X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespill.c;h=b15f46e49bba3421b2a79ed02f85b36f84317146;hb=03214ca29ad177e9d428806b442a6011f70ec6fc;hp=f8df2a8c65a4e6bdf45538a3b9e9dd7fc57d3c5b;hpb=56f231e9f79822a8421a9b8fe72aecb6239ed927;p=libfirm

diff --git a/ir/be/bespill.c b/ir/be/bespill.c
index f8df2a8c6..b15f46e49 100644
--- a/ir/be/bespill.c
+++ b/ir/be/bespill.c
@@ -14,14 +14,23 @@
 #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 "array.h"
 
-#include "besched.h"
+#include "belive_t.h"
+#include "besched_t.h"
 #include "bespill.h"
 #include "benode_t.h"
 #include "bechordal_t.h"
 
+#undef REMAT
+/* This enables re-computation of values. Current state: Unfinished and buggy. */
+#undef BUGGY_REMAT
+
 typedef struct _reloader_t reloader_t;
 typedef struct _spill_info_t spill_info_t;
 
@@ -42,15 +51,15 @@ typedef struct _spill_ctx_t {
 } spill_ctx_t;
 
 struct _spill_env_t {
-	firm_dbg_module_t *dbg;
 	const arch_register_class_t *cls;
 	const be_chordal_env_t *chordal_env;
 	struct obstack obst;
 	set *spill_ctxs;
 	set *spills;				/**< all spill_info_t's, which must be placed */
-	pset *mem_phis;				/**< set of all special spilled phis. allocated and freed seperately */
+	pset *mem_phis;				/**< set of all special spilled phis. allocated and freed separately */
 	decide_irn_t is_mem_phi;	/**< callback func to decide if a phi needs special spilling */
 	void *data;					/**< data passed to all callbacks */
+	DEBUG_ONLY(firm_dbg_module_t *dbg;)
 };
 
 static int cmp_spillctx(const void *a, const void *b, size_t n) {
@@ -65,15 +74,17 @@ static int cmp_spillinfo(const void *x, const void *y, size_t size) {
 	return ! (xx->spilled_node == yy->spilled_node);
 }
 
-spill_env_t *be_new_spill_env(firm_dbg_module_t *dbg,
-							  const be_chordal_env_t *chordal_env,
-							  decide_irn_t is_mem_phi, void *data) {
+DEBUG_ONLY(
+void be_set_spill_env_dbg_module(spill_env_t *env, firm_dbg_module_t *dbg) {
+	env->dbg = dbg;
+}
+)
 
-	spill_env_t *env = malloc(sizeof(env[0]));
+spill_env_t *be_new_spill_env(const be_chordal_env_t *chordal_env, decide_irn_t is_mem_phi, void *data) {
+	spill_env_t *env = xmalloc(sizeof(env[0]));
 	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;
@@ -104,7 +115,7 @@ static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn)
 
 	ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
 	if(!ctx->spill) {
-		const be_main_env_t *env = senv->chordal_env->main_env;
+		const be_main_env_t *env = senv->chordal_env->birg->main_env;
 		ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
 	}
 
@@ -132,11 +143,10 @@ static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn)
 	/* if not found spill the phi */
 	if(!ctx->spill) {
 		/* build a new PhiM with dummy in-array */
-		ins  = malloc(n * sizeof(ins[0]));
+    NEW_ARR_A(ir_node *, ins, n);
 		for(i=0; i<n; ++i)
 			ins[i] = new_r_Unknown(irg, mode_M);
 		ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
-		free(ins);
 
 		/* re-wire the phiM */
 		for(i=0; i<n; ++i) {
@@ -166,7 +176,7 @@ static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
 
 static void phi_walker(ir_node *irn, void *env) {
 	spill_env_t *senv = env;
-	const arch_env_t *arch = senv->chordal_env->main_env->arch_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)) {
@@ -175,13 +185,106 @@ static void phi_walker(ir_node *irn, void *env) {
 	}
 }
 
