL) {
if (! env->ignore_callers) {
- inline_irg_env *callee_env = get_irg_link(callee);
+ inline_irg_env *callee_env = (inline_irg_env*)get_irg_link(callee);
/* count all static callers */
++callee_env->n_callers;
++callee_env->n_callers_orig;
x->recursive = 1;
/* link it in the list of possible inlinable entries */
- entry = obstack_alloc(&temp_obst, sizeof(*entry));
+ entry = OALLOC(&temp_obst, call_entry);
entry->call = call;
entry->callee = callee;
entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
/**
* Returns TRUE if the number of callers is 0 in the irg's environment,
- * hence this irg is a leave.
+ * hence this irg is a leaf.
*/
-inline static int is_leave(ir_graph *irg) {
- inline_irg_env *env = get_irg_link(irg);
+inline static int is_leaf(ir_graph *irg)
+{
+ inline_irg_env *env = (inline_irg_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, unsigned size) {
- inline_irg_env *env = get_irg_link(callee);
+inline static int is_smaller(ir_graph *callee, unsigned size)
+{
+ inline_irg_env *env = (inline_irg_env*)get_irg_link(callee);
return env->n_nodes < size;
}
* delta value for the loop depth
*/
static call_entry *duplicate_call_entry(const call_entry *entry,
- ir_node *new_call, int loop_depth_delta) {
- call_entry *nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
+ ir_node *new_call, int loop_depth_delta)
+{
+ call_entry *nentry = OALLOC(&temp_obst, call_entry);
nentry->call = new_call;
nentry->callee = entry->callee;
nentry->benefice = entry->benefice;
* @param src source environment
* @param loop_depth the loop depth of the call that is replaced by the src list
*/
-static void append_call_list(inline_irg_env *dst, inline_irg_env *src, int loop_depth) {
- call_entry *entry, *nentry;
+static void append_call_list(inline_irg_env *dst, inline_irg_env *src, int loop_depth)
+{
+ call_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. */
list_for_each_entry(call_entry, entry, &src->calls, list) {
- nentry = duplicate_call_entry(entry, get_irn_link(entry->call), loop_depth);
+ nentry = duplicate_call_entry(entry, (ir_node*)get_irn_link(entry->call), loop_depth);
list_add_tail(&nentry->list, &dst->calls);
}
dst->n_call_nodes += src->n_call_nodes;
}
/*
- * Inlines small leave methods at call sites where the called address comes
+ * Inlines small leaf 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(unsigned maxsize, unsigned leavesize,
- unsigned size, int ignore_runtime)
+void inline_leaf_functions(unsigned maxsize, unsigned leafsize,
+ unsigned size, int ignore_runtime)
{
inline_irg_env *env;
ir_graph *irg;
- int i, n_irgs;
+ size_t i, n_irgs;
ir_graph *rem;
int did_inline;
wenv_t wenv;
- call_entry *entry, *next;
- const call_entry *centry;
pmap *copied_graphs;
pmap_entry *pm_entry;
assert(get_irg_phase_state(irg) != phase_building);
free_callee_info(irg);
- assure_cf_loop(irg);
- wenv.x = get_irg_link(irg);
+ assure_irg_properties(irg,
+ IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
+ wenv.x = (inline_irg_env*)get_irg_link(irg);
irg_walk_graph(irg, NULL, collect_calls2, &wenv);
+ confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
}
/* -- and now inline. -- */
- /* Inline leaves recursively -- we might construct new leaves. */
+ /* Inline leafs recursively -- we might construct new leafs. */
do {
did_inline = 0;
int phiproj_computed = 0;
current_ir_graph = get_irp_irg(i);
- env = get_irg_link(current_ir_graph);
+ env = (inline_irg_env*)get_irg_link(current_ir_graph);
ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
callee = entry->callee;
prop = get_irg_inline_property(callee);
- if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
- /* do not inline forbidden / weak graphs */
+ if (prop == irg_inline_forbidden) {
continue;
}
- if (is_leave(callee) && (
- is_smaller(callee, leavesize) || prop >= irg_inline_forced)) {
+ if (is_leaf(callee) && (
+ is_smaller(callee, leafsize) || prop >= irg_inline_forced)) {
if (!phiproj_computed) {
phiproj_computed = 1;
collect_phiprojs(current_ir_graph);
did_inline = inline_method(call, callee);
if (did_inline) {
- inline_irg_env *callee_env = get_irg_link(callee);
+ inline_irg_env *callee_env = (inline_irg_env*)get_irg_link(callee);
/* call was inlined, Phi/Projs for current graph must be recomputed */
phiproj_computed = 0;
int phiproj_computed = 0;
current_ir_graph = get_irp_irg(i);
- env = get_irg_link(current_ir_graph);
+ env = (inline_irg_env*)get_irg_link(current_ir_graph);
ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
irg_inline_property prop;
ir_graph *callee;
- pmap_entry *e;
+ ir_graph *calleee;
call = entry->call;
callee = entry->callee;
prop = get_irg_inline_property(callee);
- if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
- /* do not inline forbidden / weak graphs */
+ if (prop == irg_inline_forbidden) {
continue;
}
- e = pmap_find(copied_graphs, callee);
- if (e != NULL) {
+ calleee = pmap_get(ir_graph, copied_graphs, callee);
+ if (calleee != NULL) {
/*
* Remap callee if we have a copy.
