2 * Copyright (C) 1995-2010 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 codegenerator (transform FIRM into arm FIRM)
23 * @author Matthias Braun, Oliver Richter, Tobias Gneist, Michael Beck
29 #include "irgraph_t.h"
40 #include "../benode.h"
42 #include "../beutil.h"
43 #include "../betranshlp.h"
44 #include "../beabihelper.h"
47 #include "bearch_arm_t.h"
48 #include "arm_nodes_attr.h"
49 #include "arm_transform.h"
50 #include "arm_optimize.h"
51 #include "arm_new_nodes.h"
52 #include "arm_map_regs.h"
53 #include "arm_cconv.h"
55 #include "gen_arm_regalloc_if.h"
59 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
61 static const arch_register_t *sp_reg = &arm_registers[REG_SP];
62 static ir_mode *mode_gp;
63 static ir_mode *mode_fp;
64 static beabi_helper_env_t *abihelper;
65 static be_stackorder_t *stackorder;
66 static calling_convention_t *cconv = NULL;
67 static arm_isa_t *isa;
69 static pmap *node_to_stack;
71 static const arch_register_t *const callee_saves[] = {
72 &arm_registers[REG_R4],
73 &arm_registers[REG_R5],
74 &arm_registers[REG_R6],
75 &arm_registers[REG_R7],
76 &arm_registers[REG_R8],
77 &arm_registers[REG_R9],
78 &arm_registers[REG_R10],
79 &arm_registers[REG_R11],
80 &arm_registers[REG_LR],
83 static const arch_register_t *const caller_saves[] = {
84 &arm_registers[REG_R0],
85 &arm_registers[REG_R1],
86 &arm_registers[REG_R2],
87 &arm_registers[REG_R3],
88 &arm_registers[REG_LR],
90 &arm_registers[REG_F0],
91 &arm_registers[REG_F1],
92 &arm_registers[REG_F2],
93 &arm_registers[REG_F3],
94 &arm_registers[REG_F4],
95 &arm_registers[REG_F5],
96 &arm_registers[REG_F6],
97 &arm_registers[REG_F7],
100 static bool mode_needs_gp_reg(ir_mode *mode)
102 return mode_is_int(mode) || mode_is_reference(mode);
106 * create firm graph for a constant
108 static ir_node *create_const_graph_value(dbg_info *dbgi, ir_node *block,
115 /* We only have 8 bit immediates. So we possibly have to combine several
116 * operations to construct the desired value.
118 * we can either create the value by adding bits to 0 or by removing bits
119 * from an register with all bits set. Try which alternative needs fewer
121 arm_gen_vals_from_word(value, &v);
122 arm_gen_vals_from_word(~value, &vn);
124 if (vn.ops < v.ops) {
126 result = new_bd_arm_Mvn_imm(dbgi, block, vn.values[0], vn.rors[0]);
128 for (cnt = 1; cnt < vn.ops; ++cnt) {
129 result = new_bd_arm_Bic_imm(dbgi, block, result,
130 vn.values[cnt], vn.rors[cnt]);
134 result = new_bd_arm_Mov_imm(dbgi, block, v.values[0], v.rors[0]);
136 for (cnt = 1; cnt < v.ops; ++cnt) {
137 result = new_bd_arm_Or_imm(dbgi, block, result,
138 v.values[cnt], v.rors[cnt]);
145 * Create a DAG constructing a given Const.
147 * @param irn a Firm const
149 static ir_node *create_const_graph(ir_node *irn, ir_node *block)
151 ir_tarval *tv = get_Const_tarval(irn);
152 ir_mode *mode = get_tarval_mode(tv);
155 if (mode_is_reference(mode)) {
156 /* ARM is 32bit, so we can safely convert a reference tarval into Iu */
157 assert(get_mode_size_bits(mode) == get_mode_size_bits(mode_Iu));
158 tv = tarval_convert_to(tv, mode_Iu);
160 value = get_tarval_long(tv);
161 return create_const_graph_value(get_irn_dbg_info(irn), block, value);
165 * Create an And that will zero out upper bits.
167 * @param dbgi debug info
168 * @param block the basic block
169 * @param op the original node
170 * param src_bits number of lower bits that will remain
172 static ir_node *gen_zero_extension(dbg_info *dbgi, ir_node *block, ir_node *op,
176 return new_bd_arm_And_imm(dbgi, block, op, 0xFF, 0);
177 } else if (src_bits == 16) {
178 ir_node *lshift = new_bd_arm_Mov_reg_shift_imm(dbgi, block, op, ARM_SHF_LSL_IMM, 16);
179 ir_node *rshift = new_bd_arm_Mov_reg_shift_imm(dbgi, block, lshift, ARM_SHF_LSR_IMM, 16);
182 panic("zero extension only supported for 8 and 16 bits");
187 * Generate code for a sign extension.
189 static ir_node *gen_sign_extension(dbg_info *dbgi, ir_node *block, ir_node *op,
192 int shift_width = 32 - src_bits;
193 ir_node *lshift_node = new_bd_arm_Mov_reg_shift_imm(dbgi, block, op, ARM_SHF_LSL_IMM, shift_width);
194 ir_node *rshift_node = new_bd_arm_Mov_reg_shift_imm(dbgi, block, lshift_node, ARM_SHF_ASR_IMM, shift_width);
198 static ir_node *gen_extension(dbg_info *dbgi, ir_node *block, ir_node *op,
201 int bits = get_mode_size_bits(orig_mode);
205 if (mode_is_signed(orig_mode)) {
206 return gen_sign_extension(dbgi, block, op, bits);
208 return gen_zero_extension(dbgi, block, op, bits);
213 * returns true if it is assured, that the upper bits of a node are "clean"
214 * which means for a 16 or 8 bit value, that the upper bits in the register
215 * are 0 for unsigned and a copy of the last significant bit for signed
218 static bool upper_bits_clean(ir_node *transformed_node, ir_mode *mode)
220 (void) transformed_node;
227 * Transforms a Conv node.
229 * @return The created ia32 Conv node
231 static ir_node *gen_Conv(ir_node *node)
233 ir_node *block = be_transform_node(get_nodes_block(node));
234 ir_node *op = get_Conv_op(node);
235 ir_node *new_op = be_transform_node(op);
236 ir_mode *src_mode = get_irn_mode(op);
237 ir_mode *dst_mode = get_irn_mode(node);
238 dbg_info *dbg = get_irn_dbg_info(node);
240 if (src_mode == dst_mode)
243 if (mode_is_float(src_mode) || mode_is_float(dst_mode)) {
245 if (mode_is_float(src_mode)) {
246 if (mode_is_float(dst_mode)) {
247 /* from float to float */
248 return new_bd_arm_Mvf(dbg, block, new_op, dst_mode);
250 /* from float to int */
254 /* from int to float */
255 if (!mode_is_signed(src_mode)) {
258 return new_bd_arm_FltX(dbg, block, new_op, dst_mode);
261 } else if (USE_VFP(isa)) {
262 panic("VFP not supported yet");
264 panic("Softfloat not supported yet");
266 } else { /* complete in gp registers */
267 int src_bits = get_mode_size_bits(src_mode);
268 int dst_bits = get_mode_size_bits(dst_mode);
272 if (src_bits == dst_bits) {
273 /* kill unnecessary conv */
277 if (src_bits < dst_bits) {
285 if (upper_bits_clean(new_op, min_mode)) {
289 if (mode_is_signed(min_mode)) {
290 return gen_sign_extension(dbg, block, new_op, min_bits);
292 return gen_zero_extension(dbg, block, new_op, min_bits);
302 static bool try_encode_as_immediate(const ir_node *node, arm_immediate_t *res)
304 unsigned val, low_pos, high_pos;
309 val = get_tarval_long(get_Const_tarval(node));
321 /* arm allows to use to rotate an 8bit immediate value by a multiple of 2
323 So we determine the smallest even position with a bit set
324 and the highest even position with no bit set anymore.
325 If the difference between these 2 is <= 8, then we can encode the value
328 low_pos = ntz(val) & ~1u;
329 high_pos = (32-nlz(val)+1) & ~1u;
331 if (high_pos - low_pos <= 8) {
332 res->imm_8 = val >> low_pos;
333 res->rot = 32 - low_pos;
338 res->rot = 34 - high_pos;
339 val = val >> (32-res->rot) | val << (res->rot);
349 static bool is_downconv(const ir_node *node)
357 /* we only want to skip the conv when we're the only user
358 * (not optimal but for now...)
360 if (get_irn_n_edges(node) > 1)
363 src_mode = get_irn_mode(get_Conv_op(node));
364 dest_mode = get_irn_mode(node);
366 mode_needs_gp_reg(src_mode) &&
367 mode_needs_gp_reg(dest_mode) &&
368 get_mode_size_bits(dest_mode) <= get_mode_size_bits(src_mode);
371 static ir_node *arm_skip_downconv(ir_node *node)
373 while (is_downconv(node))
374 node = get_Conv_op(node);
380 MATCH_COMMUTATIVE = 1 << 0, /**< commutative node */
381 MATCH_REVERSE = 1 << 1, /**< support reverse opcode */
382 MATCH_SIZE_NEUTRAL = 1 << 2,
383 MATCH_SKIP_NOT = 1 << 3, /**< skip Not on ONE input */
385 ENUM_BITSET(match_flags_t)
388 * possible binop constructors.
