2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Lower double word operations, i.e. 64bit -> 32bit, 32bit -> 16bit etc.
24 * @author Michael Beck
37 #include "irnodeset.h"
38 #include "irgraph_t.h"
43 #include "dbginfo_t.h"
44 #include "iropt_dbg.h"
50 #include "iroptimize.h"
61 /** A map from (op, imode, omode) to Intrinsic functions entities. */
62 static set *intrinsic_fkt;
64 /** A map from (imode, omode) to conv function types. */
65 static set *conv_types;
67 /** A map from a method type to its lowered type. */
68 static pmap *lowered_type;
70 /** A map from a builtin type to its lower and higher type. */
71 static pmap *lowered_builtin_type_high;
72 static pmap *lowered_builtin_type_low;
74 /** The types for the binop and unop intrinsics. */
75 static ir_type *binop_tp_u, *binop_tp_s, *unop_tp_u, *unop_tp_s, *tp_s, *tp_u;
77 static ir_nodeset_t created_mux_nodes;
79 /** the debug handle */
80 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
83 * An entry in the (op, imode, omode) -> entity map.
85 typedef struct op_mode_entry {
86 const ir_op *op; /**< the op */
87 const ir_mode *imode; /**< the input mode */
88 const ir_mode *omode; /**< the output mode */
89 ir_entity *ent; /**< the associated entity of this (op, imode, omode) triple */
93 * An entry in the (imode, omode) -> tp map.
95 typedef struct conv_tp_entry {
96 const ir_mode *imode; /**< the input mode */
97 const ir_mode *omode; /**< the output mode */
98 ir_type *mtd; /**< the associated method type of this (imode, omode) pair */
102 MUST_BE_LOWERED = 1, /**< graph must be lowered */
103 CF_CHANGED = 2, /**< control flow was changed */
107 * The lower environment.
109 typedef struct lower_dw_env_t {
110 lower64_entry_t **entries; /**< entries per node */
112 struct obstack obst; /**< an obstack holding the temporary data */
113 ir_tarval *tv_mode_bytes; /**< a tarval containing the number of bytes in the lowered modes */
114 ir_tarval *tv_mode_bits; /**< a tarval containing the number of bits in the lowered modes */
115 pdeq *waitq; /**< a wait queue of all nodes that must be handled later */
116 ir_node **lowered_phis; /**< list of lowered phis */
117 ir_mode *high_signed; /**< doubleword signed type */
118 ir_mode *high_unsigned; /**< doubleword unsigned type */
119 ir_mode *low_signed; /**< word signed type */
120 ir_mode *low_unsigned; /**< word unsigned type */
121 ident *first_id; /**< .l for little and .h for big endian */
122 ident *next_id; /**< .h for little and .l for big endian */
123 const lwrdw_param_t *params; /**< transformation parameter */
124 unsigned flags; /**< some flags */
125 unsigned n_entries; /**< number of entries */
128 static lower_dw_env_t *env;
130 static void lower_node(ir_node *node);
133 * Create a method type for a Conv emulation from imode to omode.
135 static ir_type *get_conv_type(ir_mode *imode, ir_mode *omode)
137 conv_tp_entry_t key, *entry;
144 entry = (conv_tp_entry_t*)set_insert(conv_types, &key, sizeof(key), hash_ptr(imode) ^ hash_ptr(omode));
146 int n_param = 1, n_res = 1;
148 if (imode == env->high_signed || imode == env->high_unsigned)
150 if (omode == env->high_signed || omode == env->high_unsigned)
153 /* create a new one */
154 mtd = new_type_method(n_param, n_res);
156 /* set param types and result types */
158 if (imode == env->high_signed) {
159 if (env->params->little_endian) {
160 set_method_param_type(mtd, n_param++, tp_u);
161 set_method_param_type(mtd, n_param++, tp_s);
163 set_method_param_type(mtd, n_param++, tp_s);
164 set_method_param_type(mtd, n_param++, tp_u);
166 } else if (imode == env->high_unsigned) {
167 set_method_param_type(mtd, n_param++, tp_u);
168 set_method_param_type(mtd, n_param++, tp_u);
170 ir_type *tp = get_type_for_mode(imode);
171 set_method_param_type(mtd, n_param++, tp);
175 if (omode == env->high_signed) {
176 if (env->params->little_endian) {
177 set_method_res_type(mtd, n_res++, tp_u);
178 set_method_res_type(mtd, n_res++, tp_s);
180 set_method_res_type(mtd, n_res++, tp_s);
181 set_method_res_type(mtd, n_res++, tp_u);
183 } else if (omode == env->high_unsigned) {
184 set_method_res_type(mtd, n_res++, tp_u);
185 set_method_res_type(mtd, n_res++, tp_u);
187 ir_type *tp = get_type_for_mode(omode);
188 set_method_res_type(mtd, n_res++, tp);
198 * Add an additional control flow input to a block.
199 * Patch all Phi nodes. The new Phi inputs are copied from
200 * old input number nr.
202 static void add_block_cf_input_nr(ir_node *block, int nr, ir_node *cf)
204 int i, arity = get_irn_arity(block);
206 const ir_edge_t *edge;
210 NEW_ARR_A(ir_node *, in, arity + 1);
211 for (i = 0; i < arity; ++i)
212 in[i] = get_irn_n(block, i);
215 set_irn_in(block, i + 1, in);
217 foreach_out_edge(block, edge) {
218 ir_node *phi = get_edge_src_irn(edge);
222 for (i = 0; i < arity; ++i)
223 in[i] = get_irn_n(phi, i);
225 set_irn_in(phi, i + 1, in);
230 * Add an additional control flow input to a block.
231 * Patch all Phi nodes. The new Phi inputs are copied from
232 * old input from cf tmpl.
234 static void add_block_cf_input(ir_node *block, ir_node *tmpl, ir_node *cf)
236 int i, arity = get_irn_arity(block);
239 for (i = 0; i < arity; ++i) {
240 if (get_irn_n(block, i) == tmpl) {
246 add_block_cf_input_nr(block, nr, cf);
250 * Return the "operational" mode of a Firm node.
252 static ir_mode *get_irn_op_mode(ir_node *node)
254 switch (get_irn_opcode(node)) {
256 return get_Load_mode(node);
258 return get_irn_mode(get_Store_value(node));
260 return get_irn_mode(get_Div_left(node));
262 return get_irn_mode(get_Mod_left(node));
264 return get_irn_mode(get_Cmp_left(node));
266 return get_irn_mode(node);
271 * Walker, prepare the node links and determine which nodes need to be lowered
274 static void prepare_links(ir_node *node)
276 ir_mode *mode = get_irn_op_mode(node);
277 lower64_entry_t *link;
279 if (mode == env->high_signed || mode == env->high_unsigned) {
280 unsigned idx = get_irn_idx(node);
281 /* ok, found a node that will be lowered */
282 link = OALLOCZ(&env->obst, lower64_entry_t);
284 if (idx >= env->n_entries) {
285 /* enlarge: this happens only for Rotl nodes which is RARELY */
286 unsigned old = env->n_entries;
287 unsigned n_idx = idx + (idx >> 3);
289 ARR_RESIZE(lower64_entry_t *, env->entries, n_idx);
290 memset(&env->entries[old], 0, (n_idx - old) * sizeof(env->entries[0]));
291 env->n_entries = n_idx;
293 env->entries[idx] = link;
294 env->flags |= MUST_BE_LOWERED;
295 } else if (is_Conv(node)) {
296 /* Conv nodes have two modes */
297 ir_node *pred = get_Conv_op(node);
298 mode = get_irn_mode(pred);
300 if (mode == env->high_signed || mode == env->high_unsigned) {
301 /* must lower this node either but don't need a link */
302 env->flags |= MUST_BE_LOWERED;
305 } else if (is_Call(node)) {
306 /* Special case: If the result of the Call is never used, we won't
307 * find a Proj with a mode that potentially triggers MUST_BE_LOWERED
308 * to be set. Thus, if we see a call, we check its result types and
309 * decide whether MUST_BE_LOWERED has to be set.
311 ir_type *tp = get_Call_type(node);
314 n_res = get_method_n_ress(tp);
315 for (i = 0; i < n_res; ++i) {
316 ir_type *rtp = get_method_res_type(tp, i);
318 if (is_Primitive_type(rtp)) {
319 ir_mode *rmode = get_type_mode(rtp);
321 if (rmode == env->high_signed || rmode == env->high_unsigned) {
322 env->flags |= MUST_BE_LOWERED;
329 lower64_entry_t *get_node_entry(ir_node *node)
331 unsigned idx = get_irn_idx(node);
332 assert(idx < env->n_entries);
333 return env->entries[idx];
336 void ir_set_dw_lowered(ir_node *old, ir_node *new_low, ir_node *new_high)
338 lower64_entry_t *entry = get_node_entry(old);
339 entry->low_word = new_low;
340 entry->high_word = new_high;
343 ir_mode *ir_get_low_unsigned_mode(void)
345 return env->low_unsigned;
349 * Translate a Constant: create two.
351 static void lower_Const(ir_node *node, ir_mode *mode)
353 ir_graph *irg = get_irn_irg(node);
354 dbg_info *dbg = get_irn_dbg_info(node);
355 ir_mode *low_mode = env->low_unsigned;
356 ir_tarval *tv = get_Const_tarval(node);
357 ir_tarval *tv_l = tarval_convert_to(tv, low_mode);
358 ir_node *res_low = new_rd_Const(dbg, irg, tv_l);
359 ir_tarval *tv_shrs = tarval_shrs(tv, env->tv_mode_bits);
360 ir_tarval *tv_h = tarval_convert_to(tv_shrs, mode);
361 ir_node *res_high = new_rd_Const(dbg, irg, tv_h);
363 ir_set_dw_lowered(node, res_low, res_high);
367 * Translate a Load: create two.
369 static void lower_Load(ir_node *node, ir_mode *mode)
371 ir_mode *low_mode = env->low_unsigned;
372 ir_graph *irg = get_irn_irg(node);
373 ir_node *adr = get_Load_ptr(node);
374 ir_node *mem = get_Load_mem(node);
379 ir_node *block = get_nodes_block(node);
380 ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
381 ? cons_volatile : cons_none;
382 const ir_edge_t *edge;
383 const ir_edge_t *next;
385 if (env->params->little_endian) {
387 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
389 low = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
393 /* create two loads */
394 dbg = get_irn_dbg_info(node);
395 low = new_rd_Load(dbg, block, mem, low, low_mode, volatility);
396 proj_m = new_r_Proj(low, mode_M, pn_Load_M);
397 high = new_rd_Load(dbg, block, proj_m, high, mode, volatility);
399 foreach_out_edge_safe(node, edge, next) {
400 ir_node *proj = get_edge_src_irn(edge);
404 switch (get_Proj_proj(proj)) {
405 case pn_Load_M: /* Memory result. */
406 /* put it to the second one */
407 set_Proj_pred(proj, high);
409 case pn_Load_X_except: /* Execution result if exception occurred. */
410 /* put it to the first one */
411 set_Proj_pred(proj, low);
413 case pn_Load_res: { /* Result of load operation. */
414 ir_node *res_low = new_r_Proj(low, low_mode, pn_Load_res);
415 ir_node *res_high = new_r_Proj(high, mode, pn_Load_res);
416 ir_set_dw_lowered(proj, res_low, res_high);
420 assert(0 && "unexpected Proj number");
422 /* mark this proj: we have handled it already, otherwise we might fall
423 * into out new nodes. */
424 mark_irn_visited(proj);
429 * Translate a Store: create two.
