X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Flower%2Flower_switch.c;h=4ff2737980ed18c831b469641b7f8b3b0550e970;hb=0e4318adc38a54110c1b1f7f1d5edc0ee2c78170;hp=1c794554224d08a59c3d6ebae2413418c197fa49;hpb=014476c446ab6fd72461c764c018fa401e5f88c1;p=libfirm

diff --git a/ir/lower/lower_switch.c b/ir/lower/lower_switch.c
index 1c7945542..4ff273798 100644
--- a/ir/lower/lower_switch.c
+++ b/ir/lower/lower_switch.c
@@ -1,5 +1,5 @@
 /*
- * 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.
  *
@@ -23,204 +23,360 @@
  * @author  Moritz Kroll
  * @version $Id$
  */
-
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include <limits.h>
+#include <stdbool.h>
 
+#include "array_t.h"
 #include "ircons.h"
 #include "irgopt.h"
 #include "irgwalk.h"
 #include "irnode_t.h"
 #include "irouts.h"
+#include "irpass_t.h"
+#include "lowering.h"
+#include "error.h"
+#include "irnodeset.h"
 
-#define foreach_out_irn(irn, i, outirn) for(i = get_irn_n_outs(irn) - 1;\
+#define foreach_out_irn(irn, i, outirn) for (i = get_irn_n_outs(irn) - 1;\
 	i >= 0 && (outirn = get_irn_out(irn, i)); --i)
 
-typedef struct walk_env {
-	unsigned         spare_size;        /**< the allowed spare size for table switches */
-	struct obstack   obst;              /**< the obstack where data is allocated on */
-	int              changed;           /**< indicates whether a change was performed */
+typedef struct walk_env_t {
+	unsigned      spare_size; /**< the allowed spare size for table switches */
+	unsigned      small_switch;
+	bool          allow_out_of_bounds;
+	bool          changed;    /**< indicates whether a change was performed */
+	ir_nodeset_t  processed;
 } walk_env_t;
 
-typedef struct case_data {
+typedef struct case_data_t {
 	long     value;
 	ir_node *target;
 } case_data_t;
 
-typedef struct ifcas_env {
+typedef struct cond_env_t {
 	ir_node  *sel;
-	int       defindex;
-	ir_node **defusers;                 /**< the Projs pointing to the default case */
-} ifcas_env_t;
-
-/**
- * Evaluate a switch and decide whether we should build a table switch.
- *
- * @param cond       The Cond node representing the switch.
- * @param spare_size Allowed spare size for table switches in machine words.
- *                   (Default in edgfe: 128)
- */
-static int should_do_table_switch(ir_node *cond, unsigned spare_size)
+	long      switch_min;
+	long      switch_max;
+	ir_node  *default_block;
+	unsigned  num_cases;
+	ir_node **defusers;           /**< the Projs pointing to the default case */
+} cond_env_t;
+
+static void analyse_switch(cond_env_t *env, ir_node *cond)
 {
-	long     default_pn;
+	long     default_pn    = get_Cond_default_proj(cond);
+	long     switch_min    = LONG_MAX;
+	long     switch_max    = LONG_MIN;
+	ir_node *default_block = NULL;
+	unsigned num_cases     = 0;
 	int      i;
 	ir_node *proj;
-	long switch_min = LONG_MAX, switch_max = LONG_MIN;
-	unsigned long spare, num_cases = 0;
-
-	/* TODO: Minimum size for jump table? */
-	if(get_irn_n_outs(cond) <= 4)
-		return 0;
-
-	default_pn = get_Cond_defaultProj(cond);
 
 	foreach_out_irn(cond, i, proj) {
 		long pn = get_Proj_proj(proj);
-		if(pn == default_pn)
+		if (pn == default_pn) {
+			ir_node *target = get_irn_out(proj, 0);
+			assert(default_block == NULL || default_block == target);
+			default_block = target;
 			continue;
+		}
 
-		if(pn < switch_min)
+		if (pn < switch_min)
 			switch_min = pn;
-		if(pn > switch_max)
+		if (pn > switch_max)
 			switch_max = pn;
-		num_cases++;
+		++num_cases;
 	}
+	assert(default_block != NULL);
 
