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"
38 typedef struct _ia32_place_env_t {
44 IA32_AM_CAND_NONE = 0,
45 IA32_AM_CAND_LEFT = 1,
46 IA32_AM_CAND_RIGHT = 2,
51 #define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
53 typedef int is_op_func_t(const ir_node *n);
54 typedef ir_node *load_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, ir_node *mem);
57 * checks if a node represents the NOREG value
59 static int be_is_NoReg(ia32_code_gen_t *cg, const ir_node *irn) {
60 be_abi_irg_t *babi = cg->birg->abi;
61 const arch_register_t *fp_noreg = USE_SSE2(cg) ?
62 &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG];
64 return (be_abi_get_callee_save_irn(babi, &ia32_gp_regs[REG_GP_NOREG]) == irn) ||
65 (be_abi_get_callee_save_irn(babi, fp_noreg) == irn);
70 /*************************************************
73 * | | ___ _ __ ___| |_ __ _ _ __ | |_ ___
74 * | | / _ \| '_ \/ __| __/ _` | '_ \| __/ __|
75 * | |___| (_) | | | \__ \ || (_| | | | | |_\__ \
76 * \_____\___/|_| |_|___/\__\__,_|_| |_|\__|___/
78 *************************************************/
81 * creates a unique ident by adding a number to a tag
83 * @param tag the tag string, must contain a %d if a number
86 static ident *unique_id(const char *tag)
88 static unsigned id = 0;
91 snprintf(str, sizeof(str), tag, ++id);
92 return new_id_from_str(str);
96 * Transforms a SymConst.
98 * @param mod the debug module
99 * @param block the block the new node should belong to
100 * @param node the ir SymConst node
101 * @param mode mode of the SymConst
102 * @return the created ia32 Const node
104 static ir_node *gen_SymConst(ia32_transform_env_t *env) {
105 dbg_info *dbg = env->dbg;
106 ir_mode *mode = env->mode;
107 ir_graph *irg = env->irg;
108 ir_node *block = env->block;
111 if (mode_is_float(mode)) {
113 if (USE_SSE2(env->cg))
114 cnst = new_rd_ia32_xConst(dbg, irg, block, get_irg_no_mem(irg), mode);
116 cnst = new_rd_ia32_vfConst(dbg, irg, block, get_irg_no_mem(irg), mode);
119 cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), mode);
121 set_ia32_Const_attr(cnst, env->irn);
127 * Get a primitive type for a mode.
129 static ir_type *get_prim_type(pmap *types, ir_mode *mode)
131 pmap_entry *e = pmap_find(types, mode);
136 snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
137 res = new_type_primitive(new_id_from_str(buf), mode);
138 pmap_insert(types, mode, res);
146 * Get an entity that is initialized with a tarval
148 static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
150 tarval *tv = get_Const_tarval(cnst);
151 pmap_entry *e = pmap_find(cg->isa->tv_ent, tv);
156 ir_mode *mode = get_irn_mode(cnst);
157 ir_type *tp = get_Const_type(cnst);
158 if (tp == firm_unknown_type)
159 tp = get_prim_type(cg->isa->types, mode);
161 res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
163 set_entity_ld_ident(res, get_entity_ident(res));
164 set_entity_visibility(res, visibility_local);
165 set_entity_variability(res, variability_constant);
166 set_entity_allocation(res, allocation_static);
168 /* we create a new entity here: It's initialization must resist on the
170 rem = current_ir_graph;
171 current_ir_graph = get_const_code_irg();
172 set_atomic_ent_value(res, new_Const_type(tv, tp));
173 current_ir_graph = rem;
175 pmap_insert(cg->isa->tv_ent, tv, res);
183 * Transforms a Const.
185 * @param mod the debug module
186 * @param block the block the new node should belong to
187 * @param node the ir Const node
188 * @param mode mode of the Const
189 * @return the created ia32 Const node
191 static ir_node *gen_Const(ia32_transform_env_t *env) {
192 ir_node *cnst, *load;
194 ir_graph *irg = env->irg;
195 ir_node *block = env->block;
196 ir_node *node = env->irn;
197 dbg_info *dbg = env->dbg;
198 ir_mode *mode = env->mode;
200 if (mode_is_float(mode)) {
202 if (! USE_SSE2(env->cg)) {
203 cnst_classify_t clss = classify_Const(node);
205 if (clss == CNST_NULL)
206 return new_rd_ia32_vfldz(dbg, irg, block, mode);
207 else if (clss == CNST_ONE)
208 return new_rd_ia32_vfld1(dbg, irg, block, mode);
210 sym.entity_p = get_entity_for_tv(env->cg, node);
213 cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent);
214 load = new_r_Load(irg, block, get_irg_no_mem(irg), cnst, mode);
215 load = new_r_Proj(irg, block, load, mode, pn_Load_res);
218 cnst = gen_SymConst(env);
219 set_Load_ptr(get_Proj_pred(load), cnst);
223 cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), get_irn_mode(node));
224 set_ia32_Const_attr(cnst, node);
230 * Transforms (all) Const's into ia32_Const and places them in the
231 * block where they are used (or in the cfg-pred Block in case of Phi's).
232 * Additionally all reference nodes are changed into mode_Is nodes.
233 * NOTE: irn must be a firm constant!
235 static void ia32_transform_const(ir_node *irn, void *env) {
236 ia32_code_gen_t *cg = env;
237 ir_node *cnst = NULL;
238 ia32_transform_env_t tenv;
242 tenv.mode = get_irn_mode(irn);
243 tenv.dbg = get_irn_dbg_info(irn);
245 DEBUG_ONLY(tenv.mod = cg->mod;)
247 /* place const either in the smallest dominator of all its users or the original block */
248 if (cg->opt & IA32_OPT_PLACECNST)
249 tenv.block = node_users_smallest_common_dominator(irn, 1);
251 tenv.block = get_nodes_block(irn);
253 switch (get_irn_opcode(irn)) {
255 cnst = gen_Const(&tenv);
258 cnst = gen_SymConst(&tenv);
261 assert(0 && "Wrong usage of ia32_transform_const!");
264 assert(cnst && "Could not create ia32 Const");
266 /* set the new ia32 const */
271 * Transform all firm consts and assure, we visit each const only once.
