lower_dw.c

/*
 * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
 *
 * This file is part of libFirm.
 *
 * This file may be distributed and/or modified under the terms of the
 * GNU General Public License version 2 as published by the Free Software
 * Foundation and appearing in the file LICENSE.GPL included in the
 * packaging of this file.
 *
 * Licensees holding valid libFirm Professional Edition licenses may use
 * this file in accordance with the libFirm Commercial License.
 * Agreement provided with the Software.
 *
 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE.
 */

/**
 * @file
 * @brief   Lower double word operations, i.e. 64bit -> 32bit, 32bit -> 16bit etc.
 * @date    8.10.2004
 * @author  Michael Beck
 * @version $Id$
 */
#include "config.h"

#include <string.h>
#include <stdlib.h>
#include <stdbool.h>
#include <assert.h>

#include "error.h"
#include "lowering.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irmode_t.h"
#include "iropt_t.h"
#include "irgmod.h"
#include "tv_t.h"
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
#include "firmstat.h"
#include "irgwalk.h"
#include "ircons.h"
#include "irflag.h"
#include "irtools.h"
#include "debug.h"
#include "set.h"
#include "pmap.h"
#include "pdeq.h"
#include "irdump.h"
#include "array_t.h"
#include "irpass_t.h"
#include "lower_dw.h"

/** A map from (op, imode, omode) to Intrinsic functions entities. */
static set *intrinsic_fkt;

/** A map from (imode, omode) to conv function types. */
static set *conv_types;

/** A map from a method type to its lowered type. */
static pmap *lowered_type;

/** The types for the binop and unop intrinsics. */
static ir_type *binop_tp_u, *binop_tp_s, *unop_tp_u, *unop_tp_s, *shiftop_tp_u, *shiftop_tp_s, *tp_s, *tp_u;

/** the debug handle */
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)

/**
 * An entry in the (op, imode, omode) -> entity map.
 */
typedef struct op_mode_entry {
	const ir_op   *op;    /**< the op */
	const ir_mode *imode; /**< the input mode */
	const ir_mode *omode; /**< the output mode */
	ir_entity     *ent;   /**< the associated entity of this (op, imode, omode) triple */
} op_mode_entry_t;

/**
 * An entry in the (imode, omode) -> tp map.
 */
typedef struct conv_tp_entry {
	const ir_mode *imode; /**< the input mode */
	const ir_mode *omode; /**< the output mode */
	ir_type       *mtd;   /**< the associated method type of this (imode, omode) pair */
} conv_tp_entry_t;

enum lower_flags {
	MUST_BE_LOWERED = 1,  /**< graph must be lowered */
	CF_CHANGED      = 2,  /**< control flow was changed */
};

/**
 * The lower environment.
 */
typedef struct lower_dw_env_t {
	lower64_entry_t **entries;     /**< entries per node */
	ir_graph      *irg;
	struct obstack obst;           /**< an obstack holding the temporary data */
	ir_type   *l_mtp;              /**< lowered method type of the current method */
	ir_tarval *tv_mode_bytes;      /**< a tarval containing the number of bytes in the lowered modes */
	ir_tarval *tv_mode_bits;       /**< a tarval containing the number of bits in the lowered modes */
	pdeq      *waitq;              /**< a wait queue of all nodes that must be handled later */
	ir_node  **lowered_phis;       /**< list of lowered phis */
	pmap      *proj_2_block;       /**< a map from ProjX to its destination blocks */
	ir_mode   *high_signed;        /**< doubleword signed type */
	ir_mode   *high_unsigned;      /**< doubleword unsigned type */
	ir_mode   *low_signed;         /**< word signed type */
	ir_mode   *low_unsigned;       /**< word unsigned type */
	ident     *first_id;           /**< .l for little and .h for big endian */
	ident     *next_id;            /**< .h for little and .l for big endian */
	const lwrdw_param_t *params;   /**< transformation parameter */
	unsigned flags;                /**< some flags */
	unsigned n_entries;            /**< number of entries */
	ir_type  *value_param_tp;      /**< the old value param type */
} lower_dw_env_t;

static lower_dw_env_t *env;

static void lower_node(ir_node *node);
static bool mtp_must_be_lowered(ir_type *mtp);

