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 */
26 #include "pseudo_irg.h"
30 #include "iredges_t.h"
38 #include "../beabi.h" /* the general register allocator interface */
39 #include "../benode_t.h"
40 #include "../belower.h"
41 #include "../besched_t.h"
44 #include "../beirgmod.h"
45 #include "../be_dbgout.h"
46 #include "bearch_ia32_t.h"
48 #include "ia32_new_nodes.h" /* ia32 nodes interface */
49 #include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */
50 #include "ia32_gen_decls.h" /* interface declaration emitter */
51 #include "ia32_transform.h"
52 #include "ia32_emitter.h"
53 #include "ia32_map_regs.h"
54 #include "ia32_optimize.h"
56 #include "ia32_dbg_stat.h"
57 #include "ia32_finish.h"
58 #include "ia32_util.h"
60 #define DEBUG_MODULE "firm.be.ia32.isa"
63 static set *cur_reg_set = NULL;
66 #define is_Start(irn) (get_irn_opcode(irn) == iro_Start)
68 /* Creates the unique per irg GP NoReg node. */
69 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) {
70 return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_GP_NOREG]);
73 /* Creates the unique per irg FP NoReg node. */
74 ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) {
75 return be_abi_get_callee_save_irn(cg->birg->abi,
76 USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]);
79 /**************************************************
82 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
83 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
84 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
85 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
88 **************************************************/
90 static ir_node *my_skip_proj(const ir_node *n) {
98 * Return register requirements for an ia32 node.
99 * If the node returns a tuple (mode_T) then the proj's
100 * will be asked for this information.
102 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) {
103 const ia32_irn_ops_t *ops = self;
104 const ia32_register_req_t *irn_req;
105 long node_pos = pos == -1 ? 0 : pos;
106 ir_mode *mode = is_Block(irn) ? NULL : get_irn_mode(irn);
107 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, DEBUG_MODULE);
109 if (is_Block(irn) || mode == mode_M || mode == mode_X) {
110 DBG((mod, LEVEL_1, "ignoring Block, mode_M, mode_X node %+F\n", irn));
114 if (mode == mode_T && pos < 0) {
115 DBG((mod, LEVEL_1, "ignoring request OUT requirements for node %+F\n", irn));
119 DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn));
123 node_pos = ia32_translate_proj_pos(irn);
129 irn = my_skip_proj(irn);
131 DB((mod, LEVEL_1, "skipping Proj, going to %+F at pos %d ... ", irn, node_pos));
134 if (is_ia32_irn(irn)) {
136 irn_req = get_ia32_in_req(irn, pos);
139 irn_req = get_ia32_out_req(irn, node_pos);
142 DB((mod, LEVEL_1, "returning reqs for %+F at pos %d\n", irn, pos));
144 memcpy(req, &(irn_req->req), sizeof(*req));
146 if (arch_register_req_is(&(irn_req->req), should_be_same)) {
147 assert(irn_req->same_pos >= 0 && "should be same constraint for in -> out NYI");
148 req->other_same = get_irn_n(irn, irn_req->same_pos);
151 if (arch_register_req_is(&(irn_req->req), should_be_different)) {
152 assert(irn_req->different_pos >= 0 && "should be different constraint for in -> out NYI");
153 req->other_different = get_irn_n(irn, irn_req->different_pos);
157 /* treat Unknowns like Const with default requirements */
158 if (is_Unknown(irn)) {
159 DB((mod, LEVEL_1, "returning UKNWN reqs for %+F\n", irn));
160 if (mode_is_float(mode)) {
161 if (USE_SSE2(ops->cg))
162 memcpy(req, &(ia32_default_req_ia32_xmm_xmm_UKNWN), sizeof(*req));
164 memcpy(req, &(ia32_default_req_ia32_vfp_vfp_UKNWN), sizeof(*req));
166 else if (mode_is_int(mode) || mode_is_reference(mode))
167 memcpy(req, &(ia32_default_req_ia32_gp_gp_UKNWN), sizeof(*req));
168 else if (mode == mode_T || mode == mode_M) {
169 DBG((mod, LEVEL_1, "ignoring Unknown node %+F\n", irn));
173 assert(0 && "unsupported Unknown-Mode");
176 DB((mod, LEVEL_1, "returning NULL for %+F (not ia32)\n", irn));
184 static void ia32_set_irn_reg(const void *self, ir_node *irn, const arch_register_t *reg) {
186 const ia32_irn_ops_t *ops = self;
188 if (get_irn_mode(irn) == mode_X) {
192 DBG((ops->cg->mod, LEVEL_1, "ia32 assigned register %s to node %+F\n", reg->name, irn));
195 pos = ia32_translate_proj_pos(irn);
196 irn = my_skip_proj(irn);
199 if (is_ia32_irn(irn)) {
200 const arch_register_t **slots;
202 slots = get_ia32_slots(irn);
206 ia32_set_firm_reg(irn, reg, cur_reg_set);
210 static const arch_register_t *ia32_get_irn_reg(const void *self, const ir_node *irn) {
212 const arch_register_t *reg = NULL;
216 if (get_irn_mode(irn) == mode_X) {
220 pos = ia32_translate_proj_pos(irn);
221 irn = my_skip_proj(irn);
224 if (is_ia32_irn(irn)) {
225 const arch_register_t **slots;
226 slots = get_ia32_slots(irn);
230 reg = ia32_get_firm_reg(irn, cur_reg_set);
236 static arch_irn_class_t ia32_classify(const void *self, const ir_node *irn) {
237 arch_irn_class_t classification = arch_irn_class_normal;
239 irn = my_skip_proj(irn);
242 classification |= arch_irn_class_branch;
244 if (! is_ia32_irn(irn))
245 return classification & ~arch_irn_class_normal;
247 if (is_ia32_Cnst(irn))
248 classification |= arch_irn_class_const;
251 classification |= arch_irn_class_load;
253 if (is_ia32_St(irn) || is_ia32_Store8Bit(irn))
254 classification |= arch_irn_class_store;
256 if (is_ia32_got_reload(irn))
257 classification |= arch_irn_class_reload;
259 return classification;
262 static arch_irn_flags_t ia32_get_flags(const void *self, const ir_node *irn) {
265 ir_node *pred = get_Proj_pred(irn);
266 if(is_ia32_Push(pred) && get_Proj_proj(irn) == pn_ia32_Push_stack) {
267 return arch_irn_flags_modify_sp;
269 if(is_ia32_Pop(pred) && get_Proj_proj(irn) == pn_ia32_Pop_stack) {
270 return arch_irn_flags_modify_sp;
272 if(is_ia32_AddSP(pred) && get_Proj_proj(irn) == pn_ia32_AddSP_stack) {
273 return arch_irn_flags_modify_sp;
277 irn = my_skip_proj(irn);
278 if (is_ia32_irn(irn))
279 return get_ia32_flags(irn);
282 return arch_irn_flags_ignore;
287 static entity *ia32_get_frame_entity(const void *self, const ir_node *irn) {
288 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
291 static void ia32_set_frame_entity(const void *self, ir_node *irn, entity *ent) {
292 set_ia32_frame_ent(irn, ent);
295 static void ia32_set_frame_offset(const void *self, ir_node *irn, int bias) {
297 const ia32_irn_ops_t *ops = self;
