X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_common_transform.c;h=f9a40ac71f50ce45681caf86691d243afe3e4dc0;hb=85f517eca982e6a4e1d1848eb67634ae33b70de9;hp=9e5111691e6bb52f63c3c84c7a0ae05dec577dec;hpb=0e07ee89d57a00a60eb989b8fe6ffcff5f54c68d;p=libfirm

diff --git a/ir/be/ia32/ia32_common_transform.c b/ir/be/ia32/ia32_common_transform.c
index 9e5111691..f9a40ac71 100644
--- a/ir/be/ia32/ia32_common_transform.c
+++ b/ir/be/ia32/ia32_common_transform.c
@@ -24,18 +24,18 @@
  * @author      Matthias Braun, Sebastian Buchwald
  * @version     $Id: ia32_common_transform.c 21012 2008-08-06 13:35:17Z beck $
  */
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include "error.h"
-#include "irargs_t.h"
 #include "ircons.h"
 #include "irprintf.h"
 #include "typerep.h"
+#include "bitset.h"
+#include "heights.h"
 
 #include "../betranshlp.h"
-#include "../beirg_t.h"
+#include "../beirg.h"
+#include "../beabi.h"
 
 #include "ia32_architecture.h"
 #include "ia32_common_transform.h"
@@ -44,18 +44,7 @@
 #include "gen_ia32_new_nodes.h"
 #include "gen_ia32_regalloc_if.h"
 
-/** hold the current code generator during transformation */
-ia32_code_gen_t *env_cg = NULL;
-
-heights_t *heights = NULL;
-
-static const arch_register_req_t no_register_req = {
-	arch_register_req_type_none,
-	NULL,                         /* regclass */
-	NULL,                         /* limit bitset */
-	0,                            /* same pos */
-	0                             /* different pos */
-};
+ir_heights_t *ia32_heights = NULL;
 
 static int check_immediate_constraint(long val, char immediate_constraint_type)
 {
@@ -76,105 +65,73 @@ static int check_immediate_constraint(long val, char immediate_constraint_type)
 }
 
 /**
- * creates a unique ident by adding a number to a tag
- *
- * @param tag   the tag string, must contain a %d if a number
- *              should be added
- */
-static ident *unique_id(const char *tag)
-{
-	static unsigned id = 0;
-	char str[256];
-
-	snprintf(str, sizeof(str), tag, ++id);
-	return new_id_from_str(str);
-}
-
-/**
- * Get a primitive type for a mode.
+ * Get a primitive type for a mode with alignment 16.
  */
 static ir_type *ia32_get_prim_type(pmap *types, ir_mode *mode)
 {
-	pmap_entry *e = pmap_find(types, mode);
-	ir_type *res;
+	ir_type *res = (ir_type*)pmap_get(types, mode);
+	if (res != NULL)
+		return res;
 
-	if (! e) {
-		char buf[64];
-		snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
-		res = new_type_primitive(new_id_from_str(buf), mode);
+	res = new_type_primitive(mode);
+	if (get_mode_size_bits(mode) >= 80) {
 		set_type_alignment_bytes(res, 16);
-		pmap_insert(types, mode, res);
 	}
-	else
-		res = e->value;
+	pmap_insert(types, mode, res);
 	return res;
 }
 
-ir_entity *create_float_const_entity(ir_node *cnst)
+ir_entity *ia32_create_float_const_entity(ir_node *cnst)
 {
-	ia32_isa_t *isa = env_cg->isa;
-	tarval *key     = get_Const_tarval(cnst);
-	pmap_entry *e   = pmap_find(isa->tv_ent, key);
-	ir_entity *res;
-	ir_graph *rem;
-
-	if (e == NULL) {
-		tarval  *tv   = key;
-		ir_mode *mode = get_tarval_mode(tv);
-		ir_type *tp;
-
-		if (! ia32_cg_config.use_sse2) {
-			/* try to reduce the mode to produce smaller sized entities */
-			if (mode != mode_F) {
-				if (tarval_ieee754_can_conv_lossless(tv, mode_F)) {
-					mode = mode_F;
+	ir_graph         *irg      = get_irn_irg(cnst);
+	const arch_env_t *arch_env = be_get_irg_arch_env(irg);
+	ia32_isa_t       *isa      = (ia32_isa_t*) arch_env;
+	ir_tarval        *tv       = get_Const_tarval(cnst);
+	ir_entity        *res      = (ir_entity*)pmap_get(isa->tv_ent, tv);
+	ir_initializer_t *initializer;
+	ir_mode          *mode;
+	ir_type          *tp;
+
+	if (res != NULL)
+		return res;
+
+	mode = get_tarval_mode(tv);
+
+	if (! ia32_cg_config.use_sse2) {
+		/* try to reduce the mode to produce smaller sized entities */
+		if (mode != mode_F) {
+			if (tarval_ieee754_can_conv_lossless(tv, mode_F)) {
+				mode = mode_F;
+				tv = tarval_convert_to(tv, mode);
+			} else if (mode != mode_D) {
+				if (tarval_ieee754_can_conv_lossless(tv, mode_D)) {
+					mode = mode_D;
 					tv = tarval_convert_to(tv, mode);
-				} else if (mode != mode_D) {
-					if (tarval_ieee754_can_conv_lossless(tv, mode_D)) {
-						mode = mode_D;
-						tv = tarval_convert_to(tv, mode);
-					}
 				}
 			}
 		}
-
-		if (mode == get_irn_mode(cnst)) {
-			/* mode was not changed */
-			tp = get_Const_type(cnst);
-			if (tp == firm_unknown_type)
-				tp = ia32_get_prim_type(isa->types, mode);
-		} else
-			tp = ia32_get_prim_type(isa->types, mode);
-
-		res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
-
-		set_entity_ld_ident(res, get_entity_ident(res));
-		set_entity_visibility(res, visibility_local);
-		set_entity_variability(res, variability_constant);
-		set_entity_allocation(res, allocation_static);
-
-		 /* we create a new entity here: It's initialization must resist on the
-		    const code irg */
-		rem = current_ir_graph;
-		current_ir_graph = get_const_code_irg();
-		set_atomic_ent_value(res, new_Const_type(tv, tp));
-		current_ir_graph = rem;
-
-		pmap_insert(isa->tv_ent, key, res);
-	} else {
-		res = e->value;
 	}
 
