X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fifconv.c;h=f77a75dc21eb24aefc939914968e3a67c3c9dc0b;hb=edc32d1a5396331f871a0663806acb45895a1ef8;hp=a9b06fdb05434d8a01f8371169c85f47f67a09af;hpb=8c2ff434c07cfcf327664a474661db20a0cdbfb9;p=libfirm

diff --git a/ir/opt/ifconv.c b/ir/opt/ifconv.c
index a9b06fdb0..f77a75dc2 100644
--- a/ir/opt/ifconv.c
+++ b/ir/opt/ifconv.c
@@ -9,6 +9,591 @@
  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
  */
 
+#if 1
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
+#ifdef HAVE_ALLOCA_H
+#include <alloca.h>
+#endif
+
+#include <assert.h>
+
+#include "obst.h"
+#include "irnode_t.h"
+#include "cdep.h"
+#include "ircons.h"
+#include "ifconv.h"
+#include "irdom.h"
+#include "irgmod.h"
+#include "irgopt.h"
+#include "irgwalk.h"
+#include "irtools.h"
+#include "return.h"
+#include "array.h"
+
+// debug
+#include "irdump.h"
+#include "debug.h"
+
+DEBUG_ONLY(firm_dbg_module_t *dbg);
+
+static ir_node* walk_to_projx(ir_node* start)
+{
+	ir_node* pred;
+
+	pred = get_nodes_block(start);
+
+	/* if there are multiple control flow predecessors nothing sensible can be
+	 * done */
+	if (get_irn_arity(pred) > 1) return NULL;
+
+	pred = get_irn_n(pred, 0);
+	if (get_irn_op(pred) == op_Proj) {
+		assert(get_irn_mode(pred) == mode_X);
+		return pred;
+	} else {
+		return NULL;
+	}
+}
+
+/**
+ * Additional block info.
+ */
+typedef struct block_info {
+	ir_node *phi;   /**< head of the Phi list */
+	int has_pinned; /**< set if the block contains instructions that cannot be moved */
+} block_info;
+
+#define get_block_blockinfo(block) ((block_info *)get_irn_link(block))
+
+/**
+ * Returns non-zero if a Block can be emptied.
+ */
+static int can_empty_block(ir_node *block)
+{
+	return !get_block_blockinfo(block)->has_pinned;
+}
+
+
+/**
+ * Copies the DAG starting at node to the ith predecessor block of src_block
+ * -if the node isn't in the src_block, this is a nop and the node is returned
+ * -if the node is a phi in the src_block, the ith predecessor of the phi is
+ *   returned
+ * otherwise returns the copy of the passed node
+ */
+static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
+{
+	ir_node* dst_block;
+	ir_node* copy;
+	int arity;
+	int j;
+
+	if (get_nodes_block(node) != src_block) return node;
+	if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
+
+	copy_irn_to_irg(node, current_ir_graph);
+	copy = get_irn_link(node);
+	dst_block = get_nodes_block(get_irn_n(src_block, i));
+	set_nodes_block(copy, dst_block);
+
+	DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
+		node, dst_block, copy));
+
+	arity = get_irn_arity(node);
+	for (j = 0; j < arity; ++j) {
+		set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
+		DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
+	}
+	return copy;
+}
+
+
+/**
+ * Duplicate and move the contents of ith block predecessor into its
+ * predecessors if the block has multiple control dependencies and only one
+ * successor.
+ * Also bail out if the block contains non-movable nodes, because later
+ * if-conversion would be pointless.