/**
 * Create a method type for a Conv emulation from imode to omode.
 */
static ir_type *get_conv_type(ir_mode *imode, ir_mode *omode)
{
	conv_tp_entry_t key, *entry;
	ir_type *mtd;

	key.imode = imode;
	key.omode = omode;
	key.mtd   = NULL;

	entry = (conv_tp_entry_t*)set_insert(conv_types, &key, sizeof(key), HASH_PTR(imode) ^ HASH_PTR(omode));
	if (! entry->mtd) {
		int n_param = 1, n_res = 1;

		if (imode == env->high_signed || imode == env->high_unsigned)
			n_param = 2;
		if (omode == env->high_signed || omode == env->high_unsigned)
			n_res = 2;

		/* create a new one */
		mtd = new_type_method(n_param, n_res);

		/* set param types and result types */
		n_param = 0;
		if (imode == env->high_signed) {
			set_method_param_type(mtd, n_param++, tp_u);
			set_method_param_type(mtd, n_param++, tp_s);
		} else if (imode == env->high_unsigned) {
			set_method_param_type(mtd, n_param++, tp_u);
			set_method_param_type(mtd, n_param++, tp_u);
		} else {
			ir_type *tp = get_type_for_mode(imode);
			set_method_param_type(mtd, n_param++, tp);
		}

		n_res = 0;
		if (omode == env->high_signed) {
			set_method_res_type(mtd, n_res++, tp_u);
			set_method_res_type(mtd, n_res++, tp_s);
		} else if (omode == env->high_unsigned) {
			set_method_res_type(mtd, n_res++, tp_u);
			set_method_res_type(mtd, n_res++, tp_u);
		} else {
			ir_type *tp = get_type_for_mode(omode);
			set_method_res_type(mtd, n_res++, tp);
		}
		entry->mtd = mtd;
	} else {
		mtd = entry->mtd;
	}
	return mtd;
}

/**
 * Add an additional control flow input to a block.
 * Patch all Phi nodes. The new Phi inputs are copied from
 * old input number nr.
 */
static void add_block_cf_input_nr(ir_node *block, int nr, ir_node *cf)
{
	int i, arity = get_irn_arity(block);
	ir_node **in, *phi;

	assert(nr < arity);

	NEW_ARR_A(ir_node *, in, arity + 1);
	for (i = 0; i < arity; ++i)
		in[i] = get_irn_n(block, i);
	in[i] = cf;

	set_irn_in(block, i + 1, in);

	for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
		for (i = 0; i < arity; ++i)
			in[i] = get_irn_n(phi, i);
		in[i] = in[nr];
		set_irn_in(phi, i + 1, in);
	}
}

/**
 * Add an additional control flow input to a block.
 * Patch all Phi nodes. The new Phi inputs are copied from
 * old input from cf tmpl.
 */
static void add_block_cf_input(ir_node *block, ir_node *tmpl, ir_node *cf)
{
	int i, arity = get_irn_arity(block);
	int nr = 0;

	for (i = 0; i < arity; ++i) {
		if (get_irn_n(block, i) == tmpl) {
			nr = i;
			break;
		}
	}
	assert(i < arity);
	add_block_cf_input_nr(block, nr, cf);
}

/**
 * Return the "operational" mode of a Firm node.
 */
static ir_mode *get_irn_op_mode(ir_node *node)
{
	switch (get_irn_opcode(node)) {
	case iro_Load:
		return get_Load_mode(node);
	case iro_Store:
		return get_irn_mode(get_Store_value(node));
	case iro_Div:
		return get_irn_mode(get_Div_left(node));
	case iro_Mod:
		return get_irn_mode(get_Mod_left(node));
	case iro_Cmp:
		return get_irn_mode(get_Cmp_left(node));
	default:
		return get_irn_mode(node);
	}
}

