2 * This is the main ia32 firm backend driver.
3 * @author Christian Wuerdig
20 #include <libcore/lc_opts.h>
21 #include <libcore/lc_opts_enum.h>
22 #endif /* WITH_LIBCORE */
24 #include "pseudo_irg.h"
28 #include "iredges_t.h"
36 #include "../beabi.h" /* the general register allocator interface */
37 #include "../benode_t.h"
38 #include "../belower.h"
39 #include "../besched_t.h"
41 #include "bearch_ia32_t.h"
43 #include "ia32_new_nodes.h" /* ia32 nodes interface */
44 #include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */
45 #include "ia32_gen_decls.h" /* interface declaration emitter */
46 #include "ia32_transform.h"
47 #include "ia32_emitter.h"
48 #include "ia32_map_regs.h"
49 #include "ia32_optimize.h"
51 #include "ia32_dbg_stat.h"
52 #include "ia32_finish.h"
53 #include "ia32_util.h"
55 #define DEBUG_MODULE "firm.be.ia32.isa"
58 static set *cur_reg_set = NULL;
61 #define is_Start(irn) (get_irn_opcode(irn) == iro_Start)
63 /* Creates the unique per irg GP NoReg node. */
64 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) {
65 return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_GP_NOREG]);
68 /* Creates the unique per irg FP NoReg node. */
69 ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) {
70 return be_abi_get_callee_save_irn(cg->birg->abi,
71 USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]);
74 /**************************************************
77 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
78 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
79 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
80 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
83 **************************************************/
85 static ir_node *my_skip_proj(const ir_node *n) {
93 * Return register requirements for an ia32 node.
94 * If the node returns a tuple (mode_T) then the proj's
95 * will be asked for this information.
97 static const arch_register_req_t *ia32_get_irn_reg_req(const void *self, arch_register_req_t *req, const ir_node *irn, int pos) {
98 const ia32_irn_ops_t *ops = self;
99 const ia32_register_req_t *irn_req;
100 long node_pos = pos == -1 ? 0 : pos;
101 ir_mode *mode = is_Block(irn) ? NULL : get_irn_mode(irn);
102 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, DEBUG_MODULE);
104 if (is_Block(irn) || mode == mode_M || mode == mode_X) {
105 DBG((mod, LEVEL_1, "ignoring Block, mode_M, mode_X node %+F\n", irn));
109 if (mode == mode_T && pos < 0) {
110 DBG((mod, LEVEL_1, "ignoring request OUT requirements for node %+F\n", irn));
114 DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn));
118 node_pos = ia32_translate_proj_pos(irn);
124 irn = my_skip_proj(irn);
126 DB((mod, LEVEL_1, "skipping Proj, going to %+F at pos %d ... ", irn, node_pos));
129 if (is_ia32_irn(irn)) {
131 irn_req = get_ia32_in_req(irn, pos);
134 irn_req = get_ia32_out_req(irn, node_pos);
137 DB((mod, LEVEL_1, "returning reqs for %+F at pos %d\n", irn, pos));
139 memcpy(req, &(irn_req->req), sizeof(*req));
141 if (arch_register_req_is(&(irn_req->req), should_be_same)) {
142 assert(irn_req->same_pos >= 0 && "should be same constraint for in -> out NYI");
143 req->other_same = get_irn_n(irn, irn_req->same_pos);
146 if (arch_register_req_is(&(irn_req->req), should_be_different)) {
147 assert(irn_req->different_pos >= 0 && "should be different constraint for in -> out NYI");
148 req->other_different = get_irn_n(irn, irn_req->different_pos);
152 /* treat Unknowns like Const with default requirements */
153 if (is_Unknown(irn)) {
154 DB((mod, LEVEL_1, "returning UKNWN reqs for %+F\n", irn));
155 if (mode_is_float(mode)) {
156 if (USE_SSE2(ops->cg))
157 memcpy(req, &(ia32_default_req_ia32_xmm_xmm_UKNWN), sizeof(*req));
159 memcpy(req, &(ia32_default_req_ia32_vfp_vfp_UKNWN), sizeof(*req));
161 else if (mode_is_int(mode) || mode_is_reference(mode))
162 memcpy(req, &(ia32_default_req_ia32_gp_gp_UKNWN), sizeof(*req));
163 else if (mode == mode_T || mode == mode_M) {
164 DBG((mod, LEVEL_1, "ignoring Unknown node %+F\n", irn));
168 assert(0 && "unsupported Unknown-Mode");
171 DB((mod, LEVEL_1, "returning NULL for %+F (not ia32)\n", irn));
179 static void ia32_set_irn_reg(const void *self, ir_node *irn, const arch_register_t *reg) {
181 const ia32_irn_ops_t *ops = self;
183 if (get_irn_mode(irn) == mode_X) {
187 DBG((ops->cg->mod, LEVEL_1, "ia32 assigned register %s to node %+F\n", reg->name, irn));
190 pos = ia32_translate_proj_pos(irn);
191 irn = my_skip_proj(irn);
194 if (is_ia32_irn(irn)) {
195 const arch_register_t **slots;
197 slots = get_ia32_slots(irn);
201 ia32_set_firm_reg(irn, reg, cur_reg_set);
205 static const arch_register_t *ia32_get_irn_reg(const void *self, const ir_node *irn) {
207 const arch_register_t *reg = NULL;
211 if (get_irn_mode(irn) == mode_X) {
215 pos = ia32_translate_proj_pos(irn);
216 irn = my_skip_proj(irn);
219 if (is_ia32_irn(irn)) {
220 const arch_register_t **slots;
221 slots = get_ia32_slots(irn);
225 reg = ia32_get_firm_reg(irn, cur_reg_set);
231 static arch_irn_class_t ia32_classify(const void *self, const ir_node *irn) {
232 irn = my_skip_proj(irn);
234 return arch_irn_class_branch;
235 else if (is_ia32_Cnst(irn))
236 return arch_irn_class_const;
237 else if (is_ia32_Ld(irn))
238 return arch_irn_class_load;
239 else if (is_ia32_St(irn) || is_ia32_Store8Bit(irn))