-	/*
-	 * Here we have: num_cases and [switch_min, switch_max] interval.
-	 * We do an if-cascade if there are too many spare numbers.
-	 */
-	spare = (unsigned long) switch_max - (unsigned long) switch_min - num_cases + 1;
-	return spare < spare_size;
+	env->switch_min    = switch_min;
+	env->switch_max    = switch_max;
+	env->num_cases     = num_cases;
+	env->default_block = default_block;
 }
 
 static int casecmp(const void *a, const void *b)
 {
-	return ((case_data_t *) a)->value - ((case_data_t *) b)->value;
+	const case_data_t *cda = (const case_data_t*)a;
+	const case_data_t *cdb = (const case_data_t*)b;
+
+	/*
+	 * Enforce unsigned sorting. Signed comparison will behave differently for
+	 * 32-bit values, depending on sizeof(long). This will make the resulting
+	 * array deterministic.
+	 */
+	return ((unsigned long)cda->value > (unsigned long)cdb->value) -
+	       ((unsigned long)cda->value < (unsigned long)cdb->value);
 }
 
 /**
  * Creates an if cascade realizing binary search.
  */
-static void create_if_cascade(ifcas_env_t *env, ir_node *curblock,
-                              case_data_t *curcases, int numcases)
+static void create_if_cascade(cond_env_t *env, dbg_info *dbgi, ir_node *block,
+                              case_data_t *curcases, unsigned numcases)
 {
-	set_cur_block(curblock);
-
-	if(numcases == 1)
-	{
-		/* only one case: "if(sel == val) goto target else goto default;" */
-		ir_node *val  = new_Const_long(get_irn_mode(env->sel), curcases[0].value);
-		ir_node *cmp  = new_Cmp(env->sel, val);
-		ir_node *proj = new_Proj(cmp, mode_b, pn_Cmp_Eq);
-		ir_node *cond = new_Cond(proj);
-		set_Block_cfgpred(curcases[0].target, 0, new_Proj(cond, mode_X, pn_Cond_true));
-		env->defusers[env->defindex++] = new_Proj(cond, mode_X, pn_Cond_false);
-	} else if(numcases == 2) {
-		/* only two cases: "if(sel == val[0]) goto target[0];" */
-		ir_node *val  = new_Const_long(get_irn_mode(env->sel), curcases[0].value);
-		ir_node *cmp  = new_Cmp(env->sel, val);
-		ir_node *proj = new_Proj(cmp, mode_b, pn_Cmp_Eq);
-		ir_node *cond = new_Cond(proj);
+	ir_graph *irg = get_irn_irg(block);
+    ir_mode  *cmp_mode;
+    ir_node  *cmp_sel;
+    ir_node  *sel_block;
+
+    /* Get the mode and sel node for the comparison. */
+    cmp_mode  = get_irn_mode(env->sel);
+    cmp_sel   = env->sel;
+    sel_block = get_nodes_block(cmp_sel);
+
+    /*
+     * Make sure that an unsigned comparison is used, by converting the sel
+     * node to an unsigned mode and using that mode for the constants, too.
+     * This is important, because the qsort applied to the case labels uses
+     * an unsigned comparison and both comparison methods have to match.
+     */
+    if (mode_is_signed(cmp_mode)) {
+        cmp_mode = find_unsigned_mode(cmp_mode);
+        cmp_sel  = new_r_Conv(sel_block, cmp_sel, cmp_mode);
+    }
+
+	if (numcases == 0) {
+		/* zero cases: "goto default;" */
+		ARR_APP1(ir_node*, env->defusers, new_r_Jmp(block));
+	} else if (numcases == 1) {
+		/* only one case: "if (sel == val) goto target else goto default;" */
+		ir_node *val       = new_r_Const_long(irg, cmp_mode, curcases[0].value);
+		ir_node *cmp       = new_rd_Cmp(dbgi, block, cmp_sel, val,
+		                                ir_relation_equal);
+		ir_node *cond      = new_rd_Cond(dbgi, block, cmp);
+		ir_node *trueproj  = new_r_Proj(cond, mode_X, pn_Cond_true);
+		ir_node *falseproj = new_r_Proj(cond, mode_X, pn_Cond_false);
+
+		set_Block_cfgpred(curcases[0].target, 0, trueproj);
+		ARR_APP1(ir_node*, env->defusers, falseproj);
+	} else if (numcases == 2) {
+		/* only two cases: "if (sel == val[0]) goto target[0];" */
+		ir_node *val       = new_r_Const_long(irg, cmp_mode, curcases[0].value);
+		ir_node *cmp       = new_rd_Cmp(dbgi, block, cmp_sel, val,
+		                                ir_relation_equal);
+		ir_node *cond      = new_rd_Cond(dbgi, block, cmp);
+		ir_node *trueproj  = new_r_Proj(cond, mode_X, pn_Cond_true);
+		ir_node *falseproj = new_r_Proj(cond, mode_X, pn_Cond_false);
 		ir_node *in[1];
 		ir_node *neblock;
 
