2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief The main arm backend driver file.
23 * @author Matthias Braun, Oliver Richter, Tobias Gneist
29 #include "lc_opts_enum.h"
31 #include "pseudo_irg.h"
38 #include "iroptimize.h"
47 #include "../bearch.h"
48 #include "../benode.h"
49 #include "../belower.h"
50 #include "../besched.h"
53 #include "../bemachine.h"
54 #include "../beilpsched.h"
55 #include "../bemodule.h"
57 #include "../bespillslots.h"
58 #include "../begnuas.h"
59 #include "../belistsched.h"
60 #include "../beflags.h"
62 #include "bearch_arm_t.h"
64 #include "arm_new_nodes.h"
65 #include "gen_arm_regalloc_if.h"
66 #include "arm_transform.h"
67 #include "arm_optimize.h"
68 #include "arm_emitter.h"
69 #include "arm_map_regs.h"
71 static arch_irn_class_t arm_classify(const ir_node *irn)
77 static ir_entity *arm_get_frame_entity(const ir_node *irn)
79 const arm_attr_t *attr = get_arm_attr_const(irn);
81 if (is_arm_FrameAddr(irn)) {
82 const arm_SymConst_attr_t *attr = get_irn_generic_attr_const(irn);
85 if (attr->is_load_store) {
86 const arm_load_store_attr_t *load_store_attr
87 = get_arm_load_store_attr_const(irn);
88 if (load_store_attr->is_frame_entity) {
89 return load_store_attr->entity;
95 static void arm_set_frame_entity(ir_node *irn, ir_entity *ent)
99 panic("arm_set_frame_entity() called. This should not happen.");
103 * This function is called by the generic backend to correct offsets for
104 * nodes accessing the stack.
106 static void arm_set_stack_bias(ir_node *irn, int bias)
108 if (is_arm_FrameAddr(irn)) {
109 arm_SymConst_attr_t *attr = get_irn_generic_attr(irn);
110 attr->fp_offset += bias;
112 arm_load_store_attr_t *attr = get_arm_load_store_attr(irn);
113 assert(attr->base.is_load_store);
114 attr->offset += bias;
118 static int arm_get_sp_bias(const ir_node *irn)
120 /* We don't have any nodes changing the stack pointer.
121 TODO: we probably want to support post-/pre increment/decrement later */
126 /* fill register allocator interface */
128 static const arch_irn_ops_t arm_irn_ops = {
131 arm_get_frame_entity,
132 arm_set_frame_entity,
135 NULL, /* get_inverse */
136 NULL, /* get_op_estimated_cost */
137 NULL, /* possible_memory_operand */
138 NULL, /* perform_memory_operand */
142 * Transforms the standard Firm graph into
145 static void arm_prepare_graph(void *self) {
146 arm_code_gen_t *cg = self;
148 /* transform nodes into assembler instructions */
149 arm_transform_graph(cg);
151 /* do local optimizations (mainly CSE) */
152 local_optimize_graph(cg->irg);
155 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
157 /* do code placement, to optimize the position of constants */
161 be_dump(cg->irg, "-place", dump_ir_block_graph_sched);
165 * Called immediately before emit phase.
167 static void arm_finish_irg(void *self)
169 arm_code_gen_t *cg = self;
171 /* do peephole optimizations and fix stack offsets */
172 arm_peephole_optimization(cg);
175 static ir_node *arm_flags_remat(ir_node *node, ir_node *after)
180 if (is_Block(after)) {
183 block = get_nodes_block(after);
185 copy = exact_copy(node);
186 set_nodes_block(copy, block);
187 sched_add_after(after, copy);
191 static void arm_before_ra(void *self)
193 arm_code_gen_t *cg = self;
195 be_sched_fix_flags(cg->birg, &arm_reg_classes[CLASS_arm_flags],
199 static void transform_Reload(ir_node *node)
201 ir_graph *irg = get_irn_irg(node);