+	tp  = ia32_get_prim_type(isa->types, mode);
+	res = new_entity(get_glob_type(), id_unique("C%u"), tp);
+	set_entity_ld_ident(res, get_entity_ident(res));
+	set_entity_visibility(res, ir_visibility_private);
+	add_entity_linkage(res, IR_LINKAGE_CONSTANT);
+
+	initializer = create_initializer_tarval(tv);
+	set_entity_initializer(res, initializer);
+
+	pmap_insert(isa->tv_ent, tv, res);
 	return res;
 }
 
-ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long val)
+ir_node *ia32_create_Immediate(ir_entity *symconst, int symconst_sign, long val)
 {
 	ir_graph *irg         = current_ir_graph;
 	ir_node  *start_block = get_irg_start_block(irg);
-	ir_node  *immediate   = new_rd_ia32_Immediate(NULL, irg, start_block,
-	                                              symconst, symconst_sign, val);
-	arch_set_irn_register(env_cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]);
+	ir_node  *immediate   = new_bd_ia32_Immediate(NULL, start_block, symconst,
+			symconst_sign, ia32_no_pic_adjust, val);
+	arch_set_irn_register(immediate, &ia32_registers[REG_GP_NOREG]);
 
 	return immediate;
 }
@@ -188,37 +145,28 @@ const arch_register_t *ia32_get_clobber_register(const char *clobber)
 
 	/* TODO: construct a hashmap instead of doing linear search for clobber
 	 * register */
-	for(c = 0; c < N_CLASSES; ++c) {
+	for (c = 0; c < N_IA32_CLASSES; ++c) {
 		cls = & ia32_reg_classes[c];
-		for(r = 0; r < cls->n_regs; ++r) {
+		for (r = 0; r < cls->n_regs; ++r) {
 			const arch_register_t *temp_reg = arch_register_for_index(cls, r);
-			if(strcmp(temp_reg->name, clobber) == 0
+			if (strcmp(temp_reg->name, clobber) == 0
 					|| (c == CLASS_ia32_gp && strcmp(temp_reg->name+1, clobber) == 0)) {
 				reg = temp_reg;
 				break;
 			}
 		}
-		if(reg != NULL)
+		if (reg != NULL)
 			break;
 	}
 
 	return reg;
 }
 
-#ifndef NDEBUG
-const char *ia32_get_old_node_name(ia32_code_gen_t *cg, ir_node *irn) {
-	ia32_isa_t *isa = (ia32_isa_t*) cg->arch_env;
-
-	lc_eoprintf(firm_get_arg_env(), isa->name_obst, "%+F", irn);
-	obstack_1grow(isa->name_obst, 0);
- 	return obstack_finish(isa->name_obst);
-}
-#endif /* NDEBUG */
-
-int ia32_mode_needs_gp_reg(ir_mode *mode) {
-	if(mode == mode_fpcw)
+int ia32_mode_needs_gp_reg(ir_mode *mode)
+{
+	if (mode == ia32_mode_fpcw)
 		return 0;
-	if(get_mode_size_bits(mode) > 32)
+	if (get_mode_size_bits(mode) > 32)
 		return 0;
 	return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
 }
@@ -237,7 +185,7 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 	memset(constraint, 0, sizeof(constraint[0]));
 	constraint->same_as = -1;
 
-	if(*c == 0) {
+	if (*c == 0) {
 		/* a memory constraint: no need to do anything in backend about it
 		 * (the dependencies are already respected by the memory edge of
 		 * the node) */
@@ -246,8 +194,8 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 
 	/* TODO: improve error messages with node and source info. (As users can
 	 * easily hit these) */
-	while(*c != 0) {
-		switch(*c) {
+	while (*c != 0) {
+		switch (*c) {
 		case ' ':
 		case '\t':
 		case '\n':
@@ -257,43 +205,45 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 		case '=': break;
 		case '+': break;
 
