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 This file implements the ia32 node emitter.
23 * @author Christian Wuerdig, Matthias Braun
39 #include "iredges_t.h"
43 #include "raw_bitset.h"
47 #include "../besched.h"
48 #include "../benode.h"
50 #include "../be_dbgout.h"
51 #include "../beemitter.h"
52 #include "../begnuas.h"
54 #include "../be_dbgout.h"
56 #include "ia32_emitter.h"
57 #include "gen_ia32_emitter.h"
58 #include "gen_ia32_regalloc_if.h"
59 #include "ia32_nodes_attr.h"
60 #include "ia32_new_nodes.h"
61 #include "ia32_map_regs.h"
62 #include "ia32_architecture.h"
63 #include "bearch_ia32_t.h"
65 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
67 #define BLOCK_PREFIX ".L"
69 #define SNPRINTF_BUF_LEN 128
71 static const ia32_isa_t *isa;
72 static ia32_code_gen_t *cg;
73 static char pic_base_label[128];
74 static ir_label_t exc_label_id;
75 static int mark_spill_reload = 0;
78 /** Return the next block in Block schedule */
79 static ir_node *get_prev_block_sched(const ir_node *block)
81 return get_irn_link(block);
84 /** Checks if the current block is a fall-through target. */
85 static int is_fallthrough(const ir_node *cfgpred)
89 if (!is_Proj(cfgpred))
91 pred = get_Proj_pred(cfgpred);
92 if (is_ia32_SwitchJmp(pred))
99 * returns non-zero if the given block needs a label
100 * because of being a jump-target (and not a fall-through)
102 static int block_needs_label(const ir_node *block)
105 int n_cfgpreds = get_Block_n_cfgpreds(block);
107 if (has_Block_entity(block))
110 if (n_cfgpreds == 0) {
112 } else if (n_cfgpreds == 1) {
113 ir_node *cfgpred = get_Block_cfgpred(block, 0);
114 ir_node *cfgpred_block = get_nodes_block(cfgpred);
116 if (get_prev_block_sched(block) == cfgpred_block
117 && is_fallthrough(cfgpred)) {
126 * Returns the register at in position pos.
128 static const arch_register_t *get_in_reg(const ir_node *irn, int pos)
131 const arch_register_t *reg = NULL;
133 assert(get_irn_arity(irn) > pos && "Invalid IN position");
135 /* The out register of the operator at position pos is the
136 in register we need. */
137 op = get_irn_n(irn, pos);
139 reg = arch_get_irn_register(op);
141 assert(reg && "no in register found");
143 if (reg == &ia32_gp_regs[REG_GP_NOREG])
144 panic("trying to emit noreg for %+F input %d", irn, pos);
146 /* in case of unknown register: just return a valid register */
147 if (reg == &ia32_gp_regs[REG_GP_UKNWN]) {
148 const arch_register_req_t *req = arch_get_register_req(irn, pos);
150 if (arch_register_req_is(req, limited)) {
151 /* in case of limited requirements: get the first allowed register */
152 unsigned idx = rbitset_next(req->limited, 0, 1);
153 reg = arch_register_for_index(req->cls, idx);
155 /* otherwise get first register in class */
156 reg = arch_register_for_index(req->cls, 0);
164 * Returns the register at out position pos.
166 static const arch_register_t *get_out_reg(const ir_node *irn, int pos)
169 const arch_register_t *reg = NULL;
171 /* 1st case: irn is not of mode_T, so it has only */
172 /* one OUT register -> good */
173 /* 2nd case: irn is of mode_T -> collect all Projs and ask the */
174 /* Proj with the corresponding projnum for the register */
176 if (get_irn_mode(irn) != mode_T) {
178 reg = arch_get_irn_register(irn);
179 } else if (is_ia32_irn(irn)) {
180 reg = arch_irn_get_register(irn, pos);
182 const ir_edge_t *edge;
184 foreach_out_edge(irn, edge) {
185 proj = get_edge_src_irn(edge);
186 assert(is_Proj(proj) && "non-Proj from mode_T node");
187 if (get_Proj_proj(proj) == pos) {
188 reg = arch_get_irn_register(proj);
194 assert(reg && "no out register found");
199 * Add a number to a prefix. This number will not be used a second time.
