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"
39 #include "../beabi.h" /* the general register allocator interface */
40 #include "../benode_t.h"
41 #include "../belower.h"
42 #include "../besched_t.h"
45 #include "../beirgmod.h"
46 #include "../be_dbgout.h"
47 #include "../beblocksched.h"
48 #include "../bemachine.h"
49 #include "../beilpsched.h"
51 #include "bearch_ia32_t.h"
53 #include "ia32_new_nodes.h" /* ia32 nodes interface */
54 #include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */
55 #include "gen_ia32_machine.h"
56 #include "ia32_gen_decls.h" /* interface declaration emitter */
57 #include "ia32_transform.h"
58 #include "ia32_emitter.h"
59 #include "ia32_map_regs.h"
60 #include "ia32_optimize.h"
62 #include "ia32_dbg_stat.h"
63 #include "ia32_finish.h"
64 #include "ia32_util.h"
66 #define DEBUG_MODULE "firm.be.ia32.isa"
69 static set *cur_reg_set = NULL;
71 /* Creates the unique per irg GP NoReg node. */
72 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) {
73 return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_GP_NOREG]);
76 /* Creates the unique per irg FP NoReg node. */
77 ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) {
78 return be_abi_get_callee_save_irn(cg->birg->abi,
79 USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]);
83 * Returns gp_noreg or fp_noreg, depending in input requirements.
85 ir_node *ia32_get_admissible_noreg(ia32_code_gen_t *cg, ir_node *irn, int pos) {
86 arch_register_req_t req;
87 const arch_register_req_t *p_req;
89 p_req = arch_get_register_req(cg->arch_env, &req, irn, pos);
90 assert(p_req && "Missing register requirements");
91 if (p_req->cls == &ia32_reg_classes[CLASS_ia32_gp])
92 return ia32_new_NoReg_gp(cg);
94 return ia32_new_NoReg_fp(cg);
97 /**************************************************
100 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
101 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
102 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
103 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
106 **************************************************/
109 * Return register requirements for an ia32 node.
110 * If the node returns a tuple (mode_T) then the proj's
111 * will be asked for this information.
113 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) {
114 const ia32_irn_ops_t *ops = self;
115 const ia32_register_req_t *irn_req;
116 long node_pos = pos == -1 ? 0 : pos;
117 ir_mode *mode = is_Block(irn) ? NULL : get_irn_mode(irn);
118 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, DEBUG_MODULE);
120 if (is_Block(irn) || mode == mode_M || mode == mode_X) {
121 DBG((mod, LEVEL_1, "ignoring Block, mode_M, mode_X node %+F\n", irn));
125 if (mode == mode_T && pos < 0) {
126 DBG((mod, LEVEL_1, "ignoring request OUT requirements for node %+F\n", irn));
130 DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn));
134 DBG((mod, LEVEL_1, "ignoring request IN requirements for node %+F\n", irn));
138 node_pos = (pos == -1) ? get_Proj_proj(irn) : pos;
139 irn = skip_Proj_const(irn);
141 DB((mod, LEVEL_1, "skipping Proj, going to %+F at pos %d ... ", irn, node_pos));
144 if (is_ia32_irn(irn)) {
145 irn_req = (pos >= 0) ? get_ia32_in_req(irn, pos) : get_ia32_out_req(irn, node_pos);
146 if (irn_req == NULL) {
147 /* no requirements */
151 DB((mod, LEVEL_1, "returning reqs for %+F at pos %d\n", irn, pos));
153 memcpy(req, &(irn_req->req), sizeof(*req));
155 if (arch_register_req_is(&(irn_req->req), should_be_same)) {
156 assert(irn_req->same_pos >= 0 && "should be same constraint for in -> out NYI");
157 req->other_same = get_irn_n(irn, irn_req->same_pos);
160 if (arch_register_req_is(&(irn_req->req), should_be_different)) {
161 assert(irn_req->different_pos >= 0 && "should be different constraint for in -> out NYI");
162 req->other_different = get_irn_n(irn, irn_req->different_pos);
166 /* treat Unknowns like Const with default requirements */
167 if (is_Unknown(irn)) {
168 DB((mod, LEVEL_1, "returning UKNWN reqs for %+F\n", irn));
169 if (mode_is_float(mode)) {
170 if (USE_SSE2(ops->cg))
171 memcpy(req, &(ia32_default_req_ia32_xmm_xmm_UKNWN), sizeof(*req));
173 memcpy(req, &(ia32_default_req_ia32_vfp_vfp_UKNWN), sizeof(*req));
175 else if (mode_is_int(mode) || mode_is_reference(mode))
176 memcpy(req, &(ia32_default_req_ia32_gp_gp_UKNWN), sizeof(*req));
177 else if (mode == mode_T || mode == mode_M) {
178 DBG((mod, LEVEL_1, "ignoring Unknown node %+F\n", irn));
182 assert(0 && "unsupported Unknown-Mode");
185 DB((mod, LEVEL_1, "returning NULL for %+F (not ia32)\n", irn));
193 static void ia32_set_irn_reg(const void *self, ir_node *irn, const arch_register_t *reg) {
195 const ia32_irn_ops_t *ops = self;
197 if (get_irn_mode(irn) == mode_X) {
201 DBG((ops->cg->mod, LEVEL_1, "ia32 assigned register %s to node %+F\n", reg->name, irn));
204 pos = get_Proj_proj(irn);
205 irn = skip_Proj(irn);
208 if (is_ia32_irn(irn)) {
209 const arch_register_t **slots;
211 slots = get_ia32_slots(irn);
215 ia32_set_firm_reg(irn, reg, cur_reg_set);
219 static const arch_register_t *ia32_get_irn_reg(const void *self, const ir_node *irn) {
221 const arch_register_t *reg = NULL;
225 if (get_irn_mode(irn) == mode_X) {
229 pos = get_Proj_proj(irn);
230 irn = skip_Proj_const(irn);
233 if (is_ia32_irn(irn)) {
234 const arch_register_t **slots;
235 slots = get_ia32_slots(irn);
239 reg = ia32_get_firm_reg(irn, cur_reg_set);
245 static arch_irn_class_t ia32_classify(const void *self, const ir_node *irn) {
246 arch_irn_class_t classification = arch_irn_class_normal;
248 irn = skip_Proj_const(irn);
251 classification |= arch_irn_class_branch;
253 if (! is_ia32_irn(irn))
254 return classification & ~arch_irn_class_normal;
256 if (is_ia32_Cnst(irn))
257 classification |= arch_irn_class_const;
260 classification |= arch_irn_class_load;
262 if (is_ia32_St(irn) || is_ia32_Store8Bit(irn))
263 classification |= arch_irn_class_store;
265 if (is_ia32_got_reload(irn))
266 classification |= arch_irn_class_reload;
268 return classification;
271 static arch_irn_flags_t ia32_get_flags(const void *self, const ir_node *irn) {
272 arch_irn_flags_t flags;
273 ir_node *pred = is_Proj(irn) && mode_is_datab(get_irn_mode(irn)) ? get_Proj_pred(irn) : NULL;
276 flags = arch_irn_flags_ignore;
278 /* pred is only set, if we have a Proj */
279 flags = pred && is_ia32_irn(pred) ? get_ia32_out_flags(pred, get_Proj_proj(irn)) : arch_irn_flags_none;
281 irn = skip_Proj_const(irn);
282 if (is_ia32_irn(irn))
283 flags |= get_ia32_flags(irn);
290 * The IA32 ABI callback object.
