2 * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief This file implements the IR transformation from firm into
24 * @author Christian Wuerdig, Matthias Braun
35 #include "irgraph_t.h"
40 #include "iredges_t.h"
51 #include "../benode_t.h"
52 #include "../besched.h"
54 #include "../beutil.h"
55 #include "../beirg_t.h"
56 #include "../betranshlp.h"
58 #include "bearch_ia32_t.h"
59 #include "ia32_nodes_attr.h"
60 #include "ia32_transform.h"
61 #include "ia32_new_nodes.h"
62 #include "ia32_map_regs.h"
63 #include "ia32_dbg_stat.h"
64 #include "ia32_optimize.h"
65 #include "ia32_util.h"
67 #include "gen_ia32_regalloc_if.h"
69 #define SFP_SIGN "0x80000000"
70 #define DFP_SIGN "0x8000000000000000"
71 #define SFP_ABS "0x7FFFFFFF"
72 #define DFP_ABS "0x7FFFFFFFFFFFFFFF"
74 #define TP_SFP_SIGN "ia32_sfp_sign"
75 #define TP_DFP_SIGN "ia32_dfp_sign"
76 #define TP_SFP_ABS "ia32_sfp_abs"
77 #define TP_DFP_ABS "ia32_dfp_abs"
79 #define ENT_SFP_SIGN "IA32_SFP_SIGN"
80 #define ENT_DFP_SIGN "IA32_DFP_SIGN"
81 #define ENT_SFP_ABS "IA32_SFP_ABS"
82 #define ENT_DFP_ABS "IA32_DFP_ABS"
84 #define mode_vfp (ia32_reg_classes[CLASS_ia32_vfp].mode)
85 #define mode_xmm (ia32_reg_classes[CLASS_ia32_xmm].mode)
87 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
89 /** hold the current code generator during transformation */
90 static ia32_code_gen_t *env_cg = NULL;
91 static ir_node *initial_fpcw = NULL;
93 extern ir_op *get_op_Mulh(void);
95 typedef ir_node *construct_binop_func(dbg_info *db, ir_graph *irg,
96 ir_node *block, ir_node *base, ir_node *index, ir_node *op1,
97 ir_node *op2, ir_node *mem);
99 typedef ir_node *construct_binop_float_func(dbg_info *db, ir_graph *irg,
100 ir_node *block, ir_node *base, ir_node *index, ir_node *op1,
101 ir_node *op2, ir_node *mem, ir_node *fpcw);
103 typedef ir_node *construct_unop_func(dbg_info *db, ir_graph *irg,
104 ir_node *block, ir_node *base, ir_node *index, ir_node *op,
107 /****************************************************************************************************
109 * | | | | / _| | | (_)
110 * _ __ ___ __| | ___ | |_ _ __ __ _ _ __ ___| |_ ___ _ __ _ __ ___ __ _| |_ _ ___ _ __
111 * | '_ \ / _ \ / _` |/ _ \ | __| '__/ _` | '_ \/ __| _/ _ \| '__| '_ ` _ \ / _` | __| |/ _ \| '_ \
112 * | | | | (_) | (_| | __/ | |_| | | (_| | | | \__ \ || (_) | | | | | | | | (_| | |_| | (_) | | | |
113 * |_| |_|\___/ \__,_|\___| \__|_| \__,_|_| |_|___/_| \___/|_| |_| |_| |_|\__,_|\__|_|\___/|_| |_|
115 ****************************************************************************************************/
117 static ir_node *try_create_Immediate(ir_node *node,
118 char immediate_constraint_type);
120 static ir_node *create_immediate_or_transform(ir_node *node,
121 char immediate_constraint_type);
123 static ir_node *create_I2I_Conv(ir_mode *src_mode, ir_mode *tgt_mode,
124 dbg_info *dbgi, ir_node *new_block,
128 * Return true if a mode can be stored in the GP register set
130 static INLINE int mode_needs_gp_reg(ir_mode *mode) {
131 if(mode == mode_fpcw)
133 return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
137 * Returns 1 if irn is a Const representing 0, 0 otherwise
139 static INLINE int is_ia32_Const_0(ir_node *irn) {
140 return is_ia32_irn(irn) && is_ia32_Const(irn) && get_ia32_immop_type(irn) == ia32_ImmConst
141 && tarval_is_null(get_ia32_Immop_tarval(irn));
145 * Returns 1 if irn is a Const representing 1, 0 otherwise
147 static INLINE int is_ia32_Const_1(ir_node *irn) {
148 return is_ia32_irn(irn) && is_ia32_Const(irn) && get_ia32_immop_type(irn) == ia32_ImmConst
149 && tarval_is_one(get_ia32_Immop_tarval(irn));
153 * Collects all Projs of a node into the node array. Index is the projnum.
154 * BEWARE: The caller has to assure the appropriate array size!
156 static void ia32_collect_Projs(ir_node *irn, ir_node **projs, int size) {
157 const ir_edge_t *edge;
158 assert(get_irn_mode(irn) == mode_T && "need mode_T");
160 memset(projs, 0, size * sizeof(projs[0]));
162 foreach_out_edge(irn, edge) {
163 ir_node *proj = get_edge_src_irn(edge);
164 int proj_proj = get_Proj_proj(proj);
165 assert(proj_proj < size);
166 projs[proj_proj] = proj;
171 * Renumbers the proj having pn_old in the array tp pn_new
172 * and removes the proj from the array.
174 static INLINE void ia32_renumber_Proj(ir_node **projs, long pn_old, long pn_new) {
175 fprintf(stderr, "Warning: renumber_Proj used!\n");
177 set_Proj_proj(projs[pn_old], pn_new);
178 projs[pn_old] = NULL;
183 * creates a unique ident by adding a number to a tag
185 * @param tag the tag string, must contain a %d if a number
188 static ident *unique_id(const char *tag)
190 static unsigned id = 0;
193 snprintf(str, sizeof(str), tag, ++id);
194 return new_id_from_str(str);
198 * Get a primitive type for a mode.
200 static ir_type *get_prim_type(pmap *types, ir_mode *mode)
202 pmap_entry *e = pmap_find(types, mode);
207 snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
208 res = new_type_primitive(new_id_from_str(buf), mode);
209 set_type_alignment_bytes(res, 16);
210 pmap_insert(types, mode, res);
218 * Get an entity that is initialized with a tarval
220 static ir_entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
222 tarval *tv = get_Const_tarval(cnst);
223 pmap_entry *e = pmap_find(cg->isa->tv_ent, tv);
228 ir_mode *mode = get_irn_mode(cnst);
229 ir_type *tp = get_Const_type(cnst);
230 if (tp == firm_unknown_type)
231 tp = get_prim_type(cg->isa->types, mode);
233 res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
235 set_entity_ld_ident(res, get_entity_ident(res));
236 set_entity_visibility(res, visibility_local);
237 set_entity_variability(res, variability_constant);
238 set_entity_allocation(res, allocation_static);
240 /* we create a new entity here: It's initialization must resist on the
242 rem = current_ir_graph;
243 current_ir_graph = get_const_code_irg();
244 set_atomic_ent_value(res, new_Const_type(tv, tp));
245 current_ir_graph = rem;
247 pmap_insert(cg->isa->tv_ent, tv, res);
255 static int is_Const_0(ir_node *node) {
259 return classify_Const(node) == CNST_NULL;
262 static int is_Const_1(ir_node *node) {
266 return classify_Const(node) == CNST_ONE;
270 * Transforms a Const.
272 static ir_node *gen_Const(ir_node *node) {
273 ir_graph *irg = current_ir_graph;
274 ir_node *old_block = get_nodes_block(node);
275 ir_node *block = be_transform_node(old_block);
276 dbg_info *dbgi = get_irn_dbg_info(node);
277 ir_mode *mode = get_irn_mode(node);
279 if (mode_is_float(mode)) {
281 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
282 ir_node *nomem = new_NoMem();
286 if (! USE_SSE2(env_cg)) {
287 cnst_classify_t clss = classify_Const(node);
289 if (clss == CNST_NULL) {
290 load = new_rd_ia32_vfldz(dbgi, irg, block);
292 } else if (clss == CNST_ONE) {
293 load = new_rd_ia32_vfld1(dbgi, irg, block);
296 floatent = get_entity_for_tv(env_cg, node);
298 load = new_rd_ia32_vfld(dbgi, irg, block, noreg, noreg, nomem, mode);
299 set_ia32_op_type(load, ia32_AddrModeS);
300 set_ia32_am_flavour(load, ia32_am_N);
301 set_ia32_am_sc(load, floatent);
302 set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
303 res = new_r_Proj(irg, block, load, mode_vfp, pn_ia32_vfld_res);
305 set_ia32_ls_mode(load, mode);
307 floatent = get_entity_for_tv(env_cg, node);
309 load = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem);
310 set_ia32_op_type(load, ia32_AddrModeS);
311 set_ia32_am_flavour(load, ia32_am_N);
312 set_ia32_am_sc(load, floatent);
313 set_ia32_ls_mode(load, mode);
314 set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
316 res = new_r_Proj(irg, block, load, mode_xmm, pn_ia32_xLoad_res);
319 SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
321 /* Const Nodes before the initial IncSP are a bad idea, because
322 * they could be spilled and we have no SP ready at that point yet.
323 * So add a dependency to the initial frame pointer calculation to
324 * avoid that situation.
326 if (get_irg_start_block(irg) == block) {
327 add_irn_dep(load, get_irg_frame(irg));
330 SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
333 ir_node *cnst = new_rd_ia32_Const(dbgi, irg, block);
336 if (get_irg_start_block(irg) == block) {
337 add_irn_dep(cnst, get_irg_frame(irg));
340 set_ia32_Const_attr(cnst, node);
341 SET_IA32_ORIG_NODE(cnst, ia32_get_old_node_name(env_cg, node));
346 return new_r_Bad(irg);
350 * Transforms a SymConst.
352 static ir_node *gen_SymConst(ir_node *node) {
353 ir_graph *irg = current_ir_graph;
354 ir_node *old_block = get_nodes_block(node);
355 ir_node *block = be_transform_node(old_block);
356 dbg_info *dbgi = get_irn_dbg_info(node);
357 ir_mode *mode = get_irn_mode(node);
360 if (mode_is_float(mode)) {
361 if (USE_SSE2(env_cg))
362 cnst = new_rd_ia32_xConst(dbgi, irg, block);
364 cnst = new_rd_ia32_vfConst(dbgi, irg, block);
365 //set_ia32_ls_mode(cnst, mode);
366 set_ia32_ls_mode(cnst, mode_E);
368 cnst = new_rd_ia32_Const(dbgi, irg, block);
371 /* Const Nodes before the initial IncSP are a bad idea, because
372 * they could be spilled and we have no SP ready at that point yet
374 if (get_irg_start_block(irg) == block) {
375 add_irn_dep(cnst, get_irg_frame(irg));
378 set_ia32_Const_attr(cnst, node);
379 SET_IA32_ORIG_NODE(cnst, ia32_get_old_node_name(env_cg, node));
384 /* Generates an entity for a known FP const (used for FP Neg + Abs) */
385 ir_entity *ia32_gen_fp_known_const(ia32_known_const_t kct) {
386 static const struct {
388 const char *ent_name;
389 const char *cnst_str;
390 } names [ia32_known_const_max] = {
391 { TP_SFP_SIGN, ENT_SFP_SIGN, SFP_SIGN }, /* ia32_SSIGN */
392 { TP_DFP_SIGN, ENT_DFP_SIGN, DFP_SIGN }, /* ia32_DSIGN */
393 { TP_SFP_ABS, ENT_SFP_ABS, SFP_ABS }, /* ia32_SABS */
394 { TP_DFP_ABS, ENT_DFP_ABS, DFP_ABS } /* ia32_DABS */
396 static ir_entity *ent_cache[ia32_known_const_max];
398 const char *tp_name, *ent_name, *cnst_str;
406 ent_name = names[kct].ent_name;
407 if (! ent_cache[kct]) {
408 tp_name = names[kct].tp_name;
409 cnst_str = names[kct].cnst_str;
411 mode = kct == ia32_SSIGN || kct == ia32_SABS ? mode_Iu : mode_Lu;
413 tv = new_tarval_from_str(cnst_str, strlen(cnst_str), mode);
414 tp = new_type_primitive(new_id_from_str(tp_name), mode);
415 ent = new_entity(get_glob_type(), new_id_from_str(ent_name), tp);
417 set_entity_ld_ident(ent, get_entity_ident(ent));
418 set_entity_visibility(ent, visibility_local);
419 set_entity_variability(ent, variability_constant);
420 set_entity_allocation(ent, allocation_static);
422 /* we create a new entity here: It's initialization must resist on the
424 rem = current_ir_graph;
425 current_ir_graph = get_const_code_irg();
426 cnst = new_Const(mode, tv);
427 current_ir_graph = rem;
429 set_atomic_ent_value(ent, cnst);
431 /* cache the entry */
432 ent_cache[kct] = ent;
435 return ent_cache[kct];
440 * Prints the old node name on cg obst and returns a pointer to it.
442 const char *ia32_get_old_node_name(ia32_code_gen_t *cg, ir_node *irn) {
443 ia32_isa_t *isa = (ia32_isa_t *)cg->arch_env->isa;
445 lc_eoprintf(firm_get_arg_env(), isa->name_obst, "%+F", irn);
446 obstack_1grow(isa->name_obst, 0);
447 return obstack_finish(isa->name_obst);
451 /* determine if one operator is an Imm */
452 static ir_node *get_immediate_op(ir_node *op1, ir_node *op2) {
454 return is_ia32_Cnst(op1) ? op1 : (is_ia32_Cnst(op2) ? op2 : NULL);
456 return is_ia32_Cnst(op2) ? op2 : NULL;
460 /* determine if one operator is not an Imm */
461 static ir_node *get_expr_op(ir_node *op1, ir_node *op2) {
462 return !is_ia32_Cnst(op1) ? op1 : (!is_ia32_Cnst(op2) ? op2 : NULL);
465 static void fold_immediate(ir_node *node, int in1, int in2) {
469 if (!(env_cg->opt & IA32_OPT_IMMOPS))
472 left = get_irn_n(node, in1);
473 right = get_irn_n(node, in2);
474 if (! is_ia32_Cnst(right) && is_ia32_Cnst(left)) {
475 /* we can only set right operand to immediate */
476 if(!is_ia32_commutative(node))
478 /* exchange left/right */
479 set_irn_n(node, in1, right);
480 set_irn_n(node, in2, ia32_get_admissible_noreg(env_cg, node, in2));
481 copy_ia32_Immop_attr(node, left);
482 } else if(is_ia32_Cnst(right)) {
483 set_irn_n(node, in2, ia32_get_admissible_noreg(env_cg, node, in2));
484 copy_ia32_Immop_attr(node, right);
489 clear_ia32_commutative(node);
490 set_ia32_am_support(node, get_ia32_am_support(node) & ~ia32_am_Source,
491 get_ia32_am_arity(node));
495 * Construct a standard binary operation, set AM and immediate if required.
497 * @param op1 The first operand
498 * @param op2 The second operand
499 * @param func The node constructor function
500 * @return The constructed ia32 node.
502 static ir_node *gen_binop(ir_node *node, ir_node *op1, ir_node *op2,
503 construct_binop_func *func, int commutative)
505 ir_node *block = be_transform_node(get_nodes_block(node));
506 ir_graph *irg = current_ir_graph;
507 dbg_info *dbgi = get_irn_dbg_info(node);
508 ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
509 ir_node *nomem = new_NoMem();
512 ir_node *new_op1 = be_transform_node(op1);
513 ir_node *new_op2 = create_immediate_or_transform(op2, 0);
514 if (is_ia32_Immediate(new_op2)) {
518 new_node = func(dbgi, irg, block, noreg_gp, noreg_gp, new_op1, new_op2, nomem);
519 if (func == new_rd_ia32_IMul) {
520 set_ia32_am_support(new_node, ia32_am_Source, ia32_am_binary);
522 set_ia32_am_support(new_node, ia32_am_Full, ia32_am_binary);
525 SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
527 set_ia32_commutative(new_node);
534 * Construct a standard binary operation, set AM and immediate if required.
536 * @param op1 The first operand
537 * @param op2 The second operand
538 * @param func The node constructor function
539 * @return The constructed ia32 node.
541 static ir_node *gen_binop_sse_float(ir_node *node, ir_node *op1, ir_node *op2,
542 construct_binop_func *func)
544 ir_node *block = be_transform_node(get_nodes_block(node));
545 ir_node *new_op1 = be_transform_node(op1);
546 ir_node *new_op2 = be_transform_node(op2);
547 ir_node *new_node = NULL;
548 dbg_info *dbgi = get_irn_dbg_info(node);
549 ir_graph *irg = current_ir_graph;
550 ir_mode *mode = get_irn_mode(node);
551 ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
552 ir_node *nomem = new_NoMem();
554 new_node = func(dbgi, irg, block, noreg_gp, noreg_gp, new_op1, new_op2,
556 set_ia32_am_support(new_node, ia32_am_Source, ia32_am_binary);
557 if (is_op_commutative(get_irn_op(node))) {
558 set_ia32_commutative(new_node);
560 set_ia32_ls_mode(new_node, mode);
562 SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
567 static ir_node *get_fpcw(void)
570 if(initial_fpcw != NULL)
573 fpcw = be_abi_get_ignore_irn(env_cg->birg->abi,
574 &ia32_fp_cw_regs[REG_FPCW]);
575 initial_fpcw = be_transform_node(fpcw);
581 * Construct a standard binary operation, set AM and immediate if required.
583 * @param op1 The first operand
584 * @param op2 The second operand
585 * @param func The node constructor function
586 * @return The constructed ia32 node.