299 if (get_ia32_frame_ent(irn)) {
300 ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn);
302 /* Pop nodes modify the stack pointer before calculating the destination
303 * address, so fix this here
305 if(is_ia32_Pop(irn)) {
309 DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias));
311 snprintf(buf, sizeof(buf), "%d", bias);
313 if (get_ia32_op_type(irn) == ia32_Normal) {
314 set_ia32_cnst(irn, buf);
316 add_ia32_am_offs(irn, buf);
318 set_ia32_am_flavour(irn, am_flav);
323 static int ia32_get_sp_bias(const void *self, const ir_node *irn) {
325 int proj = get_Proj_proj(irn);
326 ir_node *pred = get_Proj_pred(irn);
328 if(is_ia32_Push(pred) && proj == 0)
330 if(is_ia32_Pop(pred) && proj == 1)
338 be_abi_call_flags_bits_t flags;
339 const arch_isa_t *isa;
340 const arch_env_t *aenv;
344 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
346 ia32_abi_env_t *env = xmalloc(sizeof(env[0]));
347 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
348 env->flags = fl.bits;
351 env->isa = aenv->isa;
356 * Put all registers which are saved by the prologue/epilogue in a set.
358 * @param self The callback object.
359 * @param s The result set.
361 static void ia32_abi_dont_save_regs(void *self, pset *s)
363 ia32_abi_env_t *env = self;
364 if(env->flags.try_omit_fp)
365 pset_insert_ptr(s, env->isa->bp);
369 * Generate the routine prologue.
371 * @param self The callback object.
372 * @param mem A pointer to the mem node. Update this if you define new memory.
373 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
375 * @return The register which shall be used as a stack frame base.
377 * All nodes which define registers in @p reg_map must keep @p reg_map current.
379 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
381 ia32_abi_env_t *env = self;
383 if (! env->flags.try_omit_fp) {
384 ir_node *bl = get_irg_start_block(env->irg);
385 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
386 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
390 push = new_rd_ia32_Push(NULL, env->irg, bl, curr_sp, curr_bp, *mem);
391 curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
392 *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M);
394 /* the push must have SP out register */
395 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
396 set_ia32_flags(push, arch_irn_flags_ignore);
398 /* move esp to ebp */
399 curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp);
400 be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp);
401 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
402 be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
404 /* beware: the copy must be done before any other sp use */
405 curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
406 be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp);
407 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
408 be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
410 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
411 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
420 * Generate the routine epilogue.
421 * @param self The callback object.
422 * @param bl The block for the epilog
423 * @param mem A pointer to the mem node. Update this if you define new memory.
424 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
425 * @return The register which shall be used as a stack frame base.
427 * All nodes which define registers in @p reg_map must keep @p reg_map current.
429 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
431 ia32_abi_env_t *env = self;
432 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
433 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
435 if (env->flags.try_omit_fp) {
436 /* simply remove the stack frame here */
437 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK);
438 add_irn_dep(curr_sp, *mem);
441 const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
442 ir_mode *mode_bp = env->isa->bp->reg_class->mode;
444 /* gcc always emits a leave at the end of a routine */
445 if (1 || ARCH_AMD(isa->opt_arch)) {
449 leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, curr_bp);
450 set_ia32_flags(leave, arch_irn_flags_ignore);
451 curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame);
452 curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
453 *mem = new_r_Proj(current_ir_graph, bl, leave, mode_M, pn_ia32_Leave_M);
458 /* copy ebp to esp */
459 curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
462 pop = new_rd_ia32_Pop(NULL, env->irg, bl, curr_sp, *mem);
463 set_ia32_flags(pop, arch_irn_flags_ignore);
464 curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res);
465 curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
466 *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
468 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
469 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
472 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
473 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
477 * Produces the type which sits between the stack args and the locals on the stack.
478 * it will contain the return address and space to store the old base pointer.
479 * @return The Firm type modeling the ABI between type.
481 static ir_type *ia32_abi_get_between_type(void *self)
483 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
484 static ir_type *omit_fp_between_type = NULL;
485 static ir_type *between_type = NULL;
487 ia32_abi_env_t *env = self;
489 if ( !between_type) {
491 entity *ret_addr_ent;
492 entity *omit_fp_ret_addr_ent;
494 ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P);
495 ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_P);
497 between_type = new_type_struct(IDENT("ia32_between_type"));
498 old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
499 ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
501 set_entity_offset_bytes(old_bp_ent, 0);
502 set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type));
503 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
504 set_type_state(between_type, layout_fixed);
506 omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
507 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
509 set_entity_offset_bytes(omit_fp_ret_addr_ent, 0);
510 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
511 set_type_state(omit_fp_between_type, layout_fixed);
514 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
519 * Get the estimated cycle count for @p irn.