431 static void lower_Store(ir_node *node, ir_mode *mode)
434 ir_node *block, *adr, *mem;
435 ir_node *low, *high, *proj_m;
437 ir_node *value = get_Store_value(node);
438 const lower64_entry_t *entry = get_node_entry(value);
439 ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
440 ? cons_volatile : cons_none;
441 const ir_edge_t *edge;
442 const ir_edge_t *next;
447 if (! entry->low_word) {
448 /* not ready yet, wait */
449 pdeq_putr(env->waitq, node);
453 irg = get_irn_irg(node);
454 adr = get_Store_ptr(node);
455 mem = get_Store_mem(node);
456 block = get_nodes_block(node);
458 if (env->params->little_endian) {
460 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
462 low = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
466 /* create two Stores */
467 dbg = get_irn_dbg_info(node);
468 low = new_rd_Store(dbg, block, mem, low, entry->low_word, volatility);
469 proj_m = new_r_Proj(low, mode_M, pn_Store_M);
470 high = new_rd_Store(dbg, block, proj_m, high, entry->high_word, volatility);
472 foreach_out_edge_safe(node, edge, next) {
473 ir_node *proj = get_edge_src_irn(edge);
477 switch (get_Proj_proj(proj)) {
478 case pn_Store_M: /* Memory result. */
479 /* put it to the second one */
480 set_Proj_pred(proj, high);
482 case pn_Store_X_except: /* Execution result if exception occurred. */
483 /* put it to the first one */
484 set_Proj_pred(proj, low);
487 assert(0 && "unexpected Proj number");
489 /* mark this proj: we have handled it already, otherwise we might fall into
491 mark_irn_visited(proj);
496 * Return a node containing the address of the intrinsic emulation function.
498 * @param method the method type of the emulation function
499 * @param op the emulated ir_op
500 * @param imode the input mode of the emulated opcode
501 * @param omode the output mode of the emulated opcode
502 * @param env the lower environment
504 static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
505 ir_mode *imode, ir_mode *omode)
509 op_mode_entry_t key, *entry;
516 entry = (op_mode_entry_t*)set_insert(intrinsic_fkt, &key, sizeof(key),
517 hash_ptr(op) ^ hash_ptr(imode) ^ (hash_ptr(omode) << 8));
519 /* create a new one */
520 ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);
522 assert(ent && "Intrinsic creator must return an entity");
528 return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
534 * Create an intrinsic Call.
536 static void lower_Div(ir_node *node, ir_mode *mode)
538 ir_node *left = get_Div_left(node);
539 ir_node *right = get_Div_right(node);
540 ir_node *block = get_nodes_block(node);
541 dbg_info *dbgi = get_irn_dbg_info(node);
542 ir_type *mtp = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
543 ir_mode *opmode = get_irn_op_mode(node);
545 = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
549 const ir_edge_t *edge;
550 const ir_edge_t *next;
552 if (env->params->little_endian) {
553 in[0] = get_lowered_low(left);
554 in[1] = get_lowered_high(left);
555 in[2] = get_lowered_low(right);
556 in[3] = get_lowered_high(right);
558 in[0] = get_lowered_high(left);
559 in[1] = get_lowered_low(left);
560 in[2] = get_lowered_high(right);
561 in[3] = get_lowered_low(right);
563 call = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
564 resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
565 set_irn_pinned(call, get_irn_pinned(node));
567 foreach_out_edge_safe(node, edge, next) {
568 ir_node *proj = get_edge_src_irn(edge);
572 switch (get_Proj_proj(proj)) {
573 case pn_Div_M: /* Memory result. */
574 /* reroute to the call */
575 set_Proj_pred(proj, call);
576 set_Proj_proj(proj, pn_Call_M);
578 case pn_Div_X_regular:
579 set_Proj_pred(proj, call);
580 set_Proj_proj(proj, pn_Call_X_regular);
582 case pn_Div_X_except:
583 set_Proj_pred(proj, call);
584 set_Proj_proj(proj, pn_Call_X_except);
587 if (env->params->little_endian) {
588 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 0);
589 ir_node *res_high = new_r_Proj(resproj, mode, 1);
590 ir_set_dw_lowered(proj, res_low, res_high);
592 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 1);
593 ir_node *res_high = new_r_Proj(resproj, mode, 0);
594 ir_set_dw_lowered(proj, res_low, res_high);
598 assert(0 && "unexpected Proj number");
600 /* mark this proj: we have handled it already, otherwise we might fall into
602 mark_irn_visited(proj);
609 * Create an intrinsic Call.
611 static void lower_Mod(ir_node *node, ir_mode *mode)
613 ir_node *left = get_Mod_left(node);
614 ir_node *right = get_Mod_right(node);
615 dbg_info *dbgi = get_irn_dbg_info(node);
616 ir_node *block = get_nodes_block(node);
617 ir_type *mtp = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
618 ir_mode *opmode = get_irn_op_mode(node);
620 = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
624 const ir_edge_t *edge;
625 const ir_edge_t *next;
627 if (env->params->little_endian) {
628 in[0] = get_lowered_low(left);
629 in[1] = get_lowered_high(left);
630 in[2] = get_lowered_low(right);
631 in[3] = get_lowered_high(right);
633 in[0] = get_lowered_high(left);
634 in[1] = get_lowered_low(left);
635 in[2] = get_lowered_high(right);
636 in[3] = get_lowered_low(right);
638 call = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
639 resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
640 set_irn_pinned(call, get_irn_pinned(node));
642 foreach_out_edge_safe(node, edge, next) {
643 ir_node *proj = get_edge_src_irn(edge);
647 switch (get_Proj_proj(proj)) {
648 case pn_Mod_M: /* Memory result. */
649 /* reroute to the call */
650 set_Proj_pred(proj, call);
651 set_Proj_proj(proj, pn_Call_M);
653 case pn_Div_X_regular:
654 set_Proj_pred(proj, call);
655 set_Proj_proj(proj, pn_Call_X_regular);
657 case pn_Mod_X_except:
658 set_Proj_pred(proj, call);
659 set_Proj_proj(proj, pn_Call_X_except);
662 if (env->params->little_endian) {
663 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 0);
664 ir_node *res_high = new_r_Proj(resproj, mode, 1);
665 ir_set_dw_lowered(proj, res_low, res_high);
667 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 1);
668 ir_node *res_high = new_r_Proj(resproj, mode, 0);
669 ir_set_dw_lowered(proj, res_low, res_high);
673 assert(0 && "unexpected Proj number");
675 /* mark this proj: we have handled it already, otherwise we might fall
676 * into out new nodes. */
677 mark_irn_visited(proj);
684 * Create an intrinsic Call.
686 static void lower_binop(ir_node *node, ir_mode *mode)
688 ir_node *left = get_binop_left(node);
689 ir_node *right = get_binop_right(node);
690 dbg_info *dbgi = get_irn_dbg_info(node);
691 ir_node *block = get_nodes_block(node);
692 ir_graph *irg = get_irn_irg(block);
693 ir_type *mtp = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
694 ir_node *addr = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
699 if (env->params->little_endian) {
700 in[0] = get_lowered_low(left);
701 in[1] = get_lowered_high(left);
702 in[2] = get_lowered_low(right);
703 in[3] = get_lowered_high(right);
705 in[0] = get_lowered_high(left);
706 in[1] = get_lowered_low(left);
707 in[2] = get_lowered_high(right);
708 in[3] = get_lowered_low(right);
710 call = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
711 resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
712 set_irn_pinned(call, get_irn_pinned(node));
714 if (env->params->little_endian) {
715 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 0);
716 ir_node *res_high = new_r_Proj(resproj, mode, 1);
717 ir_set_dw_lowered(node, res_low, res_high);
719 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 1);
720 ir_node *res_high = new_r_Proj(resproj, mode, 0);
721 ir_set_dw_lowered(node, res_low, res_high);
725 static ir_node *create_conv(ir_node *block, ir_node *node, ir_mode *dest_mode)
727 if (get_irn_mode(node) == dest_mode)
729 return new_r_Conv(block, node, dest_mode);
733 * Moves node and all predecessors of node from from_bl to to_bl.
734 * Does not move predecessors of Phi nodes (or block nodes).
736 static void move(ir_node *node, ir_node *from_bl, ir_node *to_bl)
741 set_nodes_block(node, to_bl);
744 if (get_irn_mode(node) == mode_T) {
745 const ir_edge_t *edge;
746 foreach_out_edge(node, edge) {
747 ir_node *proj = get_edge_src_irn(edge);
750 move(proj, from_bl, to_bl);
754 /* We must not move predecessors of Phi nodes, even if they are in
755 * from_bl. (because these are values from an earlier loop iteration
756 * which are not predecessors of node here)
762 arity = get_irn_arity(node);
763 for (i = 0; i < arity; i++) {
764 ir_node *pred = get_irn_n(node, i);
765 ir_mode *pred_mode = get_irn_mode(pred);
766 if (get_nodes_block(pred) == from_bl)
767 move(pred, from_bl, to_bl);
768 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
769 ir_node *pred_low = get_lowered_low(pred);
770 ir_node *pred_high = get_lowered_high(pred);
771 if (get_nodes_block(pred_low) == from_bl)
772 move(pred_low, from_bl, to_bl);
773 if (pred_high != NULL && get_nodes_block(pred_high) == from_bl)
774 move(pred_high, from_bl, to_bl);
780 * We need a custom version of part_block_edges because during transformation
781 * not all data-dependencies are explicit yet if a lowered nodes users are not
783 * We can fix this by modifying move to look for such implicit dependencies.