+ */
+static int fission_block(ir_node* block, int i)
+{
+	ir_node* pred = get_irn_n(block, i);
+	ir_node* pred_block;
+	block_info* info;
+	ir_node* phi;
+	int pred_arity;
+	int arity;
+	ir_node** ins;
+	int j;
+
+	if (get_irn_op(pred) != op_Jmp) return 0;
+	pred_block = get_nodes_block(pred);
+
+	if (!has_multiple_cdep(pred_block)) return 0;
+	if (!can_empty_block(pred_block)) return 0;
+
+	DB((dbg, LEVEL_1, "Fissioning block %+F\n", pred_block));
+
+	pred_arity = get_irn_arity(pred_block);
+	arity = get_irn_arity(block);
+	info = get_block_blockinfo(block);
+	NEW_ARR_A(ir_node *, ins, arity + pred_arity - 1);
+	for (phi = info->phi; phi != NULL; phi = get_irn_link(phi)) {
+		for (j = 0; j < i; ++j) ins[j] = get_irn_n(phi, j);
+		for (j = 0; j < pred_arity; ++j) {
+			ins[i + j] = copy_to(get_irn_n(phi, i), pred_block, j);
+		}
+		for (j = i + 1; j < arity; ++j) {
+			ins[pred_arity - 1 + j] = get_irn_n(phi, j);
+		}
+		set_irn_in(phi, arity + pred_arity - 1, ins);
+	}
+	for (j = 0; j < i; ++j) ins[j] = get_irn_n(block, j);
+	for (j = 0; j < pred_arity; ++j) ins[i + j] = get_irn_n(pred_block, j);
+	for (j = i + 1; j < arity; ++j) ins[pred_arity - 1 + j] = get_irn_n(block, j);
+	set_irn_in(block, arity + pred_arity - 1, ins);
+
+	/* Kill all Phis in the fissioned block
+	 * This is to make sure they're not kept alive
+	 */
+	info = get_block_blockinfo(pred_block);
+	phi = info->phi;
+	while (phi != NULL) {
+		ir_node* next = get_irn_link(phi);
+		exchange(phi, new_Bad());
+		phi = next;
+	}
+	return 1;
+}
+
+
+/**
+ * Remove predecessors i and j from node and add predecessor new_pred
+ */
+static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
+{
+	int arity = get_irn_arity(node);
+	ir_node **ins;
+	int k;
+	int l;
+
+	NEW_ARR_A(ir_node *, ins, arity - 1);
+
+	l = 0;
+	for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
+	for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
+	for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
+	ins[l++] = new_pred;
+	assert(l == arity - 1);
+	set_irn_in(node, l, ins);
+}
+
+
+static void if_conv_walker(ir_node* block, void* env)
+{
+	ir_node* phi;
+	int arity;
+	int i;
+
+	/* Bail out, if there are no Phis at all */
+	if (get_block_blockinfo(block)->phi == NULL) return;
+
+restart:
+	arity = get_irn_arity(block);
+	for (i = 0; i < arity; ++i) {
+		if (fission_block(block, i)) goto restart;
+	}
+	//return;
+
+	arity = get_irn_arity(block);
+	for (i = 0; i < arity; ++i) {
+		ir_node* pred;
+		ir_node* cond;
+		ir_node* projx0;
+		int j;
+
+		projx0 = walk_to_projx(get_irn_n(block, i));
+		if (projx0 == NULL) return;
+		pred = get_Proj_pred(projx0);
+		if (get_irn_op(pred) != op_Cond || get_irn_mode(get_Cond_selector(pred)) != mode_b) continue;
+		cond = pred;
+
+		if (!can_empty_block(get_nodes_block(get_irn_n(block, i)))) {
+			DB((dbg, LEVEL_1, "Cannot empty block %+F\n",
+				get_nodes_block(get_irn_n(block, i))
+			));
+			continue;
+		}
+
