2 * This is the main ia32 firm backend driver.
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"
53 #define DEBUG_MODULE "firm.be.ia32.isa"
56 static set *cur_reg_set = NULL;
59 #define is_Start(irn) (get_irn_opcode(irn) == iro_Start)
61 /* Creates the unique per irg GP NoReg node. */
62 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) {
63 return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_GP_NOREG]);
66 /* Creates the unique per irg FP NoReg node. */
67 ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) {
68 return be_abi_get_callee_save_irn(cg->birg->abi,
69 USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]);
72 /**************************************************
75 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
76 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
77 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
78 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
81 **************************************************/
83 static ir_node *my_skip_proj(const ir_node *n) {
91 * Return register requirements for an ia32 node.
92 * If the node returns a tuple (mode_T) then the proj's
93 * will be asked for this information.
95 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) {
96 const ia32_irn_ops_t *ops = self;
97 const ia32_register_req_t *irn_req;
98 long node_pos = pos == -1 ? 0 : pos;
99 ir_mode *mode = is_Block(irn) ? NULL : get_irn_mode(irn);
100 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, DEBUG_MODULE);
102 if (is_Block(irn) || mode == mode_M || mode == mode_X) {
103 DBG((mod, LEVEL_1, "ignoring Block, mode_M, mode_X node %+F\n", irn));
107 if (mode == mode_T && pos < 0) {
108 DBG((mod, LEVEL_1, "ignoring request OUT requirements for node %+F\n", irn));
112 DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn));
116 node_pos = ia32_translate_proj_pos(irn);
122 irn = my_skip_proj(irn);
124 DB((mod, LEVEL_1, "skipping Proj, going to %+F at pos %d ... ", irn, node_pos));
127 if (is_ia32_irn(irn)) {
129 irn_req = get_ia32_in_req(irn, pos);
132 irn_req = get_ia32_out_req(irn, node_pos);
135 DB((mod, LEVEL_1, "returning reqs for %+F at pos %d\n", irn, pos));
137 memcpy(req, &(irn_req->req), sizeof(*req));
139 if (arch_register_req_is(&(irn_req->req), should_be_same)) {
140 assert(irn_req->same_pos >= 0 && "should be same constraint for in -> out NYI");
141 req->other_same = get_irn_n(irn, irn_req->same_pos);
144 if (arch_register_req_is(&(irn_req->req), should_be_different)) {
145 assert(irn_req->different_pos >= 0 && "should be different constraint for in -> out NYI");
146 req->other_different = get_irn_n(irn, irn_req->different_pos);
150 /* treat Unknowns like Const with default requirements */
151 if (is_Unknown(irn)) {
152 DB((mod, LEVEL_1, "returning UKNWN reqs for %+F\n", irn));
153 if (mode_is_float(mode)) {
154 if (USE_SSE2(ops->cg))
155 memcpy(req, &(ia32_default_req_ia32_xmm_xmm_UKNWN), sizeof(*req));
157 memcpy(req, &(ia32_default_req_ia32_vfp_vfp_UKNWN), sizeof(*req));
159 else if (mode_is_int(mode) || mode_is_reference(mode))
160 memcpy(req, &(ia32_default_req_ia32_gp_gp_UKNWN), sizeof(*req));
161 else if (mode == mode_T || mode == mode_M) {
162 DBG((mod, LEVEL_1, "ignoring Unknown node %+F\n", irn));
166 assert(0 && "unsupported Unknown-Mode");
169 DB((mod, LEVEL_1, "returning NULL for %+F (not ia32)\n", irn));
177 static void ia32_set_irn_reg(const void *self, ir_node *irn, const arch_register_t *reg) {
179 const ia32_irn_ops_t *ops = self;
181 if (get_irn_mode(irn) == mode_X) {
185 DBG((ops->cg->mod, LEVEL_1, "ia32 assigned register %s to node %+F\n", reg->name, irn));
188 pos = ia32_translate_proj_pos(irn);
189 irn = my_skip_proj(irn);
192 if (is_ia32_irn(irn)) {
193 const arch_register_t **slots;
195 slots = get_ia32_slots(irn);
199 ia32_set_firm_reg(irn, reg, cur_reg_set);
203 static const arch_register_t *ia32_get_irn_reg(const void *self, const ir_node *irn) {
205 const arch_register_t *reg = NULL;
209 if (get_irn_mode(irn) == mode_X) {
213 pos = ia32_translate_proj_pos(irn);
214 irn = my_skip_proj(irn);
217 if (is_ia32_irn(irn)) {
218 const arch_register_t **slots;
219 slots = get_ia32_slots(irn);
223 reg = ia32_get_firm_reg(irn, cur_reg_set);
229 static arch_irn_class_t ia32_classify(const void *self, const ir_node *irn) {
230 irn = my_skip_proj(irn);
232 return arch_irn_class_branch;
233 else if (is_ia32_Cnst(irn))
234 return arch_irn_class_const;
235 else if (is_ia32_irn(irn))
236 return arch_irn_class_normal;
241 static arch_irn_flags_t ia32_get_flags(const void *self, const ir_node *irn) {
242 irn = my_skip_proj(irn);
243 if (is_ia32_irn(irn))
244 return get_ia32_flags(irn);
247 return arch_irn_flags_ignore;
252 static entity *ia32_get_frame_entity(const void *self, const ir_node *irn) {
253 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
256 static void ia32_set_stack_bias(const void *self, ir_node *irn, int bias) {