/**
 * Walker, prepare the node links and determine which nodes need to be lowered
 * at all.
 */
static void prepare_links(ir_node *node)
{
	ir_mode         *mode = get_irn_op_mode(node);
	lower64_entry_t *link;
	int              i;

	if (mode == env->high_signed || mode == env->high_unsigned) {
		unsigned idx = get_irn_idx(node);
		/* ok, found a node that will be lowered */
		link = OALLOCZ(&env->obst, lower64_entry_t);

		if (idx >= env->n_entries) {
			/* enlarge: this happens only for Rotl nodes which is RARELY */
			unsigned old   = env->n_entries;
			unsigned n_idx = idx + (idx >> 3);

			ARR_RESIZE(lower64_entry_t *, env->entries, n_idx);
			memset(&env->entries[old], 0, (n_idx - old) * sizeof(env->entries[0]));
			env->n_entries = n_idx;
		}
		env->entries[idx] = link;
		env->flags |= MUST_BE_LOWERED;
	} else if (is_Conv(node)) {
		/* Conv nodes have two modes */
		ir_node *pred = get_Conv_op(node);
		mode = get_irn_mode(pred);

		if (mode == env->high_signed || mode == env->high_unsigned) {
			/* must lower this node either but don't need a link */
			env->flags |= MUST_BE_LOWERED;
		}
		return;
	}

	if (is_Proj(node)) {
		/* link all Proj nodes to its predecessor:
		   Note that Tuple Proj's and its Projs are linked either. */
		ir_node *pred = get_Proj_pred(node);

		set_irn_link(node, get_irn_link(pred));
		set_irn_link(pred, node);
	} else if (is_Phi(node)) {
		/* link all Phi nodes to its block */
		ir_node *block = get_nodes_block(node);
		add_Block_phi(block, node);
	} else if (is_Block(node)) {
		/* fill the Proj -> Block map */
		for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
			ir_node *pred = get_Block_cfgpred(node, i);

			if (is_Proj(pred))
				pmap_insert(env->proj_2_block, pred, node);
		}
	}
}

lower64_entry_t *get_node_entry(ir_node *node)
{
	unsigned idx = get_irn_idx(node);
	assert(idx < env->n_entries);
	return env->entries[idx];
}

void ir_set_dw_lowered(ir_node *old, ir_node *new_low, ir_node *new_high)
{
	lower64_entry_t *entry = get_node_entry(old);
	entry->low_word  = new_low;
	entry->high_word = new_high;
}

/**
 * Translate a Constant: create two.
 */
static void lower_Const(ir_node *node, ir_mode *mode)
{
	ir_graph  *irg      = get_irn_irg(node);
	dbg_info  *dbg      = get_irn_dbg_info(node);
	ir_mode   *low_mode = env->low_unsigned;
	ir_tarval *tv       = get_Const_tarval(node);
	ir_tarval *tv_l     = tarval_convert_to(tv, low_mode);
	ir_node   *res_low  = new_rd_Const(dbg, irg, tv_l);
	ir_tarval *tv_shrs  = tarval_shrs(tv, env->tv_mode_bits);
	ir_tarval *tv_h     = tarval_convert_to(tv_shrs, mode);
	ir_node   *res_high = new_rd_Const(dbg, irg, tv_h);

	ir_set_dw_lowered(node, res_low, res_high);
}

/**
 * Translate a Load: create two.
 */
static void lower_Load(ir_node *node, ir_mode *mode)
{
	ir_mode    *low_mode = env->low_unsigned;
	ir_graph   *irg = get_irn_irg(node);
	ir_node    *adr = get_Load_ptr(node);
	ir_node    *mem = get_Load_mem(node);
	ir_node    *low, *high, *proj;
	dbg_info   *dbg;
	ir_node    *block = get_nodes_block(node);
	ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
	                         ? cons_volatile : cons_none;

	if (env->params->little_endian) {
		low  = adr;
		high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
	} else {
		low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
		high = adr;
	}

