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 /* Push modifies always ESP, this cannot be changed */
268 return arch_irn_flags_modify_sp | arch_irn_flags_ignore;
270 if(is_ia32_Pop(pred) && get_Proj_proj(irn) == pn_ia32_Pop_stack) {
271 return arch_irn_flags_modify_sp | arch_irn_flags_ignore;
273 if(is_ia32_AddSP(pred) && get_Proj_proj(irn) == pn_ia32_AddSP_stack) {
274 /* AddSP modifies always ESP, this cannot be changed */
275 return arch_irn_flags_modify_sp | arch_irn_flags_ignore;
277 if(is_ia32_SubSP(pred) && get_Proj_proj(irn) == pn_ia32_SubSP_stack) {
278 /* SubSP modifies always ESP, this cannot be changed */
279 return arch_irn_flags_modify_sp | arch_irn_flags_ignore;
283 irn = my_skip_proj(irn);
284 if (is_ia32_irn(irn))
285 return get_ia32_flags(irn);
288 return arch_irn_flags_ignore;
293 static entity *ia32_get_frame_entity(const void *self, const ir_node *irn) {
294 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
297 static void ia32_set_frame_entity(const void *self, ir_node *irn, entity *ent) {
298 set_ia32_frame_ent(irn, ent);
301 static void ia32_set_frame_offset(const void *self, ir_node *irn, int bias) {
303 const ia32_irn_ops_t *ops = self;
305 if (get_ia32_frame_ent(irn)) {
306 ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn);
308 /* Pop nodes modify the stack pointer before calculating the destination
309 * address, so fix this here
311 if(is_ia32_Pop(irn)) {
315 DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias));
317 snprintf(buf, sizeof(buf), "%d", bias);
319 if (get_ia32_op_type(irn) == ia32_Normal) {
320 set_ia32_cnst(irn, buf);
322 add_ia32_am_offs(irn, buf);
324 set_ia32_am_flavour(irn, am_flav);
329 static int ia32_get_sp_bias(const void *self, const ir_node *irn) {
331 int proj = get_Proj_proj(irn);
332 ir_node *pred = get_Proj_pred(irn);
334 if(is_ia32_Push(pred) && proj == 0)
336 if(is_ia32_Pop(pred) && proj == 1)
344 be_abi_call_flags_bits_t flags;
345 const arch_isa_t *isa;
346 const arch_env_t *aenv;
350 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
352 ia32_abi_env_t *env = xmalloc(sizeof(env[0]));
353 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
354 env->flags = fl.bits;
357 env->isa = aenv->isa;
362 * Put all registers which are saved by the prologue/epilogue in a set.
364 * @param self The callback object.
365 * @param s The result set.
367 static void ia32_abi_dont_save_regs(void *self, pset *s)
369 ia32_abi_env_t *env = self;
370 if(env->flags.try_omit_fp)
371 pset_insert_ptr(s, env->isa->bp);
375 * Generate the routine prologue.
377 * @param self The callback object.
378 * @param mem A pointer to the mem node. Update this if you define new memory.
379 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
381 * @return The register which shall be used as a stack frame base.
383 * All nodes which define registers in @p reg_map must keep @p reg_map current.
385 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
387 ia32_abi_env_t *env = self;
389 if (! env->flags.try_omit_fp) {
390 ir_node *bl = get_irg_start_block(env->irg);
391 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
392 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
396 push = new_rd_ia32_Push(NULL, env->irg, bl, curr_sp, curr_bp, *mem);
397 curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
398 *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M);
400 /* the push must have SP out register */
401 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
402 set_ia32_flags(push, arch_irn_flags_ignore);
404 /* move esp to ebp */
405 curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp);
406 be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp);
407 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
408 be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
410 /* beware: the copy must be done before any other sp use */
411 curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
412 be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp);
413 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
414 be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
416 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
417 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
426 * Generate the routine epilogue.
427 * @param self The callback object.
428 * @param bl The block for the epilog
429 * @param mem A pointer to the mem node. Update this if you define new memory.
430 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
431 * @return The register which shall be used as a stack frame base.
433 * All nodes which define registers in @p reg_map must keep @p reg_map current.
435 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
437 ia32_abi_env_t *env = self;
438 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
439 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
441 if (env->flags.try_omit_fp) {
442 /* simply remove the stack frame here */
443 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK);
444 add_irn_dep(curr_sp, *mem);
447 const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
448 ir_mode *mode_bp = env->isa->bp->reg_class->mode;
450 /* gcc always emits a leave at the end of a routine */
451 if (1 || ARCH_AMD(isa->opt_arch)) {
455 leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, curr_bp);
456 set_ia32_flags(leave, arch_irn_flags_ignore);
457 curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame);
458 curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
459 *mem = new_r_Proj(current_ir_graph, bl, leave, mode_M, pn_ia32_Leave_M);
464 /* copy ebp to esp */
465 curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
468 pop = new_rd_ia32_Pop(NULL, env->irg, bl, curr_sp, *mem);
469 set_ia32_flags(pop, arch_irn_flags_ignore);
470 curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res);
471 curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
472 *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
474 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
475 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
478 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
479 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
483 * Produces the type which sits between the stack args and the locals on the stack.