201 static char *get_unique_label(char *buf, size_t buflen, const char *prefix)
203 static unsigned long id = 0;
204 snprintf(buf, buflen, "%s%lu", prefix, ++id);
208 /*************************************************************
210 * (_) | | / _| | | | |
211 * _ __ _ __ _ _ __ | |_| |_ | |__ ___| |_ __ ___ _ __
212 * | '_ \| '__| | '_ \| __| _| | '_ \ / _ \ | '_ \ / _ \ '__|
213 * | |_) | | | | | | | |_| | | | | | __/ | |_) | __/ |
214 * | .__/|_| |_|_| |_|\__|_| |_| |_|\___|_| .__/ \___|_|
217 *************************************************************/
220 * Emit the name of the 8bit low register
222 static void emit_8bit_register(const arch_register_t *reg)
224 const char *reg_name = arch_register_get_name(reg);
227 be_emit_char(reg_name[1]);
232 * Emit the name of the 8bit high register
234 static void emit_8bit_register_high(const arch_register_t *reg)
236 const char *reg_name = arch_register_get_name(reg);
239 be_emit_char(reg_name[1]);
243 static void emit_16bit_register(const arch_register_t *reg)
245 const char *reg_name = ia32_get_mapped_reg_name(isa->regs_16bit, reg);
248 be_emit_string(reg_name);
252 * emit a register, possible shortened by a mode
254 * @param reg the register
255 * @param mode the mode of the register or NULL for full register
257 static void emit_register(const arch_register_t *reg, const ir_mode *mode)
259 const char *reg_name;
262 int size = get_mode_size_bits(mode);
264 case 8: emit_8bit_register(reg); return;
265 case 16: emit_16bit_register(reg); return;
267 assert(mode_is_float(mode) || size == 32);
270 reg_name = arch_register_get_name(reg);
273 be_emit_string(reg_name);
276 void ia32_emit_source_register(const ir_node *node, int pos)
278 const arch_register_t *reg = get_in_reg(node, pos);
280 emit_register(reg, NULL);
283 static void ia32_emit_entity(ir_entity *entity, int no_pic_adjust)
285 set_entity_backend_marked(entity, 1);
286 be_gas_emit_entity(entity);
288 if (get_entity_owner(entity) == get_tls_type()) {
289 if (get_entity_visibility(entity) == visibility_external_allocated) {
290 be_emit_cstring("@INDNTPOFF");
292 be_emit_cstring("@NTPOFF");
296 if (do_pic && !no_pic_adjust) {
298 be_emit_string(pic_base_label);
302 static void emit_ia32_Immediate_no_prefix(const ir_node *node)
304 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node);
306 if (attr->symconst != NULL) {
309 ia32_emit_entity(attr->symconst, attr->no_pic_adjust);
311 if (attr->symconst == NULL || attr->offset != 0) {
312 if (attr->symconst != NULL) {
313 be_emit_irprintf("%+d", attr->offset);
315 be_emit_irprintf("0x%X", attr->offset);
320 static void emit_ia32_Immediate(const ir_node *node)
323 emit_ia32_Immediate_no_prefix(node);
326 void ia32_emit_8bit_source_register_or_immediate(const ir_node *node, int pos)
328 const arch_register_t *reg;
329 const ir_node *in = get_irn_n(node, pos);
330 if (is_ia32_Immediate(in)) {
331 emit_ia32_Immediate(in);
335 reg = get_in_reg(node, pos);
336 emit_8bit_register(reg);
339 void ia32_emit_8bit_high_source_register(const ir_node *node, int pos)
341 const arch_register_t *reg = get_in_reg(node, pos);
342 emit_8bit_register_high(reg);
345 void ia32_emit_16bit_source_register_or_immediate(const ir_node *node, int pos)
347 const arch_register_t *reg;
348 const ir_node *in = get_irn_n(node, pos);
349 if (is_ia32_Immediate(in)) {