202 ir_node *block = get_nodes_block(node);
203 dbg_info *dbgi = get_irn_dbg_info(node);
204 ir_node *ptr = get_irg_frame(irg);
205 ir_node *mem = get_irn_n(node, be_pos_Reload_mem);
206 ir_mode *mode = get_irn_mode(node);
207 ir_entity *entity = be_get_frame_entity(node);
208 const arch_register_t *reg;
212 ir_node *sched_point = sched_prev(node);
214 load = new_bd_arm_Ldr(dbgi, block, ptr, mem, mode, entity, false, 0, true);
215 sched_add_after(sched_point, load);
218 proj = new_rd_Proj(dbgi, block, load, mode, pn_arm_Ldr_res);
220 reg = arch_get_irn_register(node);
221 arch_set_irn_register(proj, reg);
223 exchange(node, proj);
226 static void transform_Spill(ir_node *node)
228 ir_graph *irg = get_irn_irg(node);
229 ir_node *block = get_nodes_block(node);
230 dbg_info *dbgi = get_irn_dbg_info(node);
231 ir_node *ptr = get_irg_frame(irg);
232 ir_node *mem = new_NoMem();
233 ir_node *val = get_irn_n(node, be_pos_Spill_val);
234 ir_mode *mode = get_irn_mode(val);
235 ir_entity *entity = be_get_frame_entity(node);
236 ir_node *sched_point;
239 sched_point = sched_prev(node);
240 store = new_bd_arm_Str(dbgi, block, ptr, val, mem, mode, entity, false, 0,
244 sched_add_after(sched_point, store);
246 exchange(node, store);
249 static void arm_after_ra_walker(ir_node *block, void *data)
251 ir_node *node, *prev;
254 for (node = sched_last(block); !sched_is_begin(node); node = prev) {
255 prev = sched_prev(node);
257 if (be_is_Reload(node)) {
258 transform_Reload(node);
259 } else if (be_is_Spill(node)) {
260 transform_Spill(node);
265 static void arm_after_ra(void *self)
267 arm_code_gen_t *cg = self;
268 be_coalesce_spillslots(cg->birg);
270 irg_block_walk_graph(cg->irg, NULL, arm_after_ra_walker, NULL);
274 * Emits the code, closes the output file and frees
275 * the code generator interface.
277 static void arm_emit_and_done(void *self) {
278 arm_code_gen_t *cg = self;
279 ir_graph *irg = cg->irg;
281 arm_gen_routine(cg, irg);
283 /* de-allocate code generator */
284 del_set(cg->reg_set);
289 * Move a double floating point value into an integer register.
290 * Place the move operation into block bl.
292 * Handle some special cases here:
293 * 1.) A constant: simply split into two
294 * 2.) A load: simply split into two
296 static ir_node *convert_dbl_to_int(ir_node *bl, ir_node *arg, ir_node *mem,
297 ir_node **resH, ir_node **resL)
300 tarval *tv = get_Const_tarval(arg);
303 /* get the upper 32 bits */
304 v = get_tarval_sub_bits(tv, 7);
305 v = (v << 8) | get_tarval_sub_bits(tv, 6);
306 v = (v << 8) | get_tarval_sub_bits(tv, 5);
307 v = (v << 8) | get_tarval_sub_bits(tv, 4);
308 *resH = new_Const_long(mode_Is, v);
310 /* get the lower 32 bits */
311 v = get_tarval_sub_bits(tv, 3);
312 v = (v << 8) | get_tarval_sub_bits(tv, 2);
313 v = (v << 8) | get_tarval_sub_bits(tv, 1);
314 v = (v << 8) | get_tarval_sub_bits(tv, 0);
315 *resL = new_Const_long(mode_Is, v);
316 } else if (is_Load(skip_Proj(arg))) {
317 /* FIXME: handling of low/high depends on LE/BE here */
318 panic("Unimplemented convert_dbl_to_int() case");
323 conv = new_bd_arm_fpaDbl2GP(NULL, bl, arg, mem);
325 *resL = new_r_Proj(bl, conv, mode_Is, pn_arm_fpaDbl2GP_low);
326 *resH = new_r_Proj(bl, conv, mode_Is, pn_arm_fpaDbl2GP_high);
327 mem = new_r_Proj(bl, conv, mode_M, pn_arm_fpaDbl2GP_M);
333 * Move a single floating point value into an integer register.
334 * Place the move operation into block bl.