588 static ir_node *gen_binop_x87_float(ir_node *node, ir_node *op1, ir_node *op2,
589 construct_binop_float_func *func)
591 ir_node *block = be_transform_node(get_nodes_block(node));
592 ir_node *new_op1 = be_transform_node(op1);
593 ir_node *new_op2 = be_transform_node(op2);
594 ir_node *new_node = NULL;
595 dbg_info *dbgi = get_irn_dbg_info(node);
596 ir_graph *irg = current_ir_graph;
597 ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
598 ir_node *nomem = new_NoMem();
600 new_node = func(dbgi, irg, block, noreg_gp, noreg_gp, new_op1, new_op2,
602 set_ia32_am_support(new_node, ia32_am_Source, ia32_am_binary);
603 if (is_op_commutative(get_irn_op(node))) {
604 set_ia32_commutative(new_node);
607 SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
613 * Construct a shift/rotate binary operation, sets AM and immediate if required.
615 * @param op1 The first operand
616 * @param op2 The second operand
617 * @param func The node constructor function
618 * @return The constructed ia32 node.
620 static ir_node *gen_shift_binop(ir_node *node, ir_node *op1, ir_node *op2,
621 construct_binop_func *func)
623 ir_node *block = be_transform_node(get_nodes_block(node));
624 ir_node *new_op1 = be_transform_node(op1);
626 ir_node *new_op = NULL;
627 dbg_info *dbgi = get_irn_dbg_info(node);
628 ir_graph *irg = current_ir_graph;
629 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
630 ir_node *nomem = new_NoMem();
632 assert(! mode_is_float(get_irn_mode(node))
633 && "Shift/Rotate with float not supported");
635 new_op2 = create_immediate_or_transform(op2, 'N');
637 new_op = func(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
640 set_ia32_am_support(new_op, ia32_am_Dest, ia32_am_binary);
642 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
644 set_ia32_emit_cl(new_op);
646 /* lowered shift instruction may have a dependency operand, handle it here */
647 if (get_irn_arity(node) == 3) {
648 /* we have a dependency */
649 ir_node *new_dep = be_transform_node(get_irn_n(node, 2));
650 add_irn_dep(new_op, new_dep);
658 * Construct a standard unary operation, set AM and immediate if required.
660 * @param op The operand
661 * @param func The node constructor function
662 * @return The constructed ia32 node.
664 static ir_node *gen_unop(ir_node *node, ir_node *op, construct_unop_func *func)
666 ir_node *block = be_transform_node(get_nodes_block(node));
667 ir_node *new_op = be_transform_node(op);
668 ir_node *new_node = NULL;
669 ir_graph *irg = current_ir_graph;
670 dbg_info *dbgi = get_irn_dbg_info(node);
671 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
672 ir_node *nomem = new_NoMem();
674 new_node = func(dbgi, irg, block, noreg, noreg, new_op, nomem);
675 DB((dbg, LEVEL_1, "INT unop ..."));
676 set_ia32_am_support(new_node, ia32_am_Dest, ia32_am_unary);
678 SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
684 * Creates an ia32 Add.
686 * @return the created ia32 Add node
688 static ir_node *gen_Add(ir_node *node) {
689 ir_node *block = be_transform_node(get_nodes_block(node));
690 ir_node *op1 = get_Add_left(node);
691 ir_node *new_op1 = be_transform_node(op1);
692 ir_node *op2 = get_Add_right(node);
693 ir_node *new_op2 = be_transform_node(op2);
694 ir_node *new_op = NULL;
695 ir_graph *irg = current_ir_graph;
696 dbg_info *dbgi = get_irn_dbg_info(node);
697 ir_mode *mode = get_irn_mode(node);
698 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
699 ir_node *nomem = new_NoMem();
700 ir_node *expr_op, *imm_op;
702 /* Check if immediate optimization is on and */
703 /* if it's an operation with immediate. */
704 imm_op = (env_cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(new_op1, new_op2) : NULL;
705 expr_op = get_expr_op(new_op1, new_op2);
707 assert((expr_op || imm_op) && "invalid operands");
709 if (mode_is_float(mode)) {
710 if (USE_SSE2(env_cg))
711 return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xAdd);
713 return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfadd);
718 ia32_immop_type_t tp1 = get_ia32_immop_type(new_op1);
719 ia32_immop_type_t tp2 = get_ia32_immop_type(new_op2);
721 /* No expr_op means, that we have two const - one symconst and */
722 /* one tarval or another symconst - because this case is not */
723 /* covered by constant folding */
724 /* We need to check for: */
725 /* 1) symconst + const -> becomes a LEA */
726 /* 2) symconst + symconst -> becomes a const + LEA as the elf */
727 /* linker doesn't support two symconsts */
729 if (tp1 == ia32_ImmSymConst && tp2 == ia32_ImmSymConst) {
730 /* this is the 2nd case */
731 new_op = new_rd_ia32_Lea(dbgi, irg, block, new_op1, noreg);
732 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op2));
733 set_ia32_am_flavour(new_op, ia32_am_B);
734 set_ia32_op_type(new_op, ia32_AddrModeS);
736 DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
737 } else if (tp1 == ia32_ImmSymConst) {
738 tarval *tv = get_ia32_Immop_tarval(new_op2);
739 long offs = get_tarval_long(tv);
741 new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
742 add_irn_dep(new_op, get_irg_frame(irg));
743 DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
745 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op1));
746 add_ia32_am_offs_int(new_op, offs);
747 set_ia32_am_flavour(new_op, ia32_am_OB);
748 set_ia32_op_type(new_op, ia32_AddrModeS);
749 } else if (tp2 == ia32_ImmSymConst) {
750 tarval *tv = get_ia32_Immop_tarval(new_op1);
751 long offs = get_tarval_long(tv);
753 new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
754 add_irn_dep(new_op, get_irg_frame(irg));
755 DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
757 add_ia32_am_offs_int(new_op, offs);
758 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op2));
759 set_ia32_am_flavour(new_op, ia32_am_OB);
760 set_ia32_op_type(new_op, ia32_AddrModeS);
762 tarval *tv1 = get_ia32_Immop_tarval(new_op1);
763 tarval *tv2 = get_ia32_Immop_tarval(new_op2);
764 tarval *restv = tarval_add(tv1, tv2);
766 DEBUG_ONLY(ir_fprintf(stderr, "Warning: add with 2 consts not folded: %+F\n", node));
768 new_op = new_rd_ia32_Const(dbgi, irg, block);
769 set_ia32_Const_tarval(new_op, restv);
770 DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
773 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
776 if ((env_cg->opt & IA32_OPT_INCDEC) && get_ia32_immop_type(imm_op) == ia32_ImmConst) {
777 tarval_classification_t class_tv, class_negtv;
778 tarval *tv = get_ia32_Immop_tarval(imm_op);
780 /* optimize tarvals */
781 class_tv = classify_tarval(tv);
782 class_negtv = classify_tarval(tarval_neg(tv));
784 if (class_tv == TV_CLASSIFY_ONE) { /* + 1 == INC */
785 DB((dbg, LEVEL_2, "Add(1) to Inc ... "));
786 new_op = new_rd_ia32_Inc(dbgi, irg, block, noreg, noreg, expr_op, nomem);
787 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
789 } else if (class_tv == TV_CLASSIFY_ALL_ONE || class_negtv == TV_CLASSIFY_ONE) { /* + (-1) == DEC */
790 DB((dbg, LEVEL_2, "Add(-1) to Dec ... "));
791 new_op = new_rd_ia32_Dec(dbgi, irg, block, noreg, noreg, expr_op, nomem);
792 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
798 /* This is a normal add */
799 new_op = new_rd_ia32_Add(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
802 set_ia32_am_support(new_op, ia32_am_Full, ia32_am_binary);
803 set_ia32_commutative(new_op);
805 fold_immediate(new_op, 2, 3);
807 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
813 * Creates an ia32 Mul.
815 * @return the created ia32 Mul node
817 static ir_node *gen_Mul(ir_node *node) {
818 ir_node *op1 = get_Mul_left(node);
819 ir_node *op2 = get_Mul_right(node);
820 ir_mode *mode = get_irn_mode(node);
822 if (mode_is_float(mode)) {
823 if (USE_SSE2(env_cg))
824 return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xMul);
826 return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfmul);
830 for the lower 32bit of the result it doesn't matter whether we use
831 signed or unsigned multiplication so we use IMul as it has fewer
834 return gen_binop(node, op1, op2, new_rd_ia32_IMul, 1);
838 * Creates an ia32 Mulh.
839 * Note: Mul produces a 64Bit result and Mulh returns the upper 32 bit of
840 * this result while Mul returns the lower 32 bit.
842 * @return the created ia32 Mulh node
844 static ir_node *gen_Mulh(ir_node *node) {
845 ir_node *block = be_transform_node(get_nodes_block(node));
846 ir_node *op1 = get_irn_n(node, 0);
847 ir_node *new_op1 = be_transform_node(op1);
848 ir_node *op2 = get_irn_n(node, 1);
849 ir_node *new_op2 = be_transform_node(op2);
850 ir_graph *irg = current_ir_graph;
851 dbg_info *dbgi = get_irn_dbg_info(node);
852 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
853 ir_mode *mode = get_irn_mode(node);
854 ir_node *proj_EDX, *res;
856 assert(!mode_is_float(mode) && "Mulh with float not supported");
857 if (mode_is_signed(mode)) {
858 res = new_rd_ia32_IMul1OP(dbgi, irg, block, noreg, noreg, new_op1,
859 new_op2, new_NoMem());
861 res = new_rd_ia32_Mul(dbgi, irg, block, noreg, noreg, new_op1, new_op2,
865 set_ia32_commutative(res);
866 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
868 proj_EDX = new_rd_Proj(dbgi, irg, block, res, mode_Iu, pn_EDX);
876 * Creates an ia32 And.
878 * @return The created ia32 And node
880 static ir_node *gen_And(ir_node *node) {
881 ir_node *op1 = get_And_left(node);
882 ir_node *op2 = get_And_right(node);
883 assert(! mode_is_float(get_irn_mode(node)));
885 /* check for zero extension first */
887 tarval *tv = get_Const_tarval(op2);
888 long v = get_tarval_long(tv);
890 if (v == 0xFF || v == 0xFFFF) {
891 dbg_info *dbgi = get_irn_dbg_info(node);
892 ir_node *block = be_transform_node(get_nodes_block(node));
893 ir_node *new_op = be_transform_node(op1);
903 res = create_I2I_Conv(src_mode, mode_Iu, dbgi, block, new_op);
904 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
910 return gen_binop(node, op1, op2, new_rd_ia32_And, 1);
916 * Creates an ia32 Or.
918 * @return The created ia32 Or node
920 static ir_node *gen_Or(ir_node *node) {
921 ir_node *op1 = get_Or_left(node);
922 ir_node *op2 = get_Or_right(node);
924 assert (! mode_is_float(get_irn_mode(node)));
925 return gen_binop(node, op1, op2, new_rd_ia32_Or, 1);
931 * Creates an ia32 Eor.
933 * @return The created ia32 Eor node
935 static ir_node *gen_Eor(ir_node *node) {
936 ir_node *op1 = get_Eor_left(node);
937 ir_node *op2 = get_Eor_right(node);
939 assert(! mode_is_float(get_irn_mode(node)));
940 return gen_binop(node, op1, op2, new_rd_ia32_Xor, 1);
945 * Creates an ia32 Sub.
947 * @return The created ia32 Sub node
949 static ir_node *gen_Sub(ir_node *node) {
950 ir_node *block = be_transform_node(get_nodes_block(node));
951 ir_node *op1 = get_Sub_left(node);
952 ir_node *new_op1 = be_transform_node(op1);
953 ir_node *op2 = get_Sub_right(node);
954 ir_node *new_op2 = be_transform_node(op2);
955 ir_node *new_op = NULL;
956 ir_graph *irg = current_ir_graph;
957 dbg_info *dbgi = get_irn_dbg_info(node);
958 ir_mode *mode = get_irn_mode(node);
959 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
960 ir_node *nomem = new_NoMem();
961 ir_node *expr_op, *imm_op;
963 /* Check if immediate optimization is on and */
964 /* if it's an operation with immediate. */
965 imm_op = (env_cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, new_op2) : NULL;
966 expr_op = get_expr_op(new_op1, new_op2);
968 assert((expr_op || imm_op) && "invalid operands");
970 if (mode_is_float(mode)) {
971 if (USE_SSE2(env_cg))
972 return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xSub);
974 return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfsub);
979 ia32_immop_type_t tp1 = get_ia32_immop_type(new_op1);
980 ia32_immop_type_t tp2 = get_ia32_immop_type(new_op2);
982 /* No expr_op means, that we have two const - one symconst and */
983 /* one tarval or another symconst - because this case is not */
984 /* covered by constant folding */
985 /* We need to check for: */
986 /* 1) symconst - const -> becomes a LEA */
987 /* 2) symconst - symconst -> becomes a const - LEA as the elf */
988 /* linker doesn't support two symconsts */
989 if (tp1 == ia32_ImmSymConst && tp2 == ia32_ImmSymConst) {
990 /* this is the 2nd case */
991 new_op = new_rd_ia32_Lea(dbgi, irg, block, new_op1, noreg);
992 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(op2));
993 set_ia32_am_sc_sign(new_op);
994 set_ia32_am_flavour(new_op, ia32_am_B);
996 DBG_OPT_LEA3(op1, op2, node, new_op);
997 } else if (tp1 == ia32_ImmSymConst) {
998 tarval *tv = get_ia32_Immop_tarval(new_op2);
999 long offs = get_tarval_long(tv);
1001 new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
1002 add_irn_dep(new_op, get_irg_frame(irg));
1003 DBG_OPT_LEA3(op1, op2, node, new_op);
1005 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op1));
1006 add_ia32_am_offs_int(new_op, -offs);
1007 set_ia32_am_flavour(new_op, ia32_am_OB);
1008 set_ia32_op_type(new_op, ia32_AddrModeS);
1009 } else if (tp2 == ia32_ImmSymConst) {
1010 tarval *tv = get_ia32_Immop_tarval(new_op1);
1011 long offs = get_tarval_long(tv);
1013 new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
1014 add_irn_dep(new_op, get_irg_frame(irg));
1015 DBG_OPT_LEA3(op1, op2, node, new_op);
1017 add_ia32_am_offs_int(new_op, offs);
1018 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op2));
1019 set_ia32_am_sc_sign(new_op);
1020 set_ia32_am_flavour(new_op, ia32_am_OB);
1021 set_ia32_op_type(new_op, ia32_AddrModeS);
1023 tarval *tv1 = get_ia32_Immop_tarval(new_op1);
1024 tarval *tv2 = get_ia32_Immop_tarval(new_op2);
1025 tarval *restv = tarval_sub(tv1, tv2);
1027 DEBUG_ONLY(ir_fprintf(stderr, "Warning: sub with 2 consts not folded: %+F\n", node));
1029 new_op = new_rd_ia32_Const(dbgi, irg, block);
1030 set_ia32_Const_tarval(new_op, restv);
1031 DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
1034 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1036 } else if (imm_op) {
1037 if ((env_cg->opt & IA32_OPT_INCDEC) && get_ia32_immop_type(imm_op) == ia32_ImmConst) {
1038 tarval_classification_t class_tv, class_negtv;
1039 tarval *tv = get_ia32_Immop_tarval(imm_op);
1041 /* optimize tarvals */
1042 class_tv = classify_tarval(tv);
1043 class_negtv = classify_tarval(tarval_neg(tv));
1045 if (class_tv == TV_CLASSIFY_ONE) {
1046 DB((dbg, LEVEL_2, "Sub(1) to Dec ... "));
1047 new_op = new_rd_ia32_Dec(dbgi, irg, block, noreg, noreg, expr_op, nomem);
1048 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1050 } else if (class_tv == TV_CLASSIFY_ALL_ONE || class_negtv == TV_CLASSIFY_ONE) {
1051 DB((dbg, LEVEL_2, "Sub(-1) to Inc ... "));
1052 new_op = new_rd_ia32_Inc(dbgi, irg, block, noreg, noreg, expr_op, nomem);
1053 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1059 /* This is a normal sub */
1060 new_op = new_rd_ia32_Sub(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
1062 /* set AM support */
1063 set_ia32_am_support(new_op, ia32_am_Full, ia32_am_binary);
1065 fold_immediate(new_op, 2, 3);
1067 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1075 * Generates an ia32 DivMod with additional infrastructure for the
1076 * register allocator if needed.
1078 * @param dividend -no comment- :)
1079 * @param divisor -no comment- :)
1080 * @param dm_flav flavour_Div/Mod/DivMod
1081 * @return The created ia32 DivMod node
1083 static ir_node *generate_DivMod(ir_node *node, ir_node *dividend,
1084 ir_node *divisor, ia32_op_flavour_t dm_flav)
1086 ir_node *block = be_transform_node(get_nodes_block(node));
1087 ir_node *new_dividend = be_transform_node(dividend);
1088 ir_node *new_divisor = be_transform_node(divisor);
1089 ir_graph *irg = current_ir_graph;
1090 dbg_info *dbgi = get_irn_dbg_info(node);
1091 ir_mode *mode = get_irn_mode(node);
1092 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1093 ir_node *res, *proj_div, *proj_mod;
1094 ir_node *sign_extension;
1095 ir_node *mem, *new_mem;
1096 ir_node *projs[pn_DivMod_max];
1099 ia32_collect_Projs(node, projs, pn_DivMod_max);
1101 proj_div = proj_mod = NULL;
1105 mem = get_Div_mem(node);
1106 mode = get_Div_resmode(node);
1107 proj_div = be_get_Proj_for_pn(node, pn_Div_res);
1108 has_exc = be_get_Proj_for_pn(node, pn_Div_X_except) != NULL;
1111 mem = get_Mod_mem(node);
1112 mode = get_Mod_resmode(node);
1113 proj_mod = be_get_Proj_for_pn(node, pn_Mod_res);
1114 has_exc = be_get_Proj_for_pn(node, pn_Mod_X_except) != NULL;
1116 case flavour_DivMod:
1117 mem = get_DivMod_mem(node);
1118 mode = get_DivMod_resmode(node);
1119 proj_div = be_get_Proj_for_pn(node, pn_DivMod_res_div);
1120 proj_mod = be_get_Proj_for_pn(node, pn_DivMod_res_mod);
1121 has_exc = be_get_Proj_for_pn(node, pn_DivMod_X_except) != NULL;
1124 panic("invalid divmod flavour!");
1126 new_mem = be_transform_node(mem);
1128 if (mode_is_signed(mode)) {
1129 /* in signed mode, we need to sign extend the dividend */
1130 ir_node *produceval = new_rd_ia32_ProduceVal(dbgi, irg, block);
1131 add_irn_dep(produceval, get_irg_frame(irg));
1132 sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_dividend,
1135 sign_extension = new_rd_ia32_Const(dbgi, irg, block);
1136 set_ia32_Immop_tarval(sign_extension, get_tarval_null(mode_Iu));
1138 add_irn_dep(sign_extension, get_irg_frame(irg));
1141 if (mode_is_signed(mode)) {
1142 res = new_rd_ia32_IDiv(dbgi, irg, block, noreg, noreg, new_dividend,
1143 sign_extension, new_divisor, new_mem, dm_flav);
1145 res = new_rd_ia32_Div(dbgi, irg, block, noreg, noreg, new_dividend,
1146 sign_extension, new_divisor, new_mem, dm_flav);
1149 set_ia32_exc_label(res, has_exc);
1150 set_irn_pinned(res, get_irn_pinned(node));
1151 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
1153 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1160 * Wrapper for generate_DivMod. Sets flavour_Mod.