240 return arch_irn_class_store;
241 else if (is_ia32_irn(irn))
242 return arch_irn_class_normal;
247 static arch_irn_flags_t ia32_get_flags(const void *self, const ir_node *irn) {
248 irn = my_skip_proj(irn);
249 if (is_ia32_irn(irn))
250 return get_ia32_flags(irn);
253 return arch_irn_flags_ignore;
258 static entity *ia32_get_frame_entity(const void *self, const ir_node *irn) {
259 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
262 static void ia32_set_stack_bias(const void *self, ir_node *irn, int bias) {
264 const ia32_irn_ops_t *ops = self;
266 if (get_ia32_frame_ent(irn)) {
267 ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn);
269 DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias));
270 snprintf(buf, sizeof(buf), "%d", bias);
272 if (get_ia32_op_type(irn) == ia32_Normal) {
273 set_ia32_cnst(irn, buf);
276 add_ia32_am_offs(irn, buf);
278 set_ia32_am_flavour(irn, am_flav);
284 be_abi_call_flags_bits_t flags;
285 const arch_isa_t *isa;
286 const arch_env_t *aenv;
290 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
292 ia32_abi_env_t *env = xmalloc(sizeof(env[0]));
293 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
294 env->flags = fl.bits;
297 env->isa = aenv->isa;
302 * Put all registers which are saved by the prologue/epilogue in a set.
304 * @param self The callback object.
305 * @param s The result set.
307 static void ia32_abi_dont_save_regs(void *self, pset *s)
309 ia32_abi_env_t *env = self;
310 if(env->flags.try_omit_fp)
311 pset_insert_ptr(s, env->isa->bp);
315 * Generate the routine prologue.
317 * @param self The callback object.
318 * @param mem A pointer to the mem node. Update this if you define new memory.
319 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
321 * @return The register which shall be used as a stack frame base.
323 * All nodes which define registers in @p reg_map must keep @p reg_map current.
325 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
327 ia32_abi_env_t *env = self;
329 if (!env->flags.try_omit_fp) {
330 ir_node *bl = get_irg_start_block(env->irg);
331 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
332 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
336 push = new_rd_ia32_Push(NULL, env->irg, bl, curr_sp, curr_bp, *mem);
337 curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
338 *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M);
340 /* the push must have SP out register */
341 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
342 set_ia32_flags(push, arch_irn_flags_ignore);
344 /* move esp to ebp */
345 curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp);
346 be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp);
347 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
348 be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
350 /* beware: the copy must be done before any other sp use */
351 curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
352 be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp);
353 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
354 be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
356 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
357 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
366 * Generate the routine epilogue.
367 * @param self The callback object.
368 * @param bl The block for the epilog
369 * @param mem A pointer to the mem node. Update this if you define new memory.
370 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
371 * @return The register which shall be used as a stack frame base.
373 * All nodes which define registers in @p reg_map must keep @p reg_map current.
375 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
377 ia32_abi_env_t *env = self;
378 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
379 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
381 if (env->flags.try_omit_fp) {
382 /* simply remove the stack frame here */
383 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, *mem, BE_STACK_FRAME_SIZE, be_stack_dir_shrink);
386 const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
387 ir_mode *mode_bp = env->isa->bp->reg_class->mode;
389 /* gcc always emits a leave at the end of a routine */
390 if (1 || ARCH_AMD(isa->opt_arch)) {
394 leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, *mem);
395 set_ia32_flags(leave, arch_irn_flags_ignore);
396 curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame);
397 curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
398 *mem = new_r_Proj(current_ir_graph, bl, leave, mode_M, pn_ia32_Leave_M);
403 /* copy ebp to esp */
404 curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
407 pop = new_rd_ia32_Pop(NULL, env->irg, bl, curr_sp, *mem);
408 set_ia32_flags(pop, arch_irn_flags_ignore);
409 curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res);
410 curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
411 *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
413 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
414 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
417 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
418 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
422 * Produces the type which sits between the stack args and the locals on the stack.