784 * Additionally we have to keep the proj_2_block map updated
786 static ir_node *part_block_dw(ir_node *node)
788 ir_graph *irg = get_irn_irg(node);
789 ir_node *old_block = get_nodes_block(node);
790 int n_cfgpreds = get_Block_n_cfgpreds(old_block);
791 ir_node **cfgpreds = get_Block_cfgpred_arr(old_block);
792 ir_node *new_block = new_r_Block(irg, n_cfgpreds, cfgpreds);
793 const ir_edge_t *edge;
794 const ir_edge_t *next;
796 /* old_block has no predecessors anymore for now */
797 set_irn_in(old_block, 0, NULL);
799 /* move node and its predecessors to new_block */
800 move(node, old_block, new_block);
802 /* move Phi nodes to new_block */
803 foreach_out_edge_safe(old_block, edge, next) {
804 ir_node *phi = get_edge_src_irn(edge);
807 set_nodes_block(phi, new_block);
812 typedef ir_node* (*new_rd_shr_func)(dbg_info *dbgi, ir_node *block,
813 ir_node *left, ir_node *right,
816 static void lower_shr_helper(ir_node *node, ir_mode *mode,
817 new_rd_shr_func new_rd_shrs)
819 ir_node *right = get_binop_right(node);
820 ir_node *left = get_binop_left(node);
821 ir_mode *shr_mode = get_irn_mode(node);
822 unsigned modulo_shift = get_mode_modulo_shift(shr_mode);
823 ir_mode *low_unsigned = env->low_unsigned;
824 unsigned modulo_shift2 = get_mode_modulo_shift(mode);
825 ir_graph *irg = get_irn_irg(node);
826 ir_node *left_low = get_lowered_low(left);
827 ir_node *left_high = get_lowered_high(left);
828 dbg_info *dbgi = get_irn_dbg_info(node);
829 ir_node *lower_block;
839 ir_node *lower_in[2];
840 ir_node *phi_low_in[2];
841 ir_node *phi_high_in[2];
843 /* this version is optimized for modulo shift architectures
844 * (and can't handle anything else) */
845 if (modulo_shift != get_mode_size_bits(shr_mode)
846 || modulo_shift2<<1 != modulo_shift) {
847 panic("Shr lowering only implemented for modulo shift shr operations");
849 if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
850 panic("Shr lowering only implemented for power-of-2 modes");
852 /* without 2-complement the -x instead of (bit_width-x) trick won't work */
853 if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
854 panic("Shr lowering only implemented for two-complement modes");
857 block = get_nodes_block(node);
859 /* if the right operand is a 64bit value, we're only interested in the
861 if (get_irn_mode(right) == env->high_unsigned) {
862 right = get_lowered_low(right);
864 /* shift should never have signed mode on the right */
865 assert(get_irn_mode(right) != env->high_signed);
866 right = create_conv(block, right, low_unsigned);
869 lower_block = part_block_dw(node);
870 env->flags |= CF_CHANGED;
871 block = get_nodes_block(node);
873 /* add a Cmp to test if highest bit is set <=> whether we shift more
874 * than half the word width */
875 cnst = new_r_Const_long(irg, low_unsigned, modulo_shift2);
876 andn = new_r_And(block, right, cnst, low_unsigned);
877 cnst = new_r_Const(irg, get_mode_null(low_unsigned));
878 cmp = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
879 cond = new_rd_Cond(dbgi, block, cmp);
880 proj_true = new_r_Proj(cond, mode_X, pn_Cond_true);
881 proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
883 /* the true block => shift_width < 1word */
885 /* In theory the low value (for 64bit shifts) is:
886 * Or(High << (32-x)), Low >> x)
887 * In practice High << 32-x will fail when x is zero (since we have
888 * modulo shift and 32 will be 0). So instead we use:
889 * Or(High<<1<<~x, Low >> x)
891 ir_node *in[1] = { proj_true };
892 ir_node *block_true = new_r_Block(irg, ARRAY_SIZE(in), in);
893 ir_node *res_high = new_rd_shrs(dbgi, block_true, left_high,
895 ir_node *shift_low = new_rd_Shr(dbgi, block_true, left_low, right,
897 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
899 ir_node *conv = create_conv(block_true, left_high,
901 ir_node *one = new_r_Const(irg, get_mode_one(low_unsigned));
902 ir_node *carry0 = new_rd_Shl(dbgi, block_true, conv, one,
904 ir_node *carry1 = new_rd_Shl(dbgi, block_true, carry0,
905 not_shiftval, low_unsigned);
906 ir_node *res_low = new_rd_Or(dbgi, block_true, shift_low, carry1,
908 lower_in[0] = new_r_Jmp(block_true);
909 phi_low_in[0] = res_low;
910 phi_high_in[0] = res_high;
913 /* false block => shift_width > 1word */
915 ir_node *in[1] = { proj_false };
916 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
917 ir_node *conv = create_conv(block_false, left_high, low_unsigned);
918 ir_node *res_low = new_rd_shrs(dbgi, block_false, conv, right,
920 int cnsti = modulo_shift2-1;
921 ir_node *cnst2 = new_r_Const_long(irg, low_unsigned, cnsti);
923 if (new_rd_shrs == new_rd_Shrs) {
924 res_high = new_rd_shrs(dbgi, block_false, left_high, cnst2, mode);
926 res_high = new_r_Const(irg, get_mode_null(mode));
928 lower_in[1] = new_r_Jmp(block_false);
929 phi_low_in[1] = res_low;
930 phi_high_in[1] = res_high;
933 /* patch lower block */
934 set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
935 phi_low = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
937 phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
939 ir_set_dw_lowered(node, phi_low, phi_high);
942 static void lower_Shr(ir_node *node, ir_mode *mode)
944 lower_shr_helper(node, mode, new_rd_Shr);
947 static void lower_Shrs(ir_node *node, ir_mode *mode)
949 lower_shr_helper(node, mode, new_rd_Shrs);
952 static void lower_Shl(ir_node *node, ir_mode *mode)
954 ir_node *right = get_binop_right(node);
955 ir_node *left = get_binop_left(node);
956 ir_mode *shr_mode = get_irn_mode(node);
957 unsigned modulo_shift = get_mode_modulo_shift(shr_mode);
958 ir_mode *low_unsigned = env->low_unsigned;
959 unsigned modulo_shift2 = get_mode_modulo_shift(mode);
960 ir_graph *irg = get_irn_irg(node);
961 ir_node *left_low = get_lowered_low(left);
962 ir_node *left_high = get_lowered_high(left);
963 dbg_info *dbgi = get_irn_dbg_info(node);
964 ir_node *lower_block = get_nodes_block(node);
974 ir_node *lower_in[2];
975 ir_node *phi_low_in[2];
976 ir_node *phi_high_in[2];
978 /* this version is optimized for modulo shift architectures
979 * (and can't handle anything else) */
980 if (modulo_shift != get_mode_size_bits(shr_mode)
981 || modulo_shift2<<1 != modulo_shift) {
982 panic("Shl lowering only implemented for modulo shift shl operations");
984 if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
985 panic("Shl lowering only implemented for power-of-2 modes");
987 /* without 2-complement the -x instead of (bit_width-x) trick won't work */
988 if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
989 panic("Shl lowering only implemented for two-complement modes");
992 /* if the right operand is a 64bit value, we're only interested in the
994 if (get_irn_mode(right) == env->high_unsigned) {
995 right = get_lowered_low(right);
997 /* shift should never have signed mode on the right */
998 assert(get_irn_mode(right) != env->high_signed);
999 right = create_conv(lower_block, right, low_unsigned);
1002 part_block_dw(node);
1003 env->flags |= CF_CHANGED;
1004 block = get_nodes_block(node);
1006 /* add a Cmp to test if highest bit is set <=> whether we shift more
1007 * than half the word width */
1008 cnst = new_r_Const_long(irg, low_unsigned, modulo_shift2);
1009 andn = new_r_And(block, right, cnst, low_unsigned);
1010 cnst = new_r_Const(irg, get_mode_null(low_unsigned));
1011 cmp = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
1012 cond = new_rd_Cond(dbgi, block, cmp);
1013 proj_true = new_r_Proj(cond, mode_X, pn_Cond_true);
1014 proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
1016 /* the true block => shift_width < 1word */
1018 ir_node *in[1] = { proj_true };
1019 ir_node *block_true = new_r_Block(irg, ARRAY_SIZE(in), in);
1021 ir_node *res_low = new_rd_Shl(dbgi, block_true, left_low,
1022 right, low_unsigned);
1023 ir_node *shift_high = new_rd_Shl(dbgi, block_true, left_high, right,
1025 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
1027 ir_node *conv = create_conv(block_true, left_low, mode);
1028 ir_node *one = new_r_Const(irg, get_mode_one(low_unsigned));
1029 ir_node *carry0 = new_rd_Shr(dbgi, block_true, conv, one, mode);
1030 ir_node *carry1 = new_rd_Shr(dbgi, block_true, carry0,
1031 not_shiftval, mode);
1032 ir_node *res_high = new_rd_Or(dbgi, block_true, shift_high, carry1,
1034 lower_in[0] = new_r_Jmp(block_true);
1035 phi_low_in[0] = res_low;
1036 phi_high_in[0] = res_high;
1039 /* false block => shift_width > 1word */
1041 ir_node *in[1] = { proj_false };
1042 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
1043 ir_node *res_low = new_r_Const(irg, get_mode_null(low_unsigned));
1044 ir_node *conv = create_conv(block_false, left_low, mode);
1045 ir_node *res_high = new_rd_Shl(dbgi, block_false, conv, right, mode);
1046 lower_in[1] = new_r_Jmp(block_false);
1047 phi_low_in[1] = res_low;
1048 phi_high_in[1] = res_high;
1051 /* patch lower block */
1052 set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
1053 phi_low = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
1055 phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
1057 ir_set_dw_lowered(node, phi_low, phi_high);
1061 * Rebuild Rotl nodes into Or(Shl, Shr) and prepare all nodes.
1063 static void prepare_links_and_handle_rotl(ir_node *node, void *data)
1066 if (is_Rotl(node)) {
1067 ir_mode *mode = get_irn_op_mode(node);
1069 ir_node *left, *shl, *shr, *ornode, *block, *sub, *c;
1070 ir_mode *omode, *rmode;
1073 optimization_state_t state;
1075 if (mode != env->high_signed && mode != env->high_unsigned) {
1076 prepare_links(node);
1080 /* replace the Rotl(x,y) by an Or(Shl(x,y), Shr(x,64-y)) */
1081 right = get_Rotl_right(node);
1082 irg = get_irn_irg(node);
1083 dbg = get_irn_dbg_info(node);
1084 omode = get_irn_mode(node);
1085 left = get_Rotl_left(node);
1086 block = get_nodes_block(node);
1087 shl = new_rd_Shl(dbg, block, left, right, omode);
1088 rmode = get_irn_mode(right);
1089 c = new_r_Const_long(irg, rmode, get_mode_size_bits(omode));
1090 sub = new_rd_Sub(dbg, block, c, right, rmode);
1091 shr = new_rd_Shr(dbg, block, left, sub, omode);
1093 /* switch optimization off here, or we will get the Rotl back */
1094 save_optimization_state(&state);
1095 set_opt_algebraic_simplification(0);
1096 ornode = new_rd_Or(dbg, block, shl, shr, omode);
1097 restore_optimization_state(&state);
1099 exchange(node, ornode);
1101 /* do lowering on the new nodes */
1106 prepare_links(ornode);
1110 prepare_links(node);
1114 * Translate an Unop.
1116 * Create an intrinsic Call.
1118 static void lower_unop(ir_node *node, ir_mode *mode)
1120 ir_node *op = get_unop_op(node);
1121 dbg_info *dbgi = get_irn_dbg_info(node);
1122 ir_node *block = get_nodes_block(node);
1123 ir_graph *irg = get_irn_irg(block);
1124 ir_type *mtp = mode_is_signed(mode) ? unop_tp_s : unop_tp_u;
1125 ir_op *irop = get_irn_op(node);
1126 ir_node *addr = get_intrinsic_address(mtp, irop, mode, mode);
1127 ir_node *nomem = get_irg_no_mem(irg);
1132 if (env->params->little_endian) {
1133 in[0] = get_lowered_low(op);
1134 in[1] = get_lowered_high(op);
1136 in[0] = get_lowered_high(op);
1137 in[1] = get_lowered_low(op);
1139 call = new_rd_Call(dbgi, block, nomem, addr, 2, in, mtp);
1140 resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
1141 set_irn_pinned(call, get_irn_pinned(node));
1143 if (env->params->little_endian) {
1144 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 0);
1145 ir_node *res_high = new_r_Proj(resproj, mode, 1);
1146 ir_set_dw_lowered(node, res_low, res_high);
1148 ir_node *res_low = new_r_Proj(resproj, env->low_unsigned, 1);
1149 ir_node *res_high = new_r_Proj(resproj, mode, 0);
1150 ir_set_dw_lowered(node, res_low, res_high);
1155 * Translate a logical binop.
1157 * Create two logical binops.
1159 static void lower_binop_logical(ir_node *node, ir_mode *mode,
1160 ir_node *(*constr_rd)(dbg_info *db, ir_node *block, ir_node *op1, ir_node *op2, ir_mode *mode) )
1162 ir_node *left = get_binop_left(node);
1163 ir_node *right = get_binop_right(node);
1164 const lower64_entry_t *left_entry = get_node_entry(left);
1165 const lower64_entry_t *right_entry = get_node_entry(right);
1166 dbg_info *dbgi = get_irn_dbg_info(node);
1167 ir_node *block = get_nodes_block(node);
1169 = constr_rd(dbgi, block, left_entry->low_word, right_entry->low_word,
1172 = constr_rd(dbgi, block, left_entry->high_word, right_entry->high_word,
1174 ir_set_dw_lowered(node, res_low, res_high);
1177 static void lower_And(ir_node *node, ir_mode *mode)
1179 lower_binop_logical(node, mode, new_rd_And);
1182 static void lower_Or(ir_node *node, ir_mode *mode)
1184 lower_binop_logical(node, mode, new_rd_Or);
1187 static void lower_Eor(ir_node *node, ir_mode *mode)
1189 lower_binop_logical(node, mode, new_rd_Eor);
1195 * Create two logical Nots.