484 * it will contain the return address and space to store the old base pointer.
485 * @return The Firm type modeling the ABI between type.
487 static ir_type *ia32_abi_get_between_type(void *self)
489 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
490 static ir_type *omit_fp_between_type = NULL;
491 static ir_type *between_type = NULL;
493 ia32_abi_env_t *env = self;
495 if ( !between_type) {
497 entity *ret_addr_ent;
498 entity *omit_fp_ret_addr_ent;
500 ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P);
501 ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_P);
503 between_type = new_type_struct(IDENT("ia32_between_type"));
504 old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
505 ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
507 set_entity_offset_bytes(old_bp_ent, 0);
508 set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type));
509 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
510 set_type_state(between_type, layout_fixed);
512 omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
513 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
515 set_entity_offset_bytes(omit_fp_ret_addr_ent, 0);
516 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
517 set_type_state(omit_fp_between_type, layout_fixed);
520 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
525 * Get the estimated cycle count for @p irn.
527 * @param self The this pointer.
528 * @param irn The node.
530 * @return The estimated cycle count for this operation
532 static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
535 ia32_op_type_t op_tp;
536 const ia32_irn_ops_t *ops = self;
541 assert(is_ia32_irn(irn));
543 cost = get_ia32_latency(irn);
544 op_tp = get_ia32_op_type(irn);
546 if (is_ia32_CopyB(irn)) {
548 if (ARCH_INTEL(ops->cg->arch))
551 else if (is_ia32_CopyB_i(irn)) {
552 int size = get_tarval_long(get_ia32_Immop_tarval(irn));
553 cost = 20 + (int)ceil((4/3) * size);
554 if (ARCH_INTEL(ops->cg->arch))
557 /* in case of address mode operations add additional cycles */
558 else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) {
560 In case of stack access add 5 cycles (we assume stack is in cache),
561 other memory operations cost 20 cycles.
563 cost += is_ia32_use_frame(irn) ? 5 : 20;
570 * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
572 * @param irn The original operation
573 * @param i Index of the argument we want the inverse operation to yield
574 * @param inverse struct to be filled with the resulting inverse op
575 * @param obstack The obstack to use for allocation of the returned nodes array
576 * @return The inverse operation or NULL if operation invertible
578 static arch_inverse_t *ia32_get_inverse(const void *self, const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) {
581 ir_node *block, *noreg, *nomem;
584 /* we cannot invert non-ia32 irns */
585 if (! is_ia32_irn(irn))
588 /* operand must always be a real operand (not base, index or mem) */
589 if (i != 2 && i != 3)
592 /* we don't invert address mode operations */
593 if (get_ia32_op_type(irn) != ia32_Normal)
596 irg = get_irn_irg(irn);
597 block = get_nodes_block(irn);
598 mode = get_ia32_res_mode(irn);
599 noreg = get_irn_n(irn, 0);
600 nomem = new_r_NoMem(irg);
602 /* initialize structure */
603 inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
607 switch (get_ia32_irn_opcode(irn)) {
609 if (get_ia32_immop_type(irn) == ia32_ImmConst) {
610 /* we have an add with a const here */
611 /* invers == add with negated const */
612 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
613 pnc = pn_ia32_Add_res;
615 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
616 set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
617 set_ia32_commutative(inverse->nodes[0]);
619 else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
620 /* we have an add with a symconst here */
621 /* invers == sub with const */
622 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
623 pnc = pn_ia32_Sub_res;
625 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
628 /* normal add: inverse == sub */
629 ir_node *proj = ia32_get_res_proj(irn);
632 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem);
633 pnc = pn_ia32_Sub_res;
638 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
639 /* we have a sub with a const/symconst here */
640 /* invers == add with this const */
641 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
642 pnc = pn_ia32_Add_res;
643 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
644 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
648 ir_node *proj = ia32_get_res_proj(irn);
652 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem);
655 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, 2), proj, nomem);
657 pnc = pn_ia32_Sub_res;
662 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
663 /* xor with const: inverse = xor */
664 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
665 pnc = pn_ia32_Eor_res;
666 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
667 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
671 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, (ir_node *)irn, get_irn_n(irn, i), nomem);
672 pnc = pn_ia32_Eor_res;
677 ir_node *proj = ia32_get_res_proj(irn);
680 inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, proj, nomem);
681 pnc = pn_ia32_Not_res;
685 case iro_ia32_Minus: {
686 ir_node *proj = ia32_get_res_proj(irn);
689 inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, proj, nomem);
690 pnc = pn_ia32_Minus_res;
695 /* inverse operation not supported */
699 set_ia32_res_mode(inverse->nodes[0], mode);
700 inverse->nodes[1] = new_r_Proj(irg, block, inverse->nodes[0], mode, pnc);
706 * Check if irn can load it's operand at position i from memory (source addressmode).