258 const ia32_irn_ops_t *ops = self;
260 if (get_ia32_frame_ent(irn)) {
261 ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn);
263 DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias));
264 snprintf(buf, sizeof(buf), "%d", bias);
266 if (get_ia32_op_type(irn) == ia32_Normal) {
267 set_ia32_cnst(irn, buf);
270 add_ia32_am_offs(irn, buf);
272 set_ia32_am_flavour(irn, am_flav);
278 be_abi_call_flags_bits_t flags;
279 const arch_isa_t *isa;
280 const arch_env_t *aenv;
284 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
286 ia32_abi_env_t *env = xmalloc(sizeof(env[0]));
287 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
288 env->flags = fl.bits;
291 env->isa = aenv->isa;
296 * Put all registers which are saved by the prologue/epilogue in a set.
298 * @param self The callback object.
299 * @param s The result set.
301 static void ia32_abi_dont_save_regs(void *self, pset *s)
303 ia32_abi_env_t *env = self;
304 if(env->flags.try_omit_fp)
305 pset_insert_ptr(s, env->isa->bp);
309 * Generate the routine prologue.
311 * @param self The callback object.
312 * @param mem A pointer to the mem node. Update this if you define new memory.
313 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
315 * @return The register which shall be used as a stack frame base.
317 * All nodes which define registers in @p reg_map must keep @p reg_map current.
319 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
321 ia32_abi_env_t *env = self;
323 if (!env->flags.try_omit_fp) {
324 int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
325 ir_node *bl = get_irg_start_block(env->irg);
326 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
327 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
331 push = new_rd_ia32_Push(NULL, env->irg, bl, curr_sp, curr_bp, *mem);
332 curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
333 *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M);
335 /* the push must have SP out register */
336 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
337 set_ia32_flags(push, arch_irn_flags_ignore);
339 /* move esp to ebp */
340 curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp);
341 be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp);
342 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
343 be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
345 /* beware: the copy must be done before any other sp use */
346 curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
347 be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp);
348 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
349 be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
351 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
352 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
361 * Generate the routine epilogue.
362 * @param self The callback object.
363 * @param bl The block for the epilog
364 * @param mem A pointer to the mem node. Update this if you define new memory.
365 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
366 * @return The register which shall be used as a stack frame base.
368 * All nodes which define registers in @p reg_map must keep @p reg_map current.
370 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
372 ia32_abi_env_t *env = self;
373 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
374 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
376 if (env->flags.try_omit_fp) {
377 /* simply remove the stack frame here */
378 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, *mem, BE_STACK_FRAME_SIZE, be_stack_dir_shrink);
381 const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
382 ir_mode *mode_bp = env->isa->bp->reg_class->mode;
383 int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode);
385 /* gcc always emits a leave at the end of a routine */
386 if (1 || ARCH_AMD(isa->opt_arch)) {
390 leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, *mem);
391 set_ia32_flags(leave, arch_irn_flags_ignore);
392 curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame);
393 curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
394 *mem = new_r_Proj(current_ir_graph, bl, leave, mode_M, pn_ia32_Leave_M);
399 /* copy ebp to esp */
400 curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
403 pop = new_rd_ia32_Pop(NULL, env->irg, bl, curr_sp, *mem);
404 set_ia32_flags(pop, arch_irn_flags_ignore);
405 curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res);
406 curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
407 *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
409 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
410 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
413 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
414 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
418 * Produces the type which sits between the stack args and the locals on the stack.