1197 static void lower_Not(ir_node *node, ir_mode *mode)
1199 ir_node *op = get_Not_op(node);
1200 const lower64_entry_t *op_entry = get_node_entry(op);
1201 dbg_info *dbgi = get_irn_dbg_info(node);
1202 ir_node *block = get_nodes_block(node);
1204 = new_rd_Not(dbgi, block, op_entry->low_word, env->low_unsigned);
1206 = new_rd_Not(dbgi, block, op_entry->high_word, mode);
1207 ir_set_dw_lowered(node, res_low, res_high);
1210 static void lower_Proj(ir_node *node, ir_mode *op_mode)
1212 ir_mode *mode = get_irn_mode(node);
1215 if (mode != env->high_signed && mode != env->high_unsigned)
1218 pred = get_Proj_pred(node);
1219 if (is_Tuple(pred)) {
1220 long pn = get_Proj_proj(node);
1221 ir_node *op = get_irn_n(pred, pn);
1222 const lower64_entry_t *entry = get_node_entry(op);
1223 ir_set_dw_lowered(node, entry->low_word, entry->high_word);
1227 static bool is_equality_cmp(const ir_node *node)
1229 ir_relation relation = get_Cmp_relation(node);
1230 ir_node *left = get_Cmp_left(node);
1231 ir_node *right = get_Cmp_right(node);
1232 ir_mode *mode = get_irn_mode(left);
1234 /* this probably makes no sense if unordered is involved */
1235 assert(!mode_is_float(mode));
1237 if (relation == ir_relation_equal || relation == ir_relation_less_greater)
1240 if (!is_Const(right) || !is_Const_null(right))
1242 if (mode_is_signed(mode)) {
1243 return relation == ir_relation_less_greater;
1245 return relation == ir_relation_greater;
1249 static ir_node *get_cfop_destination(const ir_node *cfop)
1251 const ir_edge_t *first = get_irn_out_edge_first(cfop);
1252 /* we should only have 1 destination */
1253 assert(get_irn_n_edges(cfop) == 1);
1254 return get_edge_src_irn(first);
1257 static void lower_Switch(ir_node *node, ir_mode *high_mode)
1259 ir_node *selector = get_Switch_selector(node);
1260 ir_mode *mode = get_irn_mode(selector);
1262 if (mode == env->high_signed || mode == env->high_unsigned) {
1263 /* we can't really handle Switch with 64bit offsets */
1264 panic("Switch with 64bit jumptable not supported");
1266 lower_node(selector);
1272 static void lower_Cond(ir_node *node, ir_mode *high_mode)
1274 ir_node *left, *right, *block;
1275 ir_node *sel = get_Cond_selector(node);
1277 const lower64_entry_t *lentry, *rentry;
1278 ir_node *projT = NULL, *projF = NULL;
1279 ir_node *new_bl, *irn;
1280 ir_node *projHF, *projHT;
1282 ir_relation relation;
1285 const ir_edge_t *edge;
1286 const ir_edge_t *next;
1295 left = get_Cmp_left(sel);
1296 cmp_mode = get_irn_mode(left);
1297 if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned) {
1302 right = get_Cmp_right(sel);
1305 lentry = get_node_entry(left);
1306 rentry = get_node_entry(right);
1308 /* all right, build the code */
1309 foreach_out_edge_safe(node, edge, next) {
1310 ir_node *proj = get_edge_src_irn(edge);
1314 proj_nr = get_Proj_proj(proj);
1316 if (proj_nr == pn_Cond_true) {
1317 assert(projT == NULL && "more than one Proj(true)");
1320 assert(proj_nr == pn_Cond_false);
1321 assert(projF == NULL && "more than one Proj(false)");
1324 mark_irn_visited(proj);
1326 assert(projT && projF);
1328 /* create a new high compare */
1329 block = get_nodes_block(node);
1330 irg = get_Block_irg(block);
1331 dbg = get_irn_dbg_info(sel);
1332 relation = get_Cmp_relation(sel);
1334 if (is_equality_cmp(sel)) {
1335 /* x ==/!= y ==> or(x_low^y_low,x_high^y_high) ==/!= 0 */
1336 ir_mode *mode = env->low_unsigned;
1337 ir_node *low_left = new_rd_Conv(dbg, block, lentry->low_word, mode);
1338 ir_node *high_left = new_rd_Conv(dbg, block, lentry->high_word, mode);
1339 ir_node *low_right = new_rd_Conv(dbg, block, rentry->low_word, mode);
1340 ir_node *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1341 ir_node *xor_low = new_rd_Eor(dbg, block, low_left, low_right, mode);
1342 ir_node *xor_high = new_rd_Eor(dbg, block, high_left, high_right, mode);
1343 ir_node *ornode = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1344 ir_node *cmp = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1345 set_Cond_selector(node, cmp);
1349 if (relation == ir_relation_equal) {
1351 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1352 dst_blk = get_cfop_destination(projF);
1354 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1356 dbg = get_irn_dbg_info(node);
1357 irn = new_rd_Cond(dbg, block, irn);
1359 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1360 mark_irn_visited(projHF);
1361 exchange(projF, projHF);
1363 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1364 mark_irn_visited(projHT);
1366 new_bl = new_r_Block(irg, 1, &projHT);
1368 dbg = get_irn_dbg_info(sel);
1369 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1371 dbg = get_irn_dbg_info(node);
1372 irn = new_rd_Cond(dbg, new_bl, irn);
1374 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1375 mark_irn_visited(proj);
1376 add_block_cf_input(dst_blk, projHF, proj);
1378 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1379 mark_irn_visited(proj);
1380 exchange(projT, proj);
1381 } else if (relation == ir_relation_less_greater) {
1383 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1384 dst_blk = get_cfop_destination(projT);
1386 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1387 ir_relation_less_greater);
1388 dbg = get_irn_dbg_info(node);
1389 irn = new_rd_Cond(dbg, block, irn);
1391 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1392 mark_irn_visited(projHT);
1393 exchange(projT, projHT);
1395 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1396 mark_irn_visited(projHF);
1398 new_bl = new_r_Block(irg, 1, &projHF);
1400 dbg = get_irn_dbg_info(sel);
1401 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1402 ir_relation_less_greater);
1403 dbg = get_irn_dbg_info(node);
1404 irn = new_rd_Cond(dbg, new_bl, irn);
1406 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1407 mark_irn_visited(proj);
1408 add_block_cf_input(dst_blk, projHT, proj);
1410 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1411 mark_irn_visited(proj);
1412 exchange(projF, proj);
1415 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1416 ir_node *dstT, *dstF, *newbl_eq, *newbl_l;
1419 dstT = get_cfop_destination(projT);
1420 dstF = get_cfop_destination(projF);
1422 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1423 relation & ~ir_relation_equal);
1424 dbg = get_irn_dbg_info(node);
1425 irn = new_rd_Cond(dbg, block, irn);
1427 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1428 mark_irn_visited(projHT);
1430 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1431 mark_irn_visited(projHF);
1433 newbl_eq = new_r_Block(irg, 1, &projHF);
1435 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1437 irn = new_rd_Cond(dbg, newbl_eq, irn);
1439 projEqF = new_r_Proj(irn, mode_X, pn_Cond_false);
1440 mark_irn_visited(projEqF);
1442 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1443 mark_irn_visited(proj);
1445 newbl_l = new_r_Block(irg, 1, &proj);
1447 dbg = get_irn_dbg_info(sel);
1448 irn = new_rd_Cmp(dbg, newbl_l, lentry->low_word, rentry->low_word,
1450 dbg = get_irn_dbg_info(node);
1451 irn = new_rd_Cond(dbg, newbl_l, irn);
1453 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1454 mark_irn_visited(proj);
1455 add_block_cf_input(dstT, projT, proj);
1457 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1458 mark_irn_visited(proj);
1459 add_block_cf_input(dstF, projF, proj);
1461 exchange(projT, projHT);
1462 exchange(projF, projEqF);
1465 /* we have changed the control flow */
1466 env->flags |= CF_CHANGED;
1470 * Translate a Conv to higher_signed
1472 static void lower_Conv_to_Ll(ir_node *node)
1474 ir_mode *omode = get_irn_mode(node);
1475 ir_node *op = get_Conv_op(node);
1476 ir_mode *imode = get_irn_mode(op);
1477 ir_graph *irg = get_irn_irg(node);
1478 ir_node *block = get_nodes_block(node);
1479 dbg_info *dbg = get_irn_dbg_info(node);
1483 ir_mode *low_unsigned = env->low_unsigned;
1485 = mode_is_signed(omode) ? env->low_signed : low_unsigned;
1487 if (mode_is_int(imode) || mode_is_reference(imode)) {
1488 if (imode == env->high_signed || imode == env->high_unsigned) {
1489 /* a Conv from Lu to Ls or Ls to Lu */
1490 const lower64_entry_t *op_entry = get_node_entry(op);
1491 res_low = op_entry->low_word;
1492 res_high = new_rd_Conv(dbg, block, op_entry->high_word, low_signed);
1494 /* simple case: create a high word */
1495 if (imode != low_unsigned)
1496 op = new_rd_Conv(dbg, block, op, low_unsigned);
1500 if (mode_is_signed(imode)) {
1501 int c = get_mode_size_bits(low_signed) - 1;
1502 ir_node *cnst = new_r_Const_long(irg, low_unsigned, c);
1503 if (get_irn_mode(op) != low_signed)
1504 op = new_rd_Conv(dbg, block, op, low_signed);
1505 res_high = new_rd_Shrs(dbg, block, op, cnst, low_signed);
1507 res_high = new_r_Const(irg, get_mode_null(low_signed));
1510 } else if (imode == mode_b) {
1511 res_low = new_rd_Conv(dbg, block, op, low_unsigned);
1512 res_high = new_r_Const(irg, get_mode_null(low_signed));
1514 ir_node *irn, *call;
1515 ir_type *mtp = get_conv_type(imode, omode);
1517 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1518 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 1, &op, mtp);
1519 set_irn_pinned(call, get_irn_pinned(node));
1520 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1522 if (env->params->little_endian) {
1523 res_low = new_r_Proj(irn, low_unsigned, 0);
1524 res_high = new_r_Proj(irn, low_signed, 1);
1526 res_low = new_r_Proj(irn, low_unsigned, 1);
1527 res_high = new_r_Proj(irn, low_signed, 0);
1530 ir_set_dw_lowered(node, res_low, res_high);
1534 * Translate a Conv from higher_unsigned
1536 static void lower_Conv_from_Ll(ir_node *node)
1538 ir_node *op = get_Conv_op(node);
1539 ir_mode *omode = get_irn_mode(node);
1540 ir_node *block = get_nodes_block(node);
1541 dbg_info *dbg = get_irn_dbg_info(node);
1542 ir_graph *irg = get_irn_irg(node);
1543 const lower64_entry_t *entry = get_node_entry(op);
1545 if (mode_is_int(omode) || mode_is_reference(omode)) {
1546 op = entry->low_word;
1548 /* simple case: create a high word */
1549 if (omode != env->low_unsigned)
1550 op = new_rd_Conv(dbg, block, op, omode);
1552 set_Conv_op(node, op);
1553 } else if (omode == mode_b) {
1554 /* llu ? true : false <=> (low|high) ? true : false */
1555 ir_mode *mode = env->low_unsigned;
1556 ir_node *ornode = new_rd_Or(dbg, block, entry->low_word,
1557 entry->high_word, mode);
1558 set_Conv_op(node, ornode);
1560 ir_node *irn, *call, *in[2];
1561 ir_mode *imode = get_irn_mode(op);
1562 ir_type *mtp = get_conv_type(imode, omode);
1565 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1566 if (env->params->little_endian) {
1567 in[0] = entry->low_word;
1568 in[1] = entry->high_word;
1570 in[0] = entry->high_word;
1571 in[1] = entry->low_word;
1574 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 2, in, mtp);
1575 set_irn_pinned(call, get_irn_pinned(node));
1576 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1577 res = new_r_Proj(irn, omode, 0);
1579 exchange(node, res);
1586 static void lower_Cmp(ir_node *cmp, ir_mode *m)
1588 ir_node *l = get_Cmp_left(cmp);
1589 ir_mode *cmp_mode = get_irn_mode(l);
1590 ir_node *r, *low, *high, *t, *res;
1591 ir_relation relation;
1594 const lower64_entry_t *lentry;
1595 const lower64_entry_t *rentry;
1598 if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned)
1601 r = get_Cmp_right(cmp);
1602 lentry = get_node_entry(l);
1603 rentry = get_node_entry(r);
1604 relation = get_Cmp_relation(cmp);
1605 block = get_nodes_block(cmp);
1606 dbg = get_irn_dbg_info(cmp);
1608 /* easy case for x ==/!= 0 (see lower_Cond for details) */
1609 if (is_equality_cmp(cmp)) {
1610 ir_graph *irg = get_irn_irg(cmp);
1611 ir_mode *mode = env->low_unsigned;
1612 ir_node *low_left = new_rd_Conv(dbg, block, lentry->low_word, mode);
1613 ir_node *high_left = new_rd_Conv(dbg, block, lentry->high_word, mode);
1614 ir_node *low_right = new_rd_Conv(dbg, block, rentry->low_word, mode);
1615 ir_node *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1616 ir_node *xor_low = new_rd_Eor(dbg, block, low_left, low_right, mode);
1617 ir_node *xor_high = new_rd_Eor(dbg, block, high_left, high_right, mode);
1618 ir_node *ornode = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1619 ir_node *new_cmp = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1620 exchange(cmp, new_cmp);
1624 if (relation == ir_relation_equal) {
1625 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1626 low = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1628 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1630 res = new_rd_And(dbg, block, low, high, mode_b);
1631 } else if (relation == ir_relation_less_greater) {
1632 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1633 low = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1635 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1637 res = new_rd_Or(dbg, block, low, high, mode_b);
1639 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1640 ir_node *high1 = new_rd_Cmp(dbg, block, lentry->high_word,
1641 rentry->high_word, relation & ~ir_relation_equal);
1642 low = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1644 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1646 t = new_rd_And(dbg, block, low, high, mode_b);
1647 res = new_rd_Or(dbg, block, high1, t, mode_b);
1655 static void lower_Conv(ir_node *node, ir_mode *mode)
1657 mode = get_irn_mode(node);
1659 if (mode == env->high_signed || mode == env->high_unsigned) {
1660 lower_Conv_to_Ll(node);
1662 ir_mode *op_mode = get_irn_mode(get_Conv_op(node));
1664 if (op_mode == env->high_signed || op_mode == env->high_unsigned) {
1665 lower_Conv_from_Ll(node);
1670 static void fix_parameter_entities(ir_graph *irg, ir_type *orig_mtp)
1672 size_t orig_n_params = get_method_n_params(orig_mtp);
1673 ir_entity **parameter_entities;
1675 parameter_entities = ALLOCANZ(ir_entity*, orig_n_params);
1677 ir_type *frame_type = get_irg_frame_type(irg);
1678 size_t n = get_compound_n_members(frame_type);
1682 /* collect parameter entities */
1683 for (i = 0; i < n; ++i) {
1684 ir_entity *entity = get_compound_member(frame_type, i);
1686 if (!is_parameter_entity(entity))
1688 p = get_entity_parameter_number(entity);
1689 if (p == IR_VA_START_PARAMETER_NUMBER)
1691 assert(p < orig_n_params);
1692 assert(parameter_entities[p] == NULL);
1693 parameter_entities[p] = entity;
1696 /* adjust indices */
1698 for (i = 0; i < orig_n_params; ++i, ++n_param) {
1699 ir_entity *entity = parameter_entities[i];
1703 set_entity_parameter_number(entity, n_param);
1705 tp = get_method_param_type(orig_mtp, i);
1706 if (is_Primitive_type(tp)) {
1707 ir_mode *mode = get_type_mode(tp);
1708 if (mode == env->high_signed || mode == env->high_unsigned) {
1710 /* note that we do not change the type of the parameter
1711 * entities, as calling convention fixup later still needs to
1712 * know which is/was a lowered doubleword.
