3 * File name: ir/be/ia32/ia32_optimize.c
4 * Purpose: Implements several optimizations for IA32
5 * Author: Christian Wuerdig
7 * Copyright: (c) 2006 Universität Karlsruhe
8 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
18 #include "firm_types.h"
28 #include "../benode_t.h"
29 #include "../besched_t.h"
31 #include "ia32_new_nodes.h"
32 #include "bearch_ia32_t.h"
33 #include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */
34 #include "ia32_transform.h"
35 #include "ia32_dbg_stat.h"
36 #include "ia32_util.h"
39 IA32_AM_CAND_NONE = 0,
40 IA32_AM_CAND_LEFT = 1,
41 IA32_AM_CAND_RIGHT = 2,
46 #define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
48 typedef int is_op_func_t(const ir_node *n);
49 typedef ir_node *load_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, ir_node *mem);
52 * checks if a node represents the NOREG value
54 static int be_is_NoReg(ia32_code_gen_t *cg, const ir_node *irn) {
55 be_abi_irg_t *babi = cg->birg->abi;
56 const arch_register_t *fp_noreg = USE_SSE2(cg) ?
57 &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG];
59 return (be_abi_get_callee_save_irn(babi, &ia32_gp_regs[REG_GP_NOREG]) == irn) ||
60 (be_abi_get_callee_save_irn(babi, fp_noreg) == irn);
65 /*************************************************
68 * | | ___ _ __ ___| |_ __ _ _ __ | |_ ___
69 * | | / _ \| '_ \/ __| __/ _` | '_ \| __/ __|
70 * | |___| (_) | | | \__ \ || (_| | | | | |_\__ \
71 * \_____\___/|_| |_|___/\__\__,_|_| |_|\__|___/
73 *************************************************/
76 * creates a unique ident by adding a number to a tag
78 * @param tag the tag string, must contain a %d if a number
81 static ident *unique_id(const char *tag)
83 static unsigned id = 0;
86 snprintf(str, sizeof(str), tag, ++id);
87 return new_id_from_str(str);
91 * Transforms a SymConst.
93 * @param mod the debug module
94 * @param block the block the new node should belong to
95 * @param node the ir SymConst node
96 * @param mode mode of the SymConst
97 * @return the created ia32 Const node
99 static ir_node *gen_SymConst(ia32_transform_env_t *env) {
100 dbg_info *dbg = env->dbg;
101 ir_mode *mode = env->mode;
102 ir_graph *irg = env->irg;
103 ir_node *block = env->block;
106 if (mode_is_float(mode)) {
108 if (USE_SSE2(env->cg))
109 cnst = new_rd_ia32_xConst(dbg, irg, block, get_irg_no_mem(irg), mode);
111 cnst = new_rd_ia32_vfConst(dbg, irg, block, get_irg_no_mem(irg), mode);
114 cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), mode);
116 set_ia32_Const_attr(cnst, env->irn);
122 * Get a primitive type for a mode.
124 static ir_type *get_prim_type(pmap *types, ir_mode *mode)
126 pmap_entry *e = pmap_find(types, mode);
131 snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
132 res = new_type_primitive(new_id_from_str(buf), mode);
133 pmap_insert(types, mode, res);
141 * Get an entity that is initialized with a tarval
143 static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
145 tarval *tv = get_Const_tarval(cnst);
146 pmap_entry *e = pmap_find(cg->isa->tv_ent, tv);
151 ir_mode *mode = get_irn_mode(cnst);
152 ir_type *tp = get_Const_type(cnst);
153 if (tp == firm_unknown_type)
154 tp = get_prim_type(cg->isa->types, mode);
156 res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
158 set_entity_ld_ident(res, get_entity_ident(res));
159 set_entity_visibility(res, visibility_local);
160 set_entity_variability(res, variability_constant);
161 set_entity_allocation(res, allocation_static);
163 /* we create a new entity here: It's initialization must resist on the
165 rem = current_ir_graph;
166 current_ir_graph = get_const_code_irg();
167 set_atomic_ent_value(res, new_Const_type(tv, tp));
168 current_ir_graph = rem;
170 pmap_insert(cg->isa->tv_ent, tv, res);
178 * Transforms a Const.
180 * @param mod the debug module
181 * @param block the block the new node should belong to
182 * @param node the ir Const node
183 * @param mode mode of the Const
184 * @return the created ia32 Const node
186 static ir_node *gen_Const(ia32_transform_env_t *env) {
187 ir_node *cnst, *load;
189 ir_graph *irg = env->irg;
190 ir_node *block = env->block;
191 ir_node *node = env->irn;
192 dbg_info *dbg = env->dbg;
193 ir_mode *mode = env->mode;
195 if (mode_is_float(mode)) {
197 if (! USE_SSE2(env->cg)) {
198 cnst_classify_t clss = classify_Const(node);
200 if (clss == CNST_NULL)
201 return new_rd_ia32_vfldz(dbg, irg, block, mode);
202 else if (clss == CNST_ONE)
203 return new_rd_ia32_vfld1(dbg, irg, block, mode);
205 sym.entity_p = get_entity_for_tv(env->cg, node);
208 cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent);
209 load = new_r_Load(irg, block, get_irg_no_mem(irg), cnst, mode);
210 load = new_r_Proj(irg, block, load, mode, pn_Load_res);
213 cnst = gen_SymConst(env);
214 set_Load_ptr(get_Proj_pred(load), cnst);
218 cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), get_irn_mode(node));
219 set_ia32_Const_attr(cnst, node);
225 * Transforms (all) Const's into ia32_Const and places them in the
226 * block where they are used (or in the cfg-pred Block in case of Phi's).
227 * Additionally all reference nodes are changed into mode_Is nodes.
228 * NOTE: irn must be a firm constant!
230 static void ia32_transform_const(ir_node *irn, void *env) {
231 ia32_code_gen_t *cg = env;
232 ir_node *cnst = NULL;
233 ia32_transform_env_t tenv;
237 tenv.mode = get_irn_mode(irn);
238 tenv.dbg = get_irn_dbg_info(irn);
240 DEBUG_ONLY(tenv.mod = cg->mod;)
242 /* place const either in the smallest dominator of all its users or the original block */
243 if (cg->opt & IA32_OPT_PLACECNST)
244 tenv.block = node_users_smallest_common_dominator(irn, 1);
246 tenv.block = get_nodes_block(irn);
248 switch (get_irn_opcode(irn)) {
250 cnst = gen_Const(&tenv);
253 cnst = gen_SymConst(&tenv);
256 assert(0 && "Wrong usage of ia32_transform_const!");
259 assert(cnst && "Could not create ia32 Const");
261 /* set the new ia32 const */
266 * Transform all firm consts and assure, we visit each const only once.