350 emit_ia32_Immediate(in);
354 reg = get_in_reg(node, pos);
355 emit_16bit_register(reg);
358 void ia32_emit_dest_register(const ir_node *node, int pos)
360 const arch_register_t *reg = get_out_reg(node, pos);
362 emit_register(reg, NULL);
365 void ia32_emit_dest_register_size(const ir_node *node, int pos)
367 const arch_register_t *reg = get_out_reg(node, pos);
369 emit_register(reg, get_ia32_ls_mode(node));
372 void ia32_emit_8bit_dest_register(const ir_node *node, int pos)
374 const arch_register_t *reg = get_out_reg(node, pos);
376 emit_register(reg, mode_Bu);
379 void ia32_emit_x87_register(const ir_node *node, int pos)
381 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
385 be_emit_string(attr->x87[pos]->name);
388 static void ia32_emit_mode_suffix_mode(const ir_mode *mode)
390 assert(mode_is_int(mode) || mode_is_reference(mode));
391 switch (get_mode_size_bits(mode)) {
392 case 8: be_emit_char('b'); return;
393 case 16: be_emit_char('w'); return;
394 case 32: be_emit_char('l'); return;
395 /* gas docu says q is the suffix but gcc, objdump and icc use ll
397 case 64: be_emit_cstring("ll"); return;
399 panic("Can't output mode_suffix for %+F", mode);
402 void ia32_emit_mode_suffix(const ir_node *node)
404 ir_mode *mode = get_ia32_ls_mode(node);
408 ia32_emit_mode_suffix_mode(mode);
411 void ia32_emit_x87_mode_suffix(const ir_node *node)
415 /* we only need to emit the mode on address mode */
416 if (get_ia32_op_type(node) == ia32_Normal)
419 mode = get_ia32_ls_mode(node);
420 assert(mode != NULL);
422 if (mode_is_float(mode)) {
423 switch (get_mode_size_bits(mode)) {
424 case 32: be_emit_char('s'); return;
425 case 64: be_emit_char('l'); return;
427 case 96: be_emit_char('t'); return;
430 assert(mode_is_int(mode));
431 switch (get_mode_size_bits(mode)) {
432 case 16: be_emit_char('s'); return;
433 case 32: be_emit_char('l'); return;
434 /* gas docu says q is the suffix but gcc, objdump and icc use ll
436 case 64: be_emit_cstring("ll"); return;
439 panic("Can't output mode_suffix for %+F", mode);
442 static char get_xmm_mode_suffix(ir_mode *mode)
444 assert(mode_is_float(mode));
445 switch(get_mode_size_bits(mode)) {
448 default: panic("Invalid XMM mode");
452 void ia32_emit_xmm_mode_suffix(const ir_node *node)
454 ir_mode *mode = get_ia32_ls_mode(node);
455 assert(mode != NULL);
457 be_emit_char(get_xmm_mode_suffix(mode));
460 void ia32_emit_xmm_mode_suffix_s(const ir_node *node)
462 ir_mode *mode = get_ia32_ls_mode(node);
463 assert(mode != NULL);
464 be_emit_char(get_xmm_mode_suffix(mode));
467 void ia32_emit_extend_suffix(const ir_node *node)
469 ir_mode *mode = get_ia32_ls_mode(node);
470 if (get_mode_size_bits(mode) == 32)
472 be_emit_char(mode_is_signed(mode) ? 's' : 'z');
473 ia32_emit_mode_suffix_mode(mode);
476 void ia32_emit_source_register_or_immediate(const ir_node *node, int pos)
478 ir_node *in = get_irn_n(node, pos);
479 if (is_ia32_Immediate(in)) {
480 emit_ia32_Immediate(in);
482 const ir_mode *mode = get_ia32_ls_mode(node);
483 const arch_register_t *reg = get_in_reg(node, pos);
484 emit_register(reg, mode);
489 * Returns the target block for a control flow node.
491 static ir_node *get_cfop_target_block(const ir_node *irn)
493 assert(get_irn_mode(irn) == mode_X);
494 return get_irn_link(irn);
498 * Emits a block label for the given block.