419 * it will contain the return address and space to store the old base pointer.
420 * @return The Firm type modeling the ABI between type.
422 static ir_type *ia32_abi_get_between_type(void *self)
424 static ir_type *omit_fp_between_type = NULL;
425 static ir_type *between_type = NULL;
427 ia32_abi_env_t *env = self;
431 entity *ret_addr_ent;
432 entity *omit_fp_ret_addr_ent;
434 ir_type *old_bp_type = new_type_primitive(new_id_from_str("bp"), mode_P);
435 ir_type *ret_addr_type = new_type_primitive(new_id_from_str("return_addr"), mode_P);
437 between_type = new_type_class(new_id_from_str("ia32_between_type"));
438 old_bp_ent = new_entity(between_type, new_id_from_str("old_bp"), old_bp_type);
439 ret_addr_ent = new_entity(between_type, new_id_from_str("ret_addr"), ret_addr_type);
441 set_entity_offset_bytes(old_bp_ent, 0);
442 set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type));
443 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
445 omit_fp_between_type = new_type_class(new_id_from_str("ia32_between_type_omit_fp"));
446 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, new_id_from_str("ret_addr"), ret_addr_type);
448 set_entity_offset_bytes(omit_fp_ret_addr_ent, 0);
449 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
452 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
455 static const be_abi_callbacks_t ia32_abi_callbacks = {
458 ia32_abi_get_between_type,
459 ia32_abi_dont_save_regs,
464 /* fill register allocator interface */
466 static const arch_irn_ops_if_t ia32_irn_ops_if = {
467 ia32_get_irn_reg_req,
472 ia32_get_frame_entity,
476 ia32_irn_ops_t ia32_irn_ops = {
483 /**************************************************
486 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
487 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
488 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
489 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
492 **************************************************/
495 * Transforms the standard firm graph into
498 static void ia32_prepare_graph(void *self) {
499 ia32_code_gen_t *cg = self;
500 DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
502 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
504 /* 1st: transform psi condition trees */
505 irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_psi_cond_tree, cg);
507 /* 2nd: transform all remaining nodes */
508 ia32_register_transformers();
509 irg_walk_blkwise_graph(cg->irg, ia32_place_consts_set_modes, ia32_transform_node, cg);
510 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
512 /* 3rd: optimize address mode */
513 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
514 ia32_optimize_addressmode(cg);
515 be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
516 DEBUG_ONLY(cg->mod = old_mod;)
519 static INLINE int need_constraint_copy(ir_node *irn) {
521 ! is_ia32_Lea(irn) && \
522 ! is_ia32_Conv_I2I(irn) && \
523 ! is_ia32_Conv_I2I8Bit(irn) && \
524 ! is_ia32_CMov(irn) && \
529 * Insert copies for all ia32 nodes where the should_be_same requirement
531 * Transform Sub into Neg -- Add if IN2 == OUT
533 static void ia32_finish_node(ir_node *irn, void *env) {
534 ia32_code_gen_t *cg = env;
535 const ia32_register_req_t **reqs;
536 const arch_register_t *out_reg, *in_reg, *in2_reg;
538 ir_node *copy, *in_node, *block, *in2_node;
539 ia32_op_type_t op_tp;
541 if (is_ia32_irn(irn)) {
542 /* AM Dest nodes don't produce any values */
543 op_tp = get_ia32_op_type(irn);
544 if (op_tp == ia32_AddrModeD)
547 reqs = get_ia32_out_req_all(irn);
548 n_res = get_ia32_n_res(irn);
549 block = get_nodes_block(irn);
551 /* check all OUT requirements, if there is a should_be_same */
552 if ((op_tp == ia32_Normal || op_tp == ia32_AddrModeS) && need_constraint_copy(irn))
554 for (i = 0; i < n_res; i++) {
555 if (arch_register_req_is(&(reqs[i]->req), should_be_same)) {
556 /* get in and out register */
557 out_reg = get_ia32_out_reg(irn, i);
558 in_node = get_irn_n(irn, reqs[i]->same_pos);
559 in_reg = arch_get_irn_register(cg->arch_env, in_node);
561 /* don't copy ignore nodes */