	/* create two loads */
	dbg  = get_irn_dbg_info(node);
	low  = new_rd_Load(dbg, block, mem,  low,  low_mode, volatility);
	proj = new_r_Proj(low, mode_M, pn_Load_M);
	high = new_rd_Load(dbg, block, proj, high, mode, volatility);

	ir_set_dw_lowered(node, low, high);

	for (proj = (ir_node*)get_irn_link(node); proj;
	     proj = (ir_node*)get_irn_link(proj)) {
		switch (get_Proj_proj(proj)) {
		case pn_Load_M:         /* Memory result. */
			/* put it to the second one */
			set_Proj_pred(proj, high);
			break;
		case pn_Load_X_except:  /* Execution result if exception occurred. */
			/* put it to the first one */
			set_Proj_pred(proj, low);
			break;
		case pn_Load_res: {       /* Result of load operation. */
			ir_node *res_low  = new_r_Proj(low,  low_mode, pn_Load_res);
			ir_node *res_high = new_r_Proj(high, mode,     pn_Load_res);
			ir_set_dw_lowered(proj, res_low, res_high);
			break;
		}
		default:
			assert(0 && "unexpected Proj number");
		}
		/* mark this proj: we have handled it already, otherwise we might fall
		 * into out new nodes. */
		mark_irn_visited(proj);
	}
}

/**
 * Translate a Store: create two.
 */
static void lower_Store(ir_node *node, ir_mode *mode)
{
	ir_graph              *irg;
	ir_node               *block, *adr, *mem;
	ir_node               *low, *high, *proj;
	dbg_info              *dbg;
	ir_node               *value = get_Store_value(node);
	const lower64_entry_t *entry = get_node_entry(value);
	ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
	                           ? cons_volatile : cons_none;
	(void) mode;

	assert(entry);

	if (! entry->low_word) {
		/* not ready yet, wait */
		pdeq_putr(env->waitq, node);
		return;
	}

	irg = get_irn_irg(node);
	adr = get_Store_ptr(node);
	mem = get_Store_mem(node);
	block = get_nodes_block(node);

	if (env->params->little_endian) {
		low  = adr;
		high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
	} else {
		low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
		high = adr;
	}

	/* create two Stores */
	dbg = get_irn_dbg_info(node);
	low  = new_rd_Store(dbg, block, mem, low,  entry->low_word, volatility);
	proj = new_r_Proj(low, mode_M, pn_Store_M);
	high = new_rd_Store(dbg, block, proj, high, entry->high_word, volatility);

	ir_set_dw_lowered(node, low, high);

	for (proj = (ir_node*)get_irn_link(node); proj;
	     proj = (ir_node*)get_irn_link(proj)) {
		switch (get_Proj_proj(proj)) {
		case pn_Store_M:         /* Memory result. */
			/* put it to the second one */
			set_Proj_pred(proj, high);
			break;
		case pn_Store_X_except:  /* Execution result if exception occurred. */
			/* put it to the first one */
			set_Proj_pred(proj, low);
			break;
		default:
			assert(0 && "unexpected Proj number");
		}
		/* mark this proj: we have handled it already, otherwise we might fall into
		 * out new nodes. */
		mark_irn_visited(proj);
	}
}

/**
 * Return a node containing the address of the intrinsic emulation function.
 *
 * @param method  the method type of the emulation function
 * @param op      the emulated ir_op
 * @param imode   the input mode of the emulated opcode
 * @param omode   the output mode of the emulated opcode
 * @param env     the lower environment
 */
static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
                                      ir_mode *imode, ir_mode *omode)
{
	symconst_symbol sym;
	ir_entity *ent;
	op_mode_entry_t key, *entry;

	key.op    = op;
	key.imode = imode;
	key.omode = omode;
	key.ent   = NULL;

	entry = (op_mode_entry_t*)set_insert(intrinsic_fkt, &key, sizeof(key),
				HASH_PTR(op) ^ HASH_PTR(imode) ^ (HASH_PTR(omode) << 8));
	if (! entry->ent) {
		/* create a new one */
		ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);

		assert(ent && "Intrinsic creator must return an entity");
		entry->ent = ent;
	} else {
		ent = entry->ent;
	}
	sym.entity_p = ent;
	return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
}