+		case '&': break;
+
 		case '*':
 			++c;
 			break;
 		case '#':
-			while(*c != 0 && *c != ',')
+			while (*c != 0 && *c != ',')
 				++c;
 			break;
 
 		case 'a':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EAX;
+			limited |= 1 << REG_GP_EAX;
 			break;
 		case 'b':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EBX;
+			limited |= 1 << REG_GP_EBX;
 			break;
 		case 'c':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_ECX;
+			limited |= 1 << REG_GP_ECX;
 			break;
 		case 'd':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EDX;
+			limited |= 1 << REG_GP_EDX;
 			break;
 		case 'D':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EDI;
+			limited |= 1 << REG_GP_EDI;
 			break;
 		case 'S':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_ESI;
+			limited |= 1 << REG_GP_ESI;
 			break;
 		case 'Q':
 		case 'q':
@@ -301,20 +251,20 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 			 * difference to Q for us (we only assign whole registers) */
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EAX | 1 << REG_EBX | 1 << REG_ECX |
-			           1 << REG_EDX;
+			limited |= 1 << REG_GP_EAX | 1 << REG_GP_EBX | 1 << REG_GP_ECX |
+			           1 << REG_GP_EDX;
 			break;
 		case 'A':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EAX | 1 << REG_EDX;
+			limited |= 1 << REG_GP_EAX | 1 << REG_GP_EDX;
 			break;
 		case 'l':
 			assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 			cls      = &ia32_reg_classes[CLASS_ia32_gp];
-			limited |= 1 << REG_EAX | 1 << REG_EBX | 1 << REG_ECX |
-			           1 << REG_EDX | 1 << REG_ESI | 1 << REG_EDI |
-			           1 << REG_EBP;
+			limited |= 1 << REG_GP_EAX | 1 << REG_GP_EBX | 1 << REG_GP_ECX |
+			           1 << REG_GP_EDX | 1 << REG_GP_ESI | 1 << REG_GP_EDI |
+			           1 << REG_GP_EBP;
 			break;
 
 		case 'R':
@@ -394,7 +344,7 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 				panic("can only specify same constraint on input");
 
 			sscanf(c, "%d%n", &same_as, &p);
-			if(same_as >= 0) {
+			if (same_as >= 0) {
 				c += p;
 				continue;
 			}
@@ -421,16 +371,14 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 		case 'e': /* not available in 32 bit mode */
 			panic("unsupported asm constraint '%c' found in (%+F)",
 			      *c, current_ir_graph);
-			break;
 		default:
 			panic("unknown asm constraint '%c' found in (%+F)", *c,
 			      current_ir_graph);
-			break;
 		}
 		++c;
 	}
 
-	if(same_as >= 0) {
+	if (same_as >= 0) {
 		if (cls != NULL)
 			panic("same as and register constraint not supported");
 		if (immediate_type != '\0')
@@ -450,13 +398,38 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 	constraint->immediate_type        = immediate_type;
 }
 
-ir_node *gen_ASM(ir_node *node)
+static bool can_match(const arch_register_req_t *in,
+                      const arch_register_req_t *out)
+{
+	if (in->cls != out->cls)
+		return false;
+	if ( (in->type & arch_register_req_type_limited) == 0
+		|| (out->type & arch_register_req_type_limited) == 0 )
+		return true;
+
+	return (*in->limited & *out->limited) != 0;
+}
+
+static inline ir_node *get_new_node(ir_node *node)
+{
+#ifdef FIRM_GRGEN_BE
+	if (be_transformer == TRANSFORMER_DEFAULT) {
+		return be_transform_node(node);
+	} else {
+		return node;
+	}
+#else
+	return be_transform_node(node);
+#endif
+}
+
+ir_node *ia32_gen_ASM(ir_node *node)
 {
-	ir_graph                   *irg       = current_ir_graph;
-	ir_node                    *block = NULL;
-	ir_node                    *new_block = NULL;
+	ir_node                    *block     = get_nodes_block(node);
+	ir_node                    *new_block = get_new_node(block);
 	dbg_info                   *dbgi      = get_irn_dbg_info(node);
 	int                         i, arity;
+	int                         value_arity;
 	int                         out_idx;
 	ir_node                   **in;
 	ir_node                    *new_node;
@@ -471,38 +444,14 @@ ir_node *gen_ASM(ir_node *node)
 	const ir_asm_constraint    *in_constraints;
 	const ir_asm_constraint    *out_constraints;
 	ident                     **clobbers;
-	int                         clobbers_flags = 0;
-	unsigned                    clobber_bits[N_CLASSES];
+	unsigned                    clobber_bits[N_IA32_CLASSES];
+	int                         out_size;
+	backend_info_t             *info;
 
 	memset(&clobber_bits, 0, sizeof(clobber_bits));
 
-	switch (be_transformer) {
-	case TRANSFORMER_DEFAULT:
-		block     = get_nodes_block(node);
-		new_block = be_transform_node(block);
-		break;
-
-#ifdef FIRM_GRGEN_BE
-	case TRANSFORMER_PBQP:
-	case TRANSFORMER_RAND:
-		new_block = get_nodes_block(node);
-		break;
-#endif
-
-	default:
-		panic("invalid transformer");
-	}
-
-	/* workaround for lots of buggy code out there as most people think volatile
-	 * asm is enough for everything and forget the flags (linux kernel, etc.)
-	 */
-	if (get_irn_pinned(node) == op_pin_state_pinned) {
-		clobbers_flags = 1;
-	}
-
 	arity = get_irn_arity(node);
-	in    = alloca(arity * sizeof(in[0]));
-	memset(in, 0, arity * sizeof(in[0]));
+	in    = ALLOCANZ(ir_node*, arity);
 
 	clobbers   = get_ASM_clobbers(node);
 	n_clobbers = 0;
@@ -513,11 +462,10 @@ ir_node *gen_ASM(ir_node *node)
 		if (strcmp(c, "memory") == 0)
 			continue;
 		if (strcmp(c, "cc") == 0) {
-			clobbers_flags = 1;
 			continue;
 		}
 