* FIXME: Should we do this only for recursive Calls ?
*/
- callee = e->value;
+ callee = calleee;
}
if (prop >= irg_inline_forced ||
/*
* No copy yet, create one.
- * Note that recursive methods are never leaves, so it is sufficient
+ * Note that recursive methods are never leafs, so it is sufficient
* to test this condition here.
*/
copy = create_irg_copy(callee);
callee_env = alloc_inline_irg_env();
set_irg_link(copy, callee_env);
- assure_cf_loop(copy);
+ assure_irg_properties(copy,
+ IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
wenv.x = callee_env;
wenv.ignore_callers = 1;
irg_walk_graph(copy, NULL, collect_calls2, &wenv);
/* call was inlined, Phi/Projs for current graph must be recomputed */
phiproj_computed = 0;
- /* callee was inline. Append it's call list. */
+ /* callee was inline. Append its call list. */
env->got_inline = 1;
--env->n_call_nodes;
append_call_list(env, callee_env, entry->loop_depth);
/* after we have inlined callee, all called methods inside callee
are now called once more */
list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
- inline_irg_env *penv = get_irg_link(centry->callee);
+ inline_irg_env *penv = (inline_irg_env*)get_irg_link(centry->callee);
++penv->n_callers;
}
for (i = 0; i < n_irgs; ++i) {
irg = get_irp_irg(i);
- env = get_irg_link(irg);
+ env = (inline_irg_env*)get_irg_link(irg);
if (env->got_inline) {
optimize_graph_df(irg);
/* kill the copied graphs: we don't need them anymore */
foreach_pmap(copied_graphs, pm_entry) {
- ir_graph *copy = pm_entry->value;
+ ir_graph *copy = (ir_graph*)pm_entry->value;
/* reset the entity, otherwise it will be deleted in the next step ... */
set_irg_entity(copy, NULL);
current_ir_graph = rem;
}
+typedef struct inline_leaf_functions_pass_t {
+ ir_prog_pass_t pass;
+ unsigned maxsize;
+ unsigned leafsize;
+ unsigned size;
+ int ignore_runtime;
+} inline_leaf_functions_pass_t;
+
+/**
+ * Wrapper to run inline_leaf_functions() as a ir_prog pass.
+ */
+static int inline_leaf_functions_wrapper(ir_prog *irp, void *context)
+{
+ inline_leaf_functions_pass_t *pass = (inline_leaf_functions_pass_t*)context;
+
+ (void)irp;
+ inline_leaf_functions(
+ pass->maxsize, pass->leafsize,
+ pass->size, pass->ignore_runtime);
+ return 0;
+}
+
+/* create a pass for inline_leaf_functions() */
+ir_prog_pass_t *inline_leaf_functions_pass(
+ const char *name, unsigned maxsize, unsigned leafsize,
+ unsigned size, int ignore_runtime)
+{
+ inline_leaf_functions_pass_t *pass = XMALLOCZ(inline_leaf_functions_pass_t);
+
+ pass->maxsize = maxsize;
+ pass->leafsize = leafsize;
+ pass->size = size;
+ pass->ignore_runtime = ignore_runtime;
+
+ return def_prog_pass_constructor(
+ &pass->pass,
+ name ? name : "inline_leaf_functions",
+ inline_leaf_functions_wrapper);
+}
+
/**
* Calculate the parameter weights for transmitting the address of a local variable.