1163 static ir_node *gen_Mod(ir_node *node) {
1164 return generate_DivMod(node, get_Mod_left(node),
1165 get_Mod_right(node), flavour_Mod);
1169 * Wrapper for generate_DivMod. Sets flavour_Div.
1172 static ir_node *gen_Div(ir_node *node) {
1173 return generate_DivMod(node, get_Div_left(node),
1174 get_Div_right(node), flavour_Div);
1178 * Wrapper for generate_DivMod. Sets flavour_DivMod.
1180 static ir_node *gen_DivMod(ir_node *node) {
1181 return generate_DivMod(node, get_DivMod_left(node),
1182 get_DivMod_right(node), flavour_DivMod);
1188 * Creates an ia32 floating Div.
1190 * @return The created ia32 xDiv node
1192 static ir_node *gen_Quot(ir_node *node) {
1193 ir_node *block = be_transform_node(get_nodes_block(node));
1194 ir_node *op1 = get_Quot_left(node);
1195 ir_node *new_op1 = be_transform_node(op1);
1196 ir_node *op2 = get_Quot_right(node);
1197 ir_node *new_op2 = be_transform_node(op2);
1198 ir_graph *irg = current_ir_graph;
1199 dbg_info *dbgi = get_irn_dbg_info(node);
1200 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1201 ir_node *nomem = new_rd_NoMem(current_ir_graph);
1204 if (USE_SSE2(env_cg)) {
1205 ir_mode *mode = get_irn_mode(op1);
1206 if (is_ia32_xConst(new_op2)) {
1207 new_op = new_rd_ia32_xDiv(dbgi, irg, block, noreg, noreg, new_op1, noreg, nomem);
1208 set_ia32_am_support(new_op, ia32_am_None, ia32_am_arity_none);
1209 copy_ia32_Immop_attr(new_op, new_op2);
1211 new_op = new_rd_ia32_xDiv(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
1212 // Matze: disabled for now, spillslot coalescer fails
1213 set_ia32_am_support(new_op, ia32_am_Source, ia32_am_binary);
1215 set_ia32_ls_mode(new_op, mode);
1217 new_op = new_rd_ia32_vfdiv(dbgi, irg, block, noreg, noreg, new_op1,
1218 new_op2, nomem, get_fpcw());
1219 // Matze: disabled for now (spillslot coalescer fails)
1220 set_ia32_am_support(new_op, ia32_am_Source, ia32_am_binary);
1222 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1228 * Creates an ia32 Shl.
1230 * @return The created ia32 Shl node
1232 static ir_node *gen_Shl(ir_node *node) {
1233 return gen_shift_binop(node, get_Shl_left(node), get_Shl_right(node),
1240 * Creates an ia32 Shr.
1242 * @return The created ia32 Shr node
1244 static ir_node *gen_Shr(ir_node *node) {
1245 return gen_shift_binop(node, get_Shr_left(node),
1246 get_Shr_right(node), new_rd_ia32_Shr);
1252 * Creates an ia32 Sar.
1254 * @return The created ia32 Shrs node
1256 static ir_node *gen_Shrs(ir_node *node) {
1257 ir_node *left = get_Shrs_left(node);
1258 ir_node *right = get_Shrs_right(node);
1259 ir_mode *mode = get_irn_mode(node);
1260 if(is_Const(right) && mode == mode_Is) {
1261 tarval *tv = get_Const_tarval(right);
1262 long val = get_tarval_long(tv);
1264 /* this is a sign extension */
1265 ir_graph *irg = current_ir_graph;
1266 dbg_info *dbgi = get_irn_dbg_info(node);
1267 ir_node *block = be_transform_node(get_nodes_block(node));
1269 ir_node *new_op = be_transform_node(op);
1270 ir_node *pval = new_rd_ia32_ProduceVal(dbgi, irg, block);
1271 add_irn_dep(pval, get_irg_frame(irg));
1273 return new_rd_ia32_Cltd(dbgi, irg, block, new_op, pval);
1277 /* 8 or 16 bit sign extension? */
1278 if(is_Const(right) && is_Shl(left) && mode == mode_Is) {
1279 ir_node *shl_left = get_Shl_left(left);
1280 ir_node *shl_right = get_Shl_right(left);
1281 if(is_Const(shl_right)) {
1282 tarval *tv1 = get_Const_tarval(right);
1283 tarval *tv2 = get_Const_tarval(shl_right);
1284 if(tv1 == tv2 && tarval_is_long(tv1)) {
1285 long val = get_tarval_long(tv1);
1286 if(val == 16 || val == 24) {
1287 dbg_info *dbgi = get_irn_dbg_info(node);
1288 ir_node *block = be_transform_node(get_nodes_block(node));
1289 ir_node *new_op = be_transform_node(shl_left);
1299 res = create_I2I_Conv(src_mode, mode_Is, dbgi, block,
1301 SET_IA32_ORIG_NODE(res,
1302 ia32_get_old_node_name(env_cg, node));
1311 return gen_shift_binop(node, left, right, new_rd_ia32_Sar);
1317 * Creates an ia32 RotL.
1319 * @param op1 The first operator
1320 * @param op2 The second operator
1321 * @return The created ia32 RotL node
1323 static ir_node *gen_RotL(ir_node *node,
1324 ir_node *op1, ir_node *op2) {
1325 return gen_shift_binop(node, op1, op2, new_rd_ia32_Rol);
1331 * Creates an ia32 RotR.
1332 * NOTE: There is no RotR with immediate because this would always be a RotL
1333 * "imm-mode_size_bits" which can be pre-calculated.
1335 * @param op1 The first operator
1336 * @param op2 The second operator
1337 * @return The created ia32 RotR node
1339 static ir_node *gen_RotR(ir_node *node, ir_node *op1,
1341 return gen_shift_binop(node, op1, op2, new_rd_ia32_Ror);
1347 * Creates an ia32 RotR or RotL (depending on the found pattern).
1349 * @return The created ia32 RotL or RotR node
1351 static ir_node *gen_Rot(ir_node *node) {
1352 ir_node *rotate = NULL;
1353 ir_node *op1 = get_Rot_left(node);
1354 ir_node *op2 = get_Rot_right(node);
1356 /* Firm has only Rot (which is a RotL), so we are looking for a right (op2)
1357 operand "-e+mode_size_bits" (it's an already modified "mode_size_bits-e",
1358 that means we can create a RotR instead of an Add and a RotL */
1360 if (get_irn_op(op2) == op_Add) {
1362 ir_node *left = get_Add_left(add);
1363 ir_node *right = get_Add_right(add);
1364 if (is_Const(right)) {
1365 tarval *tv = get_Const_tarval(right);
1366 ir_mode *mode = get_irn_mode(node);
1367 long bits = get_mode_size_bits(mode);
1369 if (get_irn_op(left) == op_Minus &&
1370 tarval_is_long(tv) &&
1371 get_tarval_long(tv) == bits)
1373 DB((dbg, LEVEL_1, "RotL into RotR ... "));
1374 rotate = gen_RotR(node, op1, get_Minus_op(left));
1379 if (rotate == NULL) {
1380 rotate = gen_RotL(node, op1, op2);
1389 * Transforms a Minus node.
1391 * @param op The Minus operand
1392 * @return The created ia32 Minus node
1394 ir_node *gen_Minus_ex(ir_node *node, ir_node *op) {
1395 ir_node *block = be_transform_node(get_nodes_block(node));
1396 ir_graph *irg = current_ir_graph;
1397 dbg_info *dbgi = get_irn_dbg_info(node);
1398 ir_mode *mode = get_irn_mode(node);
1403 if (mode_is_float(mode)) {
1404 ir_node *new_op = be_transform_node(op);
1405 if (USE_SSE2(env_cg)) {
1406 ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
1407 ir_node *noreg_fp = ia32_new_NoReg_fp(env_cg);
1408 ir_node *nomem = new_rd_NoMem(irg);
1410 res = new_rd_ia32_xXor(dbgi, irg, block, noreg_gp, noreg_gp, new_op, noreg_fp, nomem);
1412 size = get_mode_size_bits(mode);
1413 ent = ia32_gen_fp_known_const(size == 32 ? ia32_SSIGN : ia32_DSIGN);
1415 set_ia32_am_sc(res, ent);
1416 set_ia32_op_type(res, ia32_AddrModeS);
1417 set_ia32_ls_mode(res, mode);
1419 res = new_rd_ia32_vfchs(dbgi, irg, block, new_op);
1422 res = gen_unop(node, op, new_rd_ia32_Neg);
1425 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1431 * Transforms a Minus node.
1433 * @return The created ia32 Minus node
1435 static ir_node *gen_Minus(ir_node *node) {
1436 return gen_Minus_ex(node, get_Minus_op(node));
1439 static ir_node *create_Immediate_from_int(int val)
1441 ir_graph *irg = current_ir_graph;
1442 ir_node *start_block = get_irg_start_block(irg);
1443 ir_node *immediate = new_rd_ia32_Immediate(NULL, irg, start_block, NULL, 0, val);
1444 arch_set_irn_register(env_cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]);
1449 static ir_node *gen_bin_Not(ir_node *node)
1451 ir_graph *irg = current_ir_graph;
1452 dbg_info *dbgi = get_irn_dbg_info(node);
1453 ir_node *block = be_transform_node(get_nodes_block(node));
1454 ir_node *op = get_Not_op(node);
1455 ir_node *new_op = be_transform_node(op);
1456 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1457 ir_node *nomem = new_NoMem();
1458 ir_node *one = create_Immediate_from_int(1);
1460 return new_rd_ia32_Xor(dbgi, irg, block, noreg, noreg, new_op, one, nomem);
1464 * Transforms a Not node.
1466 * @return The created ia32 Not node
1468 static ir_node *gen_Not(ir_node *node) {
1469 ir_node *op = get_Not_op(node);
1470 ir_mode *mode = get_irn_mode(node);
1472 if(mode == mode_b) {
1473 return gen_bin_Not(node);
1476 assert (! mode_is_float(get_irn_mode(node)));
1477 return gen_unop(node, op, new_rd_ia32_Not);
1483 * Transforms an Abs node.
1485 * @return The created ia32 Abs node
1487 static ir_node *gen_Abs(ir_node *node) {
1488 ir_node *block = be_transform_node(get_nodes_block(node));
1489 ir_node *op = get_Abs_op(node);
1490 ir_node *new_op = be_transform_node(op);
1491 ir_graph *irg = current_ir_graph;
1492 dbg_info *dbgi = get_irn_dbg_info(node);
1493 ir_mode *mode = get_irn_mode(node);
1494 ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
1495 ir_node *noreg_fp = ia32_new_NoReg_fp(env_cg);
1496 ir_node *nomem = new_NoMem();
1501 if (mode_is_float(mode)) {
1502 if (USE_SSE2(env_cg)) {
1503 res = new_rd_ia32_xAnd(dbgi,irg, block, noreg_gp, noreg_gp, new_op, noreg_fp, nomem);
1505 size = get_mode_size_bits(mode);
1506 ent = ia32_gen_fp_known_const(size == 32 ? ia32_SABS : ia32_DABS);
1508 set_ia32_am_sc(res, ent);
1510 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1512 set_ia32_op_type(res, ia32_AddrModeS);
1513 set_ia32_ls_mode(res, mode);
1516 res = new_rd_ia32_vfabs(dbgi, irg, block, new_op);
1517 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1521 ir_node *pval = new_rd_ia32_ProduceVal(dbgi, irg, block);
1522 ir_node *sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_op,
1525 add_irn_dep(pval, get_irg_frame(irg));
1526 SET_IA32_ORIG_NODE(sign_extension,
1527 ia32_get_old_node_name(env_cg, node));
1529 xor = new_rd_ia32_Xor(dbgi, irg, block, noreg_gp, noreg_gp, new_op,
1530 sign_extension, nomem);
1531 SET_IA32_ORIG_NODE(xor, ia32_get_old_node_name(env_cg, node));
1533 res = new_rd_ia32_Sub(dbgi, irg, block, noreg_gp, noreg_gp, xor,
1534 sign_extension, nomem);
1535 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1544 * Transforms a Load.
1546 * @return the created ia32 Load node
1548 static ir_node *gen_Load(ir_node *node) {
1549 ir_node *old_block = get_nodes_block(node);
1550 ir_node *block = be_transform_node(old_block);
1551 ir_node *ptr = get_Load_ptr(node);
1552 ir_node *new_ptr = be_transform_node(ptr);
1553 ir_node *mem = get_Load_mem(node);
1554 ir_node *new_mem = be_transform_node(mem);
1555 ir_graph *irg = current_ir_graph;
1556 dbg_info *dbgi = get_irn_dbg_info(node);
1557 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1558 ir_mode *mode = get_Load_mode(node);
1560 ir_node *lptr = new_ptr;
1563 ia32_am_flavour_t am_flav = ia32_am_B;
1565 /* address might be a constant (symconst or absolute address) */
1566 if (is_ia32_Const(new_ptr)) {
1571 if (mode_is_float(mode)) {
1572 if (USE_SSE2(env_cg)) {
1573 new_op = new_rd_ia32_xLoad(dbgi, irg, block, lptr, noreg, new_mem);
1574 res_mode = mode_xmm;
1576 new_op = new_rd_ia32_vfld(dbgi, irg, block, lptr, noreg, new_mem, mode);
1577 res_mode = mode_vfp;
1583 /* create a conv node with address mode for smaller modes */
1584 if(get_mode_size_bits(mode) < 32) {
1585 new_op = new_rd_ia32_Conv_I2I(dbgi, irg, block, lptr, noreg, noreg,
1588 new_op = new_rd_ia32_Load(dbgi, irg, block, lptr, noreg, new_mem);
1593 /* base is a constant address */
1595 if (get_ia32_immop_type(new_ptr) == ia32_ImmSymConst) {
1596 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_ptr));
1597 am_flav = ia32_am_N;
1599 tarval *tv = get_ia32_Immop_tarval(new_ptr);
1600 long offs = get_tarval_long(tv);
1602 add_ia32_am_offs_int(new_op, offs);
1603 am_flav = ia32_am_O;
1607 set_irn_pinned(new_op, get_irn_pinned(node));
1608 set_ia32_op_type(new_op, ia32_AddrModeS);
1609 set_ia32_am_flavour(new_op, am_flav);
1610 set_ia32_ls_mode(new_op, mode);
1612 /* make sure we are scheduled behind the initial IncSP/Barrier
1613 * to avoid spills being placed before it
1615 if (block == get_irg_start_block(irg)) {
1616 add_irn_dep(new_op, get_irg_frame(irg));
1619 set_ia32_exc_label(new_op, be_get_Proj_for_pn(node, pn_Load_X_except) != NULL);
1620 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1628 * Transforms a Store.