336 * Handle some special cases here:
337 * 1.) A constant: simply move
338 * 2.) A load: simply load
340 static ir_node *convert_sng_to_int(ir_node *bl, ir_node *arg)
345 tarval *tv = get_Const_tarval(arg);
348 /* get the lower 32 bits */
349 v = get_tarval_sub_bits(tv, 3);
350 v = (v << 8) | get_tarval_sub_bits(tv, 2);
351 v = (v << 8) | get_tarval_sub_bits(tv, 1);
352 v = (v << 8) | get_tarval_sub_bits(tv, 0);
353 return new_Const_long(mode_Is, v);
354 } else if (is_Load(skip_Proj(arg))) {
357 load = skip_Proj(arg);
359 panic("Unimplemented convert_sng_to_int() case");
363 * Convert the arguments of a call to support the
364 * ARM calling convention of general purpose AND floating
367 static void handle_calls(ir_node *call, void *env)
369 arm_code_gen_t *cg = env;
370 int i, j, n, size, idx, flag, n_param, n_res, first_variadic;
371 ir_type *mtp, *new_mtd, *new_tp[5];
372 ir_node *new_in[5], **in;
378 /* check, if we need conversions */
379 n = get_Call_n_params(call);
380 mtp = get_Call_type(call);
381 assert(get_method_n_params(mtp) == n);
383 /* it's always enough to handle the first 4 parameters */
386 flag = size = idx = 0;
387 bl = get_nodes_block(call);
388 for (i = 0; i < n; ++i) {
389 ir_type *param_tp = get_method_param_type(mtp, i);
391 if (is_compound_type(param_tp)) {
392 /* an aggregate parameter: bad case */
396 /* a primitive parameter */
397 ir_mode *mode = get_type_mode(param_tp);
399 if (mode_is_float(mode)) {
400 if (get_mode_size_bits(mode) > 32) {
401 ir_node *mem = get_Call_mem(call);
403 /* Beware: ARM wants the high part first */
405 new_tp[idx] = cg->int_tp;
406 new_tp[idx+1] = cg->int_tp;
407 mem = convert_dbl_to_int(bl, get_Call_param(call, i), mem, &new_in[idx], &new_in[idx+1]);
409 set_Call_mem(call, mem);
413 new_tp[idx] = cg->int_tp;
414 new_in[idx] = convert_sng_to_int(bl, get_Call_param(call, i));
421 new_tp[idx] = param_tp;
422 new_in[idx] = get_Call_param(call, i);
431 /* if flag is NOT set, no need to translate the method type */
435 /* construct a new method type */
437 n_param = get_method_n_params(mtp) - n + idx;
438 n_res = get_method_n_ress(mtp);
439 new_mtd = new_d_type_method(get_type_ident(mtp), n_param, n_res, get_type_dbg_info(mtp));
441 for (i = 0; i < idx; ++i)
442 set_method_param_type(new_mtd, i, new_tp[i]);
443 for (i = n, j = idx; i < get_method_n_params(mtp); ++i)
444 set_method_param_type(new_mtd, j++, get_method_param_type(mtp, i));
445 for (i = 0; i < n_res; ++i)
446 set_method_res_type(new_mtd, i, get_method_res_type(mtp, i));
448 set_method_calling_convention(new_mtd, get_method_calling_convention(mtp));
449 first_variadic = get_method_first_variadic_param_index(mtp);
450 if (first_variadic >= 0)
451 set_method_first_variadic_param_index(new_mtd, first_variadic);
453 if (is_lowered_type(mtp)) {
454 mtp = get_associated_type(mtp);
456 set_lowered_type(mtp, new_mtd);
458 set_Call_type(call, new_mtd);
460 /* calculate new in array of the Call */
461 NEW_ARR_A(ir_node *, in, n_param + 2);
462 for (i = 0; i < idx; ++i)
463 in[2 + i] = new_in[i];
464 for (i = n, j = idx; i < get_method_n_params(mtp); ++i)
465 in[2 + j++] = get_Call_param(call, i);
467 in[0] = get_Call_mem(call);
468 in[1] = get_Call_ptr(call);
470 /* finally, change the call inputs */
471 set_irn_in(call, n_param + 2, in);
475 * Handle graph transformations before the abi converter does its work.
477 static void arm_before_abi(void *self) {
478 arm_code_gen_t *cg = self;
480 irg_walk_graph(cg->irg, NULL, handle_calls, cg);
484 static void *arm_cg_init(be_irg_t *birg);
486 static const arch_code_generator_if_t arm_code_gen_if = {
488 NULL, /* get_pic_base */
489 arm_before_abi, /* before abi introduce */
492 arm_before_ra, /* before register allocation hook */
499 * Initializes the code generator.