268 static void ia32_place_consts_walker(ir_node *irn, void *env) {
269 ia32_code_gen_t *cg = env;
271 if(!is_Const(irn) && !is_SymConst(irn))
274 ia32_transform_const(irn, cg);
278 * Replace reference modes with mode_Iu and preserve store value modes.
280 static void ia32_set_modes(ir_node *irn, void *env) {
284 /* transform all reference nodes into mode_Iu nodes */
285 if (mode_is_reference(get_irn_mode(irn))) {
286 set_irn_mode(irn, mode_Iu);
291 * Walks over the graph, transforms all firm consts into ia32 consts
292 * and places them into the "best" block.
293 * @param cg The ia32 codegenerator object
295 static void ia32_transform_all_firm_consts(ia32_code_gen_t *cg) {
296 irg_walk_graph(cg->irg, NULL, ia32_place_consts_walker, cg);
299 /* Place all consts and change pointer arithmetics into unsigned integer arithmetics. */
300 void ia32_pre_transform_phase(ia32_code_gen_t *cg) {
302 We need to transform the consts twice:
303 - the psi condition tree transformer needs existing constants to be ia32 constants
304 - the psi condition tree transformer inserts new firm constants which need to be transformed
306 ia32_transform_all_firm_consts(cg);
307 irg_walk_graph(cg->irg, ia32_set_modes, ia32_transform_psi_cond_tree, cg);
308 ia32_transform_all_firm_consts(cg);
311 /********************************************************************************************************
312 * _____ _ _ ____ _ _ _ _ _
313 * | __ \ | | | | / __ \ | | (_) (_) | | (_)
314 * | |__) |__ ___ _ __ | |__ ___ | | ___ | | | |_ __ | |_ _ _ __ ___ _ ______ _| |_ _ ___ _ __
315 * | ___/ _ \/ _ \ '_ \| '_ \ / _ \| |/ _ \ | | | | '_ \| __| | '_ ` _ \| |_ / _` | __| |/ _ \| '_ \
316 * | | | __/ __/ |_) | | | | (_) | | __/ | |__| | |_) | |_| | | | | | | |/ / (_| | |_| | (_) | | | |
317 * |_| \___|\___| .__/|_| |_|\___/|_|\___| \____/| .__/ \__|_|_| |_| |_|_/___\__,_|\__|_|\___/|_| |_|
320 ********************************************************************************************************/
323 * NOTE: THESE PEEPHOLE OPTIMIZATIONS MUST BE CALLED AFTER SCHEDULING AND REGISTER ALLOCATION.
326 static int ia32_cnst_compare(ir_node *n1, ir_node *n2) {
327 return get_ia32_id_cnst(n1) == get_ia32_id_cnst(n2);
331 * Checks for potential CJmp/CJmpAM optimization candidates.
333 static ir_node *ia32_determine_cjmp_cand(ir_node *irn, is_op_func_t *is_op_func) {
334 ir_node *cand = NULL;
335 ir_node *prev = sched_prev(irn);
337 if (is_Block(prev)) {
338 if (get_Block_n_cfgpreds(prev) == 1)
339 prev = get_Block_cfgpred(prev, 0);
344 /* The predecessor must be a ProjX. */
345 if (prev && is_Proj(prev) && get_irn_mode(prev) == mode_X) {
346 prev = get_Proj_pred(prev);
348 if (is_op_func(prev))
355 static int is_TestJmp_cand(const ir_node *irn) {
356 return is_ia32_TestJmp(irn) || is_ia32_And(irn);
360 * Checks if two consecutive arguments of cand matches
361 * the two arguments of irn (TestJmp).
363 static int is_TestJmp_replacement(ir_node *cand, ir_node *irn) {
364 ir_node *in1 = get_irn_n(irn, 0);
365 ir_node *in2 = get_irn_n(irn, 1);
366 int i, n = get_irn_arity(cand);
369 for (i = 0; i < n - 1; i++) {
370 if (get_irn_n(cand, i) == in1 &&
371 get_irn_n(cand, i + 1) == in2)
379 return ia32_cnst_compare(cand, irn);
385 * Tries to replace a TestJmp by a CJmp or CJmpAM (in case of And)
387 static void ia32_optimize_TestJmp(ir_node *irn, ia32_code_gen_t *cg) {
388 ir_node *cand = ia32_determine_cjmp_cand(irn, is_TestJmp_cand);
391 /* we found a possible candidate */
392 replace = cand ? is_TestJmp_replacement(cand, irn) : 0;
395 DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
397 if (is_ia32_And(cand))
398 set_irn_op(irn, op_ia32_CJmpAM);
400 set_irn_op(irn, op_ia32_CJmp);
402 DB((cg->mod, LEVEL_1, "%+F\n", irn));
406 static int is_CondJmp_cand(const ir_node *irn) {
407 return is_ia32_CondJmp(irn) || is_ia32_Sub(irn);
411 * Checks if the arguments of cand are the same of irn.
413 static int is_CondJmp_replacement(ir_node *cand, ir_node *irn) {
414 int i, n = get_irn_arity(cand);
417 for (i = 0; i < n; i++) {
418 if (get_irn_n(cand, i) != get_irn_n(irn, i)) {
425 return ia32_cnst_compare(cand, irn);
431 * Tries to replace a CondJmp by a CJmpAM
433 static void ia32_optimize_CondJmp(ir_node *irn, ia32_code_gen_t *cg) {
434 ir_node *cand = ia32_determine_cjmp_cand(irn, is_CondJmp_cand);
437 /* we found a possible candidate */
438 replace = cand ? is_CondJmp_replacement(cand, irn) : 0;
441 DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
444 set_irn_op(irn, op_ia32_CJmpAM);
446 DB((cg->mod, LEVEL_1, "%+F\n", irn));
452 * Creates a Push from Store(IncSP(gp_reg_size))
454 static void ia32_create_Push(ir_node *irn, ia32_code_gen_t *cg) {
455 ir_node *sp = get_irn_n(irn, 0);
456 ir_graph *irg = cg->irg;
457 ir_node *val, *next, *push, *bl, *proj_M, *proj_res, *old_proj_M, *mem;
458 const ir_edge_t *edge;
461 /* do not create push if store has already an offset assigned or base is not a IncSP */
462 if (get_ia32_am_offs(irn) || ! be_is_IncSP(sp))
465 /* do not create push if index is not NOREG */
466 if (arch_get_irn_register(cg->arch_env, get_irn_n(irn, 1)) !=
467 &ia32_gp_regs[REG_GP_NOREG])
470 /* do not create push for floating point */
471 val = get_irn_n(irn, 2);
472 if (mode_is_float(get_irn_mode(val)))
475 /* do not create push if IncSp doesn't expand stack or expand size is different from register size */
476 if (be_get_IncSP_offset(sp) != get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_gp].mode))
479 /* do not create push, if there is a path (inside the block) from the push value to IncSP */
480 h = heights_new(cg->irg);
481 if (get_nodes_block(val) == get_nodes_block(sp) &&
482 heights_reachable_in_block(h, val, sp))
489 /* ok, translate into Push */
490 edge = get_irn_out_edge_first(irn);
491 old_proj_M = get_edge_src_irn(edge);
492 bl = get_nodes_block(irn);
494 next = sched_next(irn);
500 if the IncSP points to NoMem -> just use the memory input from store
501 if IncSP points to somewhere else -> sync memory of IncSP and Store
503 mem = get_irn_n(irn, 3);
504 push = new_rd_ia32_Push(NULL, irg, bl, be_get_IncSP_pred(sp), val, mem);
505 proj_res = new_r_Proj(irg, bl, push, get_irn_mode(sp), pn_ia32_Push_stack);
506 proj_M = new_r_Proj(irg, bl, push, mode_M, pn_ia32_Push_M);
508 /* copy a possible constant from the store */
509 set_ia32_id_cnst(push, get_ia32_id_cnst(irn));
510 set_ia32_immop_type(push, get_ia32_immop_type(irn));
512 /* the push must have SP out register */
513 arch_set_irn_register(cg->arch_env, push, arch_get_irn_register(cg->arch_env, sp));
515 exchange(old_proj_M, proj_M);
516 exchange(sp, proj_res);
517 sched_add_before(next, push);
518 sched_add_after(push, proj_res);
522 // only optimize up to 48 stores behind IncSPs
523 #define MAXPUSH_OPTIMIZE 48
526 * Tries to create pushs from IncSP,Store combinations
528 static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
533 ir_node *stores[MAXPUSH_OPTIMIZE];
534 ir_node *block = get_nodes_block(irn);
535 ir_graph *irg = cg->irg;
537 ir_mode *spmode = get_irn_mode(irn);
539 memset(stores, 0, sizeof(stores));
541 assert(be_is_IncSP(irn));
543 offset = be_get_IncSP_offset(irn);
548 * We first walk the schedule after the IncSP node as long as we find
549 * suitable stores that could be transformed to a push.