521 * @param self The this pointer.
522 * @param irn The node.
524 * @return The estimated cycle count for this operation
526 static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
529 ia32_op_type_t op_tp;
530 const ia32_irn_ops_t *ops = self;
535 assert(is_ia32_irn(irn));
537 cost = get_ia32_latency(irn);
538 op_tp = get_ia32_op_type(irn);
540 if (is_ia32_CopyB(irn)) {
542 if (ARCH_INTEL(ops->cg->arch))
545 else if (is_ia32_CopyB_i(irn)) {
546 int size = get_tarval_long(get_ia32_Immop_tarval(irn));
547 cost = 20 + (int)ceil((4/3) * size);
548 if (ARCH_INTEL(ops->cg->arch))
551 /* in case of address mode operations add additional cycles */
552 else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) {
554 In case of stack access add 5 cycles (we assume stack is in cache),
555 other memory operations cost 20 cycles.
557 cost += is_ia32_use_frame(irn) ? 5 : 20;
564 * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
566 * @param irn The original operation
567 * @param i Index of the argument we want the inverse operation to yield
568 * @param inverse struct to be filled with the resulting inverse op
569 * @param obstack The obstack to use for allocation of the returned nodes array
570 * @return The inverse operation or NULL if operation invertible
572 static arch_inverse_t *ia32_get_inverse(const void *self, const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) {
575 ir_node *block, *noreg, *nomem;
578 /* we cannot invert non-ia32 irns */
579 if (! is_ia32_irn(irn))
582 /* operand must always be a real operand (not base, index or mem) */
583 if (i != 2 && i != 3)
586 /* we don't invert address mode operations */
587 if (get_ia32_op_type(irn) != ia32_Normal)
590 irg = get_irn_irg(irn);
591 block = get_nodes_block(irn);
592 mode = get_ia32_res_mode(irn);
593 noreg = get_irn_n(irn, 0);
594 nomem = new_r_NoMem(irg);
596 /* initialize structure */
597 inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
601 switch (get_ia32_irn_opcode(irn)) {
603 if (get_ia32_immop_type(irn) == ia32_ImmConst) {
604 /* we have an add with a const here */
605 /* invers == add with negated const */
606 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
607 pnc = pn_ia32_Add_res;
609 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
610 set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
611 set_ia32_commutative(inverse->nodes[0]);
613 else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
614 /* we have an add with a symconst here */
615 /* invers == sub with const */
616 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
617 pnc = pn_ia32_Sub_res;
619 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
622 /* normal add: inverse == sub */
623 ir_node *proj = ia32_get_res_proj(irn);
626 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem);
627 pnc = pn_ia32_Sub_res;
632 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
633 /* we have a sub with a const/symconst here */
634 /* invers == add with this const */
635 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
636 pnc = pn_ia32_Add_res;
637 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
638 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
642 ir_node *proj = ia32_get_res_proj(irn);
646 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem);
649 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, 2), proj, nomem);
651 pnc = pn_ia32_Sub_res;
656 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
657 /* xor with const: inverse = xor */
658 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
659 pnc = pn_ia32_Eor_res;
660 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
661 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
665 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, (ir_node *)irn, get_irn_n(irn, i), nomem);
666 pnc = pn_ia32_Eor_res;
671 ir_node *proj = ia32_get_res_proj(irn);
674 inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, proj, nomem);
675 pnc = pn_ia32_Not_res;
679 case iro_ia32_Minus: {
680 ir_node *proj = ia32_get_res_proj(irn);
683 inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, proj, nomem);
684 pnc = pn_ia32_Minus_res;
689 /* inverse operation not supported */
693 set_ia32_res_mode(inverse->nodes[0], mode);
694 inverse->nodes[1] = new_r_Proj(irg, block, inverse->nodes[0], mode, pnc);
700 * Check if irn can load it's operand at position i from memory (source addressmode).
701 * @param self Pointer to irn ops itself
702 * @param irn The irn to be checked
703 * @param i The operands position
704 * @return Non-Zero if operand can be loaded
706 static int ia32_possible_memory_operand(const void *self, const ir_node *irn, unsigned int i) {
707 if (! is_ia32_irn(irn) || /* must be an ia32 irn */
708 get_irn_arity(irn) != 5 || /* must be a binary operation */
709 get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
710 ! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */
711 (i != 2 && i != 3) || /* a "real" operand position must be requested */
712 (i == 2 && ! is_ia32_commutative(irn)) || /* if first operand requested irn must be commutative */
713 is_ia32_use_frame(irn)) /* must not already use frame */
719 static void ia32_perform_memory_operand(const void *self, ir_node *irn, ir_node *spill, unsigned int i) {
720 assert(ia32_possible_memory_operand(self, irn, i) && "Cannot perform memory operand change");
723 ir_node *tmp = get_irn_n(irn, 3);
724 set_irn_n(irn, 3, get_irn_n(irn, 2));
725 set_irn_n(irn, 2, tmp);
728 set_ia32_am_support(irn, ia32_am_Source);
729 set_ia32_op_type(irn, ia32_AddrModeS);
730 set_ia32_am_flavour(irn, ia32_B);
731 set_ia32_ls_mode(irn, get_irn_mode(get_irn_n(irn, i)));
732 //TODO this will fail, if spill is a PhiM (give PhiMs entities?)