273 static void ia32_place_consts_walker(ir_node *irn, void *env) {
274 ia32_place_env_t *penv = env;
275 opcode opc = get_irn_opcode(irn);
277 /* transform only firm consts which are not already visited */
278 if ((opc != iro_Const && opc != iro_SymConst) || bitset_is_set(penv->visited, get_irn_idx(irn)))
281 /* mark const visited */
282 bitset_set(penv->visited, get_irn_idx(irn));
284 ia32_transform_const(irn, penv->cg);
288 * Replace reference modes with mode_Iu and preserve store value modes.
290 static void ia32_set_modes(ir_node *irn, void *env) {
294 /* transform all reference nodes into mode_Iu nodes */
295 if (mode_is_reference(get_irn_mode(irn))) {
296 set_irn_mode(irn, mode_Iu);
300 Annotate mode of stored value to link field of the Store
301 as floating point converts might be optimized and we would
304 if (get_irn_opcode(irn) == iro_Store) {
305 set_irn_link(irn, get_irn_mode(get_Store_value(irn)));
310 * Walks over the graph, transforms all firm consts into ia32 consts
311 * and places them into the "best" block.
312 * @param cg The ia32 codegenerator object
314 static void ia32_transform_all_firm_consts(ia32_code_gen_t *cg) {
315 ia32_place_env_t penv;
318 penv.visited = bitset_irg_malloc(cg->irg);
319 irg_walk_graph(cg->irg, NULL, ia32_place_consts_walker, &penv);
320 bitset_free(penv.visited);
323 /* Place all consts and change pointer arithmetics into unsigned integer arithmetics. */
324 void ia32_pre_transform_phase(ia32_code_gen_t *cg) {
326 We need to transform the consts twice:
327 - the psi condition tree transformer needs existing constants to be ia32 constants
328 - the psi condition tree transformer inserts new firm constants which need to be transformed
330 ia32_transform_all_firm_consts(cg);
331 irg_walk_graph(cg->irg, ia32_set_modes, ia32_transform_psi_cond_tree, cg);
332 ia32_transform_all_firm_consts(cg);
335 /********************************************************************************************************
336 * _____ _ _ ____ _ _ _ _ _
337 * | __ \ | | | | / __ \ | | (_) (_) | | (_)
338 * | |__) |__ ___ _ __ | |__ ___ | | ___ | | | |_ __ | |_ _ _ __ ___ _ ______ _| |_ _ ___ _ __
339 * | ___/ _ \/ _ \ '_ \| '_ \ / _ \| |/ _ \ | | | | '_ \| __| | '_ ` _ \| |_ / _` | __| |/ _ \| '_ \
340 * | | | __/ __/ |_) | | | | (_) | | __/ | |__| | |_) | |_| | | | | | | |/ / (_| | |_| | (_) | | | |
341 * |_| \___|\___| .__/|_| |_|\___/|_|\___| \____/| .__/ \__|_|_| |_| |_|_/___\__,_|\__|_|\___/|_| |_|
344 ********************************************************************************************************/
347 * NOTE: THESE PEEPHOLE OPTIMIZATIONS MUST BE CALLED AFTER SCHEDULING AND REGISTER ALLOCATION.
350 static int ia32_cnst_compare(ir_node *n1, ir_node *n2) {
351 return get_ia32_id_cnst(n1) == get_ia32_id_cnst(n2);
355 * Checks for potential CJmp/CJmpAM optimization candidates.
357 static ir_node *ia32_determine_cjmp_cand(ir_node *irn, is_op_func_t *is_op_func) {
358 ir_node *cand = NULL;
359 ir_node *prev = sched_prev(irn);
361 if (is_Block(prev)) {
362 if (get_Block_n_cfgpreds(prev) == 1)
363 prev = get_Block_cfgpred(prev, 0);
368 /* The predecessor must be a ProjX. */
369 if (prev && is_Proj(prev) && get_irn_mode(prev) == mode_X) {
370 prev = get_Proj_pred(prev);
372 if (is_op_func(prev))
379 static int is_TestJmp_cand(const ir_node *irn) {
380 return is_ia32_TestJmp(irn) || is_ia32_And(irn);
384 * Checks if two consecutive arguments of cand matches
385 * the two arguments of irn (TestJmp).
387 static int is_TestJmp_replacement(ir_node *cand, ir_node *irn) {
388 ir_node *in1 = get_irn_n(irn, 0);
389 ir_node *in2 = get_irn_n(irn, 1);
390 int i, n = get_irn_arity(cand);
393 for (i = 0; i < n - 1; i++) {
394 if (get_irn_n(cand, i) == in1 &&
395 get_irn_n(cand, i + 1) == in2)
403 return ia32_cnst_compare(cand, irn);
409 * Tries to replace a TestJmp by a CJmp or CJmpAM (in case of And)
411 static void ia32_optimize_TestJmp(ir_node *irn, ia32_code_gen_t *cg) {
412 ir_node *cand = ia32_determine_cjmp_cand(irn, is_TestJmp_cand);
415 /* we found a possible candidate */
416 replace = cand ? is_TestJmp_replacement(cand, irn) : 0;
419 DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
421 if (is_ia32_And(cand))
422 set_irn_op(irn, op_ia32_CJmpAM);
424 set_irn_op(irn, op_ia32_CJmp);
426 DB((cg->mod, LEVEL_1, "%+F\n", irn));
430 static int is_CondJmp_cand(const ir_node *irn) {
431 return is_ia32_CondJmp(irn) || is_ia32_Sub(irn);
435 * Checks if the arguments of cand are the same of irn.