1713 * So we just mark/remember it for later */
1714 if (entity != NULL) {
1715 assert(entity->attr.parameter.doubleword_low_mode == NULL);
1716 entity->attr.parameter.doubleword_low_mode
1717 = env->low_unsigned;
1725 * Lower the method type.
1727 * @param env the lower environment
1728 * @param mtp the method type to lower
1730 * @return the lowered type
1732 static ir_type *lower_mtp(ir_type *mtp)
1736 size_t orig_n_params;
1740 bool must_be_lowered;
1742 res = (ir_type*)pmap_get(lowered_type, mtp);
1745 if (type_visited(mtp))
1747 mark_type_visited(mtp);
1749 orig_n_params = get_method_n_params(mtp);
1750 orig_n_res = get_method_n_ress(mtp);
1751 n_param = orig_n_params;
1753 must_be_lowered = false;
1755 /* count new number of params */
1756 for (i = orig_n_params; i > 0;) {
1757 ir_type *tp = get_method_param_type(mtp, --i);
1759 if (is_Primitive_type(tp)) {
1760 ir_mode *mode = get_type_mode(tp);
1762 if (mode == env->high_signed || mode == env->high_unsigned) {
1764 must_be_lowered = true;
1769 /* count new number of results */
1770 for (i = orig_n_res; i > 0;) {
1771 ir_type *tp = get_method_res_type(mtp, --i);
1773 if (is_Primitive_type(tp)) {
1774 ir_mode *mode = get_type_mode(tp);
1776 if (mode == env->high_signed || mode == env->high_unsigned) {
1778 must_be_lowered = true;
1782 if (!must_be_lowered) {
1783 set_type_link(mtp, NULL);
1787 res = new_d_type_method(n_param, n_res, get_type_dbg_info(mtp));
1789 /* set param types and result types */
1790 for (i = n_param = 0; i < orig_n_params; ++i) {
1791 ir_type *tp = get_method_param_type(mtp, i);
1793 if (is_Primitive_type(tp)) {
1794 ir_mode *mode = get_type_mode(tp);
1796 if (mode == env->high_signed) {
1797 if (env->params->little_endian) {
1798 set_method_param_type(res, n_param++, tp_u);
1799 set_method_param_type(res, n_param++, tp_s);
1801 set_method_param_type(res, n_param++, tp_s);
1802 set_method_param_type(res, n_param++, tp_u);
1804 } else if (mode == env->high_unsigned) {
1805 set_method_param_type(res, n_param++, tp_u);
1806 set_method_param_type(res, n_param++, tp_u);
1808 set_method_param_type(res, n_param, tp);
1812 set_method_param_type(res, n_param, tp);
1816 for (i = n_res = 0; i < orig_n_res; ++i) {
1817 ir_type *tp = get_method_res_type(mtp, i);
1819 if (is_Primitive_type(tp)) {
1820 ir_mode *mode = get_type_mode(tp);
1822 if (mode == env->high_signed) {
1823 if (env->params->little_endian) {
1824 set_method_res_type(res, n_res++, tp_u);
1825 set_method_res_type(res, n_res++, tp_s);
1827 set_method_res_type(res, n_res++, tp_s);
1828 set_method_res_type(res, n_res++, tp_u);
1830 } else if (mode == env->high_unsigned) {
1831 set_method_res_type(res, n_res++, tp_u);
1832 set_method_res_type(res, n_res++, tp_u);
1834 set_method_res_type(res, n_res++, tp);
1837 set_method_res_type(res, n_res++, tp);
1841 set_method_variadicity(res, get_method_variadicity(mtp));
1842 set_method_calling_convention(res, get_method_calling_convention(mtp));
1843 set_method_additional_properties(res, get_method_additional_properties(mtp));
1845 set_higher_type(res, mtp);
1846 set_type_link(res, mtp);
1848 mark_type_visited(res);
1849 pmap_insert(lowered_type, mtp, res);
1854 * Translate a Return.
1856 static void lower_Return(ir_node *node, ir_mode *mode)
1858 ir_graph *irg = get_irn_irg(node);
1859 ir_entity *ent = get_irg_entity(irg);
1860 ir_type *mtp = get_entity_type(ent);
1866 /* check if this return must be lowered */
1867 for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1868 ir_node *pred = get_Return_res(node, i);
1869 ir_mode *rmode = get_irn_op_mode(pred);
1871 if (rmode == env->high_signed || rmode == env->high_unsigned)
1877 ent = get_irg_entity(irg);
1878 mtp = get_entity_type(ent);
1880 /* create a new in array */
1881 NEW_ARR_A(ir_node *, in, get_method_n_ress(mtp) + 1);
1883 in[j++] = get_Return_mem(node);
1885 for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1886 ir_node *pred = get_Return_res(node, i);
1887 ir_mode *pred_mode = get_irn_mode(pred);
1889 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1890 const lower64_entry_t *entry = get_node_entry(pred);
1891 if (env->params->little_endian) {
1892 in[j++] = entry->low_word;
1893 in[j++] = entry->high_word;
1895 in[j++] = entry->high_word;
1896 in[j++] = entry->low_word;
1902 assert(j == get_method_n_ress(mtp)+1);
1904 set_irn_in(node, j, in);
1908 * Translate the parameters.
1910 static void lower_Start(ir_node *node, ir_mode *high_mode)
1912 ir_graph *irg = get_irn_irg(node);
1913 ir_entity *ent = get_irg_entity(irg);
1914 ir_type *mtp = get_entity_type(ent);
1915 ir_type *orig_mtp = (ir_type*)get_type_link(mtp);
1918 size_t i, j, n_params;
1919 const ir_edge_t *edge;
1920 const ir_edge_t *next;
1923 /* if type link is NULL then the type was not lowered, hence no changes
1924 * at Start necessary */
1925 if (orig_mtp == NULL)
1928 n_params = get_method_n_params(orig_mtp);
1930 NEW_ARR_A(long, new_projs, n_params);
1932 /* Calculate mapping of proj numbers in new_projs */
1933 for (i = j = 0; i < n_params; ++i, ++j) {
1934 ir_type *ptp = get_method_param_type(orig_mtp, i);
1937 if (is_Primitive_type(ptp)) {
1938 ir_mode *amode = get_type_mode(ptp);
1939 if (amode == env->high_signed || amode == env->high_unsigned)
1944 /* find args Proj */
1946 foreach_out_edge(node, edge) {
1947 ir_node *proj = get_edge_src_irn(edge);
1950 if (get_Proj_proj(proj) == pn_Start_T_args) {
1958 /* fix all Proj's and create new ones */
1959 foreach_out_edge_safe(args, edge, next) {
1960 ir_node *proj = get_edge_src_irn(edge);
1961 ir_mode *mode = get_irn_mode(proj);
1962 ir_mode *mode_l = env->low_unsigned;
1973 pred = get_Proj_pred(proj);
1974 proj_nr = get_Proj_proj(proj);
1976 if (mode == env->high_signed) {
1977 mode_h = env->low_signed;
1978 } else if (mode == env->high_unsigned) {
1979 mode_h = env->low_unsigned;
1981 long new_pn = new_projs[proj_nr];
1982 set_Proj_proj(proj, new_pn);
1986 /* Switch off CSE or we might get an already existing Proj. */
1987 old_cse = get_opt_cse();
1989 dbg = get_irn_dbg_info(proj);
1990 if (env->params->little_endian) {
1991 res_low = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr]);
1992 res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr] + 1);
1994 res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr]);
1995 res_low = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr] + 1);
1997 set_opt_cse(old_cse);
1998 ir_set_dw_lowered(proj, res_low, res_high);
2005 static void lower_Call(ir_node *node, ir_mode *mode)
2007 ir_type *tp = get_Call_type(node);
2009 size_t n_params, n_res;
2010 bool need_lower = false;
2013 long *res_numbers = NULL;
2015 const ir_edge_t *edge;
2016 const ir_edge_t *next;
2019 n_params = get_method_n_params(tp);
2020 for (p = 0; p < n_params; ++p) {
2021 ir_type *ptp = get_method_param_type(tp, p);
2023 if (is_Primitive_type(ptp)) {
2024 ir_mode *pmode = get_type_mode(ptp);
2025 if (pmode == env->high_signed || pmode == env->high_unsigned) {
2031 n_res = get_method_n_ress(tp);
2033 NEW_ARR_A(long, res_numbers, n_res);
2035 for (i = j = 0; i < n_res; ++i, ++j) {
2036 ir_type *ptp = get_method_res_type(tp, i);
2039 if (is_Primitive_type(ptp)) {
2040 ir_mode *rmode = get_type_mode(ptp);
2041 if (rmode == env->high_signed || rmode == env->high_unsigned) {
2052 /* let's lower it */
2054 set_Call_type(node, tp);
2056 NEW_ARR_A(ir_node *, in, get_method_n_params(tp) + 2);
2058 in[0] = get_Call_mem(node);
2059 in[1] = get_Call_ptr(node);
2061 for (j = 2, i = 0; i < n_params; ++i) {
2062 ir_node *pred = get_Call_param(node, i);
2063 ir_mode *pred_mode = get_irn_mode(pred);
2065 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
2066 const lower64_entry_t *pred_entry = get_node_entry(pred);
2067 if (env->params->little_endian) {
2068 in[j++] = pred_entry->low_word;
2069 in[j++] = pred_entry->high_word;
2071 in[j++] = pred_entry->high_word;
2072 in[j++] = pred_entry->low_word;
2079 set_irn_in(node, j, in);
2081 /* find results T */
2083 foreach_out_edge(node, edge) {
2084 ir_node *proj = get_edge_src_irn(edge);
2087 if (get_Proj_proj(proj) == pn_Call_T_result) {
2092 if (resproj == NULL)
2095 /* fix the results */
2096 foreach_out_edge_safe(resproj, edge, next) {
2097 ir_node *proj = get_edge_src_irn(edge);
2098 ir_mode *proj_mode = get_irn_mode(proj);
2099 ir_mode *mode_l = env->low_unsigned;
2109 pred = get_Proj_pred(proj);
2110 proj_nr = get_Proj_proj(proj);
2112 if (proj_mode == env->high_signed) {
2113 mode_h = env->low_signed;
2114 } else if (proj_mode == env->high_unsigned) {
2115 mode_h = env->low_unsigned;
2117 long new_nr = res_numbers[proj_nr];
2118 set_Proj_proj(proj, new_nr);
2122 dbg = get_irn_dbg_info(proj);
2123 if (env->params->little_endian) {
2124 res_low = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr]);
2125 res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr] + 1);
2127 res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr]);
2128 res_low = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr] + 1);
2130 ir_set_dw_lowered(proj, res_low, res_high);
2135 * Translate an Unknown into two.