501 static void *arm_cg_init(be_irg_t *birg) {
502 static ir_type *int_tp = NULL;
503 arm_isa_t *isa = (arm_isa_t *)birg->main_env->arch_env;
507 /* create an integer type with machine size */
508 int_tp = new_type_primitive(new_id_from_chars("int", 3), mode_Is);
511 cg = XMALLOC(arm_code_gen_t);
512 cg->impl = &arm_code_gen_if;
514 cg->reg_set = new_set(arm_cmp_irn_reg_assoc, 1024);
518 cg->have_fp_insn = 0;
519 cg->unknown_gp = NULL;
520 cg->unknown_fpa = NULL;
521 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
523 FIRM_DBG_REGISTER(cg->mod, "firm.be.arm.cg");
525 /* enter the current code generator */
528 return (arch_code_generator_t *)cg;
533 * Maps all intrinsic calls that the backend support
534 * and map all instructions the backend did not support
537 static void arm_handle_intrinsics(void) {
538 ir_type *tp, *int_tp, *uint_tp;
542 runtime_rt rt_iDiv, rt_uDiv, rt_iMod, rt_uMod;
544 #define ID(x) new_id_from_chars(x, sizeof(x)-1)
546 int_tp = new_type_primitive(ID("int"), mode_Is);
547 uint_tp = new_type_primitive(ID("uint"), mode_Iu);
549 /* ARM has neither a signed div instruction ... */
551 i_instr_record *map_Div = &records[n_records++].i_instr;
553 tp = new_type_method(ID("rt_iDiv"), 2, 1);
554 set_method_param_type(tp, 0, int_tp);
555 set_method_param_type(tp, 1, int_tp);
556 set_method_res_type(tp, 0, int_tp);
558 rt_iDiv.ent = new_entity(get_glob_type(), ID("__divsi3"), tp);
559 set_entity_ld_ident(rt_iDiv.ent, ID("__divsi3"));
560 rt_iDiv.mode = mode_T;
561 rt_iDiv.res_mode = mode_Is;
562 rt_iDiv.mem_proj_nr = pn_Div_M;
563 rt_iDiv.regular_proj_nr = pn_Div_X_regular;
564 rt_iDiv.exc_proj_nr = pn_Div_X_except;
565 rt_iDiv.exc_mem_proj_nr = pn_Div_M;
566 rt_iDiv.res_proj_nr = pn_Div_res;
568 set_entity_visibility(rt_iDiv.ent, visibility_external_allocated);
570 map_Div->kind = INTRINSIC_INSTR;
571 map_Div->op = op_Div;
572 map_Div->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
573 map_Div->ctx = &rt_iDiv;
575 /* ... nor an unsigned div instruction ... */
577 i_instr_record *map_Div = &records[n_records++].i_instr;
579 tp = new_type_method(ID("rt_uDiv"), 2, 1);
580 set_method_param_type(tp, 0, uint_tp);
581 set_method_param_type(tp, 1, uint_tp);
582 set_method_res_type(tp, 0, uint_tp);
584 rt_uDiv.ent = new_entity(get_glob_type(), ID("__udivsi3"), tp);
585 set_entity_ld_ident(rt_uDiv.ent, ID("__udivsi3"));
586 rt_uDiv.mode = mode_T;
587 rt_uDiv.res_mode = mode_Iu;
588 rt_uDiv.mem_proj_nr = pn_Div_M;
589 rt_uDiv.regular_proj_nr = pn_Div_X_regular;
590 rt_uDiv.exc_proj_nr = pn_Div_X_except;
591 rt_uDiv.exc_mem_proj_nr = pn_Div_M;
592 rt_uDiv.res_proj_nr = pn_Div_res;
594 set_entity_visibility(rt_uDiv.ent, visibility_external_allocated);
596 map_Div->kind = INTRINSIC_INSTR;
597 map_Div->op = op_Div;
598 map_Div->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
599 map_Div->ctx = &rt_uDiv;
601 /* ... nor a signed mod instruction ... */
603 i_instr_record *map_Mod = &records[n_records++].i_instr;
605 tp = new_type_method(ID("rt_iMod"), 2, 1);
606 set_method_param_type(tp, 0, int_tp);
607 set_method_param_type(tp, 1, int_tp);
608 set_method_res_type(tp, 0, int_tp);
610 rt_iMod.ent = new_entity(get_glob_type(), ID("__modsi3"), tp);
611 set_entity_ld_ident(rt_iMod.ent, ID("__modsi3"));
612 rt_iMod.mode = mode_T;
613 rt_iMod.res_mode = mode_Is;
614 rt_iMod.mem_proj_nr = pn_Mod_M;
615 rt_iMod.regular_proj_nr = pn_Mod_X_regular;
616 rt_iMod.exc_proj_nr = pn_Mod_X_except;
617 rt_iMod.exc_mem_proj_nr = pn_Mod_M;
618 rt_iMod.res_proj_nr = pn_Mod_res;
620 set_entity_visibility(rt_iMod.ent, visibility_external_allocated);
622 map_Mod->kind = INTRINSIC_INSTR;
623 map_Mod->op = op_Mod;
624 map_Mod->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
625 map_Mod->ctx = &rt_iMod;
627 /* ... nor an unsigned mod. */
629 i_instr_record *map_Mod = &records[n_records++].i_instr;