437 static int is_CondJmp_replacement(ir_node *cand, ir_node *irn) {
438 int i, n = get_irn_arity(cand);
441 for (i = 0; i < n; i++) {
442 if (get_irn_n(cand, i) != get_irn_n(irn, i)) {
449 return ia32_cnst_compare(cand, irn);
455 * Tries to replace a CondJmp by a CJmpAM
457 static void ia32_optimize_CondJmp(ir_node *irn, ia32_code_gen_t *cg) {
458 ir_node *cand = ia32_determine_cjmp_cand(irn, is_CondJmp_cand);
461 /* we found a possible candidate */
462 replace = cand ? is_CondJmp_replacement(cand, irn) : 0;
465 DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
468 set_irn_op(irn, op_ia32_CJmpAM);
470 DB((cg->mod, LEVEL_1, "%+F\n", irn));
475 * Creates a Push from Store(IncSP(gp_reg_size))
477 static void ia32_create_Push(ir_node *irn, ia32_code_gen_t *cg) {
478 ir_node *sp = get_irn_n(irn, 0);
479 ir_graph *irg = cg->irg;
480 ir_node *val, *next, *push, *bl, *proj_M, *proj_res, *old_proj_M, *mem;
481 const ir_edge_t *edge;
484 /* do not create push if store has already an offset assigned or base is not a IncSP */
485 if (get_ia32_am_offs(irn) || ! be_is_IncSP(sp))
488 /* do not create push if index is not NOREG */
489 if (arch_get_irn_register(cg->arch_env, get_irn_n(irn, 1)) !=
490 &ia32_gp_regs[REG_GP_NOREG])
493 /* do not create push for floating point */
494 val = get_irn_n(irn, 2);
495 if (mode_is_float(get_irn_mode(val)))
498 /* do not create push if IncSp doesn't expand stack or expand size is different from register size */
499 if (be_get_IncSP_direction(sp) != be_stack_dir_expand ||
500 be_get_IncSP_offset(sp) != get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_gp].mode))
503 /* do not create push, if there is a path (inside the block) from the push value to IncSP */
504 h = heights_new(cg->irg);
505 if (get_nodes_block(val) == get_nodes_block(sp) &&
506 heights_reachable_in_block(h, val, sp))
513 /* ok, translate into Push */
514 edge = get_irn_out_edge_first(irn);
515 old_proj_M = get_edge_src_irn(edge);
516 bl = get_nodes_block(irn);
518 next = sched_next(irn);
524 if the IncSP points to NoMem -> just use the memory input from store
525 if IncSP points to somewhere else -> sync memory of IncSP and Store
527 mem = be_get_IncSP_mem(sp);
528 if (mem == get_irg_no_mem(irg))
529 mem = get_irn_n(irn, 3);
534 in[1] = get_irn_n(irn, 3);
535 mem = new_r_Sync(irg, bl, 2, in);
538 push = new_rd_ia32_Push(NULL, irg, bl, be_get_IncSP_pred(sp), val, mem);
539 proj_res = new_r_Proj(irg, bl, push, get_irn_mode(sp), pn_ia32_Push_stack);
540 proj_M = new_r_Proj(irg, bl, push, mode_M, pn_ia32_Push_M);
542 /* copy a possible constant from the store */
543 set_ia32_id_cnst(push, get_ia32_id_cnst(irn));
544 set_ia32_immop_type(push, get_ia32_immop_type(irn));
546 /* the push must have SP out register */
547 arch_set_irn_register(cg->arch_env, push, arch_get_irn_register(cg->arch_env, sp));
549 exchange(old_proj_M, proj_M);
550 exchange(sp, proj_res);
551 sched_add_before(next, push);
552 sched_add_after(push, proj_res);
556 * Creates a Pop from IncSP(Load(sp))
558 static void ia32_create_Pop(ir_node *irn, ia32_code_gen_t *cg) {
559 ir_node *old_proj_M = be_get_IncSP_mem(irn);
560 ir_node *load = skip_Proj(old_proj_M);
561 ir_node *old_proj_res = NULL;
562 ir_node *bl, *pop, *next, *proj_res, *proj_sp, *proj_M;
563 const ir_edge_t *edge;
564 const arch_register_t *reg, *sp;
566 if (! is_ia32_Load(load) || get_ia32_am_offs(load))
569 if (arch_get_irn_register(cg->arch_env, get_irn_n(load, 1)) !=
570 &ia32_gp_regs[REG_GP_NOREG])
572 if (arch_get_irn_register(cg->arch_env, get_irn_n(load, 0)) != cg->isa->arch_isa.sp)
575 /* ok, translate into pop */
576 foreach_out_edge(load, edge) {
577 ir_node *succ = get_edge_src_irn(edge);
578 if (succ != old_proj_M) {
583 if (! old_proj_res) {
585 return; /* should not happen */
588 bl = get_nodes_block(load);
590 /* IncSP is typically scheduled after the load, so remove it first */
592 next = sched_next(old_proj_res);
593 sched_remove(old_proj_res);
596 reg = arch_get_irn_register(cg->arch_env, load);
597 sp = arch_get_irn_register(cg->arch_env, irn);
599 pop = new_rd_ia32_Pop(NULL, current_ir_graph, bl, get_irn_n(irn, 0), get_irn_n(load, 2));
600 proj_res = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(old_proj_res), pn_ia32_Pop_res);
601 proj_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(irn), pn_ia32_Pop_stack);
602 proj_M = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
604 exchange(old_proj_M, proj_M);
605 exchange(old_proj_res, proj_res);
606 exchange(irn, proj_sp);
608 arch_set_irn_register(cg->arch_env, proj_res, reg);
609 arch_set_irn_register(cg->arch_env, proj_sp, sp);
611 sched_add_before(next, proj_sp);
612 sched_add_before(proj_sp, proj_res);
613 sched_add_before(proj_res,pop);
617 * Tries to optimize two following IncSP.
619 static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) {
620 ir_node *prev = be_get_IncSP_pred(irn);
621 int real_uses = get_irn_n_edges(prev);
623 if (be_is_IncSP(prev) && real_uses == 1) {
624 /* first IncSP has only one IncSP user, kill the first one */
625 unsigned prev_offs = be_get_IncSP_offset(prev);
626 be_stack_dir_t prev_dir = be_get_IncSP_direction(prev);
627 unsigned curr_offs = be_get_IncSP_offset(irn);
628 be_stack_dir_t curr_dir = be_get_IncSP_direction(irn);
630 int new_ofs = prev_offs * (prev_dir == be_stack_dir_expand ? -1 : +1) +
631 curr_offs * (curr_dir == be_stack_dir_expand ? -1 : +1);
635 curr_dir = be_stack_dir_expand;
638 curr_dir = be_stack_dir_shrink;
639 be_set_IncSP_offset(prev, 0);
640 be_set_IncSP_offset(irn, (unsigned)new_ofs);
641 be_set_IncSP_direction(irn, curr_dir);
643 /* Omit the optimized IncSP */
644 be_set_IncSP_pred(irn, be_get_IncSP_pred(prev));
649 * Performs Peephole Optimizations.