562 if (arch_irn_is(cg->arch_env, in_node, ignore) && is_Proj(in_node))
565 /* check if in and out register are equal */
566 if (! REGS_ARE_EQUAL(out_reg, in_reg)) {
567 /* in case of a commutative op: just exchange the in's */
568 /* beware: the current op could be everything, so test for ia32 */
569 /* commutativity first before getting the second in */
570 if (is_ia32_commutative(irn)) {
571 in2_node = get_irn_n(irn, reqs[i]->same_pos ^ 1);
572 in2_reg = arch_get_irn_register(cg->arch_env, in2_node);
574 if (REGS_ARE_EQUAL(out_reg, in2_reg)) {
575 set_irn_n(irn, reqs[i]->same_pos, in2_node);
576 set_irn_n(irn, reqs[i]->same_pos ^ 1, in_node);
583 DBG((cg->mod, LEVEL_1, "inserting copy for %+F in_pos %d\n", irn, reqs[i]->same_pos));
584 /* create copy from in register */
585 copy = be_new_Copy(arch_register_get_class(in_reg), cg->irg, block, in_node);
587 DBG_OPT_2ADDRCPY(copy);
589 /* destination is the out register */
590 arch_set_irn_register(cg->arch_env, copy, out_reg);
592 /* insert copy before the node into the schedule */
593 sched_add_before(irn, copy);
596 set_irn_n(irn, reqs[i]->same_pos, copy);
603 /* If we have a CondJmp with immediate, we need to */
604 /* check if it's the right operand, otherwise we have */
605 /* to change it, as CMP doesn't support immediate as */
607 if (is_ia32_CondJmp(irn) && (is_ia32_ImmConst(irn) || is_ia32_ImmSymConst(irn)) && op_tp == ia32_AddrModeS) {
608 set_ia32_op_type(irn, ia32_AddrModeD);
609 set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
612 /* check if there is a sub which need to be transformed */
613 ia32_transform_sub_to_neg_add(irn, cg);
615 /* transform a LEA into an Add if possible */
616 //ia32_transform_lea_to_add(irn, cg);
620 /* check for peephole optimization */
621 ia32_peephole_optimization(irn, cg);
624 static void ia32_finish_irg_walker(ir_node *block, void *env) {
627 for (irn = sched_first(block); !sched_is_end(irn); irn = next) {
628 next = sched_next(irn);
629 ia32_finish_node(irn, env);
634 * Add Copy nodes for not fulfilled should_be_equal constraints
636 static void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) {
637 irg_block_walk_graph(irg, NULL, ia32_finish_irg_walker, cg);
643 * Dummy functions for hooks we don't need but which must be filled.
645 static void ia32_before_sched(void *self) {
649 * Called before the register allocator.
650 * Calculate a block schedule here. We need it for the x87
651 * simulator and the emitter.
653 static void ia32_before_ra(void *self) {
654 ia32_code_gen_t *cg = self;
656 cg->blk_sched = sched_create_block_schedule(cg->irg);
661 * Transforms a be node into a Load.
663 static void transform_to_Load(ia32_transform_env_t *env) {
664 ir_node *irn = env->irn;
665 entity *ent = be_get_frame_entity(irn);
666 ir_mode *mode = env->mode;
667 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
668 ir_node *nomem = new_rd_NoMem(env->irg);
669 ir_node *sched_point = NULL;
670 ir_node *ptr = get_irn_n(irn, 0);
671 ir_node *mem = be_is_Reload(irn) ? get_irn_n(irn, 1) : nomem;
672 ir_node *new_op, *proj;
673 const arch_register_t *reg;
675 if (sched_is_scheduled(irn)) {
676 sched_point = sched_prev(irn);
679 if (mode_is_float(mode)) {
680 if (USE_SSE2(env->cg))
681 new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
683 new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
686 new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
689 set_ia32_am_support(new_op, ia32_am_Source);
690 set_ia32_op_type(new_op, ia32_AddrModeS);
691 set_ia32_am_flavour(new_op, ia32_B);
692 set_ia32_ls_mode(new_op, mode);
693 set_ia32_frame_ent(new_op, ent);
694 set_ia32_use_frame(new_op);
696 DBG_OPT_RELOAD2LD(irn, new_op);
698 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_Load_res);
701 sched_add_after(sched_point, new_op);
702 sched_add_after(new_op, proj);
707 /* copy the register from the old node to the new Load */
708 reg = arch_get_irn_register(env->cg->arch_env, irn);
709 arch_set_irn_register(env->cg->arch_env, new_op, reg);
711 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, new_op));
717 * Transforms a be node into a Store.