631 tp = new_type_method(ID("rt_uMod"), 2, 1);
632 set_method_param_type(tp, 0, uint_tp);
633 set_method_param_type(tp, 1, uint_tp);
634 set_method_res_type(tp, 0, uint_tp);
636 rt_uMod.ent = new_entity(get_glob_type(), ID("__umodsi3"), tp);
637 set_entity_ld_ident(rt_uMod.ent, ID("__umodsi3"));
638 rt_uMod.mode = mode_T;
639 rt_uMod.res_mode = mode_Iu;
640 rt_uMod.mem_proj_nr = pn_Mod_M;
641 rt_uMod.regular_proj_nr = pn_Mod_X_regular;
642 rt_uMod.exc_proj_nr = pn_Mod_X_except;
643 rt_uMod.exc_mem_proj_nr = pn_Mod_M;
644 rt_uMod.res_proj_nr = pn_Mod_res;
646 set_entity_visibility(rt_uMod.ent, visibility_external_allocated);
648 map_Mod->kind = INTRINSIC_INSTR;
649 map_Mod->op = op_Mod;
650 map_Mod->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
651 map_Mod->ctx = &rt_uMod;
655 lower_intrinsics(records, n_records, /*part_block_used=*/0);
658 /*****************************************************************
659 * ____ _ _ _____ _____
660 * | _ \ | | | | |_ _|/ ____| /\
661 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
662 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
663 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
664 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
666 *****************************************************************/
668 static arm_isa_t arm_isa_template = {
670 &arm_isa_if, /* isa interface */
671 &arm_gp_regs[REG_SP], /* stack pointer */
672 &arm_gp_regs[REG_R11], /* base pointer */
673 &arm_reg_classes[CLASS_arm_gp], /* static link pointer class */
674 -1, /* stack direction */
675 2, /* power of two stack alignment for calls, 2^2 == 4 */
676 NULL, /* main environment */
678 5, /* reload costs */
680 0, /* use generic register names instead of SP, LR, PC */
681 ARM_FPU_ARCH_FPE, /* FPU architecture */
682 NULL, /* current code generator */
686 * Initializes the backend ISA and opens the output file.
688 static arch_env_t *arm_init(FILE *file_handle) {
689 static int inited = 0;
695 isa = XMALLOC(arm_isa_t);
696 memcpy(isa, &arm_isa_template, sizeof(*isa));
701 be_emit_init(file_handle);
703 arm_create_opcodes(&arm_irn_ops);
704 arm_handle_intrinsics();
706 be_gas_emit_types = false;
708 /* needed for the debug support */
709 be_gas_emit_switch_section(GAS_SECTION_TEXT);
710 be_emit_cstring(".Ltext0:\n");
711 be_emit_write_line();
713 /* we mark referenced global entities, so we can only emit those which
714 * are actually referenced. (Note: you mustn't use the type visited flag
715 * elsewhere in the backend)
717 inc_master_type_visited();
720 return &isa->arch_env;
726 * Closes the output file and frees the ISA structure.
728 static void arm_done(void *self) {
729 arm_isa_t *isa = self;
731 be_gas_emit_decls(isa->arch_env.main_env, 1);
739 * Report the number of register classes.
740 * If we don't have fp instructions, report only GP
741 * here to speed up register allocation (and makes dumps
742 * smaller and more readable).
744 static unsigned arm_get_n_reg_class(void) {
749 * Return the register class with requested index.
751 static const arch_register_class_t *arm_get_reg_class(unsigned i) {
752 assert(i < N_CLASSES);
753 return &arm_reg_classes[i];
757 * Get the register class which shall be used to store a value of a given mode.
758 * @param self The this pointer.
759 * @param mode The mode in question.
760 * @return A register class which can hold values of the given mode.
762 const arch_register_class_t *arm_get_reg_class_for_mode(const ir_mode *mode) {
763 if (mode_is_float(mode))
764 return &arm_reg_classes[CLASS_arm_fpa];
766 return &arm_reg_classes[CLASS_arm_gp];
770 * Produces the type which sits between the stack args and the locals on the stack.
771 * it will contain the return address and space to store the old base pointer.
772 * @return The Firm type modeling the ABI between type.