293 be_abi_call_flags_bits_t flags; /**< The call flags. */
294 const arch_isa_t *isa; /**< The ISA handle. */
295 const arch_env_t *aenv; /**< The architecture environment. */
296 ir_graph *irg; /**< The associated graph. */
299 static ir_entity *ia32_get_frame_entity(const void *self, const ir_node *irn) {
300 return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL;
303 static void ia32_set_frame_entity(const void *self, ir_node *irn, ir_entity *ent) {
304 set_ia32_frame_ent(irn, ent);
307 static void ia32_set_frame_offset(const void *self, ir_node *irn, int bias) {
309 const ia32_irn_ops_t *ops = self;
311 if (get_ia32_frame_ent(irn)) {
312 ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn);
314 if(is_ia32_Pop(irn)) {
315 int omit_fp = be_abi_omit_fp(ops->cg->birg->abi);
317 /* Pop nodes modify the stack pointer before calculating the destination
318 * address, so fix this here
324 DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias));
326 snprintf(buf, sizeof(buf), "%d", bias);
328 if (get_ia32_op_type(irn) == ia32_Normal) {
329 set_ia32_cnst(irn, buf);
331 add_ia32_am_offs(irn, buf);
333 set_ia32_am_flavour(irn, am_flav);
338 static int ia32_get_sp_bias(const void *self, const ir_node *irn) {
340 long proj = get_Proj_proj(irn);
341 ir_node *pred = get_Proj_pred(irn);
343 if (proj == pn_ia32_Push_stack && is_ia32_Push(pred))
345 if (proj == pn_ia32_Pop_stack && is_ia32_Pop(pred))
353 * Put all registers which are saved by the prologue/epilogue in a set.
355 * @param self The callback object.
356 * @param s The result set.
358 static void ia32_abi_dont_save_regs(void *self, pset *s)
360 ia32_abi_env_t *env = self;
361 if(env->flags.try_omit_fp)
362 pset_insert_ptr(s, env->isa->bp);
366 * Generate the routine prologue.
368 * @param self The callback object.
369 * @param mem A pointer to the mem node. Update this if you define new memory.
370 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
372 * @return The register which shall be used as a stack frame base.
374 * All nodes which define registers in @p reg_map must keep @p reg_map current.
376 static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
378 ia32_abi_env_t *env = self;
380 if (! env->flags.try_omit_fp) {
381 ir_node *bl = get_irg_start_block(env->irg);
382 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
383 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
384 ir_node *noreg = be_abi_reg_map_get(reg_map, &ia32_gp_regs[REG_GP_NOREG]);
388 push = new_rd_ia32_Push(NULL, env->irg, bl, noreg, noreg, curr_bp, curr_sp, *mem);
389 curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
390 *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M);
392 /* the push must have SP out register */
393 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
394 set_ia32_flags(push, arch_irn_flags_ignore);
396 /* move esp to ebp */
397 curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp);
398 be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp);
399 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
400 be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
402 /* beware: the copy must be done before any other sp use */
403 curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
404 be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp);
405 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
406 be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
408 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
409 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
418 * Generate the routine epilogue.
419 * @param self The callback object.
420 * @param bl The block for the epilog
421 * @param mem A pointer to the mem node. Update this if you define new memory.
422 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
423 * @return The register which shall be used as a stack frame base.
425 * All nodes which define registers in @p reg_map must keep @p reg_map current.
427 static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
429 ia32_abi_env_t *env = self;
430 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
431 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
433 if (env->flags.try_omit_fp) {
434 /* simply remove the stack frame here */
435 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK);
436 add_irn_dep(curr_sp, *mem);
439 const ia32_isa_t *isa = (ia32_isa_t *)env->isa;
440 ir_mode *mode_bp = env->isa->bp->reg_class->mode;
442 /* gcc always emits a leave at the end of a routine */
443 if (1 || ARCH_AMD(isa->opt_arch)) {
447 leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, curr_bp);
448 set_ia32_flags(leave, arch_irn_flags_ignore);
449 curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame);
450 curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
451 *mem = new_r_Proj(current_ir_graph, bl, leave, mode_M, pn_ia32_Leave_M);
454 ir_node *noreg = be_abi_reg_map_get(reg_map, &ia32_gp_regs[REG_GP_NOREG]);
457 /* copy ebp to esp */
458 curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem);
461 pop = new_rd_ia32_Pop(NULL, env->irg, bl, noreg, noreg, curr_sp, *mem);
462 set_ia32_flags(pop, arch_irn_flags_ignore);
463 curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res);
464 curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
465 *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
467 arch_set_irn_register(env->aenv, curr_sp, env->isa->sp);
468 arch_set_irn_register(env->aenv, curr_bp, env->isa->bp);
471 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
472 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
476 * Initialize the callback object.
477 * @param call The call object.
478 * @param aenv The architecture environment.
479 * @param irg The graph with the method.
480 * @return Some pointer. This pointer is passed to all other callback functions as self object.
482 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
484 ia32_abi_env_t *env = xmalloc(sizeof(env[0]));
485 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
486 env->flags = fl.bits;
489 env->isa = aenv->isa;
494 * Destroy the callback object.
495 * @param self The callback object.
497 static void ia32_abi_done(void *self) {
502 * Produces the type which sits between the stack args and the locals on the stack.
503 * it will contain the return address and space to store the old base pointer.
504 * @return The Firm type modeling the ABI between type.
506 static ir_type *ia32_abi_get_between_type(void *self)
508 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
509 static ir_type *omit_fp_between_type = NULL;
510 static ir_type *between_type = NULL;
512 ia32_abi_env_t *env = self;
514 if (! between_type) {
515 ir_entity *old_bp_ent;
516 ir_entity *ret_addr_ent;
517 ir_entity *omit_fp_ret_addr_ent;
519 ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P);
520 ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_P);
522 between_type = new_type_struct(IDENT("ia32_between_type"));
523 old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type);
524 ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type);
526 set_entity_offset(old_bp_ent, 0);
527 set_entity_offset(ret_addr_ent, get_type_size_bytes(old_bp_type));
528 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
529 set_type_state(between_type, layout_fixed);
531 omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp"));
532 omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type);
534 set_entity_offset(omit_fp_ret_addr_ent, 0);
535 set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
536 set_type_state(omit_fp_between_type, layout_fixed);
539 return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
544 * Get the estimated cycle count for @p irn.