390 typedef struct arm_binop_factory_t {
391 /** normal reg op reg operation. */
392 ir_node *(*new_binop_reg)(dbg_info *dbgi, ir_node *block, ir_node *op1, ir_node *op2);
393 /** normal reg op imm operation. */
394 ir_node *(*new_binop_imm)(dbg_info *dbgi, ir_node *block, ir_node *op1, unsigned char imm8, unsigned char imm_rot);
395 /** barrel shifter reg op (reg shift reg operation. */
396 ir_node *(*new_binop_reg_shift_reg)(dbg_info *dbgi, ir_node *block, ir_node *left, ir_node *right, ir_node *shift, arm_shift_modifier_t shift_modifier);
397 /** barrel shifter reg op (reg shift imm operation. */
398 ir_node *(*new_binop_reg_shift_imm)(dbg_info *dbgi, ir_node *block, ir_node *left, ir_node *right, arm_shift_modifier_t shift_modifier, unsigned shift_immediate);
399 } arm_binop_factory_t;
401 static ir_node *gen_int_binop(ir_node *node, match_flags_t flags,
402 const arm_binop_factory_t *factory)
404 ir_node *block = be_transform_node(get_nodes_block(node));
405 ir_node *op1 = get_binop_left(node);
407 ir_node *op2 = get_binop_right(node);
409 dbg_info *dbgi = get_irn_dbg_info(node);
412 if (flags & MATCH_SKIP_NOT) {
414 op1 = get_Not_op(op1);
415 else if (is_Not(op2))
416 op2 = get_Not_op(op2);
418 panic("cannot execute MATCH_SKIP_NOT");
420 if (flags & MATCH_SIZE_NEUTRAL) {
421 op1 = arm_skip_downconv(op1);
422 op2 = arm_skip_downconv(op2);
424 assert(get_mode_size_bits(get_irn_mode(node)) == 32);
427 if (try_encode_as_immediate(op2, &imm)) {
428 new_op1 = be_transform_node(op1);
429 return factory->new_binop_imm(dbgi, block, new_op1, imm.imm_8, imm.rot);
431 new_op2 = be_transform_node(op2);
432 if ((flags & (MATCH_COMMUTATIVE|MATCH_REVERSE)) && try_encode_as_immediate(op1, &imm)) {
433 if (flags & MATCH_REVERSE)
434 return factory[1].new_binop_imm(dbgi, block, new_op2, imm.imm_8, imm.rot);
436 return factory[0].new_binop_imm(dbgi, block, new_op2, imm.imm_8, imm.rot);
438 new_op1 = be_transform_node(op1);
440 /* check if we can fold in a Mov */
441 if (is_arm_Mov(new_op2)) {
442 const arm_shifter_operand_t *attr = get_arm_shifter_operand_attr_const(new_op2);
444 switch (attr->shift_modifier) {
446 case ARM_SHF_ASR_IMM:
447 case ARM_SHF_LSL_IMM:
448 case ARM_SHF_LSR_IMM:
449 case ARM_SHF_ROR_IMM:
450 if (factory->new_binop_reg_shift_imm) {
451 ir_node *mov_op = get_irn_n(new_op2, 0);
452 return factory->new_binop_reg_shift_imm(dbgi, block, new_op1, mov_op,
453 attr->shift_modifier, attr->shift_immediate);
457 case ARM_SHF_ASR_REG:
458 case ARM_SHF_LSL_REG:
459 case ARM_SHF_LSR_REG:
460 case ARM_SHF_ROR_REG:
461 if (factory->new_binop_reg_shift_reg) {
462 ir_node *mov_op = get_irn_n(new_op2, 0);
463 ir_node *mov_sft = get_irn_n(new_op2, 1);
464 return factory->new_binop_reg_shift_reg(dbgi, block, new_op1, mov_op, mov_sft,
465 attr->shift_modifier);
471 case ARM_SHF_INVALID:
472 panic("invalid shift");
475 if ((flags & (MATCH_COMMUTATIVE|MATCH_REVERSE)) && is_arm_Mov(new_op1)) {
476 const arm_shifter_operand_t *attr = get_arm_shifter_operand_attr_const(new_op1);
477 int idx = flags & MATCH_REVERSE ? 1 : 0;
479 switch (attr->shift_modifier) {
480 ir_node *mov_op, *mov_sft;
483 case ARM_SHF_ASR_IMM:
484 case ARM_SHF_LSL_IMM:
485 case ARM_SHF_LSR_IMM:
486 case ARM_SHF_ROR_IMM:
487 if (factory[idx].new_binop_reg_shift_imm) {
488 mov_op = get_irn_n(new_op1, 0);
489 return factory[idx].new_binop_reg_shift_imm(dbgi, block, new_op2, mov_op,
490 attr->shift_modifier, attr->shift_immediate);
494 case ARM_SHF_ASR_REG:
495 case ARM_SHF_LSL_REG:
496 case ARM_SHF_LSR_REG:
497 case ARM_SHF_ROR_REG:
498 if (factory[idx].new_binop_reg_shift_reg) {
499 mov_op = get_irn_n(new_op1, 0);
500 mov_sft = get_irn_n(new_op1, 1);
501 return factory[idx].new_binop_reg_shift_reg(dbgi, block, new_op2, mov_op, mov_sft,
502 attr->shift_modifier);
509 case ARM_SHF_INVALID:
510 panic("invalid shift");
513 return factory->new_binop_reg(dbgi, block, new_op1, new_op2);
517 * Creates an ARM Add.
519 * @return the created arm Add node
521 static ir_node *gen_Add(ir_node *node)
523 static const arm_binop_factory_t add_factory = {
526 new_bd_arm_Add_reg_shift_reg,
527 new_bd_arm_Add_reg_shift_imm
530 ir_mode *mode = get_irn_mode(node);
532 if (mode_is_float(mode)) {
533 ir_node *block = be_transform_node(get_nodes_block(node));
534 ir_node *op1 = get_Add_left(node);
535 ir_node *op2 = get_Add_right(node);
536 dbg_info *dbgi = get_irn_dbg_info(node);
537 ir_node *new_op1 = be_transform_node(op1);
538 ir_node *new_op2 = be_transform_node(op2);
540 return new_bd_arm_Adf(dbgi, block, new_op1, new_op2, mode);
541 } else if (USE_VFP(isa)) {
542 assert(mode != mode_E && "IEEE Extended FP not supported");
543 panic("VFP not supported yet");
545 panic("Softfloat not supported yet");
550 if (is_arm_Mul(new_op1) && get_irn_n_edges(op1) == 1) {
552 new_op2 = get_irn_n(new_op1, 1);
553 new_op1 = get_irn_n(new_op1, 0);
555 return new_bd_arm_Mla(dbgi, block, new_op1, new_op2, new_op3);
557 if (is_arm_Mul(new_op2) && get_irn_n_edges(op2) == 1) {
559 new_op1 = get_irn_n(new_op2, 0);
560 new_op2 = get_irn_n(new_op2, 1);
562 return new_bd_arm_Mla(dbgi, block, new_op1, new_op2, new_op3);
566 return gen_int_binop(node, MATCH_COMMUTATIVE | MATCH_SIZE_NEUTRAL, &add_factory);
571 * Creates an ARM Mul.