-		req = parse_clobber(c);
+		req = ia32_parse_clobber(c);
 		clobber_bits[req->cls->index] |= *req->limited;
 
 		n_clobbers++;
@@ -535,31 +483,38 @@ ir_node *gen_ASM(ir_node *node)
 			reg_map_size = constraint->pos;
 	}
 	for (i = 0; i < arity; ++i) {
-		const ir_asm_constraint   *constraint = &in_constraints[i];
-		if(constraint->pos > reg_map_size)
+		const ir_asm_constraint *constraint = &in_constraints[i];
+		if (constraint->pos > reg_map_size)
 			reg_map_size = constraint->pos;
 	}
 	++reg_map_size;
 
-	obst         = get_irg_obstack(irg);
+	obst         = get_irg_obstack(current_ir_graph);
 	register_map = NEW_ARR_D(ia32_asm_reg_t, obst, reg_map_size);
 	memset(register_map, 0, reg_map_size * sizeof(register_map[0]));
 
 	/* construct output constraints */
-	out_reg_reqs = obstack_alloc(obst, out_arity * sizeof(out_reg_reqs[0]));
+	out_size = out_arity + 1;
+	out_reg_reqs = OALLOCN(obst, const arch_register_req_t*, out_size);
 
 	for (out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
 		const ir_asm_constraint   *constraint = &out_constraints[out_idx];
 		const char                *c       = get_id_str(constraint->constraint);
-		unsigned                   pos        = constraint->pos;
+		unsigned                   pos     = constraint->pos;
 		constraint_t               parsed_constraint;
 		const arch_register_req_t *req;
 
 		parse_asm_constraints(&parsed_constraint, c, 1);
-		req = make_register_req(&parsed_constraint, n_out_constraints,
+		req = ia32_make_register_req(&parsed_constraint, n_out_constraints,
 		                        out_reg_reqs, out_idx);
 		out_reg_reqs[out_idx] = req;
 
+		/* multiple constraints for same pos. This can happen for example when
+		 * a =A constraint gets lowered to two constraints: =a and =d for the
+		 * same pos */
+		if (register_map[pos].valid)
+			continue;
+
 		register_map[pos].use_input = 0;
 		register_map[pos].valid     = 1;
 		register_map[pos].memory    = 0;
@@ -568,7 +523,7 @@ ir_node *gen_ASM(ir_node *node)
 	}
 
 	/* inputs + input constraints */
-	in_reg_reqs = obstack_alloc(obst, arity * sizeof(in_reg_reqs[0]));
+	in_reg_reqs = OALLOCN(obst, const arch_register_req_t*, arity);
 	for (i = 0; i < arity; ++i) {
 		ir_node                   *pred         = get_irn_n(node, i);
 		const ir_asm_constraint   *constraint   = &in_constraints[i];
@@ -587,7 +542,7 @@ ir_node *gen_ASM(ir_node *node)
 			if (r_clobber_bits != 0) {
 				if (parsed_constraint.all_registers_allowed) {
 					parsed_constraint.all_registers_allowed = 0;
-					be_abi_set_non_ignore_regs(env_cg->birg->abi,
+					be_set_allocatable_regs(current_ir_graph,
 							parsed_constraint.cls,
 							&parsed_constraint.allowed_registers);
 				}
@@ -595,37 +550,22 @@ ir_node *gen_ASM(ir_node *node)
 			}
 		}
 
-		req = make_register_req(&parsed_constraint, n_out_constraints,
+		req = ia32_make_register_req(&parsed_constraint, n_out_constraints,
 		                        out_reg_reqs, i);
 		in_reg_reqs[i] = req;
 
 		if (parsed_constraint.immediate_type != '\0') {
 			char imm_type = parsed_constraint.immediate_type;
-			input = try_create_Immediate(pred, imm_type);
+			input = ia32_try_create_Immediate(pred, imm_type);
 		}
 
 		if (input == NULL) {
-			ir_node *pred = NULL;
-			switch (be_transformer) {
-			case TRANSFORMER_DEFAULT:
-				pred  = get_irn_n(node, i);
-				input = be_transform_node(pred);
-				break;
-
-#ifdef FIRM_GRGEN_BE
-			case TRANSFORMER_PBQP:
-			case TRANSFORMER_RAND:
-				input = get_irn_n(node, i);
-				break;
-#endif
-
-			default: panic("invalid transformer");
-			}
+			input = get_new_node(pred);
 
 			if (parsed_constraint.cls == NULL
 					&& parsed_constraint.same_as < 0) {
 				is_memory_op = 1;
-			} else if(parsed_constraint.memory_possible) {
+			} else if (parsed_constraint.memory_possible) {
 				/* TODO: match Load or Load/Store if memory possible is set */
 			}
 		}
@@ -646,146 +586,223 @@ ir_node *gen_ASM(ir_node *node)
 		if (strcmp(c, "memory") == 0 || strcmp(c, "cc") == 0)
 			continue;
 