1630 * @return the created ia32 Store node
1632 static ir_node *gen_Store(ir_node *node) {
1633 ir_node *block = be_transform_node(get_nodes_block(node));
1634 ir_node *ptr = get_Store_ptr(node);
1635 ir_node *new_ptr = be_transform_node(ptr);
1636 ir_node *val = get_Store_value(node);
1638 ir_node *mem = get_Store_mem(node);
1639 ir_node *new_mem = be_transform_node(mem);
1640 ir_graph *irg = current_ir_graph;
1641 dbg_info *dbgi = get_irn_dbg_info(node);
1642 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1643 ir_node *sptr = new_ptr;
1644 ir_mode *mode = get_irn_mode(val);
1647 ia32_am_flavour_t am_flav = ia32_am_B;
1649 /* address might be a constant (symconst or absolute address) */
1650 if (is_ia32_Const(new_ptr)) {
1655 if (mode_is_float(mode)) {
1656 new_val = be_transform_node(val);
1657 if (USE_SSE2(env_cg)) {
1658 new_op = new_rd_ia32_xStore(dbgi, irg, block, sptr, noreg, new_val,
1661 new_op = new_rd_ia32_vfst(dbgi, irg, block, sptr, noreg, new_val,
1665 new_val = create_immediate_or_transform(val, 0);
1669 if (get_mode_size_bits(mode) == 8) {
1670 new_op = new_rd_ia32_Store8Bit(dbgi, irg, block, sptr, noreg,
1673 new_op = new_rd_ia32_Store(dbgi, irg, block, sptr, noreg, new_val,
1678 /* base is an constant address */
1680 if (get_ia32_immop_type(new_ptr) == ia32_ImmSymConst) {
1681 set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_ptr));
1682 am_flav = ia32_am_N;
1684 tarval *tv = get_ia32_Immop_tarval(new_ptr);
1685 long offs = get_tarval_long(tv);
1687 add_ia32_am_offs_int(new_op, offs);
1688 am_flav = ia32_am_O;
1692 set_irn_pinned(new_op, get_irn_pinned(node));
1693 set_ia32_op_type(new_op, ia32_AddrModeD);
1694 set_ia32_am_flavour(new_op, am_flav);
1695 set_ia32_ls_mode(new_op, mode);
1697 set_ia32_exc_label(new_op, be_get_Proj_for_pn(node, pn_Store_X_except) != NULL);
1698 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
1703 static ir_node *maybe_scale_up(ir_node *new_op, ir_mode *mode, dbg_info *dbgi)
1708 if(get_mode_size_bits(mode) == 32)
1712 if(is_ia32_Immediate(new_op))
1715 if(mode_is_signed(mode))
1720 block = get_nodes_block(new_op);
1721 return create_I2I_Conv(mode, tgt_mode, dbgi, block, new_op);
1724 static ir_node *try_create_TestJmp(ir_node *block, dbg_info *dbgi, long pnc,
1725 ir_node *cmp_left, ir_node *cmp_right)
1727 ir_node *new_cmp_left;
1728 ir_node *new_cmp_right;
1735 long pure_pnc = pnc & ~ia32_pn_Cmp_Unsigned;
1737 if(cmp_right != NULL && !is_Const_0(cmp_right))
1740 if(is_And(cmp_left) && (pure_pnc == pn_Cmp_Eq || pure_pnc == pn_Cmp_Lg)) {
1741 and_left = get_And_left(cmp_left);
1742 and_right = get_And_right(cmp_left);
1744 mode = get_irn_mode(and_left);
1745 new_cmp_left = be_transform_node(and_left);
1746 new_cmp_right = create_immediate_or_transform(and_right, 0);
1748 mode = get_irn_mode(cmp_left);
1749 new_cmp_left = be_transform_node(cmp_left);
1750 new_cmp_right = be_transform_node(cmp_left);
1753 assert(get_mode_size_bits(mode) <= 32);
1754 new_cmp_left = maybe_scale_up(new_cmp_left, mode, dbgi);
1755 new_cmp_right = maybe_scale_up(new_cmp_right, mode, dbgi);
1756 noreg = ia32_new_NoReg_gp(env_cg);
1757 nomem = new_NoMem();
1759 res = new_rd_ia32_TestJmp(dbgi, current_ir_graph, block, noreg, noreg,
1760 new_cmp_left, new_cmp_right, nomem, pnc);
1761 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
1766 static ir_node *create_Switch(ir_node *node)
1768 ir_graph *irg = current_ir_graph;
1769 dbg_info *dbgi = get_irn_dbg_info(node);
1770 ir_node *block = be_transform_node(get_nodes_block(node));
1771 ir_node *sel = get_Cond_selector(node);
1772 ir_node *new_sel = be_transform_node(sel);
1774 int switch_min = INT_MAX;
1775 const ir_edge_t *edge;
1777 assert(get_mode_size_bits(get_irn_mode(sel)) == 32);
1779 /* determine the smallest switch case value */
1780 foreach_out_edge(node, edge) {
1781 ir_node *proj = get_edge_src_irn(edge);
1782 int pn = get_Proj_proj(proj);
1787 if (switch_min != 0) {
1788 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1790 /* if smallest switch case is not 0 we need an additional sub */
1791 new_sel = new_rd_ia32_Lea(dbgi, irg, block, new_sel, noreg);
1792 add_ia32_am_offs_int(new_sel, -switch_min);
1793 set_ia32_am_flavour(new_sel, ia32_am_OB);
1794 set_ia32_op_type(new_sel, ia32_AddrModeS);
1796 SET_IA32_ORIG_NODE(new_sel, ia32_get_old_node_name(env_cg, node));
1799 res = new_rd_ia32_SwitchJmp(dbgi, irg, block, new_sel);
1800 set_ia32_pncode(res, get_Cond_defaultProj(node));
1802 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1808 * Transforms a Cond -> Proj[b] -> Cmp into a CondJmp, CondJmp_i or TestJmp
1810 * @return The transformed node.
1812 static ir_node *gen_Cond(ir_node *node) {
1813 ir_node *block = be_transform_node(get_nodes_block(node));
1814 ir_graph *irg = current_ir_graph;
1815 dbg_info *dbgi = get_irn_dbg_info(node);
1816 ir_node *sel = get_Cond_selector(node);
1817 ir_mode *sel_mode = get_irn_mode(sel);
1818 ir_node *res = NULL;
1819 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1826 ir_node *nomem = new_NoMem();
1829 if (sel_mode != mode_b) {
1830 return create_Switch(node);
1833 if(!is_Proj(sel) || !is_Cmp(get_Proj_pred(sel))) {
1834 /* it's some mode_b value but not a direct comparison -> create a
1836 res = try_create_TestJmp(block, dbgi, pn_Cmp_Lg, sel, NULL);
1837 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1841 cmp = get_Proj_pred(sel);
1842 cmp_a = get_Cmp_left(cmp);
1843 cmp_b = get_Cmp_right(cmp);
1844 cmp_mode = get_irn_mode(cmp_a);
1845 pnc = get_Proj_proj(sel);
1846 if(mode_is_float(cmp_mode) || !mode_is_signed(cmp_mode)) {
1847 pnc |= ia32_pn_Cmp_Unsigned;
1850 if(mode_needs_gp_reg(cmp_mode)) {
1851 res = try_create_TestJmp(block, dbgi, pnc, cmp_a, cmp_b);
1853 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1858 new_cmp_a = be_transform_node(cmp_a);
1859 new_cmp_b = create_immediate_or_transform(cmp_b, 0);
1861 if (mode_is_float(cmp_mode)) {
1862 if (USE_SSE2(env_cg)) {
1863 res = new_rd_ia32_xCondJmp(dbgi, irg, block, noreg, noreg, cmp_a,
1865 set_ia32_commutative(res);
1866 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
1867 set_ia32_ls_mode(res, cmp_mode);
1869 res = new_rd_ia32_vfCondJmp(dbgi, irg, block, cmp_a, cmp_b, pnc);
1870 set_ia32_commutative(res);
1873 /** workaround smaller compare modes with converts...
1874 * We could easily support 16bit compares, for 8 bit we have to set
1875 * additional register constraints, which we don't do yet
1877 new_cmp_a = maybe_scale_up(new_cmp_a, cmp_mode, dbgi);
1878 new_cmp_b = maybe_scale_up(new_cmp_b, cmp_mode, dbgi);
1880 res = new_rd_ia32_CondJmp(dbgi, irg, block, noreg, noreg,
1881 new_cmp_a, new_cmp_b, nomem, pnc);
1882 set_ia32_commutative(res);
1883 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
1886 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1894 * Transforms a CopyB node.
1896 * @return The transformed node.
1898 static ir_node *gen_CopyB(ir_node *node) {
1899 ir_node *block = be_transform_node(get_nodes_block(node));
1900 ir_node *src = get_CopyB_src(node);
1901 ir_node *new_src = be_transform_node(src);
1902 ir_node *dst = get_CopyB_dst(node);
1903 ir_node *new_dst = be_transform_node(dst);
1904 ir_node *mem = get_CopyB_mem(node);
1905 ir_node *new_mem = be_transform_node(mem);
1906 ir_node *res = NULL;
1907 ir_graph *irg = current_ir_graph;
1908 dbg_info *dbgi = get_irn_dbg_info(node);
1909 int size = get_type_size_bytes(get_CopyB_type(node));
1912 /* If we have to copy more than 32 bytes, we use REP MOVSx and */
1913 /* then we need the size explicitly in ECX. */
1914 if (size >= 32 * 4) {
1915 rem = size & 0x3; /* size % 4 */
1918 res = new_rd_ia32_Const(dbgi, irg, block);
1919 add_irn_dep(res, be_abi_get_start_barrier(env_cg->birg->abi));
1920 set_ia32_Immop_tarval(res, new_tarval_from_long(size, mode_Is));
1922 res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem);
1923 set_ia32_Immop_tarval(res, new_tarval_from_long(rem, mode_Is));
1925 res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem);
1926 set_ia32_Immop_tarval(res, new_tarval_from_long(size, mode_Is));
1929 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
1935 ir_node *gen_be_Copy(ir_node *node)
1937 ir_node *result = be_duplicate_node(node);
1938 ir_mode *mode = get_irn_mode(result);
1940 if (mode_needs_gp_reg(mode)) {
1941 set_irn_mode(result, mode_Iu);
1948 static ir_node *create_set(long pnc, ir_node *cmp_left, ir_node *cmp_right,
1949 dbg_info *dbgi, ir_node *block)
1951 ir_graph *irg = current_ir_graph;
1952 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
1953 ir_node *nomem = new_rd_NoMem(irg);
1955 ir_node *new_cmp_left;
1956 ir_node *new_cmp_right;
1959 /* can we use a test instruction? */
1960 if(cmp_right == NULL || is_Const_0(cmp_right)) {
1961 long pure_pnc = pnc & ~ia32_pn_Cmp_Unsigned;
1962 if(is_And(cmp_left) &&
1963 (pure_pnc == pn_Cmp_Eq || pure_pnc == pn_Cmp_Lg)) {
1964 ir_node *and_left = get_And_left(cmp_left);
1965 ir_node *and_right = get_And_right(cmp_left);
1967 mode = get_irn_mode(and_left);
1968 new_cmp_left = be_transform_node(and_left);
1969 new_cmp_right = create_immediate_or_transform(and_right, 0);
1971 mode = get_irn_mode(cmp_left);
1972 new_cmp_left = be_transform_node(cmp_left);
1973 new_cmp_right = be_transform_node(cmp_left);
1976 assert(get_mode_size_bits(mode) <= 32);
1977 new_cmp_left = maybe_scale_up(new_cmp_left, mode, dbgi);
1978 new_cmp_right = maybe_scale_up(new_cmp_right, mode, dbgi);
1980 res = new_rd_ia32_TestSet(dbgi, current_ir_graph, block, noreg, noreg,
1981 new_cmp_left, new_cmp_right, nomem, pnc);
1982 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
1984 res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, block, noreg, noreg, res,
1990 mode = get_irn_mode(cmp_left);
1992 new_cmp_left = be_transform_node(cmp_left);
1993 new_cmp_right = create_immediate_or_transform(cmp_right, 0);
1995 assert(get_mode_size_bits(mode) <= 32);
1996 new_cmp_left = maybe_scale_up(new_cmp_left, mode, dbgi);
1997 new_cmp_right = maybe_scale_up(new_cmp_right, mode, dbgi);
1999 res = new_rd_ia32_CmpSet(dbgi, irg, block, noreg, noreg, new_cmp_left,
2000 new_cmp_right, nomem, pnc);
2001 res = new_rd_ia32_ConvI2I8Bit(dbgi, irg, block, noreg, noreg, res, nomem,
2007 static ir_node *create_cmov(long pnc, ir_node *cmp_left, ir_node *cmp_right,
2008 ir_node *val_true, ir_node *val_false,
2009 dbg_info *dbgi, ir_node *block)
2011 ir_graph *irg = current_ir_graph;
2012 ir_node *new_val_true = be_transform_node(val_true);
2013 ir_node *new_val_false = be_transform_node(val_false);
2014 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
2015 ir_node *nomem = new_NoMem();
2016 ir_node *new_cmp_left;
2017 ir_node *new_cmp_right;
2020 /* cmovs with unknowns are pointless... */
2021 if(is_Unknown(val_true)) {
2022 #ifdef DEBUG_libfirm
2023 ir_fprintf(stderr, "Optimisation warning: psi with unknown operand\n");
2025 return new_val_false;
2027 if(is_Unknown(val_false)) {
2028 #ifdef DEBUG_libfirm
2029 ir_fprintf(stderr, "Optimisation warning: psi with unknown operand\n");
2031 return new_val_true;
2034 /* can we use a test instruction? */
2035 if(is_Const_0(cmp_right)) {
2036 long pure_pnc = pnc & ~ia32_pn_Cmp_Unsigned;
2037 if(is_And(cmp_left) &&
2038 (pure_pnc == pn_Cmp_Eq || pure_pnc == pn_Cmp_Lg)) {
2039 ir_node *and_left = get_And_left(cmp_left);
2040 ir_node *and_right = get_And_right(cmp_left);
2042 new_cmp_left = be_transform_node(and_left);
2043 new_cmp_right = create_immediate_or_transform(and_right, 0);
2045 new_cmp_left = be_transform_node(cmp_left);
2046 new_cmp_right = be_transform_node(cmp_left);
2049 res = new_rd_ia32_TestCMov(dbgi, current_ir_graph, block, noreg, noreg,
2050 new_cmp_left, new_cmp_right, nomem,
2051 new_val_true, new_val_false, pnc);
2052 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
2057 new_cmp_left = be_transform_node(cmp_left);
2058 new_cmp_right = create_immediate_or_transform(cmp_right, 0);
2060 res = new_rd_ia32_CmpCMov(dbgi, irg, block, noreg, noreg, new_cmp_left,
2061 new_cmp_right, nomem, new_val_true, new_val_false,
2063 set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
2070 * Transforms a Psi node into CMov.
2072 * @return The transformed node.
2074 static ir_node *gen_Psi(ir_node *node) {
2075 ir_node *psi_true = get_Psi_val(node, 0);
2076 ir_node *psi_default = get_Psi_default(node);
2077 ia32_code_gen_t *cg = env_cg;
2078 ir_node *cond = get_Psi_cond(node, 0);
2079 ir_node *block = be_transform_node(get_nodes_block(node));
2080 dbg_info *dbgi = get_irn_dbg_info(node);
2087 assert(get_Psi_n_conds(node) == 1);
2088 assert(get_irn_mode(cond) == mode_b);
2089 assert(mode_needs_gp_reg(get_irn_mode(node)));
2091 if(!is_Proj(cond) || !is_Cmp(get_Proj_pred(cond))) {
2092 /* a mode_b value, we have to compare it against 0 */
2094 cmp_right = new_Const_long(mode_Iu, 0);
2098 ir_node *cmp = get_Proj_pred(cond);
2100 cmp_left = get_Cmp_left(cmp);
2101 cmp_right = get_Cmp_right(cmp);
2102 cmp_mode = get_irn_mode(cmp_left);
2103 pnc = get_Proj_proj(cond);
2105 assert(!mode_is_float(cmp_mode));
2107 if (!mode_is_signed(cmp_mode)) {
2108 pnc |= ia32_pn_Cmp_Unsigned;
2112 if(is_Const_1(psi_true) && is_Const_0(psi_default)) {
2113 new_op = create_set(pnc, cmp_left, cmp_right, dbgi, block);
2114 } else if(is_Const_0(psi_true) && is_Const_1(psi_default)) {
2115 pnc = get_negated_pnc(pnc, cmp_mode);
2116 new_op = create_set(pnc, cmp_left, cmp_right, dbgi, block);
2118 new_op = create_cmov(pnc, cmp_left, cmp_right, psi_true, psi_default,
2121 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, node));
2127 * Following conversion rules apply:
2131 * 1) n bit -> m bit n > m (downscale)
2133 * 2) n bit -> m bit n == m (sign change)
2135 * 3) n bit -> m bit n < m (upscale)
2136 * a) source is signed: movsx
2137 * b) source is unsigned: and with lower bits sets
2141 * SSE(1/2) convert to float or double (cvtsi2ss/sd)
2145 * SSE(1/2) convert from float or double to 32bit int (cvtss/sd2si)
2149 * SSE(1/2) convert from float or double to double or float (cvtss/sd2sd/ss)
2150 * x87 is mode_E internally, conversions happen only at load and store
2151 * in non-strict semantic
2155 * Create a conversion from x87 state register to general purpose.