500 static void ia32_emit_block_name(const ir_node *block)
502 if (has_Block_entity(block)) {
503 ir_entity *entity = get_Block_entity(block);
504 be_gas_emit_entity(entity);
506 be_emit_cstring(BLOCK_PREFIX);
507 be_emit_irprintf("%ld", get_irn_node_nr(block));
512 * Emits the target label for a control flow node.
514 static void ia32_emit_cfop_target(const ir_node *node)
516 ir_node *block = get_cfop_target_block(node);
517 ia32_emit_block_name(block);
521 * positive conditions for signed compares
523 static const char *const cmp2condition_s[] = {
524 NULL, /* always false */
531 NULL /* always true */
535 * positive conditions for unsigned compares
537 static const char *const cmp2condition_u[] = {
538 NULL, /* always false */
545 NULL /* always true */
549 * Emit the suffix for a compare instruction.
551 static void ia32_emit_cmp_suffix(int pnc)
555 if (pnc == ia32_pn_Cmp_parity) {
559 if (pnc & ia32_pn_Cmp_float || pnc & ia32_pn_Cmp_unsigned) {
560 str = cmp2condition_u[pnc & 7];
562 str = cmp2condition_s[pnc & 7];
568 typedef enum ia32_emit_mod_t {
569 EMIT_RESPECT_LS = 1U << 0,
570 EMIT_ALTERNATE_AM = 1U << 1,
575 * Emits address mode.
577 void ia32_emit_am(const ir_node *node)
579 ir_entity *ent = get_ia32_am_sc(node);
580 int offs = get_ia32_am_offs_int(node);
581 ir_node *base = get_irn_n(node, n_ia32_base);
582 int has_base = !is_ia32_NoReg_GP(base);
583 ir_node *index = get_irn_n(node, n_ia32_index);
584 int has_index = !is_ia32_NoReg_GP(index);
586 /* just to be sure... */
587 assert(!is_ia32_use_frame(node) || get_ia32_frame_ent(node) != NULL);
591 const ia32_attr_t *attr = get_ia32_attr_const(node);
592 if (is_ia32_am_sc_sign(node))
594 ia32_emit_entity(ent, attr->data.am_sc_no_pic_adjust);
597 /* also handle special case if nothing is set */
598 if (offs != 0 || (ent == NULL && !has_base && !has_index)) {
600 be_emit_irprintf("%+d", offs);
602 be_emit_irprintf("%d", offs);
606 if (has_base || has_index) {
611 const arch_register_t *reg = get_in_reg(node, n_ia32_base);
612 emit_register(reg, NULL);
615 /* emit index + scale */
617 const arch_register_t *reg = get_in_reg(node, n_ia32_index);
620 emit_register(reg, NULL);
622 scale = get_ia32_am_scale(node);
624 be_emit_irprintf(",%d", 1 << scale);
632 * fmt parameter output
633 * ---- ---------------------- ---------------------------------------------
635 * %AM <node> address mode of the node
636 * %AR const arch_register_t* address mode of the node or register
637 * %ASx <node> address mode of the node or source register x
638 * %Dx <node> destination register x
639 * %I <node> immediate of the node
640 * %L <node> control flow target of the node
641 * %M <node> mode suffix of the node
642 * %P int condition code
643 * %R const arch_register_t* register
644 * %Sx <node> source register x
645 * %s const char* string
646 * %u unsigned int unsigned int
647 * %d signed int signed int
650 * # modifier for %ASx, %D and %S uses ls mode of node to alter register width
651 * * modifier does not prefix immediates with $, but AM with *
652 * l modifier for %lu and %ld
654 static void ia32_emitf(const ir_node *node, const char *fmt, ...)