546 * @param self The this pointer.
547 * @param irn The node.
549 * @return The estimated cycle count for this operation
551 static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn)
554 ia32_op_type_t op_tp;
555 const ia32_irn_ops_t *ops = self;
559 if (!is_ia32_irn(irn))
562 assert(is_ia32_irn(irn));
564 cost = get_ia32_latency(irn);
565 op_tp = get_ia32_op_type(irn);
567 if (is_ia32_CopyB(irn)) {
569 if (ARCH_INTEL(ops->cg->arch))
572 else if (is_ia32_CopyB_i(irn)) {
573 int size = get_tarval_long(get_ia32_Immop_tarval(irn));
574 cost = 20 + (int)ceil((4/3) * size);
575 if (ARCH_INTEL(ops->cg->arch))
578 /* in case of address mode operations add additional cycles */
579 else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) {
581 In case of stack access add 5 cycles (we assume stack is in cache),
582 other memory operations cost 20 cycles.
584 cost += is_ia32_use_frame(irn) ? 5 : 20;
591 * Returns the inverse operation if @p irn, recalculating the argument at position @p i.
593 * @param irn The original operation
594 * @param i Index of the argument we want the inverse operation to yield
595 * @param inverse struct to be filled with the resulting inverse op
596 * @param obstack The obstack to use for allocation of the returned nodes array
597 * @return The inverse operation or NULL if operation invertible
599 static arch_inverse_t *ia32_get_inverse(const void *self, const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) {
602 ir_node *block, *noreg, *nomem;
605 /* we cannot invert non-ia32 irns */
606 if (! is_ia32_irn(irn))
609 /* operand must always be a real operand (not base, index or mem) */
610 if (i != 2 && i != 3)
613 /* we don't invert address mode operations */
614 if (get_ia32_op_type(irn) != ia32_Normal)
617 irg = get_irn_irg(irn);
618 block = get_nodes_block(irn);
619 mode = get_ia32_res_mode(irn);
620 noreg = get_irn_n(irn, 0);
621 nomem = new_r_NoMem(irg);
623 /* initialize structure */
624 inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0]));
628 switch (get_ia32_irn_opcode(irn)) {
630 if (get_ia32_immop_type(irn) == ia32_ImmConst) {
631 /* we have an add with a const here */
632 /* invers == add with negated const */
633 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
634 pnc = pn_ia32_Add_res;
636 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
637 set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
638 set_ia32_commutative(inverse->nodes[0]);
640 else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
641 /* we have an add with a symconst here */
642 /* invers == sub with const */
643 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
644 pnc = pn_ia32_Sub_res;
646 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
649 /* normal add: inverse == sub */
650 ir_node *proj = ia32_get_res_proj(irn);
653 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem);
654 pnc = pn_ia32_Sub_res;
659 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
660 /* we have a sub with a const/symconst here */
661 /* invers == add with this const */
662 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
663 pnc = pn_ia32_Add_res;
664 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
665 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
669 ir_node *proj = ia32_get_res_proj(irn);
673 inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem);
676 inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, 2), proj, nomem);
678 pnc = pn_ia32_Sub_res;
683 if (get_ia32_immop_type(irn) != ia32_ImmNone) {
684 /* xor with const: inverse = xor */
685 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem);
686 pnc = pn_ia32_Eor_res;
687 inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
688 copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
692 inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, (ir_node *)irn, get_irn_n(irn, i), nomem);
693 pnc = pn_ia32_Eor_res;
698 ir_node *proj = ia32_get_res_proj(irn);
701 inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, proj, nomem);
702 pnc = pn_ia32_Not_res;
706 case iro_ia32_Minus: {
707 ir_node *proj = ia32_get_res_proj(irn);
710 inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, proj, nomem);
711 pnc = pn_ia32_Minus_res;
716 /* inverse operation not supported */
720 set_ia32_res_mode(inverse->nodes[0], mode);
721 inverse->nodes[1] = new_r_Proj(irg, block, inverse->nodes[0], mode, pnc);
727 * Check if irn can load it's operand at position i from memory (source addressmode).
728 * @param self Pointer to irn ops itself
729 * @param irn The irn to be checked
730 * @param i The operands position
731 * @return Non-Zero if operand can be loaded
733 static int ia32_possible_memory_operand(const void *self, const ir_node *irn, unsigned int i) {
734 if (! is_ia32_irn(irn) || /* must be an ia32 irn */
735 get_irn_arity(irn) != 5 || /* must be a binary operation */
736 get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */
737 ! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */
738 (i != 2 && i != 3) || /* a "real" operand position must be requested */
739 (i == 2 && ! is_ia32_commutative(irn)) || /* if first operand requested irn must be commutative */
740 is_ia32_use_frame(irn)) /* must not already use frame */
746 static void ia32_perform_memory_operand(const void *self, ir_node *irn, ir_node *spill, unsigned int i) {
747 const ia32_irn_ops_t *ops = self;
748 ia32_code_gen_t *cg = ops->cg;
750 assert(ia32_possible_memory_operand(self, irn, i) && "Cannot perform memory operand change");
753 ir_node *tmp = get_irn_n(irn, 3);
754 set_irn_n(irn, 3, get_irn_n(irn, 2));
755 set_irn_n(irn, 2, tmp);
758 set_ia32_am_support(irn, ia32_am_Source);
759 set_ia32_op_type(irn, ia32_AddrModeS);
760 set_ia32_am_flavour(irn, ia32_B);
761 set_ia32_ls_mode(irn, get_irn_mode(get_irn_n(irn, i)));
762 set_ia32_use_frame(irn);
763 set_ia32_got_reload(irn);
765 set_irn_n(irn, 0, get_irg_frame(get_irn_irg(irn)));
766 set_irn_n(irn, 4, spill);
769 Input at position one is index register, which is NoReg.
770 We would need cg object to get a real noreg, but we cannot
773 set_irn_n(irn, 3, ia32_get_admissible_noreg(cg, irn, 3));
775 //FIXME DBG_OPT_AM_S(reload, irn);
778 static const be_abi_callbacks_t ia32_abi_callbacks = {
781 ia32_abi_get_between_type,
782 ia32_abi_dont_save_regs,
787 /* fill register allocator interface */
789 static const arch_irn_ops_if_t ia32_irn_ops_if = {
790 ia32_get_irn_reg_req,
795 ia32_get_frame_entity,
796 ia32_set_frame_entity,
797 ia32_set_frame_offset,
800 ia32_get_op_estimated_cost,
801 ia32_possible_memory_operand,
802 ia32_perform_memory_operand,
805 ia32_irn_ops_t ia32_irn_ops = {
812 /**************************************************
815 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
816 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
817 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
818 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
821 **************************************************/
823 static void ia32_kill_convs(ia32_code_gen_t *cg) {
826 /* BEWARE: the Projs are inserted in the set */
827 foreach_nodeset(cg->kill_conv, irn) {
828 ir_node *in = get_irn_n(get_Proj_pred(irn), 2);
829 edges_reroute(irn, in, cg->birg->irg);
834 * Transform the Thread Local Store base.