774 static ir_type *arm_get_between_type(void *self)
776 static ir_type *between_type = NULL;
779 if (between_type == NULL) {
780 between_type = new_type_class(new_id_from_str("arm_between_type"));
781 set_type_size_bytes(between_type, 0);
789 be_abi_call_flags_bits_t flags;
790 const arch_env_t *arch_env;
794 static void *arm_abi_init(const be_abi_call_t *call, const arch_env_t *arch_env, ir_graph *irg)
796 arm_abi_env_t *env = XMALLOC(arm_abi_env_t);
797 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
798 env->flags = fl.bits;
800 env->arch_env = arch_env;
805 * Generate the routine prologue.
807 * @param self The callback object.
808 * @param mem A pointer to the mem node. Update this if you define new memory.
809 * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
810 * @param stack_bias Points to the current stack bias, can be modified if needed.
812 * @return The register which shall be used as a stack frame base.
814 * All nodes which define registers in @p reg_map must keep @p reg_map current.
816 static const arch_register_t *arm_abi_prologue(void *self, ir_node **mem, pmap *reg_map, int *stack_bias) {
817 arm_abi_env_t *env = self;
821 arch_register_class_t *gp;
823 ir_node *fp, *ip, *lr, *pc;
824 ir_node *sp = be_abi_reg_map_get(reg_map, env->arch_env->sp);
828 if (env->flags.try_omit_fp)
829 return env->arch_env->sp;
831 fp = be_abi_reg_map_get(reg_map, env->arch_env->bp);
832 ip = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_R12]);
833 lr = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_LR]);
834 pc = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_PC]);
836 gp = &arm_reg_classes[CLASS_arm_gp];
838 block = get_irg_start_block(irg);
840 /* mark bp register as ignore */
841 be_set_constr_single_reg_out(get_Proj_pred(fp),
842 get_Proj_proj(fp), env->arch_env->bp,
843 arch_register_req_type_ignore);
845 /* copy SP to IP (so we can spill it */
846 ip = be_new_Copy(gp, block, sp);
847 be_set_constr_single_reg_out(ip, 0, &arm_gp_regs[REG_R12], 0);
850 store = new_bd_arm_StoreStackM4Inc(NULL, block, sp, fp, ip, lr, pc, *mem);
852 sp = new_r_Proj(block, store, env->arch_env->sp->reg_class->mode, pn_arm_StoreStackM4Inc_ptr);
853 arch_set_irn_register(sp, env->arch_env->sp);
854 *mem = new_r_Proj(block, store, mode_M, pn_arm_StoreStackM4Inc_M);
856 /* frame pointer is ip-4 (because ip is our old sp value) */
857 fp = new_bd_arm_Sub_imm(NULL, block, ip, 4, 0);
858 arch_set_irn_register(fp, env->arch_env->bp);
860 /* beware: we change the fp but the StoreStackM4Inc above wants the old
861 * fp value. We are not allowed to spill or anything in the prolog, so we
862 * have to enforce some order here. (scheduler/regalloc are too stupid
863 * to extract this order from register requirements) */
864 add_irn_dep(fp, store);
866 fp = be_new_Copy(gp, block, fp); // XXX Gammelfix: only be_ have custom register requirements
867 be_set_constr_single_reg_out(fp, 0, env->arch_env->bp,
868 arch_register_req_type_ignore);
869 arch_set_irn_register(fp, env->arch_env->bp);
871 be_abi_reg_map_set(reg_map, env->arch_env->bp, fp);
872 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_R12], ip);
873 be_abi_reg_map_set(reg_map, env->arch_env->sp, sp);
874 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_LR], lr);
875 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_PC], pc);
877 return env->arch_env->bp;
881 * Builds the ARM epilogue
883 static void arm_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map)
885 arm_abi_env_t *env = self;
886 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->arch_env->sp);
887 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->arch_env->bp);
888 ir_node *curr_pc = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_PC]);
889 ir_node *curr_lr = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_LR]);
891 // TODO: Activate Omit fp in epilogue
892 if (env->flags.try_omit_fp) {
893 ir_node *incsp = be_new_IncSP(env->arch_env->sp, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK, 0);
898 load_node = new_bd_arm_LoadStackM3Epilogue(NULL, bl, curr_bp, *mem);
900 curr_bp = new_r_Proj(bl, load_node, env->arch_env->bp->reg_class->mode, pn_arm_LoadStackM3Epilogue_res0);
901 curr_sp = new_r_Proj(bl, load_node, env->arch_env->sp->reg_class->mode, pn_arm_LoadStackM3Epilogue_res1);
902 curr_pc = new_r_Proj(bl, load_node, mode_Iu, pn_arm_LoadStackM3Epilogue_res2);
903 *mem = new_r_Proj(bl, load_node, mode_M, pn_arm_LoadStackM3Epilogue_M);
904 arch_set_irn_register(curr_bp, env->arch_env->bp);
905 arch_set_irn_register(curr_sp, env->arch_env->sp);
906 arch_set_irn_register(curr_pc, &arm_gp_regs[REG_PC]);
908 be_abi_reg_map_set(reg_map, env->arch_env->sp, curr_sp);
909 be_abi_reg_map_set(reg_map, env->arch_env->bp, curr_bp);
910 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_LR], curr_lr);
911 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_PC], curr_pc);
914 static const be_abi_callbacks_t arm_abi_callbacks = {
917 arm_get_between_type,
924 * Get the ABI restrictions for procedure calls.
