X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespill.c;h=cfc5e62b2811c478c551e5a9537ed520805d2588;hb=17e47394fa72025d14172a2acef2d258a067aa42;hp=f18206dfd8bc21efb9ad5aa2e21b9a7bfd9198a8;hpb=a7300fa5d62b1b3e244d3b2f4e1f982c92da36ed;p=libfirm

diff --git a/ir/be/bespill.c b/ir/be/bespill.c
index f18206dfd..cfc5e62b2 100644
--- a/ir/be/bespill.c
+++ b/ir/be/bespill.c
@@ -4,15 +4,26 @@
  * Copyright:   (c) Universitaet Karlsruhe
  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
  */
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <stdlib.h>
 
 #include "pset.h"
 #include "irnode_t.h"
 #include "ircons_t.h"
+#include "iredges_t.h"
+#include "ident_t.h"
+#include "type_t.h"
+#include "entity_t.h"
 #include "debug.h"
+#include "irgwalk.h"
 
 #include "besched.h"
 #include "bespill.h"
 #include "benode_t.h"
+#include "bechordal_t.h"
 
 typedef struct _reloader_t reloader_t;
 typedef struct _spill_info_t spill_info_t;
@@ -36,10 +47,13 @@ typedef struct _spill_ctx_t {
 struct _spill_env_t {
 	firm_dbg_module_t *dbg;
 	const arch_register_class_t *cls;
-	const be_main_session_env_t *session;
+	const be_chordal_env_t *chordal_env;
 	struct obstack obst;
 	set *spill_ctxs;
-	set *spills;		/**< all spill_info_t's, which must be placed */
+	set *spills;				/**< all spill_info_t's, which must be placed */
+	pset *mem_phis;				/**< set of all special spilled phis. allocated and freed seperately */
+	decide_irn_t is_mem_phi;	/**< callback func to decide if a phi needs special spilling */
+	void *data;					/**< data passed to all callbacks */
 };
 
 static int cmp_spillctx(const void *a, const void *b, size_t n) {
@@ -55,13 +69,17 @@ static int cmp_spillinfo(const void *x, const void *y, size_t size) {
 }
 
 spill_env_t *be_new_spill_env(firm_dbg_module_t *dbg,
-		const be_main_session_env_t *session, const arch_register_class_t *cls) {
+							  const be_chordal_env_t *chordal_env,
+							  decide_irn_t is_mem_phi, void *data) {
+
 	spill_env_t *env = malloc(sizeof(env[0]));
-	env->spill_ctxs = new_set(cmp_spillctx, 1024);
-	env->spills     = new_set(cmp_spillinfo, 1024);
-	env->session    = session;
-	env->cls        = cls;
-	env->dbg        = dbg;
+	env->spill_ctxs  = new_set(cmp_spillctx, 1024);
+	env->spills      = new_set(cmp_spillinfo, 1024);
+	env->cls         = chordal_env->cls;
+	env->dbg         = dbg;
+	env->is_mem_phi  = is_mem_phi;
+	env->data        = data;
+	env->chordal_env = chordal_env;
 	obstack_init(&env->obst);
 	return env;
 }
@@ -85,12 +103,12 @@ static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_ir
 
 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
 	spill_ctx_t *ctx;
-	DBG((senv->dbg, LEVEL_1, "spill_irn %+F\n", irn));
+	DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
 
 	ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
 	if(!ctx->spill) {
-		const be_main_env_t *env = senv->session->main_env;
-		ctx->spill = be_spill(env->node_factory, env->arch_env, irn);
+		const be_main_env_t *env = senv->chordal_env->birg->main_env;
+		ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
 	}
 
 	return ctx->spill;
@@ -102,15 +120,14 @@ static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn)
  * Thus we spill it and all its operands to the same spill slot.
  * Therefore the phi/dataB becomes a phi/Memory
  */
-static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn,
-		pset *mem_phis) {
+static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
 	int i, n = get_irn_arity(phi);
 	ir_node **ins, *bl = get_nodes_block(phi);
-	ir_graph *irg = senv->session->irg;
+	ir_graph *irg = senv->chordal_env->irg;
 	spill_ctx_t *ctx;
 
 	assert(is_Phi(phi));
-	DBG((senv->dbg, LEVEL_1, "spill_phi %+F\n", phi));
+	DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
 
