2 * Copyright (C) 1995-2007 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 Implements several optimizations for IA32.
23 * @author Christian Wuerdig
33 #include "firm_types.h"
43 #include "../benode_t.h"
44 #include "../besched_t.h"
46 #include "ia32_new_nodes.h"
47 #include "bearch_ia32_t.h"
48 #include "gen_ia32_regalloc_if_t.h"
49 #include "ia32_transform.h"
50 #include "ia32_dbg_stat.h"
51 #include "ia32_util.h"
53 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
58 IA32_AM_CAND_NONE = 0, /**< no addressmode possible with irn inputs */
59 IA32_AM_CAND_LEFT = 1, /**< addressmode possible with left input */
60 IA32_AM_CAND_RIGHT = 2, /**< addressmode possible with right input */
61 IA32_AM_CAND_BOTH = 3 /**< addressmode possible with both inputs */
64 typedef int is_op_func_t(const ir_node *n);
65 typedef ir_node *load_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, ir_node *mem);
68 * checks if a node represents the NOREG value
70 static INLINE int be_is_NoReg(ia32_code_gen_t *cg, const ir_node *irn) {
71 return irn == cg->noreg_gp || irn == cg->noreg_xmm || irn == cg->noreg_vfp;
74 /********************************************************************************************************
75 * _____ _ _ ____ _ _ _ _ _
76 * | __ \ | | | | / __ \ | | (_) (_) | | (_)
77 * | |__) |__ ___ _ __ | |__ ___ | | ___ | | | |_ __ | |_ _ _ __ ___ _ ______ _| |_ _ ___ _ __
78 * | ___/ _ \/ _ \ '_ \| '_ \ / _ \| |/ _ \ | | | | '_ \| __| | '_ ` _ \| |_ / _` | __| |/ _ \| '_ \
79 * | | | __/ __/ |_) | | | | (_) | | __/ | |__| | |_) | |_| | | | | | | |/ / (_| | |_| | (_) | | | |
80 * |_| \___|\___| .__/|_| |_|\___/|_|\___| \____/| .__/ \__|_|_| |_| |_|_/___\__,_|\__|_|\___/|_| |_|
83 ********************************************************************************************************/
86 * NOTE: THESE PEEPHOLE OPTIMIZATIONS MUST BE CALLED AFTER SCHEDULING AND REGISTER ALLOCATION.
89 static int ia32_const_equal(const ir_node *n1, const ir_node *n2) {
90 if(get_ia32_immop_type(n1) != get_ia32_immop_type(n2))
93 if(get_ia32_immop_type(n1) == ia32_ImmConst) {
94 return get_ia32_Immop_tarval(n1) == get_ia32_Immop_tarval(n2);
95 } else if(get_ia32_immop_type(n1) == ia32_ImmSymConst) {
96 return get_ia32_Immop_symconst(n1) == get_ia32_Immop_symconst(n2);
99 assert(get_ia32_immop_type(n1) == ia32_ImmNone);
104 * Checks for potential CJmp/CJmpAM optimization candidates.
106 static ir_node *ia32_determine_cjmp_cand(ir_node *irn, is_op_func_t *is_op_func) {
107 ir_node *cand = NULL;
108 ir_node *prev = sched_prev(irn);
110 if (is_Block(prev)) {
111 if (get_Block_n_cfgpreds(prev) == 1)
112 prev = get_Block_cfgpred(prev, 0);
117 /* The predecessor must be a ProjX. */
118 if (prev && is_Proj(prev) && get_irn_mode(prev) == mode_X) {
119 prev = get_Proj_pred(prev);
121 if (is_op_func(prev))
128 static int is_TestJmp_cand(const ir_node *irn) {
129 return is_ia32_TestJmp(irn) || is_ia32_And(irn);
133 * Checks if two consecutive arguments of cand matches
134 * the two arguments of irn (TestJmp).
136 static int is_TestJmp_replacement(ir_node *cand, ir_node *irn) {
137 ir_node *in1 = get_irn_n(irn, 0);
138 ir_node *in2 = get_irn_n(irn, 1);
139 int i, n = get_irn_arity(cand);
142 for (i = 0; i < n - 1; i++) {
143 if (get_irn_n(cand, i) == in1 &&
144 get_irn_n(cand, i + 1) == in2)
154 return ia32_const_equal(cand, irn);
158 * Tries to replace a TestJmp by a CJmp or CJmpAM (in case of And)
160 static void ia32_optimize_TestJmp(ir_node *irn)
162 ir_node *cand = ia32_determine_cjmp_cand(irn, is_TestJmp_cand);
165 /* we found a possible candidate */
166 replace = cand ? is_TestJmp_replacement(cand, irn) : 0;
169 DBG((dbg, LEVEL_1, "replacing %+F by ", irn));
171 if (is_ia32_And(cand))
172 set_irn_op(irn, op_ia32_CJmpAM);
174 set_irn_op(irn, op_ia32_CJmp);
176 DB((dbg, LEVEL_1, "%+F\n", irn));
180 static int is_CondJmp_cand(const ir_node *irn) {
181 return is_ia32_CondJmp(irn) || is_ia32_Sub(irn);
185 * Checks if the arguments of cand are the same of irn.
187 static int is_CondJmp_replacement(ir_node *cand, ir_node *irn) {
190 arity = get_irn_arity(cand);
191 for (i = 0; i < arity; i++) {
192 if (get_irn_n(cand, i) != get_irn_n(irn, i)) {
197 return ia32_const_equal(cand, irn);
201 * Tries to replace a CondJmp by a CJmpAM
203 static void ia32_optimize_CondJmp(ir_node *irn)
205 ir_node *cand = ia32_determine_cjmp_cand(irn, is_CondJmp_cand);
208 /* we found a possible candidate */
209 replace = cand ? is_CondJmp_replacement(cand, irn) : 0;
212 DBG((dbg, LEVEL_1, "replacing %+F by ", irn));
215 set_irn_op(irn, op_ia32_CJmpAM);
217 DB((dbg, LEVEL_1, "%+F\n", irn));
221 // only optimize up to 48 stores behind IncSPs
222 #define MAXPUSH_OPTIMIZE 48
225 * Tries to create pushs from IncSP,Store combinations
227 static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
231 ir_node *stores[MAXPUSH_OPTIMIZE];
232 ir_node *block = get_nodes_block(irn);
233 ir_graph *irg = cg->irg;
235 ir_mode *spmode = get_irn_mode(irn);
237 memset(stores, 0, sizeof(stores));
239 assert(be_is_IncSP(irn));
241 offset = be_get_IncSP_offset(irn);
246 * We first walk the schedule after the IncSP node as long as we find
247 * suitable stores that could be transformed to a push.