836 static void transform_tls(ir_graph *irg) {
837 ir_node *irn = get_irg_tls(irg);
840 dbg_info *dbg = get_irn_dbg_info(irn);
841 ir_node *blk = get_nodes_block(irn);
843 newn = new_rd_ia32_LdTls(dbg, irg, blk, get_irn_mode(irn));
850 * Transforms the standard firm graph into
853 static void ia32_prepare_graph(void *self) {
854 ia32_code_gen_t *cg = self;
855 DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;)
857 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform");
859 /* 1st: transform constants and psi condition trees */
860 ia32_pre_transform_phase(cg);
862 /* 2nd: transform all remaining nodes */
863 ia32_register_transformers();
865 cg->kill_conv = new_nodeset(5);
866 transform_tls(cg->irg);
867 edges_deactivate(cg->irg);
868 edges_activate(cg->irg);
869 irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg);
871 del_nodeset(cg->kill_conv);
874 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
876 /* 3rd: optimize address mode */
877 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am");
878 ia32_optimize_addressmode(cg);
881 be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
883 DEBUG_ONLY(cg->mod = old_mod;)
887 * Dummy functions for hooks we don't need but which must be filled.
889 static void ia32_before_sched(void *self) {
892 static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) {
895 ir_node *mem_proj = NULL;
900 mode = get_irn_mode(irn);
902 /* check if we already saw this node or the node has more than one user */
903 if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1) {
907 /* mark irn visited */
908 bitset_add_irn(already_visited, irn);
910 /* non-Tuple nodes with one user: ok, return */
911 if (get_irn_n_edges(irn) >= 1 && mode != mode_T) {
915 /* tuple node has one user which is not the mem proj-> ok */
916 if (mode == mode_T && get_irn_n_edges(irn) == 1) {
917 mem_proj = ia32_get_proj_for_mode(irn, mode_M);
918 if (mem_proj == NULL) {
923 arity = get_irn_arity(irn);
924 for (i = 0; i < arity; ++i) {
925 ir_node *pred = get_irn_n(irn, i);
927 /* do not follow memory edges or we will accidentally remove stores */
928 if (get_irn_mode(pred) == mode_M) {
929 if(mem_proj != NULL) {
930 edges_reroute(mem_proj, pred, get_irn_irg(mem_proj));
936 set_irn_n(irn, i, new_Bad());
939 The current node is about to be removed: if the predecessor
940 has only this node as user, it need to be removed as well.
942 if (get_irn_n_edges(pred) <= 1)
943 remove_unused_nodes(pred, already_visited);
946 // we need to set the presd to Bad again to also get the memory edges
947 arity = get_irn_arity(irn);
948 for (i = 0; i < arity; ++i) {
949 set_irn_n(irn, i, new_Bad());
952 if (sched_is_scheduled(irn)) {
957 static void remove_unused_loads_walker(ir_node *irn, void *env) {
958 bitset_t *already_visited = env;
959 if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn))
960 remove_unused_nodes(irn, env);
964 * Called before the register allocator.
965 * Calculate a block schedule here. We need it for the x87
966 * simulator and the emitter.
968 static void ia32_before_ra(void *self) {
969 ia32_code_gen_t *cg = self;
970 bitset_t *already_visited = bitset_irg_alloca(cg->irg);
974 There are sometimes unused loads, only pinned by memory.
975 We need to remove those Loads and all other nodes which won't be used
976 after removing the Load from schedule.
978 irg_walk_graph(cg->irg, NULL, remove_unused_loads_walker, already_visited);
983 * Transforms a be node into a Load.
985 static void transform_to_Load(ia32_transform_env_t *env) {
986 ir_node *irn = env->irn;
987 ir_entity *ent = be_get_frame_entity(irn);
988 ir_mode *mode = env->mode;
989 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
990 ir_node *sched_point = NULL;
991 ir_node *ptr = get_irg_frame(env->irg);
992 ir_node *mem = get_irn_n(irn, be_pos_Reload_mem);
993 ir_node *new_op, *proj;
994 const arch_register_t *reg;
996 if (sched_is_scheduled(irn)) {
997 sched_point = sched_prev(irn);
1000 if (mode_is_float(mode)) {
1001 if (USE_SSE2(env->cg))
1002 new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem);
1004 new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem);
1007 new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem);
1009 set_ia32_am_support(new_op, ia32_am_Source);
1010 set_ia32_op_type(new_op, ia32_AddrModeS);
1011 set_ia32_am_flavour(new_op, ia32_B);
1012 set_ia32_ls_mode(new_op, mode);
1013 set_ia32_frame_ent(new_op, ent);
1014 set_ia32_use_frame(new_op);
1016 DBG_OPT_RELOAD2LD(irn, new_op);
1018 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_ia32_Load_res);
1021 sched_add_after(sched_point, new_op);
1022 sched_add_after(new_op, proj);
1027 /* copy the register from the old node to the new Load */
1028 reg = arch_get_irn_register(env->cg->arch_env, irn);
1029 arch_set_irn_register(env->cg->arch_env, new_op, reg);
1031 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
1033 exchange(irn, proj);
1037 * Transforms a be node into a Store.