707 * @param self Pointer to irn ops itself
708 * @param irn The irn to be checked
709 * @param i The operands position
710 * @return Non-Zero if operand can be loaded
712 static int ia32_possible_memory_operand(const void *self, const ir_node *irn, unsigned int i) {
713 if (! is_ia32_irn(irn) || /* must be an ia32 irn */
714 get_irn_arity(irn) != 5 || /* must be a binary operation */
715 get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
716 ! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */
717 (i != 2 && i != 3) || /* a "real" operand position must be requested */
718 (i == 2 && ! is_ia32_commutative(irn)) || /* if first operand requested irn must be commutative */
719 is_ia32_use_frame(irn)) /* must not already use frame */
725 static void ia32_perform_memory_operand(const void *self, ir_node *irn, ir_node *spill, unsigned int i) {
726 assert(ia32_possible_memory_operand(self, irn, i) && "Cannot perform memory operand change");
729 ir_node *tmp = get_irn_n(irn, 3);
730 set_irn_n(irn, 3, get_irn_n(irn, 2));
731 set_irn_n(irn, 2, tmp);
734 set_ia32_am_support(irn, ia32_am_Source);
735 set_ia32_op_type(irn, ia32_AddrModeS);
736 set_ia32_am_flavour(irn, ia32_B);
737 set_ia32_ls_mode(irn, get_irn_mode(get_irn_n(irn, i)));
738 //TODO this will fail, if spill is a PhiM (give PhiMs entities?)
739 set_ia32_frame_ent(irn, be_get_frame_entity(spill));
740 set_ia32_use_frame(irn);
741 set_ia32_got_reload(irn);
743 set_irn_n(irn, 0, get_irg_frame(get_irn_irg(irn)));
744 set_irn_n(irn, 4, spill);
747 Input at position one is index register, which is NoReg.
748 We would need cg object to get a real noreg, but we cannot
751 set_irn_n(irn, 3, get_irn_n(irn, 1));
753 //FIXME DBG_OPT_AM_S(reload, irn);
756 static const be_abi_callbacks_t ia32_abi_callbacks = {
759 ia32_abi_get_between_type,
760 ia32_abi_dont_save_regs,
765 /* fill register allocator interface */
767 static const arch_irn_ops_if_t ia32_irn_ops_if = {
768 ia32_get_irn_reg_req,
773 ia32_get_frame_entity,
774 ia32_set_frame_entity,
775 ia32_set_frame_offset,
778 ia32_get_op_estimated_cost,
779 ia32_possible_memory_operand,
780 ia32_perform_memory_operand,
783 ia32_irn_ops_t ia32_irn_ops = {
790 /**************************************************
793 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
794 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
795 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
796 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
799 **************************************************/
801 static void ia32_kill_convs(ia32_code_gen_t *cg) {
804 /* BEWARE: the Projs are inserted in the set */
805 foreach_nodeset(cg->kill_conv, irn) {
806 ir_node *in = get_irn_n(get_Proj_pred(irn), 2);
807 edges_reroute(irn, in, cg->birg->irg);
812 * Transform the Thread Local Store base.
814 static void transform_tls(ir_graph *irg) {
815 ir_node *irn = get_irg_tls(irg);
818 dbg_info *dbg = get_irn_dbg_info(irn);
819 ir_node *blk = get_nodes_block(irn);
821 newn = new_rd_ia32_LdTls(dbg, irg, blk, get_irn_mode(irn));
828 * Transforms the standard firm graph into
831 static void ia32_prepare_graph(void *self) {
832 ia32_code_gen_t *cg = self;
833 dom_front_info_t *dom;
834 DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
836 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
838 /* 1st: transform constants and psi condition trees */
839 ia32_pre_transform_phase(cg);
841 /* 2nd: transform all remaining nodes */
842 ia32_register_transformers();
843 dom = be_compute_dominance_frontiers(cg->irg);
845 cg->kill_conv = new_nodeset(5);
846 transform_tls(cg->irg);
847 irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg);
849 del_nodeset(cg->kill_conv);
851 be_free_dominance_frontiers(dom);
854 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
856 /* 3rd: optimize address mode */
857 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
858 ia32_optimize_addressmode(cg);
861 be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
863 DEBUG_ONLY(cg->mod = old_mod;)
867 * Dummy functions for hooks we don't need but which must be filled.
869 static void ia32_before_sched(void *self) {
872 static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) {
875 ir_node *mem_proj = NULL;
880 mode = get_irn_mode(irn);
882 /* check if we already saw this node or the node has more than one user */
883 if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1)
886 /* mark irn visited */
887 bitset_add_irn(already_visited, irn);
889 /* non-Tuple nodes with one user: ok, return */
890 if (get_irn_n_edges(irn) >= 1 && mode != mode_T)
893 /* tuple node has one user which is not the mem proj-> ok */
894 if (mode == mode_T && get_irn_n_edges(irn) == 1) {
895 mem_proj = ia32_get_proj_for_mode(irn, mode_M);
896 if (mem_proj == NULL)
900 arity = get_irn_arity(irn);
901 for (i = 0; i < arity; ++i) {
902 ir_node *pred = get_irn_n(irn, i);
904 /* do not follow memory edges or we will accidentally remove stores */
905 if (get_irn_mode(pred) == mode_M) {
906 if(mem_proj != NULL) {
907 edges_reroute(mem_proj, pred, get_irn_irg(mem_proj));
913 set_irn_n(irn, i, new_Bad());
916 The current node is about to be removed: if the predecessor
917 has only this node as user, it need to be removed as well.
