#include "config.h"
#endif
-#ifdef HAVE_ALLOCA_H
-#include <alloca.h>
-#endif
-
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
#include "obst.h"
#include "set.h"
#include "pset.h"
#include "iredges_t.h"
#include "ircons_t.h"
#include "irprintf.h"
+#include "xmalloc.h"
#include "beutil.h"
-#include "bearch.h"
+#include "bearch_t.h"
#include "bespillbelady.h"
#include "beuses_t.h"
#include "besched_t.h"
#include "benode_t.h"
#include "bechordal_t.h"
#include "bespilloptions.h"
+#include "beloopana.h"
#define DBG_SPILL 1
#define DBG_WSETS 2
const arch_env_t *arch;
const arch_register_class_t *cls;
const be_lv_t *lv;
+ be_loopana_t *loop_ana;
int n_regs; /** number of regs in this reg-class */
workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
return p->time - q->time;
}
-void workset_print(const workset_t *w)
+static INLINE void workset_print(const workset_t *w)
{
int i;
return use.time;
}
-#if 0
-/**
- * Fix to remove dead nodes (especially don't spill nodes) from workset.
- */
-static void fix_dead_values(workset_t *ws, ir_node *irn) {
- int idx;
- ir_node *node;
- ir_node *block = get_nodes_block(irn);
-
- DBG((dbg, DBG_DECIDE, "fixing dead values at %+F:\n", irn));
-
- workset_foreach(ws, node, idx) {
- const ir_edge_t *edge;
- int fixme = 1;
-
- /* skip already fixed nodes */
- if (workset_get_time(ws, idx) == INT_MAX)
- continue;
-
- /* check all users */
- foreach_out_edge(node, edge) {
- ir_node *user = get_edge_src_irn(edge);
-
- if ((get_nodes_block(user) != block) || /* user is in a different block */
- (sched_is_scheduled(user) && sched_comes_after(irn, user)) || /* user is scheduled after irn */
- user == irn) /* irn is the user */
- { /* => don't fix distance */
- fixme = 0;
- break;
- }
- }
-
- /* all users scheduled prior to current irn in in same block as irn -> fix */
- if (fixme) {
- workset_set_time(ws, idx, INT_MAX);
- DBG((dbg, DBG_DECIDE, "\tfixing time for %+F to INT_MAX\n", node));
- }
- }
-
-}
-#endif
-
/**
* Performs the actions necessary to grant the request that:
* - new_vals can be held in registers
workset_set_time(ws, i, dist);
}
-#if 0
- /*
- FIX for don't spill nodes:
- Problem is that get_distance always returns 0 for those nodes even if they are not
- needed anymore (all their usages have already been visited).
- Even if we change this behavior, get_distance doesn't distinguish between not
- used anymore (dead) and live out of block.
- Solution: Set distances of all nodes having all their usages in schedule prior to
- current instruction to MAX_INT.
- */
- fix_dead_values(ws, env->instr);
-#endif
-
/* sort entries by increasing nextuse-distance*/
workset_sort(ws);
static void belady(ir_node *block, void *env);
-static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first, ir_node *node, ir_node *block, ir_loop *loop) {
+/** Decides whether a specific node should be in the start workset or not
+ *
+ * @param env belady environment
+ * @param first
+ * @param node the node to test
+ * @param block the block of the node
+ * @param loop the loop of the node
+ */
+static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first,
+ ir_node *node, ir_node *block,
+ ir_loop *loop)
+{
be_next_use_t next_use;
loc_t loc;
loc.time = USES_INFINITY;
+ loc.irn = node;
if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) {
loc.time = USES_INFINITY;
return loc;
}
- loc.irn = node;
-
/* We have to keep nonspillable nodes in the workingset */
if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) {
loc.time = 0;
next_use = be_get_next_use(env->uses, first, 0, node, 0);
if(USES_IS_INFINITE(next_use.time)) {
// the nodes marked as live in shouldn't be dead, so it must be a phi
+ assert(is_Phi(node));
loc.time = USES_INFINITY;
DBG((dbg, DBG_START, " %+F not taken (dead)\n", node));
- assert(is_Phi(node));
+ if(is_Phi(node)) {
+ be_spill_phi(env->senv, node);
+ }
return loc;
}
if(next_use.outermost_loop >= get_loop_depth(loop)) {
DBG((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
- return loc;
- // ARR_APP1(loc_t, starters, loc);
} else {
- loc.time = USES_INFINITY;
- DBG((dbg, DBG_START, " %+F not taken (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
- return loc;
+ loc.time = USES_PENDING;
+ DBG((dbg, DBG_START, " %+F delayed (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
}
+ return loc;
}
/*
* and notifies spill algorithm which phis need to be spilled
*/
static void compute_live_ins(ir_node *block, void *data) {
- belady_env_t *env = data;
- block_info_t *block_info;
- ir_node *first, *irn;
- loc_t loc, *starters;
- int i, len, ws_count;
- ir_loop *loop = get_irn_loop(block);
- const be_lv_t *lv = env->lv;
-
- if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
+ belady_env_t *env = data;
+ ir_loop *loop = get_irn_loop(block);
+ const be_lv_t *lv = env->lv;
+ block_info_t *block_info;
+ ir_node *first, *irn;
+ loc_t loc, *starters, *delayed;
+ int i, len, ws_count;
+ int free_slots, free_pressure_slots;
+ unsigned pressure;
+
+ if (get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
return;
block_info = new_block_info(&env->ob);
/* Collect all values living at start of block */
starters = NEW_ARR_F(loc_t, 0);
-
- /* rebuild schedule time information, because it seems to be broken */
- // Matze: is this still true?