733 set_ia32_frame_ent(irn, be_get_frame_entity(spill));
734 set_ia32_use_frame(irn);
735 set_ia32_got_reload(irn);
737 set_irn_n(irn, 0, get_irg_frame(get_irn_irg(irn)));
738 set_irn_n(irn, 4, spill);
741 Input at position one is index register, which is NoReg.
742 We would need cg object to get a real noreg, but we cannot
745 set_irn_n(irn, 3, get_irn_n(irn, 1));
747 //FIXME DBG_OPT_AM_S(reload, irn);
750 static const be_abi_callbacks_t ia32_abi_callbacks = {
753 ia32_abi_get_between_type,
754 ia32_abi_dont_save_regs,
759 /* fill register allocator interface */
761 static const arch_irn_ops_if_t ia32_irn_ops_if = {
762 ia32_get_irn_reg_req,
767 ia32_get_frame_entity,
768 ia32_set_frame_entity,
769 ia32_set_frame_offset,
772 ia32_get_op_estimated_cost,
773 ia32_possible_memory_operand,
774 ia32_perform_memory_operand,
777 ia32_irn_ops_t ia32_irn_ops = {
784 /**************************************************
787 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
788 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
789 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
790 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
793 **************************************************/
795 static void ia32_kill_convs(ia32_code_gen_t *cg) {
798 /* BEWARE: the Projs are inserted in the set */
799 foreach_nodeset(cg->kill_conv, irn) {
800 ir_node *in = get_irn_n(get_Proj_pred(irn), 2);
801 edges_reroute(irn, in, cg->birg->irg);
806 * Transform the Thread Local Store base.
808 static void transform_tls(ir_graph *irg) {
809 ir_node *irn = get_irg_tls(irg);
812 dbg_info *dbg = get_irn_dbg_info(irn);
813 ir_node *blk = get_nodes_block(irn);
815 newn = new_rd_ia32_LdTls(dbg, irg, blk, get_irn_mode(irn));
822 * Transforms the standard firm graph into
825 static void ia32_prepare_graph(void *self) {
826 ia32_code_gen_t *cg = self;
827 dom_front_info_t *dom;
828 DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
830 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
832 /* 1st: transform constants and psi condition trees */
833 ia32_pre_transform_phase(cg);
835 /* 2nd: transform all remaining nodes */
836 ia32_register_transformers();
837 dom = be_compute_dominance_frontiers(cg->irg);
839 cg->kill_conv = new_nodeset(5);
840 transform_tls(cg->irg);
841 irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg);
843 del_nodeset(cg->kill_conv);
845 be_free_dominance_frontiers(dom);
848 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
850 /* 3rd: optimize address mode */
851 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
852 ia32_optimize_addressmode(cg);
855 be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
857 DEBUG_ONLY(cg->mod = old_mod;)
861 * Dummy functions for hooks we don't need but which must be filled.
863 static void ia32_before_sched(void *self) {
866 static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) {
869 ir_node *mem_proj = NULL;
874 mode = get_irn_mode(irn);
876 /* check if we already saw this node or the node has more than one user */
877 if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1)
880 /* mark irn visited */
881 bitset_add_irn(already_visited, irn);
883 /* non-Tuple nodes with one user: ok, return */
884 if (get_irn_n_edges(irn) >= 1 && mode != mode_T)
887 /* tuple node has one user which is not the mem proj-> ok */
888 if (mode == mode_T && get_irn_n_edges(irn) == 1) {
889 mem_proj = ia32_get_proj_for_mode(irn, mode_M);
890 if (mem_proj == NULL)
894 arity = get_irn_arity(irn);
895 for (i = 0; i < arity; ++i) {
896 ir_node *pred = get_irn_n(irn, i);
898 /* do not follow memory edges or we will accidentally remove stores */
899 if (get_irn_mode(pred) == mode_M) {
900 if(mem_proj != NULL) {
901 edges_reroute(mem_proj, pred, get_irn_irg(mem_proj));
907 set_irn_n(irn, i, new_Bad());
910 The current node is about to be removed: if the predecessor
911 has only this node as user, it need to be removed as well.
913 if (get_irn_n_edges(pred) <= 1)
914 remove_unused_nodes(pred, already_visited);
917 // we need to set the presd to Bad again to also get the memory edges
918 arity = get_irn_arity(irn);
919 for (i = 0; i < arity; ++i) {
920 set_irn_n(irn, i, new_Bad());
923 if (sched_is_scheduled(irn)) {
928 static void remove_unused_loads_walker(ir_node *irn, void *env) {
929 bitset_t *already_visited = env;
930 if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn))
931 remove_unused_nodes(irn, env);
935 * Called before the register allocator.
936 * Calculate a block schedule here. We need it for the x87
937 * simulator and the emitter.
939 static void ia32_before_ra(void *self) {
940 ia32_code_gen_t *cg = self;
941 bitset_t *already_visited = bitset_irg_alloca(cg->irg);
945 There are sometimes unused loads, only pinned by memory.
946 We need to remove those Loads and all other nodes which won't be used
947 after removing the Load from schedule.
949 irg_walk_graph(cg->irg, remove_unused_loads_walker, NULL, already_visited);
954 * Transforms a be node into a Load.
956 static void transform_to_Load(ia32_transform_env_t *env) {
957 ir_node *irn = env->irn;
958 entity *ent = be_get_frame_entity(irn);
959 ir_mode *mode = env->mode;
960 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
961 ir_node *nomem = new_rd_NoMem(env->irg);
962 ir_node *sched_point = NULL;
963 ir_node *ptr = get_irn_n(irn, 0);
964 ir_node *mem = be_is_Reload(irn) ? get_irn_n(irn, 1) : nomem;
965 ir_node *new_op, *proj;
966 const arch_register_t *reg;
968 if (sched_is_scheduled(irn)) {
969 sched_point = sched_prev(irn);
972 if (mode_is_float(mode)) {
973 if (USE_SSE2(env->cg))
974 new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
976 new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
979 new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
982 set_ia32_am_support(new_op, ia32_am_Source);
983 set_ia32_op_type(new_op, ia32_AddrModeS);
984 set_ia32_am_flavour(new_op, ia32_B);
985 set_ia32_ls_mode(new_op, mode);
986 set_ia32_frame_ent(new_op, ent);
987 set_ia32_use_frame(new_op);
989 DBG_OPT_RELOAD2LD(irn, new_op);
991 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_Load_res);
994 sched_add_after(sched_point, new_op);
995 sched_add_after(new_op, proj);
1000 /* copy the register from the old node to the new Load */
1001 reg = arch_get_irn_register(env->cg->arch_env, irn);
1002 arch_set_irn_register(env->cg->arch_env, new_op, reg);
1004 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
1006 exchange(irn, proj);
1010 * Transforms a be node into a Store.