+#ifdef REMAT
+
+#ifdef BUGGY_REMAT
+
+static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
+	int pos, max;
+
+	/* check for 'normal' spill and general remat condition */
+	if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
+		return 0;
+
+	/* check availability of original arguments */
+	if (is_Block(reloader)) {
+
+		/* we want to remat at the end of a block.
+		 * thus all arguments must be alive at the end of the block
+		 */
+		for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
+			ir_node *arg = get_irn_n(spilled, pos);
+			if (!is_live_end(reloader, arg))
+				return 0;
+		}
+
+	} else {
+
+		/* we want to remat before the insn reloader
+		 * thus an arguments is alive if
+		 *   - it interferes with the reloaders result
+		 * or
+		 *   - or it is (last-) used by reloader itself
+		 */
+		for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
+			ir_node *arg = get_irn_n(spilled, pos);
+			int i, m;
+
+			if (values_interfere(reloader, arg))
+				goto is_alive;
+
+			for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
+				ir_node *rel_arg = get_irn_n(reloader, i);
+				if (rel_arg == arg)
+					goto is_alive;
+			}
+
+			/* arg is not alive before reloader */
+			return 0;
+
+is_alive:	;
+
+		}
+
+	}
+
+	return 1;
+}
+
+#else /* BUGGY_REMAT */
+
+static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
+	const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
+
+	return get_irn_arity(spilled) == 0 &&
+		   be_is_Spill(spill) &&
+		   arch_irn_is(aenv, spilled, rematerializable);
+}
+
+#endif /* BUGGY_REMAT */
+
+static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
+	ir_node *res;
+	ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
+
+	/* recompute the value */
+	res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
+		get_irn_op(spilled),
+		get_irn_mode(spilled),
+		get_irn_arity(spilled),
+		get_irn_in(spilled));
+	copy_node_attr(spilled, res);
+
+	DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
+
+	/* insert in schedule */
+	if (is_Block(reloader)) {
+		ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
+		sched_add_after(insert, res);
+	} else {
+		sched_add_before(reloader, res);
+	}
+
+	return res;
+}
+
+#endif
+
 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
-	ir_graph *irg = senv->chordal_env->irg;
+	const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
+	ir_graph *irg          = senv->chordal_env->irg;
 	ir_node *irn;
 	spill_info_t *si;
-	struct obstack ob;
-
-	obstack_init(&ob);
 
 	/* get all special spilled phis */
 	DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
@@ -208,37 +311,37 @@ void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
 	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;
-		int n_reloads = 0;
 		ir_mode *mode = get_irn_mode(si->spilled_node);
+		pset *values = pset_new_ptr(16);
 
 		/* go through all reloads for this spill */
 		for(rld = si->reloaders; rld; rld = rld->next) {
+			ir_node *new_val;
+
 			/* the spill for this reloader */
 			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  = be_new_Reload(senv->cls, irg, bl, mode, spill);
+#ifdef REMAT
+			if (check_remat_conditions(senv, spill, si->spilled_node, rld->reloader))
+				new_val = do_remat(senv, si->spilled_node, rld->reloader);
+			else
+#endif
+				/* do a reload */
+				new_val = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
 
-			DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", reload, si->spilled_node, rld->reloader));
+			DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
+			pset_insert_ptr(values, new_val);
 			if(reload_set)
-				pset_insert_ptr(reload_set, reload);
-
-			/* remember the reaload */
-			obstack_ptr_grow(&ob, reload);
-			sched_add_before(rld->reloader, reload);
-			n_reloads++;
+				pset_insert_ptr(reload_set, new_val);
 		}
 
-		assert(n_reloads > 0);
-		reloads = obstack_finish(&ob);
-		be_introduce_copies_ignore(senv->chordal_env->dom_front, si->spilled_node,
-				n_reloads, reloads, senv->mem_phis);
-		obstack_free(&ob, reloads);
-	}
+		/* introduce copies, rewire the uses */
+		assert(pset_count(values) > 0 && "???");
+		pset_insert_ptr(values, si->spilled_node);
+		be_ssa_constr_set_ignore(senv->chordal_env->dom_front, values, senv->mem_phis);
 
-	obstack_free(&ob, NULL);
+		del_pset(values);
+	}
 
 	for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
 		int i, n;
@@ -279,18 +382,23 @@ void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, in
 
 typedef struct _spill_slot_t {
 	unsigned size;
-	unsigned offset;
+	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 */
+	DEBUG_ONLY(firm_dbg_module_t *dbg;)
 } ss_env_t;
 
 
+/**
+ * Walker: compute the spill slots
+ */
 static void compute_spill_slots_walker(ir_node *spill, void *env) {
 	ss_env_t *ssenv = env;
 	ir_node *ctx;
@@ -305,26 +413,44 @@ static void compute_spill_slots_walker(ir_node *spill, void *env) {
 	entry = pmap_find(ssenv->slots, ctx);
 
 	if (!entry) {
+		struct _arch_env_t *arch_env     = ssenv->cenv->birg->main_env->arch_env;
+		const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
+		ir_mode *largest_mode            = arch_register_class_mode(cls);
+
 		/* this is a new spill context */
 		ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
-		ss->members = pset_new_ptr(8);
-		ss->size = get_mode_size_bytes(get_irn_mode(get_irn_n(spill, 0)));
+		ss->members      = pset_new_ptr(8);
+		ss->largest_mode = largest_mode;
+		ss->size         = get_mode_size_bytes(ss->largest_mode);
+		ss->align        = arch_isa_get_reg_class_alignment(arch_env->isa, cls);
 		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");
-		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!");
+
+#ifndef NDEBUG
+		/* ugly mega assert :-) */
+		{
+			ir_node *irn;
+			struct _arch_env_t *arch_env     = ssenv->cenv->birg->main_env->arch_env;
+			const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
+			int size = get_mode_size_bytes(arch_register_class_mode(cls));
+			assert(ss->size == size && "Different sizes for the same spill slot are not allowed.");
+			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!");
+			}
 		}
+#endif /* NDEBUG */
 	}
 
 	pset_insert_ptr(ss->members, spill);
 }
 
+/**
+ * qsort compare function, sort spill slots by size.
+ */
 static int ss_sorter(const void *v1, const void *v2) {
 	const spill_slot_t *ss1 = v1;
 	const spill_slot_t *ss2 = v2;
@@ -332,19 +458,20 @@ static int ss_sorter(const void *v1, const void *v2) {
 }
 