-		req = parse_clobber(c);
+		req = ia32_parse_clobber(c);
 		out_reg_reqs[out_idx] = req;
 		++out_idx;
 	}
 
-	new_node = new_rd_ia32_Asm(dbgi, irg, new_block, arity, in, out_arity,
-	                           get_ASM_text(node), register_map);
+	/* count inputs which are real values (and not memory) */
+	value_arity = 0;
+	for (i = 0; i < arity; ++i) {
+		ir_node *node_in = get_irn_n(node, i);
+		if (get_irn_mode(node_in) == mode_M)
+			continue;
+		++value_arity;
+	}
+
+	/* Attempt to make ASM node register pressure faithful.
+	 * (This does not work for complicated cases yet!)
+	 *
+	 * Algorithm: Check if there are fewer inputs or outputs (I will call this
+	 * the smaller list). Then try to match each constraint of the smaller list
+	 * to 1 of the other list. If we can't match it, then we have to add a dummy
+	 * input/output to the other list
+	 *
+	 * FIXME: This is still broken in lots of cases. But at least better than
+	 *        before...
+	 * FIXME: need to do this per register class...
+	 */
+	if (out_arity <= value_arity) {
+		int       orig_arity = arity;
+		int       in_size    = arity;
+		int       o;
+		bitset_t *used_ins = bitset_alloca(arity);
+		for (o = 0; o < out_arity; ++o) {
+			const arch_register_req_t *outreq = out_reg_reqs[o];
+
+			if (outreq->cls == NULL) {
+				continue;
+			}
+
+			for (i = 0; i < orig_arity; ++i) {
+				const arch_register_req_t *inreq;
+				if (bitset_is_set(used_ins, i))
+					continue;
+				inreq = in_reg_reqs[i];
+				if (!can_match(outreq, inreq))
+					continue;
+				bitset_set(used_ins, i);
+				break;
+			}
+			/* did we find any match? */
+			if (i < orig_arity)
+				continue;
+
+			/* we might need more space in the input arrays */
+			if (arity >= in_size) {
+				const arch_register_req_t **new_in_reg_reqs;
+				ir_node             **new_in;
+
+				in_size *= 2;
+				new_in_reg_reqs = OALLOCN(obst, const arch_register_req_t*,
+				                          in_size);
+				memcpy(new_in_reg_reqs, in_reg_reqs, arity * sizeof(new_in_reg_reqs[0]));
+				new_in = ALLOCANZ(ir_node*, in_size);
+				memcpy(new_in, in, arity*sizeof(new_in[0]));
+
+				in_reg_reqs = new_in_reg_reqs;
+				in          = new_in;
+			}
+
+			/* add a new (dummy) input which occupies the register */
+			assert(outreq->type & arch_register_req_type_limited);
+			in_reg_reqs[arity] = outreq;
+			in[arity]          = new_bd_ia32_ProduceVal(NULL, block);
+			++arity;
+		}
+	} else {
+		bitset_t *used_outs = bitset_alloca(out_arity);
+		int       orig_out_arity = out_arity;
+		for (i = 0; i < arity; ++i) {
+			int   o;
+			const arch_register_req_t *inreq = in_reg_reqs[i];
+
+			if (inreq->cls == NULL) {
+				continue;
+			}
 
-	/* Prevent the ASM node from being scheduled before the Barrier, if it has
-	 * no inputs */
-	if (arity == 0 && get_irg_start_block(irg) == new_block) {
-		add_irn_dep(new_node, get_irg_frame(irg));
+			for (o = 0; o < orig_out_arity; ++o) {
+				const arch_register_req_t *outreq;
+				if (bitset_is_set(used_outs, o))
+					continue;
+				outreq = out_reg_reqs[o];
+				if (!can_match(outreq, inreq))
+					continue;
+				bitset_set(used_outs, i);
+				break;
+			}
+			/* did we find any match? */
+			if (o < orig_out_arity)
+				continue;
+
+			/* we might need more space in the output arrays */
+			if (out_arity >= out_size) {
+				const arch_register_req_t **new_out_reg_reqs;
+
+				out_size *= 2;
+				new_out_reg_reqs
+					= OALLOCN(obst, const arch_register_req_t*, out_size);
+				memcpy(new_out_reg_reqs, out_reg_reqs,
+				       out_arity * sizeof(new_out_reg_reqs[0]));
+				out_reg_reqs = new_out_reg_reqs;
+			}
+
+			/* add a new (dummy) output which occupies the register */
+			assert(inreq->type & arch_register_req_type_limited);
+			out_reg_reqs[out_arity] = inreq;
+			++out_arity;
+		}
 	}
 
-	set_ia32_out_req_all(new_node, out_reg_reqs);
-	set_ia32_in_req_all(new_node, in_reg_reqs);
+	/* append none register requirement for the memory output */
+	if (out_arity + 1 >= out_size) {
+		const arch_register_req_t **new_out_reg_reqs;
+
+		out_size = out_arity + 1;
+		new_out_reg_reqs
+			= OALLOCN(obst, const arch_register_req_t*, out_size);
+		memcpy(new_out_reg_reqs, out_reg_reqs,
+			   out_arity * sizeof(new_out_reg_reqs[0]));
+		out_reg_reqs = new_out_reg_reqs;
+	}
 