1039 static void transform_to_Store(ia32_transform_env_t *env) {
1040 ir_node *irn = env->irn;
1041 ir_entity *ent = be_get_frame_entity(irn);
1042 ir_mode *mode = env->mode;
1043 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1044 ir_node *nomem = new_rd_NoMem(env->irg);
1045 ir_node *ptr = get_irg_frame(env->irg);
1046 ir_node *val = get_irn_n(irn, be_pos_Spill_val);
1047 ir_node *new_op, *proj;
1048 ir_node *sched_point = NULL;
1050 if (sched_is_scheduled(irn)) {
1051 sched_point = sched_prev(irn);
1054 if (mode_is_float(mode)) {
1055 if (USE_SSE2(env->cg))
1056 new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1058 new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1060 else if (get_mode_size_bits(mode) == 8) {
1061 new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1064 new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem);
1067 set_ia32_am_support(new_op, ia32_am_Dest);
1068 set_ia32_op_type(new_op, ia32_AddrModeD);
1069 set_ia32_am_flavour(new_op, ia32_B);
1070 set_ia32_ls_mode(new_op, mode);
1071 set_ia32_frame_ent(new_op, ent);
1072 set_ia32_use_frame(new_op);
1074 DBG_OPT_SPILL2ST(irn, new_op);
1076 proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, pn_ia32_Store_M);
1079 sched_add_after(sched_point, new_op);
1083 SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn));
1085 exchange(irn, proj);
1088 static ir_node *create_push(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, ir_node *mem, ir_entity *ent) {
1089 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1090 ir_node *frame = get_irg_frame(env->irg);
1092 ir_node *push = new_rd_ia32_Push(env->dbg, env->irg, env->block, frame, noreg, noreg, sp, mem);
1094 set_ia32_frame_ent(push, ent);
1095 set_ia32_use_frame(push);
1096 set_ia32_op_type(push, ia32_AddrModeS);
1097 set_ia32_am_flavour(push, ia32_B);
1098 set_ia32_ls_mode(push, mode_Is);
1100 sched_add_before(schedpoint, push);
1104 static ir_node *create_pop(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, ir_entity *ent) {
1105 ir_node *noreg = ia32_new_NoReg_gp(env->cg);
1106 ir_node *frame = get_irg_frame(env->irg);
1108 ir_node *pop = new_rd_ia32_Pop(env->dbg, env->irg, env->block, frame, noreg, sp, new_NoMem());
1110 set_ia32_frame_ent(pop, ent);
1111 set_ia32_use_frame(pop);
1112 set_ia32_op_type(pop, ia32_AddrModeD);
1113 set_ia32_am_flavour(pop, ia32_B);
1114 set_ia32_ls_mode(pop, mode_Is);
1116 sched_add_before(schedpoint, pop);
1121 static ir_node* create_spproj(ia32_transform_env_t *env, ir_node *pred, int pos, ir_node *schedpoint) {
1122 ir_mode *spmode = mode_Iu;
1123 const arch_register_t *spreg = &ia32_gp_regs[REG_ESP];
1126 sp = new_rd_Proj(env->dbg, env->irg, env->block, pred, spmode, pos);
1127 arch_set_irn_register(env->cg->arch_env, sp, spreg);
1128 sched_add_before(schedpoint, sp);
1134 * Transform memperm, currently we do this the ugly way and produce
1135 * push/pop into/from memory cascades. This is possible without using
1138 static void transform_MemPerm(ia32_transform_env_t *env) {
1139 ir_node *node = env->irn;
1141 ir_node *sp = be_abi_get_ignore_irn(env->cg->birg->abi, &ia32_gp_regs[REG_ESP]);
1142 const ir_edge_t *edge;
1143 const ir_edge_t *next;
1146 arity = be_get_MemPerm_entity_arity(node);
1147 pops = alloca(arity * sizeof(pops[0]));
1150 for(i = 0; i < arity; ++i) {
1151 ir_entity *ent = be_get_MemPerm_in_entity(node, i);
1152 ir_type *enttype = get_entity_type(ent);
1153 int entbits = get_type_size_bits(enttype);
1154 ir_node *mem = get_irn_n(node, i + 1);
1157 assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
1159 push = create_push(env, node, sp, mem, ent);
1160 sp = create_spproj(env, push, 0, node);
1162 // add another push after the first one
1163 push = create_push(env, node, sp, mem, ent);
1164 add_ia32_am_offs_int(push, 4);
1165 sp = create_spproj(env, push, 0, node);
1168 set_irn_n(node, i, new_Bad());
1172 for(i = arity - 1; i >= 0; --i) {
1173 ir_entity *ent = be_get_MemPerm_out_entity(node, i);
1174 ir_type *enttype = get_entity_type(ent);
1175 int entbits = get_type_size_bits(enttype);
1179 assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit");
1181 pop = create_pop(env, node, sp, ent);
1183 // add another pop after the first one
1184 sp = create_spproj(env, pop, 1, node);
1185 pop = create_pop(env, node, sp, ent);
1186 add_ia32_am_offs_int(pop, 4);
1188 sp = create_spproj(env, pop, 1, node);
1193 // exchange memprojs
1194 foreach_out_edge_safe(node, edge, next) {
1195 ir_node *proj = get_edge_src_irn(edge);
1196 int p = get_Proj_proj(proj);
1200 set_Proj_pred(proj, pops[p]);
1201 set_Proj_proj(proj, 3);
1205 arity = get_irn_arity(node);
1206 for(i = 0; i < arity; ++i) {
1207 set_irn_n(node, i, new_Bad());
1213 * Fix the mode of Spill/Reload
1215 static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode)
1217 if (mode_is_float(mode)) {
1229 * Block-Walker: Calls the transform functions Spill and Reload.
1231 static void ia32_after_ra_walker(ir_node *block, void *env) {
1232 ir_node *node, *prev;
1233 ia32_code_gen_t *cg = env;
1234 ia32_transform_env_t tenv;
1237 tenv.irg = current_ir_graph;
1239 DEBUG_ONLY(tenv.mod = cg->mod;)
1241 /* beware: the schedule is changed here */
1242 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
1243 prev = sched_prev(node);
1244 if (be_is_Reload(node)) {
1245 /* we always reload the whole register */
1246 tenv.dbg = get_irn_dbg_info(node);
1248 tenv.mode = fix_spill_mode(cg, get_irn_mode(node));
1249 transform_to_Load(&tenv);
1251 else if (be_is_Spill(node)) {
1252 ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
1253 /* we always spill the whole register */
1254 tenv.dbg = get_irn_dbg_info(node);
1256 tenv.mode = fix_spill_mode(cg, get_irn_mode(spillval));
1257 transform_to_Store(&tenv);
1259 else if(be_is_MemPerm(node)) {
1260 tenv.dbg = get_irn_dbg_info(node);
1262 transform_MemPerm(&tenv);
1268 * We transform Spill and Reload here. This needs to be done before
1269 * stack biasing otherwise we would miss the corrected offset for these nodes.
1271 * If x87 instruction should be emitted, run the x87 simulator and patch
1272 * the virtual instructions. This must obviously be done after register allocation.
1274 static void ia32_after_ra(void *self) {
1275 ia32_code_gen_t *cg = self;
1276 ir_graph *irg = cg->irg;
1278 irg_block_walk_graph(irg, NULL, ia32_after_ra_walker, cg);
1280 ia32_finish_irg(irg, cg);
1284 * Last touchups for the graph before emit
1286 static void ia32_finish(void *self) {
1287 ia32_code_gen_t *cg = self;
1288 ir_graph *irg = cg->irg;
1290 /* if we do x87 code generation, rewrite all the virtual instructions and registers */
1291 if (cg->used_fp == fp_x87 || cg->force_sim) {
1292 x87_simulate_graph(cg->arch_env, cg->birg);
1295 /* create block schedule, this also removes empty blocks which might
1296 * produce critical edges */
1297 cg->blk_sched = be_create_block_schedule(irg, cg->birg->exec_freq);
1299 /* do peephole optimisations */
1300 ia32_peephole_optimization(irg, cg);
1304 * Emits the code, closes the output file and frees
1305 * the code generator interface.
