ia32_transform.c

/**
 * This file implements the IR transformation from firm into ia32-Firm.
 * @author Christian Wuerdig
 * $Id$
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include <limits.h>

#include "irargs_t.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irmode_t.h"
#include "iropt_t.h"
#include "irop_t.h"
#include "irprog_t.h"
#include "iredges_t.h"
#include "irgmod.h"
#include "irvrfy.h"
#include "ircons.h"
#include "dbginfo.h"
#include "irprintf.h"
#include "debug.h"
#include "irdom.h"
#include "type.h"
#include "entity.h"
#include "archop.h"     /* we need this for Min and Max nodes */

#include "../benode_t.h"
#include "../besched.h"
#include "../beabi.h"

#include "bearch_ia32_t.h"
#include "ia32_nodes_attr.h"
#include "ia32_transform.h"
#include "ia32_new_nodes.h"
#include "ia32_map_regs.h"
#include "ia32_dbg_stat.h"
#include "ia32_optimize.h"
#include "ia32_util.h"

#include "gen_ia32_regalloc_if.h"

#define SFP_SIGN "0x80000000"
#define DFP_SIGN "0x8000000000000000"
#define SFP_ABS  "0x7FFFFFFF"
#define DFP_ABS  "0x7FFFFFFFFFFFFFFF"

#define TP_SFP_SIGN "ia32_sfp_sign"
#define TP_DFP_SIGN "ia32_dfp_sign"
#define TP_SFP_ABS  "ia32_sfp_abs"
#define TP_DFP_ABS  "ia32_dfp_abs"

#define ENT_SFP_SIGN "IA32_SFP_SIGN"
#define ENT_DFP_SIGN "IA32_DFP_SIGN"
#define ENT_SFP_ABS  "IA32_SFP_ABS"
#define ENT_DFP_ABS  "IA32_DFP_ABS"

extern ir_op *get_op_Mulh(void);

typedef ir_node *construct_binop_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \
									  ir_node *op1, ir_node *op2, ir_node *mem);

typedef ir_node *construct_unop_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \
									 ir_node *op, ir_node *mem);

typedef enum {
	ia32_SSIGN, ia32_DSIGN, ia32_SABS, ia32_DABS, ia32_known_const_max
} ia32_known_const_t;

/****************************************************************************************************
 *                  _        _                        __                           _   _
 *                 | |      | |                      / _|                         | | (_)
 *  _ __   ___   __| | ___  | |_ _ __ __ _ _ __  ___| |_ ___  _ __ _ __ ___   __ _| |_ _  ___  _ __
 * | '_ \ / _ \ / _` |/ _ \ | __| '__/ _` | '_ \/ __|  _/ _ \| '__| '_ ` _ \ / _` | __| |/ _ \| '_ \
 * | | | | (_) | (_| |  __/ | |_| | | (_| | | | \__ \ || (_) | |  | | | | | | (_| | |_| | (_) | | | |
 * |_| |_|\___/ \__,_|\___|  \__|_|  \__,_|_| |_|___/_| \___/|_|  |_| |_| |_|\__,_|\__|_|\___/|_| |_|
 *
 ****************************************************************************************************/

/**
 * Returns 1 if irn is a Const representing 0, 0 otherwise
 */
static INLINE int is_ia32_Const_0(ir_node *irn) {
	return (is_ia32_irn(irn) && get_ia32_op_type(irn) == ia32_Const) ?
		classify_tarval(get_ia32_Immop_tarval(irn)) == TV_CLASSIFY_NULL : 0;
}

/**
 * Returns 1 if irn is a Const representing 1, 0 otherwise
 */
static INLINE int is_ia32_Const_1(ir_node *irn) {
	return (is_ia32_irn(irn) && get_ia32_op_type(irn) == ia32_Const) ?
		classify_tarval(get_ia32_Immop_tarval(irn)) == TV_CLASSIFY_ONE : 0;
}