719 static void transform_to_Store(ia32_transform_env_t *env) {
720 ir_node *irn = env->irn;
721 entity *ent = be_get_frame_entity(irn);
722 ir_mode *mode = env->mode;
723 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
724 ir_node *nomem = new_rd_NoMem(env->irg);
725 ir_node *ptr = get_irn_n(irn, 0);
726 ir_node *val = get_irn_n(irn, 1);
727 ir_node *new_op, *proj;
728 ir_node *sched_point = NULL;
730 if (sched_is_scheduled(irn)) {
731 sched_point = sched_prev(irn);
734 if (mode_is_float(mode)) {
735 if (USE_SSE2(env->cg))
736 new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
738 new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
740 else if (get_mode_size_bits(mode) == 8) {
741 new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
744 new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
747 set_ia32_am_support(new_op, ia32_am_Dest);
748 set_ia32_op_type(new_op, ia32_AddrModeD);
749 set_ia32_am_flavour(new_op, ia32_B);
750 set_ia32_ls_mode(new_op, mode);
751 set_ia32_frame_ent(new_op, ent);
752 set_ia32_use_frame(new_op);
754 DBG_OPT_SPILL2ST(irn, new_op);
756 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, pn_ia32_Store_M);
759 sched_add_after(sched_point, new_op);
760 sched_add_after(new_op, proj);
765 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, new_op));
771 * Fix the mode of Spill/Reload
773 static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode)
775 if (mode_is_float(mode)) {
787 * Block-Walker: Calls the transform functions Spill and Reload.
789 static void ia32_after_ra_walker(ir_node *block, void *env) {
790 ir_node *node, *prev;
791 ia32_code_gen_t *cg = env;
792 ia32_transform_env_t tenv;
795 tenv.irg = current_ir_graph;
797 DEBUG_ONLY(tenv.mod = cg->mod;)
799 /* beware: the schedule is changed here */
800 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
801 prev = sched_prev(node);
802 if (be_is_Reload(node)) {
803 /* we always reload the whole register */
804 tenv.dbg = get_irn_dbg_info(node);
806 tenv.mode = fix_spill_mode(cg, get_irn_mode(node));
807 transform_to_Load(&tenv);
809 else if (be_is_Spill(node)) {
810 /* we always spill the whole register */
811 tenv.dbg = get_irn_dbg_info(node);
813 tenv.mode = fix_spill_mode(cg, get_irn_mode(be_get_Spill_context(node)));
814 transform_to_Store(&tenv);
820 * We transform Spill and Reload here. This needs to be done before
821 * stack biasing otherwise we would miss the corrected offset for these nodes.
823 * If x87 instruction should be emitted, run the x87 simulator and patch
824 * the virtual instructions. This must obviously be done after register allocation.
826 static void ia32_after_ra(void *self) {
827 ia32_code_gen_t *cg = self;
828 irg_block_walk_graph(cg->irg, NULL, ia32_after_ra_walker, self);
830 /* if we do x87 code generation, rewrite all the virtual instructions and registers */
831 if (cg->used_fp == fp_x87) {
832 x87_simulate_graph(cg->arch_env, cg->irg, cg->blk_sched);
838 * Emits the code, closes the output file and frees
839 * the code generator interface.