423 * it will contain the return address and space to store the old base pointer.
424 * @return The Firm type modeling the ABI between type.
426 static ir_type *ia32_abi_get_between_type(void *self)
428 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
429 static ir_type *omit_fp_between_type = NULL;
430 static ir_type *between_type = NULL;
432 ia32_abi_env_t *env = self;
434 if ( !between_type) {
436 entity *ret_addr_ent;
437 entity *omit_fp_ret_addr_ent;
439 ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P);
440 ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_P);
442 between_type = new_type_struct(IDENT("ia32_between_type"));
443 old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
444 ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
446 set_entity_offset_bytes(old_bp_ent, 0);
447 set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type));
448 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
449 set_type_state(between_type, layout_fixed);
451 omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
452 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
454 set_entity_offset_bytes(omit_fp_ret_addr_ent, 0);
455 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
456 set_type_state(omit_fp_between_type, layout_fixed);
459 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
464 * Get the estimated cycle count for @p irn.
466 * @param self The this pointer.
467 * @param irn The node.
469 * @return The estimated cycle count for this operation
471 static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
478 switch (get_ia32_irn_opcode(irn)) {
480 case iro_ia32_DivMod:
485 case iro_ia32_l_Load:
495 case iro_ia32_xStore:
496 case iro_ia32_l_Store:
498 case iro_ia32_Store8Bit:
506 case iro_ia32_l_MulS:
519 * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
521 * @param irn The original operation
522 * @param i Index of the argument we want the inverse operation to yield
523 * @param inverse struct to be filled with the resulting inverse op
524 * @param obstack The obstack to use for allocation of the returned nodes array
525 * @return The inverse operation or NULL if operation invertible
527 static arch_inverse_t *ia32_get_inverse(const void *self, const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) {
530 ir_node *block, *noreg, *nomem;
533 /* we cannot invert non-ia32 irns */
534 if (! is_ia32_irn(irn))
537 /* operand must always be a real operand (not base, index or mem) */
538 if (i != 2 && i != 3)
541 /* we don't invert address mode operations */
542 if (get_ia32_op_type(irn) != ia32_Normal)
545 irg = get_irn_irg(irn);
546 block = get_nodes_block(irn);
547 mode = get_ia32_res_mode(irn);
548 noreg = get_irn_n(irn, 0);
549 nomem = new_r_NoMem(irg);
551 /* initialize structure */
552 inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
556 switch (get_ia32_irn_opcode(irn)) {
558 if (get_ia32_immop_type(irn) == ia32_ImmConst) {
559 /* we have an add with a const here */
560 /* invers == add with negated const */
561 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
562 pnc = pn_ia32_Add_res;
564 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
565 set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
566 set_ia32_commutative(inverse->nodes[0]);
568 else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
569 /* we have an add with a symconst here */
570 /* invers == sub with const */
571 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
572 pnc = pn_ia32_Sub_res;
574 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
577 /* normal add: inverse == sub */
578 ir_node *proj = ia32_get_res_proj(irn);
581 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem);
582 pnc = pn_ia32_Sub_res;
587 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
588 /* we have a sub with a const/symconst here */
589 /* invers == add with this const */
590 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
591 pnc = pn_ia32_Add_res;
592 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
593 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
597 ir_node *proj = ia32_get_res_proj(irn);
601 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem);
604 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, 2), proj, nomem);
606 pnc = pn_ia32_Sub_res;
611 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
612 /* xor with const: inverse = xor */
613 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
614 pnc = pn_ia32_Eor_res;
615 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
616 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
620 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, (ir_node *)irn, get_irn_n(irn, i), nomem);
621 pnc = pn_ia32_Eor_res;
626 ir_node *proj = ia32_get_res_proj(irn);
629 inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, proj, nomem);
630 pnc = pn_ia32_Not_res;
634 case iro_ia32_Minus: {
635 ir_node *proj = ia32_get_res_proj(irn);
638 inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, proj, nomem);
639 pnc = pn_ia32_Minus_res;
644 /* inverse operation not supported */
648 set_ia32_res_mode(inverse->nodes[0], mode);
649 inverse->nodes[1] = new_r_Proj(irg, block, inverse->nodes[0], mode, pnc);
655 * Check if irn can load it's operand at position i from memory (source addressmode).