573 * @return the created arm Mul node
575 static ir_node *gen_Mul(ir_node *node)
577 ir_node *block = be_transform_node(get_nodes_block(node));
578 ir_node *op1 = get_Mul_left(node);
579 ir_node *new_op1 = be_transform_node(op1);
580 ir_node *op2 = get_Mul_right(node);
581 ir_node *new_op2 = be_transform_node(op2);
582 ir_mode *mode = get_irn_mode(node);
583 dbg_info *dbg = get_irn_dbg_info(node);
585 if (mode_is_float(mode)) {
587 return new_bd_arm_Muf(dbg, block, new_op1, new_op2, mode);
588 } else if (USE_VFP(isa)) {
589 assert(mode != mode_E && "IEEE Extended FP not supported");
590 panic("VFP not supported yet");
592 panic("Softfloat not supported yet");
595 assert(mode_is_data(mode));
596 return new_bd_arm_Mul(dbg, block, new_op1, new_op2);
599 static ir_node *gen_Div(ir_node *node)
601 ir_node *block = be_transform_node(get_nodes_block(node));
602 ir_node *op1 = get_Div_left(node);
603 ir_node *new_op1 = be_transform_node(op1);
604 ir_node *op2 = get_Div_right(node);
605 ir_node *new_op2 = be_transform_node(op2);
606 ir_mode *mode = get_Div_resmode(node);
607 dbg_info *dbg = get_irn_dbg_info(node);
609 assert(mode != mode_E && "IEEE Extended FP not supported");
610 /* integer division should be replaced by builtin call */
611 assert(mode_is_float(mode));
614 return new_bd_arm_Dvf(dbg, block, new_op1, new_op2, mode);
615 } else if (USE_VFP(isa)) {
616 assert(mode != mode_E && "IEEE Extended FP not supported");
617 panic("VFP not supported yet");
619 panic("Softfloat not supported yet");
623 static ir_node *gen_And(ir_node *node)
625 static const arm_binop_factory_t and_factory = {
628 new_bd_arm_And_reg_shift_reg,
629 new_bd_arm_And_reg_shift_imm
631 static const arm_binop_factory_t bic_factory = {
634 new_bd_arm_Bic_reg_shift_reg,
635 new_bd_arm_Bic_reg_shift_imm
638 /* check for and not */
639 ir_node *left = get_And_left(node);
640 ir_node *right = get_And_right(node);
642 if (is_Not(left) || is_Not(right)) {
643 return gen_int_binop(node, MATCH_COMMUTATIVE | MATCH_SIZE_NEUTRAL | MATCH_SKIP_NOT,
647 return gen_int_binop(node, MATCH_COMMUTATIVE | MATCH_SIZE_NEUTRAL, &and_factory);
650 static ir_node *gen_Or(ir_node *node)
652 static const arm_binop_factory_t or_factory = {
655 new_bd_arm_Or_reg_shift_reg,
656 new_bd_arm_Or_reg_shift_imm
659 return gen_int_binop(node, MATCH_COMMUTATIVE | MATCH_SIZE_NEUTRAL, &or_factory);
662 static ir_node *gen_Eor(ir_node *node)
664 static const arm_binop_factory_t eor_factory = {
667 new_bd_arm_Eor_reg_shift_reg,
668 new_bd_arm_Eor_reg_shift_imm
671 return gen_int_binop(node, MATCH_COMMUTATIVE | MATCH_SIZE_NEUTRAL, &eor_factory);
674 static ir_node *gen_Sub(ir_node *node)
676 static const arm_binop_factory_t sub_rsb_factory[2] = {
680 new_bd_arm_Sub_reg_shift_reg,
681 new_bd_arm_Sub_reg_shift_imm
686 new_bd_arm_Rsb_reg_shift_reg,
687 new_bd_arm_Rsb_reg_shift_imm
691 ir_node *block = be_transform_node(get_nodes_block(node));
692 ir_node *op1 = get_Sub_left(node);
693 ir_node *new_op1 = be_transform_node(op1);
694 ir_node *op2 = get_Sub_right(node);
695 ir_node *new_op2 = be_transform_node(op2);
696 ir_mode *mode = get_irn_mode(node);
697 dbg_info *dbgi = get_irn_dbg_info(node);
699 if (mode_is_float(mode)) {
701 return new_bd_arm_Suf(dbgi, block, new_op1, new_op2, mode);
702 } else if (USE_VFP(isa)) {
703 assert(mode != mode_E && "IEEE Extended FP not supported");
704 panic("VFP not supported yet");
706 panic("Softfloat not supported yet");
709 return gen_int_binop(node, MATCH_SIZE_NEUTRAL | MATCH_REVERSE, sub_rsb_factory);
714 * Checks if a given value can be used as an immediate for the given
717 static bool can_use_shift_constant(unsigned int val,
718 arm_shift_modifier_t modifier)
722 if (val == 32 && modifier != ARM_SHF_LSL_REG && modifier != ARM_SHF_ROR_REG)
728 * generate an ARM shift instruction.
730 * @param node the node
731 * @param flags matching flags
732 * @param shift_modifier initial encoding of the desired shift operation
734 static ir_node *make_shift(ir_node *node, match_flags_t flags,
735 arm_shift_modifier_t shift_modifier)
737 ir_node *block = be_transform_node(get_nodes_block(node));
738 ir_node *op1 = get_binop_left(node);
739 ir_node *op2 = get_binop_right(node);
740 dbg_info *dbgi = get_irn_dbg_info(node);
741 ir_mode *mode = get_irn_mode(node);
745 if (get_mode_modulo_shift(mode) != 32)
746 panic("modulo shift!=32 not supported by arm backend");
748 if (flags & MATCH_SIZE_NEUTRAL) {
749 op1 = arm_skip_downconv(op1);
750 op2 = arm_skip_downconv(op2);
753 new_op1 = be_transform_node(op1);
755 ir_tarval *tv = get_Const_tarval(op2);
756 unsigned int val = get_tarval_long(tv);
757 assert(tarval_is_long(tv));
758 if (can_use_shift_constant(val, shift_modifier)) {
759 switch (shift_modifier) {
760 case ARM_SHF_LSL_REG: shift_modifier = ARM_SHF_LSL_IMM; break;
761 case ARM_SHF_LSR_REG: shift_modifier = ARM_SHF_LSR_IMM; break;
762 case ARM_SHF_ASR_REG: shift_modifier = ARM_SHF_ASR_IMM; break;
763 case ARM_SHF_ROR_REG: shift_modifier = ARM_SHF_ROR_IMM; break;
764 default: panic("unexpected shift modifier");
766 return new_bd_arm_Mov_reg_shift_imm(dbgi, block, new_op1,
767 shift_modifier, val);
771 new_op2 = be_transform_node(op2);
772 return new_bd_arm_Mov_reg_shift_reg(dbgi, block, new_op1, new_op2,
776 static ir_node *gen_Shl(ir_node *node)
778 return make_shift(node, MATCH_SIZE_NEUTRAL, ARM_SHF_LSL_REG);
781 static ir_node *gen_Shr(ir_node *node)
783 return make_shift(node, MATCH_NONE, ARM_SHF_LSR_REG);
786 static ir_node *gen_Shrs(ir_node *node)
788 return make_shift(node, MATCH_NONE, ARM_SHF_ASR_REG);
791 static ir_node *gen_Ror(ir_node *node, ir_node *op1, ir_node *op2)
793 ir_node *block = be_transform_node(get_nodes_block(node));
794 ir_node *new_op1 = be_transform_node(op1);
795 dbg_info *dbgi = get_irn_dbg_info(node);
796 ir_node *new_op2 = be_transform_node(op2);
798 return new_bd_arm_Mov_reg_shift_reg(dbgi, block, new_op1, new_op2,
802 static ir_node *gen_Rol(ir_node *node, ir_node *op1, ir_node *op2)
804 ir_node *block = be_transform_node(get_nodes_block(node));
805 ir_node *new_op1 = be_transform_node(op1);
806 dbg_info *dbgi = get_irn_dbg_info(node);
807 ir_node *new_op2 = be_transform_node(op2);
809 /* Note: there is no Rol on arm, we have to use Ror */
810 new_op2 = new_bd_arm_Rsb_imm(dbgi, block, new_op2, 32, 0);
811 return new_bd_arm_Mov_reg_shift_reg(dbgi, block, new_op1, new_op2,
815 static ir_node *gen_Rotl(ir_node *node)
817 ir_node *rotate = NULL;
818 ir_node *op1 = get_Rotl_left(node);
819 ir_node *op2 = get_Rotl_right(node);
821 /* Firm has only RotL, so we are looking for a right (op2)
822 operand "-e+mode_size_bits" (it's an already modified "mode_size_bits-e",
823 that means we can create a RotR. */
826 ir_node *right = get_Add_right(op2);
827 if (is_Const(right)) {
828 ir_tarval *tv = get_Const_tarval(right);
829 ir_mode *mode = get_irn_mode(node);
830 long bits = get_mode_size_bits(mode);
831 ir_node *left = get_Add_left(op2);
833 if (is_Minus(left) &&
834 tarval_is_long(tv) &&
835 get_tarval_long(tv) == bits &&
837 rotate = gen_Ror(node, op1, get_Minus_op(left));
839 } else if (is_Sub(op2)) {
840 ir_node *left = get_Sub_left(op2);
841 if (is_Const(left)) {
842 ir_tarval *tv = get_Const_tarval(left);
843 ir_mode *mode = get_irn_mode(node);
844 long bits = get_mode_size_bits(mode);
845 ir_node *right = get_Sub_right(op2);
847 if (tarval_is_long(tv) &&
848 get_tarval_long(tv) == bits &&
850 rotate = gen_Ror(node, op1, right);
852 } else if (is_Const(op2)) {