919 if (get_irn_n_edges(pred) <= 1)
920 remove_unused_nodes(pred, already_visited);
923 // we need to set the presd to Bad again to also get the memory edges
924 arity = get_irn_arity(irn);
925 for (i = 0; i < arity; ++i) {
926 set_irn_n(irn, i, new_Bad());
929 if (sched_is_scheduled(irn)) {
934 static void remove_unused_loads_walker(ir_node *irn, void *env) {
935 bitset_t *already_visited = env;
936 if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn))
937 remove_unused_nodes(irn, env);
941 * Called before the register allocator.
942 * Calculate a block schedule here. We need it for the x87
943 * simulator and the emitter.
945 static void ia32_before_ra(void *self) {
946 ia32_code_gen_t *cg = self;
947 bitset_t *already_visited = bitset_irg_alloca(cg->irg);
951 There are sometimes unused loads, only pinned by memory.
952 We need to remove those Loads and all other nodes which won't be used
953 after removing the Load from schedule.
955 irg_walk_graph(cg->irg, remove_unused_loads_walker, NULL, already_visited);
960 * Transforms a be node into a Load.
962 static void transform_to_Load(ia32_transform_env_t *env) {
963 ir_node *irn = env->irn;
964 entity *ent = be_get_frame_entity(irn);
965 ir_mode *mode = env->mode;
966 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
967 ir_node *nomem = new_rd_NoMem(env->irg);
968 ir_node *sched_point = NULL;
969 ir_node *ptr = get_irn_n(irn, 0);
970 ir_node *mem = be_is_Reload(irn) ? get_irn_n(irn, 1) : nomem;
971 ir_node *new_op, *proj;
972 const arch_register_t *reg;
974 if (sched_is_scheduled(irn)) {
975 sched_point = sched_prev(irn);
978 if (mode_is_float(mode)) {
979 if (USE_SSE2(env->cg))
980 new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
982 new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
985 new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
988 set_ia32_am_support(new_op, ia32_am_Source);
989 set_ia32_op_type(new_op, ia32_AddrModeS);
990 set_ia32_am_flavour(new_op, ia32_B);
991 set_ia32_ls_mode(new_op, mode);
992 set_ia32_frame_ent(new_op, ent);
993 set_ia32_use_frame(new_op);
995 DBG_OPT_RELOAD2LD(irn, new_op);
997 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_Load_res);
1000 sched_add_after(sched_point, new_op);
1001 sched_add_after(new_op, proj);
1006 /* copy the register from the old node to the new Load */
1007 reg = arch_get_irn_register(env->cg->arch_env, irn);
1008 arch_set_irn_register(env->cg->arch_env, new_op, reg);
1010 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
1012 exchange(irn, proj);
1016 * Transforms a be node into a Store.
1018 static void transform_to_Store(ia32_transform_env_t *env) {
1019 ir_node *irn = env->irn;
1020 entity *ent = be_get_frame_entity(irn);
1021 ir_mode *mode = env->mode;
1022 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1023 ir_node *nomem = new_rd_NoMem(env->irg);
1024 ir_node *ptr = get_irn_n(irn, 0);
1025 ir_node *val = get_irn_n(irn, 1);
1026 ir_node *new_op, *proj;
1027 ir_node *sched_point = NULL;
1029 if (sched_is_scheduled(irn)) {
1030 sched_point = sched_prev(irn);
1033 if (mode_is_float(mode)) {
1034 if (USE_SSE2(env->cg))
1035 new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1037 new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1039 else if (get_mode_size_bits(mode) == 8) {
1040 new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1043 new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1046 set_ia32_am_support(new_op, ia32_am_Dest);
1047 set_ia32_op_type(new_op, ia32_AddrModeD);
1048 set_ia32_am_flavour(new_op, ia32_B);
1049 set_ia32_ls_mode(new_op, mode);
1050 set_ia32_frame_ent(new_op, ent);
1051 set_ia32_use_frame(new_op);
1053 DBG_OPT_SPILL2ST(irn, new_op);
1055 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, pn_ia32_Store_M);
1058 sched_add_after(sched_point, new_op);
1062 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
1064 exchange(irn, proj);
1067 static ir_node *create_push(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, ir_node *mem, entity *ent, const char *offset) {
1068 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1070 ir_node *push = new_rd_ia32_Push(env->dbg, env->irg, env->block, sp, noreg, mem);
1072 set_ia32_frame_ent(push, ent);
1073 set_ia32_use_frame(push);
1074 set_ia32_op_type(push, ia32_AddrModeS);
1075 set_ia32_am_flavour(push, ia32_B);
1076 set_ia32_ls_mode(push, mode_Is);
1078 add_ia32_am_offs(push, offset);
1080 sched_add_before(schedpoint, push);
1084 static ir_node *create_pop(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, entity *ent, const char *offset) {
1085 ir_node *pop = new_rd_ia32_Pop(env->dbg, env->irg, env->block, sp, new_NoMem());
1087 set_ia32_frame_ent(pop, ent);
1088 set_ia32_use_frame(pop);
1089 set_ia32_op_type(pop, ia32_AddrModeD);
1090 set_ia32_am_flavour(pop, ia32_B);