2137 static void lower_Unknown(ir_node *node, ir_mode *mode)
2139 ir_mode *low_mode = env->low_unsigned;
2140 ir_graph *irg = get_irn_irg(node);
2141 ir_node *res_low = new_r_Unknown(irg, low_mode);
2142 ir_node *res_high = new_r_Unknown(irg, mode);
2143 ir_set_dw_lowered(node, res_low, res_high);
2147 * Translate a Bad into two.
2149 static void lower_Bad(ir_node *node, ir_mode *mode)
2151 ir_mode *low_mode = env->low_unsigned;
2152 ir_graph *irg = get_irn_irg(node);
2153 ir_node *res_low = new_r_Bad(irg, low_mode);
2154 ir_node *res_high = new_r_Bad(irg, mode);
2155 ir_set_dw_lowered(node, res_low, res_high);
2161 * First step: just create two templates
2163 static void lower_Phi(ir_node *phi)
2165 ir_mode *mode = get_irn_mode(phi);
2180 /* enqueue predecessors */
2181 arity = get_Phi_n_preds(phi);
2182 for (i = 0; i < arity; ++i) {
2183 ir_node *pred = get_Phi_pred(phi, i);
2184 pdeq_putr(env->waitq, pred);
2187 if (mode != env->high_signed && mode != env->high_unsigned)
2190 /* first create a new in array */
2191 NEW_ARR_A(ir_node *, in_l, arity);
2192 NEW_ARR_A(ir_node *, in_h, arity);
2193 irg = get_irn_irg(phi);
2194 mode_l = env->low_unsigned;
2195 mode_h = mode == env->high_signed ? env->low_signed : env->low_unsigned;
2196 unk_l = new_r_Dummy(irg, mode_l);
2197 unk_h = new_r_Dummy(irg, mode_h);
2198 for (i = 0; i < arity; ++i) {
2203 dbg = get_irn_dbg_info(phi);
2204 block = get_nodes_block(phi);
2205 phi_l = new_rd_Phi(dbg, block, arity, in_l, mode_l);
2206 phi_h = new_rd_Phi(dbg, block, arity, in_h, mode_h);
2208 ir_set_dw_lowered(phi, phi_l, phi_h);
2210 /* remember that we need to fixup the predecessors later */
2211 ARR_APP1(ir_node*, env->lowered_phis, phi);
2214 static void fixup_phi(ir_node *phi)
2216 const lower64_entry_t *entry = get_node_entry(phi);
2217 ir_node *phi_l = entry->low_word;
2218 ir_node *phi_h = entry->high_word;
2219 int arity = get_Phi_n_preds(phi);
2222 /* exchange phi predecessors which are lowered by now */
2223 for (i = 0; i < arity; ++i) {
2224 ir_node *pred = get_Phi_pred(phi, i);
2225 const lower64_entry_t *pred_entry = get_node_entry(pred);
2227 set_Phi_pred(phi_l, i, pred_entry->low_word);
2228 set_Phi_pred(phi_h, i, pred_entry->high_word);
2235 static void lower_Mux(ir_node *mux, ir_mode *mode)
2237 ir_node *truen = get_Mux_true(mux);
2238 ir_node *falsen = get_Mux_false(mux);
2239 ir_node *sel = get_Mux_sel(mux);
2240 const lower64_entry_t *true_entry = get_node_entry(truen);
2241 const lower64_entry_t *false_entry = get_node_entry(falsen);
2242 ir_node *true_l = true_entry->low_word;
2243 ir_node *true_h = true_entry->high_word;
2244 ir_node *false_l = false_entry->low_word;
2245 ir_node *false_h = false_entry->high_word;
2246 dbg_info *dbgi = get_irn_dbg_info(mux);
2247 ir_node *block = get_nodes_block(mux);
2249 = new_rd_Mux(dbgi, block, sel, false_l, true_l, env->low_unsigned);
2251 = new_rd_Mux(dbgi, block, sel, false_h, true_h, mode);
2252 ir_set_dw_lowered(mux, res_low, res_high);
2256 * Translate an ASM node.
2258 static void lower_ASM(ir_node *asmn, ir_mode *mode)
2260 ir_mode *high_signed = env->high_signed;
2261 ir_mode *high_unsigned = env->high_unsigned;
2262 int n_outs = get_ASM_n_output_constraints(asmn);
2263 ir_asm_constraint *output_constraints = get_ASM_output_constraints(asmn);
2264 ir_asm_constraint *input_constraints = get_ASM_input_constraints(asmn);
2265 unsigned n_64bit_outs = 0;
2270 for (i = get_irn_arity(asmn) - 1; i >= 0; --i) {
2271 ir_node *op = get_irn_n(asmn, i);
2272 ir_mode *op_mode = get_irn_mode(op);
2273 if (op_mode == high_signed || op_mode == high_unsigned) {
2274 panic("lowering ASM 64bit input unimplemented");
2278 for (i = 0; i < n_outs; ++i) {
2279 const ir_asm_constraint *constraint = &output_constraints[i];
2280 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2281 const char *constr = get_id_str(constraint->constraint);
2283 /* TODO: How to do this architecture neutral? This is very
2284 * i386 specific... */
2285 if (constr[0] != '=' || constr[1] != 'A') {
2286 panic("lowering ASM 64bit output only supports '=A' currently");
2291 if (n_64bit_outs == 0)
2295 dbg_info *dbgi = get_irn_dbg_info(asmn);
2296 ir_node *block = get_nodes_block(asmn);
2297 int arity = get_irn_arity(asmn);
2298 ir_node **in = get_irn_in(asmn) + 1;
2300 int n_clobber = get_ASM_n_clobbers(asmn);
2301 long *proj_map = ALLOCAN(long, n_outs);
2302 ident **clobbers = get_ASM_clobbers(asmn);
2303 ident *asm_text = get_ASM_text(asmn);
2304 ir_asm_constraint *new_outputs
2305 = ALLOCAN(ir_asm_constraint, n_outs+n_64bit_outs);
2307 const ir_edge_t *edge;
2308 const ir_edge_t *next;
2310 for (i = 0; i < n_outs; ++i) {
2311 const ir_asm_constraint *constraint = &output_constraints[i];
2312 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2313 new_outputs[new_n_outs].pos = constraint->pos;
2314 new_outputs[new_n_outs].constraint = new_id_from_str("=a");
2315 new_outputs[new_n_outs].mode = env->low_unsigned;
2316 proj_map[i] = new_n_outs;
2318 new_outputs[new_n_outs].pos = constraint->pos;
2319 new_outputs[new_n_outs].constraint = new_id_from_str("=d");
2320 if (constraint->mode == high_signed)
2321 new_outputs[new_n_outs].mode = env->low_signed;
2323 new_outputs[new_n_outs].mode = env->low_unsigned;
2326 new_outputs[new_n_outs] = *constraint;
2327 proj_map[i] = new_n_outs;
2331 assert(new_n_outs == n_outs+(int)n_64bit_outs);
2333 new_asm = new_rd_ASM(dbgi, block, arity, in, input_constraints,
2334 new_n_outs, new_outputs, n_clobber, clobbers,
2337 foreach_out_edge_safe(asmn, edge, next) {
2338 ir_node *proj = get_edge_src_irn(edge);
2339 ir_mode *proj_mode = get_irn_mode(proj);
2344 pn = get_Proj_proj(proj);
2349 pn = new_n_outs + pn - n_outs;
2351 if (proj_mode == high_signed || proj_mode == high_unsigned) {
2353 = proj_mode == high_signed ? env->low_signed : env->low_unsigned;
2354 ir_node *np_low = new_r_Proj(new_asm, env->low_unsigned, pn);
2355 ir_node *np_high = new_r_Proj(new_asm, high_mode, pn+1);
2356 ir_set_dw_lowered(proj, np_low, np_high);
2358 ir_node *np = new_r_Proj(new_asm, proj_mode, pn);
2366 * Lower the builtin type to its higher part.
2368 * @param mtp the builtin type to lower
2370 * @return the lowered type
2372 static ir_type *lower_Builtin_type_high(ir_type *mtp)
2378 bool must_be_lowered;
2380 res = (ir_type*)pmap_get(lowered_builtin_type_high, mtp);
2384 n_params = get_method_n_params(mtp);
2385 n_results = get_method_n_ress(mtp);
2386 must_be_lowered = false;
2388 /* check for double word parameter */
2389 for (i = n_params; i > 0;) {
2390 ir_type *tp = get_method_param_type(mtp, --i);
2392 if (is_Primitive_type(tp)) {
2393 ir_mode *mode = get_type_mode(tp);
2395 if (mode == env->high_signed || mode == env->high_unsigned) {
2396 must_be_lowered = true;
2402 if (!must_be_lowered) {
2403 set_type_link(mtp, NULL);
2407 res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2409 /* set param types and result types */
2410 for (i = 0; i < n_params; ++i) {
2411 ir_type *tp = get_method_param_type(mtp, i);
2413 if (is_Primitive_type(tp)) {
2414 ir_mode *mode = get_type_mode(tp);
2416 if (mode == env->high_signed) {
2417 if (env->params->little_endian) {
2418 set_method_param_type(res, i, tp_u);
2420 set_method_param_type(res, i, tp_s);
2422 } else if (mode == env->high_unsigned) {
2423 set_method_param_type(res, i, tp_u);
2425 set_method_param_type(res, i, tp);
2428 set_method_param_type(res, i, tp);
2431 for (i = n_results = 0; i < n_results; ++i) {
2432 ir_type *tp = get_method_res_type(mtp, i);
2434 if (is_Primitive_type(tp)) {
2435 ir_mode *mode = get_type_mode(tp);
2437 if (mode == env->high_signed) {
2438 if (env->params->little_endian) {
2439 set_method_res_type(res, i, tp_u);
2441 set_method_res_type(res, i, tp_s);
2443 } else if (mode == env->high_unsigned) {
2444 set_method_res_type(res, i, tp_u);
2446 set_method_res_type(res, i, tp);
2449 set_method_res_type(res, i, tp);
2453 set_method_variadicity(res, get_method_variadicity(mtp));
2454 set_method_calling_convention(res, get_method_calling_convention(mtp));
2455 set_method_additional_properties(res, get_method_additional_properties(mtp));
2457 pmap_insert(lowered_builtin_type_high, mtp, res);
2462 * Lower the builtin type to its lower part.