1307 static void ia32_codegen(void *self) {
1308 ia32_code_gen_t *cg = self;
1309 ir_graph *irg = cg->irg;
1311 ia32_gen_routine(cg->isa->out, irg, cg);
1315 /* remove it from the isa */
1318 /* de-allocate code generator */
1319 del_set(cg->reg_set);
1323 static void *ia32_cg_init(be_irg_t *birg);
1325 static const arch_code_generator_if_t ia32_code_gen_if = {
1327 NULL, /* before abi introduce hook */
1330 ia32_before_sched, /* before scheduling hook */
1331 ia32_before_ra, /* before register allocation hook */
1332 ia32_after_ra, /* after register allocation hook */
1333 ia32_finish, /* called before codegen */
1334 ia32_codegen /* emit && done */
1338 * Initializes a IA32 code generator.
1340 static void *ia32_cg_init(be_irg_t *birg) {
1341 ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa;
1342 ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg));
1344 cg->impl = &ia32_code_gen_if;
1345 cg->irg = birg->irg;
1346 cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024);
1347 cg->arch_env = birg->main_env->arch_env;
1350 cg->blk_sched = NULL;
1351 cg->fp_to_gp = NULL;
1352 cg->gp_to_fp = NULL;
1353 cg->fp_kind = isa->fp_kind;
1354 cg->used_fp = fp_none;
1355 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
1357 FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg");
1359 /* copy optimizations from isa for easier access */
1361 cg->arch = isa->arch;
1362 cg->opt_arch = isa->opt_arch;
1368 if (isa->name_obst) {
1369 obstack_free(isa->name_obst, NULL);
1370 obstack_init(isa->name_obst);
1374 cur_reg_set = cg->reg_set;
1376 ia32_irn_ops.cg = cg;
1378 return (arch_code_generator_t *)cg;
1383 /*****************************************************************
1384 * ____ _ _ _____ _____
1385 * | _ \ | | | | |_ _|/ ____| /\
1386 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
1387 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
1388 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
1389 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
1391 *****************************************************************/
1394 * Set output modes for GCC
1396 static const tarval_mode_info mo_integer = {
1403 * set the tarval output mode of all integer modes to decimal
1405 static void set_tarval_output_modes(void)
1409 for (i = get_irp_n_modes() - 1; i >= 0; --i) {
1410 ir_mode *mode = get_irp_mode(i);
1412 if (mode_is_int(mode))
1413 set_tarval_mode_output_option(mode, &mo_integer);
1419 * The template that generates a new ISA object.
1420 * Note that this template can be changed by command line
1423 static ia32_isa_t ia32_isa_template = {
1425 &ia32_isa_if, /* isa interface implementation */
1426 &ia32_gp_regs[REG_ESP], /* stack pointer register */
1427 &ia32_gp_regs[REG_EBP], /* base pointer register */
1428 -1, /* stack direction */
1429 NULL, /* main environment */
1431 NULL, /* 16bit register names */
1432 NULL, /* 8bit register names */
1436 IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */
1437 IA32_OPT_DOAM | /* optimize address mode default: on */
1438 IA32_OPT_LEA | /* optimize for LEAs default: on */
1439 IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */
1440 IA32_OPT_IMMOPS | /* operations can use immediates, default: on */
1441 IA32_OPT_EXTBB | /* use extended basic block scheduling, default: on */
1442 IA32_OPT_PUSHARGS), /* create pushs for function argument passing, default: on */
1443 arch_pentium_4, /* instruction architecture */
1444 arch_pentium_4, /* optimize for architecture */
1445 fp_sse2, /* use sse2 unit */
1446 NULL, /* current code generator */
1447 NULL, /* output file */
1449 NULL, /* name obstack */
1450 0 /* name obst size */
1455 * Initializes the backend ISA.
1457 static void *ia32_init(FILE *file_handle) {
1458 static int inited = 0;
1464 set_tarval_output_modes();
1466 isa = xmalloc(sizeof(*isa));
1467 memcpy(isa, &ia32_isa_template, sizeof(*isa));
1469 ia32_register_init(isa);
1470 ia32_create_opcodes();
1471 ia32_register_copy_attr_func();
1473 if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) ||
1474 (ARCH_AMD(isa->arch) && isa->arch < arch_athlon))
1475 /* no SSE2 for these cpu's */
1476 isa->fp_kind = fp_x87;
1478 if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) {
1479 /* Pentium 4 don't like inc and dec instructions */
1480 isa->opt &= ~IA32_OPT_INCDEC;
1483 isa->regs_16bit = pmap_create();
1484 isa->regs_8bit = pmap_create();
1485 isa->types = pmap_create();
1486 isa->tv_ent = pmap_create();
1487 isa->out = file_handle;
1488 isa->cpu = ia32_init_machine_description();
1490 ia32_build_16bit_reg_map(isa->regs_16bit);
1491 ia32_build_8bit_reg_map(isa->regs_8bit);
1493 /* patch register names of x87 registers */
1494 ia32_st_regs[0].name = "st";
1495 ia32_st_regs[1].name = "st(1)";
1496 ia32_st_regs[2].name = "st(2)";
1497 ia32_st_regs[3].name = "st(3)";
1498 ia32_st_regs[4].name = "st(4)";
1499 ia32_st_regs[5].name = "st(5)";
1500 ia32_st_regs[6].name = "st(6)";
1501 ia32_st_regs[7].name = "st(7)";
1504 isa->name_obst = xmalloc(sizeof(*isa->name_obst));
1505 obstack_init(isa->name_obst);
1508 ia32_handle_intrinsics();
1509 ia32_switch_section(isa->out, NO_SECTION);
1510 fprintf(isa->out, "\t.intel_syntax\n");
1512 /* needed for the debug support */
1513 ia32_switch_section(isa->out, SECTION_TEXT);
1514 fprintf(isa->out, ".Ltext0:\n");
1524 * Closes the output file and frees the ISA structure.
1526 static void ia32_done(void *self) {
1527 ia32_isa_t *isa = self;
1529 /* emit now all global declarations */
1530 ia32_gen_decls(isa->out, isa->arch_isa.main_env);
1532 pmap_destroy(isa->regs_16bit);
1533 pmap_destroy(isa->regs_8bit);
1534 pmap_destroy(isa->tv_ent);
1535 pmap_destroy(isa->types);
1538 obstack_free(isa->name_obst, NULL);
1546 * Return the number of register classes for this architecture.
1547 * We report always these:
1548 * - the general purpose registers
1549 * - the SSE floating point register set
1550 * - the virtual floating point registers
1552 static int ia32_get_n_reg_class(const void *self) {
1557 * Return the register class for index i.