1012 static void transform_to_Store(ia32_transform_env_t *env) {
1013 ir_node *irn = env->irn;
1014 entity *ent = be_get_frame_entity(irn);
1015 ir_mode *mode = env->mode;
1016 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1017 ir_node *nomem = new_rd_NoMem(env->irg);
1018 ir_node *ptr = get_irn_n(irn, 0);
1019 ir_node *val = get_irn_n(irn, 1);
1020 ir_node *new_op, *proj;
1021 ir_node *sched_point = NULL;
1023 if (sched_is_scheduled(irn)) {
1024 sched_point = sched_prev(irn);
1027 if (mode_is_float(mode)) {
1028 if (USE_SSE2(env->cg))
1029 new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1031 new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1033 else if (get_mode_size_bits(mode) == 8) {
1034 new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1037 new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1040 set_ia32_am_support(new_op, ia32_am_Dest);
1041 set_ia32_op_type(new_op, ia32_AddrModeD);
1042 set_ia32_am_flavour(new_op, ia32_B);
1043 set_ia32_ls_mode(new_op, mode);
1044 set_ia32_frame_ent(new_op, ent);
1045 set_ia32_use_frame(new_op);
1047 DBG_OPT_SPILL2ST(irn, new_op);
1049 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, pn_ia32_Store_M);
1052 sched_add_after(sched_point, new_op);
1056 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
1058 exchange(irn, proj);
1061 static ir_node *create_push(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, ir_node *mem, entity *ent, const char *offset) {
1062 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1064 ir_node *push = new_rd_ia32_Push(env->dbg, env->irg, env->block, sp, noreg, mem);
1066 set_ia32_frame_ent(push, ent);
1067 set_ia32_use_frame(push);
1068 set_ia32_op_type(push, ia32_AddrModeS);
1069 set_ia32_am_flavour(push, ia32_B);
1070 set_ia32_ls_mode(push, mode_Is);
1072 add_ia32_am_offs(push, offset);
1074 sched_add_before(schedpoint, push);
1078 static ir_node *create_pop(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, entity *ent, const char *offset) {
1079 ir_node *pop = new_rd_ia32_Pop(env->dbg, env->irg, env->block, sp, new_NoMem());
1081 set_ia32_frame_ent(pop, ent);
1082 set_ia32_use_frame(pop);
1083 set_ia32_op_type(pop, ia32_AddrModeD);
1084 set_ia32_am_flavour(pop, ia32_B);
1085 set_ia32_ls_mode(pop, mode_Is);
1087 add_ia32_am_offs(pop, offset);
1089 sched_add_before(schedpoint, pop);
1094 static ir_node* create_spproj(ia32_transform_env_t *env, ir_node *pred, int pos, ir_node *schedpoint, const ir_node *oldsp) {
1095 ir_mode *spmode = get_irn_mode(oldsp);
1096 const arch_register_t *spreg = arch_get_irn_register(env->cg->arch_env, oldsp);
1099 sp = new_rd_Proj(env->dbg, env->irg, env->block, pred, spmode, pos);
1100 arch_set_irn_register(env->cg->arch_env, sp, spreg);
1101 sched_add_before(schedpoint, sp);
1106 static void transform_MemPerm(ia32_transform_env_t *env) {
1108 * Transform memperm, currently we do this the ugly way and produce
1109 * push/pop into/from memory cascades. This is possible without using
1112 ir_node *node = env->irn;
1114 ir_node *sp = get_irn_n(node, 0);
1115 const ir_edge_t *edge;
1116 const ir_edge_t *next;
1119 arity = be_get_MemPerm_entity_arity(node);
1120 pops = alloca(arity * sizeof(pops[0]));
1123 for(i = 0; i < arity; ++i) {
1124 entity *ent = be_get_MemPerm_in_entity(node, i);
1125 ir_type *enttype = get_entity_type(ent);
1126 int entbits = get_type_size_bits(enttype);
1127 ir_node *mem = get_irn_n(node, i + 1);
1130 assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
1132 push = create_push(env, node, sp, mem, ent, NULL);
1133 sp = create_spproj(env, push, 0, node, sp);
1135 // add another push after the first one
1136 push = create_push(env, node, sp, mem, ent, "4");
1137 sp = create_spproj(env, push, 0, node, sp);
1140 set_irn_n(node, i, new_Bad());
1144 for(i = arity - 1; i >= 0; --i) {
1145 entity *ent = be_get_MemPerm_out_entity(node, i);
1146 ir_type *enttype = get_entity_type(ent);
1147 int entbits = get_type_size_bits(enttype);
1151 assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
1153 pop = create_pop(env, node, sp, ent, NULL);
1155 // add another pop after the first one
1156 sp = create_spproj(env, pop, 1, node, sp);
1157 pop = create_pop(env, node, sp, ent, "4");
1160 sp = create_spproj(env, pop, 1, node, sp);
1166 // exchange memprojs
1167 foreach_out_edge_safe(node, edge, next) {
1168 ir_node *proj = get_edge_src_irn(edge);
1169 int p = get_Proj_proj(proj);
1173 set_Proj_pred(proj, pops[p]);
1174 set_Proj_proj(proj, 3);
1178 arity = get_irn_arity(node);
1179 for(i = 0; i < arity; ++i) {
1180 set_irn_n(node, i, new_Bad());
1186 * Fix the mode of Spill/Reload
1188 static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode)
1190 if (mode_is_float(mode)) {
1202 * Block-Walker: Calls the transform functions Spill and Reload.