550 * We save them into the stores array which is sorted by the frame offset/4
551 * attached to the node
553 for(node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) {
560 // it has to be a store
561 if(!is_ia32_Store(node))
564 // it has to use our sp value
565 if(get_irn_n(node, 0) != irn)
567 // store has to be attached to NoMem
568 mem = get_irn_n(node, 3);
573 if( (get_ia32_am_flavour(node) & ia32_am_IS) != 0)
576 am_offs = get_ia32_am_offs(node);
577 if(am_offs == NULL) {
580 // the am_offs has to be of the form "+NUMBER"
581 if(sscanf(am_offs, "+%d%n", &offset, &n) != 1 || am_offs[n] != '\0') {
582 // we shouldn't have any cases in the compiler at the moment
583 // that produce something different from esp+XX
589 storeslot = offset / 4;
590 if(storeslot >= MAXPUSH_OPTIMIZE)
593 // storing into the same slot twice is bad (and shouldn't happen...)
594 if(stores[storeslot] != NULL)
597 // storing at half-slots is bad
601 stores[storeslot] = node;
604 offset = be_get_IncSP_offset(irn);
607 for(i = 0; i < MAXPUSH_OPTIMIZE; ++i) {
608 ir_node *store = stores[i];
609 if(store == NULL || is_Bad(store))
619 curr_sp = get_irn_n(irn, 0);
621 // walk the stores in inverse order and create pushs for them
622 for(i = firststore; i >= 0; --i) {
623 const ir_edge_t *edge, *next;
624 const arch_register_t *spreg;
627 ir_node *store = stores[i];
629 val = get_irn_n(store, 2);
630 mem = get_irn_n(store, 3);
631 spreg = arch_get_irn_register(cg->arch_env, curr_sp);
634 push = new_rd_ia32_Push(NULL, irg, block, curr_sp, val, mem);
635 if(get_ia32_immop_type(store) != ia32_ImmNone) {
636 copy_ia32_Immop_attr(push, store);
638 sched_add_before(irn, push);
640 // create stackpointer proj
641 curr_sp = new_r_Proj(irg, block, push, spmode, pn_ia32_Push_stack);
642 arch_set_irn_register(cg->arch_env, curr_sp, spreg);
643 sched_add_before(irn, curr_sp);
645 // rewire memprojs of the store
646 foreach_out_edge_safe(store, edge, next) {
647 ir_node *succ = get_edge_src_irn(edge);
649 assert(is_Proj(succ) && get_Proj_proj(succ) == pn_ia32_Store_M);
650 set_irn_n(succ, 0, push);
653 // we can remove the store from schedule now
657 be_set_IncSP_offset(irn, offset);
659 // can we remove the IncSP now?
661 const ir_edge_t *edge, *next;
664 set_irn_n(irn, 0, new_Bad());
666 foreach_out_edge_safe(irn, edge, next) {
667 ir_node *arg = get_edge_src_irn(edge);
668 int pos = get_edge_src_pos(edge);
670 set_irn_n(arg, pos, curr_sp);
674 set_irn_n(irn, 0, new_Bad());
676 set_irn_n(irn, 0, curr_sp);
681 * Tries to optimize two following IncSP.
683 static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) {
684 ir_node *prev = be_get_IncSP_pred(irn);
685 int real_uses = get_irn_n_edges(prev);
687 if (be_is_IncSP(prev) && real_uses == 1) {
688 /* first IncSP has only one IncSP user, kill the first one */
689 int prev_offs = be_get_IncSP_offset(prev);
690 int curr_offs = be_get_IncSP_offset(irn);
692 be_set_IncSP_offset(prev, prev_offs + curr_offs);
694 /* Omit the optimized IncSP */
695 be_set_IncSP_pred(irn, be_get_IncSP_pred(prev));
701 * Performs Peephole Optimizations.
703 static void ia32_peephole_optimize_node(ir_node *irn, void *env) {
704 ia32_code_gen_t *cg = env;
706 /* AMD CPUs want explicit compare before conditional jump */
707 if (! ARCH_AMD(cg->opt_arch)) {
708 if (is_ia32_TestJmp(irn))
709 ia32_optimize_TestJmp(irn, cg);
710 else if (is_ia32_CondJmp(irn))
711 ia32_optimize_CondJmp(irn, cg);
714 if (be_is_IncSP(irn)) {
715 // optimize_IncSP doesn't respect dependency edges yet...