 	/* search an existing spill for this context */
 	ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
@@ -121,7 +138,7 @@ static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn,
 		ins  = malloc(n * sizeof(ins[0]));
 		for(i=0; i<n; ++i)
 			ins[i] = new_r_Unknown(irg, mode_M);
-		ctx->spill = new_r_Phi(senv->session->irg, bl, n, ins, mode_M);
+		ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
 		free(ins);
 
 		/* re-wire the phiM */
@@ -129,8 +146,8 @@ static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn,
 			ir_node *arg = get_irn_n(phi, i);
 			ir_node *sub_res;
 
-			if(is_Phi(arg) && pset_find_ptr(mem_phis, arg))
-				sub_res = be_spill_phi(senv, arg, ctx_irn, mem_phis);
+			if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
+				sub_res = be_spill_phi(senv, arg, ctx_irn);
 			else
 				sub_res = be_spill_irn(senv, arg, ctx_irn);
 
@@ -140,33 +157,58 @@ static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn,
 	return ctx->spill;
 }
 
-static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill, pset *mem_phis) {
+static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
 	ir_node *res;
-	if (pset_find_ptr(mem_phis, to_spill))
-		res = be_spill_phi(senv, to_spill, to_spill, mem_phis);
+	if (pset_find_ptr(senv->mem_phis, to_spill))
+		res = be_spill_phi(senv, to_spill, to_spill);
 	else
 		res = be_spill_irn(senv, to_spill, to_spill);
 
 	return res;
 }
 
-void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set, decide_irn_t is_mem_phi, void *data) {
-	ir_graph *irg = senv->session->irg;
+static void phi_walker(ir_node *irn, void *env) {
+	spill_env_t *senv = env;
+	const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
+
+	if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
+			&& senv->is_mem_phi(irn, senv->data)) {
+		DBG((senv->dbg, LEVEL_1, "  %+F\n", irn));
+		pset_insert_ptr(senv->mem_phis, irn);
+	}
+}
+
+void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
+	ir_graph *irg = senv->chordal_env->irg;
 	ir_node *irn;
 	spill_info_t *si;
 	struct obstack ob;
-	pset *mem_phis = pset_new_ptr_default();
 
 	obstack_init(&ob);
 
 	/* get all special spilled phis */
-	for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
-		irn = si->spilled_node;
-		if (is_Phi(irn) && is_mem_phi(irn, data))
-			pset_insert_ptr(mem_phis, irn);
+	DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
+	senv->mem_phis = pset_new_ptr_default();
+	irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
+
+	/* Add reloads for mem_phis */
+	/* BETTER: These reloads (1) should only be inserted, if they are really needed */
+	DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
+	for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
+		const ir_edge_t *e;
+		DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
+		foreach_out_edge(irn, e) {
+			ir_node *user = e->src;
+			if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
+					ir_node *use_bl = get_nodes_block(user);
+					DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
+					be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
+			}
+		}
 	}
 
 	/* process each spilled node */
+	DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
 	for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
 		reloader_t *rld;
 		ir_node **reloads;
@@ -176,15 +218,13 @@ void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set, decide_irn_t
 		/* go through all reloads for this spill */
 		for(rld = si->reloaders; rld; rld = rld->next) {
 			/* the spill for this reloader */
-			ir_node *spill   = be_spill_node(senv, si->spilled_node, mem_phis);
+			ir_node *spill   = be_spill_node(senv, si->spilled_node);
 
 			/* the reload */
 			ir_node *bl      = is_Block(rld->reloader) ? rld->reloader : get_nodes_block(rld->reloader);
-			ir_node *reload  = new_Reload(senv->session->main_env->node_factory,
-														senv->cls, irg, bl, mode, spill);
+			ir_node *reload  = be_new_Reload(senv->cls, irg, bl, mode, spill);
 
-			DBG((senv->dbg, LEVEL_2, " RELOADER %+F   Reload %+F of %+F\n",
-						rld->reloader, reload, si->spilled_node));
+			DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", reload, si->spilled_node, rld->reloader));
 			if(reload_set)
 				pset_insert_ptr(reload_set, reload);
 