/**
 * Returns the Proj representing the UNKNOWN register for given mode.
 */
static ir_node *be_get_unknown_for_mode(ia32_code_gen_t *cg, ir_mode *mode) {
	be_abi_irg_t          *babi       = cg->birg->abi;
	const arch_register_t *unknwn_reg = NULL;

	if (mode_is_float(mode)) {
		unknwn_reg = USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_UKNWN] : &ia32_vfp_regs[REG_VFP_UKNWN];
	}
	else {
		unknwn_reg = &ia32_gp_regs[REG_GP_UKNWN];
	}

	return be_abi_get_callee_save_irn(babi, unknwn_reg);
}

/**
 * Gets the Proj with number pn from irn.
 */
static ir_node *get_proj_for_pn(const ir_node *irn, long pn) {
	const ir_edge_t *edge;
	ir_node   *proj;
	assert(get_irn_mode(irn) == mode_T && "need mode_T");

	foreach_out_edge(irn, edge) {
		proj = get_edge_src_irn(edge);

		if (get_Proj_proj(proj) == pn)
			return proj;
	}

	return NULL;
}

/**
 * SSE convert of an integer node into a floating point node.
 */
static ir_node *gen_sse_conv_int2float(ia32_code_gen_t *cg, dbg_info *dbg, ir_graph *irg, ir_node *block,
                                       ir_node *in, ir_node *old_node, ir_mode *tgt_mode)
{
	ir_node *noreg = ia32_new_NoReg_gp(cg);
	ir_node *nomem = new_rd_NoMem(irg);

	ir_node *conv = new_rd_ia32_Conv_I2FP(dbg, irg, block, noreg, noreg, in, nomem);
	set_ia32_src_mode(conv, get_irn_mode(in));
	set_ia32_tgt_mode(conv, tgt_mode);
	set_ia32_am_support(conv, ia32_am_Source);
	SET_IA32_ORIG_NODE(conv, ia32_get_old_node_name(cg, old_node));

	return new_rd_Proj(dbg, irg, block, conv, tgt_mode, pn_ia32_Conv_I2FP_res);
}

/**
* SSE convert of an float node into a double node.
*/
static ir_node *gen_sse_conv_f2d(ia32_code_gen_t *cg, dbg_info *dbg, ir_graph *irg, ir_node *block,
									   ir_node *in, ir_node *old_node)
{
	ir_node *noreg = ia32_new_NoReg_gp(cg);
	ir_node *nomem = new_rd_NoMem(irg);

	ir_node *conv = new_rd_ia32_Conv_FP2FP(dbg, irg, block, noreg, noreg, in, nomem);
	set_ia32_src_mode(conv, mode_F);
	set_ia32_tgt_mode(conv, mode_D);
	set_ia32_am_support(conv, ia32_am_Source);
	SET_IA32_ORIG_NODE(conv, ia32_get_old_node_name(cg, old_node));

	return new_rd_Proj(dbg, irg, block, conv, mode_D, pn_ia32_Conv_FP2FP_res);
}

/* Generates an entity for a known FP const (used for FP Neg + Abs) */
static ident *gen_fp_known_const(ia32_known_const_t kct) {
	static const struct {
		const char *tp_name;
		const char *ent_name;
		const char *cnst_str;
	} names [ia32_known_const_max] = {
		{ TP_SFP_SIGN, ENT_SFP_SIGN, SFP_SIGN },	/* ia32_SSIGN */
		{ TP_DFP_SIGN, ENT_DFP_SIGN, DFP_SIGN },	/* ia32_DSIGN */
		{ TP_SFP_ABS,  ENT_SFP_ABS,  SFP_ABS },		/* ia32_SABS */
		{ TP_DFP_ABS,  ENT_DFP_ABS,  DFP_ABS }		/* ia32_DABS */
	};
	static struct entity *ent_cache[ia32_known_const_max];

	const char    *tp_name, *ent_name, *cnst_str;
	ir_type       *tp;
	ir_node       *cnst;
	ir_graph      *rem;
	entity        *ent;
	tarval        *tv;
	ir_mode       *mode;

	ent_name = names[kct].ent_name;
	if (! ent_cache[kct]) {
		tp_name  = names[kct].tp_name;
		cnst_str = names[kct].cnst_str;

		mode = kct == ia32_SSIGN || kct == ia32_SABS ? mode_Iu : mode_Lu;
		tv  = new_tarval_from_str(cnst_str, strlen(cnst_str), mode);
		tp  = new_type_primitive(new_id_from_str(tp_name), mode);
		ent = new_entity(get_glob_type(), new_id_from_str(ent_name), tp);

		set_entity_ld_ident(ent, get_entity_ident(ent));
		set_entity_visibility(ent, visibility_local);
		set_entity_variability(ent, variability_constant);
		set_entity_allocation(ent, 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();
		cnst = new_Const(mode, tv);
		current_ir_graph = rem;

		set_atomic_ent_value(ent, cnst);