656 * @param self Pointer to irn ops itself
657 * @param irn The irn to be checked
658 * @param i The operands position
659 * @return Non-Zero if operand can be loaded
661 static int ia32_possible_memory_operand(const void *self, const ir_node *irn, unsigned int i) {
662 if (! is_ia32_irn(irn) || /* must be an ia32 irn */
663 get_irn_arity(irn) != 5 || /* must be a binary operation */
664 get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
665 ! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */
666 (i != 2 && i != 3) || /* a "real" operand position must be requested */
667 (i == 2 && ! is_ia32_commutative(irn)) || /* if first operand requested irn must be commutative */
668 is_ia32_use_frame(irn)) /* must not already use frame */
674 static void ia32_perform_memory_operand(const void *self, ir_node *irn, ir_node *reload, unsigned int i) {
675 assert(ia32_possible_memory_operand(self, irn, i) && "Cannot perform memory operand change");
676 assert(get_nodes_block(reload) == get_nodes_block(irn) && "Reload must be in same block as irn.");
678 if (get_irn_n_edges(reload) > 1)
682 ir_node *tmp = get_irn_n(irn, 3);
683 set_irn_n(irn, 3, get_irn_n(irn, 2));
684 set_irn_n(irn, 2, tmp);
687 set_ia32_am_support(irn, ia32_am_Source);
688 set_ia32_op_type(irn, ia32_AddrModeS);
689 set_ia32_am_flavour(irn, ia32_B);
690 set_ia32_ls_mode(irn, get_irn_mode(reload));
691 set_ia32_frame_ent(irn, be_get_frame_entity(reload));
692 set_ia32_use_frame(irn);
694 set_irn_n(irn, 0, be_get_Reload_frame(reload));
695 set_irn_n(irn, 4, be_get_Reload_mem(reload));
698 Input at position one is index register, which is NoReg.
699 We would need cg object to get a real noreg, but we cannot
702 set_irn_n(irn, 3, get_irn_n(irn, 1));
704 DBG_OPT_AM_S(reload, irn);
707 static const be_abi_callbacks_t ia32_abi_callbacks = {
710 ia32_abi_get_between_type,
711 ia32_abi_dont_save_regs,
716 /* fill register allocator interface */
718 static const arch_irn_ops_if_t ia32_irn_ops_if = {
719 ia32_get_irn_reg_req,
724 ia32_get_frame_entity,
727 ia32_get_op_estimated_cost,
728 ia32_possible_memory_operand,
729 ia32_perform_memory_operand,
732 ia32_irn_ops_t ia32_irn_ops = {
739 /**************************************************
742 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
743 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
744 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
745 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
748 **************************************************/
751 * Transforms the standard firm graph into
754 static void ia32_prepare_graph(void *self) {
755 ia32_code_gen_t *cg = self;
756 dom_front_info_t *dom;
757 DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
759 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
761 /* 1st: transform constants and psi condition trees */
762 ia32_pre_transform_phase(cg);
764 /* 2nd: transform all remaining nodes */
765 ia32_register_transformers();
766 dom = be_compute_dominance_frontiers(cg->irg);
767 irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg);
768 be_free_dominance_frontiers(dom);
771 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
773 /* 3rd: optimize address mode */
774 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
775 ia32_optimize_addressmode(cg);
778 be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
780 DEBUG_ONLY(cg->mod = old_mod;)
784 * Dummy functions for hooks we don't need but which must be filled.
786 static void ia32_before_sched(void *self) {
789 static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) {
797 mode = get_irn_mode(irn);
799 /* check if we already saw this node or the node has more than one user */
800 if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1)
803 /* mark irn visited */
804 bitset_add_irn(already_visited, irn);
806 /* non-Tuple nodes with one user: ok, return */
807 if (get_irn_n_edges(irn) >= 1 && mode != mode_T)
810 /* tuple node has one user which is not the mem proj-> ok */
811 if (mode == mode_T && get_irn_n_edges(irn) == 1) {
812 mem_proj = ia32_get_proj_for_mode(irn, mode_M);
817 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
818 ir_node *pred = get_irn_n(irn, i);
820 /* do not follow memory edges or we will accidentally remove stores */
821 if (is_Proj(pred) && get_irn_mode(pred) == mode_M)
824 set_irn_n(irn, i, new_Bad());
827 The current node is about to be removed: if the predecessor
828 has only this node as user, it need to be removed as well.
830 if (get_irn_n_edges(pred) <= 1)
831 remove_unused_nodes(pred, already_visited);
834 if (sched_is_scheduled(irn))
838 static void remove_unused_loads_walker(ir_node *irn, void *env) {
839 bitset_t *already_visited = env;
840 if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn))
841 remove_unused_nodes(irn, env);
845 * Called before the register allocator.
846 * Calculate a block schedule here. We need it for the x87
847 * simulator and the emitter.
849 static void ia32_before_ra(void *self) {
850 ia32_code_gen_t *cg = self;
851 bitset_t *already_visited = bitset_irg_malloc(cg->irg);
853 cg->blk_sched = sched_create_block_schedule(cg->irg);
857 There are sometimes unused loads, only pinned by memory.
858 We need to remove those Loads and all other nodes which won't be used
859 after removing the Load from schedule.
861 irg_walk_graph(cg->irg, remove_unused_loads_walker, NULL, already_visited);
862 bitset_free(already_visited);
867 * Transforms a be node into a Load.