651 void ia32_peephole_optimization(ir_node *irn, void *env) {
652 ia32_code_gen_t *cg = env;
654 /* AMD CPUs want explicit compare before conditional jump */
655 if (! ARCH_AMD(cg->opt_arch)) {
656 if (is_ia32_TestJmp(irn))
657 ia32_optimize_TestJmp(irn, cg);
658 else if (is_ia32_CondJmp(irn))
659 ia32_optimize_CondJmp(irn, cg);
661 /* seems to be buggy when using Pushes */
662 // else if (be_is_IncSP(irn))
663 // ia32_optimize_IncSP(irn, cg);
664 else if (is_ia32_Store(irn))
665 ia32_create_Push(irn, cg);
670 /******************************************************************
672 * /\ | | | | | \/ | | |
673 * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___
674 * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \
675 * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/
676 * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___|
678 ******************************************************************/
685 static int node_is_ia32_comm(const ir_node *irn) {
686 return is_ia32_irn(irn) ? is_ia32_commutative(irn) : 0;
689 static int ia32_get_irn_n_edges(const ir_node *irn) {
690 const ir_edge_t *edge;
693 foreach_out_edge(irn, edge) {
701 * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
703 * @param pred The node to be checked
704 * @param is_op_func The check-function
705 * @return 1 if conditions are fulfilled, 0 otherwise
707 static int pred_is_specific_node(const ir_node *pred, is_op_func_t *is_op_func) {
708 if (is_Proj(pred) && is_op_func(get_Proj_pred(pred))) {
716 * Determines if pred is a Proj and if is_op_func returns true for it's predecessor
717 * and if the predecessor is in block bl.
719 * @param bl The block
720 * @param pred The node to be checked
721 * @param is_op_func The check-function
722 * @return 1 if conditions are fulfilled, 0 otherwise
724 static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred,
725 int (*is_op_func)(const ir_node *n))
728 pred = get_Proj_pred(pred);
729 if ((bl == get_nodes_block(pred)) && is_op_func(pred)) {
738 * Checks if irn is a candidate for address calculation.
740 * - none of the operand must be a Load within the same block OR
741 * - all Loads must have more than one user OR
742 * - the irn has a frame entity (it's a former FrameAddr)
744 * @param block The block the Loads must/mustnot be in
745 * @param irn The irn to check
746 * return 1 if irn is a candidate, 0 otherwise
748 static int is_addr_candidate(const ir_node *block, const ir_node *irn) {
749 ir_node *in, *left, *right;
752 left = get_irn_n(irn, 2);
753 right = get_irn_n(irn, 3);
757 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
758 n = ia32_get_irn_n_edges(in);
759 is_cand = (n == 1) ? 0 : is_cand; /* load with only one user: don't create LEA */
764 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
765 n = ia32_get_irn_n_edges(in);
766 is_cand = (n == 1) ? 0 : is_cand; /* load with only one user: don't create LEA */
769 is_cand = get_ia32_frame_ent(irn) ? 1 : is_cand;
775 * Checks if irn is a candidate for address mode.
778 * - at least one operand has to be a Load within the same block AND
779 * - the load must not have other users than the irn AND
780 * - the irn must not have a frame entity set
782 * @param cg The ia32 code generator
783 * @param h The height information of the irg
784 * @param block The block the Loads must/mustnot be in
785 * @param irn The irn to check
786 * return 0 if irn is no candidate, 1 if left load can be used, 2 if right one, 3 for both
788 static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const ir_node *block, ir_node *irn) {
789 ir_node *in, *load, *other, *left, *right;
790 int n, is_cand = 0, cand;
792 if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn))
795 left = get_irn_n(irn, 2);
796 right = get_irn_n(irn, 3);
800 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
801 n = ia32_get_irn_n_edges(in);
802 is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */
804 load = get_Proj_pred(in);
807 /* If there is a data dependency of other irn from load: cannot use AM */
808 if (get_nodes_block(other) == block) {
809 other = skip_Proj(other);
810 is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
811 /* this could happen in loops */
812 is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
816 cand = is_cand ? IA32_AM_CAND_LEFT : IA32_AM_CAND_NONE;
820 if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
821 n = ia32_get_irn_n_edges(in);
822 is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */
824 load = get_Proj_pred(in);
827 /* If there is a data dependency of other irn from load: cannot use load */
828 if (get_nodes_block(other) == block) {
829 other = skip_Proj(other);
830 is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
831 /* this could happen in loops */
832 is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
836 cand = is_cand ? (cand | IA32_AM_CAND_RIGHT) : cand;
838 /* check some special cases */
839 if (USE_SSE2(cg) && is_ia32_Conv_I2FP(irn)) {
840 /* SSE Conv I -> FP cvtsi2s(s|d) can only load 32 bit values */
841 if (get_mode_size_bits(get_ia32_tgt_mode(irn)) != 32)
842 cand = IA32_AM_CAND_NONE;
844 else if (is_ia32_Conv_I2I(irn)) {
845 /* we cannot load an N bit value and implicitly convert it into an M bit value if N > M */
846 if (get_mode_size_bits(get_ia32_src_mode(irn)) > get_mode_size_bits(get_ia32_tgt_mode(irn)))
847 cand = IA32_AM_CAND_NONE;
850 /* if the irn has a frame entity: we do not use address mode */
851 return get_ia32_frame_ent(irn) ? IA32_AM_CAND_NONE : cand;
855 * Compares the base and index addr and the load/store entities
856 * and returns 1 if they are equal.
858 static int load_store_addr_is_equal(const ir_node *load, const ir_node *store,
859 const ir_node *addr_b, const ir_node *addr_i)
861 int is_equal = (addr_b == get_irn_n(load, 0)) && (addr_i == get_irn_n(load, 1));
862 entity *lent = get_ia32_frame_ent(load);
863 entity *sent = get_ia32_frame_ent(store);
864 ident *lid = get_ia32_am_sc(load);
865 ident *sid = get_ia32_am_sc(store);
866 char *loffs = get_ia32_am_offs(load);
867 char *soffs = get_ia32_am_offs(store);
869 /* are both entities set and equal? */
870 if (is_equal && (lent || sent))
871 is_equal = lent && sent && (lent == sent);
873 /* are address mode idents set and equal? */
874 if (is_equal && (lid || sid))
875 is_equal = lid && sid && (lid == sid);
877 /* are offsets set and equal */
878 if (is_equal && (loffs || soffs))
879 is_equal = loffs && soffs && strcmp(loffs, soffs) == 0;
881 /* are the load and the store of the same mode? */
882 is_equal = is_equal ? get_ia32_ls_mode(load) == get_ia32_ls_mode(store) : 0;
887 typedef enum _ia32_take_lea_attr {
888 IA32_LEA_ATTR_NONE = 0,
889 IA32_LEA_ATTR_BASE = (1 << 0),
890 IA32_LEA_ATTR_INDEX = (1 << 1),
891 IA32_LEA_ATTR_OFFS = (1 << 2),
892 IA32_LEA_ATTR_SCALE = (1 << 3),
893 IA32_LEA_ATTR_AMSC = (1 << 4),
894 IA32_LEA_ATTR_FENT = (1 << 5)
895 } ia32_take_lea_attr;
898 * Decides if we have to keep the LEA operand or if we can assimilate it.