@@ -195,22 +235,22 @@ void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set, decide_irn_t
 		}
 
 		assert(n_reloads > 0);
+		obstack_ptr_grow(&ob, si->spilled_node);
 		reloads = obstack_finish(&ob);
-		be_introduce_copies_ignore(senv->session->dom_front, si->spilled_node,
-				n_reloads, reloads, mem_phis);
+		be_ssa_constr_ignore(senv->chordal_env->dom_front, n_reloads + 1, reloads, senv->mem_phis);
 		obstack_free(&ob, reloads);
 	}
 
 	obstack_free(&ob, NULL);
 
-	for(irn = pset_first(mem_phis); irn; irn = pset_next(mem_phis)) {
+	for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
 		int i, n;
 		for(i = 0, n = get_irn_arity(irn); i < n; ++i)
-			set_irn_n(irn, i, new_r_Bad(senv->session->irg));
+			set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
 		sched_remove(irn);
 	}
 
-	del_pset(mem_phis);
+	del_pset(senv->mem_phis);
 }
 
 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
@@ -228,7 +268,240 @@ void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
 }
 
 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
-	ir_node *insert_bl = get_irn_arity(bl) == 1
-		? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
+	ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
 	be_add_reload(senv, to_spill, insert_bl);
 }
+
+
+
+/****************************************
+
+	SPILL SLOT MANAGEMENT AND OPTS
+
+****************************************/
+
+typedef struct _spill_slot_t {
+	unsigned size;
+	unsigned align;
+	pset *members;
+	ir_mode *largest_mode;	/* the mode of all members with largest size */
+} spill_slot_t;
+
+typedef struct _ss_env_t {
+	firm_dbg_module_t *dbg;
+	struct obstack ob;
+	be_chordal_env_t *cenv;
+	pmap *slots;		/* maps spill_contexts to spill_slots */
+  pmap *types;    /* maps modes to types */
+} ss_env_t;
+
+
+static void compute_spill_slots_walker(ir_node *spill, void *env) {
+	ss_env_t *ssenv = env;
+	ir_node *ctx;
+	pmap_entry *entry;
+	spill_slot_t *ss;
+
+	if (!be_is_Spill(spill))
+		return;
+
+	/* check, if this spill is for a context already known */
+	ctx = be_get_Spill_context(spill);
+	entry = pmap_find(ssenv->slots, ctx);
+
+	if (!entry) {
+		/* this is a new spill context */
+		ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
+		ss->members = pset_new_ptr(8);
+		ss->largest_mode = get_irn_mode(get_irn_n(spill, 0));
+		ss->size = get_mode_size_bytes(ss->largest_mode);
+		ss->align = ss->size; /* TODO Assumed for now */
+		pmap_insert(ssenv->slots, ctx, ss);
+	} else {
+		ir_node *irn;
+		/* values with the same spill_ctx must go into the same spill slot */
+		ss = entry->value;
+		assert(ss->size == (unsigned)get_mode_size_bytes(get_irn_mode(get_irn_n(spill, 0))) && "Different sizes for the same spill slot are not allowed yet.");
+		for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
+			/* use values_interfere here, because it uses the dominance check,
+			   which does work for values in memory */
+			assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
+		}
+	}
+
+	pset_insert_ptr(ss->members, spill);
+}
+
+static int ss_sorter(const void *v1, const void *v2) {
+	const spill_slot_t *ss1 = v1;
+	const spill_slot_t *ss2 = v2;
+	return ((int) ss2->size) - ((int) ss1->size);
+}
+
+
+/**
+ * This function should optimize the spill slots.
+ *  - Coalescing of multiple slots
+ *  - Ordering the slots
+ *
+ * Input slots are in @p ssenv->slots
+ * @p size The count of initial spill slots in @p ssenv->slots
+ *         This also is the size of the preallocated array @p ass
+ *
+ * @return An array of spill slots @p ass in specific order
+ **/
+static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
+	int i, o, used_slots;
+	pmap_entry *entr;
+
+	i=0;
+	pmap_foreach(ssenv->slots, entr)
+		ass[i++] = entr->value;
+
+	/* Sort the array to minimize fragmentation and cache footprint.
+	   Large slots come first */
+	qsort(ass, size, sizeof(ass[0]), ss_sorter);
+
+	/* For each spill slot:
+		- assign a new offset to this slot