248 * We save them into the stores array which is sorted by the frame offset/4
249 * attached to the node
251 for(node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) {
256 // it has to be a store
257 if(!is_ia32_Store(node))
260 // it has to use our sp value
261 if(get_irn_n(node, 0) != irn)
263 // store has to be attached to NoMem
264 mem = get_irn_n(node, 3);
269 if( (get_ia32_am_flavour(node) & ia32_am_IS) != 0)
272 offset = get_ia32_am_offs_int(node);
274 storeslot = offset / 4;
275 if(storeslot >= MAXPUSH_OPTIMIZE)
278 // storing into the same slot twice is bad (and shouldn't happen...)
279 if(stores[storeslot] != NULL)
282 // storing at half-slots is bad
286 stores[storeslot] = node;
289 curr_sp = get_irn_n(irn, 0);
291 // walk the stores in inverse order and create pushs for them
292 i = (offset / 4) - 1;
293 if(i >= MAXPUSH_OPTIMIZE) {
294 i = MAXPUSH_OPTIMIZE - 1;
297 for( ; i >= 0; --i) {
298 const arch_register_t *spreg;
300 ir_node *val, *mem, *mem_proj;
301 ir_node *store = stores[i];
302 ir_node *noreg = ia32_new_NoReg_gp(cg);
304 if(store == NULL || is_Bad(store))
307 val = get_irn_n(store, 2);
308 mem = get_irn_n(store, 3);
309 spreg = arch_get_irn_register(cg->arch_env, curr_sp);
312 push = new_rd_ia32_Push(NULL, irg, block, noreg, noreg, val, curr_sp, mem);
314 set_ia32_am_support(push, ia32_am_Source);
315 copy_ia32_Immop_attr(push, store);
317 sched_add_before(irn, push);
319 // create stackpointer proj
320 curr_sp = new_r_Proj(irg, block, push, spmode, pn_ia32_Push_stack);
321 arch_set_irn_register(cg->arch_env, curr_sp, spreg);
322 #ifdef SCHEDULE_PROJS
323 sched_add_before(irn, curr_sp);
325 // create memory proj
326 mem_proj = new_r_Proj(irg, block, push, mode_M, pn_ia32_Push_M);
328 // use the memproj now
329 exchange(store, mem_proj);
331 // we can remove the store now
337 be_set_IncSP_offset(irn, offset);
339 // can we remove the IncSP now?
341 const ir_edge_t *edge, *next;
343 foreach_out_edge_safe(irn, edge, next) {
344 ir_node *arg = get_edge_src_irn(edge);
345 int pos = get_edge_src_pos(edge);
347 set_irn_n(arg, pos, curr_sp);
350 set_irn_n(irn, 0, new_Bad());
353 set_irn_n(irn, 0, curr_sp);
359 * Tries to optimize two following IncSP.
361 static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) {
362 ir_node *prev = be_get_IncSP_pred(irn);
363 int real_uses = get_irn_n_edges(prev);
365 if (be_is_IncSP(prev) && real_uses == 1) {
366 /* first IncSP has only one IncSP user, kill the first one */
367 int prev_offs = be_get_IncSP_offset(prev);
368 int curr_offs = be_get_IncSP_offset(irn);
370 be_set_IncSP_offset(prev, prev_offs + curr_offs);
372 /* Omit the optimized IncSP */
373 be_set_IncSP_pred(irn, be_get_IncSP_pred(prev));
375 set_irn_n(prev, 0, new_Bad());
382 * Performs Peephole Optimizations.
384 static void ia32_peephole_optimize_node(ir_node *irn, void *env) {
385 ia32_code_gen_t *cg = env;
387 /* AMD CPUs want explicit compare before conditional jump */
388 if (! ARCH_AMD(cg->opt_arch)) {
389 if (is_ia32_TestJmp(irn))
390 ia32_optimize_TestJmp(irn);
391 else if (is_ia32_CondJmp(irn))
392 ia32_optimize_CondJmp(irn);
395 if (be_is_IncSP(irn)) {
396 // optimize_IncSP doesn't respect dependency edges yet...
397 //ia32_optimize_IncSP(irn, cg);
399 if (cg->opt & IA32_OPT_PUSHARGS)
400 ia32_create_Pushs(irn, cg);
404 void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *cg) {
405 irg_walk_graph(irg, ia32_peephole_optimize_node, NULL, cg);
408 /******************************************************************
410 * /\ | | | | | \/ | | |
411 * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___
412 * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \
413 * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/
414 * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___|
416 ******************************************************************/
423 static int node_is_ia32_comm(const ir_node *irn) {
424 return is_ia32_irn(irn) ? is_ia32_commutative(irn) : 0;
427 static int ia32_get_irn_n_edges(const ir_node *irn) {
428 const ir_edge_t *edge;
431 foreach_out_edge(irn, edge) {
439 * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
441 * @param pred The node to be checked
442 * @param is_op_func The check-function
443 * @return 1 if conditions are fulfilled, 0 otherwise
445 static int pred_is_specific_node(const ir_node *pred, is_op_func_t *is_op_func) {
446 return is_op_func(pred);
450 * Determines if pred is a Proj and if is_op_func returns true for it's predecessor
451 * and if the predecessor is in block bl.