/**
 * Translate a Div.
 *
 * Create an intrinsic Call.
 */
static void lower_Div(ir_node *node, ir_mode *mode)
{
	ir_node         *left   = get_Div_left(node);
	ir_node         *right  = get_Div_right(node);
	ir_node         *block  = get_nodes_block(node);
	dbg_info        *dbgi   = get_irn_dbg_info(node);
	ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
	ir_mode         *opmode = get_irn_op_mode(node);
	ir_node         *addr
	    = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
	ir_node         *in[4];
	ir_node         *call;
	ir_node         *resproj;
	const ir_edge_t *edge;
	const ir_edge_t *next;

	if (env->params->little_endian) {
		in[0] = get_lowered_low(left);
		in[1] = get_lowered_high(left);
		in[2] = get_lowered_low(right);
		in[3] = get_lowered_high(right);
	} else {
		in[0] = get_lowered_high(left);
		in[1] = get_lowered_low(left);
		in[2] = get_lowered_high(right);
		in[3] = get_lowered_low(right);
	}
	call    = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
	set_irn_pinned(call, get_irn_pinned(node));

	for (proj = (ir_node*)get_irn_link(node); proj;
	     proj = (ir_node*)get_irn_link(proj)) {
		switch (get_Proj_proj(proj)) {
		case pn_Div_M:         /* Memory result. */
			/* reroute to the call */
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_M);
			break;
		case pn_Div_X_except:  /* Execution result if exception occurred. */
			/* reroute to the call */
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_X_except);
			break;
		case pn_Div_res:
			if (env->params->little_endian) {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
				ir_node *res_high = new_r_Proj(resproj, mode,              1);
				ir_set_dw_lowered(proj, res_low, res_high);
			} else {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
				ir_node *res_high = new_r_Proj(resproj, mode,              0);
				ir_set_dw_lowered(proj, res_low, res_high);
			}
			break;
		default:
			assert(0 && "unexpected Proj number");
		}
		/* mark this proj: we have handled it already, otherwise we might fall into
		 * out new nodes. */
		mark_irn_visited(proj);
	}
}

/**
 * Translate a Mod.
 *
 * Create an intrinsic Call.
 */
static void lower_Mod(ir_node *node, ir_mode *mode)
{
	ir_node         *left   = get_Mod_left(node);
	ir_node         *right  = get_Mod_right(node);
	dbg_info        *dbgi   = get_irn_dbg_info(node);
	ir_node         *block  = get_nodes_block(node);
	ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
	ir_mode         *opmode = get_irn_op_mode(node);
	ir_node         *addr
	    = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
	ir_node         *in[4];
	ir_node         *call;
	ir_node         *resproj;
	const ir_edge_t *edge;
	const ir_edge_t *next;

	if (env->params->little_endian) {
		in[0] = get_lowered_low(left);
		in[1] = get_lowered_high(left);
		in[2] = get_lowered_low(right);
		in[3] = get_lowered_high(right);
	} else {
		in[0] = get_lowered_high(left);
		in[1] = get_lowered_low(left);
		in[2] = get_lowered_high(right);
		in[3] = get_lowered_low(right);
	}
	call    = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
	set_irn_pinned(call, get_irn_pinned(node));

	for (proj = (ir_node*)get_irn_link(node); proj;
	     proj = (ir_node*)get_irn_link(proj)) {
		switch (get_Proj_proj(proj)) {
		case pn_Mod_M:         /* Memory result. */
			/* reroute to the call */
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_M);
			break;
		case pn_Mod_X_except:  /* Execution result if exception occurred. */
			/* reroute to the call */
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_X_except);
			break;
		case pn_Mod_res:
			if (env->params->little_endian) {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
				ir_node *res_high = new_r_Proj(resproj, mode,              1);
				ir_set_dw_lowered(proj, res_low, res_high);
			} else {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
				ir_node *res_high = new_r_Proj(resproj, mode,              0);
				ir_set_dw_lowered(proj, res_low, res_high);
			}
			break;
		default:
			assert(0 && "unexpected Proj number");
		}
		/* mark this proj: we have handled it already, otherwise we might fall
		 * into out new nodes. */
		mark_irn_visited(proj);
	}
}