716 //ia32_optimize_IncSP(irn, cg);
717 ia32_create_Pushs(irn, cg);
721 void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *cg) {
722 irg_walk_graph(irg, ia32_peephole_optimize_node, NULL, cg);
726 /******************************************************************
728 * /\ | | | | | \/ | | |
729 * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___
730 * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \
731 * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/
732 * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___|
734 ******************************************************************/
741 static int node_is_ia32_comm(const ir_node *irn) {
742 return is_ia32_irn(irn) ? is_ia32_commutative(irn) : 0;
745 static int ia32_get_irn_n_edges(const ir_node *irn) {
746 const ir_edge_t *edge;
749 foreach_out_edge(irn, edge) {
757 * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
759 * @param pred The node to be checked
760 * @param is_op_func The check-function
761 * @return 1 if conditions are fulfilled, 0 otherwise
763 static int pred_is_specific_node(const ir_node *pred, is_op_func_t *is_op_func) {
764 if (is_Proj(pred) && is_op_func(get_Proj_pred(pred))) {
772 * Determines if pred is a Proj and if is_op_func returns true for it's predecessor
773 * and if the predecessor is in block bl.
775 * @param bl The block
776 * @param pred The node to be checked
777 * @param is_op_func The check-function
778 * @return 1 if conditions are fulfilled, 0 otherwise
780 static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred,
781 int (*is_op_func)(const ir_node *n))
784 pred = get_Proj_pred(pred);
785 if ((bl == get_nodes_block(pred)) && is_op_func(pred)) {
794 * Checks if irn is a candidate for address calculation.
796 * - none of the operand must be a Load within the same block OR
797 * - all Loads must have more than one user OR
798 * - the irn has a frame entity (it's a former FrameAddr)
800 * @param block The block the Loads must/mustnot be in
801 * @param irn The irn to check
802 * return 1 if irn is a candidate, 0 otherwise
804 static int is_addr_candidate(const ir_node *block, const ir_node *irn) {
805 ir_node *in, *left, *right;
808 left = get_irn_n(irn, 2);
809 right = get_irn_n(irn, 3);
813 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
814 n = ia32_get_irn_n_edges(in);
815 is_cand = (n == 1) ? 0 : is_cand; /* load with only one user: don't create LEA */
820 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
821 n = ia32_get_irn_n_edges(in);
822 is_cand = (n == 1) ? 0 : is_cand; /* load with only one user: don't create LEA */
825 is_cand = get_ia32_frame_ent(irn) ? 1 : is_cand;
831 * Checks if irn is a candidate for address mode.
834 * - at least one operand has to be a Load within the same block AND
835 * - the load must not have other users than the irn AND
836 * - the irn must not have a frame entity set
838 * @param cg The ia32 code generator
839 * @param h The height information of the irg
840 * @param block The block the Loads must/mustnot be in
841 * @param irn The irn to check
842 * return 0 if irn is no candidate, 1 if left load can be used, 2 if right one, 3 for both
844 static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const ir_node *block, ir_node *irn) {
845 ir_node *in, *load, *other, *left, *right;
846 int n, is_cand = 0, cand;
848 if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn) ||
849 is_ia32_GetST0(irn) || is_ia32_SetST0(irn) || is_ia32_xStoreSimple(irn))
852 left = get_irn_n(irn, 2);
853 right = get_irn_n(irn, 3);
857 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
858 n = ia32_get_irn_n_edges(in);
859 is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */
861 load = get_Proj_pred(in);
864 /* 8bit Loads are not supported, they cannot be used with every register */
865 if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16)
868 /* If there is a data dependency of other irn from load: cannot use AM */
869 if (is_cand && get_nodes_block(other) == block) {
870 other = skip_Proj(other);
871 is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
872 /* this could happen in loops */
873 is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
877 cand = is_cand ? IA32_AM_CAND_LEFT : IA32_AM_CAND_NONE;
881 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
882 n = ia32_get_irn_n_edges(in);
883 is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */
885 load = get_Proj_pred(in);
888 /* 8bit Loads are not supported, they cannot be used with every register */
889 if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16)
892 /* If there is a data dependency of other irn from load: cannot use load */
893 if (is_cand && get_nodes_block(other) == block) {
894 other = skip_Proj(other);
895 is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
896 /* this could happen in loops */
897 is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
901 cand = is_cand ? (cand | IA32_AM_CAND_RIGHT) : cand;
903 /* check some special cases */
904 if (USE_SSE2(cg) && is_ia32_Conv_I2FP(irn)) {
905 /* SSE Conv I -> FP cvtsi2s(s|d) can only load 32 bit values */
906 if (get_mode_size_bits(get_ia32_tgt_mode(irn)) != 32)
907 cand = IA32_AM_CAND_NONE;
909 else if (is_ia32_Conv_I2I(irn)) {
910 /* we cannot load an N bit value and implicitly convert it into an M bit value if N > M */
911 if (get_mode_size_bits(get_ia32_src_mode(irn)) > get_mode_size_bits(get_ia32_tgt_mode(irn)))
912 cand = IA32_AM_CAND_NONE;
915 /* if the irn has a frame entity: we do not use address mode */
916 return get_ia32_frame_ent(irn) ? IA32_AM_CAND_NONE : cand;
920 * Compares the base and index addr and the load/store entities
921 * and returns 1 if they are equal.
923 static int load_store_addr_is_equal(const ir_node *load, const ir_node *store,
924 const ir_node *addr_b, const ir_node *addr_i)
926 int is_equal = (addr_b == get_irn_n(load, 0)) && (addr_i == get_irn_n(load, 1));
927 entity *lent = get_ia32_frame_ent(load);
928 entity *sent = get_ia32_frame_ent(store);
929 ident *lid = get_ia32_am_sc(load);
930 ident *sid = get_ia32_am_sc(store);
931 char *loffs = get_ia32_am_offs(load);
932 char *soffs = get_ia32_am_offs(store);
934 /* are both entities set and equal? */
935 if (is_equal && (lent || sent))
936 is_equal = lent && sent && (lent == sent);
938 /* are address mode idents set and equal? */
939 if (is_equal && (lid || sid))
940 is_equal = lid && sid && (lid == sid);
942 /* are offsets set and equal */
943 if (is_equal && (loffs || soffs))
944 is_equal = loffs && soffs && strcmp(loffs, soffs) == 0;
946 /* are the load and the store of the same mode? */
947 is_equal = is_equal ? get_ia32_ls_mode(load) == get_ia32_ls_mode(store) : 0;
952 typedef enum _ia32_take_lea_attr {
953 IA32_LEA_ATTR_NONE = 0,
954 IA32_LEA_ATTR_BASE = (1 << 0),
955 IA32_LEA_ATTR_INDEX = (1 << 1),
956 IA32_LEA_ATTR_OFFS = (1 << 2),
957 IA32_LEA_ATTR_SCALE = (1 << 3),
958 IA32_LEA_ATTR_AMSC = (1 << 4),
959 IA32_LEA_ATTR_FENT = (1 << 5)
960 } ia32_take_lea_attr;
963 * Decides if we have to keep the LEA operand or if we can assimilate it.