453 * @param bl The block
454 * @param pred The node to be checked
455 * @param is_op_func The check-function
456 * @return 1 if conditions are fulfilled, 0 otherwise
458 static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred,
459 int (*is_op_func)(const ir_node *n))
462 pred = get_Proj_pred(pred);
463 if ((bl == get_nodes_block(pred)) && is_op_func(pred)) {
472 * Checks if irn is a candidate for address calculation. We avoid transforming
473 * adds to leas if they have a load as pred, because then we can use AM mode
476 * - none of the operand must be a Load within the same block OR
477 * - all Loads must have more than one user OR
479 * @param block The block the Loads must/mustnot be in
480 * @param irn The irn to check
481 * return 1 if irn is a candidate, 0 otherwise
483 static int is_addr_candidate(const ir_node *irn)
485 #ifndef AGGRESSIVE_AM
486 const ir_node *block = get_nodes_block(irn);
487 ir_node *left, *right;
490 left = get_irn_n(irn, 2);
491 right = get_irn_n(irn, 3);
493 if (pred_is_specific_nodeblock(block, left, is_ia32_Ld)) {
494 n = ia32_get_irn_n_edges(left);
495 /* load with only one user: don't create LEA */
500 if (pred_is_specific_nodeblock(block, right, is_ia32_Ld)) {
501 n = ia32_get_irn_n_edges(right);
512 * Checks if irn is a candidate for address mode.
515 * - at least one operand has to be a Load within the same block AND
516 * - the load must not have other users than the irn AND
517 * - the irn must not have a frame entity set
519 * @param cg The ia32 code generator
520 * @param h The height information of the irg
521 * @param block The block the Loads must/mustnot be in
522 * @param irn The irn to check
523 * return 0 if irn is no candidate, 1 if left load can be used, 2 if right one, 3 for both
525 static ia32_am_cand_t is_am_candidate(heights_t *h, const ir_node *block, ir_node *irn) {
526 ir_node *in, *load, *other, *left, *right;
527 int is_cand = 0, cand;
530 if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn) ||
531 is_ia32_GetST0(irn) || is_ia32_SetST0(irn) || is_ia32_xStoreSimple(irn))
534 if(get_ia32_frame_ent(irn) != NULL)
535 return IA32_AM_CAND_NONE;
537 left = get_irn_n(irn, 2);
538 arity = get_irn_arity(irn);
539 assert(arity == 5 || arity == 4);
542 right = get_irn_n(irn, 3);
550 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
551 #ifndef AGGRESSIVE_AM
553 n = ia32_get_irn_n_edges(in);
554 is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */
559 load = get_Proj_pred(in);
562 /* 8bit Loads are not supported (for binary ops),
563 * they cannot be used with every register */
564 if (get_irn_arity(irn) != 4 && get_mode_size_bits(get_ia32_ls_mode(load)) < 16) {
565 assert(get_irn_arity(irn) == 5);
569 /* If there is a data dependency of other irn from load: cannot use AM */
570 if (is_cand && get_nodes_block(other) == block) {
571 other = skip_Proj(other);
572 is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
573 /* this could happen in loops */
574 is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
578 cand = is_cand ? IA32_AM_CAND_LEFT : IA32_AM_CAND_NONE;
582 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
583 #ifndef AGGRESSIVE_AM
585 n = ia32_get_irn_n_edges(in);
586 is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */
591 load = get_Proj_pred(in);
594 /* 8bit Loads are not supported, they cannot be used with every register */
595 if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16)
598 /* If there is a data dependency of other irn from load: cannot use load */
599 if (is_cand && get_nodes_block(other) == block) {
600 other = skip_Proj(other);
601 is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
602 /* this could happen in loops */
603 is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
607 cand = is_cand ? (cand | IA32_AM_CAND_RIGHT) : cand;
609 /* if the irn has a frame entity: we do not use address mode */
614 * Compares the base and index addr and the load/store entities
615 * and returns 1 if they are equal.
617 static int load_store_addr_is_equal(const ir_node *load, const ir_node *store,
618 const ir_node *addr_b, const ir_node *addr_i)
620 if(get_irn_n(load, 0) != addr_b)
622 if(get_irn_n(load, 1) != addr_i)
625 if(get_ia32_frame_ent(load) != get_ia32_frame_ent(store))
628 if(get_ia32_am_sc(load) != get_ia32_am_sc(store))
630 if(is_ia32_am_sc_sign(load) != is_ia32_am_sc_sign(store))
632 if(get_ia32_am_offs_int(load) != get_ia32_am_offs_int(store))
634 if(get_ia32_ls_mode(load) != get_ia32_ls_mode(store))
640 typedef enum _ia32_take_lea_attr {
641 IA32_LEA_ATTR_NONE = 0,
642 IA32_LEA_ATTR_BASE = (1 << 0),
643 IA32_LEA_ATTR_INDEX = (1 << 1),
644 IA32_LEA_ATTR_OFFS = (1 << 2),
645 IA32_LEA_ATTR_SCALE = (1 << 3),
646 IA32_LEA_ATTR_AMSC = (1 << 4),
647 IA32_LEA_ATTR_FENT = (1 << 5)
648 } ia32_take_lea_attr;
651 * Decides if we have to keep the LEA operand or if we can assimilate it.