/**
 * Translate a binop.
 *
 * Create an intrinsic Call.
 */
static void lower_binop(ir_node *node, ir_mode *mode)
{
	ir_node  *left  = get_binop_left(node);
	ir_node  *right = get_binop_right(node);
	dbg_info *dbgi  = get_irn_dbg_info(node);
	ir_node  *block = get_nodes_block(node);
	ir_graph *irg   = get_irn_irg(block);
	ir_type  *mtp   = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
	ir_node  *addr  = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
	ir_node  *in[4];
	ir_node  *call;
	ir_node  *resproj;

	if (env->params->little_endian) {
		in[0] = get_lowered_low(left);
		in[1] = get_lowered_high(left);
		in[2] = get_lowered_low(right);
		in[3] = get_lowered_high(right);
	} else {
		in[0] = get_lowered_high(left);
		in[1] = get_lowered_low(left);
		in[2] = get_lowered_high(right);
		in[3] = get_lowered_low(right);
	}
	call    = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
	set_irn_pinned(call, get_irn_pinned(node));

	if (env->params->little_endian) {
		ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
		ir_node *res_high = new_r_Proj(resproj, mode,              1);
		ir_set_dw_lowered(node, res_low, res_high);
	} else {
		ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
		ir_node *res_high = new_r_Proj(resproj, mode,              0);
		ir_set_dw_lowered(node, res_low, res_high);
	}
}

/**
 * Translate a Shiftop.
 *
 * Create an intrinsic Call.
 */
static void lower_Shiftop(ir_node *node, ir_mode *mode)
{
	ir_node               *block      = get_nodes_block(node);
	ir_node               *left       = get_binop_left(node);
	const lower64_entry_t *left_entry = get_node_entry(left);
	ir_node               *right      = get_binop_right(node);
	ir_node               *in[3]      = {
		left_entry->low_word, left_entry->high_word,
		/* it should be safe to conv to low_unsigned */
		new_r_Conv(block, right, env->low_unsigned)
	};
	dbg_info           *dbgi       = get_irn_dbg_info(node);
	ir_graph           *irg        = get_irn_irg(block);
	ir_type            *mtp
		= mode_is_signed(mode) ? shiftop_tp_s : shiftop_tp_u;
	ir_node            *addr
		= get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
	ir_node            *call
		= new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 3, in, mtp);
	ir_node            *resproj  = new_r_Proj(call, mode_T, pn_Call_T_result);
	ir_node            *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
	ir_node            *res_high = new_r_Proj(resproj, mode,              1);

	set_irn_pinned(call, get_irn_pinned(node));
	ir_set_dw_lowered(node, res_low, res_high);
}

/**
 * Translate a Shr and handle special cases.
 */
static void lower_Shr(ir_node *node, ir_mode *mode)
{
	ir_graph *irg   = get_irn_irg(node);
	ir_node  *right = get_Shr_right(node);

	if (get_mode_arithmetic(mode) == irma_twos_complement && is_Const(right)) {
		ir_tarval *tv = get_Const_tarval(right);

		if (tarval_is_long(tv) &&
		    get_tarval_long(tv) >= (long)get_mode_size_bits(mode)) {
			ir_node *block        = get_nodes_block(node);
			ir_node *left         = get_Shr_left(node);
			ir_mode *low_unsigned = env->low_unsigned;
			long shf_cnt = get_tarval_long(tv) - get_mode_size_bits(mode);
			const lower64_entry_t *left_entry = get_node_entry(left);
			ir_node *res_low;
			ir_node *res_high;

			left = left_entry->high_word;

			/* convert high word into low_unsigned mode if necessary */
			if (get_irn_mode(left) != low_unsigned)
				left = new_r_Conv(block, left, low_unsigned);

			if (shf_cnt > 0) {
				ir_node *c = new_r_Const_long(irg, low_unsigned, shf_cnt);
				res_low = new_r_Shr(block, left, c, low_unsigned);
			} else {
				res_low = left;
			}
			res_high = new_r_Const(irg, get_mode_null(mode));
			ir_set_dw_lowered(node, res_low, res_high);

			return;
		}
	}
	lower_Shiftop(node, mode);
}

/**
 * Translate a Shl and handle special cases.
 */
static void lower_Shl(ir_node *node, ir_mode *mode)
{
	ir_graph *irg   = get_irn_irg(node);
	ir_node  *right = get_Shl_right(node);

	if (get_mode_arithmetic(mode) == irma_twos_complement && is_Const(right)) {
		ir_tarval *tv = get_Const_tarval(right);