+		for (j = i + 1; j < arity; ++j) {
+			ir_node* projx1;
+			ir_node* psi_block;
+			ir_node* conds[1];
+			ir_node* vals[2];
+			ir_node* psi;
+
+			projx1 = walk_to_projx(get_irn_n(block, j));
+			if (projx1 == NULL) continue;
+			pred = get_Proj_pred(projx1);
+			if (get_irn_op(pred) != op_Cond || get_irn_mode(get_Cond_selector(pred)) != mode_b) continue;
+			if (pred != cond) continue;
+			DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n", cond, projx0, projx1));
+
+			if (!can_empty_block(get_nodes_block(get_irn_n(block, j)))) {
+				DB((dbg, LEVEL_1, "Cannot empty %+F\n",	get_nodes_block(get_irn_n(block, j))));
+				continue;
+			}
+
+			conds[0] = get_Cond_selector(cond);
+
+			psi_block = get_nodes_block(cond);
+			phi = get_block_blockinfo(block)->phi;
+			do {
+				ir_node* val_i = get_irn_n(phi, i);
+				ir_node* val_j = get_irn_n(phi, j);
+
+				if (val_i == val_j) {
+					psi = val_i;
+					DB((dbg, LEVEL_2,  "Generating no psi, because both values are equal\n"));
+				} else {
+					/* Something is very fishy if two predecessors of a PhiM point into
+					 * one block, but not at the same memory node
+					 */
+					assert(get_irn_mode(phi) != mode_M);
+					if (get_Proj_proj(projx0) == pn_Cond_true) {
+						vals[0] = val_i;
+						vals[1] = val_j;
+					} else {
+						vals[0] = val_j;
+						vals[1] = val_i;
+					}
+					psi = new_r_Psi(
+						current_ir_graph, psi_block, 1, conds, vals, get_irn_mode(phi)
+					);
+					DB((dbg, LEVEL_2, "Generating %+F for %+F\n", psi, phi));
+				}
+
+				if (arity == 2) {
+					exchange(phi, psi);
+				} else {
+					rewire(phi, i, j, psi);
+				}
+
+				phi = get_irn_link(phi);
+			} while (phi != NULL);
+
+			exchange(get_nodes_block(get_irn_n(block, i)), psi_block);
+			exchange(get_nodes_block(get_irn_n(block, j)), psi_block);
+
+			if (arity == 2) {
+				DB((dbg, LEVEL_1,  "Welding block %+F to %+F\n", block, psi_block));
+				get_block_blockinfo(psi_block)->has_pinned |=	get_block_blockinfo(block)->has_pinned;
+				exchange(block, psi_block);
+				return;
+			} else {
+				rewire(block, i, j, new_r_Jmp(current_ir_graph, psi_block));
+				goto restart;
+			}
+		}
+	}
+}
+
+/**
+ * Block walker: add additional data
+ */
+static void init_block_link(ir_node *block, void *env)
+{
+	struct obstack *obst = env;
+	block_info *bi = obstack_alloc(obst, sizeof(*bi));
+
+	bi->phi = NULL;
+	bi->has_pinned = 0;
+	set_irn_link(block, bi);
+}
+
+
+/**
+ * Daisy-chain all phis in a block
+ * If a non-movable node is encountered set the has_pinned flag
+ */
+static void collect_phis(ir_node *node, void *env)
+{
+	if (is_Phi(node)) {
+		ir_node *block = get_nodes_block(node);
+		block_info *bi = get_block_blockinfo(block);
+
+		set_irn_link(node, bi->phi);
+		bi->phi = node;
+	}
+	else {
+		if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
+			/*
+			 * Ignore control flow nodes, these will be removed.
+			 * This ignores Raise. That is surely bad. FIXME.