653 static int do_new_lea(ir_node *irn, ir_node *base, ir_node *index, ir_node *lea,
654 int have_am_sc, ia32_code_gen_t *cg)
656 ir_entity *irn_ent = get_ia32_frame_ent(irn);
657 ir_entity *lea_ent = get_ia32_frame_ent(lea);
659 int is_noreg_base = be_is_NoReg(cg, base);
660 int is_noreg_index = be_is_NoReg(cg, index);
661 ia32_am_flavour_t am_flav = get_ia32_am_flavour(lea);
663 /* If the Add and the LEA both have a different frame entity set: keep */
664 if (irn_ent && lea_ent && (irn_ent != lea_ent))
665 return IA32_LEA_ATTR_NONE;
666 else if (! irn_ent && lea_ent)
667 ret_val |= IA32_LEA_ATTR_FENT;
669 /* If the Add and the LEA both have already an address mode symconst: keep */
670 if (have_am_sc && get_ia32_am_sc(lea))
671 return IA32_LEA_ATTR_NONE;
672 else if (get_ia32_am_sc(lea))
673 ret_val |= IA32_LEA_ATTR_AMSC;
675 /* Check the different base-index combinations */
677 if (! is_noreg_base && ! is_noreg_index) {
678 /* Assimilate if base is the lea and the LEA is just a Base + Offset calculation */
679 if ((base == lea) && ! (am_flav & ia32_I ? 1 : 0)) {
680 if (am_flav & ia32_O)
681 ret_val |= IA32_LEA_ATTR_OFFS;
683 ret_val |= IA32_LEA_ATTR_BASE;
686 return IA32_LEA_ATTR_NONE;
688 else if (! is_noreg_base && is_noreg_index) {
689 /* Base is set but index not */
691 /* Base points to LEA: assimilate everything */
692 if (am_flav & ia32_O)
693 ret_val |= IA32_LEA_ATTR_OFFS;
694 if (am_flav & ia32_S)
695 ret_val |= IA32_LEA_ATTR_SCALE;
696 if (am_flav & ia32_I)
697 ret_val |= IA32_LEA_ATTR_INDEX;
699 ret_val |= IA32_LEA_ATTR_BASE;
701 else if (am_flav & ia32_B ? 0 : 1) {
702 /* Base is not the LEA but the LEA is an index only calculation: assimilate */
703 if (am_flav & ia32_O)
704 ret_val |= IA32_LEA_ATTR_OFFS;
705 if (am_flav & ia32_S)
706 ret_val |= IA32_LEA_ATTR_SCALE;
708 ret_val |= IA32_LEA_ATTR_INDEX;
711 return IA32_LEA_ATTR_NONE;
713 else if (is_noreg_base && ! is_noreg_index) {
714 /* Index is set but not base */
716 /* Index points to LEA: assimilate everything */
717 if (am_flav & ia32_O)
718 ret_val |= IA32_LEA_ATTR_OFFS;
719 if (am_flav & ia32_S)
720 ret_val |= IA32_LEA_ATTR_SCALE;
721 if (am_flav & ia32_B)
722 ret_val |= IA32_LEA_ATTR_BASE;
724 ret_val |= IA32_LEA_ATTR_INDEX;
726 else if (am_flav & ia32_I ? 0 : 1) {
727 /* Index is not the LEA but the LEA is a base only calculation: assimilate */
728 if (am_flav & ia32_O)
729 ret_val |= IA32_LEA_ATTR_OFFS;
730 if (am_flav & ia32_S)
731 ret_val |= IA32_LEA_ATTR_SCALE;
733 ret_val |= IA32_LEA_ATTR_BASE;
736 return IA32_LEA_ATTR_NONE;
739 assert(0 && "There must have been set base or index");
746 * Adds res before irn into schedule if irn was scheduled.
747 * @param irn The schedule point
748 * @param res The node to be scheduled
750 static INLINE void try_add_to_sched(ir_node *irn, ir_node *res) {
751 if (sched_is_scheduled(irn))
752 sched_add_before(irn, res);
756 * Removes node from schedule if it is not used anymore. If irn is a mode_T node
757 * all it's Projs are removed as well.
758 * @param irn The irn to be removed from schedule
760 static INLINE void try_remove_from_sched(ir_node *node) {
763 if(get_irn_mode(node) == mode_T) {
764 const ir_edge_t *edge, *next;
765 foreach_out_edge_safe(node, edge, next) {
766 ir_node *proj = get_edge_src_irn(edge);
767 try_remove_from_sched(proj);
771 if(get_irn_n_edges(node) != 0)
774 if (sched_is_scheduled(node)) {
778 arity = get_irn_arity(node);
779 for(i = 0; i < arity; ++i) {
780 set_irn_n(node, i, new_Bad());
785 * Folds Add or Sub to LEA if possible
787 static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn) {
788 ir_graph *irg = get_irn_irg(irn);
789 dbg_info *dbg_info = get_irn_dbg_info(irn);
790 ir_node *block = get_nodes_block(irn);
792 ir_node *shift = NULL;
793 ir_node *lea_o = NULL;
803 ir_entity *am_sc = NULL;
804 ir_entity *lea_ent = NULL;
805 ir_node *noreg = ia32_new_NoReg_gp(cg);
806 ir_node *left, *right, *temp;
807 ir_node *base, *index;
808 int consumed_left_shift;
809 ia32_am_flavour_t am_flav;
811 if (is_ia32_Add(irn))
814 left = get_irn_n(irn, 2);
815 right = get_irn_n(irn, 3);
817 /* "normalize" arguments in case of add with two operands */
818 if (isadd && ! be_is_NoReg(cg, right)) {
819 /* put LEA == ia32_am_O as right operand */
820 if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) {
821 set_irn_n(irn, 2, right);
822 set_irn_n(irn, 3, left);
828 /* put LEA != ia32_am_O as left operand */
829 if (is_ia32_Lea(right) && get_ia32_am_flavour(right) != ia32_am_O) {
830 set_irn_n(irn, 2, right);
831 set_irn_n(irn, 3, left);
837 /* put SHL as left operand iff left is NOT a LEA */
838 if (! is_ia32_Lea(left) && pred_is_specific_node(right, is_ia32_Shl)) {
839 set_irn_n(irn, 2, right);
840 set_irn_n(irn, 3, left);
853 /* check for operation with immediate */
854 if (is_ia32_ImmConst(irn)) {
855 tarval *tv = get_ia32_Immop_tarval(irn);
857 DBG((dbg, LEVEL_1, "\tfound op with imm const"));
859 offs_cnst = get_tarval_long(tv);
862 else if (isadd && is_ia32_ImmSymConst(irn)) {
863 DBG((dbg, LEVEL_1, "\tfound op with imm symconst"));
867 am_sc = get_ia32_Immop_symconst(irn);
868 am_sc_sign = is_ia32_am_sc_sign(irn);