-	SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
+	/* add a new (dummy) output which occupies the register */
+	out_reg_reqs[out_arity] = arch_no_register_req;
+	++out_arity;
+
+	new_node = new_bd_ia32_Asm(dbgi, new_block, arity, in, out_arity,
+	                           get_ASM_text(node), register_map);
+
+	info = be_get_info(new_node);
+	for (i = 0; i < out_arity; ++i) {
+		info->out_infos[i].req = out_reg_reqs[i];
+	}
+	arch_set_in_register_reqs(new_node, in_reg_reqs);
+
+	SET_IA32_ORIG_NODE(new_node, node);
 
 	return new_node;
 }
 
-ir_node *gen_CopyB(ir_node *node) {
-	ir_node  *block    = NULL;
-	ir_node  *src      = NULL;
-	ir_node  *new_src  = NULL;
-	ir_node  *dst      = NULL;
-	ir_node  *new_dst  = NULL;
-	ir_node  *mem      = NULL;
-	ir_node  *new_mem  = NULL;
+ir_node *ia32_gen_CopyB(ir_node *node)
+{
+	ir_node  *block    = get_new_node(get_nodes_block(node));
+	ir_node  *src      = get_CopyB_src(node);
+	ir_node  *new_src  = get_new_node(src);
+	ir_node  *dst      = get_CopyB_dst(node);
+	ir_node  *new_dst  = get_new_node(dst);
+	ir_node  *mem      = get_CopyB_mem(node);
+	ir_node  *new_mem  = get_new_node(mem);
 	ir_node  *res      = NULL;
-	ir_graph *irg      = current_ir_graph;
 	dbg_info *dbgi     = get_irn_dbg_info(node);
 	int      size      = get_type_size_bytes(get_CopyB_type(node));
+	int      throws_exception = ir_throws_exception(node);
 	int      rem;
 
-	switch (be_transformer) {
-		case TRANSFORMER_DEFAULT:
-			block    = be_transform_node(get_nodes_block(node));
-			src      = get_CopyB_src(node);
-			new_src  = be_transform_node(src);
-			dst      = get_CopyB_dst(node);
-			new_dst  = be_transform_node(dst);
-			mem      = get_CopyB_mem(node);
-			new_mem  = be_transform_node(mem);
-			break;
-
-#ifdef FIRM_GRGEN_BE
-		case TRANSFORMER_PBQP:
-		case TRANSFORMER_RAND:
-			block    = get_nodes_block(node);
-			new_src  = get_CopyB_src(node);
-			new_dst  = get_CopyB_dst(node);
-			new_mem  = get_CopyB_mem(node);
-			break;
-#endif
-
-		default: panic("invalid transformer");
-	}
-
 	/* If we have to copy more than 32 bytes, we use REP MOVSx and */
 	/* then we need the size explicitly in ECX.                    */
 	if (size >= 32 * 4) {
 		rem = size & 0x3; /* size % 4 */
 		size >>= 2;
 
-		res = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, size);
-		add_irn_dep(res, get_irg_frame(irg));
+		res = new_bd_ia32_Const(dbgi, block, NULL, 0, 0, size);
 
-		res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem, rem);
+		res = new_bd_ia32_CopyB(dbgi, block, new_dst, new_src, res, new_mem, rem);
 	} else {
-		if(size == 0) {
+		if (size == 0) {
 			ir_fprintf(stderr, "Optimization warning copyb %+F with size <4\n",
 			           node);
 		}
-		res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem, size);
+		res = new_bd_ia32_CopyB_i(dbgi, block, new_dst, new_src, new_mem, size);
 	}
+	ir_set_throws_exception(res, throws_exception);
 
-	SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
+	SET_IA32_ORIG_NODE(res, node);
 
 	return res;
 }
 
-ir_node *gen_Proj_tls(ir_node *node) {
-	ir_node  *block = NULL;
-	ir_graph *irg   = current_ir_graph;
-	dbg_info *dbgi  = NULL;
-	ir_node  *res   = NULL;
-
-	switch (be_transformer) {
-		case TRANSFORMER_DEFAULT:
-			block = be_transform_node(get_nodes_block(node));
-			break;
-
-#ifdef FIRM_GRGEN_BE
-		case TRANSFORMER_PBQP:
-		case TRANSFORMER_RAND:
-			block = get_nodes_block(node);
-			break;
-#endif
-
-		default: panic("invalid transformer");
-	}
-
-	res   = new_rd_ia32_LdTls(dbgi, irg, block, mode_Iu);
-
+ir_node *ia32_gen_Proj_tls(ir_node *node)
+{
+	ir_node *block = get_new_node(get_nodes_block(node));
+	ir_node *res   = new_bd_ia32_LdTls(NULL, block);
 	return res;
 }
 