2157 static ir_node *gen_x87_fp_to_gp(ir_node *node) {
2158 ir_node *block = be_transform_node(get_nodes_block(node));
2159 ir_node *op = get_Conv_op(node);
2160 ir_node *new_op = be_transform_node(op);
2161 ia32_code_gen_t *cg = env_cg;
2162 ir_graph *irg = current_ir_graph;
2163 dbg_info *dbgi = get_irn_dbg_info(node);
2164 ir_node *noreg = ia32_new_NoReg_gp(cg);
2165 ir_node *trunc_mode = ia32_new_Fpu_truncate(cg);
2166 ir_mode *mode = get_irn_mode(node);
2167 ir_node *fist, *load;
2170 fist = new_rd_ia32_vfist(dbgi, irg, block,
2171 get_irg_frame(irg), noreg, new_op, trunc_mode, new_NoMem());
2173 set_irn_pinned(fist, op_pin_state_floats);
2174 set_ia32_use_frame(fist);
2175 set_ia32_op_type(fist, ia32_AddrModeD);
2176 set_ia32_am_flavour(fist, ia32_am_B);
2178 assert(get_mode_size_bits(mode) <= 32);
2179 /* exception we can only store signed 32 bit integers, so for unsigned
2180 we store a 64bit (signed) integer and load the lower bits */
2181 if(get_mode_size_bits(mode) == 32 && !mode_is_signed(mode)) {
2182 set_ia32_ls_mode(fist, mode_Ls);
2184 set_ia32_ls_mode(fist, mode_Is);
2186 SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(cg, node));
2189 load = new_rd_ia32_Load(dbgi, irg, block, get_irg_frame(irg), noreg, fist);
2191 set_irn_pinned(load, op_pin_state_floats);
2192 set_ia32_use_frame(load);
2193 set_ia32_op_type(load, ia32_AddrModeS);
2194 set_ia32_am_flavour(load, ia32_am_B);
2195 set_ia32_ls_mode(load, mode_Is);
2196 if(get_ia32_ls_mode(fist) == mode_Ls) {
2197 ia32_attr_t *attr = get_ia32_attr(load);
2198 attr->data.need_64bit_stackent = 1;
2200 ia32_attr_t *attr = get_ia32_attr(load);
2201 attr->data.need_32bit_stackent = 1;
2203 SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(cg, node));
2205 return new_r_Proj(irg, block, load, mode_Iu, pn_ia32_Load_res);
2208 static ir_node *create_strict_conv(ir_mode *tgt_mode, ir_node *node)
2210 ir_node *block = get_nodes_block(node);
2211 ir_graph *irg = current_ir_graph;
2212 dbg_info *dbgi = get_irn_dbg_info(node);
2213 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
2214 ir_node *nomem = new_NoMem();
2215 ir_node *frame = get_irg_frame(irg);
2216 ir_node *store, *load;
2219 store = new_rd_ia32_vfst(dbgi, irg, block, frame, noreg, node, nomem,
2221 set_ia32_use_frame(store);
2222 set_ia32_op_type(store, ia32_AddrModeD);
2223 set_ia32_am_flavour(store, ia32_am_OB);
2224 SET_IA32_ORIG_NODE(store, ia32_get_old_node_name(env_cg, node));
2226 load = new_rd_ia32_vfld(dbgi, irg, block, frame, noreg, store,
2228 set_ia32_use_frame(load);
2229 set_ia32_op_type(load, ia32_AddrModeS);
2230 set_ia32_am_flavour(load, ia32_am_OB);
2231 SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
2233 res = new_r_Proj(irg, block, load, mode_E, pn_ia32_vfld_res);
2238 * Create a conversion from general purpose to x87 register
2240 static ir_node *gen_x87_gp_to_fp(ir_node *node, ir_mode *src_mode) {
2241 ir_node *block = be_transform_node(get_nodes_block(node));
2242 ir_node *op = get_Conv_op(node);
2243 ir_node *new_op = be_transform_node(op);
2244 ir_graph *irg = current_ir_graph;
2245 dbg_info *dbgi = get_irn_dbg_info(node);
2246 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
2247 ir_node *nomem = new_NoMem();
2248 ir_mode *mode = get_irn_mode(op);
2249 ir_mode *store_mode;
2250 ir_node *fild, *store;
2254 /* first convert to 32 bit signed if necessary */
2255 src_bits = get_mode_size_bits(src_mode);
2256 if (src_bits == 8) {
2257 new_op = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, block, noreg, noreg, new_op, nomem,
2259 set_ia32_am_support(new_op, ia32_am_Source, ia32_am_unary);
2260 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
2262 } else if (src_bits < 32) {
2263 new_op = new_rd_ia32_Conv_I2I(dbgi, irg, block, noreg, noreg, new_op, nomem, src_mode);
2264 set_ia32_am_support(new_op, ia32_am_Source, ia32_am_unary);
2265 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
2269 assert(get_mode_size_bits(mode) == 32);
2272 store = new_rd_ia32_Store(dbgi, irg, block, get_irg_frame(irg), noreg, new_op, nomem);
2274 set_ia32_use_frame(store);
2275 set_ia32_op_type(store, ia32_AddrModeD);
2276 set_ia32_am_flavour(store, ia32_am_OB);
2277 set_ia32_ls_mode(store, mode_Iu);
2279 /* exception for 32bit unsigned, do a 64bit spill+load */
2280 if(!mode_is_signed(mode)) {
2283 ir_node *zero_const = create_Immediate_from_int(0);
2285 ir_node *zero_store = new_rd_ia32_Store(dbgi, irg, block, get_irg_frame(irg), noreg,
2288 set_ia32_use_frame(zero_store);
2289 set_ia32_op_type(zero_store, ia32_AddrModeD);
2290 add_ia32_am_offs_int(zero_store, 4);
2291 set_ia32_ls_mode(zero_store, mode_Iu);
2296 store = new_rd_Sync(dbgi, irg, block, 2, in);
2297 store_mode = mode_Ls;
2299 store_mode = mode_Is;
2303 fild = new_rd_ia32_vfild(dbgi, irg, block, get_irg_frame(irg), noreg, store);
2305 set_ia32_use_frame(fild);
2306 set_ia32_op_type(fild, ia32_AddrModeS);
2307 set_ia32_am_flavour(fild, ia32_am_OB);
2308 set_ia32_ls_mode(fild, store_mode);
2310 res = new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res);
2316 * Crete a conversion from one integer mode into another one
2318 static ir_node *create_I2I_Conv(ir_mode *src_mode, ir_mode *tgt_mode,
2319 dbg_info *dbgi, ir_node *new_block,
2322 ir_graph *irg = current_ir_graph;
2323 int src_bits = get_mode_size_bits(src_mode);
2324 int tgt_bits = get_mode_size_bits(tgt_mode);
2325 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
2326 ir_node *nomem = new_rd_NoMem(irg);
2328 ir_mode *smaller_mode;
2331 if (src_bits < tgt_bits) {
2332 smaller_mode = src_mode;
2333 smaller_bits = src_bits;
2335 smaller_mode = tgt_mode;
2336 smaller_bits = tgt_bits;
2339 DB((dbg, LEVEL_1, "create Conv(int, int) ...", src_mode, tgt_mode));
2340 if (smaller_bits == 8) {
2341 res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg,
2342 new_op, nomem, smaller_mode);
2344 res = new_rd_ia32_Conv_I2I(dbgi, irg, new_block, noreg, noreg, new_op,
2345 nomem, smaller_mode);
2347 set_ia32_am_support(res, ia32_am_Source, ia32_am_unary);
2353 * Transforms a Conv node.
2355 * @return The created ia32 Conv node
2357 static ir_node *gen_Conv(ir_node *node) {
2358 ir_node *block = be_transform_node(get_nodes_block(node));
2359 ir_node *op = get_Conv_op(node);
2360 ir_node *new_op = be_transform_node(op);
2361 ir_graph *irg = current_ir_graph;
2362 dbg_info *dbgi = get_irn_dbg_info(node);
2363 ir_mode *src_mode = get_irn_mode(op);
2364 ir_mode *tgt_mode = get_irn_mode(node);
2365 int src_bits = get_mode_size_bits(src_mode);
2366 int tgt_bits = get_mode_size_bits(tgt_mode);
2367 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
2368 ir_node *nomem = new_rd_NoMem(irg);
2371 if (src_mode == mode_b) {
2372 assert(mode_is_int(tgt_mode));
2373 /* nothing to do, we already model bools as 0/1 ints */
2377 if (src_mode == tgt_mode) {
2378 if (get_Conv_strict(node)) {
2379 if (USE_SSE2(env_cg)) {
2380 /* when we are in SSE mode, we can kill all strict no-op conversion */
2384 /* this should be optimized already, but who knows... */
2385 DEBUG_ONLY(ir_fprintf(stderr, "Debug warning: conv %+F is pointless\n", node));
2386 DB((dbg, LEVEL_1, "killed Conv(mode, mode) ..."));
2391 if (mode_is_float(src_mode)) {
2392 /* we convert from float ... */
2393 if (mode_is_float(tgt_mode)) {
2394 if(src_mode == mode_E && tgt_mode == mode_D
2395 && !get_Conv_strict(node)) {
2396 DB((dbg, LEVEL_1, "killed Conv(mode, mode) ..."));
2401 if (USE_SSE2(env_cg)) {
2402 DB((dbg, LEVEL_1, "create Conv(float, float) ..."));
2403 res = new_rd_ia32_Conv_FP2FP(dbgi, irg, block, noreg, noreg, new_op, nomem);
2404 set_ia32_ls_mode(res, tgt_mode);
2406 if(get_Conv_strict(node)) {
2407 res = create_strict_conv(tgt_mode, new_op);
2408 SET_IA32_ORIG_NODE(get_Proj_pred(res), ia32_get_old_node_name(env_cg, node));
2411 DB((dbg, LEVEL_1, "killed Conv(float, float) ..."));
2416 DB((dbg, LEVEL_1, "create Conv(float, int) ..."));
2417 if (USE_SSE2(env_cg)) {
2418 res = new_rd_ia32_Conv_FP2I(dbgi, irg, block, noreg, noreg, new_op, nomem);
2419 set_ia32_ls_mode(res, src_mode);
2421 return gen_x87_fp_to_gp(node);
2425 /* we convert from int ... */
2426 if (mode_is_float(tgt_mode)) {
2428 DB((dbg, LEVEL_1, "create Conv(int, float) ..."));
2429 if (USE_SSE2(env_cg)) {
2430 res = new_rd_ia32_Conv_I2FP(dbgi, irg, block, noreg, noreg, new_op, nomem);
2431 set_ia32_ls_mode(res, tgt_mode);
2432 if(src_bits == 32) {
2433 set_ia32_am_support(res, ia32_am_Source, ia32_am_unary);
2436 res = gen_x87_gp_to_fp(node, src_mode);
2437 if(get_Conv_strict(node)) {
2438 res = create_strict_conv(tgt_mode, res);
2439 SET_IA32_ORIG_NODE(get_Proj_pred(res),
2440 ia32_get_old_node_name(env_cg, node));
2444 } else if(tgt_mode == mode_b) {
2445 /* mode_b lowering already took care that we only have 0/1 values */
2446 DB((dbg, LEVEL_1, "omitting unnecessary Conv(%+F, %+F) ...",
2447 src_mode, tgt_mode));
2451 if (src_bits == tgt_bits) {
2452 DB((dbg, LEVEL_1, "omitting unnecessary Conv(%+F, %+F) ...",
2453 src_mode, tgt_mode));
2457 res = create_I2I_Conv(src_mode, tgt_mode, dbgi, block, new_op);
2461 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
2467 int check_immediate_constraint(long val, char immediate_constraint_type)
2469 switch (immediate_constraint_type) {
2473 return val >= 0 && val <= 32;
2475 return val >= 0 && val <= 63;
2477 return val >= -128 && val <= 127;
2479 return val == 0xff || val == 0xffff;
2481 return val >= 0 && val <= 3;
2483 return val >= 0 && val <= 255;
2485 return val >= 0 && val <= 127;
2489 panic("Invalid immediate constraint found");
2494 ir_node *try_create_Immediate(ir_node *node, char immediate_constraint_type)
2497 tarval *offset = NULL;
2498 int offset_sign = 0;
2500 ir_entity *symconst_ent = NULL;
2501 int symconst_sign = 0;
2503 ir_node *cnst = NULL;
2504 ir_node *symconst = NULL;
2510 mode = get_irn_mode(node);
2511 if(!mode_is_int(mode) && !mode_is_reference(mode)) {
2515 if(is_Minus(node)) {
2517 node = get_Minus_op(node);
2520 if(is_Const(node)) {
2523 offset_sign = minus;
2524 } else if(is_SymConst(node)) {
2527 symconst_sign = minus;
2528 } else if(is_Add(node)) {
2529 ir_node *left = get_Add_left(node);
2530 ir_node *right = get_Add_right(node);
2531 if(is_Const(left) && is_SymConst(right)) {
2534 symconst_sign = minus;
2535 offset_sign = minus;
2536 } else if(is_SymConst(left) && is_Const(right)) {
2539 symconst_sign = minus;
2540 offset_sign = minus;
2542 } else if(is_Sub(node)) {
2543 ir_node *left = get_Sub_left(node);
2544 ir_node *right = get_Sub_right(node);
2545 if(is_Const(left) && is_SymConst(right)) {
2548 symconst_sign = !minus;
2549 offset_sign = minus;
2550 } else if(is_SymConst(left) && is_Const(right)) {
2553 symconst_sign = minus;
2554 offset_sign = !minus;
2561 offset = get_Const_tarval(cnst);
2562 if(tarval_is_long(offset)) {
2563 val = get_tarval_long(offset);
2564 } else if(tarval_is_null(offset)) {
2567 ir_fprintf(stderr, "Optimisation Warning: tarval from %+F is not a "
2572 if(!check_immediate_constraint(val, immediate_constraint_type))
2575 if(symconst != NULL) {
2576 if(immediate_constraint_type != 0) {
2577 /* we need full 32bits for symconsts */
2581 if(get_SymConst_kind(symconst) != symconst_addr_ent)
2583 symconst_ent = get_SymConst_entity(symconst);
2585 if(cnst == NULL && symconst == NULL)
2588 if(offset_sign && offset != NULL) {
2589 offset = tarval_neg(offset);
2592 irg = current_ir_graph;
2593 dbgi = get_irn_dbg_info(node);
2594 block = get_irg_start_block(irg);
2595 res = new_rd_ia32_Immediate(dbgi, irg, block, symconst_ent,
2596 symconst_sign, val);
2597 arch_set_irn_register(env_cg->arch_env, res, &ia32_gp_regs[REG_GP_NOREG]);
2603 ir_node *create_immediate_or_transform(ir_node *node, char immediate_constraint_type)
2605 ir_node *new_node = try_create_Immediate(node, immediate_constraint_type);
2606 if (new_node == NULL) {
2607 new_node = be_transform_node(node);
2612 typedef struct constraint_t constraint_t;
2613 struct constraint_t {
2616 const arch_register_req_t **out_reqs;
2618 const arch_register_req_t *req;
2619 unsigned immediate_possible;
2620 char immediate_type;
2623 void parse_asm_constraint(int pos, constraint_t *constraint, const char *c)
2625 int immediate_possible = 0;
2626 char immediate_type = 0;
2627 unsigned limited = 0;
2628 const arch_register_class_t *cls = NULL;
2630 struct obstack *obst;
2631 arch_register_req_t *req;
2632 unsigned *limited_ptr;
2636 /* TODO: replace all the asserts with nice error messages */
2638 printf("Constraint: %s\n", c);
2648 assert(cls == NULL ||
2649 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2650 cls = &ia32_reg_classes[CLASS_ia32_gp];
2651 limited |= 1 << REG_EAX;
2654 assert(cls == NULL ||
2655 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2656 cls = &ia32_reg_classes[CLASS_ia32_gp];
2657 limited |= 1 << REG_EBX;
2660 assert(cls == NULL ||
2661 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2662 cls = &ia32_reg_classes[CLASS_ia32_gp];
2663 limited |= 1 << REG_ECX;
2666 assert(cls == NULL ||
2667 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2668 cls = &ia32_reg_classes[CLASS_ia32_gp];
2669 limited |= 1 << REG_EDX;
2672 assert(cls == NULL ||
2673 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2674 cls = &ia32_reg_classes[CLASS_ia32_gp];
2675 limited |= 1 << REG_EDI;
2678 assert(cls == NULL ||
2679 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2680 cls = &ia32_reg_classes[CLASS_ia32_gp];
2681 limited |= 1 << REG_ESI;
2684 case 'q': /* q means lower part of the regs only, this makes no
2685 * difference to Q for us (we only assigne whole registers) */
2686 assert(cls == NULL ||
2687 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2688 cls = &ia32_reg_classes[CLASS_ia32_gp];
2689 limited |= 1 << REG_EAX | 1 << REG_EBX | 1 << REG_ECX |
2693 assert(cls == NULL ||
2694 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2695 cls = &ia32_reg_classes[CLASS_ia32_gp];
2696 limited |= 1 << REG_EAX | 1 << REG_EDX;
2699 assert(cls == NULL ||
2700 (cls == &ia32_reg_classes[CLASS_ia32_gp] && limited != 0));
2701 cls = &ia32_reg_classes[CLASS_ia32_gp];
2702 limited |= 1 << REG_EAX | 1 << REG_EBX | 1 << REG_ECX |
2703 1 << REG_EDX | 1 << REG_ESI | 1 << REG_EDI |
2710 assert(cls == NULL);
2711 cls = &ia32_reg_classes[CLASS_ia32_gp];
2717 /* TODO: mark values so the x87 simulator knows about t and u */
2718 assert(cls == NULL);
2719 cls = &ia32_reg_classes[CLASS_ia32_vfp];
2724 assert(cls == NULL);
2725 /* TODO: check that sse2 is supported */
2726 cls = &ia32_reg_classes[CLASS_ia32_xmm];
2736 assert(!immediate_possible);
2737 immediate_possible = 1;
2738 immediate_type = *c;
2742 assert(!immediate_possible);
2743 immediate_possible = 1;
2747 assert(!immediate_possible && cls == NULL);
2748 immediate_possible = 1;
2749 cls = &ia32_reg_classes[CLASS_ia32_gp];
2762 assert(constraint->is_in && "can only specify same constraint "
2765 sscanf(c, "%d%n", &same_as, &p);
2772 case 'E': /* no float consts yet */
2773 case 'F': /* no float consts yet */
2774 case 's': /* makes no sense on x86 */
2775 case 'X': /* we can't support that in firm */
2779 case '<': /* no autodecrement on x86 */
2780 case '>': /* no autoincrement on x86 */
2781 case 'C': /* sse constant not supported yet */
2782 case 'G': /* 80387 constant not supported yet */
2783 case 'y': /* we don't support mmx registers yet */
2784 case 'Z': /* not available in 32 bit mode */
2785 case 'e': /* not available in 32 bit mode */
2786 assert(0 && "asm constraint not supported");
2789 assert(0 && "unknown asm constraint found");
2796 const arch_register_req_t *other_constr;
2798 assert(cls == NULL && "same as and register constraint not supported");
2799 assert(!immediate_possible && "same as and immediate constraint not "
2801 assert(same_as < constraint->n_outs && "wrong constraint number in "
2802 "same_as constraint");
2804 other_constr = constraint->out_reqs[same_as];
2806 req = obstack_alloc(obst, sizeof(req[0]));
2807 req->cls = other_constr->cls;
2808 req->type = arch_register_req_type_should_be_same;
2809 req->limited = NULL;
2810 req->other_same = pos;
2811 req->other_different = -1;
2813 /* switch constraints. This is because in firm we have same_as
2814 * constraints on the output constraints while in the gcc asm syntax
2815 * they are specified on the input constraints */
2816 constraint->req = other_constr;
2817 constraint->out_reqs[same_as] = req;
2818 constraint->immediate_possible = 0;
2822 if(immediate_possible && cls == NULL) {