-		set_Block_cfgpred(curcases[0].target, 0, new_Proj(cond, mode_X, pn_Cond_true));
-		in[0] = new_Proj(cond, mode_X, pn_Cond_false);
-		neblock = new_Block(1, in);
-
-		/* second part: "else if(sel == val[1]) goto target[1] else goto default;" */
-		val  = new_Const_long(get_irn_mode(env->sel), curcases[1].value);
-		cmp  = new_Cmp(env->sel, val);
-		proj = new_Proj(cmp, mode_b, pn_Cmp_Eq);
-		cond = new_Cond(proj);
-		set_Block_cfgpred(curcases[1].target, 0, new_Proj(cond, mode_X, pn_Cond_true));
-		env->defusers[env->defindex++] = new_Proj(cond, mode_X, pn_Cond_false);
+		set_Block_cfgpred(curcases[0].target, 0, trueproj);
+
+		in[0] = falseproj;
+		neblock = new_r_Block(irg, 1, in);
+
+		/* second part: "else if (sel == val[1]) goto target[1] else goto default;" */
+		val       = new_r_Const_long(irg, cmp_mode, curcases[1].value);
+		cmp       = new_rd_Cmp(dbgi, neblock, cmp_sel, val, ir_relation_equal);
+		cond      = new_rd_Cond(dbgi, neblock, cmp);
+		trueproj  = new_r_Proj(cond, mode_X, pn_Cond_true);
+		falseproj = new_r_Proj(cond, mode_X, pn_Cond_false);
+		set_Block_cfgpred(curcases[1].target, 0, trueproj);
+		ARR_APP1(ir_node*, env->defusers, falseproj);
 	} else {
 		/* recursive case: split cases in the middle */
 		int midcase = numcases / 2;
-		ir_node *val  = new_Const_long(get_irn_mode(env->sel), curcases[midcase].value);
-		ir_node *cmp  = new_Cmp(env->sel, val);
-		ir_node *proj = new_Proj(cmp, mode_b, pn_Cmp_Lt);
-		ir_node *cond = new_Cond(proj);
+		ir_node *val  = new_r_Const_long(irg, cmp_mode,
+		                                 curcases[midcase].value);
+		ir_node *cmp  = new_rd_Cmp(dbgi, block, cmp_sel, val, ir_relation_less);
+		ir_node *cond = new_rd_Cond(dbgi, block, cmp);
 		ir_node *in[1];
 		ir_node *ltblock;
 		ir_node *geblock;
 
-		in[0] = new_Proj(cond, mode_X, pn_Cond_true);
-		ltblock = new_Block(1, in);
+		in[0]   = new_r_Proj(cond, mode_X, pn_Cond_true);
+		ltblock = new_r_Block(irg, 1, in);
 
-		set_cur_block(curblock);
-		in[0] = new_Proj(cond, mode_X, pn_Cond_false);
-		geblock = new_Block(1, in);
+		in[0]   = new_r_Proj(cond, mode_X, pn_Cond_false);
+		geblock = new_r_Block(irg, 1, in);
 
-		create_if_cascade(env, ltblock, curcases, midcase);
-		create_if_cascade(env, geblock, curcases + midcase, numcases - midcase);
+		create_if_cascade(env, dbgi, ltblock, curcases, midcase);
+		create_if_cascade(env, dbgi, geblock, curcases + midcase,
+		                  numcases - midcase);
 	}
 }
 