-ir_node *gen_Unknown(ir_node *node)
+ir_node *ia32_gen_Unknown(ir_node *node)
 {
-	ir_mode *mode = get_irn_mode(node);
+	ir_mode  *mode  = get_irn_mode(node);
+	ir_graph *irg   = current_ir_graph;
+	dbg_info *dbgi  = get_irn_dbg_info(node);
+	ir_node  *block = get_irg_start_block(irg);
+	ir_node  *res   = NULL;
 
 	if (mode_is_float(mode)) {
 		if (ia32_cg_config.use_sse2) {
-			return ia32_new_Unknown_xmm(env_cg);
+			res = new_bd_ia32_xUnknown(dbgi, block);
 		} else {
-			/* Unknown nodes are buggy in x87 simulator, use zero for now... */
-			ir_graph *irg   = current_ir_graph;
-			dbg_info *dbgi  = get_irn_dbg_info(node);
-			ir_node  *block = get_irg_start_block(irg);
-			ir_node  *ret   = new_rd_ia32_vfldz(dbgi, irg, block);
-
-			/* Const Nodes before the initial IncSP are a bad idea, because
-			 * they could be spilled and we have no SP ready at that point yet.
-			 * So add a dependency to the initial frame pointer calculation to
-			 * avoid that situation.
-			 */
-			add_irn_dep(ret, get_irg_frame(irg));
-			return ret;
+			res = new_bd_ia32_vfldz(dbgi, block);
 		}
 	} else if (ia32_mode_needs_gp_reg(mode)) {
-		return ia32_new_Unknown_gp(env_cg);
+		res = new_bd_ia32_Unknown(dbgi, block);
 	} else {
 		panic("unsupported Unknown-Mode");
 	}
-	return NULL;
+
+	return res;
 }
 
-const arch_register_req_t *make_register_req(const constraint_t *constraint,
+const arch_register_req_t *ia32_make_register_req(const constraint_t *constraint,
 		int n_outs, const arch_register_req_t **out_reqs, int pos)
 {
 	struct obstack      *obst    = get_irg_obstack(current_ir_graph);
@@ -800,10 +817,11 @@ const arch_register_req_t *make_register_req(const constraint_t *constraint,
 
 		other_constr     = out_reqs[same_as];
 
-		req              = obstack_alloc(obst, sizeof(req[0]));
+		req              = OALLOC(obst, arch_register_req_t);
 		*req             = *other_constr;
 		req->type       |= arch_register_req_type_should_be_same;
 		req->other_same  = 1U << pos;
+		req->width       = 1;
 
 		/* switch constraints. This is because in firm we have same_as
 		 * constraints on the output constraints while in the gcc asm syntax
@@ -814,14 +832,14 @@ const arch_register_req_t *make_register_req(const constraint_t *constraint,
 
 	/* pure memory ops */
 	if (constraint->cls == NULL) {
-		return &no_register_req;
+		return arch_no_register_req;
 	}
 
 	if (constraint->allowed_registers != 0
 			&& !constraint->all_registers_allowed) {
 		unsigned *limited_ptr;
 
-		req         = obstack_alloc(obst, sizeof(req[0]) + sizeof(unsigned));
+		req         = (arch_register_req_t*)obstack_alloc(obst, sizeof(req[0]) + sizeof(unsigned));
 		memset(req, 0, sizeof(req[0]));
 		limited_ptr = (unsigned*) (req+1);
 
@@ -829,138 +847,136 @@ const arch_register_req_t *make_register_req(const constraint_t *constraint,
 		*limited_ptr = constraint->allowed_registers;
 		req->limited = limited_ptr;
 	} else {
-		req       = obstack_alloc(obst, sizeof(req[0]));
-		memset(req, 0, sizeof(req[0]));
+		req       = OALLOCZ(obst, arch_register_req_t);
 		req->type = arch_register_req_type_normal;
 	}
-	req->cls = constraint->cls;
+	req->cls   = constraint->cls;
+	req->width = 1;
 
 	return req;
 }
 
-const arch_register_req_t *parse_clobber(const char *clobber)
+const arch_register_req_t *ia32_parse_clobber(const char *clobber)
 {
 	struct obstack        *obst = get_irg_obstack(current_ir_graph);
 	const arch_register_t *reg  = ia32_get_clobber_register(clobber);
 	arch_register_req_t   *req;
 	unsigned              *limited;
 
-	if(reg == NULL) {
+	if (reg == NULL) {
 		panic("Register '%s' mentioned in asm clobber is unknown", clobber);
 	}
 
 	assert(reg->index < 32);
 
-	limited  = obstack_alloc(obst, sizeof(limited[0]));
+	limited  = OALLOC(obst, unsigned);
 	*limited = 1 << reg->index;
 
-	req          = obstack_alloc(obst, sizeof(req[0]));
-	memset(req, 0, sizeof(req[0]));
+	req          = OALLOCZ(obst, arch_register_req_t);
 	req->type    = arch_register_req_type_limited;
 	req->cls     = arch_register_get_class(reg);
 	req->limited = limited;
+	req->width   = 1;
 
 	return req;
 }
 
-ir_node *try_create_Immediate(ir_node *node, char immediate_constraint_type)
+
+int ia32_prevents_AM(ir_node *const block, ir_node *const am_candidate,
+                       ir_node *const other)
 {
-	int          minus         = 0;
-	tarval      *offset        = NULL;
-	int          offset_sign   = 0;
-	long         val = 0;
-	ir_entity   *symconst_ent  = NULL;
-	int          symconst_sign = 0;
-	ir_mode     *mode;
-	ir_node     *cnst          = NULL;
-	ir_node     *symconst      = NULL;
-	ir_node     *new_node;
+	if (get_nodes_block(other) != block)
+		return 0;
 