2464 * @param mtp the builtin type to lower
2466 * @return the lowered type
2468 static ir_type *lower_Builtin_type_low(ir_type *mtp)
2474 bool must_be_lowered;
2476 res = (ir_type*)pmap_get(lowered_builtin_type_low, mtp);
2480 n_params = get_method_n_params(mtp);
2481 n_results = get_method_n_ress(mtp);
2482 must_be_lowered = false;
2484 /* check for double word parameter */
2485 for (i = n_params; i > 0;) {
2486 ir_type *tp = get_method_param_type(mtp, --i);
2488 if (is_Primitive_type(tp)) {
2489 ir_mode *mode = get_type_mode(tp);
2491 if (mode == env->high_signed || mode == env->high_unsigned) {
2492 must_be_lowered = true;
2498 if (!must_be_lowered) {
2499 set_type_link(mtp, NULL);
2503 res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2505 /* set param types and result types */
2506 for (i = 0; i < n_params; ++i) {
2507 ir_type *tp = get_method_param_type(mtp, i);
2509 if (is_Primitive_type(tp)) {
2510 ir_mode *mode = get_type_mode(tp);
2512 if (mode == env->high_signed) {
2513 if (env->params->little_endian) {
2514 set_method_param_type(res, i, tp_s);
2516 set_method_param_type(res, i, tp_u);
2518 } else if (mode == env->high_unsigned) {
2519 set_method_param_type(res, i, tp_u);
2521 set_method_param_type(res, i, tp);
2524 set_method_param_type(res, i, tp);
2527 for (i = 0; i < n_results; ++i) {
2528 ir_type *tp = get_method_res_type(mtp, i);
2530 if (is_Primitive_type(tp)) {
2531 ir_mode *mode = get_type_mode(tp);
2533 if (mode == env->high_signed) {
2534 if (env->params->little_endian) {
2535 set_method_res_type(res, i, tp_s);
2537 set_method_res_type(res, i, tp_u);
2539 } else if (mode == env->high_unsigned) {
2540 set_method_res_type(res, i, tp_u);
2542 set_method_res_type(res, i, tp);
2545 set_method_res_type(res, i, tp);
2549 set_method_variadicity(res, get_method_variadicity(mtp));
2550 set_method_calling_convention(res, get_method_calling_convention(mtp));
2551 set_method_additional_properties(res, get_method_additional_properties(mtp));
2553 pmap_insert(lowered_builtin_type_low, mtp, res);
2558 * lowers a builtin which reduces a 64bit value to a simple summary value
2559 * (popcount, ffs, ...)
2561 static void lower_reduce_builtin(ir_node *builtin, ir_mode *mode)
2563 ir_builtin_kind kind = get_Builtin_kind(builtin);
2564 ir_node *operand = get_Builtin_param(builtin, 0);
2565 ir_mode *operand_mode = get_irn_mode(operand);
2566 if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2570 arch_allow_ifconv_func allow_ifconv = be_get_backend_param()->allow_ifconv;
2571 int arity = get_irn_arity(builtin);
2572 dbg_info *dbgi = get_irn_dbg_info(builtin);
2573 ir_graph *irg = get_irn_irg(builtin);
2574 ir_type *type = get_Builtin_type(builtin);
2575 ir_type *lowered_type_high = lower_Builtin_type_high(type);
2576 ir_type *lowered_type_low = lower_Builtin_type_low(type);
2577 ir_type *result_type = get_method_res_type(lowered_type_low, 0);
2578 ir_mode *result_mode = get_type_mode(result_type);
2579 ir_node *block = get_nodes_block(builtin);
2580 ir_node *mem = get_Builtin_mem(builtin);
2581 const lower64_entry_t *entry = get_node_entry(operand);
2582 ir_mode *high_mode = get_irn_mode(entry->high_word);
2583 ir_node *in_high[1] = {entry->high_word};
2584 ir_node *in_low[1] = {entry->low_word};
2587 assert(is_NoMem(mem));
2592 ir_node *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2593 ir_node *zero_high = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2594 ir_node *zero_unsigned = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2595 ir_node *zero_result = new_rd_Const(dbgi, irg, get_mode_null(result_mode));
2596 ir_node *cmp_low = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2597 ir_node *cmp_high = new_rd_Cmp(dbgi, block, entry->high_word, zero_high, ir_relation_equal);
2598 ir_node *ffs_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2599 ir_node *high_proj = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2600 ir_node *high = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2601 ir_node *ffs_low = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2602 ir_node *low = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2603 ir_node *mux_high = new_rd_Mux(dbgi, block, cmp_high, high, zero_result, result_mode);
2605 if (! allow_ifconv(cmp_high, high, zero_result))
2606 ir_nodeset_insert(&created_mux_nodes, mux_high);
2608 res = new_rd_Mux(dbgi, block, cmp_low, low, mux_high, result_mode);
2610 if (! allow_ifconv(cmp_low, low, mux_high))
2611 ir_nodeset_insert(&created_mux_nodes, res);
2615 ir_node *zero = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2616 ir_node *cmp_high = new_rd_Cmp(dbgi, block, entry->high_word, zero, ir_relation_equal);
2617 ir_node *clz_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2618 ir_node *high = new_r_Proj(clz_high, result_mode, pn_Builtin_max+1);
2619 ir_node *clz_low = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2620 ir_node *low_proj = new_r_Proj(clz_low, result_mode, pn_Builtin_max+1);
2621 ir_node *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(mode));
2622 ir_node *low = new_rd_Add(dbgi, block, low_proj, number_of_bits, result_mode);
2624 res = new_rd_Mux(dbgi, block, cmp_high, high, low, result_mode);
2626 if (! allow_ifconv(cmp_high, high, low))
2627 ir_nodeset_insert(&created_mux_nodes, res);
2631 ir_node *zero_unsigned = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2632 ir_node *cmp_low = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2633 ir_node *ffs_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2634 ir_node *high_proj = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2635 ir_node *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2636 ir_node *high = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2637 ir_node *ffs_low = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2638 ir_node *low = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2640 res = new_rd_Mux(dbgi, block, cmp_low, low, high, result_mode);
2642 if (! allow_ifconv(cmp_low, low, high))
2643 ir_nodeset_insert(&created_mux_nodes, res);
2646 case ir_bk_popcount: {
2647 ir_node *popcount_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2648 ir_node *popcount_low = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2649 ir_node *high = new_r_Proj(popcount_high, result_mode, pn_Builtin_max+1);
2650 ir_node *low = new_r_Proj(popcount_low, result_mode, pn_Builtin_max+1);
2652 res = new_rd_Add(dbgi, block, high, low, result_mode);
2655 case ir_bk_parity: {
2656 ir_node *parity_high;
2657 ir_node *parity_low;
2663 parity_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2664 high = new_r_Proj(parity_high, result_mode, pn_Builtin_max+1);
2665 parity_low = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2666 low = new_r_Proj(parity_low, result_mode, pn_Builtin_max+1);
2667 res = new_rd_Eor(dbgi, block, high, low, result_mode);
2671 panic("unexpected builtin");
2674 turn_into_tuple(builtin, 2);
2675 set_irn_n(builtin, pn_Builtin_M, mem);
2676 set_irn_n(builtin, pn_Builtin_max+1, res);
2681 * lowers builtins performing arithmetic (bswap)
2683 static void lower_arithmetic_builtin(ir_node *builtin, ir_mode *mode)
2685 ir_builtin_kind kind = get_Builtin_kind(builtin);
2686 ir_node *operand = get_Builtin_param(builtin, 0);
2687 ir_mode *operand_mode = get_irn_mode(operand);
2689 if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2693 dbg_info *dbgi = get_irn_dbg_info(builtin);
2694 ir_type *type = get_Builtin_type(builtin);
2695 ir_type *lowered_type_high = lower_Builtin_type_high(type);
2696 ir_type *lowered_type_low = lower_Builtin_type_low(type);
2697 ir_node *block = get_nodes_block(builtin);
2698 ir_node *mem = get_Builtin_mem(builtin);
2699 const lower64_entry_t *entry = get_node_entry(operand);
2700 ir_mode *mode_high = get_irn_mode(entry->high_word);
2701 const ir_edge_t *edge;
2702 const ir_edge_t *next;
2708 ir_node *in_high[1] = { entry->high_word };
2709 ir_node *in_low[1] = { entry->low_word };
2710 ir_node *swap_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2711 ir_node *swap_low = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2712 ir_node *high = new_r_Proj(swap_high, mode_high, pn_Builtin_max+1);
2713 ir_node *low = new_r_Proj(swap_low, env->low_unsigned, pn_Builtin_max+1);
2714 if (mode_high == env->low_signed) {
2715 res_high = new_rd_Conv(dbgi, block, low, env->low_signed);
2716 res_low = new_rd_Conv(dbgi, block, high, env->low_unsigned);
2724 panic("unexpected builtin");
2727 /* search result Proj */
2728 foreach_out_edge_safe(builtin, edge, next) {
2729 ir_node *proj = get_edge_src_irn(edge);
2733 if (get_Proj_proj(proj) == pn_Builtin_max+1) {
2734 ir_set_dw_lowered(proj, res_low, res_high);
2741 * Lower double word builtins.
2743 static void lower_Builtin(ir_node *builtin, ir_mode *mode)
2745 ir_builtin_kind kind = get_Builtin_kind(builtin);
2749 case ir_bk_debugbreak:
2750 case ir_bk_return_address:
2751 case ir_bk_frame_address:
2752 case ir_bk_prefetch:
2755 case ir_bk_inner_trampoline:
2756 /* Nothing to do. */
2759 lower_arithmetic_builtin(builtin, mode);
2764 case ir_bk_popcount:
2766 lower_reduce_builtin(builtin, mode);
2769 panic("unknown builtin");
2773 * check for opcodes that must always be lowered.
2775 static bool always_lower(unsigned code)
2794 * Compare two op_mode_entry_t's.
2796 static int cmp_op_mode(const void *elt, const void *key, size_t size)
2798 const op_mode_entry_t *e1 = (const op_mode_entry_t*)elt;
2799 const op_mode_entry_t *e2 = (const op_mode_entry_t*)key;
2802 return (e1->op != e2->op) | (e1->imode != e2->imode) | (e1->omode != e2->omode);
2806 * Compare two conv_tp_entry_t's.
2808 static int cmp_conv_tp(const void *elt, const void *key, size_t size)
2810 const conv_tp_entry_t *e1 = (const conv_tp_entry_t*)elt;
2811 const conv_tp_entry_t *e2 = (const conv_tp_entry_t*)key;
2814 return (e1->imode != e2->imode) | (e1->omode != e2->omode);
2818 * Enter a lowering function into an ir_op.
2820 void ir_register_dw_lower_function(ir_op *op, lower_dw_func func)
2822 op->ops.generic = (op_func)func;
2825 /* Determine which modes need to be lowered */
2826 static void setup_modes(void)
2828 unsigned size_bits = env->params->doubleword_size;
2829 ir_mode *doubleword_signed = NULL;
2830 ir_mode *doubleword_unsigned = NULL;
2831 size_t n_modes = ir_get_n_modes();
2832 ir_mode_arithmetic arithmetic;
2833 unsigned modulo_shift;
2836 /* search for doubleword modes... */
2837 for (i = 0; i < n_modes; ++i) {
2838 ir_mode *mode = ir_get_mode(i);
2839 if (!mode_is_int(mode))
2841 if (get_mode_size_bits(mode) != size_bits)
2843 if (mode_is_signed(mode)) {
2844 if (doubleword_signed != NULL) {
2845 /* sigh - the lowerer should really just lower all mode with
2846 * size_bits it finds. Unfortunately this required a bigger
2848 panic("multiple double word signed modes found");
2850 doubleword_signed = mode;
2852 if (doubleword_unsigned != NULL) {
2853 /* sigh - the lowerer should really just lower all mode with
2854 * size_bits it finds. Unfortunately this required a bigger
2856 panic("multiple double word unsigned modes found");
2858 doubleword_unsigned = mode;
2861 if (doubleword_signed == NULL || doubleword_unsigned == NULL) {
2862 panic("Couldn't find doubleword modes");
2865 arithmetic = get_mode_arithmetic(doubleword_signed);
2866 modulo_shift = get_mode_modulo_shift(doubleword_signed);
2868 assert(get_mode_size_bits(doubleword_unsigned) == size_bits);
2869 assert(size_bits % 2 == 0);
2870 assert(get_mode_sign(doubleword_signed) == 1);
2871 assert(get_mode_sign(doubleword_unsigned) == 0);
2872 assert(get_mode_sort(doubleword_signed) == irms_int_number);
2873 assert(get_mode_sort(doubleword_unsigned) == irms_int_number);
2874 assert(get_mode_arithmetic(doubleword_unsigned) == arithmetic);
2875 assert(get_mode_modulo_shift(doubleword_unsigned) == modulo_shift);
2877 /* try to guess a sensible modulo shift for the new mode.