		/* cache the entry */
		ent_cache[kct] = ent;
	}

	return get_entity_ident(ent_cache[kct]);
}

#ifndef NDEBUG
/**
 * Prints the old node name on cg obst and returns a pointer to it.
 */
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->isa;

	lc_eoprintf(firm_get_arg_env(), isa->name_obst, "%+F", irn);
	obstack_1grow(isa->name_obst, 0);
	isa->name_obst_size += obstack_object_size(isa->name_obst);
 	return obstack_finish(isa->name_obst);
}
#endif /* NDEBUG */

/* determine if one operator is an Imm */
static ir_node *get_immediate_op(ir_node *op1, ir_node *op2) {
	if (op1)
		return is_ia32_Cnst(op1) ? op1 : (is_ia32_Cnst(op2) ? op2 : NULL);
	else return is_ia32_Cnst(op2) ? op2 : NULL;
}

/* determine if one operator is not an Imm */
static ir_node *get_expr_op(ir_node *op1, ir_node *op2) {
	return !is_ia32_Cnst(op1) ? op1 : (!is_ia32_Cnst(op2) ? op2 : NULL);
}


/**
 * Construct a standard binary operation, set AM and immediate if required.
 *
 * @param env   The transformation environment
 * @param op1   The first operand
 * @param op2   The second operand
 * @param func  The node constructor function
 * @return The constructed ia32 node.
 */
static ir_node *gen_binop(ia32_transform_env_t *env, ir_node *op1, ir_node *op2, construct_binop_func *func) {
	ir_node  *new_op   = NULL;
	ir_mode  *mode     = env->mode;
	dbg_info *dbg      = env->dbg;
	ir_graph *irg      = env->irg;
	ir_node  *block    = env->block;
	ir_node  *noreg_gp = ia32_new_NoReg_gp(env->cg);
	ir_node  *noreg_fp = ia32_new_NoReg_fp(env->cg);
	ir_node  *nomem    = new_NoMem();
	int      is_mul    = 0;
	ir_node  *expr_op, *imm_op;
	DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)

	/* Check if immediate optimization is on and */
	/* if it's an operation with immediate.      */
	/* Mul/MulS/Mulh don't support immediates    */
	if (! (env->cg->opt & IA32_OPT_IMMOPS) ||
		func == new_rd_ia32_Mul            ||
		func == new_rd_ia32_Mulh           ||
		func == new_rd_ia32_MulS)
	{
		expr_op = op1;
		imm_op  = NULL;
		/* immediate operations are requested, but we are here: it a mul */
		if (env->cg->opt & IA32_OPT_IMMOPS)
			is_mul = 1;
	}
	else if (is_op_commutative(get_irn_op(env->irn))) {
		imm_op  = get_immediate_op(op1, op2);
		expr_op = get_expr_op(op1, op2);
	}
	else {
		imm_op  = get_immediate_op(NULL, op2);
		expr_op = get_expr_op(op1, op2);
	}

	assert((expr_op || imm_op) && "invalid operands");

	if (!expr_op) {
		/* We have two consts here: not yet supported */
		imm_op = NULL;
	}

	if (mode_is_float(mode)) {
		/* floating point operations */
		if (imm_op) {
			DB((mod, LEVEL_1, "FP with immediate ..."));
			new_op = func(dbg, irg, block, noreg_gp, noreg_gp, expr_op, noreg_fp, nomem);
			set_ia32_Immop_attr(new_op, imm_op);
			set_ia32_am_support(new_op, ia32_am_None);
		}
		else {
			DB((mod, LEVEL_1, "FP binop ..."));
			new_op = func(dbg, irg, block, noreg_gp, noreg_gp, op1, op2, nomem);
			set_ia32_am_support(new_op, ia32_am_Source);
		}
		set_ia32_ls_mode(new_op, mode);
	}
	else {
		/* integer operations */
		if (imm_op) {
			/* This is expr + const */
			DB((mod, LEVEL_1, "INT with immediate ..."));
			new_op = func(dbg, irg, block, noreg_gp, noreg_gp, expr_op, noreg_gp, nomem);
			set_ia32_Immop_attr(new_op, imm_op);

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Dest);
		}
		else {
			DB((mod, LEVEL_1, "INT binop ..."));
			/* This is a normal operation */
			new_op = func(dbg, irg, block, noreg_gp, noreg_gp, op1, op2, nomem);

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Full);
		}

		/* Muls can only have AM source */
		if (is_mul)
			set_ia32_am_support(new_op, ia32_am_Source);
	}

	SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));

	set_ia32_res_mode(new_op, mode);

	if (is_op_commutative(get_irn_op(env->irn))) {
		set_ia32_commutative(new_op);
	}

	return new_rd_Proj(dbg, irg, block, new_op, mode, 0);
}



/**
 * Construct a shift/rotate binary operation, sets AM and immediate if required.
 *
 * @param env   The transformation environment
 * @param op1   The first operand
 * @param op2   The second operand
 * @param func  The node constructor function
 * @return The constructed ia32 node.
 */
static ir_node *gen_shift_binop(ia32_transform_env_t *env, ir_node *op1, ir_node *op2, construct_binop_func *func) {
	ir_node           *new_op = NULL;
	ir_mode           *mode   = env->mode;
	dbg_info          *dbg    = env->dbg;
	ir_graph          *irg    = env->irg;
	ir_node           *block  = env->block;
	ir_node           *noreg  = ia32_new_NoReg_gp(env->cg);
	ir_node           *nomem  = new_NoMem();
	ir_node           *expr_op, *imm_op;
	tarval            *tv;
	DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)

	assert(! mode_is_float(mode) && "Shift/Rotate with float not supported");

	/* Check if immediate optimization is on and */
	/* if it's an operation with immediate.      */
	imm_op  = (env->cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, op2) : NULL;
	expr_op = get_expr_op(op1, op2);

	assert((expr_op || imm_op) && "invalid operands");

	if (!expr_op) {
		/* We have two consts here: not yet supported */
		imm_op = NULL;
	}

	/* Limit imm_op within range imm8 */
	if (imm_op) {
		tv = get_ia32_Immop_tarval(imm_op);

		if (tv) {
			tv = tarval_mod(tv, new_tarval_from_long(32, mode_Iu));
			set_ia32_Immop_tarval(imm_op, tv);
		}
		else {
			imm_op = NULL;
		}
	}

	/* integer operations */
	if (imm_op) {
		/* This is shift/rot with const */
		DB((mod, LEVEL_1, "Shift/Rot with immediate ..."));

		new_op = func(dbg, irg, block, noreg, noreg, expr_op, noreg, nomem);
		set_ia32_Immop_attr(new_op, imm_op);
	}
	else {
		/* This is a normal shift/rot */
		DB((mod, LEVEL_1, "Shift/Rot binop ..."));
		new_op = func(dbg, irg, block, noreg, noreg, op1, op2, nomem);
	}

	/* set AM support */
	set_ia32_am_support(new_op, ia32_am_Dest);

	SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));

	set_ia32_res_mode(new_op, mode);
	set_ia32_emit_cl(new_op);

	return new_rd_Proj(dbg, irg, block, new_op, mode, 0);
}


/**
 * Construct a standard unary operation, set AM and immediate if required.
 *
 * @param env   The transformation environment
 * @param op    The operand
 * @param func  The node constructor function
 * @return The constructed ia32 node.
 */
static ir_node *gen_unop(ia32_transform_env_t *env, ir_node *op, construct_unop_func *func) {
	ir_node           *new_op = NULL;
	ir_mode           *mode   = env->mode;
	dbg_info          *dbg    = env->dbg;
	ir_graph          *irg    = env->irg;
	ir_node           *block  = env->block;
	ir_node           *noreg  = ia32_new_NoReg_gp(env->cg);
	ir_node           *nomem  = new_NoMem();
	DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)

	new_op = func(dbg, irg, block, noreg, noreg, op, nomem);

	if (mode_is_float(mode)) {
		DB((mod, LEVEL_1, "FP unop ..."));
		/* floating point operations don't support implicit store */
		set_ia32_am_support(new_op, ia32_am_None);
	}
	else {
		DB((mod, LEVEL_1, "INT unop ..."));
		set_ia32_am_support(new_op, ia32_am_Dest);
	}

	SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));

	set_ia32_res_mode(new_op, mode);

	return new_rd_Proj(dbg, irg, block, new_op, mode, 0);
}



/**
 * Creates an ia32 Add with immediate.
 *
 * @param env       The transformation environment
 * @param expr_op   The expression operator
 * @param const_op  The constant
 * @return the created ia32 Add node
 */
static ir_node *gen_imm_Add(ia32_transform_env_t *env, ir_node *expr_op, ir_node *const_op) {
	ir_node                *new_op     = NULL;
	tarval                 *tv         = get_ia32_Immop_tarval(const_op);
	dbg_info               *dbg        = env->dbg;
	ir_graph               *irg        = env->irg;
	ir_node                *block      = env->block;
	ir_node                *noreg      = ia32_new_NoReg_gp(env->cg);
	ir_node                *nomem      = new_NoMem();
	int                     normal_add = 1;
	tarval_classification_t class_tv, class_negtv;
	DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)