841 static void ia32_codegen(void *self) {
842 ia32_code_gen_t *cg = self;
843 ir_graph *irg = cg->irg;
845 ia32_finish_irg(irg, cg);
846 be_dump(irg, "-finished", dump_ir_block_graph_sched);
847 ia32_gen_routine(cg->isa->out, irg, cg);
851 /* remove it from the isa */
854 /* de-allocate code generator */
855 del_set(cg->reg_set);
860 static void *ia32_cg_init(const be_irg_t *birg);
862 static const arch_code_generator_if_t ia32_code_gen_if = {
864 NULL, /* before abi introduce hook */
866 ia32_before_sched, /* before scheduling hook */
867 ia32_before_ra, /* before register allocation hook */
868 ia32_after_ra, /* after register allocation hook */
869 ia32_codegen /* emit && done */
873 * Initializes a IA32 code generator.
875 static void *ia32_cg_init(const be_irg_t *birg) {
876 ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa;
877 ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg));
879 cg->impl = &ia32_code_gen_if;
881 cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
882 cg->arch_env = birg->main_env->arch_env;
885 cg->blk_sched = NULL;
888 cg->fp_kind = isa->fp_kind;
889 cg->used_fp = fp_none;
891 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
893 /* copy optimizations from isa for easier access */
900 if (isa->name_obst_size) {
901 //printf("freed %d bytes from name obst\n", isa->name_obst_size);
902 isa->name_obst_size = 0;
903 obstack_free(isa->name_obst, NULL);
904 obstack_init(isa->name_obst);
908 cur_reg_set = cg->reg_set;
910 ia32_irn_ops.cg = cg;
912 return (arch_code_generator_t *)cg;
917 /*****************************************************************
918 * ____ _ _ _____ _____
919 * | _ \ | | | | |_ _|/ ____| /\
920 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
921 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
922 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
923 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
925 *****************************************************************/
928 * The template that generates a new ISA object.
929 * Note that this template can be changed by command line
932 static ia32_isa_t ia32_isa_template = {
934 &ia32_isa_if, /* isa interface implementation */
935 &ia32_gp_regs[REG_ESP], /* stack pointer register */
936 &ia32_gp_regs[REG_EBP], /* base pointer register */
937 -1, /* stack direction */
939 NULL, /* 16bit register names */
940 NULL, /* 8bit register names */
944 IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
945 IA32_OPT_DOAM | /* optimize address mode default: on */
946 IA32_OPT_LEA | /* optimize for LEAs default: on */
947 IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */
948 IA32_OPT_IMMOPS | /* operations can use immediates, default: on */
949 IA32_OPT_EXTBB), /* use extended basic block scheduling, default: on */
950 arch_pentium_4, /* instruction architecture */
951 arch_pentium_4, /* optimize for architecture */
952 fp_sse2, /* use sse2 unit */
953 NULL, /* current code generator */
955 NULL, /* name obstack */
956 0 /* name obst size */
961 * Initializes the backend ISA.
963 static void *ia32_init(FILE *file_handle) {
964 static int inited = 0;
970 isa = xmalloc(sizeof(*isa));
971 memcpy(isa, &ia32_isa_template, sizeof(*isa));
973 ia32_register_init(isa);
974 ia32_create_opcodes();
976 if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
977 (ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
978 /* no SSE2 for these cpu's */
979 isa->fp_kind = fp_x87;
981 if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) {
982 /* Pentium 4 don't like inc and dec instructions */
983 isa->opt &= ~IA32_OPT_INCDEC;
986 isa->regs_16bit = pmap_create();
987 isa->regs_8bit = pmap_create();
988 isa->types = pmap_create();
989 isa->tv_ent = pmap_create();
990 isa->out = file_handle;
992 ia32_build_16bit_reg_map(isa->regs_16bit);
993 ia32_build_8bit_reg_map(isa->regs_8bit);
995 /* patch register names of x87 registers */
997 ia32_st_regs[0].name = "st";
998 ia32_st_regs[1].name = "st(1)";
999 ia32_st_regs[2].name = "st(2)";
1000 ia32_st_regs[3].name = "st(3)";
1001 ia32_st_regs[4].name = "st(4)";
1002 ia32_st_regs[5].name = "st(5)";
1003 ia32_st_regs[6].name = "st(6)";
1004 ia32_st_regs[7].name = "st(7)";
1008 isa->name_obst = xmalloc(sizeof(*isa->name_obst));
1009 obstack_init(isa->name_obst);
1010 isa->name_obst_size = 0;
1013 fprintf(isa->out, "\t.intel_syntax\n");
1023 * Closes the output file and frees the ISA structure.