1559 static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) {
1560 assert(i >= 0 && i < 3 && "Invalid ia32 register class requested.");
1562 return &ia32_reg_classes[CLASS_ia32_gp];
1564 return &ia32_reg_classes[CLASS_ia32_xmm];
1566 return &ia32_reg_classes[CLASS_ia32_vfp];
1570 * Get the register class which shall be used to store a value of a given mode.
1571 * @param self The this pointer.
1572 * @param mode The mode in question.
1573 * @return A register class which can hold values of the given mode.
1575 const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
1576 const ia32_isa_t *isa = self;
1577 if (mode_is_float(mode)) {
1578 return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp];
1581 return &ia32_reg_classes[CLASS_ia32_gp];
1585 * Get the ABI restrictions for procedure calls.
1586 * @param self The this pointer.
1587 * @param method_type The type of the method (procedure) in question.
1588 * @param abi The abi object to be modified
1590 static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1591 const ia32_isa_t *isa = self;
1594 unsigned cc = get_method_calling_convention(method_type);
1595 int n = get_method_n_params(method_type);
1598 int i, ignore_1, ignore_2;
1600 const arch_register_t *reg;
1601 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
1603 unsigned use_push = !IS_P6_ARCH(isa->opt_arch);
1605 /* set abi flags for calls */
1606 call_flags.bits.left_to_right = 0; /* always last arg first on stack */
1607 call_flags.bits.store_args_sequential = use_push;
1608 /* call_flags.bits.try_omit_fp not changed: can handle both settings */
1609 call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */
1610 call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
1612 /* set stack parameter passing style */
1613 be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
1615 /* collect the mode for each type */
1616 modes = alloca(n * sizeof(modes[0]));
1618 for (i = 0; i < n; i++) {
1619 tp = get_method_param_type(method_type, i);
1620 modes[i] = get_type_mode(tp);
1623 /* set register parameters */
1624 if (cc & cc_reg_param) {
1625 /* determine the number of parameters passed via registers */
1626 biggest_n = ia32_get_n_regparam_class(n, modes, &ignore_1, &ignore_2);
1628 /* loop over all parameters and set the register requirements */
1629 for (i = 0; i <= biggest_n; i++) {
1630 reg = ia32_get_RegParam_reg(n, modes, i, cc);
1631 assert(reg && "kaputt");
1632 be_abi_call_param_reg(abi, i, reg);
1639 /* set stack parameters */
1640 for (i = stack_idx; i < n; i++) {
1641 /* parameters on the stack are 32 bit aligned */
1642 be_abi_call_param_stack(abi, i, 4, 0, 0);
1646 /* set return registers */
1647 n = get_method_n_ress(method_type);
1649 assert(n <= 2 && "more than two results not supported");
1651 /* In case of 64bit returns, we will have two 32bit values */
1653 tp = get_method_res_type(method_type, 0);
1654 mode = get_type_mode(tp);
1656 assert(!mode_is_float(mode) && "two FP results not supported");
1658 tp = get_method_res_type(method_type, 1);
1659 mode = get_type_mode(tp);
1661 assert(!mode_is_float(mode) && "mixed INT, FP results not supported");
1663 be_abi_call_res_reg(abi, 0, &ia32_gp_regs[REG_EAX]);
1664 be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]);
1667 const arch_register_t *reg;
1669 tp = get_method_res_type(method_type, 0);
1670 assert(is_atomic_type(tp));
1671 mode = get_type_mode(tp);
1673 reg = mode_is_float(mode) ? &ia32_vfp_regs[REG_VF0] : &ia32_gp_regs[REG_EAX];
1675 be_abi_call_res_reg(abi, 0, reg);
1680 static const void *ia32_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1681 return &ia32_irn_ops;
1684 const arch_irn_handler_t ia32_irn_handler = {
1688 const arch_irn_handler_t *ia32_get_irn_handler(const void *self) {
1689 return &ia32_irn_handler;
1692 int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1693 return is_ia32_irn(irn) ? 1 : -1;
1697 * Initializes the code generator interface.
1699 static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self) {
1700 return &ia32_code_gen_if;
1704 * Returns the estimated execution time of an ia32 irn.
1706 static sched_timestep_t ia32_sched_exectime(void *env, const ir_node *irn) {
1707 const arch_env_t *arch_env = env;
1708 return is_ia32_irn(irn) ? ia32_get_op_estimated_cost(arch_get_irn_ops(arch_env, irn), irn) : 1;
1711 list_sched_selector_t ia32_sched_selector;
1714 * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded
1716 static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self, list_sched_selector_t *selector) {
1717 memcpy(&ia32_sched_selector, selector, sizeof(ia32_sched_selector));
1718 ia32_sched_selector.exectime = ia32_sched_exectime;
1719 ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule;
1720 return &ia32_sched_selector;
1723 static const ilp_sched_selector_t *ia32_get_ilp_sched_selector(const void *self) {
1728 * Returns the necessary byte alignment for storing a register of given class.
1730 static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1731 ir_mode *mode = arch_register_class_mode(cls);
1732 int bytes = get_mode_size_bytes(mode);
1734 if (mode_is_float(mode) && bytes > 8)
1739 static const be_execution_unit_t ***ia32_get_allowed_execution_units(const void *self, const ir_node *irn) {
1740 static const be_execution_unit_t *_allowed_units_BRANCH[] = {
1741 &ia32_execution_units_BRANCH[IA32_EXECUNIT_TP_BRANCH_BRANCH1],
1742 &ia32_execution_units_BRANCH[IA32_EXECUNIT_TP_BRANCH_BRANCH2],
1745 static const be_execution_unit_t *_allowed_units_ALU[] = {
1746 &ia32_execution_units_ALU[IA32_EXECUNIT_TP_ALU_ALU1],
1747 &ia32_execution_units_ALU[IA32_EXECUNIT_TP_ALU_ALU2],
1748 &ia32_execution_units_ALU[IA32_EXECUNIT_TP_ALU_ALU3],
1749 &ia32_execution_units_ALU[IA32_EXECUNIT_TP_ALU_ALU4],
1752 static const be_execution_unit_t *_allowed_units_DUMMY[] = {
1753 &ia32_execution_units_DUMMY[IA32_EXECUNIT_TP_DUMMY_DUMMY1],
1754 &ia32_execution_units_DUMMY[IA32_EXECUNIT_TP_DUMMY_DUMMY2],
1755 &ia32_execution_units_DUMMY[IA32_EXECUNIT_TP_DUMMY_DUMMY3],
1756 &ia32_execution_units_DUMMY[IA32_EXECUNIT_TP_DUMMY_DUMMY4],
1759 static const be_execution_unit_t **_units_callret[] = {
1760 _allowed_units_BRANCH,
1763 static const be_execution_unit_t **_units_other[] = {
1767 static const be_execution_unit_t **_units_dummy[] = {
1768 _allowed_units_DUMMY,
1771 const be_execution_unit_t ***ret;
1773 if (is_ia32_irn(irn)) {
1774 ret = get_ia32_exec_units(irn);
1776 else if (is_be_node(irn)) {
1777 if (be_is_Call(irn) || be_is_Return(irn)) {
1778 ret = _units_callret;
1780 else if (be_is_Barrier(irn)) {
1795 * Return the abstract ia32 machine.