965 static int do_new_lea(ir_node *irn, ir_node *base, ir_node *index, ir_node *lea,
966 int have_am_sc, ia32_code_gen_t *cg)
968 entity *irn_ent = get_ia32_frame_ent(irn);
969 entity *lea_ent = get_ia32_frame_ent(lea);
971 int is_noreg_base = be_is_NoReg(cg, base);
972 int is_noreg_index = be_is_NoReg(cg, index);
973 ia32_am_flavour_t am_flav = get_ia32_am_flavour(lea);
975 /* If the Add and the LEA both have a different frame entity set: keep */
976 if (irn_ent && lea_ent && (irn_ent != lea_ent))
977 return IA32_LEA_ATTR_NONE;
978 else if (! irn_ent && lea_ent)
979 ret_val |= IA32_LEA_ATTR_FENT;
981 /* If the Add and the LEA both have already an address mode symconst: keep */
982 if (have_am_sc && get_ia32_am_sc(lea))
983 return IA32_LEA_ATTR_NONE;
984 else if (get_ia32_am_sc(lea))
985 ret_val |= IA32_LEA_ATTR_AMSC;
987 /* Check the different base-index combinations */
989 if (! is_noreg_base && ! is_noreg_index) {
990 /* Assimilate if base is the lea and the LEA is just a Base + Offset calculation */
991 if ((base == lea) && ! (am_flav & ia32_I ? 1 : 0)) {
992 if (am_flav & ia32_O)
993 ret_val |= IA32_LEA_ATTR_OFFS;
995 ret_val |= IA32_LEA_ATTR_BASE;
998 return IA32_LEA_ATTR_NONE;
1000 else if (! is_noreg_base && is_noreg_index) {
1001 /* Base is set but index not */
1003 /* Base points to LEA: assimilate everything */
1004 if (am_flav & ia32_O)
1005 ret_val |= IA32_LEA_ATTR_OFFS;
1006 if (am_flav & ia32_S)
1007 ret_val |= IA32_LEA_ATTR_SCALE;
1008 if (am_flav & ia32_I)
1009 ret_val |= IA32_LEA_ATTR_INDEX;
1011 ret_val |= IA32_LEA_ATTR_BASE;
1013 else if (am_flav & ia32_B ? 0 : 1) {
1014 /* Base is not the LEA but the LEA is an index only calculation: assimilate */
1015 if (am_flav & ia32_O)
1016 ret_val |= IA32_LEA_ATTR_OFFS;
1017 if (am_flav & ia32_S)
1018 ret_val |= IA32_LEA_ATTR_SCALE;
1020 ret_val |= IA32_LEA_ATTR_INDEX;
1023 return IA32_LEA_ATTR_NONE;
1025 else if (is_noreg_base && ! is_noreg_index) {
1026 /* Index is set but not base */
1028 /* Index points to LEA: assimilate everything */
1029 if (am_flav & ia32_O)
1030 ret_val |= IA32_LEA_ATTR_OFFS;
1031 if (am_flav & ia32_S)
1032 ret_val |= IA32_LEA_ATTR_SCALE;
1033 if (am_flav & ia32_B)
1034 ret_val |= IA32_LEA_ATTR_BASE;
1036 ret_val |= IA32_LEA_ATTR_INDEX;
1038 else if (am_flav & ia32_I ? 0 : 1) {
1039 /* Index is not the LEA but the LEA is a base only calculation: assimilate */
1040 if (am_flav & ia32_O)
1041 ret_val |= IA32_LEA_ATTR_OFFS;
1042 if (am_flav & ia32_S)
1043 ret_val |= IA32_LEA_ATTR_SCALE;
1045 ret_val |= IA32_LEA_ATTR_BASE;
1048 return IA32_LEA_ATTR_NONE;
1051 assert(0 && "There must have been set base or index");
1058 * Adds res before irn into schedule if irn was scheduled.
1059 * @param irn The schedule point
1060 * @param res The node to be scheduled
1062 static INLINE void try_add_to_sched(ir_node *irn, ir_node *res) {
1063 if (sched_is_scheduled(irn))
1064 sched_add_before(irn, res);
1068 * Removes irn from schedule if it was scheduled. If irn is a mode_T node
1069 * all it's Projs are removed as well.
1070 * @param irn The irn to be removed from schedule
1072 static INLINE void try_remove_from_sched(ir_node *irn) {
1073 if (sched_is_scheduled(irn)) {
1074 if (get_irn_mode(irn) == mode_T) {
1075 const ir_edge_t *edge;
1076 foreach_out_edge(irn, edge) {
1077 ir_node *proj = get_edge_src_irn(edge);
1078 if (sched_is_scheduled(proj))
1087 * Folds Add or Sub to LEA if possible
1089 static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
1090 ir_graph *irg = get_irn_irg(irn);
1091 dbg_info *dbg = get_irn_dbg_info(irn);
1092 ir_node *block = get_nodes_block(irn);
1094 ir_node *shift = NULL;
1095 ir_node *lea_o = NULL;
1096 ir_node *lea = NULL;
1098 const char *offs_cnst = NULL;
1099 char *offs_lea = NULL;
1105 ident *am_sc = NULL;
1106 entity *lea_ent = NULL;
1107 ir_node *left, *right, *temp;
1108 ir_node *base, *index;
1109 int consumed_left_shift;
1110 ia32_am_flavour_t am_flav;
1111 DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
1113 if (is_ia32_Add(irn))
1116 left = get_irn_n(irn, 2);
1117 right = get_irn_n(irn, 3);
1119 /* "normalize" arguments in case of add with two operands */
1120 if (isadd && ! be_is_NoReg(cg, right)) {
1121 /* put LEA == ia32_am_O as right operand */
1122 if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) {
1123 set_irn_n(irn, 2, right);
1124 set_irn_n(irn, 3, left);
1130 /* put LEA != ia32_am_O as left operand */
1131 if (is_ia32_Lea(right) && get_ia32_am_flavour(right) != ia32_am_O) {
1132 set_irn_n(irn, 2, right);
1133 set_irn_n(irn, 3, left);
1139 /* put SHL as left operand iff left is NOT a LEA */
1140 if (! is_ia32_Lea(left) && pred_is_specific_node(right, is_ia32_Shl)) {
1141 set_irn_n(irn, 2, right);
1142 set_irn_n(irn, 3, left);
1155 /* check for operation with immediate */
1156 if (is_ia32_ImmConst(irn)) {
1157 DBG((mod, LEVEL_1, "\tfound op with imm const"));
1159 offs_cnst = get_ia32_cnst(irn);
1162 else if (is_ia32_ImmSymConst(irn)) {
1163 DBG((mod, LEVEL_1, "\tfound op with imm symconst"));
1167 am_sc = get_ia32_id_cnst(irn);
1168 am_sc_sign = is_ia32_am_sc_sign(irn);
1171 /* determine the operand which needs to be checked */
1172 temp = be_is_NoReg(cg, right) ? left : right;
1174 /* check if right operand is AMConst (LEA with ia32_am_O) */
1175 /* but we can only eat it up if there is no other symconst */
1176 /* because the linker won't accept two symconsts */
1177 if (! have_am_sc && is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) {
1178 DBG((mod, LEVEL_1, "\tgot op with LEA am_O"));
1180 offs_lea = get_ia32_am_offs(temp);
1181 am_sc = get_ia32_am_sc(temp);