1025 static void ia32_done(void *self) {
1026 ia32_isa_t *isa = self;
1028 /* emit now all global declarations */
1029 ia32_gen_decls(isa->out);
1031 pmap_destroy(isa->regs_16bit);
1032 pmap_destroy(isa->regs_8bit);
1033 pmap_destroy(isa->tv_ent);
1034 pmap_destroy(isa->types);
1037 //printf("name obst size = %d bytes\n", isa->name_obst_size);
1038 obstack_free(isa->name_obst, NULL);
1046 * Return the number of register classes for this architecture.
1047 * We report always these:
1048 * - the general purpose registers
1049 * - the floating point register set (depending on the unit used for FP)
1050 * - MMX/SSE registers (currently not supported)
1052 static int ia32_get_n_reg_class(const void *self) {
1057 * Return the register class for index i.
1059 static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
1060 const ia32_isa_t *isa = self;
1061 assert(i >= 0 && i < 2 && "Invalid ia32 register class requested.");
1063 return &ia32_reg_classes[CLASS_ia32_gp];
1064 return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1068 * Get the register class which shall be used to store a value of a given mode.
1069 * @param self The this pointer.
1070 * @param mode The mode in question.
1071 * @return A register class which can hold values of the given mode.
1073 const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
1074 const ia32_isa_t *isa = self;
1075 if (mode_is_float(mode)) {
1076 return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1079 return &ia32_reg_classes[CLASS_ia32_gp];
1083 * Get the ABI restrictions for procedure calls.
1084 * @param self The this pointer.
1085 * @param method_type The type of the method (procedure) in question.
1086 * @param abi The abi object to be modified
1088 static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1089 const ia32_isa_t *isa = self;
1092 unsigned cc = get_method_calling_convention(method_type);
1093 int n = get_method_n_params(method_type);
1096 int i, ignore_1, ignore_2;
1098 const arch_register_t *reg;
1099 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1101 unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
1103 /* set abi flags for calls */
1104 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1105 call_flags.bits.store_args_sequential = use_push;
1106 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1107 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1108 call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
1110 /* set stack parameter passing style */
1111 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1113 /* collect the mode for each type */
1114 modes = alloca(n * sizeof(modes[0]));
1116 for (i = 0; i < n; i++) {
1117 tp = get_method_param_type(method_type, i);
1118 modes[i] = get_type_mode(tp);
1121 /* set register parameters */
1122 if (cc & cc_reg_param) {
1123 /* determine the number of parameters passed via registers */
1124 biggest_n = ia32_get_n_regparam_class(n, modes, &ignore_1, &ignore_2);
1126 /* loop over all parameters and set the register requirements */
1127 for (i = 0; i <= biggest_n; i++) {
1128 reg = ia32_get_RegParam_reg(n, modes, i, cc);
1129 assert(reg && "kaputt");
1130 be_abi_call_param_reg(abi, i, reg);
1137 /* set stack parameters */
1138 for (i = stack_idx; i < n; i++) {
1139 be_abi_call_param_stack(abi, i, 1, 0, 0);
1143 /* set return registers */
1144 n = get_method_n_ress(method_type);
1146 assert(n <= 2 && "more than two results not supported");
1148 /* In case of 64bit returns, we will have two 32bit values */
1150 tp = get_method_res_type(method_type, 0);
1151 mode = get_type_mode(tp);
1153 assert(!mode_is_float(mode) && "two FP results not supported");
1155 tp = get_method_res_type(method_type, 1);
1156 mode = get_type_mode(tp);
1158 assert(!mode_is_float(mode) && "two FP results not supported");
1160 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX]);
1161 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]);
1164 const arch_register_t *reg;
1166 tp = get_method_res_type(method_type, 0);
1167 assert(is_atomic_type(tp));
1168 mode = get_type_mode(tp);
1170 reg = mode_is_float(mode) ?
1171 (USE_SSE2(isa) ? &ia32_xmm_regs[REG_XMM0] : &ia32_vfp_regs[REG_VF0]) :
1172 &ia32_gp_regs[REG_EAX];
1174 be_abi_call_res_reg(abi, 0, reg);
1179 static const void *ia32_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1180 return &ia32_irn_ops;
1183 const arch_irn_handler_t ia32_irn_handler = {
1187 const arch_irn_handler_t *ia32_get_irn_handler(const void *self) {
1188 return &ia32_irn_handler;
1191 int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1192 return is_ia32_irn(irn);
1196 * Initializes the code generator interface.