+static void create_out_of_bounds_check(cond_env_t *env, ir_node *cond)
+{
+	ir_graph      *irg           = get_irn_irg(cond);
+	dbg_info      *dbgi          = get_irn_dbg_info(cond);
+	ir_node       *sel           = get_Cond_selector(cond);
+	ir_node       *block         = get_nodes_block(cond);
+	ir_mode       *cmp_mode      = get_irn_mode(sel);
+	ir_node      **default_preds = NEW_ARR_F(ir_node*, 0);
+	long           default_pn    = get_Cond_default_proj(cond);
+	long           delta         = 0;
+	ir_node       *max_const;
+	ir_node       *proj_true;
+	ir_node       *proj_false;
+	ir_node       *cmp;
+	ir_node       *oob_cond;
+	ir_node       *in[1];
+	ir_node       *new_block;
+	size_t         n_default_preds;
+	int            i;
+	ir_node       *proj;
+
+	if (mode_is_signed(cmp_mode)) {
+		cmp_mode = find_unsigned_mode(cmp_mode);
+		sel      = new_r_Conv(block, sel, cmp_mode);
+	}
+
+	/* normalize so switch_min is at 0 */
+	if (env->switch_min != 0) {
+		ir_node *min_const  = new_r_Const_long(irg, cmp_mode, env->switch_min);
+		sel = new_rd_Sub(dbgi, block, sel, min_const, cmp_mode);
+
+		delta            = env->switch_min;
+		env->switch_min  = 0;
+		env->switch_max -= delta;
+		set_Cond_selector(cond, sel);
+	}
+
+	/* check for out-of-bounds */
+	max_const  = new_r_Const_long(irg, cmp_mode, env->switch_max);
+	cmp        = new_rd_Cmp(dbgi, block, sel, max_const, ir_relation_less_equal);
+	oob_cond   = new_rd_Cond(dbgi, block, cmp);
+	proj_true  = new_r_Proj(oob_cond, mode_X, pn_Cond_true);
+	proj_false = new_r_Proj(oob_cond, mode_X, pn_Cond_false);
+
+	ARR_APP1(ir_node*, default_preds, proj_false);
+
+	/* create new block containing the cond */
+	in[0] = proj_true;
+	new_block = new_r_Block(irg, 1, in);
+	set_nodes_block(cond, new_block);
+
+	/* adapt projs */
+	foreach_out_irn(cond, i, proj) {
+		long pn     = get_Proj_proj(proj);
+		long new_pn = pn - delta;
+		if (pn == default_pn) {
+			set_Cond_default_proj(cond, new_pn);
+			ARR_APP1(ir_node*, default_preds, proj);
+		}
+
+		set_Proj_proj(proj, new_pn);
+		set_nodes_block(proj, new_block);
+	}
+
+	/* adapt default block */
+	n_default_preds = ARR_LEN(default_preds);
+	if (n_default_preds > 1) {
+		size_t p;
+
+		/* create new intermediate blocks so we don't have critical edges */
+		for (p = 0; p < n_default_preds; ++p) {
+			ir_node *pred = default_preds[p];
+			ir_node *split_block;
+			ir_node *block_in[1];
+
+			block_in[0] = pred;
+			split_block = new_r_Block(irg, 1, block_in);
+
+			default_preds[p] = new_r_Jmp(split_block);
+		}
+	}
+	set_irn_in(env->default_block, n_default_preds, default_preds);
+
+	DEL_ARR_F(default_preds);
+}
+
 /**
  * Block-Walker: searches for Cond nodes with a non-boolean mode
  */
 static void find_cond_nodes(ir_node *block, void *ctx)
 {
-	walk_env_t  *env = ctx;
+	walk_env_t  *env = (walk_env_t *)ctx;
 	ir_node     *projx;
 	ir_node     *cond;
 	ir_node     *sel;
 	ir_mode     *sel_mode;
 	long         default_pn;
-	int          i, j = 0, numcases;
+	int          i;
+	unsigned     j = 0;
+	unsigned     numcases;
 	ir_node     *proj;
 	case_data_t *cases;
 	ir_node     *condblock;
 	ir_node     *defblock = NULL;
-	ifcas_env_t  ifcas_env;
-
-	if(get_Block_n_cfgpreds(block) != 1)
+	dbg_info    *dbgi;
+	cond_env_t   cond_env;
+	unsigned long spare;
+	bool         lower_switch = false;
+
+	/* because we split critical blocks only blocks with 1 predecessors may
+	 * contain Proj->Cond nodes */
+	if (get_Block_n_cfgpreds(block) != 1)
 		return;
 
 	projx = get_Block_cfgpred(block, 0);
-	if(!is_Proj(projx))
+	if (!is_Proj(projx))
 		return;
 	assert(get_irn_mode(projx) == mode_X);
 
 	cond = get_Proj_pred(projx);
-	if(!is_Cond(cond))
+	if (!is_Cond(cond))
 		return;
 
 	sel      = get_Cond_selector(cond);
 	sel_mode = get_irn_mode(sel);
 
-	if(sel_mode == mode_b)    /* not a switch? */
+	if (sel_mode == mode_b)
 		return;
 
-	if(should_do_table_switch(cond, env->spare_size))
+	if (ir_nodeset_contains(&env->processed, cond))
+		return;
+	ir_nodeset_insert(&env->processed, cond);
+
+	/* the algorithms here don't work reliable for modes bigger than 32
+	 * since we operate with long numbers */
+	assert(get_mode_size_bits(sel_mode) <= 32);
+
+	analyse_switch(&cond_env, cond);
+
+	/*
+	 * Here we have: num_cases and [switch_min, switch_max] interval.
+	 * We do an if-cascade if there are too many spare numbers.
+	 */
+	spare = (unsigned long) cond_env.switch_max
+		- (unsigned long) cond_env.switch_min
+		- (unsigned long) cond_env.num_cases + 1;
+	lower_switch |= spare >= env->spare_size;
+	lower_switch |= cond_env.num_cases <= env->small_switch;
+
+	if (!lower_switch) {
+		/* we won't decompose the switch. But we might have to add
+		 * out-of-bounds checking */
+		if (!env->allow_out_of_bounds) {
+			create_out_of_bounds_check(&cond_env, cond);
+			env->changed = true;
+		}
 		return;
+	}
 