925 * @param self The this pointer.
926 * @param method_type The type of the method (procedure) in question.
927 * @param abi The abi object to be modified
929 void arm_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi)
934 int n = get_method_n_params(method_type);
935 be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
938 /* set abi flags for calls */
939 call_flags.bits.left_to_right = 0;
940 call_flags.bits.store_args_sequential = 0;
941 /* call_flags.bits.try_omit_fp don't change this we can handle both */
942 call_flags.bits.fp_free = 0;
943 call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */
945 /* set stack parameter passing style */
946 be_abi_call_set_flags(abi, call_flags, &arm_abi_callbacks);
948 for (i = 0; i < n; i++) {
949 /* reg = get reg for param i; */
950 /* be_abi_call_param_reg(abi, i, reg); */
952 be_abi_call_param_reg(abi, i, arm_get_RegParam_reg(i));
954 tp = get_method_param_type(method_type, i);
955 mode = get_type_mode(tp);
956 be_abi_call_param_stack(abi, i, mode, 4, 0, 0);
960 /* set return registers */
961 n = get_method_n_ress(method_type);
963 assert(n <= 2 && "more than two results not supported");
965 /* In case of 64bit returns, we will have two 32bit values */
967 tp = get_method_res_type(method_type, 0);
968 mode = get_type_mode(tp);
970 assert(!mode_is_float(mode) && "two FP results not supported");
972 tp = get_method_res_type(method_type, 1);
973 mode = get_type_mode(tp);
975 assert(!mode_is_float(mode) && "mixed INT, FP results not supported");
977 be_abi_call_res_reg(abi, 0, &arm_gp_regs[REG_R0]);
978 be_abi_call_res_reg(abi, 1, &arm_gp_regs[REG_R1]);
980 const arch_register_t *reg;
982 tp = get_method_res_type(method_type, 0);
983 assert(is_atomic_type(tp));
984 mode = get_type_mode(tp);
986 reg = mode_is_float(mode) ? &arm_fpa_regs[REG_F0] : &arm_gp_regs[REG_R0];
987 be_abi_call_res_reg(abi, 0, reg);
991 int arm_to_appear_in_schedule(void *block_env, const ir_node *irn)
1001 * Initializes the code generator interface.
1003 static const arch_code_generator_if_t *arm_get_code_generator_if(void *self)
1006 return &arm_code_gen_if;
1009 list_sched_selector_t arm_sched_selector;
1012 * Returns the reg_pressure scheduler with to_appear_in_schedule() over\loaded
1014 static const list_sched_selector_t *arm_get_list_sched_selector(const void *self, list_sched_selector_t *selector)
1017 memcpy(&arm_sched_selector, selector, sizeof(arm_sched_selector));
1018 /* arm_sched_selector.exectime = arm_sched_exectime; */
1019 arm_sched_selector.to_appear_in_schedule = arm_to_appear_in_schedule;
1020 return &arm_sched_selector;
1024 static const ilp_sched_selector_t *arm_get_ilp_sched_selector(const void *self) {
1030 * Returns the necessary byte alignment for storing a register of given class.
1032 static int arm_get_reg_class_alignment(const arch_register_class_t *cls)
1035 /* ARM is a 32 bit CPU, no need for other alignment */
1039 static const be_execution_unit_t ***arm_get_allowed_execution_units(const ir_node *irn) {
1042 panic("Unimplemented arm_get_allowed_execution_units()");
1045 static const be_machine_t *arm_get_machine(const void *self) {
1048 panic("Unimplemented arm_get_machine()");
1052 * Return irp irgs in the desired order.
1054 static ir_graph **arm_get_irg_list(const void *self, ir_graph ***irg_list) {
1061 * Allows or disallows the creation of Psi nodes for the given Phi nodes.