*/
-static unsigned calc_method_local_weight(ir_node *arg) {
- int i, j, k;
+static unsigned calc_method_local_weight(ir_node *arg)
+{
+ int j;
unsigned v, weight = 0;
- for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
+ for (unsigned i = get_irn_n_outs(arg); i-- > 0; ) {
ir_node *succ = get_irn_out(arg, i);
switch (get_irn_opcode(succ)) {
ir_node *pred = get_Tuple_pred(succ, j);
if (pred == arg) {
/* look for Proj(j) */
- for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
+ for (unsigned k = get_irn_n_outs(succ); k-- > 0; ) {
ir_node *succ_succ = get_irn_out(succ, k);
if (is_Proj(succ_succ)) {
if (get_Proj_proj(succ_succ) == j) {
/**
* Calculate the parameter weights for transmitting the address of a local variable.
*/
-static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
+static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg)
+{
ir_entity *ent = get_irg_entity(irg);
ir_type *mtp;
- int nparams, i, proj_nr;
+ size_t nparams;
+ long proj_nr;
ir_node *irg_args, *arg;
mtp = get_entity_type(ent);
assure_irg_outs(irg);
irg_args = get_irg_args(irg);
- for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
+ for (unsigned i = get_irn_n_outs(irg_args); i-- > 0; ) {
arg = get_irn_out(irg_args, i);
proj_nr = get_Proj_proj(arg);
env->local_weights[proj_nr] = calc_method_local_weight(arg);
* After inlining, the local variable might be transformed into a
* SSA variable by scalar_replacement().
*/
-static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
- inline_irg_env *env = get_irg_link(callee);
-
- if (env->local_weights != NULL) {
- if (pos < ARR_LEN(env->local_weights))
- return env->local_weights[pos];
- return 0;
- }
+static unsigned get_method_local_adress_weight(ir_graph *callee, size_t pos)
+{
+ inline_irg_env *env = (inline_irg_env*)get_irg_link(callee);
- analyze_irg_local_weights(env, callee);
+ if (env->local_weights == NULL)
+ analyze_irg_local_weights(env, callee);
if (pos < ARR_LEN(env->local_weights))
return env->local_weights[pos];
{
ir_node *call = entry->call;
ir_entity *ent = get_irg_entity(callee);
+ ir_type *callee_frame;
+ size_t i, n_members, n_params;
ir_node *frame_ptr;
ir_type *mtp;
int weight = 0;
- int i, n_params, all_const;
+ int all_const;
unsigned cc, v;
irg_inline_property prop;
return entry->benefice = INT_MIN;
}
- if (get_irg_additional_properties(callee) & (mtp_property_noreturn | mtp_property_weak)) {
+ callee_frame = get_irg_frame_type(callee);
+ n_members = get_class_n_members(callee_frame);
+ for (i = 0; i < n_members; ++i) {
+ ir_entity *frame_ent = get_class_member(callee_frame, i);
+ if (is_parameter_entity(frame_ent)) {
+ // TODO inliner should handle parameter entities by inserting Store operations
+ DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden due to parameter entity\n", call, callee));
+ set_irg_inline_property(callee, irg_inline_forbidden);
+ return entry->benefice = INT_MIN;
+ }
+ }
+
+ if (get_irg_additional_properties(callee) & mtp_property_noreturn) {
DB((dbg, LEVEL_2, "In %+F Call to %+F: not inlining noreturn or weak\n",
call, callee));
return entry->benefice = INT_MIN;
cc = get_method_calling_convention(mtp);
if (cc & cc_reg_param) {
/* register parameter, smaller costs for register parameters */
- int max_regs = cc & ~cc_bits;
+ size_t max_regs = cc & ~cc_bits;
if (max_regs < n_params)
weight += max_regs * 2 + (n_params - max_regs) * 5;
}
entry->all_const = all_const;
- callee_env = get_irg_link(callee);
+ callee_env = (inline_irg_env*)get_irg_link(callee);
if (callee_env->n_callers == 1 &&
callee != current_ir_graph &&
- get_entity_visibility(ent) == visibility_local) {
+ !entity_is_externally_visible(ent)) {
weight += 700;
}
if (callee_env->n_nodes < 30 && !callee_env->recursive)
weight += 2000;
- /* and finally for leaves: they do not increase the register pressure
+ /* and finally for leafs: they do not increase the register pressure
because of callee safe registers */
if (callee_env->n_call_nodes == 0)
weight += 400;
return entry->benefice = weight;
}
-static ir_graph **irgs;
-static int last_irg;
+typedef struct walk_env_t {
+ ir_graph **irgs;
+ size_t last_irg;
+} walk_env_t;
/**
* Callgraph walker, collect all visited graphs.