900 static int do_new_lea(ir_node *irn, ir_node *base, ir_node *index, ir_node *lea,
901 int have_am_sc, ia32_code_gen_t *cg)
903 entity *irn_ent = get_ia32_frame_ent(irn);
904 entity *lea_ent = get_ia32_frame_ent(lea);
906 int is_noreg_base = be_is_NoReg(cg, base);
907 int is_noreg_index = be_is_NoReg(cg, index);
908 ia32_am_flavour_t am_flav = get_ia32_am_flavour(lea);
910 /* If the Add and the LEA both have a different frame entity set: keep */
911 if (irn_ent && lea_ent && (irn_ent != lea_ent))
912 return IA32_LEA_ATTR_NONE;
913 else if (! irn_ent && lea_ent)
914 ret_val |= IA32_LEA_ATTR_FENT;
916 /* If the Add and the LEA both have already an address mode symconst: keep */
917 if (have_am_sc && get_ia32_am_sc(lea))
918 return IA32_LEA_ATTR_NONE;
919 else if (get_ia32_am_sc(lea))
920 ret_val |= IA32_LEA_ATTR_AMSC;
922 /* Check the different base-index combinations */
924 if (! is_noreg_base && ! is_noreg_index) {
925 /* Assimilate if base is the lea and the LEA is just a Base + Offset calculation */
926 if ((base == lea) && ! (am_flav & ia32_I ? 1 : 0)) {
927 if (am_flav & ia32_O)
928 ret_val |= IA32_LEA_ATTR_OFFS;
930 ret_val |= IA32_LEA_ATTR_BASE;
933 return IA32_LEA_ATTR_NONE;
935 else if (! is_noreg_base && is_noreg_index) {
936 /* Base is set but index not */
938 /* Base points to LEA: assimilate everything */
939 if (am_flav & ia32_O)
940 ret_val |= IA32_LEA_ATTR_OFFS;
941 if (am_flav & ia32_S)
942 ret_val |= IA32_LEA_ATTR_SCALE;
943 if (am_flav & ia32_I)
944 ret_val |= IA32_LEA_ATTR_INDEX;
946 ret_val |= IA32_LEA_ATTR_BASE;
948 else if (am_flav & ia32_B ? 0 : 1) {
949 /* Base is not the LEA but the LEA is an index only calculation: assimilate */
950 if (am_flav & ia32_O)
951 ret_val |= IA32_LEA_ATTR_OFFS;
952 if (am_flav & ia32_S)
953 ret_val |= IA32_LEA_ATTR_SCALE;
955 ret_val |= IA32_LEA_ATTR_INDEX;
958 return IA32_LEA_ATTR_NONE;
960 else if (is_noreg_base && ! is_noreg_index) {
961 /* Index is set but not base */
963 /* Index points to LEA: assimilate everything */
964 if (am_flav & ia32_O)
965 ret_val |= IA32_LEA_ATTR_OFFS;
966 if (am_flav & ia32_S)
967 ret_val |= IA32_LEA_ATTR_SCALE;
968 if (am_flav & ia32_B)
969 ret_val |= IA32_LEA_ATTR_BASE;
971 ret_val |= IA32_LEA_ATTR_INDEX;
973 else if (am_flav & ia32_I ? 0 : 1) {
974 /* Index is not the LEA but the LEA is a base only calculation: assimilate */
975 if (am_flav & ia32_O)
976 ret_val |= IA32_LEA_ATTR_OFFS;
977 if (am_flav & ia32_S)
978 ret_val |= IA32_LEA_ATTR_SCALE;
980 ret_val |= IA32_LEA_ATTR_BASE;
983 return IA32_LEA_ATTR_NONE;
986 assert(0 && "There must have been set base or index");
994 * Folds Add or Sub to LEA if possible
996 static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
997 ir_graph *irg = get_irn_irg(irn);
998 dbg_info *dbg = get_irn_dbg_info(irn);
999 ir_node *block = get_nodes_block(irn);
1001 ir_node *shift = NULL;
1002 ir_node *lea_o = NULL;
1003 ir_node *lea = NULL;
1005 const char *offs_cnst = NULL;
1006 char *offs_lea = NULL;
1012 ident *am_sc = NULL;
1013 entity *lea_ent = NULL;
1014 ir_node *left, *right, *temp;
1015 ir_node *base, *index;
1016 ia32_am_flavour_t am_flav;
1017 DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
1019 if (is_ia32_Add(irn))
1022 left = get_irn_n(irn, 2);
1023 right = get_irn_n(irn, 3);
1025 /* "normalize" arguments in case of add with two operands */
1026 if (isadd && ! be_is_NoReg(cg, right)) {
1027 /* put LEA == ia32_am_O as right operand */
1028 if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) {
1029 set_irn_n(irn, 2, right);
1030 set_irn_n(irn, 3, left);
1036 /* put LEA != ia32_am_O as left operand */
1037 if (is_ia32_Lea(right) && get_ia32_am_flavour(right) != ia32_am_O) {
1038 set_irn_n(irn, 2, right);
1039 set_irn_n(irn, 3, left);
1045 /* put SHL as left operand iff left is NOT a LEA */
1046 if (! is_ia32_Lea(left) && pred_is_specific_node(right, is_ia32_Shl)) {
1047 set_irn_n(irn, 2, right);
1048 set_irn_n(irn, 3, left);
1061 /* check for operation with immediate */
1062 if (is_ia32_ImmConst(irn)) {
1063 DBG((mod, LEVEL_1, "\tfound op with imm const"));
1065 offs_cnst = get_ia32_cnst(irn);
1068 else if (is_ia32_ImmSymConst(irn)) {
1069 DBG((mod, LEVEL_1, "\tfound op with imm symconst"));
1073 am_sc = get_ia32_id_cnst(irn);
1074 am_sc_sign = is_ia32_am_sc_sign(irn);
1077 /* determine the operand which needs to be checked */
1078 temp = be_is_NoReg(cg, right) ? left : right;
1080 /* check if right operand is AMConst (LEA with ia32_am_O) */
1081 /* but we can only eat it up if there is no other symconst */
1082 /* because the linker won't accept two symconsts */
1083 if (! have_am_sc && is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) {
1084 DBG((mod, LEVEL_1, "\tgot op with LEA am_O"));
1086 offs_lea = get_ia32_am_offs(temp);
1087 am_sc = get_ia32_am_sc(temp);
1088 am_sc_sign = is_ia32_am_sc_sign(temp);
1095 else if (temp == right)
1100 /* default for add -> make right operand to index */