1091 set_ia32_ls_mode(pop, mode_Is);
1093 add_ia32_am_offs(pop, offset);
1095 sched_add_before(schedpoint, pop);
1100 static ir_node* create_spproj(ia32_transform_env_t *env, ir_node *pred, int pos, ir_node *schedpoint, const ir_node *oldsp) {
1101 ir_mode *spmode = get_irn_mode(oldsp);
1102 const arch_register_t *spreg = arch_get_irn_register(env->cg->arch_env, oldsp);
1105 sp = new_rd_Proj(env->dbg, env->irg, env->block, pred, spmode, pos);
1106 arch_set_irn_register(env->cg->arch_env, sp, spreg);
1107 sched_add_before(schedpoint, sp);
1112 static void transform_MemPerm(ia32_transform_env_t *env) {
1114 * Transform memperm, currently we do this the ugly way and produce
1115 * push/pop into/from memory cascades. This is possible without using
1118 ir_node *node = env->irn;
1120 ir_node *sp = get_irn_n(node, 0);
1121 const ir_edge_t *edge;
1122 const ir_edge_t *next;
1125 arity = be_get_MemPerm_entity_arity(node);
1126 pops = alloca(arity * sizeof(pops[0]));
1129 for(i = 0; i < arity; ++i) {
1130 entity *ent = be_get_MemPerm_in_entity(node, i);
1131 ir_type *enttype = get_entity_type(ent);
1132 int entbits = get_type_size_bits(enttype);
1133 ir_node *mem = get_irn_n(node, i + 1);
1136 assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
1138 push = create_push(env, node, sp, mem, ent, NULL);
1139 sp = create_spproj(env, push, 0, node, sp);
1141 // add another push after the first one
1142 push = create_push(env, node, sp, mem, ent, "4");
1143 sp = create_spproj(env, push, 0, node, sp);
1146 set_irn_n(node, i, new_Bad());
1150 for(i = arity - 1; i >= 0; --i) {
1151 entity *ent = be_get_MemPerm_out_entity(node, i);
1152 ir_type *enttype = get_entity_type(ent);
1153 int entbits = get_type_size_bits(enttype);
1157 assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
1159 pop = create_pop(env, node, sp, ent, NULL);
1161 // add another pop after the first one
1162 sp = create_spproj(env, pop, 1, node, sp);
1163 pop = create_pop(env, node, sp, ent, "4");
1166 sp = create_spproj(env, pop, 1, node, sp);
1172 // exchange memprojs
1173 foreach_out_edge_safe(node, edge, next) {
1174 ir_node *proj = get_edge_src_irn(edge);
1175 int p = get_Proj_proj(proj);
1179 set_Proj_pred(proj, pops[p]);
1180 set_Proj_proj(proj, 3);
1184 arity = get_irn_arity(node);
1185 for(i = 0; i < arity; ++i) {
1186 set_irn_n(node, i, new_Bad());
1192 * Fix the mode of Spill/Reload
1194 static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode)
1196 if (mode_is_float(mode)) {
1208 * Block-Walker: Calls the transform functions Spill and Reload.
1210 static void ia32_after_ra_walker(ir_node *block, void *env) {
1211 ir_node *node, *prev;
1212 ia32_code_gen_t *cg = env;
1213 ia32_transform_env_t tenv;
1216 tenv.irg = current_ir_graph;
1218 DEBUG_ONLY(tenv.mod = cg->mod;)
1220 /* beware: the schedule is changed here */
1221 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
1222 prev = sched_prev(node);
1223 if (be_is_Reload(node)) {
1224 /* we always reload the whole register */
1225 tenv.dbg = get_irn_dbg_info(node);
1227 tenv.mode = fix_spill_mode(cg, get_irn_mode(node));
1228 transform_to_Load(&tenv);
1230 else if (be_is_Spill(node)) {
1231 ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
1232 /* we always spill the whole register */
1233 tenv.dbg = get_irn_dbg_info(node);
1235 tenv.mode = fix_spill_mode(cg, get_irn_mode(spillval));
1236 transform_to_Store(&tenv);
1238 else if(be_is_MemPerm(node)) {
1239 tenv.dbg = get_irn_dbg_info(node);
1241 transform_MemPerm(&tenv);
1247 * We transform Spill and Reload here. This needs to be done before
1248 * stack biasing otherwise we would miss the corrected offset for these nodes.
1250 * If x87 instruction should be emitted, run the x87 simulator and patch
1251 * the virtual instructions. This must obviously be done after register allocation.
1253 static void ia32_after_ra(void *self) {
1254 ia32_code_gen_t *cg = self;
1255 ir_graph *irg = cg->irg;
1257 irg_block_walk_graph(irg, NULL, ia32_after_ra_walker, cg);
1259 ia32_finish_irg(irg, cg);
1263 * Last touchups for the graph before emit
1265 static void ia32_finish(void *self) {
1266 ia32_code_gen_t *cg = self;
1267 ir_graph *irg = cg->irg;
1269 // Matze: disabled for now, as the irextbb algo sometimes returns extbb in
1270 // the wrong order if the graph has critical edges
1271 be_remove_empty_blocks(irg);
1273 cg->blk_sched = sched_create_block_schedule(cg->irg, cg->birg->execfreqs);
1275 /* if we do x87 code generation, rewrite all the virtual instructions and registers */
1276 if (cg->used_fp == fp_x87 || cg->force_sim) {
1277 x87_simulate_graph(cg->arch_env, irg, cg->blk_sched);
1280 ia32_peephole_optimization(irg, cg);
1284 * Emits the code, closes the output file and frees
1285 * the code generator interface.