+			 */
+			if (! is_cfop(node)) {
+				ir_node *block = get_nodes_block(node);
+				block_info *bi = get_block_blockinfo(block);
+
+				DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
+				bi->has_pinned = 1;
+			}
+		}
+	}
+}
+
+
+/*
+ * Transform multiple cascaded Psis into one Psi
+ */
+static ir_node* fold_psi(ir_node* psi)
+{
+	int arity = get_Psi_n_conds(psi);
+	int new_arity = 0;
+	int i;
+	ir_node* n;
+	ir_node** conds;
+	ir_node** vals;
+	int j;
+	int k;
+	int a;
+	ir_node* new_psi;
+
+	for (i = 0; i < arity; ++i) {
+		n = get_Psi_val(psi, i);
+		if (get_irn_op(n) == op_Psi) {
+			new_arity += get_Psi_n_conds(n) + 1;
+		} else {
+			++new_arity;
+		}
+	}
+	n = get_Psi_default(psi);
+	if (get_irn_op(n) == op_Psi) {
+		new_arity += get_Psi_n_conds(n);
+	}
+
+	if (arity == new_arity) return psi; // no attached Psis found
+	DB((dbg, LEVEL_1, "Folding %+F from %d to %d conds\n", psi, arity, new_arity));
+
+	NEW_ARR_A(ir_node *, conds, new_arity);
+	NEW_ARR_A(ir_node *, vals, new_arity + 1);
+	j = 0;
+	for (i = 0; i < arity; ++i) {
+		ir_node* c = get_Psi_cond(psi, i);
+
+		n = get_Psi_val(psi, i);
+		if (get_irn_op(n) == op_Psi) {
+			a = get_Psi_n_conds(n);
+			for (k = 0; k < a; ++k) {
+				conds[j] = new_r_And(
+					current_ir_graph, get_nodes_block(psi),
+					c, get_Psi_cond(n, k), mode_b
+				);
+				vals[j] = get_Psi_val(n, k);
+				++j;
+			}
+			conds[j] = c;
+			vals[j] = get_Psi_default(n);
+		} else {
+			conds[j] = c;
+			vals[j] = n;
+		}
+		++j;
+	}
+	n = get_Psi_default(psi);
+	if (get_irn_op(n) == op_Psi) {
+		a = get_Psi_n_conds(n);
+		for (k = 0; k < a; ++k) {
+			conds[j] = get_Psi_cond(n, k);
+			vals[j] = get_Psi_val(n, k);
+			++j;
+		}
+		vals[j] = get_Psi_default(n);
+	} else {
+		vals[j] = n;
+	}
+	assert(j == new_arity);
+	new_psi = new_r_Psi(
+		current_ir_graph, get_nodes_block(psi),
+		new_arity, conds, vals, get_irn_mode(psi)
+	);
+	DB((dbg, LEVEL_1, "Folded %+F into new %+F\n", psi, new_psi));
+	exchange(psi, new_psi);
+	return new_psi;
+}
+
+
+/*
+ * Merge consecutive psi inputs if the data inputs are the same
+ */
+static ir_node* meld_psi(ir_node* psi)
+{
+	int arity = get_Psi_n_conds(psi);
+	int new_arity;
+	ir_node** conds;
+	ir_node** vals;
+	ir_node* cond;
+	ir_node* val;
+	int i;
+	int j;
+	ir_node* new_psi;
+
+	new_arity = 1;
+	val = get_Psi_val(psi, 0);
+	DB((dbg, LEVEL_1, "Pred  0 of %+F is %+F\n", psi, val));
+	for (i = 1; i < arity; ++i) {
+		ir_node* v = get_Psi_val(psi, i);
+		DB((dbg, LEVEL_1, "Pred %2d of %+F is %+F\n", i, psi, v));
+		if (val != v) {
+			val = v;
+			++new_arity;
+		}
+	}
+	DB((dbg, LEVEL_1, "Default of %+F is %+F\n", psi, get_Psi_default(psi)));
+	if (val == get_Psi_default(psi)) --new_arity;
+
+	DB((dbg, LEVEL_1, "Melding Psi %+F from %d conds to %d\n", psi, arity, new_arity));
+
+	if (new_arity == arity) return psi;
+
+	/* If all data inputs of the Psi are equal, exchange the Psi with that value */
+	if (new_arity == 0) {
+		exchange(psi, val);
+		return val;
+	}
+
+	NEW_ARR_A(ir_node *, conds, new_arity);
+	NEW_ARR_A(ir_node *, vals, new_arity + 1);
+	cond = get_Psi_cond(psi, 0);
+	val = get_Psi_val(psi, 0);
+	j = 0;
+	for (i = 1; i < arity; ++i) {
+		ir_node* v = get_Psi_val(psi, i);
+
+		if (v == val) {
+			cond = new_r_Or(
+				current_ir_graph, get_nodes_block(psi),
+				cond, get_Psi_cond(psi, i), mode_b
+			);
+		} else {