1104 DBG((mod, LEVEL_1, "\tgot LEA candidate with index %+F\n", index));
1106 /* determine the operand which needs to be checked */
1108 if (is_ia32_Lea(left)) {
1112 /* check for SHL 1,2,3 */
1113 if (pred_is_specific_node(temp, is_ia32_Shl)) {
1114 temp = get_Proj_pred(temp);
1117 if (get_ia32_Immop_tarval(temp)) {
1118 scale = get_tarval_long(get_ia32_Immop_tarval(temp));
1121 index = get_irn_n(temp, 2);
1123 DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index));
1133 if (! be_is_NoReg(cg, index)) {
1134 /* if we have index, but left == right -> no base */
1135 if (left == right) {
1138 else if (! is_ia32_Lea(left) && (index != right)) {
1139 /* index != right -> we found a good Shl */
1140 /* left != LEA -> this Shl was the left operand */
1141 /* -> base is right operand */
1142 base = (right == lea_o) ? noreg : right;
1147 /* Try to assimilate a LEA as left operand */
1148 if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
1149 /* check if we can assimilate the LEA */
1150 int take_attr = do_new_lea(irn, base, index, left, have_am_sc, cg);
1152 if (take_attr == IA32_LEA_ATTR_NONE) {
1153 DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n"));
1156 DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
1157 lea = left; /* for statistics */
1159 if (take_attr & IA32_LEA_ATTR_OFFS)
1160 offs = get_ia32_am_offs(left);
1162 if (take_attr & IA32_LEA_ATTR_AMSC) {
1163 am_sc = get_ia32_am_sc(left);
1165 am_sc_sign = is_ia32_am_sc_sign(left);
1168 if (take_attr & IA32_LEA_ATTR_SCALE)
1169 scale = get_ia32_am_scale(left);
1171 if (take_attr & IA32_LEA_ATTR_BASE)
1172 base = get_irn_n(left, 0);
1174 if (take_attr & IA32_LEA_ATTR_INDEX)
1175 index = get_irn_n(left, 1);
1177 if (take_attr & IA32_LEA_ATTR_FENT)
1178 lea_ent = get_ia32_frame_ent(left);
1182 /* ok, we can create a new LEA */
1184 res = new_rd_ia32_Lea(dbg, irg, block, base, index, mode_Is);
1186 /* add the old offset of a previous LEA */
1188 add_ia32_am_offs(res, offs);
1191 /* add the new offset */
1194 add_ia32_am_offs(res, offs_cnst);
1197 add_ia32_am_offs(res, offs_lea);
1201 /* either lea_O-cnst, -cnst or -lea_O */
1204 add_ia32_am_offs(res, offs_lea);
1207 sub_ia32_am_offs(res, offs_cnst);
1210 sub_ia32_am_offs(res, offs_lea);
1214 /* set the address mode symconst */
1216 set_ia32_am_sc(res, am_sc);
1218 set_ia32_am_sc_sign(res);
1221 /* copy the frame entity (could be set in case of Add */
1222 /* which was a FrameAddr) */
1224 set_ia32_frame_ent(res, lea_ent);
1226 set_ia32_frame_ent(res, get_ia32_frame_ent(irn));
1228 if (get_ia32_frame_ent(res))
1229 set_ia32_use_frame(res);
1232 set_ia32_am_scale(res, scale);
1234 am_flav = ia32_am_N;
1235 /* determine new am flavour */
1236 if (offs || offs_cnst || offs_lea || have_am_sc) {
1239 if (! be_is_NoReg(cg, base)) {
1242 if (! be_is_NoReg(cg, index)) {
1248 set_ia32_am_flavour(res, am_flav);
1250 set_ia32_op_type(res, ia32_AddrModeS);
1252 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn));
1254 DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res)));
1256 /* we will exchange it, report here before the Proj is created */
1257 if (shift && lea && lea_o)
1258 DBG_OPT_LEA4(irn, lea_o, lea, shift, res);
1259 else if (shift && lea)
1260 DBG_OPT_LEA3(irn, lea, shift, res);
1261 else if (shift && lea_o)
1262 DBG_OPT_LEA3(irn, lea_o, shift, res);
1263 else if (lea && lea_o)
1264 DBG_OPT_LEA3(irn, lea_o, lea, res);
1266 DBG_OPT_LEA2(irn, shift, res);
1268 DBG_OPT_LEA2(irn, lea, res);
1270 DBG_OPT_LEA2(irn, lea_o, res);
1272 DBG_OPT_LEA1(irn, res);
1274 /* get the result Proj of the Add/Sub */
1275 irn = ia32_get_res_proj(irn);
1277 assert(irn && "Couldn't find result proj");
1279 /* exchange the old op with the new LEA */
1288 * Merges a Load/Store node with a LEA.
1289 * @param irn The Load/Store node
1290 * @param lea The LEA
1292 static void merge_loadstore_lea(ir_node *irn, ir_node *lea) {
1293 entity *irn_ent = get_ia32_frame_ent(irn);
1294 entity *lea_ent = get_ia32_frame_ent(lea);
1296 /* If the irn and the LEA both have a different frame entity set: do not merge */
1297 if (irn_ent && lea_ent && (irn_ent != lea_ent))
1299 else if (! irn_ent && lea_ent) {
1300 set_ia32_frame_ent(irn, lea_ent);
1301 set_ia32_use_frame(irn);
1304 /* get the AM attributes from the LEA */
1305 add_ia32_am_offs(irn, get_ia32_am_offs(lea));
1306 set_ia32_am_scale(irn, get_ia32_am_scale(lea));
1307 set_ia32_am_flavour(irn, get_ia32_am_flavour(lea));
1309 set_ia32_am_sc(irn, get_ia32_am_sc(lea));
1310 if (is_ia32_am_sc_sign(lea))
1311 set_ia32_am_sc_sign(irn);
1313 set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
1315 /* set base and index */
1316 set_irn_n(irn, 0, get_irn_n(lea, 0));
1317 set_irn_n(irn, 1, get_irn_n(lea, 1));
1319 /* clear remat flag */
1320 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1322 if (is_ia32_Ld(irn))
1323 DBG_OPT_LOAD_LEA(lea, irn);
1325 DBG_OPT_STORE_LEA(lea, irn);
1330 * Sets new_right index of irn to right and new_left index to left.