2823 cls = &ia32_reg_classes[CLASS_ia32_gp];
2825 assert(!immediate_possible || cls == &ia32_reg_classes[CLASS_ia32_gp]);
2826 assert(cls != NULL);
2828 if(immediate_possible) {
2829 assert(constraint->is_in
2830 && "imeediates make no sense for output constraints");
2832 /* todo: check types (no float input on 'r' constrainted in and such... */
2834 irg = current_ir_graph;
2835 obst = get_irg_obstack(irg);
2838 req = obstack_alloc(obst, sizeof(req[0]) + sizeof(unsigned));
2839 limited_ptr = (unsigned*) (req+1);
2841 req = obstack_alloc(obst, sizeof(req[0]));
2843 memset(req, 0, sizeof(req[0]));
2846 req->type = arch_register_req_type_limited;
2847 *limited_ptr = limited;
2848 req->limited = limited_ptr;
2850 req->type = arch_register_req_type_normal;
2854 constraint->req = req;
2855 constraint->immediate_possible = immediate_possible;
2856 constraint->immediate_type = immediate_type;
2860 void parse_clobber(ir_node *node, int pos, constraint_t *constraint,
2867 panic("Clobbers not supported yet");
2870 ir_node *gen_ASM(ir_node *node)
2873 ir_graph *irg = current_ir_graph;
2874 ir_node *block = be_transform_node(get_nodes_block(node));
2875 dbg_info *dbgi = get_irn_dbg_info(node);
2882 ia32_asm_attr_t *attr;
2883 const arch_register_req_t **out_reqs;
2884 const arch_register_req_t **in_reqs;
2885 struct obstack *obst;
2886 constraint_t parsed_constraint;
2888 /* transform inputs */
2889 arity = get_irn_arity(node);
2890 in = alloca(arity * sizeof(in[0]));
2891 memset(in, 0, arity * sizeof(in[0]));
2893 n_outs = get_ASM_n_output_constraints(node);
2894 n_clobbers = get_ASM_n_clobbers(node);
2895 out_arity = n_outs + n_clobbers;
2897 /* construct register constraints */
2898 obst = get_irg_obstack(irg);
2899 out_reqs = obstack_alloc(obst, out_arity * sizeof(out_reqs[0]));
2900 parsed_constraint.out_reqs = out_reqs;
2901 parsed_constraint.n_outs = n_outs;
2902 parsed_constraint.is_in = 0;
2903 for(i = 0; i < out_arity; ++i) {
2907 const ir_asm_constraint *constraint;
2908 constraint = & get_ASM_output_constraints(node) [i];
2909 c = get_id_str(constraint->constraint);
2910 parse_asm_constraint(i, &parsed_constraint, c);
2912 ident *glob_id = get_ASM_clobbers(node) [i - n_outs];
2913 c = get_id_str(glob_id);
2914 parse_clobber(node, i, &parsed_constraint, c);
2916 out_reqs[i] = parsed_constraint.req;
2919 in_reqs = obstack_alloc(obst, arity * sizeof(in_reqs[0]));
2920 parsed_constraint.is_in = 1;
2921 for(i = 0; i < arity; ++i) {
2922 const ir_asm_constraint *constraint;
2926 constraint = & get_ASM_input_constraints(node) [i];
2927 constr_id = constraint->constraint;
2928 c = get_id_str(constr_id);
2929 parse_asm_constraint(i, &parsed_constraint, c);
2930 in_reqs[i] = parsed_constraint.req;
2932 if(parsed_constraint.immediate_possible) {
2933 ir_node *pred = get_irn_n(node, i);
2934 char imm_type = parsed_constraint.immediate_type;
2935 ir_node *immediate = try_create_Immediate(pred, imm_type);
2937 if(immediate != NULL) {
2943 /* transform inputs */
2944 for(i = 0; i < arity; ++i) {
2946 ir_node *transformed;
2951 pred = get_irn_n(node, i);
2952 transformed = be_transform_node(pred);
2953 in[i] = transformed;
2956 res = new_rd_ia32_Asm(dbgi, irg, block, arity, in, out_arity);
2958 generic_attr = get_irn_generic_attr(res);
2959 attr = CAST_IA32_ATTR(ia32_asm_attr_t, generic_attr);
2960 attr->asm_text = get_ASM_text(node);
2961 set_ia32_out_req_all(res, out_reqs);
2962 set_ia32_in_req_all(res, in_reqs);
2964 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
2969 /********************************************
2972 * | |__ ___ _ __ ___ __| | ___ ___
2973 * | '_ \ / _ \ '_ \ / _ \ / _` |/ _ \/ __|
2974 * | |_) | __/ | | | (_) | (_| | __/\__ \
2975 * |_.__/ \___|_| |_|\___/ \__,_|\___||___/
2977 ********************************************/
2979 static ir_node *gen_be_StackParam(ir_node *node) {
2980 ir_node *block = be_transform_node(get_nodes_block(node));
2981 ir_node *ptr = get_irn_n(node, be_pos_StackParam_ptr);
2982 ir_node *new_ptr = be_transform_node(ptr);
2983 ir_node *new_op = NULL;
2984 ir_graph *irg = current_ir_graph;
2985 dbg_info *dbgi = get_irn_dbg_info(node);
2986 ir_node *nomem = new_rd_NoMem(current_ir_graph);
2987 ir_entity *ent = arch_get_frame_entity(env_cg->arch_env, node);
2988 ir_mode *load_mode = get_irn_mode(node);
2989 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
2993 if (mode_is_float(load_mode)) {
2994 if (USE_SSE2(env_cg)) {
2995 new_op = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, nomem);
2996 pn_res = pn_ia32_xLoad_res;
2997 proj_mode = mode_xmm;
2999 new_op = new_rd_ia32_vfld(dbgi, irg, block, new_ptr, noreg, nomem, load_mode);
3000 pn_res = pn_ia32_vfld_res;
3001 proj_mode = mode_vfp;
3004 new_op = new_rd_ia32_Load(dbgi, irg, block, new_ptr, noreg, nomem);
3005 proj_mode = mode_Iu;
3006 pn_res = pn_ia32_Load_res;
3009 set_irn_pinned(new_op, op_pin_state_floats);
3010 set_ia32_frame_ent(new_op, ent);
3011 set_ia32_use_frame(new_op);
3013 set_ia32_op_type(new_op, ia32_AddrModeS);
3014 set_ia32_am_flavour(new_op, ia32_am_B);
3015 set_ia32_ls_mode(new_op, load_mode);
3016 set_ia32_flags(new_op, get_ia32_flags(new_op) | arch_irn_flags_rematerializable);
3018 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3020 return new_rd_Proj(dbgi, irg, block, new_op, proj_mode, pn_res);
3024 * Transforms a FrameAddr into an ia32 Add.
3026 static ir_node *gen_be_FrameAddr(ir_node *node) {
3027 ir_node *block = be_transform_node(get_nodes_block(node));
3028 ir_node *op = be_get_FrameAddr_frame(node);
3029 ir_node *new_op = be_transform_node(op);
3030 ir_graph *irg = current_ir_graph;
3031 dbg_info *dbgi = get_irn_dbg_info(node);
3032 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3035 res = new_rd_ia32_Lea(dbgi, irg, block, new_op, noreg);
3036 set_ia32_frame_ent(res, arch_get_frame_entity(env_cg->arch_env, node));
3037 set_ia32_use_frame(res);
3038 set_ia32_am_flavour(res, ia32_am_OB);
3040 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
3046 * Transforms a FrameLoad into an ia32 Load.
3048 static ir_node *gen_be_FrameLoad(ir_node *node) {
3049 ir_node *block = be_transform_node(get_nodes_block(node));
3050 ir_node *mem = get_irn_n(node, be_pos_FrameLoad_mem);
3051 ir_node *new_mem = be_transform_node(mem);
3052 ir_node *ptr = get_irn_n(node, be_pos_FrameLoad_ptr);
3053 ir_node *new_ptr = be_transform_node(ptr);
3054 ir_node *new_op = NULL;
3055 ir_graph *irg = current_ir_graph;
3056 dbg_info *dbgi = get_irn_dbg_info(node);
3057 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3058 ir_entity *ent = arch_get_frame_entity(env_cg->arch_env, node);
3059 ir_mode *mode = get_type_mode(get_entity_type(ent));
3060 ir_node *projs[pn_Load_max];
3062 ia32_collect_Projs(node, projs, pn_Load_max);
3064 if (mode_is_float(mode)) {
3065 if (USE_SSE2(env_cg)) {
3066 new_op = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, new_mem);
3069 new_op = new_rd_ia32_vfld(dbgi, irg, block, new_ptr, noreg, new_mem, mode);
3073 new_op = new_rd_ia32_Load(dbgi, irg, block, new_ptr, noreg, new_mem);
3076 set_irn_pinned(new_op, op_pin_state_floats);
3077 set_ia32_frame_ent(new_op, ent);
3078 set_ia32_use_frame(new_op);
3080 set_ia32_op_type(new_op, ia32_AddrModeS);
3081 set_ia32_am_flavour(new_op, ia32_am_B);
3082 set_ia32_ls_mode(new_op, mode);
3083 set_ia32_flags(new_op, get_ia32_flags(new_op) | arch_irn_flags_rematerializable);
3085 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3092 * Transforms a FrameStore into an ia32 Store.
3094 static ir_node *gen_be_FrameStore(ir_node *node) {
3095 ir_node *block = be_transform_node(get_nodes_block(node));
3096 ir_node *mem = get_irn_n(node, be_pos_FrameStore_mem);
3097 ir_node *new_mem = be_transform_node(mem);
3098 ir_node *ptr = get_irn_n(node, be_pos_FrameStore_ptr);
3099 ir_node *new_ptr = be_transform_node(ptr);
3100 ir_node *val = get_irn_n(node, be_pos_FrameStore_val);
3101 ir_node *new_val = be_transform_node(val);
3102 ir_node *new_op = NULL;
3103 ir_graph *irg = current_ir_graph;
3104 dbg_info *dbgi = get_irn_dbg_info(node);
3105 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3106 ir_entity *ent = arch_get_frame_entity(env_cg->arch_env, node);
3107 ir_mode *mode = get_irn_mode(val);
3109 if (mode_is_float(mode)) {
3110 if (USE_SSE2(env_cg)) {
3111 new_op = new_rd_ia32_xStore(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
3113 new_op = new_rd_ia32_vfst(dbgi, irg, block, new_ptr, noreg, new_val, new_mem, mode);
3115 } else if (get_mode_size_bits(mode) == 8) {
3116 new_op = new_rd_ia32_Store8Bit(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
3118 new_op = new_rd_ia32_Store(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
3121 set_ia32_frame_ent(new_op, ent);
3122 set_ia32_use_frame(new_op);
3124 set_ia32_op_type(new_op, ia32_AddrModeD);
3125 set_ia32_am_flavour(new_op, ia32_am_B);
3126 set_ia32_ls_mode(new_op, mode);
3128 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3134 * In case SSE is used we need to copy the result from XMM0 to FPU TOS before return.
3136 static ir_node *gen_be_Return(ir_node *node) {
3137 ir_graph *irg = current_ir_graph;
3138 ir_node *ret_val = get_irn_n(node, be_pos_Return_val);
3139 ir_node *ret_mem = get_irn_n(node, be_pos_Return_mem);
3140 ir_entity *ent = get_irg_entity(irg);
3141 ir_type *tp = get_entity_type(ent);
3146 ir_node *frame, *sse_store, *fld, *mproj, *barrier;
3147 ir_node *new_barrier, *new_ret_val, *new_ret_mem;
3150 int pn_ret_val, pn_ret_mem, arity, i;
3152 assert(ret_val != NULL);
3153 if (be_Return_get_n_rets(node) < 1 || ! USE_SSE2(env_cg)) {
3154 return be_duplicate_node(node);
3157 res_type = get_method_res_type(tp, 0);
3159 if (! is_Primitive_type(res_type)) {
3160 return be_duplicate_node(node);
3163 mode = get_type_mode(res_type);
3164 if (! mode_is_float(mode)) {
3165 return be_duplicate_node(node);
3168 assert(get_method_n_ress(tp) == 1);
3170 pn_ret_val = get_Proj_proj(ret_val);
3171 pn_ret_mem = get_Proj_proj(ret_mem);
3173 /* get the Barrier */
3174 barrier = get_Proj_pred(ret_val);
3176 /* get result input of the Barrier */
3177 ret_val = get_irn_n(barrier, pn_ret_val);
3178 new_ret_val = be_transform_node(ret_val);
3180 /* get memory input of the Barrier */
3181 ret_mem = get_irn_n(barrier, pn_ret_mem);
3182 new_ret_mem = be_transform_node(ret_mem);
3184 frame = get_irg_frame(irg);
3186 dbgi = get_irn_dbg_info(barrier);
3187 block = be_transform_node(get_nodes_block(barrier));
3189 noreg = ia32_new_NoReg_gp(env_cg);
3191 /* store xmm0 onto stack */
3192 sse_store = new_rd_ia32_xStoreSimple(dbgi, irg, block, frame, noreg,
3193 new_ret_val, new_ret_mem);
3194 set_ia32_ls_mode(sse_store, mode);
3195 set_ia32_op_type(sse_store, ia32_AddrModeD);
3196 set_ia32_use_frame(sse_store);
3197 set_ia32_am_flavour(sse_store, ia32_am_B);
3199 /* load into x87 register */
3200 fld = new_rd_ia32_vfld(dbgi, irg, block, frame, noreg, sse_store, mode);
3201 set_ia32_op_type(fld, ia32_AddrModeS);
3202 set_ia32_use_frame(fld);
3203 set_ia32_am_flavour(fld, ia32_am_B);
3205 mproj = new_r_Proj(irg, block, fld, mode_M, pn_ia32_vfld_M);
3206 fld = new_r_Proj(irg, block, fld, mode_vfp, pn_ia32_vfld_res);
3208 /* create a new barrier */
3209 arity = get_irn_arity(barrier);
3210 in = alloca(arity * sizeof(in[0]));
3211 for (i = 0; i < arity; ++i) {
3214 if (i == pn_ret_val) {
3216 } else if (i == pn_ret_mem) {
3219 ir_node *in = get_irn_n(barrier, i);
3220 new_in = be_transform_node(in);
3225 new_barrier = new_ir_node(dbgi, irg, block,
3226 get_irn_op(barrier), get_irn_mode(barrier),
3228 copy_node_attr(barrier, new_barrier);
3229 be_duplicate_deps(barrier, new_barrier);
3230 be_set_transformed_node(barrier, new_barrier);
3231 mark_irn_visited(barrier);
3233 /* transform normally */
3234 return be_duplicate_node(node);
3238 * Transform a be_AddSP into an ia32_AddSP. Eat up const sizes.
3240 static ir_node *gen_be_AddSP(ir_node *node) {
3241 ir_node *block = be_transform_node(get_nodes_block(node));
3242 ir_node *sz = get_irn_n(node, be_pos_AddSP_size);
3244 ir_node *sp = get_irn_n(node, be_pos_AddSP_old_sp);
3245 ir_node *new_sp = be_transform_node(sp);
3246 ir_graph *irg = current_ir_graph;
3247 dbg_info *dbgi = get_irn_dbg_info(node);
3248 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3249 ir_node *nomem = new_NoMem();
3252 new_sz = create_immediate_or_transform(sz, 0);
3254 /* ia32 stack grows in reverse direction, make a SubSP */
3255 new_op = new_rd_ia32_SubSP(dbgi, irg, block, noreg, noreg, new_sp, new_sz,
3257 set_ia32_am_support(new_op, ia32_am_Source, ia32_am_binary);
3258 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3264 * Transform a be_SubSP into an ia32_SubSP. Eat up const sizes.
3266 static ir_node *gen_be_SubSP(ir_node *node) {
3267 ir_node *block = be_transform_node(get_nodes_block(node));
3268 ir_node *sz = get_irn_n(node, be_pos_SubSP_size);
3270 ir_node *sp = get_irn_n(node, be_pos_SubSP_old_sp);
3271 ir_node *new_sp = be_transform_node(sp);
3272 ir_graph *irg = current_ir_graph;
3273 dbg_info *dbgi = get_irn_dbg_info(node);
3274 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3275 ir_node *nomem = new_NoMem();
3278 new_sz = create_immediate_or_transform(sz, 0);
3280 /* ia32 stack grows in reverse direction, make an AddSP */
3281 new_op = new_rd_ia32_AddSP(dbgi, irg, block, noreg, noreg, new_sp, new_sz, nomem);
3282 set_ia32_am_support(new_op, ia32_am_Source, ia32_am_binary);
3283 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3289 * This function just sets the register for the Unknown node
3290 * as this is not done during register allocation because Unknown
3291 * is an "ignore" node.
3293 static ir_node *gen_Unknown(ir_node *node) {
3294 ir_mode *mode = get_irn_mode(node);
3296 if (mode_is_float(mode)) {
3298 /* Unknown nodes are buggy in x87 sim, use zero for now... */
3299 if (USE_SSE2(env_cg))
3301 return ia32_new_Unknown_vfp(env_cg);
3303 if (!USE_SSE2(env_cg)) {
3304 ir_graph *irg = current_ir_graph;
3305 dbg_info *dbgi = get_irn_dbg_info(node);
3306 ir_node *block = get_irg_start_block(irg);
3307 return new_rd_ia32_vfldz(dbgi, irg, block);
3309 return ia32_new_Unknown_xmm(env_cg);
3312 } else if (mode_needs_gp_reg(mode)) {
3313 return ia32_new_Unknown_gp(env_cg);
3315 assert(0 && "unsupported Unknown-Mode");
3322 * Change some phi modes
3324 static ir_node *gen_Phi(ir_node *node) {
3325 ir_node *block = be_transform_node(get_nodes_block(node));
3326 ir_graph *irg = current_ir_graph;
3327 dbg_info *dbgi = get_irn_dbg_info(node);
3328 ir_mode *mode = get_irn_mode(node);
3331 if(mode_needs_gp_reg(mode)) {
3332 /* we shouldn't have any 64bit stuff around anymore */
3333 assert(get_mode_size_bits(mode) <= 32);
3334 /* all integer operations are on 32bit registers now */
3336 } else if(mode_is_float(mode)) {
3337 if (USE_SSE2(env_cg)) {
3344 /* phi nodes allow loops, so we use the old arguments for now
3345 * and fix this later */
3346 phi = new_ir_node(dbgi, irg, block, op_Phi, mode, get_irn_arity(node), get_irn_in(node) + 1);
3347 copy_node_attr(node, phi);
3348 be_duplicate_deps(node, phi);
3350 be_set_transformed_node(node, phi);
3351 be_enqueue_preds(node);
3359 static ir_node *gen_IJmp(ir_node *node) {
3360 ir_node *block = be_transform_node(get_nodes_block(node));
3361 ir_graph *irg = current_ir_graph;
3362 dbg_info *dbgi = get_irn_dbg_info(node);
3363 ir_node *new_op = be_transform_node(get_IJmp_target(node));
3364 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3365 ir_node *nomem = new_NoMem();
3368 new_node = new_rd_ia32_IJmp(dbgi, irg, block, noreg, noreg, new_op, nomem);
3369 set_ia32_am_support(new_node, ia32_am_Source, ia32_am_unary);
3371 SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
3377 /**********************************************************************
3380 * | | _____ _____ _ __ ___ __| | _ __ ___ __| | ___ ___
3381 * | |/ _ \ \ /\ / / _ \ '__/ _ \/ _` | | '_ \ / _ \ / _` |/ _ \/ __|
3382 * | | (_) \ V V / __/ | | __/ (_| | | | | | (_) | (_| | __/\__ \
3383 * |_|\___/ \_/\_/ \___|_| \___|\__,_| |_| |_|\___/ \__,_|\___||___/
3385 **********************************************************************/
3387 /* These nodes are created in intrinsic lowering (64bit -> 32bit) */
3389 typedef ir_node *construct_load_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \
3392 typedef ir_node *construct_store_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \
3393 ir_node *val, ir_node *mem);
3396 * Transforms a lowered Load into a "real" one.