+			conds[j] = cond;
+			vals[j] = val;
+			++j;
+			val = v;
+		}
+	}
+	if (val != get_Psi_default(psi)) {
+		conds[j] = cond;
+		vals[j] = val;
+		++j;
+	}
+	vals[j] = get_Psi_default(psi);
+	assert(j == new_arity);
+	new_psi = new_r_Psi(
+		current_ir_graph, get_nodes_block(psi),
+		new_arity, conds, vals, get_irn_mode(psi)
+	);
+	DB((dbg, LEVEL_1, "Molded %+F into %+F\n", psi, new_psi));
+	exchange(psi, new_psi);
+	return new_psi;
+}
+
+
+/**
+ * Split a Psi with multiple conditions into multiple Psis with one condtition
+ * each
+ */
+static ir_node* split_psi(ir_node* psi)
+{
+	int arity = get_Psi_n_conds(psi);
+	ir_mode* mode;
+	ir_node* block;
+	ir_node* rval;
+	int i;
+
+	if (arity == 1) return psi;
+
+	mode = get_irn_mode(psi);
+	block = get_nodes_block(psi);
+	rval = get_Psi_default(psi);
+	for (i = arity - 1; i >= 0; --i) {
+		ir_node* conds[1];
+		ir_node* vals[2];
+
+		conds[0] = get_Psi_cond(psi, i);
+		vals[0] = get_Psi_val(psi, i);
+		vals[1] = rval;
+		rval = new_r_Psi(
+			current_ir_graph, block, 1, conds, vals, mode
+		);
+	}
+	exchange(psi, rval);
+	return rval;
+}
+
+
+static void optimise_psis(ir_node* node, void* env)
+{
+	if (get_irn_op(node) != op_Psi) return;
+#if 1
+	node = fold_psi(node);
+#endif
+#if 1
+	node = meld_psi(node);
+#endif
+#if 1
+	node = split_psi(node);
+#endif
+}
+
+
+void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params)
+{
+	struct obstack obst;
+
+	if (!get_opt_if_conversion())
+		return;
+
+	FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
+
+	DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
+
+	dump_ir_block_graph(irg, "_00_pre");
+
+	normalize_one_return(irg);
+	remove_critical_cf_edges(irg);
+
+	dump_ir_block_graph(irg, "_01_normal");
+
+	compute_cdep(irg);
+	assure_doms(irg);
+
+	obstack_init(&obst);
+	irg_block_walk_graph(irg, init_block_link, NULL, &obst);
+	irg_walk_graph(irg, collect_phis, NULL, NULL);
+	irg_block_walk_graph(irg, NULL, if_conv_walker, NULL);
+
+	local_optimize_graph(irg);
+	dump_ir_block_graph(irg, "_02_ifconv");
+
+	irg_walk_graph(irg, NULL, optimise_psis, NULL);
+
+	dump_ir_block_graph(irg, "_03_postifconv");
+
+	obstack_free(&obst, NULL);
+
+	free_dom(irg);
+	free_cdep(irg);
+}
+
+#else
+
 /**
  * @file ifconv.c
  * If conversion.
@@ -29,9 +614,6 @@
 #ifdef HAVE_ALLOCA_H
 #include <alloca.h>
 #endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
 
 #include "irgraph_t.h"
 #include "irnode_t.h"
@@ -840,7 +1422,7 @@ static void annotate_cond_info_post(ir_node *irn, void *data)
 
       /*
        * Add this cond info to the list of all cond infos
-       * in this graph. This is just done to free the
+       * in this graph. This is just done to xfree the
        * set easier afterwards (we save an irg_walk_graph).
        */
       list_add(&cwi->cond_info_head, &ci->list);
@@ -964,7 +1546,7 @@ void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params)
   normalize_one_return(irg);
 
   /* Ensure, that the dominators are computed. */
-  compute_doms(irg);
+  assure_doms(irg);
 
   DBG((dbg, LEVEL_1, "if conversion for irg %s(%p)\n",
         get_entity_name(get_irg_entity(irg)), irg));
@@ -995,4 +1577,8 @@ void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params)
   DBG((dbg, LEVEL_1, "muxes made: %d\n", muxes_made));
 
   obstack_free(&obst, NULL);
+
+	dump_ir_block_graph(irg, "_ifconv_hack");
 }
+
+#endif