 
-/* NOTE/TODO: This function assumes, that all spill slot sizes are a power of 2.
-   Further it assumes, that the alignment is equal to the size and the
-   baseaddr of the spill area is aligned sufficiently for all possible aligments.
-*/
-static void coalesce_slots(ss_env_t *ssenv) {
+/**
+ * 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;
-	unsigned curr_offset;
 	pmap_entry *entr;
-	spill_slot_t **ass;
-
-	/* Build an array of all spill slots */
-	int count = pmap_count(ssenv->slots);
-	ass = obstack_alloc(&ssenv->ob, count * sizeof(*ass));
 
 	i=0;
 	pmap_foreach(ssenv->slots, entr)
@@ -352,14 +479,13 @@ static void coalesce_slots(ss_env_t *ssenv) {
 
 	/* Sort the array to minimize fragmentation and cache footprint.
 	   Large slots come first */
-	qsort(ass, count, sizeof(ass[0]), ss_sorter);
+	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 */
-	curr_offset = 0;
 	used_slots = 0;
-	for (i=0; i<count; ++i) { /* for each spill slot */
+	for (i=0; i<size; ++i) { /* for each spill slot */
 		ir_node *n1;
 		int tgt_slot = -1;
 
@@ -393,9 +519,6 @@ interf_detected: /*nothing*/ ;
 			tgt_slot = used_slots;
 			used_slots++;
 
-			ass[tgt_slot]->offset = curr_offset;
-			curr_offset += ass[i]->size;
-
 			/* init slot */
 			if (tgt_slot != i) {
 				ass[tgt_slot]->size = ass[i]->size;
@@ -405,33 +528,119 @@ interf_detected: /*nothing*/ ;
 		}
 
 		/* copy the members to the target pset */
-		for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members)) {
-			/* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
-			if (tgt_slot != i)
-				pset_insert_ptr(ass[tgt_slot]->members, n1);
+		/* 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);
+	}
+}
 
-			be_set_Spill_offset(n1, ass[tgt_slot]->offset);
-			DBG((ssenv->dbg, LEVEL_1, "    Offset %+F  %d\n", n1, ass[tgt_slot]->offset));
-		}
+#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.
+ *
+ * @param types  a map containing all created types
+ * @param ss     the spill slot
+ *
+ * Note that type types should are identical for every mode.
+ * This rule might break if two different register classes return the same
+ * mode but different alignments.
+ */
+static ir_type *get_spill_type(pmap *types, spill_slot_t *ss) {
+  pmap_entry *e = pmap_find(types, ss->largest_mode);
+  ir_type *res;
+
+  if (! e) {
+		char buf[64];
+    snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(ss->largest_mode));
+    res = new_type_primitive(new_id_from_str(buf), ss->largest_mode);
+		set_type_alignment_bytes(res, ss->align);
+    pmap_insert(types, ss->largest_mode, res);
+  }
+  else {
+    res = e->value;
+		assert(get_type_alignment_bytes(res) == (int)ss->align);
 	}
+  return res;
+}
+
+/**
+ * Create spill slot entities on the frame type.
+ *
+ * @param ssenv   the spill environment
+ * @param n       number of spill slots
+ * @param ss      array of spill slots
+ */
+static void assign_entities(ss_env_t *ssenv, int n_slots, spill_slot_t *ss[]) {
+	int i, offset, frame_align;
+	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_slots; ++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);
 
-	/* free all used psets, all other stuff is on the ssenv-obstack */
-	for (i=0; i<count; ++i)
-		del_pset(ass[i]->members);
+		spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]));
 
+		/* align */
+		offset = round_up2(offset, ss[i]->align);
+		/* set */
+		set_entity_offset_bytes(spill_ent, offset);
+		/* next possible offset */
+		offset += round_up2(ss[i]->size, ss[i]->align);
+
+		pset_foreach(ss[i]->members, irn)
+			be_set_Spill_entity(irn, spill_ent);
+	}
+
+	/* set final size of stack frame */
+	frame_align = get_type_alignment_bytes(frame);
+	set_type_size_bytes(frame, round_up2(offset, frame_align));
 }
 
 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.dbg   = firm_dbg_register("ir.be.spillslots");
+	ssenv.types = pmap_create();
+	FIRM_DBG_REGISTER(ssenv.dbg, "ir.be.spillslots");
 
+	/* Get initial spill slots */
 	irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
-	coalesce_slots(&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);
+
+	be_copy_entities_to_reloads(cenv->irg);
 }