- //sched_renumber(block);
+ delayed = NEW_ARR_F(loc_t, 0);
DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
first = sched_first(block);
+ /* check all Phis first */
sched_foreach(block, irn) {
- if(!is_Phi(irn))
+ if (! is_Phi(irn))
break;
loc = to_take_or_not_to_take(env, first, irn, block, loop);
- if(!USES_IS_INFINITE(loc.time)) {
- ARR_APP1(loc_t, starters, loc);
- } else {
- be_spill_phi(env->senv, irn);
+ if (! USES_IS_INFINITE(loc.time)) {
+ if (USES_IS_PENDING(loc.time))
+ ARR_APP1(loc_t, delayed, loc);
+ else
+ ARR_APP1(loc_t, starters, loc);
}
}
+ /* check all Live-Ins */
be_lv_foreach(lv, block, be_lv_state_in, i) {
ir_node *node = be_lv_get_irn(lv, block, i);
loc = to_take_or_not_to_take(env, first, node, block, loop);
- if(!USES_IS_INFINITE(loc.time)) {
- ARR_APP1(loc_t, starters, loc);
+ if (! USES_IS_INFINITE(loc.time)) {
+ if (USES_IS_PENDING(loc.time))
+ ARR_APP1(loc_t, delayed, loc);
+ else
+ ARR_APP1(loc_t, starters, loc);
}
}
- // Sort start values by first use
+ pressure = be_get_loop_pressure(env->loop_ana, env->cls, loop);
+ assert(ARR_LEN(delayed) <= pressure);
+ free_slots = env->n_regs - ARR_LEN(starters);
+ free_pressure_slots = env->n_regs - (pressure - ARR_LEN(delayed));
+ free_slots = MIN(free_slots, free_pressure_slots);
+ /* append nodes delayed due to loop structure until start set is full */
+ for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
+ DBG((dbg, DBG_START, " delayed %+F taken\n", delayed[i].irn));
+ ARR_APP1(loc_t, starters, delayed[i]);
+ delayed[i].irn = NULL;
+ }
+
+ /* spill all delayed phis which didn't make it into start workset */
+ for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
+ ir_node *irn = delayed[i].irn;
+ if (irn && is_Phi(irn)) {
+ DBG((dbg, DBG_START, " spilling delayed phi %+F\n", irn));
+ be_spill_phi(env->senv, irn);
+ }
+ }
+ DEL_ARR_F(delayed);
+
+ /* Sort start values by first use */
qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
/* Copy the best ones from starters to start workset */
- ws_count = MIN(ARR_LEN(starters), env->n_regs);
+ ws_count = MIN(ARR_LEN(starters), env->n_regs);
block_info->ws_start = new_workset(env, &env->ob);
workset_bulk_fill(block_info->ws_start, ws_count, starters);
len = ARR_LEN(starters);
for (i = ws_count; i < len; ++i) {
irn = starters[i].irn;
- if (!is_Phi(irn) || get_nodes_block(irn) != block)
+ if (! is_Phi(irn) || get_nodes_block(irn) != block)
continue;
be_spill_phi(env->senv, irn);
}
}
-void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) {
+/**
+ * Do spilling for a register class on a graph using the belady heuristic.
+ * In the transformed graph, the register pressure never exceeds the number
+ * of available registers.
+ *
+ * @param birg The backend graph
+ * @param cls The register class to spill
+ */
+static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) {
be_spill_belady_spill_env(birg, cls, NULL);
}
void be_spill_belady_spill_env(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) {
belady_env_t env;
ir_graph *irg = be_get_birg_irg(birg);
+ int n_regs;
- FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
- //firm_dbg_set_mask(dbg, DBG_SPILL);
+ /* some special classes contain only ignore regs, nothing to do then */
+ n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
+ if(n_regs == 0)
+ return;
+ be_invalidate_liveness(birg);
be_assure_liveness(birg);
+ /* construct control flow loop tree */
+ if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
+ construct_cf_backedges(irg);
+ }
/* init belady env */
obstack_init(&env.ob);
env.arch = birg->main_env->arch_env;
env.cls = cls;
env.lv = be_get_birg_liveness(birg);
- env.n_regs = env.cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
+ env.n_regs = n_regs;
env.ws = new_workset(&env, &env.ob);
env.uses = be_begin_uses(irg, env.lv);
+ env.loop_ana = be_new_loop_pressure(birg);
if(spill_env == NULL) {
env.senv = be_new_spill_env(birg);
} else {
if(spill_env == NULL)
be_delete_spill_env(env.senv);
be_end_uses(env.uses);
+ be_free_loop_pressure(env.loop_ana);
obstack_free(&env.ob, NULL);
}
-static void be_spill_belady_oldinterface(const be_chordal_env_t *chordal_env)
-{
- be_spill_belady(chordal_env->birg, chordal_env->cls);
-}
-
void be_init_spillbelady(void)
{
static be_spiller_t belady_spiller = {
- be_spill_belady_oldinterface
+ be_spill_belady
};
be_register_spiller("belady", &belady_spiller);
+ FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
}
BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);
projects / libfirm / blobdiff