871 /* determine the operand which needs to be checked */
872 temp = be_is_NoReg(cg, right) ? left : right;
874 /* check if right operand is AMConst (LEA with ia32_am_O) */
875 /* but we can only eat it up if there is no other symconst */
876 /* because the linker won't accept two symconsts */
877 if (! have_am_sc && is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) {
878 DBG((dbg, LEVEL_1, "\tgot op with LEA am_O"));
880 offs_lea = get_ia32_am_offs_int(temp);
881 am_sc = get_ia32_am_sc(temp);
882 am_sc_sign = is_ia32_am_sc_sign(temp);
889 else if (temp == right)
894 /* default for add -> make right operand to index */
897 consumed_left_shift = -1;
899 DBG((dbg, LEVEL_1, "\tgot LEA candidate with index %+F\n", index));
901 /* determine the operand which needs to be checked */
903 if (is_ia32_Lea(left)) {
905 consumed_left_shift = 0;
908 /* check for SHL 1,2,3 */
909 if (pred_is_specific_node(temp, is_ia32_Shl)) {
910 ir_node *right = get_irn_n(temp, n_ia32_Shl_right);
912 if (is_ia32_Immediate(right)) {
913 const ia32_immediate_attr_t *attr
914 = get_ia32_immediate_attr_const(right);
915 long shiftval = attr->offset;
918 index = get_irn_n(temp, 2);
919 consumed_left_shift = consumed_left_shift < 0 ? 1 : 0;
923 DBG((dbg, LEVEL_1, "\tgot scaled index %+F\n", index));
929 if (! be_is_NoReg(cg, index)) {
930 /* if we have index, but left == right -> no base */
934 else if (consumed_left_shift == 1) {
935 /* -> base is right operand */
936 base = (right == lea_o) ? noreg : right;
941 /* Try to assimilate a LEA as left operand */
942 if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
943 /* check if we can assimilate the LEA */
944 int take_attr = do_new_lea(irn, base, index, left, have_am_sc, cg);
946 if (take_attr == IA32_LEA_ATTR_NONE) {
947 DBG((dbg, LEVEL_1, "\tleave old LEA, creating new one\n"));
950 DBG((dbg, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
951 lea = left; /* for statistics */
953 if (take_attr & IA32_LEA_ATTR_OFFS)
954 offs = get_ia32_am_offs_int(left);
956 if (take_attr & IA32_LEA_ATTR_AMSC) {
957 am_sc = get_ia32_am_sc(left);
959 am_sc_sign = is_ia32_am_sc_sign(left);
962 if (take_attr & IA32_LEA_ATTR_SCALE)
963 scale = get_ia32_am_scale(left);
965 if (take_attr & IA32_LEA_ATTR_BASE)
966 base = get_irn_n(left, 0);
968 if (take_attr & IA32_LEA_ATTR_INDEX)
969 index = get_irn_n(left, 1);
971 if (take_attr & IA32_LEA_ATTR_FENT)
972 lea_ent = get_ia32_frame_ent(left);
976 /* ok, we can create a new LEA */
978 res = new_rd_ia32_Lea(dbg_info, irg, block, base, index);
979 /* we don't want stuff before the barrier... */
980 if(be_is_NoReg(cg, base) && be_is_NoReg(cg, index)) {
981 add_irn_dep(res, get_irg_frame(irg));
984 /* add the old offset of a previous LEA */
985 add_ia32_am_offs_int(res, offs);
987 /* add the new offset */
989 add_ia32_am_offs_int(res, offs_cnst);
990 add_ia32_am_offs_int(res, offs_lea);
992 /* either lea_O-cnst, -cnst or -lea_O */
993 if (offs_cnst != 0) {
994 add_ia32_am_offs_int(res, offs_lea);
995 add_ia32_am_offs_int(res, -offs_cnst);
997 add_ia32_am_offs_int(res, offs_lea);
1001 /* set the address mode symconst */
1003 set_ia32_am_sc(res, am_sc);
1005 set_ia32_am_sc_sign(res);
1008 /* copy the frame entity (could be set in case of Add */
1009 /* which was a FrameAddr) */
1010 if (lea_ent != NULL) {
1011 set_ia32_frame_ent(res, lea_ent);
1012 set_ia32_use_frame(res);
1014 set_ia32_frame_ent(res, get_ia32_frame_ent(irn));
1015 if(is_ia32_use_frame(irn))
1016 set_ia32_use_frame(res);
1020 set_ia32_am_scale(res, scale);
1022 am_flav = ia32_am_N;
1023 /* determine new am flavour */
1024 if (offs || offs_cnst || offs_lea || have_am_sc) {
1027 if (! be_is_NoReg(cg, base)) {
1030 if (! be_is_NoReg(cg, index)) {
1036 set_ia32_am_flavour(res, am_flav);
1038 set_ia32_op_type(res, ia32_AddrModeS);
1040 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn));
1042 DBG((dbg, LEVEL_1, "\tLEA [%+F + %+F * %d + %d]\n", base, index, scale, get_ia32_am_offs_int(res)));
1044 assert(irn && "Couldn't find result proj");
1046 /* get the result Proj of the Add/Sub */
1047 try_add_to_sched(irn, res);
1049 /* exchange the old op with the new LEA */
1050 try_remove_from_sched(irn);
1053 /* we will exchange it, report here before the Proj is created */
1054 if (shift && lea && lea_o) {
1055 try_remove_from_sched(shift);
1056 try_remove_from_sched(lea);
1057 try_remove_from_sched(lea_o);
1058 DBG_OPT_LEA4(irn, lea_o, lea, shift, res);
1059 } else if (shift && lea) {
1060 try_remove_from_sched(shift);
1061 try_remove_from_sched(lea);
1062 DBG_OPT_LEA3(irn, lea, shift, res);
1063 } else if (shift && lea_o) {
1064 try_remove_from_sched(shift);
1065 try_remove_from_sched(lea_o);
1066 DBG_OPT_LEA3(irn, lea_o, shift, res);
1067 } else if (lea && lea_o) {
1068 try_remove_from_sched(lea);
1069 try_remove_from_sched(lea_o);
1070 DBG_OPT_LEA3(irn, lea_o, lea, res);
1072 try_remove_from_sched(shift);
1073 DBG_OPT_LEA2(irn, shift, res);
1075 try_remove_from_sched(lea);
1076 DBG_OPT_LEA2(irn, lea, res);
1078 try_remove_from_sched(lea_o);
1079 DBG_OPT_LEA2(irn, lea_o, res);
1081 DBG_OPT_LEA1(irn, res);
1090 * Merges a Load/Store node with a LEA.