1204 static void ia32_after_ra_walker(ir_node *block, void *env) {
1205 ir_node *node, *prev;
1206 ia32_code_gen_t *cg = env;
1207 ia32_transform_env_t tenv;
1210 tenv.irg = current_ir_graph;
1212 DEBUG_ONLY(tenv.mod = cg->mod;)
1214 /* beware: the schedule is changed here */
1215 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
1216 prev = sched_prev(node);
1217 if (be_is_Reload(node)) {
1218 /* we always reload the whole register */
1219 tenv.dbg = get_irn_dbg_info(node);
1221 tenv.mode = fix_spill_mode(cg, get_irn_mode(node));
1222 transform_to_Load(&tenv);
1224 else if (be_is_Spill(node)) {
1225 ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
1226 /* we always spill the whole register */
1227 tenv.dbg = get_irn_dbg_info(node);
1229 tenv.mode = fix_spill_mode(cg, get_irn_mode(spillval));
1230 transform_to_Store(&tenv);
1232 else if(be_is_MemPerm(node)) {
1233 tenv.dbg = get_irn_dbg_info(node);
1235 transform_MemPerm(&tenv);
1241 * We transform Spill and Reload here. This needs to be done before
1242 * stack biasing otherwise we would miss the corrected offset for these nodes.
1244 * If x87 instruction should be emitted, run the x87 simulator and patch
1245 * the virtual instructions. This must obviously be done after register allocation.
1247 static void ia32_after_ra(void *self) {
1248 ia32_code_gen_t *cg = self;
1249 ir_graph *irg = cg->irg;
1251 irg_block_walk_graph(irg, NULL, ia32_after_ra_walker, cg);
1253 ia32_finish_irg(irg, cg);
1257 * Last touchups for the graph before emit
1259 static void ia32_finish(void *self) {
1260 ia32_code_gen_t *cg = self;
1261 ir_graph *irg = cg->irg;
1263 // Matze: disabled for now, as the irextbb algo sometimes returns extbb in
1264 // the wrong order if the graph has critical edges
1265 be_remove_empty_blocks(irg);
1267 cg->blk_sched = sched_create_block_schedule(cg->irg, cg->birg->execfreqs);
1269 /* if we do x87 code generation, rewrite all the virtual instructions and registers */
1270 if (cg->used_fp == fp_x87 || cg->force_sim) {
1271 x87_simulate_graph(cg->arch_env, irg, cg->blk_sched);
1274 ia32_peephole_optimization(irg, cg);
1278 * Emits the code, closes the output file and frees
1279 * the code generator interface.
1281 static void ia32_codegen(void *self) {
1282 ia32_code_gen_t *cg = self;
1283 ir_graph *irg = cg->irg;
1285 ia32_gen_routine(cg->isa->out, irg, cg);
1289 /* remove it from the isa */
1292 /* de-allocate code generator */
1293 del_set(cg->reg_set);
1297 static void *ia32_cg_init(const be_irg_t *birg);
1299 static const arch_code_generator_if_t ia32_code_gen_if = {
1301 NULL, /* before abi introduce hook */
1303 ia32_before_sched, /* before scheduling hook */
1304 ia32_before_ra, /* before register allocation hook */
1305 ia32_after_ra, /* after register allocation hook */
1306 ia32_finish, /* called before codegen */
1307 ia32_codegen /* emit && done */
1311 * Initializes a IA32 code generator.
1313 static void *ia32_cg_init(const be_irg_t *birg) {
1314 ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa;
1315 ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg));
1317 cg->impl = &ia32_code_gen_if;
1318 cg->irg = birg->irg;
1319 cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
1320 cg->arch_env = birg->main_env->arch_env;
1323 cg->blk_sched = NULL;
1324 cg->fp_to_gp = NULL;
1325 cg->gp_to_fp = NULL;
1326 cg->fp_kind = isa->fp_kind;
1327 cg->used_fp = fp_none;
1328 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
1330 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
1332 /* copy optimizations from isa for easier access */
1334 cg->arch = isa->arch;
1335 cg->opt_arch = isa->opt_arch;
1341 if (isa->name_obst_size) {
1342 //printf("freed %d bytes from name obst\n", isa->name_obst_size);
1343 isa->name_obst_size = 0;
1344 obstack_free(isa->name_obst, NULL);
1345 obstack_init(isa->name_obst);
1349 cur_reg_set = cg->reg_set;
1351 ia32_irn_ops.cg = cg;
1353 return (arch_code_generator_t *)cg;
1358 /*****************************************************************
1359 * ____ _ _ _____ _____
1360 * | _ \ | | | | |_ _|/ ____| /\
1361 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
1362 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
1363 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
1364 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
1366 *****************************************************************/
1369 * Set output modes for GCC
1371 static const tarval_mode_info mo_integer = {
1378 * set the tarval output mode of all integer modes to decimal
1380 static void set_tarval_output_modes(void)
1384 for (i = get_irp_n_modes() - 1; i >= 0; --i) {
1385 ir_mode *mode = get_irp_mode(i);
1387 if (mode_is_int(mode))
1388 set_tarval_mode_output_option(mode, &mo_integer);
1394 * The template that generates a new ISA object.
1395 * Note that this template can be changed by command line
1398 static ia32_isa_t ia32_isa_template = {
1400 &ia32_isa_if, /* isa interface implementation */
1401 &ia32_gp_regs[REG_ESP], /* stack pointer register */
1402 &ia32_gp_regs[REG_EBP], /* base pointer register */
1403 -1, /* stack direction */
1404 NULL, /* main environment */
1406 NULL, /* 16bit register names */
1407 NULL, /* 8bit register names */
1411 IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
1412 IA32_OPT_DOAM | /* optimize address mode default: on */
1413 IA32_OPT_LEA | /* optimize for LEAs default: on */
1414 IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */
1415 IA32_OPT_IMMOPS | /* operations can use immediates, default: on */
1416 IA32_OPT_EXTBB), /* use extended basic block scheduling, default: on */
1417 arch_pentium_4, /* instruction architecture */
1418 arch_pentium_4, /* optimize for architecture */
1419 fp_sse2, /* use sse2 unit */
1420 NULL, /* current code generator */
1421 NULL, /* output file */
1423 NULL, /* name obstack */
1424 0 /* name obst size */
1429 * Initializes the backend ISA.