 	/*
 	 * Switch should be transformed into an if cascade.
 	 * So first order the cases, so we can do a binary search on them.
 	 */
+	env->changed = true;
 
-	numcases = get_irn_n_outs(cond) - 1;      // does not contain default case
-	cases    = obstack_alloc(&env->obst, numcases * sizeof(*cases));
+	numcases = get_irn_n_outs(cond) - 1; /* does not contain default case */
+	cases    = XMALLOCN(case_data_t, numcases);
 
-	default_pn = get_Cond_defaultProj(cond);
-	ifcas_env.sel = sel;
-	ifcas_env.defindex = 0;
-	NEW_ARR_A(ir_node*, ifcas_env.defusers, numcases);
+	default_pn        = get_Cond_default_proj(cond);
+	cond_env.sel      = sel;
+	cond_env.defusers = NEW_ARR_F(ir_node*, 0);
 
 	foreach_out_irn(cond, i, proj) {
 		long pn = get_Proj_proj(proj);
 		ir_node *target = get_irn_out(proj, 0);
-		assert(get_Block_n_cfgpreds(target) == 1 && "Encountered critical edge in switch");
+		assert(get_Block_n_cfgpreds(target) == 1);
 
-		if(pn == default_pn)
-		{
+		if (pn == default_pn) {
 			defblock = target;
 			continue;
 		}
@@ -229,51 +385,44 @@ static void find_cond_nodes(ir_node *block, void *ctx)
 		cases[j].target = target;
 		j++;
 	}
+	assert(j == numcases);
 
-	assert(defblock != NULL && "Switch without default proj");
-	qsort(cases, numcases, sizeof(*cases), casecmp);
+	if (defblock == NULL)
+		panic("Switch %+F has no default proj", cond);
+
+	qsort(cases, numcases, sizeof(cases[0]), casecmp);
 
 	/* Now create the if cascade */
 	condblock = get_nodes_block(cond);
-	create_if_cascade(&ifcas_env, condblock, cases, numcases);
+	dbgi      = get_irn_dbg_info(cond);
+	create_if_cascade(&cond_env, dbgi, condblock, cases, numcases);
 
 	/* Connect new default case users */
-	set_irn_in(defblock, ifcas_env.defindex, ifcas_env.defusers);
+	set_irn_in(defblock, ARR_LEN(cond_env.defusers), cond_env.defusers);
 
-	obstack_free(&env->obst, cases);
+	xfree(cases);
+	DEL_ARR_F(cond_env.defusers);
 }
 
-/**
- * Lowers all Switches (Cond nodes with non-boolean mode) depending on spare_size.
- * They will either remain the same or be converted into if-cascades.
- *
- * @param irg        The ir graph to be lowered.
- * @param spare_size Allowed spare size for table switches in machine words.
- *                   (Default in edgfe: 128)
- */
-void lower_switch(ir_graph *irg, unsigned spare_size)
+void lower_switch(ir_graph *irg, unsigned small_switch, unsigned spare_size,
+                  int allow_out_of_bounds)
 {
 	walk_env_t env;
-	ir_graph *rem = current_ir_graph;
-
-	current_ir_graph = irg;
-
-	obstack_init(&env.obst);
-	env.spare_size = spare_size;
+	env.changed             = false;
+	env.spare_size          = spare_size;
+	env.small_switch        = small_switch;
+	env.allow_out_of_bounds = allow_out_of_bounds;
+	ir_nodeset_init(&env.processed);
 
 	remove_critical_cf_edges(irg);
 	assure_irg_outs(irg);
 
 	irg_block_walk_graph(irg, find_cond_nodes, NULL, &env);
+	ir_nodeset_destroy(&env.processed);
 
-	if(env.changed) {
+	if (env.changed) {
 		/* control flow changed */
-		set_irg_outs_inconsistent(irg);
 		set_irg_doms_inconsistent(irg);
 		set_irg_extblk_inconsistent(irg);
-		set_irg_loopinfo_inconsistent(irg);
 	}
-
-	obstack_free(&env.obst, NULL);
-	current_ir_graph = rem;
 }