853 ir_tarval *tv = get_Const_tarval(op2);
854 ir_mode *mode = get_irn_mode(node);
855 long bits = get_mode_size_bits(mode);
857 if (tarval_is_long(tv) && bits == 32) {
858 ir_node *block = be_transform_node(get_nodes_block(node));
859 ir_node *new_op1 = be_transform_node(op1);
860 dbg_info *dbgi = get_irn_dbg_info(node);
862 bits = (bits - get_tarval_long(tv)) & 31;
863 rotate = new_bd_arm_Mov_reg_shift_imm(dbgi, block, new_op1, ARM_SHF_ROR_IMM, bits);
867 if (rotate == NULL) {
868 rotate = gen_Rol(node, op1, op2);
874 static ir_node *gen_Not(ir_node *node)
876 ir_node *block = be_transform_node(get_nodes_block(node));
877 ir_node *op = get_Not_op(node);
878 ir_node *new_op = be_transform_node(op);
879 dbg_info *dbgi = get_irn_dbg_info(node);
881 /* check if we can fold in a Mov */
882 if (is_arm_Mov(new_op)) {
883 const arm_shifter_operand_t *attr = get_arm_shifter_operand_attr_const(new_op);
885 switch (attr->shift_modifier) {
886 ir_node *mov_op, *mov_sft;
889 case ARM_SHF_ASR_IMM:
890 case ARM_SHF_LSL_IMM:
891 case ARM_SHF_LSR_IMM:
892 case ARM_SHF_ROR_IMM:
893 mov_op = get_irn_n(new_op, 0);
894 return new_bd_arm_Mvn_reg_shift_imm(dbgi, block, mov_op,
895 attr->shift_modifier, attr->shift_immediate);
897 case ARM_SHF_ASR_REG:
898 case ARM_SHF_LSL_REG:
899 case ARM_SHF_LSR_REG:
900 case ARM_SHF_ROR_REG:
901 mov_op = get_irn_n(new_op, 0);
902 mov_sft = get_irn_n(new_op, 1);
903 return new_bd_arm_Mvn_reg_shift_reg(dbgi, block, mov_op, mov_sft,
904 attr->shift_modifier);
909 case ARM_SHF_INVALID:
910 panic("invalid shift");
914 return new_bd_arm_Mvn_reg(dbgi, block, new_op);
917 static ir_node *gen_Minus(ir_node *node)
919 ir_node *block = be_transform_node(get_nodes_block(node));
920 ir_node *op = get_Minus_op(node);
921 ir_node *new_op = be_transform_node(op);
922 dbg_info *dbgi = get_irn_dbg_info(node);
923 ir_mode *mode = get_irn_mode(node);
925 if (mode_is_float(mode)) {
927 return new_bd_arm_Mvf(dbgi, block, op, mode);
928 } else if (USE_VFP(isa)) {
929 assert(mode != mode_E && "IEEE Extended FP not supported");
930 panic("VFP not supported yet");
932 panic("Softfloat not supported yet");
935 assert(mode_is_data(mode));
936 return new_bd_arm_Rsb_imm(dbgi, block, new_op, 0, 0);
939 static ir_node *gen_Load(ir_node *node)
941 ir_node *block = be_transform_node(get_nodes_block(node));
942 ir_node *ptr = get_Load_ptr(node);
943 ir_node *new_ptr = be_transform_node(ptr);
944 ir_node *mem = get_Load_mem(node);
945 ir_node *new_mem = be_transform_node(mem);
946 ir_mode *mode = get_Load_mode(node);
947 dbg_info *dbgi = get_irn_dbg_info(node);
948 ir_node *new_load = NULL;
950 if (get_Load_unaligned(node) == align_non_aligned)
951 panic("arm: unaligned Loads not supported yet");
953 if (mode_is_float(mode)) {
955 new_load = new_bd_arm_Ldf(dbgi, block, new_ptr, new_mem, mode,
957 } else if (USE_VFP(isa)) {
958 assert(mode != mode_E && "IEEE Extended FP not supported");
959 panic("VFP not supported yet");
961 panic("Softfloat not supported yet");
964 assert(mode_is_data(mode) && "unsupported mode for Load");
966 new_load = new_bd_arm_Ldr(dbgi, block, new_ptr, new_mem, mode, NULL, 0, 0, false);
968 set_irn_pinned(new_load, get_irn_pinned(node));
970 /* check for special case: the loaded value might not be used */
971 if (be_get_Proj_for_pn(node, pn_Load_res) == NULL) {
972 /* add a result proj and a Keep to produce a pseudo use */
973 ir_node *proj = new_r_Proj(new_load, mode_Iu, pn_arm_Ldr_res);
974 be_new_Keep(block, 1, &proj);
980 static ir_node *gen_Store(ir_node *node)
982 ir_node *block = be_transform_node(get_nodes_block(node));
983 ir_node *ptr = get_Store_ptr(node);
984 ir_node *new_ptr = be_transform_node(ptr);
985 ir_node *mem = get_Store_mem(node);
986 ir_node *new_mem = be_transform_node(mem);
987 ir_node *val = get_Store_value(node);
988 ir_node *new_val = be_transform_node(val);
989 ir_mode *mode = get_irn_mode(val);
990 dbg_info *dbgi = get_irn_dbg_info(node);
991 ir_node *new_store = NULL;
993 if (get_Store_unaligned(node) == align_non_aligned)
994 panic("arm: unaligned Stores not supported yet");
996 if (mode_is_float(mode)) {
998 new_store = new_bd_arm_Stf(dbgi, block, new_ptr, new_val,
999 new_mem, mode, NULL, 0, 0, false);
1000 } else if (USE_VFP(isa)) {
1001 assert(mode != mode_E && "IEEE Extended FP not supported");
1002 panic("VFP not supported yet");
1004 panic("Softfloat not supported yet");
1007 assert(mode_is_data(mode) && "unsupported mode for Store");
1008 new_store = new_bd_arm_Str(dbgi, block, new_ptr, new_val, new_mem, mode,
1011 set_irn_pinned(new_store, get_irn_pinned(node));
1015 static ir_node *gen_Jmp(ir_node *node)
1017 ir_node *block = get_nodes_block(node);
1018 ir_node *new_block = be_transform_node(block);
1019 dbg_info *dbgi = get_irn_dbg_info(node);
1021 return new_bd_arm_Jmp(dbgi, new_block);
1024 static ir_node *gen_SwitchJmp(ir_node *node)
1026 ir_node *block = be_transform_node(get_nodes_block(node));
1027 ir_node *selector = get_Cond_selector(node);
1028 dbg_info *dbgi = get_irn_dbg_info(node);
1029 ir_node *new_op = be_transform_node(selector);
1030 ir_node *const_graph;
1034 const ir_edge_t *edge;
1041 foreach_out_edge(node, edge) {
1042 proj = get_edge_src_irn(edge);
1043 assert(is_Proj(proj) && "Only proj allowed at SwitchJmp");
1045 pn = get_Proj_proj(proj);
1047 min = pn<min ? pn : min;
1048 max = pn>max ? pn : max;
1051 n_projs = max - translation + 1;
1053 foreach_out_edge(node, edge) {
1054 proj = get_edge_src_irn(edge);
1055 assert(is_Proj(proj) && "Only proj allowed at SwitchJmp");
1057 pn = get_Proj_proj(proj) - translation;
1058 set_Proj_proj(proj, pn);
1061 const_graph = create_const_graph_value(dbgi, block, translation);
1062 sub = new_bd_arm_Sub_reg(dbgi, block, new_op, const_graph);
1063 return new_bd_arm_SwitchJmp(dbgi, block, sub, n_projs, get_Cond_default_proj(node) - translation);
1066 static ir_node *gen_Cmp(ir_node *node)
1068 ir_node *block = be_transform_node(get_nodes_block(node));
1069 ir_node *op1 = get_Cmp_left(node);
1070 ir_node *op2 = get_Cmp_right(node);
1071 ir_mode *cmp_mode = get_irn_mode(op1);
1072 dbg_info *dbgi = get_irn_dbg_info(node);
1077 if (mode_is_float(cmp_mode)) {
1078 /* TODO: this is broken... */
1079 new_op1 = be_transform_node(op1);
1080 new_op2 = be_transform_node(op2);
1082 return new_bd_arm_Cmfe(dbgi, block, new_op1, new_op2, false);
1085 assert(get_irn_mode(op2) == cmp_mode);
1086 is_unsigned = !mode_is_signed(cmp_mode);
1088 /* integer compare, TODO: use shifter_op in all its combinations */
1089 new_op1 = be_transform_node(op1);
1090 new_op1 = gen_extension(dbgi, block, new_op1, cmp_mode);
1091 new_op2 = be_transform_node(op2);
1092 new_op2 = gen_extension(dbgi, block, new_op2, cmp_mode);
1093 return new_bd_arm_Cmp_reg(dbgi, block, new_op1, new_op2, false,
1097 static ir_node *gen_Cond(ir_node *node)
1099 ir_node *selector = get_Cond_selector(node);
1100 ir_mode *mode = get_irn_mode(selector);
1101 ir_relation relation;
1106 if (mode != mode_b) {
1107 return gen_SwitchJmp(node);
1109 assert(is_Cmp(selector));
1111 block = be_transform_node(get_nodes_block(node));
1112 dbgi = get_irn_dbg_info(node);
1113 flag_node = be_transform_node(selector);
1114 relation = get_Cmp_relation(selector);
1116 return new_bd_arm_B(dbgi, block, flag_node, relation);
1122 FPA_IMM_EXTENDED = 2,
1123 FPA_IMM_MAX = FPA_IMM_EXTENDED
1126 static ir_tarval *fpa_imm[FPA_IMM_MAX + 1][fpa_max];
1130 * Check, if a floating point tarval is an fpa immediate, i.e.
1131 * one of 0, 1, 2, 3, 4, 5, 10, or 0.5.