1331 * Also exchange left and right
1333 static void exchange_left_right(ir_node *irn, ir_node **left, ir_node **right, int new_left, int new_right) {
1336 set_irn_n(irn, new_right, *right);
1337 set_irn_n(irn, new_left, *left);
1343 /* this is only needed for Compares, but currently ALL nodes
1344 * have this attribute :-) */
1345 set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
1349 * Performs address calculation optimization (create LEAs if possible)
1351 static void optimize_lea(ir_node *irn, void *env) {
1352 ia32_code_gen_t *cg = env;
1353 ir_node *block, *noreg_gp, *left, *right;
1355 if (! is_ia32_irn(irn))
1358 /* Following cases can occur: */
1359 /* - Sub (l, imm) -> LEA [base - offset] */
1360 /* - Sub (l, r == LEA with ia32_am_O) -> LEA [base - offset] */
1361 /* - Add (l, imm) -> LEA [base + offset] */
1362 /* - Add (l, r == LEA with ia32_am_O) -> LEA [base + offset] */
1363 /* - Add (l == LEA with ia32_am_O, r) -> LEA [base + offset] */
1364 /* - Add (l, r) -> LEA [base + index * scale] */
1365 /* with scale > 1 iff l/r == shl (1,2,3) */
1367 if (is_ia32_Sub(irn) || is_ia32_Add(irn)) {
1368 left = get_irn_n(irn, 2);
1369 right = get_irn_n(irn, 3);
1370 block = get_nodes_block(irn);
1371 noreg_gp = ia32_new_NoReg_gp(cg);
1373 /* Do not try to create a LEA if one of the operands is a Load. */
1374 /* check is irn is a candidate for address calculation */
1375 if (is_addr_candidate(block, irn)) {
1378 DBG((cg->mod, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn));
1379 res = fold_addr(cg, irn, noreg_gp);
1382 DB((cg->mod, LEVEL_1, "transformed into %+F\n", res));
1384 DB((cg->mod, LEVEL_1, "not transformed\n"));
1387 else if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) {
1388 /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */
1389 /* - Store -> LEA into Store } it might be better to keep the LEA */
1390 left = get_irn_n(irn, 0);
1392 if (is_ia32_Lea(left)) {
1393 const ir_edge_t *edge, *ne;
1396 /* merge all Loads/Stores connected to this LEA with the LEA */
1397 foreach_out_edge_safe(left, edge, ne) {
1398 src = get_edge_src_irn(edge);
1400 if (src && (get_edge_src_pos(edge) == 0) && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) {
1401 DBG((cg->mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
1402 if (! is_ia32_got_lea(src))
1403 merge_loadstore_lea(src, left);
1404 set_ia32_got_lea(src);
1413 * Checks for address mode patterns and performs the
1414 * necessary transformations.
1415 * This function is called by a walker.
1417 static void optimize_am(ir_node *irn, void *env) {
1418 ia32_am_opt_env_t *am_opt_env = env;
1419 ia32_code_gen_t *cg = am_opt_env->cg;
1420 heights_t *h = am_opt_env->h;
1421 ir_node *block, *noreg_gp, *noreg_fp;
1422 ir_node *left, *right;
1423 ir_node *store, *load, *mem_proj;
1424 ir_node *succ, *addr_b, *addr_i;
1425 int check_am_src = 0;
1426 int need_exchange_on_fail = 0;
1427 DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
1429 if (! is_ia32_irn(irn) || is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn))
1432 block = get_nodes_block(irn);
1433 noreg_gp = ia32_new_NoReg_gp(cg);
1434 noreg_fp = ia32_new_NoReg_fp(cg);
1436 DBG((mod, LEVEL_1, "checking for AM\n"));
1438 /* fold following patterns: */
1439 /* - op -> Load into AMop with am_Source */
1441 /* - op is am_Source capable AND */
1442 /* - the Load is only used by this op AND */
1443 /* - the Load is in the same block */
1444 /* - Store -> op -> Load into AMop with am_Dest */
1446 /* - op is am_Dest capable AND */
1447 /* - the Store uses the same address as the Load AND */
1448 /* - the Load is only used by this op AND */
1449 /* - the Load and Store are in the same block AND */
1450 /* - nobody else uses the result of the op */
1452 if ((get_ia32_am_support(irn) != ia32_am_None) && ! is_ia32_Lea(irn)) {
1453 ia32_am_cand_t cand = is_am_candidate(cg, h, block, irn);
1454 ia32_am_cand_t orig_cand = cand;
1456 /* cand == 1: load is left; cand == 2: load is right; */
1458 if (cand == IA32_AM_CAND_NONE)
1461 DBG((mod, LEVEL_1, "\tfound address mode candidate %+F ... ", irn));
1463 left = get_irn_n(irn, 2);
1464 if (get_irn_arity(irn) == 4) {
1465 /* it's an "unary" operation */
1469 right = get_irn_n(irn, 3);
1472 /* normalize commutative ops */
1473 if (node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_RIGHT)) {
1475 /* Assure that left operand is always a Load if there is one */
1476 /* because non-commutative ops can only use Dest AM if the left */
1477 /* operand is a load, so we only need to check left operand. */
1479 exchange_left_right(irn, &left, &right, 3, 2);
1480 need_exchange_on_fail = 1;
1482 /* now: load is right */
1483 cand = IA32_AM_CAND_LEFT;
1486 /* check for Store -> op -> Load */
1488 /* Store -> op -> Load optimization is only possible if supported by op */
1489 /* and if right operand is a Load */
1490 if ((get_ia32_am_support(irn) & ia32_am_Dest) && (cand & IA32_AM_CAND_LEFT))
1492 /* An address mode capable op always has a result Proj. */
1493 /* If this Proj is used by more than one other node, we don't need to */
1494 /* check further, otherwise we check for Store and remember the address, */
1495 /* the Store points to. */
1497 succ = ia32_get_res_proj(irn);
1498 assert(succ && "Couldn't find result proj");
1504 /* now check for users and Store */
1505 if (ia32_get_irn_n_edges(succ) == 1) {
1506 succ = get_edge_src_irn(get_irn_out_edge_first(succ));
1508 if (is_ia32_xStore(succ) || is_ia32_Store(succ)) {
1510 addr_b = get_irn_n(store, 0);
1511 addr_i = get_irn_n(store, 1);
1516 /* we found a Store as single user: Now check for Load */
1518 /* skip the Proj for easier access */