1431 static void *ia32_init(FILE *file_handle) {
1432 static int inited = 0;
1438 set_tarval_output_modes();
1440 isa = xmalloc(sizeof(*isa));
1441 memcpy(isa, &ia32_isa_template, sizeof(*isa));
1443 ia32_register_init(isa);
1444 ia32_create_opcodes();
1446 if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
1447 (ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
1448 /* no SSE2 for these cpu's */
1449 isa->fp_kind = fp_x87;
1451 if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) {
1452 /* Pentium 4 don't like inc and dec instructions */
1453 isa->opt &= ~IA32_OPT_INCDEC;
1456 isa->regs_16bit = pmap_create();
1457 isa->regs_8bit = pmap_create();
1458 isa->types = pmap_create();
1459 isa->tv_ent = pmap_create();
1460 isa->out = file_handle;
1462 ia32_build_16bit_reg_map(isa->regs_16bit);
1463 ia32_build_8bit_reg_map(isa->regs_8bit);
1465 /* patch register names of x87 registers */
1467 ia32_st_regs[0].name = "st";
1468 ia32_st_regs[1].name = "st(1)";
1469 ia32_st_regs[2].name = "st(2)";
1470 ia32_st_regs[3].name = "st(3)";
1471 ia32_st_regs[4].name = "st(4)";
1472 ia32_st_regs[5].name = "st(5)";
1473 ia32_st_regs[6].name = "st(6)";
1474 ia32_st_regs[7].name = "st(7)";
1478 isa->name_obst = xmalloc(sizeof(*isa->name_obst));
1479 obstack_init(isa->name_obst);
1480 isa->name_obst_size = 0;
1483 ia32_handle_intrinsics();
1484 ia32_switch_section(isa->out, NO_SECTION);
1485 fprintf(isa->out, "\t.intel_syntax\n");
1487 /* needed for the debug support */
1488 ia32_switch_section(isa->out, SECTION_TEXT);
1489 fprintf(isa->out, ".Ltext0:\n");
1499 * Closes the output file and frees the ISA structure.
1501 static void ia32_done(void *self) {
1502 ia32_isa_t *isa = self;
1504 /* emit now all global declarations */
1505 ia32_gen_decls(isa->out, isa->arch_isa.main_env);
1507 pmap_destroy(isa->regs_16bit);
1508 pmap_destroy(isa->regs_8bit);
1509 pmap_destroy(isa->tv_ent);
1510 pmap_destroy(isa->types);
1513 //printf("name obst size = %d bytes\n", isa->name_obst_size);
1514 obstack_free(isa->name_obst, NULL);
1522 * Return the number of register classes for this architecture.
1523 * We report always these:
1524 * - the general purpose registers
1525 * - the SSE floating point register set
1526 * - the virtual floating point registers
1528 static int ia32_get_n_reg_class(const void *self) {
1533 * Return the register class for index i.
1535 static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
1536 assert(i >= 0 && i < 3 && "Invalid ia32 register class requested.");
1538 return &ia32_reg_classes[CLASS_ia32_gp];
1540 return &ia32_reg_classes[CLASS_ia32_xmm];
1542 return &ia32_reg_classes[CLASS_ia32_vfp];
1546 * Get the register class which shall be used to store a value of a given mode.
1547 * @param self The this pointer.
1548 * @param mode The mode in question.
1549 * @return A register class which can hold values of the given mode.
1551 const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
1552 const ia32_isa_t *isa = self;
1553 if (mode_is_float(mode)) {
1554 return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1557 return &ia32_reg_classes[CLASS_ia32_gp];
1561 * Get the ABI restrictions for procedure calls.
1562 * @param self The this pointer.
1563 * @param method_type The type of the method (procedure) in question.
1564 * @param abi The abi object to be modified
1566 static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1567 const ia32_isa_t *isa = self;
1570 unsigned cc = get_method_calling_convention(method_type);
1571 int n = get_method_n_params(method_type);
1574 int i, ignore_1, ignore_2;
1576 const arch_register_t *reg;
1577 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1579 unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
1581 /* set abi flags for calls */
1582 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1583 call_flags.bits.store_args_sequential = use_push;
1584 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1585 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1586 call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
1588 /* set stack parameter passing style */
1589 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1591 /* collect the mode for each type */
1592 modes = alloca(n * sizeof(modes[0]));
1594 for (i = 0; i < n; i++) {
1595 tp = get_method_param_type(method_type, i);
1596 modes[i] = get_type_mode(tp);
1599 /* set register parameters */
1600 if (cc & cc_reg_param) {
1601 /* determine the number of parameters passed via registers */
1602 biggest_n = ia32_get_n_regparam_class(n, modes, &ignore_1, &ignore_2);
1604 /* loop over all parameters and set the register requirements */
1605 for (i = 0; i <= biggest_n; i++) {
1606 reg = ia32_get_RegParam_reg(n, modes, i, cc);
1607 assert(reg && "kaputt");
1608 be_abi_call_param_reg(abi, i, reg);
1615 /* set stack parameters */
1616 for (i = stack_idx; i < n; i++) {
1617 be_abi_call_param_stack(abi, i, 1, 0, 0);
1621 /* set return registers */
1622 n = get_method_n_ress(method_type);
1624 assert(n <= 2 && "more than two results not supported");
1626 /* In case of 64bit returns, we will have two 32bit values */
1628 tp = get_method_res_type(method_type, 0);
1629 mode = get_type_mode(tp);
1631 assert(!mode_is_float(mode) && "two FP results not supported");
1633 tp = get_method_res_type(method_type, 1);
1634 mode = get_type_mode(tp);
1636 assert(!mode_is_float(mode) && "two FP results not supported");
1638 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX]);
1639 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]);
1642 const arch_register_t *reg;
1644 tp = get_method_res_type(method_type, 0);
1645 assert(is_atomic_type(tp));
1646 mode = get_type_mode(tp);
1648 reg = mode_is_float(mode) ?