1133 static int is_fpa_immediate(tarval *tv)
1135 ir_mode *mode = get_tarval_mode(tv);
1138 switch (get_mode_size_bits(mode)) {
1146 i = FPA_IMM_EXTENDED;
1149 if (tarval_is_negative(tv)) {
1150 tv = tarval_neg(tv);
1154 for (j = 0; j < fpa_max; ++j) {
1155 if (tv == fpa_imm[i][j])
1162 static ir_node *gen_Const(ir_node *node)
1164 ir_node *block = be_transform_node(get_nodes_block(node));
1165 ir_mode *mode = get_irn_mode(node);
1166 dbg_info *dbg = get_irn_dbg_info(node);
1168 if (mode_is_float(mode)) {
1170 ir_tarval *tv = get_Const_tarval(node);
1171 node = new_bd_arm_fConst(dbg, block, tv);
1173 } else if (USE_VFP(isa)) {
1174 assert(mode != mode_E && "IEEE Extended FP not supported");
1175 panic("VFP not supported yet");
1177 panic("Softfloat not supported yet");
1180 return create_const_graph(node, block);
1183 static ir_node *gen_SymConst(ir_node *node)
1185 ir_node *block = be_transform_node(get_nodes_block(node));
1186 ir_entity *entity = get_SymConst_entity(node);
1187 dbg_info *dbgi = get_irn_dbg_info(node);
1190 new_node = new_bd_arm_SymConst(dbgi, block, entity, 0);
1194 static ir_node *ints_to_double(dbg_info *dbgi, ir_node *block, ir_node *node0,
1197 /* the good way to do this would be to use the stm (store multiple)
1198 * instructions, since our input is nearly always 2 consecutive 32bit
1200 ir_graph *irg = current_ir_graph;
1201 ir_node *stack = get_irg_frame(irg);
1202 ir_node *nomem = get_irg_no_mem(irg);
1203 ir_node *str0 = new_bd_arm_Str(dbgi, block, stack, node0, nomem, mode_gp,
1205 ir_node *str1 = new_bd_arm_Str(dbgi, block, stack, node1, nomem, mode_gp,
1207 ir_node *in[2] = { str0, str1 };
1208 ir_node *sync = new_r_Sync(block, 2, in);
1210 set_irn_pinned(str0, op_pin_state_floats);
1211 set_irn_pinned(str1, op_pin_state_floats);
1213 ldf = new_bd_arm_Ldf(dbgi, block, stack, sync, mode_D, NULL, 0, 0, true);
1214 set_irn_pinned(ldf, op_pin_state_floats);
1216 return new_r_Proj(ldf, mode_fp, pn_arm_Ldf_res);
1219 static ir_node *int_to_float(dbg_info *dbgi, ir_node *block, ir_node *node)
1221 ir_graph *irg = current_ir_graph;
1222 ir_node *stack = get_irg_frame(irg);
1223 ir_node *nomem = get_irg_no_mem(irg);
1224 ir_node *str = new_bd_arm_Str(dbgi, block, stack, node, nomem, mode_gp,
1227 set_irn_pinned(str, op_pin_state_floats);
1229 ldf = new_bd_arm_Ldf(dbgi, block, stack, str, mode_F, NULL, 0, 0, true);
1230 set_irn_pinned(ldf, op_pin_state_floats);
1232 return new_r_Proj(ldf, mode_fp, pn_arm_Ldf_res);
1235 static ir_node *float_to_int(dbg_info *dbgi, ir_node *block, ir_node *node)
1237 ir_graph *irg = current_ir_graph;
1238 ir_node *stack = get_irg_frame(irg);
1239 ir_node *nomem = get_irg_no_mem(irg);
1240 ir_node *stf = new_bd_arm_Stf(dbgi, block, stack, node, nomem, mode_F,
1243 set_irn_pinned(stf, op_pin_state_floats);
1245 ldr = new_bd_arm_Ldr(dbgi, block, stack, stf, mode_gp, NULL, 0, 0, true);
1246 set_irn_pinned(ldr, op_pin_state_floats);
1248 return new_r_Proj(ldr, mode_gp, pn_arm_Ldr_res);
1251 static void double_to_ints(dbg_info *dbgi, ir_node *block, ir_node *node,
1252 ir_node **out_value0, ir_node **out_value1)
1254 ir_graph *irg = current_ir_graph;
1255 ir_node *stack = get_irg_frame(irg);
1256 ir_node *nomem = get_irg_no_mem(irg);
1257 ir_node *stf = new_bd_arm_Stf(dbgi, block, stack, node, nomem, mode_D,
1259 ir_node *ldr0, *ldr1;
1260 set_irn_pinned(stf, op_pin_state_floats);
1262 ldr0 = new_bd_arm_Ldr(dbgi, block, stack, stf, mode_gp, NULL, 0, 0, true);
1263 set_irn_pinned(ldr0, op_pin_state_floats);
1264 ldr1 = new_bd_arm_Ldr(dbgi, block, stack, stf, mode_gp, NULL, 0, 4, true);
1265 set_irn_pinned(ldr1, op_pin_state_floats);
1267 *out_value0 = new_r_Proj(ldr0, mode_gp, pn_arm_Ldr_res);
1268 *out_value1 = new_r_Proj(ldr1, mode_gp, pn_arm_Ldr_res);
1271 static ir_node *gen_CopyB(ir_node *node)
1273 ir_node *block = be_transform_node(get_nodes_block(node));
1274 ir_node *src = get_CopyB_src(node);
1275 ir_node *new_src = be_transform_node(src);
1276 ir_node *dst = get_CopyB_dst(node);
1277 ir_node *new_dst = be_transform_node(dst);
1278 ir_node *mem = get_CopyB_mem(node);
1279 ir_node *new_mem = be_transform_node(mem);
1280 dbg_info *dbg = get_irn_dbg_info(node);
1281 int size = get_type_size_bytes(get_CopyB_type(node));
1285 src_copy = be_new_Copy(block, new_src);
1286 dst_copy = be_new_Copy(block, new_dst);
1288 return new_bd_arm_CopyB(dbg, block, dst_copy, src_copy,
1289 new_bd_arm_EmptyReg(dbg, block),
1290 new_bd_arm_EmptyReg(dbg, block),
1291 new_bd_arm_EmptyReg(dbg, block),
1296 * Transform builtin clz.
1298 static ir_node *gen_clz(ir_node *node)
1300 ir_node *block = be_transform_node(get_nodes_block(node));
1301 dbg_info *dbg = get_irn_dbg_info(node);
1302 ir_node *op = get_irn_n(node, 1);
1303 ir_node *new_op = be_transform_node(op);
1305 /* TODO armv5 instruction, otherwise create a call */
1306 return new_bd_arm_Clz(dbg, block, new_op);
1310 * Transform Builtin node.
1312 static ir_node *gen_Builtin(ir_node *node)
1314 ir_builtin_kind kind = get_Builtin_kind(node);
1318 case ir_bk_debugbreak:
1319 case ir_bk_return_address:
1320 case ir_bk_frame_address:
1321 case ir_bk_prefetch:
1325 return gen_clz(node);
1328 case ir_bk_popcount:
1332 case ir_bk_inner_trampoline:
1335 panic("Builtin %s not implemented in ARM", get_builtin_kind_name(kind));
1339 * Transform Proj(Builtin) node.
1341 static ir_node *gen_Proj_Builtin(ir_node *proj)
1343 ir_node *node = get_Proj_pred(proj);
1344 ir_node *new_node = be_transform_node(node);
1345 ir_builtin_kind kind = get_Builtin_kind(node);
1348 case ir_bk_return_address:
1349 case ir_bk_frame_address:
1354 case ir_bk_popcount:
1356 assert(get_Proj_proj(proj) == pn_Builtin_1_result);
1359 case ir_bk_debugbreak:
1360 case ir_bk_prefetch:
1362 assert(get_Proj_proj(proj) == pn_Builtin_M);
1365 case ir_bk_inner_trampoline:
1368 panic("Builtin %s not implemented in ARM", get_builtin_kind_name(kind));
1371 static ir_node *gen_Proj_Load(ir_node *node)
1373 ir_node *load = get_Proj_pred(node);
1374 ir_node *new_load = be_transform_node(load);
1375 dbg_info *dbgi = get_irn_dbg_info(node);
1376 long proj = get_Proj_proj(node);
1378 /* renumber the proj */
1379 switch (get_arm_irn_opcode(new_load)) {
1381 /* handle all gp loads equal: they have the same proj numbers. */
1382 if (proj == pn_Load_res) {
1383 return new_rd_Proj(dbgi, new_load, mode_Iu, pn_arm_Ldr_res);
1384 } else if (proj == pn_Load_M) {
1385 return new_rd_Proj(dbgi, new_load, mode_M, pn_arm_Ldr_M);
1389 if (proj == pn_Load_res) {
1390 ir_mode *mode = get_Load_mode(load);
1391 return new_rd_Proj(dbgi, new_load, mode, pn_arm_Ldf_res);
1392 } else if (proj == pn_Load_M) {
1393 return new_rd_Proj(dbgi, new_load, mode_M, pn_arm_Ldf_M);
1399 panic("Unsupported Proj from Load");
1402 static ir_node *gen_Proj_CopyB(ir_node *node)
1404 ir_node *pred = get_Proj_pred(node);
1405 ir_node *new_pred = be_transform_node(pred);
1406 dbg_info *dbgi = get_irn_dbg_info(node);
1407 long proj = get_Proj_proj(node);
1411 if (is_arm_CopyB(new_pred)) {
1412 return new_rd_Proj(dbgi, new_pred, mode_M, pn_arm_CopyB_M);
1418 panic("Unsupported Proj from CopyB");
1421 static ir_node *gen_Proj_Div(ir_node *node)
1423 ir_node *pred = get_Proj_pred(node);
1424 ir_node *new_pred = be_transform_node(pred);
1425 dbg_info *dbgi = get_irn_dbg_info(node);
1426 ir_mode *mode = get_irn_mode(node);
1427 long proj = get_Proj_proj(node);
1431 return new_rd_Proj(dbgi, new_pred, mode_M, pn_arm_Dvf_M);
1433 return new_rd_Proj(dbgi, new_pred, mode, pn_arm_Dvf_res);
1437 panic("Unsupported Proj from Div");
1441 * Transform the Projs from a Cmp.