869 static void transform_to_Load(ia32_transform_env_t *env) {
870 ir_node *irn = env->irn;
871 entity *ent = be_get_frame_entity(irn);
872 ir_mode *mode = env->mode;
873 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
874 ir_node *nomem = new_rd_NoMem(env->irg);
875 ir_node *sched_point = NULL;
876 ir_node *ptr = get_irn_n(irn, 0);
877 ir_node *mem = be_is_Reload(irn) ? get_irn_n(irn, 1) : nomem;
878 ir_node *new_op, *proj;
879 const arch_register_t *reg;
881 if (sched_is_scheduled(irn)) {
882 sched_point = sched_prev(irn);
885 if (mode_is_float(mode)) {
886 if (USE_SSE2(env->cg))
887 new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
889 new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
892 new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
895 set_ia32_am_support(new_op, ia32_am_Source);
896 set_ia32_op_type(new_op, ia32_AddrModeS);
897 set_ia32_am_flavour(new_op, ia32_B);
898 set_ia32_ls_mode(new_op, mode);
899 set_ia32_frame_ent(new_op, ent);
900 set_ia32_use_frame(new_op);
902 DBG_OPT_RELOAD2LD(irn, new_op);
904 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_Load_res);
907 sched_add_after(sched_point, new_op);
908 sched_add_after(new_op, proj);
913 /* copy the register from the old node to the new Load */
914 reg = arch_get_irn_register(env->cg->arch_env, irn);
915 arch_set_irn_register(env->cg->arch_env, new_op, reg);
917 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
923 * Transforms a be node into a Store.
925 static void transform_to_Store(ia32_transform_env_t *env) {
926 ir_node *irn = env->irn;
927 entity *ent = be_get_frame_entity(irn);
928 ir_mode *mode = env->mode;
929 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
930 ir_node *nomem = new_rd_NoMem(env->irg);
931 ir_node *ptr = get_irn_n(irn, 0);
932 ir_node *val = get_irn_n(irn, 1);
933 ir_node *new_op, *proj;
934 ir_node *sched_point = NULL;
936 if (sched_is_scheduled(irn)) {
937 sched_point = sched_prev(irn);
940 if (mode_is_float(mode)) {
941 if (USE_SSE2(env->cg))
942 new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
944 new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
946 else if (get_mode_size_bits(mode) == 8) {
947 new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
950 new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
953 set_ia32_am_support(new_op, ia32_am_Dest);
954 set_ia32_op_type(new_op, ia32_AddrModeD);
955 set_ia32_am_flavour(new_op, ia32_B);
956 set_ia32_ls_mode(new_op, mode);
957 set_ia32_frame_ent(new_op, ent);
958 set_ia32_use_frame(new_op);
960 DBG_OPT_SPILL2ST(irn, new_op);
962 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, pn_ia32_Store_M);
965 sched_add_after(sched_point, new_op);
969 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
975 * Fix the mode of Spill/Reload
977 static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode)
979 if (mode_is_float(mode)) {
991 * Block-Walker: Calls the transform functions Spill and Reload.
993 static void ia32_after_ra_walker(ir_node *block, void *env) {
994 ir_node *node, *prev;
995 ia32_code_gen_t *cg = env;
996 ia32_transform_env_t tenv;
999 tenv.irg = current_ir_graph;
1001 DEBUG_ONLY(tenv.mod = cg->mod;)
1003 /* beware: the schedule is changed here */
1004 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
1005 prev = sched_prev(node);
1006 if (be_is_Reload(node)) {
1007 /* we always reload the whole register */
1008 tenv.dbg = get_irn_dbg_info(node);
1010 tenv.mode = fix_spill_mode(cg, get_irn_mode(node));
1011 transform_to_Load(&tenv);
1013 else if (be_is_Spill(node)) {
1014 ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
1015 /* we always spill the whole register */
1016 tenv.dbg = get_irn_dbg_info(node);
1018 tenv.mode = fix_spill_mode(cg, get_irn_mode(spillval));
1019 transform_to_Store(&tenv);
1025 * We transform Spill and Reload here. This needs to be done before
1026 * stack biasing otherwise we would miss the corrected offset for these nodes.
1028 * If x87 instruction should be emitted, run the x87 simulator and patch
1029 * the virtual instructions. This must obviously be done after register allocation.
1031 static void ia32_after_ra(void *self) {
1032 ia32_code_gen_t *cg = self;
1034 irg_block_walk_graph(cg->irg, NULL, ia32_after_ra_walker, self);
1036 /* if we do x87 code generation, rewrite all the virtual instructions and registers */
1037 if (cg->used_fp == fp_x87 || cg->force_sim) {
1038 x87_simulate_graph(cg->arch_env, cg->irg, cg->blk_sched);
1043 * Last touchups for the graph before emit
1045 static void ia32_finish(void *self) {
1046 ia32_code_gen_t *cg = self;
1047 ir_graph *irg = cg->irg;
1049 ia32_finish_irg(irg, cg);
1051 be_dump(irg, "-finished", dump_ir_block_graph_sched);
1055 * Emits the code, closes the output file and frees
1056 * the code generator interface.