1182 am_sc_sign = is_ia32_am_sc_sign(temp);
1189 else if (temp == right)
1194 /* default for add -> make right operand to index */
1197 consumed_left_shift = -1;
1199 DBG((mod, LEVEL_1, "\tgot LEA candidate with index %+F\n", index));
1201 /* determine the operand which needs to be checked */
1203 if (is_ia32_Lea(left)) {
1205 consumed_left_shift = 0;
1208 /* check for SHL 1,2,3 */
1209 if (pred_is_specific_node(temp, is_ia32_Shl)) {
1210 temp = get_Proj_pred(temp);
1213 if (get_ia32_Immop_tarval(temp)) {
1214 scale = get_tarval_long(get_ia32_Immop_tarval(temp));
1217 index = get_irn_n(temp, 2);
1218 consumed_left_shift = consumed_left_shift < 0 ? 1 : 0;
1220 DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index));
1230 if (! be_is_NoReg(cg, index)) {
1231 /* if we have index, but left == right -> no base */
1232 if (left == right) {
1235 else if (consumed_left_shift == 1) {
1236 /* -> base is right operand */
1237 base = (right == lea_o) ? noreg : right;
1242 /* Try to assimilate a LEA as left operand */
1243 if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
1244 /* check if we can assimilate the LEA */
1245 int take_attr = do_new_lea(irn, base, index, left, have_am_sc, cg);
1247 if (take_attr == IA32_LEA_ATTR_NONE) {
1248 DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n"));
1251 DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
1252 lea = left; /* for statistics */
1254 if (take_attr & IA32_LEA_ATTR_OFFS)
1255 offs = get_ia32_am_offs(left);
1257 if (take_attr & IA32_LEA_ATTR_AMSC) {
1258 am_sc = get_ia32_am_sc(left);
1260 am_sc_sign = is_ia32_am_sc_sign(left);
1263 if (take_attr & IA32_LEA_ATTR_SCALE)
1264 scale = get_ia32_am_scale(left);
1266 if (take_attr & IA32_LEA_ATTR_BASE)
1267 base = get_irn_n(left, 0);
1269 if (take_attr & IA32_LEA_ATTR_INDEX)
1270 index = get_irn_n(left, 1);
1272 if (take_attr & IA32_LEA_ATTR_FENT)
1273 lea_ent = get_ia32_frame_ent(left);
1277 /* ok, we can create a new LEA */
1279 res = new_rd_ia32_Lea(dbg, irg, block, base, index, mode_Is);
1281 /* add the old offset of a previous LEA */
1283 add_ia32_am_offs(res, offs);
1286 /* add the new offset */
1289 add_ia32_am_offs(res, offs_cnst);
1292 add_ia32_am_offs(res, offs_lea);
1296 /* either lea_O-cnst, -cnst or -lea_O */
1299 add_ia32_am_offs(res, offs_lea);
1302 sub_ia32_am_offs(res, offs_cnst);
1305 sub_ia32_am_offs(res, offs_lea);
1309 /* set the address mode symconst */
1311 set_ia32_am_sc(res, am_sc);
1313 set_ia32_am_sc_sign(res);
1316 /* copy the frame entity (could be set in case of Add */
1317 /* which was a FrameAddr) */
1319 set_ia32_frame_ent(res, lea_ent);
1321 set_ia32_frame_ent(res, get_ia32_frame_ent(irn));
1323 if (get_ia32_frame_ent(res))
1324 set_ia32_use_frame(res);
1327 set_ia32_am_scale(res, scale);
1329 am_flav = ia32_am_N;
1330 /* determine new am flavour */
1331 if (offs || offs_cnst || offs_lea || have_am_sc) {
1334 if (! be_is_NoReg(cg, base)) {
1337 if (! be_is_NoReg(cg, index)) {
1343 set_ia32_am_flavour(res, am_flav);
1345 set_ia32_op_type(res, ia32_AddrModeS);
1347 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn));
1349 DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res)));
1351 /* we will exchange it, report here before the Proj is created */
1352 if (shift && lea && lea_o) {
1353 try_remove_from_sched(shift);
1354 try_remove_from_sched(lea);
1355 try_remove_from_sched(lea_o);
1356 DBG_OPT_LEA4(irn, lea_o, lea, shift, res);
1358 else if (shift && lea) {
1359 try_remove_from_sched(shift);
1360 try_remove_from_sched(lea);
1361 DBG_OPT_LEA3(irn, lea, shift, res);
1363 else if (shift && lea_o) {
1364 try_remove_from_sched(shift);
1365 try_remove_from_sched(lea_o);
1366 DBG_OPT_LEA3(irn, lea_o, shift, res);
1368 else if (lea && lea_o) {
1369 try_remove_from_sched(lea);
1370 try_remove_from_sched(lea_o);
1371 DBG_OPT_LEA3(irn, lea_o, lea, res);
1374 try_remove_from_sched(shift);
1375 DBG_OPT_LEA2(irn, shift, res);
1378 try_remove_from_sched(lea);
1379 DBG_OPT_LEA2(irn, lea, res);
1382 try_remove_from_sched(lea_o);
1383 DBG_OPT_LEA2(irn, lea_o, res);
1386 DBG_OPT_LEA1(irn, res);
1388 /* get the result Proj of the Add/Sub */
1389 try_add_to_sched(irn, res);
1390 try_remove_from_sched(irn);
1391 irn = ia32_get_res_proj(irn);
1393 assert(irn && "Couldn't find result proj");
1395 /* exchange the old op with the new LEA */
1404 * Merges a Load/Store node with a LEA.
1405 * @param irn The Load/Store node
1406 * @param lea The LEA
1408 static void merge_loadstore_lea(ir_node *irn, ir_node *lea) {
1409 entity *irn_ent = get_ia32_frame_ent(irn);
1410 entity *lea_ent = get_ia32_frame_ent(lea);
1412 /* If the irn and the LEA both have a different frame entity set: do not merge */
1413 if (irn_ent && lea_ent && (irn_ent != lea_ent))
1415 else if (! irn_ent && lea_ent) {
1416 set_ia32_frame_ent(irn, lea_ent);
1417 set_ia32_use_frame(irn);
1420 /* get the AM attributes from the LEA */
1421 add_ia32_am_offs(irn, get_ia32_am_offs(lea));
1422 set_ia32_am_scale(irn, get_ia32_am_scale(lea));
1423 set_ia32_am_flavour(irn, get_ia32_am_flavour(lea));
1425 set_ia32_am_sc(irn, get_ia32_am_sc(lea));
1426 if (is_ia32_am_sc_sign(lea))
1427 set_ia32_am_sc_sign(irn);
1429 set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
1431 /* set base and index */
1432 set_irn_n(irn, 0, get_irn_n(lea, 0));
1433 set_irn_n(irn, 1, get_irn_n(lea, 1));
1435 try_remove_from_sched(lea);
1437 /* clear remat flag */
1438 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1440 if (is_ia32_Ld(irn))
1441 DBG_OPT_LOAD_LEA(lea, irn);
1443 DBG_OPT_STORE_LEA(lea, irn);
1448 * Sets new_right index of irn to right and new_left index to left.