1287 static void ia32_codegen(void *self) {
1288 ia32_code_gen_t *cg = self;
1289 ir_graph *irg = cg->irg;
1291 ia32_gen_routine(cg->isa->out, irg, cg);
1295 /* remove it from the isa */
1298 /* de-allocate code generator */
1299 del_set(cg->reg_set);
1303 static void *ia32_cg_init(const be_irg_t *birg);
1305 static const arch_code_generator_if_t ia32_code_gen_if = {
1307 NULL, /* before abi introduce hook */
1309 ia32_before_sched, /* before scheduling hook */
1310 ia32_before_ra, /* before register allocation hook */
1311 ia32_after_ra, /* after register allocation hook */
1312 ia32_finish, /* called before codegen */
1313 ia32_codegen /* emit && done */
1317 * Initializes a IA32 code generator.
1319 static void *ia32_cg_init(const be_irg_t *birg) {
1320 ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa;
1321 ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg));
1323 cg->impl = &ia32_code_gen_if;
1324 cg->irg = birg->irg;
1325 cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
1326 cg->arch_env = birg->main_env->arch_env;
1329 cg->blk_sched = NULL;
1330 cg->fp_to_gp = NULL;
1331 cg->gp_to_fp = NULL;
1332 cg->fp_kind = isa->fp_kind;
1333 cg->used_fp = fp_none;
1334 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
1336 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
1338 /* copy optimizations from isa for easier access */
1340 cg->arch = isa->arch;
1341 cg->opt_arch = isa->opt_arch;
1347 if (isa->name_obst_size) {
1348 //printf("freed %d bytes from name obst\n", isa->name_obst_size);
1349 isa->name_obst_size = 0;
1350 obstack_free(isa->name_obst, NULL);
1351 obstack_init(isa->name_obst);
1355 cur_reg_set = cg->reg_set;
1357 ia32_irn_ops.cg = cg;
1359 return (arch_code_generator_t *)cg;
1364 /*****************************************************************
1365 * ____ _ _ _____ _____
1366 * | _ \ | | | | |_ _|/ ____| /\
1367 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
1368 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
1369 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
1370 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
1372 *****************************************************************/
1375 * Set output modes for GCC
1377 static const tarval_mode_info mo_integer = {
1384 * set the tarval output mode of all integer modes to decimal
1386 static void set_tarval_output_modes(void)
1390 for (i = get_irp_n_modes() - 1; i >= 0; --i) {
1391 ir_mode *mode = get_irp_mode(i);
1393 if (mode_is_int(mode))
1394 set_tarval_mode_output_option(mode, &mo_integer);
1400 * The template that generates a new ISA object.
1401 * Note that this template can be changed by command line
1404 static ia32_isa_t ia32_isa_template = {
1406 &ia32_isa_if, /* isa interface implementation */
1407 &ia32_gp_regs[REG_ESP], /* stack pointer register */
1408 &ia32_gp_regs[REG_EBP], /* base pointer register */
1409 -1, /* stack direction */
1410 NULL, /* main environment */
1412 NULL, /* 16bit register names */
1413 NULL, /* 8bit register names */
1417 IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
1418 IA32_OPT_DOAM | /* optimize address mode default: on */
1419 IA32_OPT_LEA | /* optimize for LEAs default: on */
1420 IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */
1421 IA32_OPT_IMMOPS | /* operations can use immediates, default: on */
1422 IA32_OPT_EXTBB), /* use extended basic block scheduling, default: on */
1423 arch_pentium_4, /* instruction architecture */
1424 arch_pentium_4, /* optimize for architecture */
1425 fp_sse2, /* use sse2 unit */
1426 NULL, /* current code generator */
1427 NULL, /* output file */
1429 NULL, /* name obstack */
1430 0 /* name obst size */
1435 * Initializes the backend ISA.
1437 static void *ia32_init(FILE *file_handle) {
1438 static int inited = 0;
1444 set_tarval_output_modes();
1446 isa = xmalloc(sizeof(*isa));
1447 memcpy(isa, &ia32_isa_template, sizeof(*isa));
1449 ia32_register_init(isa);
1450 ia32_create_opcodes();
1452 if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
1453 (ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
1454 /* no SSE2 for these cpu's */
1455 isa->fp_kind = fp_x87;
1457 if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) {
1458 /* Pentium 4 don't like inc and dec instructions */
1459 isa->opt &= ~IA32_OPT_INCDEC;
1462 isa->regs_16bit = pmap_create();
1463 isa->regs_8bit = pmap_create();
1464 isa->types = pmap_create();
1465 isa->tv_ent = pmap_create();
1466 isa->out = file_handle;
1468 ia32_build_16bit_reg_map(isa->regs_16bit);
1469 ia32_build_8bit_reg_map(isa->regs_8bit);
1471 /* patch register names of x87 registers */
1473 ia32_st_regs[0].name = "st";
1474 ia32_st_regs[1].name = "st(1)";
1475 ia32_st_regs[2].name = "st(2)";
1476 ia32_st_regs[3].name = "st(3)";
1477 ia32_st_regs[4].name = "st(4)";
1478 ia32_st_regs[5].name = "st(5)";
1479 ia32_st_regs[6].name = "st(6)";
1480 ia32_st_regs[7].name = "st(7)";
1484 isa->name_obst = xmalloc(sizeof(*isa->name_obst));
1485 obstack_init(isa->name_obst);
1486 isa->name_obst_size = 0;
1489 ia32_handle_intrinsics();
1490 ia32_switch_section(isa->out, NO_SECTION);
1491 fprintf(isa->out, "\t.intel_syntax\n");
1493 /* needed for the debug support */
1494 ia32_switch_section(isa->out, SECTION_TEXT);
1495 fprintf(isa->out, ".Ltext0:\n");
1505 * Closes the output file and frees the ISA structure.