3398 static ir_node *gen_lowered_Load(ir_node *node, construct_load_func func)
3400 ir_node *block = be_transform_node(get_nodes_block(node));
3401 ir_node *ptr = get_irn_n(node, 0);
3402 ir_node *new_ptr = be_transform_node(ptr);
3403 ir_node *mem = get_irn_n(node, 1);
3404 ir_node *new_mem = be_transform_node(mem);
3405 ir_graph *irg = current_ir_graph;
3406 dbg_info *dbgi = get_irn_dbg_info(node);
3407 ir_mode *mode = get_ia32_ls_mode(node);
3408 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3411 new_op = func(dbgi, irg, block, new_ptr, noreg, new_mem);
3413 set_ia32_op_type(new_op, ia32_AddrModeS);
3414 set_ia32_am_flavour(new_op, ia32_am_OB);
3415 set_ia32_am_offs_int(new_op, 0);
3416 set_ia32_am_scale(new_op, 1);
3417 set_ia32_am_sc(new_op, get_ia32_am_sc(node));
3418 if (is_ia32_am_sc_sign(node))
3419 set_ia32_am_sc_sign(new_op);
3420 set_ia32_ls_mode(new_op, mode);
3421 if (is_ia32_use_frame(node)) {
3422 set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
3423 set_ia32_use_frame(new_op);
3426 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3432 * Transforms a lowered Store into a "real" one.
3434 static ir_node *gen_lowered_Store(ir_node *node, construct_store_func func)
3436 ir_node *block = be_transform_node(get_nodes_block(node));
3437 ir_node *ptr = get_irn_n(node, 0);
3438 ir_node *new_ptr = be_transform_node(ptr);
3439 ir_node *val = get_irn_n(node, 1);
3440 ir_node *new_val = be_transform_node(val);
3441 ir_node *mem = get_irn_n(node, 2);
3442 ir_node *new_mem = be_transform_node(mem);
3443 ir_graph *irg = current_ir_graph;
3444 dbg_info *dbgi = get_irn_dbg_info(node);
3445 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3446 ir_mode *mode = get_ia32_ls_mode(node);
3449 ia32_am_flavour_t am_flav = ia32_B;
3451 new_op = func(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
3453 if ((am_offs = get_ia32_am_offs_int(node)) != 0) {
3455 add_ia32_am_offs_int(new_op, am_offs);
3458 set_ia32_op_type(new_op, ia32_AddrModeD);
3459 set_ia32_am_flavour(new_op, am_flav);
3460 set_ia32_ls_mode(new_op, mode);
3461 set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
3462 set_ia32_use_frame(new_op);
3464 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3471 * Transforms an ia32_l_XXX into a "real" XXX node
3473 * @param env The transformation environment
3474 * @return the created ia32 XXX node
3476 #define GEN_LOWERED_OP(op) \
3477 static ir_node *gen_ia32_l_##op(ir_node *node) { \
3478 return gen_binop(node, get_binop_left(node), \
3479 get_binop_right(node), new_rd_ia32_##op,0); \
3482 #define GEN_LOWERED_x87_OP(op) \
3483 static ir_node *gen_ia32_l_##op(ir_node *node) { \
3485 new_op = gen_binop_x87_float(node, get_binop_left(node), \
3486 get_binop_right(node), new_rd_ia32_##op); \
3490 #define GEN_LOWERED_UNOP(op) \
3491 static ir_node *gen_ia32_l_##op(ir_node *node) {\
3492 return gen_unop(node, get_unop_op(node), new_rd_ia32_##op); \
3495 #define GEN_LOWERED_SHIFT_OP(l_op, op) \
3496 static ir_node *gen_ia32_##l_op(ir_node *node) { \
3497 return gen_shift_binop(node, get_irn_n(node, 0), \
3498 get_irn_n(node, 1), new_rd_ia32_##op); \
3501 #define GEN_LOWERED_LOAD(op) \
3502 static ir_node *gen_ia32_l_##op(ir_node *node) { \
3503 return gen_lowered_Load(node, new_rd_ia32_##op); \
3506 #define GEN_LOWERED_STORE(op) \
3507 static ir_node *gen_ia32_l_##op(ir_node *node) { \
3508 return gen_lowered_Store(node, new_rd_ia32_##op); \
3515 GEN_LOWERED_OP(IMul)
3517 GEN_LOWERED_x87_OP(vfprem)
3518 GEN_LOWERED_x87_OP(vfmul)
3519 GEN_LOWERED_x87_OP(vfsub)
3521 GEN_LOWERED_UNOP(Neg)
3523 GEN_LOWERED_LOAD(vfild)
3524 GEN_LOWERED_LOAD(Load)
3525 GEN_LOWERED_STORE(Store)
3528 * Transforms a l_vfist into a "real" vfist node.
3530 * @param env The transformation environment
3531 * @return the created ia32 vfist node
3533 static ir_node *gen_ia32_l_vfist(ir_node *node) {
3534 ir_node *block = be_transform_node(get_nodes_block(node));
3535 ir_node *ptr = get_irn_n(node, 0);
3536 ir_node *new_ptr = be_transform_node(ptr);
3537 ir_node *val = get_irn_n(node, 1);
3538 ir_node *new_val = be_transform_node(val);
3539 ir_node *mem = get_irn_n(node, 2);
3540 ir_node *new_mem = be_transform_node(mem);
3541 ir_graph *irg = current_ir_graph;
3542 dbg_info *dbgi = get_irn_dbg_info(node);
3543 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3544 ir_mode *mode = get_ia32_ls_mode(node);
3545 ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg);
3548 ia32_am_flavour_t am_flav = ia32_B;
3550 new_op = new_rd_ia32_vfist(dbgi, irg, block, new_ptr, noreg, new_val,
3551 trunc_mode, new_mem);
3553 if ((am_offs = get_ia32_am_offs_int(node)) != 0) {
3555 add_ia32_am_offs_int(new_op, am_offs);
3558 set_ia32_op_type(new_op, ia32_AddrModeD);
3559 set_ia32_am_flavour(new_op, am_flav);
3560 set_ia32_ls_mode(new_op, mode);
3561 set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
3562 set_ia32_use_frame(new_op);
3564 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3570 * Transforms a l_vfdiv into a "real" vfdiv node.
3572 * @param env The transformation environment
3573 * @return the created ia32 vfdiv node
3575 static ir_node *gen_ia32_l_vfdiv(ir_node *node) {
3576 ir_node *block = be_transform_node(get_nodes_block(node));
3577 ir_node *left = get_binop_left(node);
3578 ir_node *new_left = be_transform_node(left);
3579 ir_node *right = get_binop_right(node);
3580 ir_node *new_right = be_transform_node(right);
3581 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3582 ir_graph *irg = current_ir_graph;
3583 dbg_info *dbgi = get_irn_dbg_info(node);
3584 ir_node *fpcw = get_fpcw();
3587 vfdiv = new_rd_ia32_vfdiv(dbgi, irg, block, noreg, noreg, new_left,
3588 new_right, new_NoMem(), fpcw);
3589 clear_ia32_commutative(vfdiv);
3590 set_ia32_am_support(vfdiv, ia32_am_Source, ia32_am_binary);
3592 SET_IA32_ORIG_NODE(vfdiv, ia32_get_old_node_name(env_cg, node));
3598 * Transforms a l_MulS into a "real" MulS node.
3600 * @param env The transformation environment
3601 * @return the created ia32 Mul node
3603 static ir_node *gen_ia32_l_Mul(ir_node *node) {
3604 ir_node *block = be_transform_node(get_nodes_block(node));
3605 ir_node *left = get_binop_left(node);
3606 ir_node *new_left = be_transform_node(left);
3607 ir_node *right = get_binop_right(node);
3608 ir_node *new_right = be_transform_node(right);
3609 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3610 ir_graph *irg = current_ir_graph;
3611 dbg_info *dbgi = get_irn_dbg_info(node);
3613 /* l_Mul is already a mode_T node, so we create the Mul in the normal way */
3614 /* and then skip the result Proj, because all needed Projs are already there. */
3615 ir_node *muls = new_rd_ia32_Mul(dbgi, irg, block, noreg, noreg, new_left,
3616 new_right, new_NoMem());
3617 clear_ia32_commutative(muls);
3618 set_ia32_am_support(muls, ia32_am_Source, ia32_am_binary);
3620 SET_IA32_ORIG_NODE(muls, ia32_get_old_node_name(env_cg, node));
3625 GEN_LOWERED_SHIFT_OP(l_ShlDep, Shl)
3626 GEN_LOWERED_SHIFT_OP(l_ShrDep, Shr)
3627 GEN_LOWERED_SHIFT_OP(l_Sar, Sar)
3628 GEN_LOWERED_SHIFT_OP(l_SarDep, Sar)
3631 * Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
3632 * op1 - target to be shifted
3633 * op2 - contains bits to be shifted into target
3635 * Only op3 can be an immediate.
3637 static ir_node *gen_lowered_64bit_shifts(ir_node *node, ir_node *op1,
3638 ir_node *op2, ir_node *count)
3640 ir_node *block = be_transform_node(get_nodes_block(node));
3641 ir_node *new_op1 = be_transform_node(op1);
3642 ir_node *new_op2 = be_transform_node(op2);
3643 ir_node *new_op = NULL;
3644 ir_node *new_count = be_transform_node(count);
3645 ir_graph *irg = current_ir_graph;
3646 dbg_info *dbgi = get_irn_dbg_info(node);
3647 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
3648 ir_node *nomem = new_NoMem();
3652 assert(! mode_is_float(get_irn_mode(node)) && "Shift/Rotate with float not supported");
3654 /* Check if immediate optimization is on and */
3655 /* if it's an operation with immediate. */
3656 imm_op = (env_cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, new_count) : NULL;
3658 /* Limit imm_op within range imm8 */
3660 tv = get_ia32_Immop_tarval(imm_op);
3663 tv = tarval_mod(tv, new_tarval_from_long(32, get_tarval_mode(tv)));
3664 set_ia32_Immop_tarval(imm_op, tv);
3671 /* integer operations */
3673 /* This is ShiftD with const */
3674 DB((dbg, LEVEL_1, "ShiftD with immediate ..."));
3676 if (is_ia32_l_ShlD(node))
3677 new_op = new_rd_ia32_ShlD(dbgi, irg, block, noreg, noreg,
3678 new_op1, new_op2, noreg, nomem);
3680 new_op = new_rd_ia32_ShrD(dbgi, irg, block, noreg, noreg,
3681 new_op1, new_op2, noreg, nomem);
3682 copy_ia32_Immop_attr(new_op, imm_op);
3685 /* This is a normal ShiftD */
3686 DB((dbg, LEVEL_1, "ShiftD binop ..."));
3687 if (is_ia32_l_ShlD(node))
3688 new_op = new_rd_ia32_ShlD(dbgi, irg, block, noreg, noreg,
3689 new_op1, new_op2, new_count, nomem);
3691 new_op = new_rd_ia32_ShrD(dbgi, irg, block, noreg, noreg,
3692 new_op1, new_op2, new_count, nomem);
3695 /* set AM support */
3696 set_ia32_am_support(new_op, ia32_am_Dest, ia32_am_binary);
3698 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
3700 set_ia32_emit_cl(new_op);
3705 static ir_node *gen_ia32_l_ShlD(ir_node *node) {
3706 return gen_lowered_64bit_shifts(node, get_irn_n(node, 0),
3707 get_irn_n(node, 1), get_irn_n(node, 2));
3710 static ir_node *gen_ia32_l_ShrD(ir_node *node) {
3711 return gen_lowered_64bit_shifts(node, get_irn_n(node, 0),
3712 get_irn_n(node, 1), get_irn_n(node, 2));
3716 * In case SSE Unit is used, the node is transformed into a vfst + xLoad.
3718 static ir_node *gen_ia32_l_X87toSSE(ir_node *node) {
3719 ir_node *block = be_transform_node(get_nodes_block(node));
3720 ir_node *val = get_irn_n(node, 1);
3721 ir_node *new_val = be_transform_node(val);
3722 ia32_code_gen_t *cg = env_cg;
3723 ir_node *res = NULL;
3724 ir_graph *irg = current_ir_graph;
3726 ir_node *noreg, *new_ptr, *new_mem;
3733 mem = get_irn_n(node, 2);
3734 new_mem = be_transform_node(mem);
3735 ptr = get_irn_n(node, 0);
3736 new_ptr = be_transform_node(ptr);
3737 noreg = ia32_new_NoReg_gp(cg);
3738 dbgi = get_irn_dbg_info(node);
3740 /* Store x87 -> MEM */
3741 res = new_rd_ia32_vfst(dbgi, irg, block, new_ptr, noreg, new_val, new_mem, get_ia32_ls_mode(node));
3742 set_ia32_frame_ent(res, get_ia32_frame_ent(node));
3743 set_ia32_use_frame(res);
3744 set_ia32_ls_mode(res, get_ia32_ls_mode(node));
3745 set_ia32_am_flavour(res, ia32_B);
3746 set_ia32_op_type(res, ia32_AddrModeD);
3748 /* Load MEM -> SSE */
3749 res = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, res);
3750 set_ia32_frame_ent(res, get_ia32_frame_ent(node));
3751 set_ia32_use_frame(res);
3752 set_ia32_ls_mode(res, get_ia32_ls_mode(node));
3753 set_ia32_am_flavour(res, ia32_B);
3754 set_ia32_op_type(res, ia32_AddrModeS);
3755 res = new_rd_Proj(dbgi, irg, block, res, mode_xmm, pn_ia32_xLoad_res);
3761 * In case SSE Unit is used, the node is transformed into a xStore + vfld.
3763 static ir_node *gen_ia32_l_SSEtoX87(ir_node *node) {
3764 ir_node *block = be_transform_node(get_nodes_block(node));
3765 ir_node *val = get_irn_n(node, 1);
3766 ir_node *new_val = be_transform_node(val);
3767 ia32_code_gen_t *cg = env_cg;
3768 ir_graph *irg = current_ir_graph;
3769 ir_node *res = NULL;
3770 ir_entity *fent = get_ia32_frame_ent(node);
3771 ir_mode *lsmode = get_ia32_ls_mode(node);
3773 ir_node *noreg, *new_ptr, *new_mem;
3777 if (! USE_SSE2(cg)) {
3778 /* SSE unit is not used -> skip this node. */
3782 ptr = get_irn_n(node, 0);
3783 new_ptr = be_transform_node(ptr);
3784 mem = get_irn_n(node, 2);
3785 new_mem = be_transform_node(mem);
3786 noreg = ia32_new_NoReg_gp(cg);
3787 dbgi = get_irn_dbg_info(node);
3789 /* Store SSE -> MEM */
3790 if (is_ia32_xLoad(skip_Proj(new_val))) {
3791 ir_node *ld = skip_Proj(new_val);
3793 /* we can vfld the value directly into the fpu */
3794 fent = get_ia32_frame_ent(ld);
3795 ptr = get_irn_n(ld, 0);
3796 offs = get_ia32_am_offs_int(ld);
3798 res = new_rd_ia32_xStore(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
3799 set_ia32_frame_ent(res, fent);
3800 set_ia32_use_frame(res);
3801 set_ia32_ls_mode(res, lsmode);
3802 set_ia32_am_flavour(res, ia32_B);
3803 set_ia32_op_type(res, ia32_AddrModeD);
3807 /* Load MEM -> x87 */
3808 res = new_rd_ia32_vfld(dbgi, irg, block, new_ptr, noreg, new_mem, lsmode);
3809 set_ia32_frame_ent(res, fent);
3810 set_ia32_use_frame(res);
3811 add_ia32_am_offs_int(res, offs);
3812 set_ia32_am_flavour(res, ia32_B);
3813 set_ia32_op_type(res, ia32_AddrModeS);
3814 res = new_rd_Proj(dbgi, irg, block, res, mode_vfp, pn_ia32_vfld_res);
3819 /*********************************************************
3822 * _ __ ___ __ _ _ _ __ __| |_ __ ___ _____ _ __
3823 * | '_ ` _ \ / _` | | '_ \ / _` | '__| \ \ / / _ \ '__|
3824 * | | | | | | (_| | | | | | | (_| | | | |\ V / __/ |
3825 * |_| |_| |_|\__,_|_|_| |_| \__,_|_| |_| \_/ \___|_|
3827 *********************************************************/
3830 * the BAD transformer.
3832 static ir_node *bad_transform(ir_node *node) {
3833 panic("No transform function for %+F available.\n", node);
3838 * Transform the Projs of an AddSP.