1797 static const be_machine_t *ia32_get_machine(const void *self) {
1798 const ia32_isa_t *isa = self;
1803 * Allows or disallows the creation of Psi nodes for the given Phi nodes.
1804 * @return 1 if allowed, 0 otherwise
1806 static int ia32_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j)
1808 ir_node *cmp, *cmp_a, *phi;
1811 /* we don't want long long an floating point Psi */
1812 #define IS_BAD_PSI_MODE(mode) (mode_is_float(mode) || get_mode_size_bits(mode) > 32)
1814 if (get_irn_mode(sel) != mode_b)
1817 cmp = get_Proj_pred(sel);
1818 cmp_a = get_Cmp_left(cmp);
1819 mode = get_irn_mode(cmp_a);
1821 if (IS_BAD_PSI_MODE(mode))
1824 /* check the Phi nodes */
1825 for (phi = phi_list; phi; phi = get_irn_link(phi)) {
1826 ir_node *pred_i = get_irn_n(phi, i);
1827 ir_node *pred_j = get_irn_n(phi, j);
1828 ir_mode *mode_i = get_irn_mode(pred_i);
1829 ir_mode *mode_j = get_irn_mode(pred_j);
1831 if (IS_BAD_PSI_MODE(mode_i) || IS_BAD_PSI_MODE(mode_j))
1835 #undef IS_BAD_PSI_MODE
1840 static ia32_intrinsic_env_t intrinsic_env = {
1841 NULL, /**< the irg, these entities belong to */
1842 NULL, /**< entity for first div operand (move into FPU) */
1843 NULL, /**< entity for second div operand (move into FPU) */
1844 NULL, /**< entity for converts ll -> d */
1845 NULL, /**< entity for converts d -> ll */
1849 * Returns the libFirm configuration parameter for this backend.
1851 static const backend_params *ia32_get_libfirm_params(void) {
1852 static const opt_if_conv_info_t ifconv = {
1853 4, /* maxdepth, doesn't matter for Psi-conversion */
1854 ia32_is_psi_allowed /* allows or disallows Psi creation for given selector */
1856 static const arch_dep_params_t ad = {
1857 1, /* also use subs */
1858 4, /* maximum shifts */
1859 31, /* maximum shift amount */
1861 1, /* allow Mulhs */
1862 1, /* allow Mulus */
1863 32 /* Mulh allowed up to 32 bit */
1865 static backend_params p = {
1866 NULL, /* no additional opcodes */
1867 NULL, /* will be set later */
1868 1, /* need dword lowering */
1869 ia32_create_intrinsic_fkt,
1870 &intrinsic_env, /* context for ia32_create_intrinsic_fkt */
1871 NULL, /* will be set later */
1875 p.if_conv_info = &ifconv;
1880 /* instruction set architectures. */
1881 static const lc_opt_enum_int_items_t arch_items[] = {
1882 { "386", arch_i386, },
1883 { "486", arch_i486, },
1884 { "pentium", arch_pentium, },
1885 { "586", arch_pentium, },
1886 { "pentiumpro", arch_pentium_pro, },
1887 { "686", arch_pentium_pro, },
1888 { "pentiummmx", arch_pentium_mmx, },
1889 { "pentium2", arch_pentium_2, },
1890 { "p2", arch_pentium_2, },
1891 { "pentium3", arch_pentium_3, },
1892 { "p3", arch_pentium_3, },
1893 { "pentium4", arch_pentium_4, },
1894 { "p4", arch_pentium_4, },
1895 { "pentiumm", arch_pentium_m, },
1896 { "pm", arch_pentium_m, },
1897 { "core", arch_core, },
1899 { "athlon", arch_athlon, },
1900 { "athlon64", arch_athlon_64, },
1901 { "opteron", arch_opteron, },
1905 static lc_opt_enum_int_var_t arch_var = {
1906 &ia32_isa_template.arch, arch_items
1909 static lc_opt_enum_int_var_t opt_arch_var = {
1910 &ia32_isa_template.opt_arch, arch_items
1913 static const lc_opt_enum_int_items_t fp_unit_items[] = {
1915 { "sse2", fp_sse2 },
1919 static lc_opt_enum_int_var_t fp_unit_var = {
1920 &ia32_isa_template.fp_kind, fp_unit_items
1923 static const lc_opt_enum_int_items_t gas_items[] = {
1924 { "linux", ASM_LINUX_GAS },
1925 { "mingw", ASM_MINGW_GAS },
1929 static lc_opt_enum_int_var_t gas_var = {
1930 (int *)&asm_flavour, gas_items
1933 static const lc_opt_table_entry_t ia32_options[] = {
1934 LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var),
1935 LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var),
1936 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var),
1937 LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM),
1938 LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA),
1939 LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST),
1940 LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS),
1941 LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB),
1942 LC_OPT_ENT_NEGBIT("nopushargs", "do not create pushs for function arguments", &ia32_isa_template.opt, IA32_OPT_PUSHARGS),
1943 LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
1948 * Register command line options for the ia32 backend.
1952 * ia32-arch=arch create instruction for arch
1953 * ia32-opt=arch optimize for run on arch
1954 * ia32-fpunit=unit select floating point unit (x87 or SSE2)
1955 * ia32-incdec optimize for inc/dec
1956 * ia32-noaddrmode do not use address mode
1957 * ia32-nolea do not optimize for LEAs
1958 * ia32-noplacecnst do not place constants,
1959 * ia32-noimmop no operations with immediates
1960 * ia32-noextbb do not use extended basic block scheduling
1961 * ia32-nopushargs do not create pushs for function argument passing
1962 * ia32-gasmode set the GAS compatibility mode
1964 static void ia32_register_options(lc_opt_entry_t *ent)
1966 lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32");
1967 lc_opt_add_table(be_grp_ia32, ia32_options);
1969 #endif /* WITH_LIBCORE */
1971 const arch_isa_if_t ia32_isa_if = {
1974 ia32_get_n_reg_class,
1976 ia32_get_reg_class_for_mode,
1978 ia32_get_irn_handler,
1979 ia32_get_code_generator_if,
1980 ia32_get_list_sched_selector,
1981 ia32_get_ilp_sched_selector,
1982 ia32_get_reg_class_alignment,
1983 ia32_get_libfirm_params,
1984 ia32_get_allowed_execution_units,
1987 ia32_register_options