1649 (USE_SSE2(isa) ? &ia32_xmm_regs[REG_XMM0] : &ia32_vfp_regs[REG_VF0]) :
1650 &ia32_gp_regs[REG_EAX];
1652 be_abi_call_res_reg(abi, 0, reg);
1657 static const void *ia32_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1658 return &ia32_irn_ops;
1661 const arch_irn_handler_t ia32_irn_handler = {
1665 const arch_irn_handler_t *ia32_get_irn_handler(const void *self) {
1666 return &ia32_irn_handler;
1669 int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1670 return is_ia32_irn(irn) ? 1 : -1;
1674 * Initializes the code generator interface.
1676 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self) {
1677 return &ia32_code_gen_if;
1681 * Returns the estimated execution time of an ia32 irn.
1683 static sched_timestep_t ia32_sched_exectime(void *env, const ir_node *irn) {
1684 const arch_env_t *arch_env = env;
1685 return is_ia32_irn(irn) ? ia32_get_op_estimated_cost(arch_get_irn_ops(arch_env, irn), irn) : 1;
1688 list_sched_selector_t ia32_sched_selector;
1691 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1693 static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self, list_sched_selector_t *selector) {
1694 memcpy(&ia32_sched_selector, selector, sizeof(ia32_sched_selector));
1695 ia32_sched_selector.exectime = ia32_sched_exectime;
1696 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
1697 return &ia32_sched_selector;
1701 * Returns the necessary byte alignment for storing a register of given class.
1703 static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1704 ir_mode *mode = arch_register_class_mode(cls);
1705 int bytes = get_mode_size_bytes(mode);
1707 if (mode_is_float(mode) && bytes > 8)
1712 static ia32_intrinsic_env_t intrinsic_env = { NULL, NULL };
1715 * Returns the libFirm configuration parameter for this backend.
1717 static const backend_params *ia32_get_libfirm_params(void) {
1718 static const arch_dep_params_t ad = {
1719 1, /* also use subs */
1720 4, /* maximum shifts */
1721 31, /* maximum shift amount */
1723 1, /* allow Mulhs */
1724 1, /* allow Mulus */
1725 32 /* Mulh allowed up to 32 bit */
1727 static backend_params p = {
1728 NULL, /* no additional opcodes */
1729 NULL, /* will be set later */
1730 1, /* need dword lowering */
1731 ia32_create_intrinsic_fkt,
1732 &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
1740 /* instruction set architectures. */
1741 static const lc_opt_enum_int_items_t arch_items[] = {
1742 { "386", arch_i386, },
1743 { "486", arch_i486, },
1744 { "pentium", arch_pentium, },
1745 { "586", arch_pentium, },
1746 { "pentiumpro", arch_pentium_pro, },
1747 { "686", arch_pentium_pro, },
1748 { "pentiummmx", arch_pentium_mmx, },
1749 { "pentium2", arch_pentium_2, },
1750 { "p2", arch_pentium_2, },
1751 { "pentium3", arch_pentium_3, },
1752 { "p3", arch_pentium_3, },
1753 { "pentium4", arch_pentium_4, },
1754 { "p4", arch_pentium_4, },
1755 { "pentiumm", arch_pentium_m, },
1756 { "pm", arch_pentium_m, },
1757 { "core", arch_core, },
1759 { "athlon", arch_athlon, },
1760 { "athlon64", arch_athlon_64, },
1761 { "opteron", arch_opteron, },
1765 static lc_opt_enum_int_var_t arch_var = {
1766 &ia32_isa_template.arch, arch_items
1769 static lc_opt_enum_int_var_t opt_arch_var = {
1770 &ia32_isa_template.opt_arch, arch_items
1773 static const lc_opt_enum_int_items_t fp_unit_items[] = {
1775 { "sse2", fp_sse2 },
1779 static lc_opt_enum_int_var_t fp_unit_var = {
1780 &ia32_isa_template.fp_kind, fp_unit_items
1783 static const lc_opt_enum_int_items_t gas_items[] = {
1784 { "linux", ASM_LINUX_GAS },
1785 { "mingw", ASM_MINGW_GAS },
1789 static lc_opt_enum_int_var_t gas_var = {
1790 (int *)&asm_flavour, gas_items
1793 static const lc_opt_table_entry_t ia32_options[] = {
1794 LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var),
1795 LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var),
1796 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var),
1797 LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
1798 LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA),
1799 LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST),
1800 LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS),
1801 LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB),
1802 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
1807 * Register command line options for the ia32 backend.
1811 * ia32-arch=arch create instruction for arch
1812 * ia32-opt=arch optimize for run on arch
1813 * ia32-fpunit=unit select floating point unit (x87 or SSE2)
1814 * ia32-incdec optimize for inc/dec
1815 * ia32-noaddrmode do not use address mode
1816 * ia32-nolea do not optimize for LEAs
1817 * ia32-noplacecnst do not place constants,
1818 * ia32-noimmop no operations with immediates
1819 * ia32-noextbb do not use extended basic block scheduling
1820 * ia32-gasmode set the GAS compatibility mode
1822 static void ia32_register_options(lc_opt_entry_t *ent)
1824 lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32");
1825 lc_opt_add_table(be_grp_ia32, ia32_options);
1827 #endif /* WITH_LIBCORE */
1829 const arch_isa_if_t ia32_isa_if = {
1832 ia32_get_n_reg_class,
1834 ia32_get_reg_class_for_mode,
1836 ia32_get_irn_handler,
1837 ia32_get_code_generator_if,
1838 ia32_get_list_sched_selector,
1839 ia32_get_reg_class_alignment,
1840 ia32_get_libfirm_params,
1842 ia32_register_options