	/* try to optimize to inc/dec  */
	if ((env->cg->opt & IA32_OPT_INCDEC) && (get_ia32_op_type(const_op) == ia32_Const)) {
		/* optimize tarvals */
		class_tv    = classify_tarval(tv);
		class_negtv = classify_tarval(tarval_neg(tv));

		if (class_tv == TV_CLASSIFY_ONE) { /* + 1 == INC */
			DB((env->mod, LEVEL_2, "Add(1) to Inc ... "));
			new_op     = new_rd_ia32_Inc(dbg, irg, block, noreg, noreg, expr_op, nomem);
			normal_add = 0;
		}
		else if (class_tv == TV_CLASSIFY_ALL_ONE || class_negtv == TV_CLASSIFY_ONE) { /* + (-1) == DEC */
			DB((mod, LEVEL_2, "Add(-1) to Dec ... "));
			new_op     = new_rd_ia32_Dec(dbg, irg, block, noreg, noreg, expr_op, nomem);
			normal_add = 0;
		}
	}

	if (normal_add) {
		new_op = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, expr_op, noreg, nomem);
		set_ia32_Immop_attr(new_op, const_op);
		set_ia32_commutative(new_op);
	}

	return new_op;
}

/**
 * Creates an ia32 Add.
 *
 * @param env   The transformation environment
 * @return the created ia32 Add node
 */
static ir_node *gen_Add(ia32_transform_env_t *env) {
	ir_node  *new_op = NULL;
	dbg_info *dbg    = env->dbg;
	ir_mode  *mode   = env->mode;
	ir_graph *irg    = env->irg;
	ir_node  *block  = env->block;
	ir_node  *noreg  = ia32_new_NoReg_gp(env->cg);
	ir_node  *nomem  = new_NoMem();
	ir_node  *expr_op, *imm_op;
	ir_node  *op1    = get_Add_left(env->irn);
	ir_node  *op2    = get_Add_right(env->irn);

	/* Check if immediate optimization is on and */
	/* if it's an operation with immediate.      */
	imm_op  = (env->cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(op1, op2) : NULL;
	expr_op = get_expr_op(op1, op2);

	assert((expr_op || imm_op) && "invalid operands");

	if (mode_is_float(mode)) {
		FP_USED(env->cg);
		if (USE_SSE2(env->cg))
			return gen_binop(env, op1, op2, new_rd_ia32_xAdd);
		else
			return gen_binop(env, op1, op2, new_rd_ia32_vfadd);
	}
	else {
		/* integer ADD */
		if (!expr_op) {
			/* No expr_op means, that we have two const - one symconst and */
			/* one tarval or another symconst - because this case is not   */
			/* covered by constant folding                                 */
			/* We need to check for:                                       */
			/*  1) symconst + const    -> becomes a LEA                    */
			/*  2) symconst + symconst -> becomes a const + LEA as the elf */
			/*        linker doesn't support two symconsts                 */

			if (get_ia32_op_type(op1) == ia32_SymConst && get_ia32_op_type(op2) == ia32_SymConst) {
				/* this is the 2nd case */
				new_op = new_rd_ia32_Lea(dbg, irg, block, op1, noreg, mode);
				set_ia32_am_sc(new_op, get_ia32_id_cnst(op2));
				set_ia32_am_flavour(new_op, ia32_am_OB);

				DBG_OPT_LEA3(op1, op2, env->irn, new_op);
			}
			else {
				/* this is the 1st case */
				new_op = new_rd_ia32_Lea(dbg, irg, block, noreg, noreg, mode);

				DBG_OPT_LEA3(op1, op2, env->irn, new_op);

				if (get_ia32_op_type(op1) == ia32_SymConst) {
					set_ia32_am_sc(new_op, get_ia32_id_cnst(op1));
					add_ia32_am_offs(new_op, get_ia32_cnst(op2));
				}
				else {
					add_ia32_am_offs(new_op, get_ia32_cnst(op1));
					set_ia32_am_sc(new_op, get_ia32_id_cnst(op2));
				}
				set_ia32_am_flavour(new_op, ia32_am_O);
			}

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Source);
			set_ia32_op_type(new_op, ia32_AddrModeS);

			/* Lea doesn't need a Proj */
			return new_op;
		}
		else if (imm_op) {
			/* This is expr + const */
			new_op = gen_imm_Add(env, expr_op, imm_op);

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Dest);
		}
		else {
			/* This is a normal add */
			new_op = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, op1, op2, nomem);

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Full);
			set_ia32_commutative(new_op);
		}
	}

	SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));

	set_ia32_res_mode(new_op, mode);

	return new_rd_Proj(dbg, irg, block, new_op, mode, pn_ia32_Add_res);
}



/**
 * Creates an ia32 Mul.
 *
 * @param env   The transformation environment
 * @return the created ia32 Mul node
 */
static ir_node *gen_Mul(ia32_transform_env_t *env) {
	ir_node *op1 = get_Mul_left(env->irn);
	ir_node *op2 = get_Mul_right(env->irn);
	ir_node *new_op;

	if (mode_is_float(env->mode)) {
		FP_USED(env->cg);
		if (USE_SSE2(env->cg))
			new_op = gen_binop(env, op1, op2, new_rd_ia32_xMul);
		else
			new_op = gen_binop(env, op1, op2, new_rd_ia32_vfmul);
	}
	else {
		new_op = gen_binop(env, op1, op2, new_rd_ia32_Mul);
	}

	return new_op;
}



/**
 * Creates an ia32 Mulh.
 * Note: Mul produces a 64Bit result and Mulh returns the upper 32 bit of
 * this result while Mul returns the lower 32 bit.
 *
 * @param env   The transformation environment
 * @return the created ia32 Mulh node
 */
static ir_node *gen_Mulh(ia32_transform_env_t *env) {
	ir_node *op1 = get_irn_n(env->irn, 0);
	ir_node *op2 = get_irn_n(env->irn, 1);
	ir_node *proj_EAX, *proj_EDX, *mulh;
	ir_node *in[1];

	assert(!mode_is_float(env->mode) && "Mulh with float not supported");
	proj_EAX = gen_binop(env, op1, op2, new_rd_ia32_Mulh);
	mulh     = get_Proj_pred(proj_EAX);
	proj_EDX = new_rd_Proj(env->dbg, env->irg, env->block, mulh, env->mode, pn_EDX);

	/* to be on the save side */
	set_Proj_proj(proj_EAX, pn_EAX);

	if (is_ia32_ImmConst(mulh) || is_ia32_ImmSymConst(mulh)) {
		/* Mulh with const cannot have AM */
		set_ia32_am_support(mulh, ia32_am_None);
	}
	else {
		/* Mulh cannot have AM for destination */
		set_ia32_am_support(mulh, ia32_am_Source);
	}

	in[0] = proj_EAX;

	/* keep EAX */
	be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], env->irg, env->block, 1, in);

	return proj_EDX;
}



/**
 * Creates an ia32 And.
 *
 * @param env   The transformation environment
 * @return The created ia32 And node
 */
static ir_node *gen_And(ia32_transform_env_t *env) {
	ir_node *op1 = get_And_left(env->irn);
	ir_node *op2 = get_And_right(env->irn);

	assert (! mode_is_float(env->mode));
	return gen_binop(env, op1, op2, new_rd_ia32_And);
}



/**
 * Creates an ia32 Or.
 *
 * @param env   The transformation environment
 * @return The created ia32 Or node
 */
static ir_node *gen_Or(ia32_transform_env_t *env) {
	ir_node *op1 = get_Or_left(env->irn);
	ir_node *op2 = get_Or_right(env->irn);

	assert (! mode_is_float(env->mode));
	return gen_binop(env, op1, op2, new_rd_ia32_Or);
}



/**
 * Creates an ia32 Eor.
 *
 * @param env   The transformation environment
 * @return The created ia32 Eor node
 */
static ir_node *gen_Eor(ia32_transform_env_t *env) {
	ir_node *op1 = get_Eor_left(env->irn);
	ir_node *op2 = get_Eor_right(env->irn);

	assert(! mode_is_float(env->mode));
	return gen_binop(env, op1, op2, new_rd_ia32_Eor);
}



/**
 * Creates an ia32 Max.
 *
 * @param env      The transformation environment
 * @return the created ia32 Max node
 */
static ir_node *gen_Max(ia32_transform_env_t *env) {
	ir_node *op1 = get_irn_n(env->irn, 0);
	ir_node *op2 = get_irn_n(env->irn, 1);
	ir_node *new_op;

	if (mode_is_float(env->mode)) {
		FP_USED(env->cg);
		if (USE_SSE2(env->cg))
			new_op = gen_binop(env, op1, op2, new_rd_ia32_xMax);
		else {
			assert(0);
		}
	}
	else {
		new_op = new_rd_ia32_Max(env->dbg, env->irg, env->block, op1, op2, env->mode);
		set_ia32_am_support(new_op, ia32_am_None);
		SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
	}