1058 static void ia32_codegen(void *self) {
1059 ia32_code_gen_t *cg = self;
1060 ir_graph *irg = cg->irg;
1062 ia32_gen_routine(cg->isa->out, irg, cg);
1066 /* remove it from the isa */
1069 /* de-allocate code generator */
1070 del_set(cg->reg_set);
1074 static void *ia32_cg_init(const be_irg_t *birg);
1076 static const arch_code_generator_if_t ia32_code_gen_if = {
1078 NULL, /* before abi introduce hook */
1080 ia32_before_sched, /* before scheduling hook */
1081 ia32_before_ra, /* before register allocation hook */
1082 ia32_after_ra, /* after register allocation hook */
1083 ia32_finish, /* called before codegen */
1084 ia32_codegen /* emit && done */
1088 * Initializes a IA32 code generator.
1090 static void *ia32_cg_init(const be_irg_t *birg) {
1091 ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa;
1092 ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg));
1094 cg->impl = &ia32_code_gen_if;
1095 cg->irg = birg->irg;
1096 cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
1097 cg->arch_env = birg->main_env->arch_env;
1100 cg->blk_sched = NULL;
1101 cg->fp_to_gp = NULL;
1102 cg->gp_to_fp = NULL;
1103 cg->fp_kind = isa->fp_kind;
1104 cg->used_fp = fp_none;
1105 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
1107 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
1109 /* copy optimizations from isa for easier access */
1111 cg->arch = isa->arch;
1112 cg->opt_arch = isa->opt_arch;
1118 if (isa->name_obst_size) {
1119 //printf("freed %d bytes from name obst\n", isa->name_obst_size);
1120 isa->name_obst_size = 0;
1121 obstack_free(isa->name_obst, NULL);
1122 obstack_init(isa->name_obst);
1126 cur_reg_set = cg->reg_set;
1128 ia32_irn_ops.cg = cg;
1130 return (arch_code_generator_t *)cg;
1135 /*****************************************************************
1136 * ____ _ _ _____ _____
1137 * | _ \ | | | | |_ _|/ ____| /\
1138 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
1139 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
1140 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
1141 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
1143 *****************************************************************/
1146 * Set output modes for GCC
1148 static const tarval_mode_info mo_integer = {
1155 * set the tarval output mode to C-semantics
1157 static void set_tarval_output_modes(void)
1159 set_tarval_mode_output_option(get_modeLs(), &mo_integer);
1160 set_tarval_mode_output_option(get_modeLu(), &mo_integer);
1161 set_tarval_mode_output_option(get_modeIs(), &mo_integer);
1162 set_tarval_mode_output_option(get_modeIu(), &mo_integer);
1163 set_tarval_mode_output_option(get_modeHs(), &mo_integer);
1164 set_tarval_mode_output_option(get_modeHu(), &mo_integer);
1165 set_tarval_mode_output_option(get_modeBs(), &mo_integer);
1166 set_tarval_mode_output_option(get_modeBu(), &mo_integer);
1167 set_tarval_mode_output_option(get_modeC(), &mo_integer);
1168 set_tarval_mode_output_option(get_modeU(), &mo_integer);
1169 set_tarval_mode_output_option(get_modeIu(), &mo_integer);
1174 * The template that generates a new ISA object.
1175 * Note that this template can be changed by command line
1178 static ia32_isa_t ia32_isa_template = {
1180 &ia32_isa_if, /* isa interface implementation */
1181 &ia32_gp_regs[REG_ESP], /* stack pointer register */
1182 &ia32_gp_regs[REG_EBP], /* base pointer register */
1183 -1, /* stack direction */
1185 NULL, /* 16bit register names */
1186 NULL, /* 8bit register names */
1190 IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
1191 IA32_OPT_DOAM | /* optimize address mode default: on */
1192 IA32_OPT_LEA | /* optimize for LEAs default: on */
1193 IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */
1194 IA32_OPT_IMMOPS | /* operations can use immediates, default: on */
1195 IA32_OPT_EXTBB), /* use extended basic block scheduling, default: on */
1196 arch_pentium_4, /* instruction architecture */
1197 arch_pentium_4, /* optimize for architecture */
1198 fp_sse2, /* use sse2 unit */
1199 NULL, /* current code generator */
1201 NULL, /* name obstack */
1202 0 /* name obst size */
1207 * Initializes the backend ISA.
1209 static void *ia32_init(FILE *file_handle) {
1210 static int inited = 0;
1216 set_tarval_output_modes();
1218 isa = xmalloc(sizeof(*isa));
1219 memcpy(isa, &ia32_isa_template, sizeof(*isa));
1221 ia32_register_init(isa);
1222 ia32_create_opcodes();
1224 if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
1225 (ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
1226 /* no SSE2 for these cpu's */
1227 isa->fp_kind = fp_x87;
1229 if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) {
1230 /* Pentium 4 don't like inc and dec instructions */
1231 isa->opt &= ~IA32_OPT_INCDEC;
1234 isa->regs_16bit = pmap_create();
1235 isa->regs_8bit = pmap_create();
1236 isa->types = pmap_create();
1237 isa->tv_ent = pmap_create();
1238 isa->out = file_handle;
1240 ia32_build_16bit_reg_map(isa->regs_16bit);
1241 ia32_build_8bit_reg_map(isa->regs_8bit);
1243 /* patch register names of x87 registers */
1245 ia32_st_regs[0].name = "st";
1246 ia32_st_regs[1].name = "st(1)";
1247 ia32_st_regs[2].name = "st(2)";
1248 ia32_st_regs[3].name = "st(3)";
1249 ia32_st_regs[4].name = "st(4)";
1250 ia32_st_regs[5].name = "st(5)";
1251 ia32_st_regs[6].name = "st(6)";
1252 ia32_st_regs[7].name = "st(7)";
1256 isa->name_obst = xmalloc(sizeof(*isa->name_obst));
1257 obstack_init(isa->name_obst);
1258 isa->name_obst_size = 0;
1261 ia32_handle_intrinsics();
1262 ia32_switch_section(NULL, NO_SECTION);
1263 fprintf(isa->out, "\t.intel_syntax\n");
1273 * Closes the output file and frees the ISA structure.