1091 * @param irn The Load/Store node
1092 * @param lea The LEA
1094 static void merge_loadstore_lea(ir_node *irn, ir_node *lea) {
1095 ir_entity *irn_ent = get_ia32_frame_ent(irn);
1096 ir_entity *lea_ent = get_ia32_frame_ent(lea);
1098 /* If the irn and the LEA both have a different frame entity set: do not merge */
1099 if (irn_ent != NULL && lea_ent != NULL && (irn_ent != lea_ent))
1101 else if (irn_ent == NULL && lea_ent != NULL) {
1102 set_ia32_frame_ent(irn, lea_ent);
1103 set_ia32_use_frame(irn);
1106 /* get the AM attributes from the LEA */
1107 add_ia32_am_offs_int(irn, get_ia32_am_offs_int(lea));
1108 set_ia32_am_scale(irn, get_ia32_am_scale(lea));
1109 set_ia32_am_flavour(irn, get_ia32_am_flavour(lea));
1111 set_ia32_am_sc(irn, get_ia32_am_sc(lea));
1112 if (is_ia32_am_sc_sign(lea))
1113 set_ia32_am_sc_sign(irn);
1115 set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
1117 /* set base and index */
1118 set_irn_n(irn, 0, get_irn_n(lea, 0));
1119 set_irn_n(irn, 1, get_irn_n(lea, 1));
1121 try_remove_from_sched(lea);
1123 /* clear remat flag */
1124 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1126 if (is_ia32_Ld(irn))
1127 DBG_OPT_LOAD_LEA(lea, irn);
1129 DBG_OPT_STORE_LEA(lea, irn);
1134 * Sets new_right index of irn to right and new_left index to left.
1135 * Also exchange left and right
1137 static void exchange_left_right(ir_node *irn, ir_node **left, ir_node **right, int new_left, int new_right) {
1140 set_irn_n(irn, new_right, *right);
1141 set_irn_n(irn, new_left, *left);
1147 /* this is only needed for Compares, but currently ALL nodes
1148 * have this attribute :-) */
1149 set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
1153 * Performs address calculation optimization (create LEAs if possible)
1155 static void optimize_lea(ia32_code_gen_t *cg, ir_node *irn) {
1156 if (! is_ia32_irn(irn))
1159 /* Following cases can occur: */
1160 /* - Sub (l, imm) -> LEA [base - offset] */
1161 /* - Sub (l, r == LEA with ia32_am_O) -> LEA [base - offset] */
1162 /* - Add (l, imm) -> LEA [base + offset] */
1163 /* - Add (l, r == LEA with ia32_am_O) -> LEA [base + offset] */
1164 /* - Add (l == LEA with ia32_am_O, r) -> LEA [base + offset] */
1165 /* - Add (l, r) -> LEA [base + index * scale] */
1166 /* with scale > 1 iff l/r == shl (1,2,3) */
1167 if (is_ia32_Sub(irn) || is_ia32_Add(irn)) {
1170 if(!is_addr_candidate(irn))
1173 DBG((dbg, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn));
1174 res = fold_addr(cg, irn);
1177 DB((dbg, LEVEL_1, "transformed into %+F\n", res));
1179 DB((dbg, LEVEL_1, "not transformed\n"));
1180 } else if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) {
1181 /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */
1182 /* - Store -> LEA into Store } it might be better to keep the LEA */
1183 ir_node *left = get_irn_n(irn, 0);
1185 if (is_ia32_Lea(left)) {
1186 const ir_edge_t *edge, *ne;
1189 /* merge all Loads/Stores connected to this LEA with the LEA */
1190 foreach_out_edge_safe(left, edge, ne) {
1191 src = get_edge_src_irn(edge);
1193 if (src && (get_edge_src_pos(edge) == 0) && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) {
1194 DBG((dbg, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
1195 if (! is_ia32_got_lea(src))
1196 merge_loadstore_lea(src, left);
1197 set_ia32_got_lea(src);
1204 static void optimize_conv_store(ir_node *node)
1208 ir_mode *store_mode;
1210 if(!is_ia32_Store(node) && !is_ia32_Store8Bit(node))
1213 pred = get_irn_n(node, 2);
1214 if(!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred))
1217 /* the store only stores the lower bits, so we only need the conv
1218 * it it shrinks the mode */
1219 conv_mode = get_ia32_ls_mode(pred);
1220 store_mode = get_ia32_ls_mode(node);
1221 if(get_mode_size_bits(conv_mode) < get_mode_size_bits(store_mode))
1224 set_irn_n(node, 2, get_irn_n(pred, 2));
1225 if(get_irn_n_edges(pred) == 0) {
1230 static void optimize_load_conv(ir_node *node)
1232 ir_node *pred, *predpred;
1236 if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node))
1239 pred = get_irn_n(node, 2);
1243 predpred = get_Proj_pred(pred);
1244 if(!is_ia32_Load(predpred))
1247 /* the load is sign extending the upper bits, so we only need the conv
1248 * if it shrinks the mode */
1249 load_mode = get_ia32_ls_mode(predpred);
1250 conv_mode = get_ia32_ls_mode(node);
1251 if(get_mode_size_bits(conv_mode) < get_mode_size_bits(load_mode))
1254 if(get_mode_sign(conv_mode) != get_mode_sign(load_mode)) {
1255 /* change the load if it has only 1 user */
1256 if(get_irn_n_edges(pred) == 1) {
1258 if(get_mode_sign(conv_mode)) {
1259 newmode = find_signed_mode(load_mode);
1261 newmode = find_unsigned_mode(load_mode);
1263 assert(newmode != NULL);
1264 set_ia32_ls_mode(predpred, newmode);
1266 /* otherwise we have to keep the conv */
1272 exchange(node, pred);
1275 static void optimize_conv_conv(ir_node *node)
1281 if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node))
1284 assert(n_ia32_Conv_I2I_val == n_ia32_Conv_I2I8Bit_val);
1285 pred = get_irn_n(node, n_ia32_Conv_I2I_val);
1286 if(!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred))
1289 /* we know that after a conv, the upper bits are sign extended
1290 * so we only need the 2nd conv if it shrinks the mode */
1291 conv_mode = get_ia32_ls_mode(node);
1292 pred_mode = get_ia32_ls_mode(pred);
1293 if(get_mode_size_bits(conv_mode) < get_mode_size_bits(pred_mode))
1296 /* we can't eliminate an upconv signed->unsigned */
1297 if (get_mode_size_bits(conv_mode) != get_mode_size_bits(pred_mode) &&
1298 !get_mode_sign(conv_mode) && get_mode_sign(pred_mode))
1302 exchange(node, pred);
1305 static void optimize_node(ir_node *node, void *env)
1307 ia32_code_gen_t *cg = env;
1309 optimize_load_conv(node);
1310 optimize_conv_store(node);
1311 optimize_conv_conv(node);
1312 optimize_lea(cg, node);
1316 * Checks for address mode patterns and performs the
1317 * necessary transformations.