1198 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self) {
1199 return &ia32_code_gen_if;
1202 list_sched_selector_t ia32_sched_selector;
1205 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1207 static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self) {
1208 // memcpy(&ia32_sched_selector, reg_pressure_selector, sizeof(list_sched_selector_t));
1209 memcpy(&ia32_sched_selector, trivial_selector, sizeof(list_sched_selector_t));
1210 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
1211 return &ia32_sched_selector;
1215 * Returns the necessary byte alignment for storing a register of given class.
1217 static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1218 ir_mode *mode = arch_register_class_mode(cls);
1219 int bytes = get_mode_size_bytes(mode);
1221 if (mode_is_float(mode) && bytes > 8)
1228 /* instruction set architectures. */
1229 static const lc_opt_enum_int_items_t arch_items[] = {
1230 { "386", arch_i386, },
1231 { "486", arch_i486, },
1232 { "pentium", arch_pentium, },
1233 { "586", arch_pentium, },
1234 { "pentiumpro", arch_pentium_pro, },
1235 { "686", arch_pentium_pro, },
1236 { "pentiummmx", arch_pentium_mmx, },
1237 { "pentium2", arch_pentium_2, },
1238 { "p2", arch_pentium_2, },
1239 { "pentium3", arch_pentium_3, },
1240 { "p3", arch_pentium_3, },
1241 { "pentium4", arch_pentium_4, },
1242 { "p4", arch_pentium_4, },
1243 { "pentiumm", arch_pentium_m, },
1244 { "pm", arch_pentium_m, },
1245 { "core", arch_core, },
1247 { "athlon", arch_athlon, },
1248 { "athlon64", arch_athlon_64, },
1249 { "opteron", arch_opteron, },
1253 static lc_opt_enum_int_var_t arch_var = {
1254 &ia32_isa_template.arch, arch_items
1257 static lc_opt_enum_int_var_t opt_arch_var = {
1258 &ia32_isa_template.opt_arch, arch_items
1261 static const lc_opt_enum_int_items_t fp_unit_items[] = {
1263 { "sse2", fp_sse2 },
1267 static lc_opt_enum_int_var_t fp_unit_var = {
1268 &ia32_isa_template.fp_kind, fp_unit_items
1271 static const lc_opt_enum_int_items_t gas_items[] = {
1272 { "linux", ASM_LINUX_GAS },
1273 { "mingw", ASM_MINGW_GAS },
1277 static lc_opt_enum_int_var_t gas_var = {
1278 (int *)&asm_flavour, gas_items
1281 static const lc_opt_table_entry_t ia32_options[] = {
1282 LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var),
1283 LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var),
1284 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var),
1285 LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
1286 LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA),
1287 LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST),
1288 LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS),
1289 LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB),
1290 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
1295 * Register command line options for the ia32 backend.
1299 * ia32-arch=arch create instruction for arch
1300 * ia32-opt=arch optimize for run on arch
1301 * ia32-fpunit=unit select floating point unit (x87 or SSE2)
1302 * ia32-incdec optimize for inc/dec
1303 * ia32-noaddrmode do not use address mode
1304 * ia32-nolea do not optimize for LEAs
1305 * ia32-noplacecnst do not place constants,
1306 * ia32-noimmop no operations with immediates
1307 * ia32-noextbb do not use extended basic block scheduling
1308 * ia32-gasmode set the GAS compatibility mode
1310 static void ia32_register_options(lc_opt_entry_t *ent)
1312 lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32");
1313 lc_opt_add_table(be_grp_ia32, ia32_options);
1315 #endif /* WITH_LIBCORE */
1317 const arch_isa_if_t ia32_isa_if = {
1320 ia32_get_n_reg_class,
1322 ia32_get_reg_class_for_mode,
1324 ia32_get_irn_handler,
1325 ia32_get_code_generator_if,
1326 ia32_get_list_sched_selector,
1327 ia32_get_reg_class_alignment,
1329 ia32_register_options