1443 static ir_node *gen_Proj_Cmp(ir_node *node)
1446 /* we should only be here in case of a Mux node */
1450 static ir_node *gen_Proj_Start(ir_node *node)
1452 ir_node *block = get_nodes_block(node);
1453 ir_node *new_block = be_transform_node(block);
1454 long proj = get_Proj_proj(node);
1456 switch ((pn_Start) proj) {
1457 case pn_Start_X_initial_exec:
1458 /* we exchange the ProjX with a jump */
1459 return new_bd_arm_Jmp(NULL, new_block);
1462 return be_prolog_get_memory(abihelper);
1464 case pn_Start_T_args:
1465 return new_r_Bad(get_irn_irg(block), mode_T);
1467 case pn_Start_P_frame_base:
1468 return be_prolog_get_reg_value(abihelper, sp_reg);
1470 panic("unexpected start proj: %ld\n", proj);
1473 static ir_node *gen_Proj_Proj_Start(ir_node *node)
1475 long pn = get_Proj_proj(node);
1476 ir_node *block = get_nodes_block(node);
1477 ir_node *new_block = be_transform_node(block);
1478 ir_entity *entity = get_irg_entity(current_ir_graph);
1479 ir_type *method_type = get_entity_type(entity);
1480 ir_type *param_type = get_method_param_type(method_type, pn);
1481 const reg_or_stackslot_t *param;
1483 /* Proj->Proj->Start must be a method argument */
1484 assert(get_Proj_proj(get_Proj_pred(node)) == pn_Start_T_args);
1486 param = &cconv->parameters[pn];
1488 if (param->reg0 != NULL) {
1489 /* argument transmitted in register */
1490 ir_mode *mode = get_type_mode(param_type);
1491 ir_node *value = be_prolog_get_reg_value(abihelper, param->reg0);
1493 if (mode_is_float(mode)) {
1494 ir_node *value1 = NULL;
1496 if (param->reg1 != NULL) {
1497 value1 = be_prolog_get_reg_value(abihelper, param->reg1);
1498 } else if (param->entity != NULL) {
1499 ir_graph *irg = get_irn_irg(node);
1500 ir_node *fp = get_irg_frame(irg);
1501 ir_node *mem = be_prolog_get_memory(abihelper);
1502 ir_node *ldr = new_bd_arm_Ldr(NULL, new_block, fp, mem,
1503 mode_gp, param->entity,
1505 value1 = new_r_Proj(ldr, mode_gp, pn_arm_Ldr_res);
1508 /* convert integer value to float */
1509 if (value1 == NULL) {
1510 value = int_to_float(NULL, new_block, value);
1512 value = ints_to_double(NULL, new_block, value, value1);
1517 /* argument transmitted on stack */
1518 ir_graph *irg = get_irn_irg(node);
1519 ir_node *fp = get_irg_frame(irg);
1520 ir_node *mem = be_prolog_get_memory(abihelper);
1521 ir_mode *mode = get_type_mode(param->type);
1525 if (mode_is_float(mode)) {
1526 load = new_bd_arm_Ldf(NULL, new_block, fp, mem, mode,
1527 param->entity, 0, 0, true);
1528 value = new_r_Proj(load, mode_fp, pn_arm_Ldf_res);
1530 load = new_bd_arm_Ldr(NULL, new_block, fp, mem, mode,
1531 param->entity, 0, 0, true);
1532 value = new_r_Proj(load, mode_gp, pn_arm_Ldr_res);
1534 set_irn_pinned(load, op_pin_state_floats);
1541 * Finds number of output value of a mode_T node which is constrained to
1542 * a single specific register.
1544 static int find_out_for_reg(ir_node *node, const arch_register_t *reg)
1546 int n_outs = arch_get_irn_n_outs(node);
1549 for (o = 0; o < n_outs; ++o) {
1550 const arch_register_req_t *req = arch_get_irn_register_req_out(node, o);
1551 if (req == reg->single_req)
1557 static ir_node *gen_Proj_Proj_Call(ir_node *node)
1559 long pn = get_Proj_proj(node);
1560 ir_node *call = get_Proj_pred(get_Proj_pred(node));
1561 ir_node *new_call = be_transform_node(call);
1562 ir_type *function_type = get_Call_type(call);
1563 calling_convention_t *cconv
1564 = arm_decide_calling_convention(NULL, function_type);
1565 const reg_or_stackslot_t *res = &cconv->results[pn];
1569 /* TODO 64bit modes */
1570 assert(res->reg0 != NULL && res->reg1 == NULL);
1571 regn = find_out_for_reg(new_call, res->reg0);
1573 panic("Internal error in calling convention for return %+F", node);
1575 mode = res->reg0->reg_class->mode;
1577 arm_free_calling_convention(cconv);
1579 return new_r_Proj(new_call, mode, regn);
1582 static ir_node *gen_Proj_Call(ir_node *node)
1584 long pn = get_Proj_proj(node);
1585 ir_node *call = get_Proj_pred(node);
1586 ir_node *new_call = be_transform_node(call);
1588 switch ((pn_Call) pn) {
1590 return new_r_Proj(new_call, mode_M, 0);
1591 case pn_Call_X_regular:
1592 case pn_Call_X_except:
1593 case pn_Call_T_result:
1596 panic("Unexpected Call proj %ld\n", pn);
1600 * Transform a Proj node.
1602 static ir_node *gen_Proj(ir_node *node)
1604 ir_node *pred = get_Proj_pred(node);
1605 long proj = get_Proj_proj(node);
1607 switch (get_irn_opcode(pred)) {
1609 if (proj == pn_Store_M) {
1610 return be_transform_node(pred);
1612 panic("Unsupported Proj from Store");
1615 return gen_Proj_Load(node);
1617 return gen_Proj_Call(node);
1619 return gen_Proj_CopyB(node);
1621 return gen_Proj_Div(node);
1623 return gen_Proj_Cmp(node);
1625 return gen_Proj_Start(node);
1628 return be_duplicate_node(node);
1630 ir_node *pred_pred = get_Proj_pred(pred);
1631 if (is_Call(pred_pred)) {
1632 return gen_Proj_Proj_Call(node);
1633 } else if (is_Start(pred_pred)) {
1634 return gen_Proj_Proj_Start(node);
1639 return gen_Proj_Builtin(node);
1641 panic("code selection didn't expect Proj after %+F\n", pred);
1645 typedef ir_node *(*create_const_node_func)(dbg_info *db, ir_node *block);
1647 static inline ir_node *create_const(ir_graph *irg, ir_node **place,
1648 create_const_node_func func,
1649 const arch_register_t* reg)
1651 ir_node *block, *res;
1656 block = get_irg_start_block(irg);
1657 res = func(NULL, block);
1658 arch_set_irn_register(res, reg);
1663 static ir_node *gen_Unknown(ir_node *node)
1665 ir_node *block = get_nodes_block(node);
1666 ir_node *new_block = be_transform_node(block);
1667 dbg_info *dbgi = get_irn_dbg_info(node);
1669 /* just produce a 0 */
1670 ir_mode *mode = get_irn_mode(node);
1671 if (mode_is_float(mode)) {
1672 ir_tarval *tv = get_mode_null(mode);
1673 ir_node *fconst = new_bd_arm_fConst(dbgi, new_block, tv);
1675 } else if (mode_needs_gp_reg(mode)) {
1676 return create_const_graph_value(dbgi, new_block, 0);
1679 panic("Unexpected Unknown mode");
1683 * Produces the type which sits between the stack args and the locals on the
1684 * stack. It will contain the return address and space to store the old base
1686 * @return The Firm type modeling the ABI between type.