1318 * This function is called by a walker.
1320 static void optimize_am(ir_node *irn, void *env) {
1321 ia32_am_opt_env_t *am_opt_env = env;
1322 ia32_code_gen_t *cg = am_opt_env->cg;
1323 ir_graph *irg = get_irn_irg(irn);
1324 heights_t *h = am_opt_env->h;
1325 ir_node *block, *left, *right;
1326 ir_node *store, *load, *mem_proj;
1327 ir_node *addr_b, *addr_i;
1328 int need_exchange_on_fail = 0;
1329 ia32_am_type_t am_support;
1330 ia32_am_cand_t cand;
1331 ia32_am_cand_t orig_cand;
1333 int source_possible;
1335 static const arch_register_req_t dest_out_reg_req_0 = {
1336 arch_register_req_type_none,
1337 NULL, /* regclass */
1338 NULL, /* limit bitset */
1340 -1 /* different pos */
1342 static const arch_register_req_t *dest_am_out_reqs[] = {
1346 if (!is_ia32_irn(irn) || is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn))
1348 if (is_ia32_Lea(irn))
1351 am_support = get_ia32_am_support(irn);
1352 block = get_nodes_block(irn);
1354 /* fold following patterns: */
1355 /* - op -> Load into AMop with am_Source */
1357 /* - op is am_Source capable AND */
1358 /* - the Load is only used by this op AND */
1359 /* - the Load is in the same block */
1360 /* - Store -> op -> Load into AMop with am_Dest */
1362 /* - op is am_Dest capable AND */
1363 /* - the Store uses the same address as the Load AND */
1364 /* - the Load is only used by this op AND */
1365 /* - the Load and Store are in the same block AND */
1366 /* - nobody else uses the result of the op */
1367 if (get_ia32_am_support(irn) == ia32_am_None)
1370 cand = is_am_candidate(h, block, irn);
1371 if (cand == IA32_AM_CAND_NONE)
1375 DBG((dbg, LEVEL_1, "\tfound address mode candidate %+F (candleft %d candright %d)... \n", irn,
1376 cand & IA32_AM_CAND_LEFT, cand & IA32_AM_CAND_RIGHT));
1378 left = get_irn_n(irn, 2);
1379 if (get_irn_arity(irn) == 4) {
1380 /* it's an "unary" operation */
1382 assert(cand == IA32_AM_CAND_BOTH);
1384 right = get_irn_n(irn, 3);
1387 dest_possible = am_support & ia32_am_Dest ? 1 : 0;
1388 source_possible = am_support & ia32_am_Source ? 1 : 0;
1390 DBG((dbg, LEVEL_2, "\tdest_possible %d source_possible %d ... \n", dest_possible, source_possible));
1392 if (dest_possible) {
1397 /* we should only have 1 user which is a store */
1398 if (ia32_get_irn_n_edges(irn) == 1) {
1399 ir_node *succ = get_edge_src_irn(get_irn_out_edge_first(irn));
1401 if (is_ia32_xStore(succ) || is_ia32_Store(succ)) {
1403 addr_b = get_irn_n(store, 0);
1404 addr_i = get_irn_n(store, 1);
1408 if (store == NULL) {
1409 DBG((dbg, LEVEL_2, "\tno store found, not using dest_mode\n"));
1414 if (dest_possible) {
1415 /* normalize nodes, we need the interesting load on the left side */
1416 if (cand & IA32_AM_CAND_RIGHT) {
1417 load = get_Proj_pred(right);
1418 if (load_store_addr_is_equal(load, store, addr_b, addr_i)
1419 && node_is_ia32_comm(irn)) {
1420 DBG((dbg, LEVEL_2, "\texchanging left/right\n"));
1421 exchange_left_right(irn, &left, &right, 3, 2);
1422 need_exchange_on_fail ^= 1;
1423 if (cand == IA32_AM_CAND_RIGHT)
1424 cand = IA32_AM_CAND_LEFT;
1429 if (dest_possible) {
1430 if(cand & IA32_AM_CAND_LEFT && is_Proj(left)) {
1431 load = get_Proj_pred(left);
1433 #ifndef AGGRESSIVE_AM
1434 /* we have to be the only user of the load */
1435 if (get_irn_n_edges(left) > 1) {
1436 DBG((dbg, LEVEL_2, "\tmatching load has too may users, not using dest_mode\n"));
1441 DBG((dbg, LEVEL_2, "\tno matching load found, not using dest_mode"));
1446 if (dest_possible) {
1447 /* the store has to use the loads memory or the same memory
1449 ir_node *loadmem = get_irn_n(load, 2);
1450 ir_node *storemem = get_irn_n(store, 3);
1451 assert(get_irn_mode(loadmem) == mode_M);
1452 assert(get_irn_mode(storemem) == mode_M);
1453 /* TODO there could be a sync between store and load... */
1454 if(storemem != loadmem && (!is_Proj(storemem) || get_Proj_pred(storemem) != load)) {
1455 DBG((dbg, LEVEL_2, "\tload/store using different memories, not using dest_mode"));
1460 if (dest_possible) {
1461 /* Compare Load and Store address */
1462 if (!load_store_addr_is_equal(load, store, addr_b, addr_i)) {
1463 DBG((dbg, LEVEL_2, "\taddresses not equal, not using dest_mode"));
1468 if (dest_possible) {
1469 ir_mode *lsmode = get_ia32_ls_mode(load);
1470 if(get_mode_size_bits(lsmode) != 32) {
1475 if (dest_possible) {
1476 /* all conditions fullfilled, do the transformation */
1477 assert(cand & IA32_AM_CAND_LEFT);
1479 /* set new base, index and attributes */
1480 set_irn_n(irn, 0, addr_b);
1481 set_irn_n(irn, 1, addr_i);
1482 add_ia32_am_offs_int(irn, get_ia32_am_offs_int(load));
1483 set_ia32_am_scale(irn, get_ia32_am_scale(load));
1484 set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
1485 set_ia32_op_type(irn, ia32_AddrModeD);
1486 set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
1487 set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
1489 set_ia32_am_sc(irn, get_ia32_am_sc(load));