-	mode = get_irn_mode(node);
-	if(!mode_is_int(mode) && !mode_is_reference(mode)) {
-		return NULL;
+	if (is_Sync(other)) {
+		int i;
+
+		for (i = get_Sync_n_preds(other) - 1; i >= 0; --i) {
+			ir_node *const pred = get_Sync_pred(other, i);
+
+			if (get_nodes_block(pred) != block)
+				continue;
+
+			/* Do not block ourselves from getting eaten */
+			if (is_Proj(pred) && get_Proj_pred(pred) == am_candidate)
+				continue;
+
+			if (!heights_reachable_in_block(ia32_heights, pred, am_candidate))
+				continue;
+
+			return 1;
+		}
+
+		return 0;
+	} else {
+		/* Do not block ourselves from getting eaten */
+		if (is_Proj(other) && get_Proj_pred(other) == am_candidate)
+			return 0;
+
+		if (!heights_reachable_in_block(ia32_heights, other, am_candidate))
+			return 0;
+
+		return 1;
 	}
+}
+
+ir_node *ia32_try_create_Immediate(ir_node *node, char immediate_constraint_type)
+{
+	long       val = 0;
+	ir_entity *symconst_ent  = NULL;
+	ir_mode   *mode;
+	ir_node   *cnst          = NULL;
+	ir_node   *symconst      = NULL;
+	ir_node   *new_node;
 
-	if(is_Minus(node)) {
-		minus = 1;
-		node  = get_Minus_op(node);
+	mode = get_irn_mode(node);
+	if (!mode_is_int(mode) && !mode_is_reference(mode)) {
+		return NULL;
 	}
 
-	if(is_Const(node)) {
-		cnst        = node;
-		symconst    = NULL;
-		offset_sign = minus;
-	} else if(is_SymConst(node)) {
-		cnst          = NULL;
-		symconst      = node;
-		symconst_sign = minus;
-	} else if(is_Add(node)) {
+	if (is_Const(node)) {
+		cnst     = node;
+		symconst = NULL;
+	} else if (is_SymConst_addr_ent(node)
+			&& get_entity_owner(get_SymConst_entity(node)) != get_tls_type()) {
+		cnst     = NULL;
+		symconst = node;
+	} else if (is_Add(node)) {
 		ir_node *left  = get_Add_left(node);
 		ir_node *right = get_Add_right(node);
-		if(is_Const(left) && is_SymConst(right)) {
-			cnst          = left;
-			symconst      = right;
-			symconst_sign = minus;
-			offset_sign   = minus;
-		} else if(is_SymConst(left) && is_Const(right)) {
-			cnst          = right;
-			symconst      = left;
-			symconst_sign = minus;
-			offset_sign   = minus;
-		}
-	} else if(is_Sub(node)) {
-		ir_node *left  = get_Sub_left(node);
-		ir_node *right = get_Sub_right(node);
-		if(is_Const(left) && is_SymConst(right)) {
-			cnst          = left;
-			symconst      = right;
-			symconst_sign = !minus;
-			offset_sign   = minus;
-		} else if(is_SymConst(left) && is_Const(right)) {
-			cnst          = right;
-			symconst      = left;
-			symconst_sign = minus;
-			offset_sign   = !minus;
+		if (is_Const(left) && is_SymConst_addr_ent(right)) {
+			cnst     = left;
+			symconst = right;
+		} else if (is_SymConst_addr_ent(left) && is_Const(right)) {
+			cnst     = right;
+			symconst = left;
 		}
 	} else {
 		return NULL;
 	}
 
-	if(cnst != NULL) {
-		offset = get_Const_tarval(cnst);
-		if(tarval_is_long(offset)) {
-			val = get_tarval_long(offset);
-		} else {
-			ir_fprintf(stderr, "Optimisation Warning: tarval from %+F is not a "
-			           "long?\n", cnst);
+	if (cnst != NULL) {
+		ir_tarval *offset = get_Const_tarval(cnst);
+		if (!tarval_is_long(offset)) {
+			ir_fprintf(stderr, "Optimisation Warning: tarval of %+F is not a long?\n", cnst);
 			return NULL;
 		}
 
-		if(!check_immediate_constraint(val, immediate_constraint_type))
+		val = get_tarval_long(offset);
+		if (!check_immediate_constraint(val, immediate_constraint_type))
 			return NULL;
 	}
-	if(symconst != NULL) {
-		if(immediate_constraint_type != 0) {
+	if (symconst != NULL) {
+		if (immediate_constraint_type != 0) {
 			/* we need full 32bits for symconsts */
 			return NULL;
 		}
 
-		/* unfortunately the assembler/linker doesn't support -symconst */
-		if(symconst_sign)
-			return NULL;
-
-		if(get_SymConst_kind(symconst) != symconst_addr_ent)
-			return NULL;
-		symconst_ent = get_SymConst_entity(symconst);
+		symconst_ent = get_Global_entity(symconst);
 	}
-	if(cnst == NULL && symconst == NULL)
+	if (cnst == NULL && symconst == NULL)
 		return NULL;
 
-	if(offset_sign && offset != NULL) {
-		offset = tarval_neg(offset);
-	}
-
-	new_node = create_Immediate(symconst_ent, symconst_sign, val);
-
+	new_node = ia32_create_Immediate(symconst_ent, 0, val);
 	return new_node;
 }