1519 load = is_Proj(right) ? (is_ia32_Load(get_Proj_pred(right)) ? get_Proj_pred(right) : NULL) : NULL;
1521 /* Extra check for commutative ops with two Loads */
1522 /* -> put the interesting Load left */
1523 if (load && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_BOTH)) {
1524 if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
1525 /* We exchange left and right, so it's easier to kill */
1526 /* the correct Load later and to handle unary operations. */
1527 exchange_left_right(irn, &left, &right, 3, 2);
1528 need_exchange_on_fail ^= 1;
1532 /* skip the Proj for easier access */
1533 load = get_Proj_pred(left);
1535 /* Compare Load and Store address */
1536 if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
1537 /* Left Load is from same address, so we can */
1538 /* disconnect the Load and Store here */
1540 /* set new base, index and attributes */
1541 set_irn_n(irn, 0, addr_b);
1542 set_irn_n(irn, 1, addr_i);
1543 add_ia32_am_offs(irn, get_ia32_am_offs(load));
1544 set_ia32_am_scale(irn, get_ia32_am_scale(load));
1545 set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
1546 set_ia32_op_type(irn, ia32_AddrModeD);
1547 set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
1548 set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
1550 set_ia32_am_sc(irn, get_ia32_am_sc(load));
1551 if (is_ia32_am_sc_sign(load))
1552 set_ia32_am_sc_sign(irn);
1554 if (is_ia32_use_frame(load))
1555 set_ia32_use_frame(irn);
1557 /* connect to Load memory and disconnect Load */
1558 if (get_irn_arity(irn) == 5) {
1560 set_irn_n(irn, 4, get_irn_n(load, 2));
1561 set_irn_n(irn, 2, noreg_gp);
1565 set_irn_n(irn, 3, get_irn_n(load, 2));
1566 set_irn_n(irn, 2, noreg_gp);
1569 /* connect the memory Proj of the Store to the op */
1570 mem_proj = ia32_get_proj_for_mode(store, mode_M);
1571 set_Proj_pred(mem_proj, irn);
1572 set_Proj_proj(mem_proj, 1);
1574 /* clear remat flag */
1575 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1577 DBG_OPT_AM_D(load, store, irn);
1579 DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
1581 need_exchange_on_fail = 0;
1584 else if (get_ia32_am_support(irn) & ia32_am_Source) {
1585 /* There was no store, check if we still can optimize for source address mode */
1588 } /* if (support AM Dest) */
1589 else if (get_ia32_am_support(irn) & ia32_am_Source) {
1590 /* op doesn't support am AM Dest -> check for AM Source */
1594 /* was exchanged but optimize failed: exchange back */
1595 if (need_exchange_on_fail) {
1596 exchange_left_right(irn, &left, &right, 3, 2);
1600 need_exchange_on_fail = 0;
1602 /* normalize commutative ops */
1603 if (check_am_src && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_LEFT)) {
1605 /* Assure that right operand is always a Load if there is one */
1606 /* because non-commutative ops can only use Source AM if the */
1607 /* right operand is a Load, so we only need to check the right */
1608 /* operand afterwards. */
1610 exchange_left_right(irn, &left, &right, 3, 2);
1611 need_exchange_on_fail = 1;
1613 /* now: load is left */
1614 cand = IA32_AM_CAND_RIGHT;
1617 /* optimize op -> Load iff Load is only used by this op */
1618 /* and right operand is a Load which only used by this irn */
1620 (cand & IA32_AM_CAND_RIGHT) &&
1621 (get_irn_arity(irn) == 5) &&
1622 (ia32_get_irn_n_edges(right) == 1))
1624 right = get_Proj_pred(right);
1626 addr_b = get_irn_n(right, 0);
1627 addr_i = get_irn_n(right, 1);
1629 /* set new base, index and attributes */
1630 set_irn_n(irn, 0, addr_b);
1631 set_irn_n(irn, 1, addr_i);
1632 add_ia32_am_offs(irn, get_ia32_am_offs(right));
1633 set_ia32_am_scale(irn, get_ia32_am_scale(right));
1634 set_ia32_am_flavour(irn, get_ia32_am_flavour(right));
1635 set_ia32_op_type(irn, ia32_AddrModeS);
1636 set_ia32_frame_ent(irn, get_ia32_frame_ent(right));
1637 set_ia32_ls_mode(irn, get_ia32_ls_mode(right));
1639 set_ia32_am_sc(irn, get_ia32_am_sc(right));
1640 if (is_ia32_am_sc_sign(right))
1641 set_ia32_am_sc_sign(irn);
1643 /* clear remat flag */
1644 set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
1646 if (is_ia32_use_frame(right))
1647 set_ia32_use_frame(irn);
1649 /* connect to Load memory */
1650 set_irn_n(irn, 4, get_irn_n(right, 2));
1652 /* this is only needed for Compares, but currently ALL nodes
1653 * have this attribute :-) */
1654 set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
1656 /* disconnect from Load */
1657 set_irn_n(irn, 3, noreg_gp);
1659 DBG_OPT_AM_S(right, irn);
1661 /* If Load has a memory Proj, connect it to the op */
1662 mem_proj = ia32_get_proj_for_mode(right, mode_M);
1664 set_Proj_pred(mem_proj, irn);
1665 set_Proj_proj(mem_proj, 1);
1668 DB((mod, LEVEL_1, "merged with %+F into source AM\n", right));
1671 /* was exchanged but optimize failed: exchange back */
1672 if (need_exchange_on_fail)
1673 exchange_left_right(irn, &left, &right, 3, 2);
1679 * Performs address mode optimization.
1681 void ia32_optimize_addressmode(ia32_code_gen_t *cg) {
1682 /* if we are supposed to do AM or LEA optimization: recalculate edges */
1683 if (cg->opt & (IA32_OPT_DOAM | IA32_OPT_LEA)) {
1684 edges_deactivate(cg->irg);
1685 edges_activate(cg->irg);
1688 /* no optimizations at all */
1692 /* beware: we cannot optimize LEA and AM in one run because */
1693 /* LEA optimization adds new nodes to the irg which */
1694 /* invalidates the phase data */
1696 if (cg->opt & IA32_OPT_LEA) {
1697 irg_walk_blkwise_graph(cg->irg, NULL, optimize_lea, cg);
1701 be_dump(cg->irg, "-lea", dump_ir_block_graph_sched);
1703 if (cg->opt & IA32_OPT_DOAM) {
1704 /* we need height information for am optimization */
1705 heights_t *h = heights_new(cg->irg);
1706 ia32_am_opt_env_t env;
1711 irg_walk_blkwise_graph(cg->irg, NULL, optimize_am, &env);