2878 * (This is IMO another indication that this should really be a node
2879 * attribute instead of a mode thing) */
2880 if (modulo_shift == size_bits) {
2881 modulo_shift = modulo_shift / 2;
2882 } else if (modulo_shift == 0) {
2885 panic("Don't know what new modulo shift to use for lowered doubleword mode");
2889 /* produce lowered modes */
2890 env->high_signed = doubleword_signed;
2891 env->high_unsigned = doubleword_unsigned;
2892 env->low_signed = new_int_mode("WS", arithmetic, size_bits, 1,
2894 env->low_unsigned = new_int_mode("WU", arithmetic, size_bits, 0,
2898 static void enqueue_preds(ir_node *node)
2900 int arity = get_irn_arity(node);
2903 for (i = 0; i < arity; ++i) {
2904 ir_node *pred = get_irn_n(node, i);
2905 pdeq_putr(env->waitq, pred);
2909 static void lower_node(ir_node *node)
2917 lower64_entry_t *entry;
2919 if (irn_visited_else_mark(node))
2922 /* cycles are always broken at Phi and Block nodes. So we don't need special
2923 * magic in all the other lower functions */
2924 if (is_Block(node)) {
2925 enqueue_preds(node);
2927 } else if (is_Phi(node)) {
2932 /* depth-first: descend into operands */
2933 if (!is_Block(node)) {
2934 ir_node *block = get_nodes_block(node);
2938 if (!is_Cond(node)) {
2939 arity = get_irn_arity(node);
2940 for (i = 0; i < arity; ++i) {
2941 ir_node *pred = get_irn_n(node, i);
2946 op = get_irn_op(node);
2947 func = (lower_dw_func) op->ops.generic;
2951 idx = get_irn_idx(node);
2952 entry = idx < env->n_entries ? env->entries[idx] : NULL;
2953 if (entry != NULL || always_lower(get_irn_opcode(node))) {
2954 mode = get_irn_op_mode(node);
2955 if (mode == env->high_signed) {
2956 mode = env->low_signed;
2958 mode = env->low_unsigned;
2960 DB((dbg, LEVEL_1, " %+F\n", node));
2965 static void clear_node_and_phi_links(ir_node *node, void *data)
2968 if (get_irn_mode(node) == mode_T) {
2969 set_irn_link(node, node);
2971 set_irn_link(node, NULL);
2974 set_Block_phis(node, NULL);
2975 else if (is_Phi(node))
2976 set_Phi_next(node, NULL);
2979 static void lower_irg(ir_graph *irg)
2983 ir_type *lowered_mtp;
2986 obstack_init(&env->obst);
2988 /* just here for debugging */
2989 current_ir_graph = irg;
2992 n_idx = get_irg_last_idx(irg);
2993 n_idx = n_idx + (n_idx >> 2); /* add 25% */
2994 env->n_entries = n_idx;
2995 env->entries = NEW_ARR_F(lower64_entry_t*, n_idx);
2996 memset(env->entries, 0, sizeof(env->entries[0]) * n_idx);
3001 ent = get_irg_entity(irg);
3002 mtp = get_entity_type(ent);
3003 lowered_mtp = lower_mtp(mtp);
3005 if (lowered_mtp != mtp) {
3006 set_entity_type(ent, lowered_mtp);
3007 env->flags |= MUST_BE_LOWERED;
3009 fix_parameter_entities(irg, mtp);
3012 /* first step: link all nodes and allocate data */
3013 ir_reserve_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
3014 visit_all_identities(irg, clear_node_and_phi_links, NULL);
3015 irg_walk_graph(irg, NULL, prepare_links_and_handle_rotl, env);
3017 if (env->flags & MUST_BE_LOWERED) {
3019 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
3020 inc_irg_visited(irg);
3022 assert(pdeq_empty(env->waitq));
3023 pdeq_putr(env->waitq, get_irg_end(irg));
3025 env->lowered_phis = NEW_ARR_F(ir_node*, 0);
3026 while (!pdeq_empty(env->waitq)) {
3027 ir_node *node = (ir_node*)pdeq_getl(env->waitq);
3031 /* we need to fixup phis */
3032 for (i = 0; i < ARR_LEN(env->lowered_phis); ++i) {
3033 ir_node *phi = env->lowered_phis[i];
3036 DEL_ARR_F(env->lowered_phis);
3039 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
3041 if (env->flags & CF_CHANGED) {
3042 /* control flow changed, dominance info is invalid */
3043 clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
3045 edges_deactivate(irg);
3048 ir_free_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
3050 DEL_ARR_F(env->entries);
3051 obstack_free(&env->obst, NULL);
3054 static const lwrdw_param_t *param;
3056 void ir_prepare_dw_lowering(const lwrdw_param_t *new_param)
3058 assert(new_param != NULL);
3059 FIRM_DBG_REGISTER(dbg, "firm.lower.dw");
3063 ir_clear_opcodes_generic_func();
3064 ir_register_dw_lower_function(op_ASM, lower_ASM);
3065 ir_register_dw_lower_function(op_Add, lower_binop);
3066 ir_register_dw_lower_function(op_And, lower_And);
3067 ir_register_dw_lower_function(op_Bad, lower_Bad);
3068 ir_register_dw_lower_function(op_Builtin, lower_Builtin);
3069 ir_register_dw_lower_function(op_Call, lower_Call);
3070 ir_register_dw_lower_function(op_Cmp, lower_Cmp);
3071 ir_register_dw_lower_function(op_Cond, lower_Cond);
3072 ir_register_dw_lower_function(op_Const, lower_Const);
3073 ir_register_dw_lower_function(op_Conv, lower_Conv);
3074 ir_register_dw_lower_function(op_Div, lower_Div);
3075 ir_register_dw_lower_function(op_Eor, lower_Eor);
3076 ir_register_dw_lower_function(op_Load, lower_Load);
3077 ir_register_dw_lower_function(op_Minus, lower_unop);
3078 ir_register_dw_lower_function(op_Mod, lower_Mod);
3079 ir_register_dw_lower_function(op_Mul, lower_binop);
3080 ir_register_dw_lower_function(op_Mux, lower_Mux);
3081 ir_register_dw_lower_function(op_Not, lower_Not);
3082 ir_register_dw_lower_function(op_Or, lower_Or);
3083 ir_register_dw_lower_function(op_Proj, lower_Proj);
3084 ir_register_dw_lower_function(op_Return, lower_Return);
3085 ir_register_dw_lower_function(op_Shl, lower_Shl);
3086 ir_register_dw_lower_function(op_Shr, lower_Shr);
3087 ir_register_dw_lower_function(op_Shrs, lower_Shrs);
3088 ir_register_dw_lower_function(op_Start, lower_Start);
3089 ir_register_dw_lower_function(op_Store, lower_Store);
3090 ir_register_dw_lower_function(op_Sub, lower_binop);
3091 ir_register_dw_lower_function(op_Switch, lower_Switch);
3092 ir_register_dw_lower_function(op_Unknown, lower_Unknown);
3096 * Callback to lower only the Mux nodes we created.
3098 static int lower_mux_cb(ir_node *mux)
3100 return ir_nodeset_contains(&created_mux_nodes, mux);
3106 void ir_lower_dw_ops(void)
3108 lower_dw_env_t lenv;
3111 memset(&lenv, 0, sizeof(lenv));
3112 lenv.params = param;
3117 /* create the necessary maps */
3118 if (! intrinsic_fkt)
3119 intrinsic_fkt = new_set(cmp_op_mode, iro_Last + 1);
3121 conv_types = new_set(cmp_conv_tp, 16);
3123 lowered_type = pmap_create();
3124 if (! lowered_builtin_type_low)
3125 lowered_builtin_type_low = pmap_create();
3126 if (! lowered_builtin_type_high)
3127 lowered_builtin_type_high = pmap_create();
3129 /* create a primitive unsigned and signed type */
3131 tp_u = get_type_for_mode(lenv.low_unsigned);
3133 tp_s = get_type_for_mode(lenv.low_signed);
3135 /* create method types for the created binop calls */
3137 binop_tp_u = new_type_method(4, 2);
3138 set_method_param_type(binop_tp_u, 0, tp_u);
3139 set_method_param_type(binop_tp_u, 1, tp_u);
3140 set_method_param_type(binop_tp_u, 2, tp_u);
3141 set_method_param_type(binop_tp_u, 3, tp_u);
3142 set_method_res_type(binop_tp_u, 0, tp_u);
3143 set_method_res_type(binop_tp_u, 1, tp_u);
3146 binop_tp_s = new_type_method(4, 2);
3147 if (env->params->little_endian) {
3148 set_method_param_type(binop_tp_s, 0, tp_u);
3149 set_method_param_type(binop_tp_s, 1, tp_s);
3150 set_method_param_type(binop_tp_s, 2, tp_u);
3151 set_method_param_type(binop_tp_s, 3, tp_s);
3152 set_method_res_type(binop_tp_s, 0, tp_u);
3153 set_method_res_type(binop_tp_s, 1, tp_s);
3155 set_method_param_type(binop_tp_s, 0, tp_s);
3156 set_method_param_type(binop_tp_s, 1, tp_u);
3157 set_method_param_type(binop_tp_s, 2, tp_s);
3158 set_method_param_type(binop_tp_s, 3, tp_u);
3159 set_method_res_type(binop_tp_s, 0, tp_s);
3160 set_method_res_type(binop_tp_s, 1, tp_u);
3164 unop_tp_u = new_type_method(2, 2);
3165 set_method_param_type(unop_tp_u, 0, tp_u);
3166 set_method_param_type(unop_tp_u, 1, tp_u);
3167 set_method_res_type(unop_tp_u, 0, tp_u);
3168 set_method_res_type(unop_tp_u, 1, tp_u);
3171 unop_tp_s = new_type_method(2, 2);
3172 if (env->params->little_endian) {
3173 set_method_param_type(unop_tp_s, 0, tp_u);
3174 set_method_param_type(unop_tp_s, 1, tp_s);
3175 set_method_res_type(unop_tp_s, 0, tp_u);
3176 set_method_res_type(unop_tp_s, 1, tp_s);
3178 set_method_param_type(unop_tp_s, 0, tp_s);
3179 set_method_param_type(unop_tp_s, 1, tp_u);
3180 set_method_res_type(unop_tp_s, 0, tp_s);
3181 set_method_res_type(unop_tp_s, 1, tp_u);
3185 lenv.tv_mode_bytes = new_tarval_from_long(param->doubleword_size/(2*8), lenv.low_unsigned);
3186 lenv.tv_mode_bits = new_tarval_from_long(param->doubleword_size/2, lenv.low_unsigned);
3187 lenv.waitq = new_pdeq();
3188 lenv.first_id = new_id_from_chars(param->little_endian ? ".l" : ".h", 2);
3189 lenv.next_id = new_id_from_chars(param->little_endian ? ".h" : ".l", 2);
3191 irp_reserve_resources(irp, IRP_RESOURCE_TYPE_LINK | IRP_RESOURCE_TYPE_VISITED);
3192 inc_master_type_visited();
3193 /* transform all graphs */
3194 for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
3195 ir_graph *irg = get_irp_irg(i);
3197 ir_nodeset_init(&created_mux_nodes);
3201 if (ir_nodeset_size(&created_mux_nodes) > 0)
3202 lower_mux(irg, lower_mux_cb);
3204 ir_nodeset_destroy(&created_mux_nodes);
3206 irp_free_resources(irp, IRP_RESOURCE_TYPE_LINK | IRP_RESOURCE_TYPE_VISITED);
3207 del_pdeq(lenv.waitq);
3212 /* Default implementation. */
3213 ir_entity *def_create_intrinsic_fkt(ir_type *method, const ir_op *op,
3214 const ir_mode *imode, const ir_mode *omode,
3222 if (imode == omode) {
3223 snprintf(buf, sizeof(buf), "__l%s%s", get_op_name(op), get_mode_name(imode));
3225 snprintf(buf, sizeof(buf), "__l%s%s%s", get_op_name(op),
3226 get_mode_name(imode), get_mode_name(omode));
3228 id = new_id_from_str(buf);
3230 ent = new_entity(get_glob_type(), id, method);
3231 set_entity_ld_ident(ent, get_entity_ident(ent));
3232 set_entity_visibility(ent, ir_visibility_external);