	return new_op;
}



/**
 * Creates an ia32 Min.
 *
 * @param env      The transformation environment
 * @return the created ia32 Min node
 */
static ir_node *gen_Min(ia32_transform_env_t *env) {
	ir_node *op1 = get_irn_n(env->irn, 0);
	ir_node *op2 = get_irn_n(env->irn, 1);
	ir_node *new_op;

	if (mode_is_float(env->mode)) {
		FP_USED(env->cg);
		if (USE_SSE2(env->cg))
			new_op = gen_binop(env, op1, op2, new_rd_ia32_xMin);
		else {
			assert(0);
		}
	}
	else {
		new_op = new_rd_ia32_Min(env->dbg, env->irg, env->block, op1, op2, env->mode);
		set_ia32_am_support(new_op, ia32_am_None);
		SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
	}

	return new_op;
}



/**
 * Creates an ia32 Sub with immediate.
 *
 * @param env        The transformation environment
 * @param expr_op    The first operator
 * @param const_op   The constant operator
 * @return The created ia32 Sub node
 */
static ir_node *gen_imm_Sub(ia32_transform_env_t *env, ir_node *expr_op, ir_node *const_op) {
	ir_node                *new_op     = NULL;
	tarval                 *tv         = get_ia32_Immop_tarval(const_op);
	dbg_info               *dbg        = env->dbg;
	ir_graph               *irg        = env->irg;
	ir_node                *block      = env->block;
	ir_node                *noreg      = ia32_new_NoReg_gp(env->cg);
	ir_node                *nomem      = new_NoMem();
	int                     normal_sub = 1;
	tarval_classification_t class_tv, class_negtv;
	DEBUG_ONLY(firm_dbg_module_t *mod = env->mod;)

	/* try to optimize to inc/dec  */
	if ((env->cg->opt & IA32_OPT_INCDEC) && tv) {
		/* optimize tarvals */
		class_tv    = classify_tarval(tv);
		class_negtv = classify_tarval(tarval_neg(tv));

		if (class_tv == TV_CLASSIFY_ONE) { /* - 1 == DEC */
			DB((mod, LEVEL_2, "Sub(1) to Dec ... "));
			new_op     = new_rd_ia32_Dec(dbg, irg, block, noreg, noreg, expr_op, nomem);
			normal_sub = 0;
		}
		else if (class_negtv == TV_CLASSIFY_ONE) { /* - (-1) == Sub */
			DB((mod, LEVEL_2, "Sub(-1) to Inc ... "));
			new_op     = new_rd_ia32_Inc(dbg, irg, block, noreg, noreg, expr_op, nomem);
			normal_sub = 0;
		}
	}

	if (normal_sub) {
		new_op = new_rd_ia32_Sub(dbg, irg, block, noreg, noreg, expr_op, noreg, nomem);
		set_ia32_Immop_attr(new_op, const_op);
	}

	return new_op;
}

/**
 * Creates an ia32 Sub.
 *
 * @param env   The transformation environment
 * @return The created ia32 Sub node
 */
static ir_node *gen_Sub(ia32_transform_env_t *env) {
	ir_node  *new_op = NULL;
	dbg_info *dbg    = env->dbg;
	ir_mode  *mode   = env->mode;
	ir_graph *irg    = env->irg;
	ir_node  *block  = env->block;
	ir_node  *noreg  = ia32_new_NoReg_gp(env->cg);
	ir_node  *nomem  = new_NoMem();
	ir_node  *op1    = get_Sub_left(env->irn);
	ir_node  *op2    = get_Sub_right(env->irn);
	ir_node  *expr_op, *imm_op;

	/* Check if immediate optimization is on and */
	/* if it's an operation with immediate.      */
	imm_op  = (env->cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, op2) : NULL;
	expr_op = get_expr_op(op1, op2);

	assert((expr_op || imm_op) && "invalid operands");

	if (mode_is_float(mode)) {
		FP_USED(env->cg);
		if (USE_SSE2(env->cg))
			return gen_binop(env, op1, op2, new_rd_ia32_xSub);
		else
			return gen_binop(env, op1, op2, new_rd_ia32_vfsub);
	}
	else {
		/* integer SUB */
		if (! expr_op) {
			/* No expr_op means, that we have two const - one symconst and */
			/* one tarval or another symconst - because this case is not   */
			/* covered by constant folding                                 */
			/* We need to check for:                                       */
			/*  1) symconst - const    -> becomes a LEA                    */
			/*  2) symconst - symconst -> becomes a const - LEA as the elf */
			/*        linker doesn't support two symconsts                 */

			if (get_ia32_op_type(op1) == ia32_SymConst && get_ia32_op_type(op2) == ia32_SymConst) {
				/* this is the 2nd case */
				new_op = new_rd_ia32_Lea(dbg, irg, block, op1, noreg, mode);
				set_ia32_am_sc(new_op, get_ia32_id_cnst(op2));
				set_ia32_am_sc_sign(new_op);
				set_ia32_am_flavour(new_op, ia32_am_OB);