1275 static void ia32_done(void *self) {
1276 ia32_isa_t *isa = self;
1278 /* emit now all global declarations */
1279 ia32_gen_decls(isa->out);
1281 pmap_destroy(isa->regs_16bit);
1282 pmap_destroy(isa->regs_8bit);
1283 pmap_destroy(isa->tv_ent);
1284 pmap_destroy(isa->types);
1287 //printf("name obst size = %d bytes\n", isa->name_obst_size);
1288 obstack_free(isa->name_obst, NULL);
1296 * Return the number of register classes for this architecture.
1297 * We report always these:
1298 * - the general purpose registers
1299 * - the SSE floating point register set
1300 * - the virtual floating point registers
1302 static int ia32_get_n_reg_class(const void *self) {
1307 * Return the register class for index i.
1309 static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
1310 assert(i >= 0 && i < 3 && "Invalid ia32 register class requested.");
1312 return &ia32_reg_classes[CLASS_ia32_gp];
1314 return &ia32_reg_classes[CLASS_ia32_xmm];
1316 return &ia32_reg_classes[CLASS_ia32_vfp];
1320 * Get the register class which shall be used to store a value of a given mode.
1321 * @param self The this pointer.
1322 * @param mode The mode in question.
1323 * @return A register class which can hold values of the given mode.
1325 const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
1326 const ia32_isa_t *isa = self;
1327 if (mode_is_float(mode)) {
1328 return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1331 return &ia32_reg_classes[CLASS_ia32_gp];
1335 * Get the ABI restrictions for procedure calls.
1336 * @param self The this pointer.
1337 * @param method_type The type of the method (procedure) in question.
1338 * @param abi The abi object to be modified
1340 static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1341 const ia32_isa_t *isa = self;
1344 unsigned cc = get_method_calling_convention(method_type);
1345 int n = get_method_n_params(method_type);
1348 int i, ignore_1, ignore_2;
1350 const arch_register_t *reg;
1351 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1353 unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
1355 /* set abi flags for calls */
1356 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1357 call_flags.bits.store_args_sequential = use_push;
1358 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1359 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1360 call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
1362 /* set stack parameter passing style */
1363 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1365 /* collect the mode for each type */
1366 modes = alloca(n * sizeof(modes[0]));
1368 for (i = 0; i < n; i++) {
1369 tp = get_method_param_type(method_type, i);
1370 modes[i] = get_type_mode(tp);
1373 /* set register parameters */
1374 if (cc & cc_reg_param) {
1375 /* determine the number of parameters passed via registers */
1376 biggest_n = ia32_get_n_regparam_class(n, modes, &ignore_1, &ignore_2);
1378 /* loop over all parameters and set the register requirements */
1379 for (i = 0; i <= biggest_n; i++) {
1380 reg = ia32_get_RegParam_reg(n, modes, i, cc);
1381 assert(reg && "kaputt");
1382 be_abi_call_param_reg(abi, i, reg);
1389 /* set stack parameters */
1390 for (i = stack_idx; i < n; i++) {
1391 be_abi_call_param_stack(abi, i, 1, 0, 0);
1395 /* set return registers */
1396 n = get_method_n_ress(method_type);
1398 assert(n <= 2 && "more than two results not supported");
1400 /* In case of 64bit returns, we will have two 32bit values */
1402 tp = get_method_res_type(method_type, 0);
1403 mode = get_type_mode(tp);
1405 assert(!mode_is_float(mode) && "two FP results not supported");
1407 tp = get_method_res_type(method_type, 1);
1408 mode = get_type_mode(tp);
1410 assert(!mode_is_float(mode) && "two FP results not supported");
1412 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX]);
1413 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]);
1416 const arch_register_t *reg;
1418 tp = get_method_res_type(method_type, 0);
1419 assert(is_atomic_type(tp));
1420 mode = get_type_mode(tp);
1422 reg = mode_is_float(mode) ?