*/
-static void callgraph_walker(ir_graph *irg, void *data) {
- (void) data;
- irgs[last_irg++] = irg;
+static void callgraph_walker(ir_graph *irg, void *data)
+{
+ walk_env_t *env = (walk_env_t *)data;
+ env->irgs[env->last_irg++] = irg;
}
/**
*
* @return the list of graphs.
*/
-static ir_graph **create_irg_list(void) {
- ir_entity **free_methods;
- int arr_len;
- int n_irgs = get_irp_n_irgs();
+static ir_graph **create_irg_list(void)
+{
+ ir_entity **free_methods;
+ size_t n_irgs = get_irp_n_irgs();
+ walk_env_t env;
- cgana(&arr_len, &free_methods);
+ cgana(&free_methods);
xfree(free_methods);
compute_callgraph();
- last_irg = 0;
- irgs = XMALLOCNZ(ir_graph*, n_irgs);
+ env.irgs = XMALLOCNZ(ir_graph*, n_irgs);
+ env.last_irg = 0;
+
+ callgraph_walk(NULL, callgraph_walker, &env);
+ assert(n_irgs == env.last_irg);
- callgraph_walk(NULL, callgraph_walker, NULL);
- assert(n_irgs == last_irg);
+ free_callgraph();
- return irgs;
+ return env.irgs;
}
/**
int inline_threshold, pmap *copied_graphs)
{
int phiproj_computed = 0;
- inline_irg_env *env = get_irg_link(irg);
- call_entry *curr_call;
+ inline_irg_env *env = (inline_irg_env*)get_irg_link(irg);
wenv_t wenv;
pqueue_t *pqueue;
/* note that the list of possible calls is updated during the process */
while (!pqueue_empty(pqueue)) {
int did_inline;
- call_entry *curr_call = pqueue_pop_front(pqueue);
+ call_entry *curr_call = (call_entry*)pqueue_pop_front(pqueue);
ir_graph *callee = curr_call->callee;
ir_node *call_node = curr_call->call;
- inline_irg_env *callee_env = get_irg_link(callee);
+ inline_irg_env *callee_env = (inline_irg_env*)get_irg_link(callee);
irg_inline_property prop = get_irg_inline_property(callee);
+ ir_graph *calleee;
int loop_depth;
- const call_entry *centry;
- pmap_entry *e;
if ((prop < irg_inline_forced) && env->n_nodes + callee_env->n_nodes > maxsize) {
DB((dbg, LEVEL_2, "%+F: too big (%d) + %+F (%d)\n", irg,
continue;
}
- e = pmap_find(copied_graphs, callee);
- if (e != NULL) {
+ calleee = pmap_get(ir_graph, copied_graphs, callee);
+ if (calleee != NULL) {
int benefice = curr_call->benefice;
/*
* Reduce the weight for recursive function IFF not all arguments are const.
/*
* Remap callee if we have a copy.
*/
- callee = e->value;
- callee_env = get_irg_link(callee);
+ callee = calleee;
+ callee_env = (inline_irg_env*)get_irg_link(callee);
}
if (current_ir_graph == callee) {
/*
* No copy yet, create one.
- * Note that recursive methods are never leaves, so it is
+ * Note that recursive methods are never leafs, so it is
* sufficient to test this condition here.