3840 static ir_node *gen_Proj_be_AddSP(ir_node *node) {
3841 ir_node *block = be_transform_node(get_nodes_block(node));
3842 ir_node *pred = get_Proj_pred(node);
3843 ir_node *new_pred = be_transform_node(pred);
3844 ir_graph *irg = current_ir_graph;
3845 dbg_info *dbgi = get_irn_dbg_info(node);
3846 long proj = get_Proj_proj(node);
3848 if (proj == pn_be_AddSP_sp) {
3849 ir_node *res = new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
3850 pn_ia32_SubSP_stack);
3851 arch_set_irn_register(env_cg->arch_env, res, &ia32_gp_regs[REG_ESP]);
3853 } else if(proj == pn_be_AddSP_res) {
3854 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
3855 pn_ia32_SubSP_addr);
3856 } else if (proj == pn_be_AddSP_M) {
3857 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_SubSP_M);
3861 return new_rd_Unknown(irg, get_irn_mode(node));
3865 * Transform the Projs of a SubSP.
3867 static ir_node *gen_Proj_be_SubSP(ir_node *node) {
3868 ir_node *block = be_transform_node(get_nodes_block(node));
3869 ir_node *pred = get_Proj_pred(node);
3870 ir_node *new_pred = be_transform_node(pred);
3871 ir_graph *irg = current_ir_graph;
3872 dbg_info *dbgi = get_irn_dbg_info(node);
3873 long proj = get_Proj_proj(node);
3875 if (proj == pn_be_SubSP_sp) {
3876 ir_node *res = new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
3877 pn_ia32_AddSP_stack);
3878 arch_set_irn_register(env_cg->arch_env, res, &ia32_gp_regs[REG_ESP]);
3880 } else if (proj == pn_be_SubSP_M) {
3881 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_AddSP_M);
3885 return new_rd_Unknown(irg, get_irn_mode(node));
3889 * Transform and renumber the Projs from a Load.
3891 static ir_node *gen_Proj_Load(ir_node *node) {
3892 ir_node *block = be_transform_node(get_nodes_block(node));
3893 ir_node *pred = get_Proj_pred(node);
3894 ir_node *new_pred = be_transform_node(pred);
3895 ir_graph *irg = current_ir_graph;
3896 dbg_info *dbgi = get_irn_dbg_info(node);
3897 long proj = get_Proj_proj(node);
3899 /* renumber the proj */
3900 if (is_ia32_Load(new_pred)) {
3901 if (proj == pn_Load_res) {
3902 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Load_res);
3903 } else if (proj == pn_Load_M) {
3904 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Load_M);
3906 } else if(is_ia32_Conv_I2I(new_pred)) {
3907 set_irn_mode(new_pred, mode_T);
3908 if (proj == pn_Load_res) {
3909 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, 0);
3910 } else if (proj == pn_Load_M) {
3911 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, 1);
3913 } else if (is_ia32_xLoad(new_pred)) {
3914 if (proj == pn_Load_res) {
3915 return new_rd_Proj(dbgi, irg, block, new_pred, mode_xmm, pn_ia32_xLoad_res);
3916 } else if (proj == pn_Load_M) {
3917 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_xLoad_M);
3919 } else if (is_ia32_vfld(new_pred)) {
3920 if (proj == pn_Load_res) {
3921 return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfld_res);
3922 } else if (proj == pn_Load_M) {
3923 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_vfld_M);
3928 return new_rd_Unknown(irg, get_irn_mode(node));
3932 * Transform and renumber the Projs from a DivMod like instruction.
3934 static ir_node *gen_Proj_DivMod(ir_node *node) {
3935 ir_node *block = be_transform_node(get_nodes_block(node));
3936 ir_node *pred = get_Proj_pred(node);
3937 ir_node *new_pred = be_transform_node(pred);
3938 ir_graph *irg = current_ir_graph;
3939 dbg_info *dbgi = get_irn_dbg_info(node);
3940 ir_mode *mode = get_irn_mode(node);
3941 long proj = get_Proj_proj(node);
3943 assert(is_ia32_Div(new_pred) || is_ia32_IDiv(new_pred));
3945 switch (get_irn_opcode(pred)) {
3949 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Div_M);
3951 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_div_res);
3959 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Div_M);
3961 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_mod_res);
3969 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Div_M);
3970 case pn_DivMod_res_div:
3971 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_div_res);
3972 case pn_DivMod_res_mod:
3973 return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_mod_res);
3983 return new_rd_Unknown(irg, mode);
3987 * Transform and renumber the Projs from a CopyB.
3989 static ir_node *gen_Proj_CopyB(ir_node *node) {
3990 ir_node *block = be_transform_node(get_nodes_block(node));
3991 ir_node *pred = get_Proj_pred(node);
3992 ir_node *new_pred = be_transform_node(pred);
3993 ir_graph *irg = current_ir_graph;
3994 dbg_info *dbgi = get_irn_dbg_info(node);
3995 ir_mode *mode = get_irn_mode(node);
3996 long proj = get_Proj_proj(node);
3999 case pn_CopyB_M_regular:
4000 if (is_ia32_CopyB_i(new_pred)) {
4001 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_CopyB_i_M);
4002 } else if (is_ia32_CopyB(new_pred)) {
4003 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_CopyB_M);
4011 return new_rd_Unknown(irg, mode);
4015 * Transform and renumber the Projs from a vfdiv.
4017 static ir_node *gen_Proj_l_vfdiv(ir_node *node) {
4018 ir_node *block = be_transform_node(get_nodes_block(node));
4019 ir_node *pred = get_Proj_pred(node);
4020 ir_node *new_pred = be_transform_node(pred);
4021 ir_graph *irg = current_ir_graph;
4022 dbg_info *dbgi = get_irn_dbg_info(node);
4023 ir_mode *mode = get_irn_mode(node);
4024 long proj = get_Proj_proj(node);
4027 case pn_ia32_l_vfdiv_M:
4028 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_vfdiv_M);
4029 case pn_ia32_l_vfdiv_res:
4030 return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfdiv_res);
4035 return new_rd_Unknown(irg, mode);
4039 * Transform and renumber the Projs from a Quot.
4041 static ir_node *gen_Proj_Quot(ir_node *node) {
4042 ir_node *block = be_transform_node(get_nodes_block(node));
4043 ir_node *pred = get_Proj_pred(node);
4044 ir_node *new_pred = be_transform_node(pred);
4045 ir_graph *irg = current_ir_graph;
4046 dbg_info *dbgi = get_irn_dbg_info(node);
4047 ir_mode *mode = get_irn_mode(node);
4048 long proj = get_Proj_proj(node);
4052 if (is_ia32_xDiv(new_pred)) {
4053 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_xDiv_M);
4054 } else if (is_ia32_vfdiv(new_pred)) {
4055 return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_vfdiv_M);
4059 if (is_ia32_xDiv(new_pred)) {
4060 return new_rd_Proj(dbgi, irg, block, new_pred, mode_xmm, pn_ia32_xDiv_res);
4061 } else if (is_ia32_vfdiv(new_pred)) {
4062 return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfdiv_res);
4070 return new_rd_Unknown(irg, mode);
4074 * Transform the Thread Local Storage Proj.
4076 static ir_node *gen_Proj_tls(ir_node *node) {
4077 ir_node *block = be_transform_node(get_nodes_block(node));
4078 ir_graph *irg = current_ir_graph;
4079 dbg_info *dbgi = NULL;
4080 ir_node *res = new_rd_ia32_LdTls(dbgi, irg, block, mode_Iu);
4086 * Transform the Projs from a be_Call.
4088 static ir_node *gen_Proj_be_Call(ir_node *node) {
4089 ir_node *block = be_transform_node(get_nodes_block(node));
4090 ir_node *call = get_Proj_pred(node);
4091 ir_node *new_call = be_transform_node(call);
4092 ir_graph *irg = current_ir_graph;
4093 dbg_info *dbgi = get_irn_dbg_info(node);
4094 ir_type *method_type = be_Call_get_type(call);
4095 int n_res = get_method_n_ress(method_type);
4096 long proj = get_Proj_proj(node);
4097 ir_mode *mode = get_irn_mode(node);
4099 const arch_register_class_t *cls;
4101 /* The following is kinda tricky: If we're using SSE, then we have to
4102 * move the result value of the call in floating point registers to an
4103 * xmm register, we therefore construct a GetST0 -> xLoad sequence
4104 * after the call, we have to make sure to correctly make the
4105 * MemProj and the result Proj use these 2 nodes
4107 if (proj == pn_be_Call_M_regular) {
4108 // get new node for result, are we doing the sse load/store hack?
4109 ir_node *call_res = be_get_Proj_for_pn(call, pn_be_Call_first_res);
4110 ir_node *call_res_new;
4111 ir_node *call_res_pred = NULL;
4113 if (call_res != NULL) {
4114 call_res_new = be_transform_node(call_res);
4115 call_res_pred = get_Proj_pred(call_res_new);
4118 if (call_res_pred == NULL || be_is_Call(call_res_pred)) {
4119 return new_rd_Proj(dbgi, irg, block, new_call, mode_M,
4120 pn_be_Call_M_regular);
4122 assert(is_ia32_xLoad(call_res_pred));
4123 return new_rd_Proj(dbgi, irg, block, call_res_pred, mode_M,
4127 if (USE_SSE2(env_cg) && proj >= pn_be_Call_first_res
4128 && proj < (pn_be_Call_first_res + n_res) && mode_is_float(mode)
4129 && USE_SSE2(env_cg)) {
4131 ir_node *frame = get_irg_frame(irg);
4132 ir_node *noreg = ia32_new_NoReg_gp(env_cg);
4134 ir_node *call_mem = be_get_Proj_for_pn(call, pn_be_Call_M_regular);
4137 /* in case there is no memory output: create one to serialize the copy
4139 call_mem = new_rd_Proj(dbgi, irg, block, new_call, mode_M,
4140 pn_be_Call_M_regular);
4141 call_res = new_rd_Proj(dbgi, irg, block, new_call, mode,
4142 pn_be_Call_first_res);
4144 /* store st(0) onto stack */
4145 fstp = new_rd_ia32_vfst(dbgi, irg, block, frame, noreg, call_res,
4147 set_ia32_op_type(fstp, ia32_AddrModeD);
4148 set_ia32_use_frame(fstp);
4149 set_ia32_am_flavour(fstp, ia32_am_B);
4151 /* load into SSE register */
4152 sse_load = new_rd_ia32_xLoad(dbgi, irg, block, frame, noreg, fstp);
4153 set_ia32_ls_mode(sse_load, mode);
4154 set_ia32_op_type(sse_load, ia32_AddrModeS);
4155 set_ia32_use_frame(sse_load);
4156 set_ia32_am_flavour(sse_load, ia32_am_B);
4158 sse_load = new_rd_Proj(dbgi, irg, block, sse_load, mode_xmm,
4162 /* now: create new Keep whith all former ins and one additional in - the result Proj */
4164 /* get a Proj representing a caller save register */
4165 p = be_get_Proj_for_pn(call, pn_be_Call_first_res + 1);
4166 assert(is_Proj(p) && "Proj expected.");
4168 /* user of the the proj is the Keep */
4169 p = get_edge_src_irn(get_irn_out_edge_first(p));
4170 assert(be_is_Keep(p) && "Keep expected.");
4176 /* transform call modes */
4177 if (mode_is_data(mode)) {
4178 cls = arch_get_irn_reg_class(env_cg->arch_env, node, -1);
4182 return new_rd_Proj(dbgi, irg, block, new_call, mode, proj);
4186 * Transform the Projs from a Cmp.
4188 static ir_node *gen_Proj_Cmp(ir_node *node)
4190 /* normally Cmps are processed when looking at Cond nodes, but this case
4191 * can happen in complicated Psi conditions */
4193 ir_node *cmp = get_Proj_pred(node);
4194 long pnc = get_Proj_proj(node);
4195 ir_node *cmp_left = get_Cmp_left(cmp);
4196 ir_node *cmp_right = get_Cmp_right(cmp);
4197 ir_mode *cmp_mode = get_irn_mode(cmp_left);
4198 dbg_info *dbgi = get_irn_dbg_info(cmp);
4199 ir_node *block = be_transform_node(get_nodes_block(node));
4202 assert(!mode_is_float(cmp_mode));
4204 if(!mode_is_signed(cmp_mode)) {
4205 pnc |= ia32_pn_Cmp_Unsigned;
4208 res = create_set(pnc, cmp_left, cmp_right, dbgi, block);
4209 SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, cmp));
4215 * Transform and potentially renumber Proj nodes.
4217 static ir_node *gen_Proj(ir_node *node) {
4218 ir_graph *irg = current_ir_graph;
4219 dbg_info *dbgi = get_irn_dbg_info(node);
4220 ir_node *pred = get_Proj_pred(node);
4221 long proj = get_Proj_proj(node);
4223 if (is_Store(pred) || be_is_FrameStore(pred)) {
4224 if (proj == pn_Store_M) {
4225 return be_transform_node(pred);
4228 return new_r_Bad(irg);
4230 } else if (is_Load(pred) || be_is_FrameLoad(pred)) {
4231 return gen_Proj_Load(node);
4232 } else if (is_Div(pred) || is_Mod(pred) || is_DivMod(pred)) {
4233 return gen_Proj_DivMod(node);
4234 } else if (is_CopyB(pred)) {
4235 return gen_Proj_CopyB(node);
4236 } else if (is_Quot(pred)) {
4237 return gen_Proj_Quot(node);
4238 } else if (is_ia32_l_vfdiv(pred)) {
4239 return gen_Proj_l_vfdiv(node);
4240 } else if (be_is_SubSP(pred)) {
4241 return gen_Proj_be_SubSP(node);
4242 } else if (be_is_AddSP(pred)) {
4243 return gen_Proj_be_AddSP(node);
4244 } else if (be_is_Call(pred)) {
4245 return gen_Proj_be_Call(node);
4246 } else if (is_Cmp(pred)) {
4247 return gen_Proj_Cmp(node);
4248 } else if (get_irn_op(pred) == op_Start) {
4249 if (proj == pn_Start_X_initial_exec) {
4250 ir_node *block = get_nodes_block(pred);
4253 /* we exchange the ProjX with a jump */
4254 block = be_transform_node(block);
4255 jump = new_rd_Jmp(dbgi, irg, block);
4258 if (node == be_get_old_anchor(anchor_tls)) {
4259 return gen_Proj_tls(node);
4262 } else if(!is_ia32_irn(pred)) { // Quick hack for SIMD optimization
4266 ir_node *new_pred = be_transform_node(pred);
4267 ir_node *block = be_transform_node(get_nodes_block(node));
4268 ir_mode *mode = get_irn_mode(node);
4269 if (mode_needs_gp_reg(mode)) {
4270 ir_node *new_proj = new_r_Proj(irg, block, new_pred, mode_Iu,
4271 get_Proj_proj(node));
4272 #ifdef DEBUG_libfirm
4273 new_proj->node_nr = node->node_nr;
4279 return be_duplicate_node(node);
4283 * Enters all transform functions into the generic pointer
4285 static void register_transformers(void)
4289 /* first clear the generic function pointer for all ops */
4290 clear_irp_opcodes_generic_func();
4292 #define GEN(a) { be_transform_func *func = gen_##a; op_##a->ops.generic = (op_func) func; }
4293 #define BAD(a) op_##a->ops.generic = (op_func)bad_transform
4330 /* transform ops from intrinsic lowering */
4353 GEN(ia32_l_X87toSSE);
4354 GEN(ia32_l_SSEtoX87);
4359 /* we should never see these nodes */
4374 /* handle generic backend nodes */
4385 /* set the register for all Unknown nodes */
4388 op_Mulh = get_op_Mulh();
4397 * Pre-transform all unknown and noreg nodes.
4399 static void ia32_pretransform_node(void *arch_cg) {
4400 ia32_code_gen_t *cg = arch_cg;
4402 cg->unknown_gp = be_pre_transform_node(cg->unknown_gp);
4403 cg->unknown_vfp = be_pre_transform_node(cg->unknown_vfp);
4404 cg->unknown_xmm = be_pre_transform_node(cg->unknown_xmm);
4405 cg->noreg_gp = be_pre_transform_node(cg->noreg_gp);
4406 cg->noreg_vfp = be_pre_transform_node(cg->noreg_vfp);
4407 cg->noreg_xmm = be_pre_transform_node(cg->noreg_xmm);
4412 * Walker, checks if all ia32 nodes producing more than one result have
4413 * its Projs, other wise creates new projs and keep them using a be_Keep node.
4416 void add_missing_keep_walker(ir_node *node, void *data)
4419 unsigned found_projs = 0;
4420 const ir_edge_t *edge;
4421 ir_mode *mode = get_irn_mode(node);
4426 if(!is_ia32_irn(node))
4429 n_outs = get_ia32_n_res(node);
4432 if(is_ia32_SwitchJmp(node))
4435 assert(n_outs < (int) sizeof(unsigned) * 8);
4436 foreach_out_edge(node, edge) {
4437 ir_node *proj = get_edge_src_irn(edge);
4438 int pn = get_Proj_proj(proj);
4440 assert(pn < n_outs);
4441 found_projs |= 1 << pn;
4445 /* are keeps missing? */
4447 for(i = 0; i < n_outs; ++i) {
4450 const arch_register_req_t *req;
4451 const arch_register_class_t *class;
4453 if(found_projs & (1 << i)) {
4457 req = get_ia32_out_req(node, i);
4463 block = get_nodes_block(node);
4464 in[0] = new_r_Proj(current_ir_graph, block, node,
4465 arch_register_class_mode(class), i);
4466 if(last_keep != NULL) {
4467 be_Keep_add_node(last_keep, class, in[0]);
4469 last_keep = be_new_Keep(class, current_ir_graph, block, 1, in);
4475 * Adds missing keeps to nodes. Adds missing Proj nodes for unused outputs
4479 void add_missing_keeps(ia32_code_gen_t *cg)
4481 ir_graph *irg = be_get_birg_irg(cg->birg);
4482 irg_walk_graph(irg, add_missing_keep_walker, NULL, NULL);
4485 /* do the transformation */
4486 void ia32_transform_graph(ia32_code_gen_t *cg) {
4487 register_transformers();
4489 initial_fpcw = NULL;
4490 be_transform_graph(cg->birg, ia32_pretransform_node, cg);
4491 edges_verify(cg->irg);
4492 add_missing_keeps(cg);
4493 edges_verify(cg->irg);
4496 void ia32_init_transform(void)
4498 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.transform");