+	    - xor find another slot to coalesce with */
+	used_slots = 0;
+	for (i=0; i<size; ++i) { /* for each spill slot */
+		ir_node *n1;
+		int tgt_slot = -1;
+
+		DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
+		for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
+			DBG((ssenv->dbg, LEVEL_1, "  %+F\n", n1));
+
+
+		for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
+			/* check inter-slot-pairs for interference */
+			ir_node *n2;
+			for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
+				for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
+					if(values_interfere(n1, n2)) {
+						pset_break(ass[i]->members);
+						pset_break(ass[o]->members);
+						DBG((ssenv->dbg, LEVEL_1, "    Interf %+F -- %+F\n", n1, n2));
+						goto interf_detected;
+					}
+
+			/* if we are here, there is no interference between ass[i] and ass[o] */
+			tgt_slot = o;
+
+interf_detected: /*nothing*/ ;
+		}
+
+		/* now the members of ass[i] join the members of ass[tgt_slot] */
+
+		/* do we need a new slot? */
+		if (tgt_slot == -1) {
+			tgt_slot = used_slots;
+			used_slots++;
+
+			/* init slot */
+			if (tgt_slot != i) {
+				ass[tgt_slot]->size = ass[i]->size;
+				del_pset(ass[tgt_slot]->members);
+				ass[tgt_slot]->members = pset_new_ptr(8);
+			}
+		}
+
+		/* copy the members to the target pset */
+		/* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
+		if (tgt_slot != i)
+			for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
+					pset_insert_ptr(ass[tgt_slot]->members, n1);
+	}
+}
+
+#define ALIGN_SPILL_AREA 16
+#define pset_foreach(pset, elm)  for(elm=pset_first(pset); elm; elm=pset_next(pset))
+
+/**
+ * Returns a spill type for a mode. Keep them in a map to reduce
+ * the number of types.
+ */
+static ir_type *get_spill_type(pmap *types, ir_mode *mode) {
+  pmap_entry *e = pmap_find(types, mode);
+  ir_type *res;
+
+  if (! e) {
+		char buf[64];
+    snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(mode));
+    res = new_type_primitive(new_id_from_str(buf), mode);
+    pmap_insert(types, mode, res);
+  }
+  else
+    res = e->value;
+  return res;
+}
+
+static void assign_entities(ss_env_t *ssenv, int n, spill_slot_t **ss) {
+	int i, offset;
+	ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
+
+	/* aligning by increasing frame size */
+	offset = get_type_size_bits(frame) / 8;
+	offset = round_up2(offset, ALIGN_SPILL_AREA);
+	set_type_size_bytes(frame, -1);
+
+	/* create entities and assign offsets according to size and alignment*/
+	for (i=0; i<n; ++i) {
+		char buf[64];
+		ident *name;
+		entity *spill_ent;
+		ir_node *irn;
+
+		/* build entity */
+		snprintf(buf, sizeof(buf), "spill_slot_%d", i);
+		name = new_id_from_str(buf);
+
+		spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]->largest_mode));
+
+		/* align */
+		offset = round_up2(offset, ss[i]->align);
+		/* set */
+		set_entity_offset_bytes(spill_ent, offset);
+		/* next possible offset */
+		offset += ss[i]->size;
+
+		pset_foreach(ss[i]->members, irn)
+			be_set_Spill_entity(irn, spill_ent);
+	}
+
+	/* set final size of stack frame */
+	set_type_size_bytes(frame, offset);
+}
+
+void be_compute_spill_offsets(be_chordal_env_t *cenv) {
+	ss_env_t ssenv;
+	spill_slot_t **ss;
+	int ss_size;
+	pmap_entry *pme;
+
+	obstack_init(&ssenv.ob);
+	ssenv.cenv  = cenv;
+	ssenv.slots = pmap_create();
+	ssenv.types = pmap_create();
+	ssenv.dbg   = firm_dbg_register("ir.be.spillslots");
+
+	/* Get initial spill slots */
+	irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
+
+	/* Build an empty array for optimized spill slots */
+	ss_size = pmap_count(ssenv.slots);
+	ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
+	optimize_slots(&ssenv, ss_size, ss);
+
+	/* Integrate slots into the stack frame entity */
+	assign_entities(&ssenv, ss_size, ss);
+
+	/* Clean up */
+	pmap_foreach(ssenv.slots, pme)
+		del_pset(((spill_slot_t *)pme->value)->members);
+	pmap_destroy(ssenv.slots);
+  pmap_destroy(ssenv.types);
+	obstack_free(&ssenv.ob, NULL);
+}