1449 * Also exchange left and right
1451 static void exchange_left_right(ir_node *irn, ir_node **left, ir_node **right, int new_left, int new_right) {
1454 set_irn_n(irn, new_right, *right);
1455 set_irn_n(irn, new_left, *left);
1461 /* this is only needed for Compares, but currently ALL nodes
1462 * have this attribute :-) */
1463 set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
1467 * Performs address calculation optimization (create LEAs if possible)
1469 static void optimize_lea(ir_node *irn, void *env) {
1470 ia32_code_gen_t *cg = env;
1471 ir_node *block, *noreg_gp, *left, *right;
1473 if (! is_ia32_irn(irn))
1476 /* Following cases can occur: */
1477 /* - Sub (l, imm) -> LEA [base - offset] */
1478 /* - Sub (l, r == LEA with ia32_am_O) -> LEA [base - offset] */
1479 /* - Add (l, imm) -> LEA [base + offset] */
1480 /* - Add (l, r == LEA with ia32_am_O) -> LEA [base + offset] */
1481 /* - Add (l == LEA with ia32_am_O, r) -> LEA [base + offset] */
1482 /* - Add (l, r) -> LEA [base + index * scale] */
1483 /* with scale > 1 iff l/r == shl (1,2,3) */
1485 if (is_ia32_Sub(irn) || is_ia32_Add(irn)) {
1486 left = get_irn_n(irn, 2);
1487 right = get_irn_n(irn, 3);
1488 block = get_nodes_block(irn);
1489 noreg_gp = ia32_new_NoReg_gp(cg);
1491 /* Do not try to create a LEA if one of the operands is a Load. */
1492 /* check is irn is a candidate for address calculation */
1493 if (is_addr_candidate(block, irn)) {
1496 DBG((cg->mod, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn));
1497 res = fold_addr(cg, irn, noreg_gp);
1500 DB((cg->mod, LEVEL_1, "transformed into %+F\n", res));
1502 DB((cg->mod, LEVEL_1, "not transformed\n"));
1505 else if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) {
1506 /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */
1507 /* - Store -> LEA into Store } it might be better to keep the LEA */
1508 left = get_irn_n(irn, 0);
1510 if (is_ia32_Lea(left)) {
1511 const ir_edge_t *edge, *ne;
1514 /* merge all Loads/Stores connected to this LEA with the LEA */
1515 foreach_out_edge_safe(left, edge, ne) {
1516 src = get_edge_src_irn(edge);
1518 if (src && (get_edge_src_pos(edge) == 0) && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) {
1519 DBG((cg->mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
1520 if (! is_ia32_got_lea(src))
1521 merge_loadstore_lea(src, left);
1522 set_ia32_got_lea(src);
1531 * Checks for address mode patterns and performs the
1532 * necessary transformations.
1533 * This function is called by a walker.
1535 static void optimize_am(ir_node *irn, void *env) {
1536 ia32_am_opt_env_t *am_opt_env = env;
1537 ia32_code_gen_t *cg = am_opt_env->cg;
1538 heights_t *h = am_opt_env->h;
1539 ir_node *block, *noreg_gp, *noreg_fp;
1540 ir_node *left, *right;
1541 ir_node *store, *load, *mem_proj;
1542 ir_node *succ, *addr_b, *addr_i;
1543 int check_am_src = 0;
1544 int need_exchange_on_fail = 0;
1545 DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
1547 if (! is_ia32_irn(irn) || is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn))
1550 block = get_nodes_block(irn);
1551 noreg_gp = ia32_new_NoReg_gp(cg);
1552 noreg_fp = ia32_new_NoReg_fp(cg);
1554 DBG((mod, LEVEL_1, "checking for AM\n"));
1556 /* fold following patterns: */
1557 /* - op -> Load into AMop with am_Source */
1559 /* - op is am_Source capable AND */
1560 /* - the Load is only used by this op AND */
1561 /* - the Load is in the same block */
1562 /* - Store -> op -> Load into AMop with am_Dest */
1564 /* - op is am_Dest capable AND */
1565 /* - the Store uses the same address as the Load AND */
1566 /* - the Load is only used by this op AND */
1567 /* - the Load and Store are in the same block AND */
1568 /* - nobody else uses the result of the op */
1570 if ((get_ia32_am_support(irn) != ia32_am_None) && ! is_ia32_Lea(irn)) {
1571 ia32_am_cand_t cand = is_am_candidate(cg, h, block, irn);
1572 ia32_am_cand_t orig_cand = cand;
1574 /* cand == 1: load is left; cand == 2: load is right; */
1576 if (cand == IA32_AM_CAND_NONE)
1579 DBG((mod, LEVEL_1, "\tfound address mode candidate %+F ... ", irn));
1581 left = get_irn_n(irn, 2);
1582 if (get_irn_arity(irn) == 4) {
1583 /* it's an "unary" operation */
1587 right = get_irn_n(irn, 3);
1590 /* normalize commutative ops */
1591 if (node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_RIGHT)) {
1593 /* Assure that left operand is always a Load if there is one */
1594 /* because non-commutative ops can only use Dest AM if the left */
1595 /* operand is a load, so we only need to check left operand. */
1597 exchange_left_right(irn, &left, &right, 3, 2);
1598 need_exchange_on_fail = 1;
1600 /* now: load is right */
1601 cand = IA32_AM_CAND_LEFT;
1604 /* check for Store -> op -> Load */
1606 /* Store -> op -> Load optimization is only possible if supported by op */
1607 /* and if right operand is a Load */
1608 if ((get_ia32_am_support(irn) & ia32_am_Dest) && (cand & IA32_AM_CAND_LEFT))
1610 /* An address mode capable op always has a result Proj. */
1611 /* If this Proj is used by more than one other node, we don't need to */
1612 /* check further, otherwise we check for Store and remember the address, */
1613 /* the Store points to. */
1615 succ = ia32_get_res_proj(irn);
1616 assert(succ && "Couldn't find result proj");
1622 /* now check for users and Store */
1623 if (ia32_get_irn_n_edges(succ) == 1) {
1624 succ = get_edge_src_irn(get_irn_out_edge_first(succ));
1626 if (is_ia32_xStore(succ) || is_ia32_Store(succ)) {
1628 addr_b = get_irn_n(store, 0);
1629 addr_i = get_irn_n(store, 1);
1634 /* we found a Store as single user: Now check for Load */
1636 /* skip the Proj for easier access */
1637 load = is_Proj(right) ? (is_ia32_Load(get_Proj_pred(right)) ? get_Proj_pred(right) : NULL) : NULL;