1062 * @return 1 if allowed, 0 otherwise
1064 static int arm_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j) {
1065 ir_node *cmp, *cmp_a, *phi;
1069 /* currently Psi support is not implemented */
1072 /* we don't want long long Psi */
1073 #define IS_BAD_PSI_MODE(mode) (!mode_is_float(mode) && get_mode_size_bits(mode) > 32)
1075 if (get_irn_mode(sel) != mode_b)
1078 cmp = get_Proj_pred(sel);
1079 cmp_a = get_Cmp_left(cmp);
1080 mode = get_irn_mode(cmp_a);
1082 if (IS_BAD_PSI_MODE(mode))
1085 /* check the Phi nodes */
1086 for (phi = phi_list; phi; phi = get_irn_link(phi)) {
1087 ir_node *pred_i = get_irn_n(phi, i);
1088 ir_node *pred_j = get_irn_n(phi, j);
1089 ir_mode *mode_i = get_irn_mode(pred_i);
1090 ir_mode *mode_j = get_irn_mode(pred_j);
1092 if (IS_BAD_PSI_MODE(mode_i) || IS_BAD_PSI_MODE(mode_j))
1096 #undef IS_BAD_PSI_MODE
1101 static asm_constraint_flags_t arm_parse_asm_constraint(const char **c)
1103 /* asm not supported */
1105 return ASM_CONSTRAINT_FLAG_INVALID;
1108 static int arm_is_valid_clobber(const char *clobber)
1115 * Returns the libFirm configuration parameter for this backend.
1117 static const backend_params *arm_get_libfirm_params(void) {
1118 static const ir_settings_if_conv_t ifconv = {
1119 4, /* maxdepth, doesn't matter for Psi-conversion */
1120 arm_is_psi_allowed /* allows or disallows Psi creation for given selector */
1122 static ir_settings_arch_dep_t ad = {
1124 1, /* Muls are fast enough on ARM but ... */
1125 31, /* ... one shift would be possible better */
1126 NULL, /* no evaluator function */
1127 0, /* SMUL is needed, only in Arch M */
1128 0, /* UMUL is needed, only in Arch M */
1129 32, /* SMUL & UMUL available for 32 bit */
1131 static backend_params p = {
1132 1, /* need dword lowering */
1133 0, /* don't support inline assembler yet */
1134 NULL, /* will be set later */
1135 NULL, /* but yet no creator function */
1136 NULL, /* context for create_intrinsic_fkt */
1137 NULL, /* ifconv_info will be set below */
1138 NULL, /* float arithmetic mode (TODO) */
1139 0, /* no trampoline support: size 0 */
1140 0, /* no trampoline support: align 0 */
1141 NULL, /* no trampoline support: no trampoline builder */
1142 4 /* alignment of stack parameter */
1146 p.if_conv_info = &ifconv;
1150 /* fpu set architectures. */
1151 static const lc_opt_enum_int_items_t arm_fpu_items[] = {
1152 { "softfloat", ARM_FPU_ARCH_SOFTFLOAT },
1153 { "fpe", ARM_FPU_ARCH_FPE },
1154 { "fpa", ARM_FPU_ARCH_FPA },
1155 { "vfp1xd", ARM_FPU_ARCH_VFP_V1xD },
1156 { "vfp1", ARM_FPU_ARCH_VFP_V1 },
1157 { "vfp2", ARM_FPU_ARCH_VFP_V2 },
1161 static lc_opt_enum_int_var_t arch_fpu_var = {
1162 &arm_isa_template.fpu_arch, arm_fpu_items
1165 static const lc_opt_table_entry_t arm_options[] = {
1166 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &arch_fpu_var),
1167 LC_OPT_ENT_BOOL("gen_reg_names", "use generic register names", &arm_isa_template.gen_reg_names),
1171 const arch_isa_if_t arm_isa_if = {
1174 NULL, /* handle_intrinsics */
1175 arm_get_n_reg_class,
1177 arm_get_reg_class_for_mode,
1179 arm_get_code_generator_if,
1180 arm_get_list_sched_selector,
1181 arm_get_ilp_sched_selector,
1182 arm_get_reg_class_alignment,
1183 arm_get_libfirm_params,
1184 arm_get_allowed_execution_units,
1187 NULL, /* mark remat */
1188 arm_parse_asm_constraint,
1189 arm_is_valid_clobber
1192 void be_init_arch_arm(void)
1194 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1195 lc_opt_entry_t *arm_grp = lc_opt_get_grp(be_grp, "arm");
1197 lc_opt_add_table(arm_grp, arm_options);
1199 be_register_isa_if("arm", &arm_isa_if);
1201 arm_init_transform();
1205 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_arm);