		if (tarval_is_long(tv)) {
			long value = get_tarval_long(tv);
		    if (value >= (long)get_mode_size_bits(mode)) {
				/* simple case: shift above the lower word */
				ir_mode *mode_l;
				ir_node *block = get_nodes_block(node);
				ir_node *left = get_Shl_left(node);
				ir_node *c;
				long shf_cnt = get_tarval_long(tv) - get_mode_size_bits(mode);
				const lower64_entry_t *left_entry = get_node_entry(left);
				ir_node  *res_low;
				ir_node  *res_high;

				left = left_entry->low_word;
				left = new_r_Conv(block, left, mode);

				mode_l = env->low_unsigned;
				if (shf_cnt > 0) {
					c        = new_r_Const_long(irg, mode_l, shf_cnt);
					res_high = new_r_Shl(block, left, c, mode);
				} else {
					res_high = left;
				}
				res_low = new_r_Const(irg, get_mode_null(mode_l));
				ir_set_dw_lowered(node, res_low, res_high);

				return;
			}
			if (value == 1) {
				/* left << 1 == left + left */
				ir_node               *left        = get_binop_left(node);
				const lower64_entry_t *left_entry  = get_node_entry(left);
				ir_node               *in[4]       = {
					left_entry->low_word, left_entry->high_word,
					left_entry->low_word, left_entry->high_word,
				};
				dbg_info           *dbgi        = get_irn_dbg_info(node);
				ir_node            *block       = get_nodes_block(node);
				ir_graph           *irg         = get_irn_irg(block);
				ir_type            *mtp
					= mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
				ir_node            *addr
					= get_intrinsic_address(mtp, op_Add, mode, mode);
				ir_node            *call
					= new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
				ir_node            *resproj  = new_r_Proj(call, mode_T, pn_Call_T_result);
				ir_node            *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
				ir_node            *res_high = new_r_Proj(resproj, mode,              1);
				set_irn_pinned(call, get_irn_pinned(node));
				ir_set_dw_lowered(node, res_low, res_high);

				return;
			}
		}
	}
	lower_Shiftop(node, mode);
}

/**
 * Translate a Shrs and handle special cases.
 */
static void lower_Shrs(ir_node *node, ir_mode *mode)
{
	ir_graph *irg   = get_irn_irg(node);
	ir_node  *right = get_Shrs_right(node);

	if (get_mode_arithmetic(mode) == irma_twos_complement && is_Const(right)) {
		ir_tarval *tv = get_Const_tarval(right);

		if (tarval_is_long(tv) &&
		    get_tarval_long(tv) >= (long)get_mode_size_bits(mode)) {
			ir_node *block         = get_nodes_block(node);
			ir_node *left          = get_Shrs_left(node);
			ir_mode *low_unsigned  = env->low_unsigned;
			long     shf_cnt       = get_tarval_long(tv) - get_mode_size_bits(mode);
			const lower64_entry_t *left_entry = get_node_entry(left);
			ir_node *left_unsigned = left;
			ir_node *res_low;
			ir_node *res_high;
			ir_node *c;

			left = left_entry->high_word;

			/* convert high word into low_unsigned mode if necessary */
			if (get_irn_mode(left_unsigned) != low_unsigned)
				left_unsigned = new_r_Conv(block, left, low_unsigned);

			if (shf_cnt > 0) {
				c       = new_r_Const_long(irg, low_unsigned, shf_cnt);
				res_low = new_r_Shrs(block, left_unsigned, c, low_unsigned);
			} else {
				res_low = left_unsigned;
			}

			c        = new_r_Const(irg, get_mode_all_one(low_unsigned));
			res_high = new_r_Shrs(block, left, c, mode);
			ir_set_dw_lowered(node, res_low, res_high);
			return;
		}
	}
	lower_Shiftop(node, mode);
}

/**
 * Rebuild Rotl nodes into Or(Shl, Shr) and prepare all nodes.
 */
static void prepare_links_and_handle_rotl(ir_node *node, void *data)
{
	(void) data;
	if (is_Rotl(node)) {
		ir_mode  *mode = get_irn_op_mode(node);
		ir_node  *right;
		ir_node  *left, *shl, *shr, *ornode, *block, *sub, *c;
		ir_mode  *omode, *rmode;
		ir_graph *irg;
		dbg_info *dbg;
		optimization_state_t state;

		if (mode != env->high_signed && mode != env->high_unsigned) {
			prepare_links(node);
			return;
		}