1639 /* Extra check for commutative ops with two Loads */
1640 /* -> put the interesting Load left */
1641 if (load && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_BOTH)) {
1642 if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
1643 /* We exchange left and right, so it's easier to kill */
1644 /* the correct Load later and to handle unary operations. */
1645 exchange_left_right(irn, &left, &right, 3, 2);
1646 need_exchange_on_fail ^= 1;
1650 /* skip the Proj for easier access */
1651 load = get_Proj_pred(left);
1653 /* Compare Load and Store address */
1654 if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
1655 /* Left Load is from same address, so we can */
1656 /* disconnect the Load and Store here */
1658 /* set new base, index and attributes */
1659 set_irn_n(irn, 0, addr_b);
1660 set_irn_n(irn, 1, addr_i);
1661 add_ia32_am_offs(irn, get_ia32_am_offs(load));
1662 set_ia32_am_scale(irn, get_ia32_am_scale(load));
1663 set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
1664 set_ia32_op_type(irn, ia32_AddrModeD);
1665 set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
1666 set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
1668 set_ia32_am_sc(irn, get_ia32_am_sc(load));
1669 if (is_ia32_am_sc_sign(load))
1670 set_ia32_am_sc_sign(irn);
1672 if (is_ia32_use_frame(load))
1673 set_ia32_use_frame(irn);
1675 /* connect to Load memory and disconnect Load */
1676 if (get_irn_arity(irn) == 5) {
1678 set_irn_n(irn, 4, get_irn_n(load, 2));
1679 set_irn_n(irn, 2, noreg_gp);
1683 set_irn_n(irn, 3, get_irn_n(load, 2));
1684 set_irn_n(irn, 2, noreg_gp);
1687 /* connect the memory Proj of the Store to the op */
1688 mem_proj = ia32_get_proj_for_mode(store, mode_M);
1689 set_Proj_pred(mem_proj, irn);
1690 set_Proj_proj(mem_proj, 1);
1692 /* clear remat flag */
1693 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1695 try_remove_from_sched(load);
1696 try_remove_from_sched(store);
1697 DBG_OPT_AM_D(load, store, irn);
1699 DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
1701 need_exchange_on_fail = 0;
1704 else if (get_ia32_am_support(irn) & ia32_am_Source) {
1705 /* There was no store, check if we still can optimize for source address mode */
1708 } /* if (support AM Dest) */
1709 else if (get_ia32_am_support(irn) & ia32_am_Source) {
1710 /* op doesn't support am AM Dest -> check for AM Source */
1714 /* was exchanged but optimize failed: exchange back */
1715 if (need_exchange_on_fail) {
1716 exchange_left_right(irn, &left, &right, 3, 2);
1720 need_exchange_on_fail = 0;
1722 /* normalize commutative ops */
1723 if (check_am_src && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_LEFT)) {
1725 /* Assure that right operand is always a Load if there is one */
1726 /* because non-commutative ops can only use Source AM if the */
1727 /* right operand is a Load, so we only need to check the right */
1728 /* operand afterwards. */
1730 exchange_left_right(irn, &left, &right, 3, 2);
1731 need_exchange_on_fail = 1;
1733 /* now: load is left */
1734 cand = IA32_AM_CAND_RIGHT;
1737 /* optimize op -> Load iff Load is only used by this op */
1738 /* and right operand is a Load which only used by this irn */
1740 (cand & IA32_AM_CAND_RIGHT) &&
1741 (get_irn_arity(irn) == 5) &&
1742 (ia32_get_irn_n_edges(right) == 1))
1744 right = get_Proj_pred(right);
1746 addr_b = get_irn_n(right, 0);
1747 addr_i = get_irn_n(right, 1);
1749 /* set new base, index and attributes */
1750 set_irn_n(irn, 0, addr_b);
1751 set_irn_n(irn, 1, addr_i);
1752 add_ia32_am_offs(irn, get_ia32_am_offs(right));
1753 set_ia32_am_scale(irn, get_ia32_am_scale(right));
1754 set_ia32_am_flavour(irn, get_ia32_am_flavour(right));
1755 set_ia32_op_type(irn, ia32_AddrModeS);
1756 set_ia32_frame_ent(irn, get_ia32_frame_ent(right));
1757 set_ia32_ls_mode(irn, get_ia32_ls_mode(right));
1759 set_ia32_am_sc(irn, get_ia32_am_sc(right));
1760 if (is_ia32_am_sc_sign(right))
1761 set_ia32_am_sc_sign(irn);
1763 /* clear remat flag */
1764 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1766 if (is_ia32_use_frame(right))
1767 set_ia32_use_frame(irn);
1769 /* connect to Load memory */
1770 set_irn_n(irn, 4, get_irn_n(right, 2));
1772 /* this is only needed for Compares, but currently ALL nodes
1773 * have this attribute :-) */
1774 set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
1776 /* disconnect from Load */
1777 set_irn_n(irn, 3, noreg_gp);
1779 DBG_OPT_AM_S(right, irn);
1781 /* If Load has a memory Proj, connect it to the op */
1782 mem_proj = ia32_get_proj_for_mode(right, mode_M);
1784 set_Proj_pred(mem_proj, irn);
1785 set_Proj_proj(mem_proj, 1);
1788 try_remove_from_sched(right);
1790 DB((mod, LEVEL_1, "merged with %+F into source AM\n", right));
1793 /* was exchanged but optimize failed: exchange back */
1794 if (need_exchange_on_fail)
1795 exchange_left_right(irn, &left, &right, 3, 2);
1801 * Performs address mode optimization.
1803 void ia32_optimize_addressmode(ia32_code_gen_t *cg) {
1804 /* if we are supposed to do AM or LEA optimization: recalculate edges */
1805 if (cg->opt & (IA32_OPT_DOAM | IA32_OPT_LEA)) {
1806 edges_deactivate(cg->irg);
1807 edges_activate(cg->irg);
1810 /* no optimizations at all */
1814 /* beware: we cannot optimize LEA and AM in one run because */
1815 /* LEA optimization adds new nodes to the irg which */
1816 /* invalidates the phase data */
1818 if (cg->opt & IA32_OPT_LEA) {
1819 irg_walk_blkwise_graph(cg->irg, NULL, optimize_lea, cg);
1823 be_dump(cg->irg, "-lea", dump_ir_block_graph_sched);
1825 if (cg->opt & IA32_OPT_DOAM) {
1826 /* we need height information for am optimization */
1827 heights_t *h = heights_new(cg->irg);
1828 ia32_am_opt_env_t env;
1833 irg_walk_blkwise_graph(cg->irg, NULL, optimize_am, &env);