1507 static void ia32_done(void *self) {
1508 ia32_isa_t *isa = self;
1510 /* emit now all global declarations */
1511 ia32_gen_decls(isa->out, isa->arch_isa.main_env);
1513 pmap_destroy(isa->regs_16bit);
1514 pmap_destroy(isa->regs_8bit);
1515 pmap_destroy(isa->tv_ent);
1516 pmap_destroy(isa->types);
1519 //printf("name obst size = %d bytes\n", isa->name_obst_size);
1520 obstack_free(isa->name_obst, NULL);
1528 * Return the number of register classes for this architecture.
1529 * We report always these:
1530 * - the general purpose registers
1531 * - the SSE floating point register set
1532 * - the virtual floating point registers
1534 static int ia32_get_n_reg_class(const void *self) {
1539 * Return the register class for index i.
1541 static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
1542 assert(i >= 0 && i < 3 && "Invalid ia32 register class requested.");
1544 return &ia32_reg_classes[CLASS_ia32_gp];
1546 return &ia32_reg_classes[CLASS_ia32_xmm];
1548 return &ia32_reg_classes[CLASS_ia32_vfp];
1552 * Get the register class which shall be used to store a value of a given mode.
1553 * @param self The this pointer.
1554 * @param mode The mode in question.
1555 * @return A register class which can hold values of the given mode.
1557 const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
1558 const ia32_isa_t *isa = self;
1559 if (mode_is_float(mode)) {
1560 return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1563 return &ia32_reg_classes[CLASS_ia32_gp];
1567 * Get the ABI restrictions for procedure calls.
1568 * @param self The this pointer.
1569 * @param method_type The type of the method (procedure) in question.
1570 * @param abi The abi object to be modified
1572 static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1573 const ia32_isa_t *isa = self;
1576 unsigned cc = get_method_calling_convention(method_type);
1577 int n = get_method_n_params(method_type);
1580 int i, ignore_1, ignore_2;
1582 const arch_register_t *reg;
1583 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1585 unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
1587 /* set abi flags for calls */
1588 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1589 call_flags.bits.store_args_sequential = use_push;
1590 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1591 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1592 call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
1594 /* set stack parameter passing style */
1595 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1597 /* collect the mode for each type */
1598 modes = alloca(n * sizeof(modes[0]));
1600 for (i = 0; i < n; i++) {
1601 tp = get_method_param_type(method_type, i);
1602 modes[i] = get_type_mode(tp);
1605 /* set register parameters */
1606 if (cc & cc_reg_param) {
1607 /* determine the number of parameters passed via registers */
1608 biggest_n = ia32_get_n_regparam_class(n, modes, &ignore_1, &ignore_2);
1610 /* loop over all parameters and set the register requirements */
1611 for (i = 0; i <= biggest_n; i++) {
1612 reg = ia32_get_RegParam_reg(n, modes, i, cc);
1613 assert(reg && "kaputt");
1614 be_abi_call_param_reg(abi, i, reg);
1621 /* set stack parameters */
1622 for (i = stack_idx; i < n; i++) {
1623 be_abi_call_param_stack(abi, i, 1, 0, 0);
1627 /* set return registers */
1628 n = get_method_n_ress(method_type);
1630 assert(n <= 2 && "more than two results not supported");
1632 /* In case of 64bit returns, we will have two 32bit values */
1634 tp = get_method_res_type(method_type, 0);
1635 mode = get_type_mode(tp);
1637 assert(!mode_is_float(mode) && "two FP results not supported");
1639 tp = get_method_res_type(method_type, 1);
1640 mode = get_type_mode(tp);
1642 assert(!mode_is_float(mode) && "two FP results not supported");
1644 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX]);
1645 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]);
1648 const arch_register_t *reg;
1650 tp = get_method_res_type(method_type, 0);
1651 assert(is_atomic_type(tp));
1652 mode = get_type_mode(tp);
1654 reg = mode_is_float(mode) ?
1655 (USE_SSE2(isa) ? &ia32_xmm_regs[REG_XMM0] : &ia32_vfp_regs[REG_VF0]) :
1656 &ia32_gp_regs[REG_EAX];
1658 be_abi_call_res_reg(abi, 0, reg);
1663 static const void *ia32_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1664 return &ia32_irn_ops;
1667 const arch_irn_handler_t ia32_irn_handler = {
1671 const arch_irn_handler_t *ia32_get_irn_handler(const void *self) {
1672 return &ia32_irn_handler;
1675 int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1676 return is_ia32_irn(irn) ? 1 : -1;
1680 * Initializes the code generator interface.