				DBG_OPT_LEA3(op1, op2, env->irn, new_op);
			}
			else {
				/* this is the 1st case */
				new_op = new_rd_ia32_Lea(dbg, irg, block, noreg, noreg, mode);

				DBG_OPT_LEA3(op1, op2, env->irn, new_op);

				if (get_ia32_op_type(op1) == ia32_SymConst) {
					set_ia32_am_sc(new_op, get_ia32_id_cnst(op1));
					sub_ia32_am_offs(new_op, get_ia32_cnst(op2));
				}
				else {
					add_ia32_am_offs(new_op, get_ia32_cnst(op1));
					set_ia32_am_sc(new_op, get_ia32_id_cnst(op2));
					set_ia32_am_sc_sign(new_op);
				}
				set_ia32_am_flavour(new_op, ia32_am_O);
			}

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Source);
			set_ia32_op_type(new_op, ia32_AddrModeS);

			/* Lea doesn't need a Proj */
			return new_op;
		}
		else if (imm_op) {
			/* This is expr - const */
			new_op = gen_imm_Sub(env, expr_op, imm_op);

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Dest);
		}
		else {
			/* This is a normal sub */
			new_op = new_rd_ia32_Sub(dbg, irg, block, noreg, noreg, op1, op2, nomem);

			/* set AM support */
			set_ia32_am_support(new_op, ia32_am_Full);
		}
	}

	SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));

	set_ia32_res_mode(new_op, mode);

	return new_rd_Proj(dbg, irg, block, new_op, mode, pn_ia32_Sub_res);
}



/**
 * Generates an ia32 DivMod with additional infrastructure for the
 * register allocator if needed.
 *
 * @param env      The transformation environment
 * @param dividend -no comment- :)
 * @param divisor  -no comment- :)
 * @param dm_flav  flavour_Div/Mod/DivMod
 * @return The created ia32 DivMod node
 */
static ir_node *generate_DivMod(ia32_transform_env_t *env, ir_node *dividend, ir_node *divisor, ia32_op_flavour_t dm_flav) {
	ir_node  *res, *proj;
	ir_node  *edx_node, *cltd;
	ir_node  *in_keep[1];
	dbg_info *dbg   = env->dbg;
	ir_graph *irg   = env->irg;
	ir_node  *block = env->block;
	ir_mode  *mode  = env->mode;
	ir_node  *irn   = env->irn;
	ir_node  *mem;
	int      n;

	switch (dm_flav) {
		case flavour_Div:
			mem  = get_Div_mem(irn);
			mode = get_irn_mode(get_proj_for_pn(irn, pn_Div_res));
			break;
		case flavour_Mod:
			mem  = get_Mod_mem(irn);
			mode = get_irn_mode(get_proj_for_pn(irn, pn_Mod_res));
			break;
		case flavour_DivMod:
			mem  = get_DivMod_mem(irn);
			mode = get_irn_mode(get_proj_for_pn(irn, pn_DivMod_res_div));
			break;
		default:
			assert(0);
	}

	if (mode_is_signed(mode)) {
		/* in signed mode, we need to sign extend the dividend */
		cltd     = new_rd_ia32_Cdq(dbg, irg, block, dividend);
		dividend = new_rd_Proj(dbg, irg, block, cltd, mode_Is, pn_ia32_Cdq_EAX);
		edx_node = new_rd_Proj(dbg, irg, block, cltd, mode_Is, pn_ia32_Cdq_EDX);
	}
	else {
		edx_node = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), mode_Iu);
		set_ia32_Const_type(edx_node, ia32_Const);
		set_ia32_Immop_tarval(edx_node, get_tarval_null(mode_Iu));
	}

	res = new_rd_ia32_DivMod(dbg, irg, block, dividend, divisor, edx_node, mem, dm_flav);

	set_ia32_n_res(res, 2);

	/* Only one proj is used -> We must add a second proj and */
	/* connect this one to a Keep node to eat up the second   */
	/* destroyed register.                                    */
	n    = get_irn_n_edges(irn);
	proj = NULL;