*/
copy = create_irg_copy(callee);
callee_env = alloc_inline_irg_env();
set_irg_link(copy, callee_env);
- assure_cf_loop(copy);
+ assure_irg_properties(copy, IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
+ memset(&wenv, 0, sizeof(wenv));
wenv.x = callee_env;
wenv.ignore_callers = 1;
irg_walk_graph(copy, NULL, collect_calls2, &wenv);
/* remove it from the caller list */
list_del(&curr_call->list);
- /* callee was inline. Append it's call list. */
+ /* callee was inline. Append its call list. */
env->got_inline = 1;
- if (curr_call->local_adr)
- env->local_vars = 1;
--env->n_call_nodes;
/* we just generate a bunch of new calls */
loop_depth = curr_call->loop_depth;
list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
- inline_irg_env *penv = get_irg_link(centry->callee);
+ inline_irg_env *penv = (inline_irg_env*)get_irg_link(centry->callee);
ir_node *new_call;
call_entry *new_entry;
/* 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. */
- new_call = get_irn_link(centry->call);
+ new_call = (ir_node*)get_irn_link(centry->call);
+ if (get_irn_irg(new_call) != irg) {
+ /* centry->call has not been copied, which means it is dead.
+ * This might happen during inlining, if a const function,
+ * which cannot be inlined is only used as an unused argument
+ * of another function, which is inlined. */
+ continue;
+ }
assert(is_Call(new_call));
new_entry = duplicate_call_entry(centry, new_call, loop_depth);
* Heuristic inliner. Calculates a benefice value for every call and inlines
* those calls with a value higher than the threshold.
*/
-void inline_functions(unsigned maxsize, int inline_threshold) {
+void inline_functions(unsigned maxsize, int inline_threshold,
+ opt_ptr after_inline_opt)
+{
inline_irg_env *env;
- int i, n_irgs;
+ size_t i, n_irgs;
ir_graph *rem;
wenv_t wenv;
pmap *copied_graphs;
free_callee_info(irg);
- wenv.x = get_irg_link(irg);
- assure_cf_loop(irg);
+ wenv.x = (inline_irg_env*)get_irg_link(irg);
+ assure_loopinfo(irg);
irg_walk_graph(irg, NULL, collect_calls2, &wenv);
}
for (i = 0; i < n_irgs; ++i) {
ir_graph *irg = irgs[i];
- env = get_irg_link(irg);
- if (env->got_inline) {
+ env = (inline_irg_env*)get_irg_link(irg);
+ if (env->got_inline && after_inline_opt != NULL) {
/* this irg got calls inlined: optimize it */
- if (get_opt_combo()) {
- if (env->local_vars) {
- scalar_replacement_opt(irg);
- }
- combo(irg);
- } else {
- if (env->local_vars) {
- if (scalar_replacement_opt(irg)) {
- optimize_graph_df(irg);
- }
- }
- optimize_cf(irg);
- }
+ after_inline_opt(irg);
}
if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
/* kill the copied graphs: we don't need them anymore */
foreach_pmap(copied_graphs, pm_entry) {
- ir_graph *copy = pm_entry->value;
+ ir_graph *copy = (ir_graph*)pm_entry->value;
/* reset the entity, otherwise it will be deleted in the next step ... */
set_irg_entity(copy, NULL);
current_ir_graph = rem;
}
-void firm_init_inline(void) {
+typedef struct inline_functions_pass_t {
+ ir_prog_pass_t pass;
+ unsigned maxsize;
+ int inline_threshold;
+ opt_ptr after_inline_opt;
+} inline_functions_pass_t;
+
+/**
+ * Wrapper to run inline_functions() as a ir_prog pass.
+ */
+static int inline_functions_wrapper(ir_prog *irp, void *context)
+{
+ inline_functions_pass_t *pass = (inline_functions_pass_t*)context;
+
+ (void)irp;
+ inline_functions(pass->maxsize, pass->inline_threshold,
+ pass->after_inline_opt);
+ return 0;
+}
+
+/* create a ir_prog pass for inline_functions */
+ir_prog_pass_t *inline_functions_pass(
+ const char *name, unsigned maxsize, int inline_threshold,
+ opt_ptr after_inline_opt)
+{
+ inline_functions_pass_t *pass = XMALLOCZ(inline_functions_pass_t);
+
+ pass->maxsize = maxsize;
+ pass->inline_threshold = inline_threshold;
+ pass->after_inline_opt = after_inline_opt;
+
+ return def_prog_pass_constructor(
+ &pass->pass, name ? name : "inline_functions",
+ inline_functions_wrapper);
+}
+
+void firm_init_inline(void)
+{
FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
}
projects / libfirm / blobdiff