		/* replace the Rotl(x,y) by an Or(Shl(x,y), Shr(x,64-y)) */
		right = get_Rotl_right(node);
		irg   = get_irn_irg(node);
		dbg   = get_irn_dbg_info(node);
		omode = get_irn_mode(node);
		left  = get_Rotl_left(node);
		block = get_nodes_block(node);
		shl   = new_rd_Shl(dbg, block, left, right, omode);
		rmode = get_irn_mode(right);
		c     = new_r_Const_long(irg, rmode, get_mode_size_bits(omode));
		sub   = new_rd_Sub(dbg, block, c, right, rmode);
		shr   = new_rd_Shr(dbg, block, left, sub, omode);

		/* switch optimization off here, or we will get the Rotl back */
		save_optimization_state(&state);
		set_opt_algebraic_simplification(0);
		ornode = new_rd_Or(dbg, block, shl, shr, omode);
		restore_optimization_state(&state);

		exchange(node, ornode);

		/* do lowering on the new nodes */
		prepare_links(shl);
		prepare_links(c);
		prepare_links(sub);
		prepare_links(shr);
		prepare_links(ornode);
		return;
	}

	prepare_links(node);
}

/**
 * Translate an Unop.
 *
 * Create an intrinsic Call.
 */
static void lower_unop(ir_node *node, ir_mode *mode)
{
	ir_node  *op       = get_unop_op(node);
	dbg_info *dbgi     = get_irn_dbg_info(node);
	ir_node  *block    = get_nodes_block(node);
	ir_graph *irg      = get_irn_irg(block);
	ir_type  *mtp      = mode_is_signed(mode) ? unop_tp_s : unop_tp_u;
	ir_op    *irop     = get_irn_op(node);
	ir_node  *addr     = get_intrinsic_address(mtp, irop, mode, mode);
	ir_node  *nomem    = get_irg_no_mem(irg);
	ir_node  *in[2];
	ir_node  *call;
	ir_node  *resproj;

	if (env->params->little_endian) {
		in[0] = get_lowered_low(op);
		in[1] = get_lowered_high(op);
	} else {
		in[0] = get_lowered_high(op);
		in[1] = get_lowered_low(op);
	}
	call    = new_rd_Call(dbgi, block, nomem, addr, 2, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
	set_irn_pinned(call, get_irn_pinned(node));

	if (env->params->little_endian) {
		ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
		ir_node *res_high = new_r_Proj(resproj, mode,              1);
		ir_set_dw_lowered(node, res_low, res_high);
	} else {
		ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
		ir_node *res_high = new_r_Proj(resproj, mode,              0);
		ir_set_dw_lowered(node, res_low, res_high);
	}
}

/**
 * Translate a logical binop.
 *
 * Create two logical binops.
 */
static void lower_binop_logical(ir_node *node, ir_mode *mode,
								ir_node *(*constr_rd)(dbg_info *db, ir_node *block, ir_node *op1, ir_node *op2, ir_mode *mode) )
{
	ir_node               *left        = get_binop_left(node);
	ir_node               *right       = get_binop_right(node);
	const lower64_entry_t *left_entry  = get_node_entry(left);
	const lower64_entry_t *right_entry = get_node_entry(right);
	dbg_info              *dbgi        = get_irn_dbg_info(node);
	ir_node               *block       = get_nodes_block(node);
	ir_node               *res_low
		= constr_rd(dbgi, block, left_entry->low_word, right_entry->low_word,
		            env->low_unsigned);
	ir_node               *res_high
		= constr_rd(dbgi, block, left_entry->high_word, right_entry->high_word,
		            mode);
	ir_set_dw_lowered(node, res_low, res_high);
}

static void lower_And(ir_node *node, ir_mode *mode)
{
	lower_binop_logical(node, mode, new_rd_And);
}

static void lower_Or(ir_node *node, ir_mode *mode)
{
	lower_binop_logical(node, mode, new_rd_Or);