1490 if (is_ia32_am_sc_sign(load))
1491 set_ia32_am_sc_sign(irn);
1493 /* connect to Load memory and disconnect Load */
1494 if (get_irn_arity(irn) == 5) {
1496 set_irn_n(irn, 4, get_irn_n(load, 2));
1497 set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2));
1500 set_irn_n(irn, 3, get_irn_n(load, 2));
1501 set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2));
1504 /* change node mode and out register requirements */
1505 set_irn_mode(irn, mode_M);
1506 set_ia32_out_req_all(irn, dest_am_out_reqs);
1508 /* connect the memory Proj of the Store to the op */
1509 edges_reroute(store, irn, irg);
1511 /* clear remat flag */
1512 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1514 try_remove_from_sched(store);
1515 try_remove_from_sched(load);
1516 DBG_OPT_AM_D(load, store, irn);
1518 DB((dbg, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
1519 need_exchange_on_fail = 0;
1520 source_possible = 0;
1523 if (source_possible) {
1524 /* normalize ops, we need the load on the right */
1525 if(cand == IA32_AM_CAND_LEFT) {
1526 if(node_is_ia32_comm(irn)) {
1527 exchange_left_right(irn, &left, &right, 3, 2);
1528 need_exchange_on_fail ^= 1;
1529 cand = IA32_AM_CAND_RIGHT;
1531 source_possible = 0;
1536 if (source_possible) {
1537 /* all conditions fullfilled, do transform */
1538 assert(cand & IA32_AM_CAND_RIGHT);
1539 load = get_Proj_pred(right);
1541 if(get_irn_n_edges(load) > 1) {
1542 source_possible = 0;
1546 if (source_possible) {
1547 ir_mode *ls_mode = get_ia32_ls_mode(load);
1548 if(get_mode_size_bits(ls_mode) != 32)
1549 source_possible = 0;
1553 if (source_possible) {
1554 addr_b = get_irn_n(load, 0);
1555 addr_i = get_irn_n(load, 1);
1557 /* set new base, index and attributes */
1558 set_irn_n(irn, 0, addr_b);
1559 set_irn_n(irn, 1, addr_i);
1560 add_ia32_am_offs_int(irn, get_ia32_am_offs_int(load));
1561 set_ia32_am_scale(irn, get_ia32_am_scale(load));
1562 set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
1563 set_ia32_op_type(irn, ia32_AddrModeS);
1564 set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
1565 set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
1567 set_ia32_am_sc(irn, get_ia32_am_sc(load));
1568 if (is_ia32_am_sc_sign(load))
1569 set_ia32_am_sc_sign(irn);
1571 /* clear remat flag */
1572 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1574 if (is_ia32_use_frame(load)) {
1575 if(get_ia32_frame_ent(load) == NULL) {
1576 set_ia32_need_stackent(irn);
1578 set_ia32_use_frame(irn);
1581 /* connect to Load memory and disconnect Load */
1582 if (get_irn_arity(irn) == 5) {
1584 set_irn_n(irn, 3, ia32_get_admissible_noreg(cg, irn, 3));
1585 set_irn_n(irn, 4, get_irn_n(load, 2));
1587 assert(get_irn_arity(irn) == 4);
1589 set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2));
1590 set_irn_n(irn, 3, get_irn_n(load, 2));
1593 DBG_OPT_AM_S(load, irn);
1595 /* If Load has a memory Proj, connect it to the op */
1596 mem_proj = ia32_get_proj_for_mode(load, mode_M);
1597 if (mem_proj != NULL) {
1599 ir_mode *mode = get_irn_mode(irn);
1601 res_proj = new_rd_Proj(get_irn_dbg_info(irn), irg,
1602 get_nodes_block(irn), new_Unknown(mode_T),
1604 set_irn_mode(irn, mode_T);
1605 edges_reroute(irn, res_proj, irg);
1606 set_Proj_pred(res_proj, irn);
1608 set_Proj_pred(mem_proj, irn);
1609 set_Proj_proj(mem_proj, 1);
1611 if(sched_is_scheduled(irn)) {
1612 sched_add_after(irn, res_proj);
1613 sched_add_after(irn, mem_proj);
1617 if(get_irn_n_edges(load) == 0) {
1618 try_remove_from_sched(load);
1620 need_exchange_on_fail = 0;
1622 DB((dbg, LEVEL_1, "merged with %+F into source AM\n", load));
1625 /* was exchanged but optimize failed: exchange back */
1626 if (need_exchange_on_fail) {
1627 exchange_left_right(irn, &left, &right, 3, 2);
1632 * Performs conv and address mode optimization.
1634 void ia32_optimize_graph(ia32_code_gen_t *cg) {
1635 /* if we are supposed to do AM or LEA optimization: recalculate edges */
1636 if (! (cg->opt & (IA32_OPT_DOAM | IA32_OPT_LEA))) {
1637 /* no optimizations at all */
1641 /* beware: we cannot optimize LEA and AM in one run because */
1642 /* LEA optimization adds new nodes to the irg which */
1643 /* invalidates the phase data */
1645 if (cg->opt & IA32_OPT_LEA) {
1646 irg_walk_blkwise_graph(cg->irg, NULL, optimize_node, cg);
1650 be_dump(cg->irg, "-lea", dump_ir_block_graph_sched);
1652 /* hack for now, so these don't get created during optimize, because then
1653 * they will be unknown to the heights module
1655 ia32_new_NoReg_gp(cg);
1656 ia32_new_NoReg_fp(cg);
1657 ia32_new_NoReg_vfp(cg);
1659 if (cg->opt & IA32_OPT_DOAM) {
1660 /* we need height information for am optimization */
1661 heights_t *h = heights_new(cg->irg);
1662 ia32_am_opt_env_t env;
1667 irg_walk_blkwise_graph(cg->irg, NULL, optimize_am, &env);
1673 void ia32_init_optimize(void)
1675 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.optimize");