1688 static ir_type *arm_get_between_type(void)
1690 static ir_type *between_type = NULL;
1692 if (between_type == NULL) {
1693 between_type = new_type_class(new_id_from_str("arm_between_type"));
1694 set_type_size_bytes(between_type, 0);
1697 return between_type;
1700 static void create_stacklayout(ir_graph *irg)
1702 ir_entity *entity = get_irg_entity(irg);
1703 ir_type *function_type = get_entity_type(entity);
1704 be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
1709 /* calling conventions must be decided by now */
1710 assert(cconv != NULL);
1712 /* construct argument type */
1713 arg_type = new_type_struct(id_mangle_u(get_entity_ident(entity), new_id_from_chars("arg_type", 8)));
1714 n_params = get_method_n_params(function_type);
1715 for (p = 0; p < n_params; ++p) {
1716 reg_or_stackslot_t *param = &cconv->parameters[p];
1720 if (param->type == NULL)
1723 snprintf(buf, sizeof(buf), "param_%d", p);
1724 id = new_id_from_str(buf);
1725 param->entity = new_entity(arg_type, id, param->type);
1726 set_entity_offset(param->entity, param->offset);
1729 /* TODO: what about external functions? we don't know most of the stack
1730 * layout for them. And probably don't need all of this... */
1731 memset(layout, 0, sizeof(*layout));
1733 layout->frame_type = get_irg_frame_type(irg);
1734 layout->between_type = arm_get_between_type();
1735 layout->arg_type = arg_type;
1736 layout->param_map = NULL; /* TODO */
1737 layout->initial_offset = 0;
1738 layout->initial_bias = 0;
1739 layout->sp_relative = true;
1741 assert(N_FRAME_TYPES == 3);
1742 layout->order[0] = layout->frame_type;
1743 layout->order[1] = layout->between_type;
1744 layout->order[2] = layout->arg_type;
1748 * transform the start node to the prolog code
1750 static ir_node *gen_Start(ir_node *node)
1752 ir_graph *irg = get_irn_irg(node);
1753 ir_entity *entity = get_irg_entity(irg);
1754 ir_type *function_type = get_entity_type(entity);
1755 ir_node *block = get_nodes_block(node);
1756 ir_node *new_block = be_transform_node(block);
1757 dbg_info *dbgi = get_irn_dbg_info(node);
1761 /* stackpointer is important at function prolog */
1762 be_prolog_add_reg(abihelper, sp_reg,
1763 arch_register_req_type_produces_sp | arch_register_req_type_ignore);
1764 /* function parameters in registers */
1765 for (i = 0; i < get_method_n_params(function_type); ++i) {
1766 const reg_or_stackslot_t *param = &cconv->parameters[i];
1767 if (param->reg0 != NULL)
1768 be_prolog_add_reg(abihelper, param->reg0, arch_register_req_type_none);
1769 if (param->reg1 != NULL)
1770 be_prolog_add_reg(abihelper, param->reg1, arch_register_req_type_none);
1772 /* announce that we need the values of the callee save regs */
1773 for (i = 0; i < (sizeof(callee_saves)/sizeof(callee_saves[0])); ++i) {
1774 be_prolog_add_reg(abihelper, callee_saves[i], arch_register_req_type_none);
1777 start = be_prolog_create_start(abihelper, dbgi, new_block);
1781 static ir_node *get_stack_pointer_for(ir_node *node)
1783 /* get predecessor in stack_order list */
1784 ir_node *stack_pred = be_get_stack_pred(stackorder, node);
1787 if (stack_pred == NULL) {
1788 /* first stack user in the current block. We can simply use the
1789 * initial sp_proj for it */
1790 ir_node *sp_proj = be_prolog_get_reg_value(abihelper, sp_reg);
1794 be_transform_node(stack_pred);
1795 stack = (ir_node*)pmap_get(node_to_stack, stack_pred);
1796 if (stack == NULL) {
1797 return get_stack_pointer_for(stack_pred);
1804 * transform a Return node into epilogue code + return statement
1806 static ir_node *gen_Return(ir_node *node)
1808 ir_node *block = get_nodes_block(node);
1809 ir_node *new_block = be_transform_node(block);
1810 dbg_info *dbgi = get_irn_dbg_info(node);
1811 ir_node *mem = get_Return_mem(node);
1812 ir_node *new_mem = be_transform_node(mem);
1813 int n_callee_saves = sizeof(callee_saves)/sizeof(callee_saves[0]);
1814 ir_node *sp_proj = get_stack_pointer_for(node);
1815 int n_res = get_Return_n_ress(node);
1819 be_epilog_begin(abihelper);
1820 be_epilog_set_memory(abihelper, new_mem);
1821 /* connect stack pointer with initial stack pointer. fix_stack phase
1822 will later serialize all stack pointer adjusting nodes */
1823 be_epilog_add_reg(abihelper, sp_reg,
1824 arch_register_req_type_produces_sp | arch_register_req_type_ignore,
1828 for (i = 0; i < n_res; ++i) {
1829 ir_node *res_value = get_Return_res(node, i);
1830 ir_node *new_res_value = be_transform_node(res_value);
1831 const reg_or_stackslot_t *slot = &cconv->results[i];
1832 const arch_register_t *reg = slot->reg0;
1833 assert(slot->reg1 == NULL);
1834 be_epilog_add_reg(abihelper, reg, arch_register_req_type_none, new_res_value);
1837 /* connect callee saves with their values at the function begin */
1838 for (i = 0; i < n_callee_saves; ++i) {
1839 const arch_register_t *reg = callee_saves[i];
1840 ir_node *value = be_prolog_get_reg_value(abihelper, reg);
1841 be_epilog_add_reg(abihelper, reg, arch_register_req_type_none, value);
1844 /* epilog code: an incsp */
1845 bereturn = be_epilog_create_return(abihelper, dbgi, new_block);
1850 static ir_node *gen_Call(ir_node *node)
1852 ir_graph *irg = get_irn_irg(node);
1853 ir_node *callee = get_Call_ptr(node);
1854 ir_node *block = get_nodes_block(node);
1855 ir_node *new_block = be_transform_node(block);
1856 ir_node *mem = get_Call_mem(node);
1857 ir_node *new_mem = be_transform_node(mem);
1858 dbg_info *dbgi = get_irn_dbg_info(node);
1859 ir_type *type = get_Call_type(node);
1860 calling_convention_t *cconv = arm_decide_calling_convention(NULL, type);
1861 size_t n_params = get_Call_n_params(node);
1862 size_t n_param_regs = cconv->n_reg_params;
1863 /* max inputs: memory, callee, register arguments */
1864 int max_inputs = 2 + n_param_regs;
1865 ir_node **in = ALLOCAN(ir_node*, max_inputs);
1866 ir_node **sync_ins = ALLOCAN(ir_node*, max_inputs);
1867 struct obstack *obst = be_get_be_obst(irg);
1868 const arch_register_req_t **in_req
1869 = OALLOCNZ(obst, const arch_register_req_t*, max_inputs);
1873 = sizeof(caller_saves)/sizeof(caller_saves[0]);
1874 ir_entity *entity = NULL;
1875 ir_node *incsp = NULL;
1882 assert(n_params == get_method_n_params(type));
1884 /* construct arguments */
1887 in_req[in_arity] = arch_no_register_req;
1891 for (p = 0; p < n_params; ++p) {
1892 ir_node *value = get_Call_param(node, p);
1893 ir_node *new_value = be_transform_node(value);
1894 ir_node *new_value1 = NULL;
1895 const reg_or_stackslot_t *param = &cconv->parameters[p];
1896 ir_type *param_type = get_method_param_type(type, p);
1897 ir_mode *mode = get_type_mode(param_type);
1900 if (mode_is_float(mode) && param->reg0 != NULL) {
1901 unsigned size_bits = get_mode_size_bits(mode);
1902 if (size_bits == 64) {
1903 double_to_ints(dbgi, new_block, new_value, &new_value,
1906 assert(size_bits == 32);
1907 new_value = float_to_int(dbgi, new_block, new_value);
1911 /* put value into registers */
1912 if (param->reg0 != NULL) {
1913 in[in_arity] = new_value;
1914 in_req[in_arity] = param->reg0->single_req;
1916 if (new_value1 == NULL)
1919 if (param->reg1 != NULL) {
1920 assert(new_value1 != NULL);
1921 in[in_arity] = new_value1;
1922 in_req[in_arity] = param->reg1->single_req;
1927 /* we need a store if we're here */
1928 if (new_value1 != NULL) {
1929 new_value = new_value1;
1933 /* create a parameter frame if necessary */
1934 if (incsp == NULL) {
1935 ir_node *new_frame = get_stack_pointer_for(node);
1936 incsp = be_new_IncSP(sp_reg, new_block, new_frame,
1937 cconv->param_stack_size, 1);
1939 if (mode_is_float(mode)) {
1940 str = new_bd_arm_Stf(dbgi, new_block, incsp, new_value, new_mem,
1941 mode, NULL, 0, param->offset, true);
1943 str = new_bd_arm_Str(dbgi, new_block, incsp, new_value, new_mem,
1944 mode, NULL, 0, param->offset, true);
1946 sync_ins[sync_arity++] = str;
1948 assert(in_arity <= max_inputs);
1950 /* construct memory input */
1951 if (sync_arity == 0) {
1952 in[mem_pos] = new_mem;
1953 } else if (sync_arity == 1) {
1954 in[mem_pos] = sync_ins[0];
1956 in[mem_pos] = new_rd_Sync(NULL, new_block, sync_arity, sync_ins);