1423 (USE_SSE2(isa) ? &ia32_xmm_regs[REG_XMM0] : &ia32_vfp_regs[REG_VF0]) :
1424 &ia32_gp_regs[REG_EAX];
1426 be_abi_call_res_reg(abi, 0, reg);
1431 static const void *ia32_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1432 return &ia32_irn_ops;
1435 const arch_irn_handler_t ia32_irn_handler = {
1439 const arch_irn_handler_t *ia32_get_irn_handler(const void *self) {
1440 return &ia32_irn_handler;
1443 int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1444 return is_ia32_irn(irn) ? 1 : -1;
1448 * Initializes the code generator interface.
1450 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self) {
1451 return &ia32_code_gen_if;
1454 list_sched_selector_t ia32_sched_selector;
1457 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1459 static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self) {
1460 // memcpy(&ia32_sched_selector, reg_pressure_selector, sizeof(list_sched_selector_t));
1461 memcpy(&ia32_sched_selector, trivial_selector, sizeof(list_sched_selector_t));
1462 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
1463 return &ia32_sched_selector;
1467 * Returns the necessary byte alignment for storing a register of given class.
1469 static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1470 ir_mode *mode = arch_register_class_mode(cls);
1471 int bytes = get_mode_size_bytes(mode);
1473 if (mode_is_float(mode) && bytes > 8)
1478 static ia32_intrinsic_env_t intrinsic_env = { NULL, NULL };
1481 * Returns the libFirm configuration parameter for this backend.
1483 static const backend_params *ia32_get_libfirm_params(void) {
1484 static const arch_dep_params_t ad = {
1485 1, /* also use subs */
1486 4, /* maximum shifts */
1487 31, /* maximum shift amount */
1489 1, /* allow Mulhs */
1490 1, /* allow Mulus */
1491 32 /* Mulh allowed up to 32 bit */
1493 static backend_params p = {
1494 NULL, /* no additional opcodes */
1495 NULL, /* will be set later */
1496 1, /* need dword lowering */
1497 ia32_create_intrinsic_fkt,
1498 &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
1506 /* instruction set architectures. */
1507 static const lc_opt_enum_int_items_t arch_items[] = {
1508 { "386", arch_i386, },
1509 { "486", arch_i486, },
1510 { "pentium", arch_pentium, },
1511 { "586", arch_pentium, },
1512 { "pentiumpro", arch_pentium_pro, },
1513 { "686", arch_pentium_pro, },
1514 { "pentiummmx", arch_pentium_mmx, },
1515 { "pentium2", arch_pentium_2, },
1516 { "p2", arch_pentium_2, },
1517 { "pentium3", arch_pentium_3, },
1518 { "p3", arch_pentium_3, },
1519 { "pentium4", arch_pentium_4, },
1520 { "p4", arch_pentium_4, },
1521 { "pentiumm", arch_pentium_m, },
1522 { "pm", arch_pentium_m, },
1523 { "core", arch_core, },
1525 { "athlon", arch_athlon, },
1526 { "athlon64", arch_athlon_64, },
1527 { "opteron", arch_opteron, },
1531 static lc_opt_enum_int_var_t arch_var = {
1532 &ia32_isa_template.arch, arch_items
1535 static lc_opt_enum_int_var_t opt_arch_var = {
1536 &ia32_isa_template.opt_arch, arch_items
1539 static const lc_opt_enum_int_items_t fp_unit_items[] = {
1541 { "sse2", fp_sse2 },
1545 static lc_opt_enum_int_var_t fp_unit_var = {
1546 &ia32_isa_template.fp_kind, fp_unit_items
1549 static const lc_opt_enum_int_items_t gas_items[] = {
1550 { "linux", ASM_LINUX_GAS },
1551 { "mingw", ASM_MINGW_GAS },
1555 static lc_opt_enum_int_var_t gas_var = {
1556 (int *)&asm_flavour, gas_items
1559 static const lc_opt_table_entry_t ia32_options[] = {
1560 LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var),
1561 LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var),
1562 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var),
1563 LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
1564 LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA),
1565 LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST),
1566 LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS),
1567 LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB),
1568 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
1573 * Register command line options for the ia32 backend.
1577 * ia32-arch=arch create instruction for arch
1578 * ia32-opt=arch optimize for run on arch
1579 * ia32-fpunit=unit select floating point unit (x87 or SSE2)
1580 * ia32-incdec optimize for inc/dec
1581 * ia32-noaddrmode do not use address mode
1582 * ia32-nolea do not optimize for LEAs
1583 * ia32-noplacecnst do not place constants,
1584 * ia32-noimmop no operations with immediates
1585 * ia32-noextbb do not use extended basic block scheduling
1586 * ia32-gasmode set the GAS compatibility mode
1588 static void ia32_register_options(lc_opt_entry_t *ent)
1590 lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32");
1591 lc_opt_add_table(be_grp_ia32, ia32_options);
1593 #endif /* WITH_LIBCORE */
1595 const arch_isa_if_t ia32_isa_if = {
1598 ia32_get_n_reg_class,
1600 ia32_get_reg_class_for_mode,
1602 ia32_get_irn_handler,
1603 ia32_get_code_generator_if,
1604 ia32_get_list_sched_selector,
1605 ia32_get_reg_class_alignment,
1606 ia32_get_libfirm_params,
1608 ia32_register_options