1682 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self) {
1683 return &ia32_code_gen_if;
1687 * Returns the estimated execution time of an ia32 irn.
1689 static sched_timestep_t ia32_sched_exectime(void *env, const ir_node *irn) {
1690 const arch_env_t *arch_env = env;
1691 return is_ia32_irn(irn) ? ia32_get_op_estimated_cost(arch_get_irn_ops(arch_env, irn), irn) : 1;
1694 list_sched_selector_t ia32_sched_selector;
1697 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1699 static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self, list_sched_selector_t *selector) {
1700 memcpy(&ia32_sched_selector, selector, sizeof(ia32_sched_selector));
1701 ia32_sched_selector.exectime = ia32_sched_exectime;
1702 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
1703 return &ia32_sched_selector;
1707 * Returns the necessary byte alignment for storing a register of given class.
1709 static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1710 ir_mode *mode = arch_register_class_mode(cls);
1711 int bytes = get_mode_size_bytes(mode);
1713 if (mode_is_float(mode) && bytes > 8)
1718 static ia32_intrinsic_env_t intrinsic_env = { NULL, NULL };
1721 * Returns the libFirm configuration parameter for this backend.
1723 static const backend_params *ia32_get_libfirm_params(void) {
1724 static const arch_dep_params_t ad = {
1725 1, /* also use subs */
1726 4, /* maximum shifts */
1727 31, /* maximum shift amount */
1729 1, /* allow Mulhs */
1730 1, /* allow Mulus */
1731 32 /* Mulh allowed up to 32 bit */
1733 static backend_params p = {
1734 NULL, /* no additional opcodes */
1735 NULL, /* will be set later */
1736 1, /* need dword lowering */
1737 ia32_create_intrinsic_fkt,
1738 &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
1746 /* instruction set architectures. */
1747 static const lc_opt_enum_int_items_t arch_items[] = {
1748 { "386", arch_i386, },
1749 { "486", arch_i486, },
1750 { "pentium", arch_pentium, },
1751 { "586", arch_pentium, },
1752 { "pentiumpro", arch_pentium_pro, },
1753 { "686", arch_pentium_pro, },
1754 { "pentiummmx", arch_pentium_mmx, },
1755 { "pentium2", arch_pentium_2, },
1756 { "p2", arch_pentium_2, },
1757 { "pentium3", arch_pentium_3, },
1758 { "p3", arch_pentium_3, },
1759 { "pentium4", arch_pentium_4, },
1760 { "p4", arch_pentium_4, },
1761 { "pentiumm", arch_pentium_m, },
1762 { "pm", arch_pentium_m, },
1763 { "core", arch_core, },
1765 { "athlon", arch_athlon, },
1766 { "athlon64", arch_athlon_64, },
1767 { "opteron", arch_opteron, },
1771 static lc_opt_enum_int_var_t arch_var = {
1772 &ia32_isa_template.arch, arch_items
1775 static lc_opt_enum_int_var_t opt_arch_var = {
1776 &ia32_isa_template.opt_arch, arch_items
1779 static const lc_opt_enum_int_items_t fp_unit_items[] = {
1781 { "sse2", fp_sse2 },
1785 static lc_opt_enum_int_var_t fp_unit_var = {
1786 &ia32_isa_template.fp_kind, fp_unit_items
1789 static const lc_opt_enum_int_items_t gas_items[] = {
1790 { "linux", ASM_LINUX_GAS },
1791 { "mingw", ASM_MINGW_GAS },
1795 static lc_opt_enum_int_var_t gas_var = {
1796 (int *)&asm_flavour, gas_items
1799 static const lc_opt_table_entry_t ia32_options[] = {
1800 LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var),
1801 LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var),
1802 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var),
1803 LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
1804 LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA),
1805 LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST),
1806 LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS),
1807 LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB),
1808 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
1813 * Register command line options for the ia32 backend.
1817 * ia32-arch=arch create instruction for arch
1818 * ia32-opt=arch optimize for run on arch
1819 * ia32-fpunit=unit select floating point unit (x87 or SSE2)
1820 * ia32-incdec optimize for inc/dec
1821 * ia32-noaddrmode do not use address mode
1822 * ia32-nolea do not optimize for LEAs
1823 * ia32-noplacecnst do not place constants,
1824 * ia32-noimmop no operations with immediates
1825 * ia32-noextbb do not use extended basic block scheduling
1826 * ia32-gasmode set the GAS compatibility mode
1828 static void ia32_register_options(lc_opt_entry_t *ent)
1830 lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32");
1831 lc_opt_add_table(be_grp_ia32, ia32_options);
1833 #endif /* WITH_LIBCORE */
1835 const arch_isa_if_t ia32_isa_if = {
1838 ia32_get_n_reg_class,
1840 ia32_get_reg_class_for_mode,
1842 ia32_get_irn_handler,
1843 ia32_get_code_generator_if,
1844 ia32_get_list_sched_selector,
1845 ia32_get_reg_class_alignment,
1846 ia32_get_libfirm_params,
1848 ia32_register_options