1959 /* TODO: use a generic symconst matcher here */
1960 if (is_SymConst(callee)) {
1961 entity = get_SymConst_entity(callee);
1963 /* TODO: finish load matcher here */
1966 if (is_Proj(callee) && is_Load(get_Proj_pred(callee))) {
1967 ir_node *load = get_Proj_pred(callee);
1968 ir_node *ptr = get_Load_ptr(load);
1969 ir_node *new_ptr = be_transform_node(ptr);
1970 ir_node *mem = get_Load_mem(load);
1971 ir_node *new_mem = be_transform_node(mem);
1972 ir_mode *mode = get_Load_mode(node);
1976 in[in_arity] = be_transform_node(callee);
1977 in_req[in_arity] = arm_reg_classes[CLASS_arm_gp].class_req;
1986 out_arity = 1 + n_caller_saves;
1988 if (entity != NULL) {
1989 /* TODO: use a generic symconst matcher here
1990 * so we can also handle entity+offset, etc. */
1991 res = new_bd_arm_Bl(dbgi, new_block, in_arity, in, out_arity,entity, 0);
1994 * - use a proper shifter_operand matcher
1995 * - we could also use LinkLdrPC
1997 res = new_bd_arm_LinkMovPC(dbgi, new_block, in_arity, in, out_arity,
2001 if (incsp != NULL) {
2002 /* IncSP to destroy the call stackframe */
2003 incsp = be_new_IncSP(sp_reg, new_block, incsp, -cconv->param_stack_size,
2005 /* if we are the last IncSP producer in a block then we have to keep
2007 * Note: This here keeps all producers which is more than necessary */
2008 add_irn_dep(incsp, res);
2011 pmap_insert(node_to_stack, node, incsp);
2014 arch_set_irn_register_reqs_in(res, in_req);
2016 /* create output register reqs */
2017 arch_set_irn_register_req_out(res, 0, arch_no_register_req);
2018 for (o = 0; o < n_caller_saves; ++o) {
2019 const arch_register_t *reg = caller_saves[o];
2020 arch_set_irn_register_req_out(res, o+1, reg->single_req);
2023 /* copy pinned attribute */
2024 set_irn_pinned(res, get_irn_pinned(node));
2026 arm_free_calling_convention(cconv);
2030 static ir_node *gen_Sel(ir_node *node)
2032 dbg_info *dbgi = get_irn_dbg_info(node);
2033 ir_node *block = get_nodes_block(node);
2034 ir_node *new_block = be_transform_node(block);
2035 ir_node *ptr = get_Sel_ptr(node);
2036 ir_node *new_ptr = be_transform_node(ptr);
2037 ir_entity *entity = get_Sel_entity(node);
2039 /* must be the frame pointer all other sels must have been lowered
2041 assert(is_Proj(ptr) && is_Start(get_Proj_pred(ptr)));
2043 return new_bd_arm_FrameAddr(dbgi, new_block, new_ptr, entity, 0);
2047 * Change some phi modes
2049 static ir_node *gen_Phi(ir_node *node)
2051 const arch_register_req_t *req;
2052 ir_node *block = be_transform_node(get_nodes_block(node));
2053 ir_graph *irg = current_ir_graph;
2054 dbg_info *dbgi = get_irn_dbg_info(node);
2055 ir_mode *mode = get_irn_mode(node);
2058 if (mode_needs_gp_reg(mode)) {
2059 /* we shouldn't have any 64bit stuff around anymore */
2060 assert(get_mode_size_bits(mode) <= 32);
2061 /* all integer operations are on 32bit registers now */
2063 req = arm_reg_classes[CLASS_arm_gp].class_req;
2065 req = arch_no_register_req;
2068 /* phi nodes allow loops, so we use the old arguments for now
2069 * and fix this later */
2070 phi = new_ir_node(dbgi, irg, block, op_Phi, mode, get_irn_arity(node),
2071 get_irn_in(node) + 1);
2072 copy_node_attr(irg, node, phi);
2073 be_duplicate_deps(node, phi);
2075 arch_set_irn_register_req_out(phi, 0, req);
2077 be_enqueue_preds(node);
2084 * Enters all transform functions into the generic pointer
2086 static void arm_register_transformers(void)
2088 be_start_transform_setup();
2090 be_set_transform_function(op_Add, gen_Add);
2091 be_set_transform_function(op_And, gen_And);
2092 be_set_transform_function(op_Call, gen_Call);
2093 be_set_transform_function(op_Cmp, gen_Cmp);
2094 be_set_transform_function(op_Cond, gen_Cond);
2095 be_set_transform_function(op_Const, gen_Const);
2096 be_set_transform_function(op_Conv, gen_Conv);
2097 be_set_transform_function(op_CopyB, gen_CopyB);
2098 be_set_transform_function(op_Div, gen_Div);
2099 be_set_transform_function(op_Eor, gen_Eor);
2100 be_set_transform_function(op_Jmp, gen_Jmp);
2101 be_set_transform_function(op_Load, gen_Load);
2102 be_set_transform_function(op_Minus, gen_Minus);
2103 be_set_transform_function(op_Mul, gen_Mul);
2104 be_set_transform_function(op_Not, gen_Not);
2105 be_set_transform_function(op_Or, gen_Or);
2106 be_set_transform_function(op_Phi, gen_Phi);
2107 be_set_transform_function(op_Proj, gen_Proj);
2108 be_set_transform_function(op_Return, gen_Return);
2109 be_set_transform_function(op_Rotl, gen_Rotl);
2110 be_set_transform_function(op_Sel, gen_Sel);
2111 be_set_transform_function(op_Shl, gen_Shl);
2112 be_set_transform_function(op_Shr, gen_Shr);
2113 be_set_transform_function(op_Shrs, gen_Shrs);
2114 be_set_transform_function(op_Start, gen_Start);
2115 be_set_transform_function(op_Store, gen_Store);
2116 be_set_transform_function(op_Sub, gen_Sub);
2117 be_set_transform_function(op_SymConst, gen_SymConst);
2118 be_set_transform_function(op_Unknown, gen_Unknown);
2119 be_set_transform_function(op_Builtin, gen_Builtin);
2123 * Initialize fpa Immediate support.
2125 static void arm_init_fpa_immediate(void)
2127 /* 0, 1, 2, 3, 4, 5, 10, or 0.5. */
2128 fpa_imm[FPA_IMM_FLOAT][fpa_null] = get_mode_null(mode_F);
2129 fpa_imm[FPA_IMM_FLOAT][fpa_one] = get_mode_one(mode_F);
2130 fpa_imm[FPA_IMM_FLOAT][fpa_two] = new_tarval_from_str("2", 1, mode_F);
2131 fpa_imm[FPA_IMM_FLOAT][fpa_three] = new_tarval_from_str("3", 1, mode_F);
2132 fpa_imm[FPA_IMM_FLOAT][fpa_four] = new_tarval_from_str("4", 1, mode_F);
2133 fpa_imm[FPA_IMM_FLOAT][fpa_five] = new_tarval_from_str("5", 1, mode_F);
2134 fpa_imm[FPA_IMM_FLOAT][fpa_ten] = new_tarval_from_str("10", 2, mode_F);
2135 fpa_imm[FPA_IMM_FLOAT][fpa_half] = new_tarval_from_str("0.5", 3, mode_F);
2137 fpa_imm[FPA_IMM_DOUBLE][fpa_null] = get_mode_null(mode_D);
2138 fpa_imm[FPA_IMM_DOUBLE][fpa_one] = get_mode_one(mode_D);
2139 fpa_imm[FPA_IMM_DOUBLE][fpa_two] = new_tarval_from_str("2", 1, mode_D);
2140 fpa_imm[FPA_IMM_DOUBLE][fpa_three] = new_tarval_from_str("3", 1, mode_D);
2141 fpa_imm[FPA_IMM_DOUBLE][fpa_four] = new_tarval_from_str("4", 1, mode_D);
2142 fpa_imm[FPA_IMM_DOUBLE][fpa_five] = new_tarval_from_str("5", 1, mode_D);
2143 fpa_imm[FPA_IMM_DOUBLE][fpa_ten] = new_tarval_from_str("10", 2, mode_D);
2144 fpa_imm[FPA_IMM_DOUBLE][fpa_half] = new_tarval_from_str("0.5", 3, mode_D);
2146 fpa_imm[FPA_IMM_EXTENDED][fpa_null] = get_mode_null(mode_E);
2147 fpa_imm[FPA_IMM_EXTENDED][fpa_one] = get_mode_one(mode_E);
2148 fpa_imm[FPA_IMM_EXTENDED][fpa_two] = new_tarval_from_str("2", 1, mode_E);
2149 fpa_imm[FPA_IMM_EXTENDED][fpa_three] = new_tarval_from_str("3", 1, mode_E);
2150 fpa_imm[FPA_IMM_EXTENDED][fpa_four] = new_tarval_from_str("4", 1, mode_E);
2151 fpa_imm[FPA_IMM_EXTENDED][fpa_five] = new_tarval_from_str("5", 1, mode_E);
2152 fpa_imm[FPA_IMM_EXTENDED][fpa_ten] = new_tarval_from_str("10", 2, mode_E);
2153 fpa_imm[FPA_IMM_EXTENDED][fpa_half] = new_tarval_from_str("0.5", 3, mode_E);
2157 * Transform a Firm graph into an ARM graph.
2159 void arm_transform_graph(ir_graph *irg)
2161 static int imm_initialized = 0;
2162 ir_entity *entity = get_irg_entity(irg);
2163 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2164 ir_type *frame_type;
2169 if (! imm_initialized) {
2170 arm_init_fpa_immediate();
2171 imm_initialized = 1;
2173 arm_register_transformers();
2175 isa = (arm_isa_t*) arch_env;
2177 node_to_stack = pmap_create();
2179 assert(abihelper == NULL);
2180 abihelper = be_abihelper_prepare(irg);
2181 stackorder = be_collect_stacknodes(irg);
2182 assert(cconv == NULL);
2183 cconv = arm_decide_calling_convention(irg, get_entity_type(entity));
2184 create_stacklayout(irg);
2186 be_transform_graph(irg, NULL);
2188 be_abihelper_finish(abihelper);
2190 be_free_stackorder(stackorder);
2193 arm_free_calling_convention(cconv);
2196 frame_type = get_irg_frame_type(irg);
2197 if (get_type_state(frame_type) == layout_undefined) {
2198 default_layout_compound_type(frame_type);
2201 pmap_destroy(node_to_stack);
2202 node_to_stack = NULL;
2204 be_add_missing_keeps(irg);
2207 void arm_init_transform(void)
2209 FIRM_DBG_REGISTER(dbg, "firm.be.arm.transform");