660 const char *start = fmt;
661 ia32_emit_mod_t mod = 0;
663 while (*fmt != '%' && *fmt != '\n' && *fmt != '\0')
666 be_emit_string_len(start, fmt - start);
670 be_emit_finish_line_gas(node);
682 mod |= EMIT_ALTERNATE_AM;
687 mod |= EMIT_RESPECT_LS;
704 if (mod & EMIT_ALTERNATE_AM)
711 const arch_register_t *reg = va_arg(ap, const arch_register_t*);
712 if (mod & EMIT_ALTERNATE_AM)
714 if (get_ia32_op_type(node) == ia32_AddrModeS) {
717 emit_register(reg, NULL);
723 if (get_ia32_op_type(node) == ia32_AddrModeS) {
724 if (mod & EMIT_ALTERNATE_AM)
729 assert(get_ia32_op_type(node) == ia32_Normal);
734 default: goto unknown;
741 const arch_register_t *reg;
743 if (*fmt < '0' || '9' <= *fmt)
747 reg = get_out_reg(node, pos);
748 emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL);
753 if (!(mod & EMIT_ALTERNATE_AM))
755 emit_ia32_Immediate_no_prefix(node);
759 ia32_emit_cfop_target(node);
763 ia32_emit_mode_suffix_mode(get_ia32_ls_mode(node));
768 int pnc = va_arg(ap, int);
769 ia32_emit_cmp_suffix(pnc);
774 const arch_register_t *reg = va_arg(ap, const arch_register_t*);
775 emit_register(reg, NULL);
784 if (*fmt < '0' || '9' <= *fmt)
788 in = get_irn_n(node, pos);
789 if (is_ia32_Immediate(in)) {
790 if (!(mod & EMIT_ALTERNATE_AM))
792 emit_ia32_Immediate_no_prefix(in);
794 const arch_register_t *reg;
796 if (mod & EMIT_ALTERNATE_AM)
798 reg = get_in_reg(node, pos);
799 emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL);
805 const char *str = va_arg(ap, const char*);
811 if (mod & EMIT_LONG) {
812 unsigned long num = va_arg(ap, unsigned long);
813 be_emit_irprintf("%lu", num);
815 unsigned num = va_arg(ap, unsigned);
816 be_emit_irprintf("%u", num);
821 if (mod & EMIT_LONG) {
822 long num = va_arg(ap, long);
823 be_emit_irprintf("%ld", num);
825 int num = va_arg(ap, int);
826 be_emit_irprintf("%d", num);
832 panic("unknown format conversion in ia32_emitf()");
840 * Emits registers and/or address mode of a binary operation.
842 void ia32_emit_binop(const ir_node *node)
844 if (is_ia32_Immediate(get_irn_n(node, n_ia32_binary_right))) {
845 ia32_emitf(node, "%#S4, %#AS3");
847 ia32_emitf(node, "%#AS4, %#S3");
852 * Emits registers and/or address mode of a binary operation.
854 void ia32_emit_x87_binop(const ir_node *node)
856 switch(get_ia32_op_type(node)) {
859 const ia32_x87_attr_t *x87_attr = get_ia32_x87_attr_const(node);
860 const arch_register_t *in1 = x87_attr->x87[0];
861 const arch_register_t *in = x87_attr->x87[1];
862 const arch_register_t *out = x87_attr->x87[2];
866 } else if (out == in) {
871 be_emit_string(arch_register_get_name(in));
872 be_emit_cstring(", %");
873 be_emit_string(arch_register_get_name(out));
881 assert(0 && "unsupported op type");
886 * Emits registers and/or address mode of a unary operation.
888 void ia32_emit_unop(const ir_node *node, int pos)
892 ia32_emitf(node, fmt);
895 static void emit_ia32_IMul(const ir_node *node)
897 ir_node *left = get_irn_n(node, n_ia32_IMul_left);
898 const arch_register_t *out_reg = get_out_reg(node, pn_ia32_IMul_res);
900 /* do we need the 3-address form? */
901 if (is_ia32_NoReg_GP(left) ||
902 get_in_reg(node, n_ia32_IMul_left) != out_reg) {
903 ia32_emitf(node, "\timul%M %#S4, %#AS3, %#D0\n");
905 ia32_emitf(node, "\timul%M %#AS4, %#S3\n");
910 * walks up a tree of copies/perms/spills/reloads to find the original value
911 * that is moved around
913 static ir_node *find_original_value(ir_node *node)
915 if (irn_visited(node))
918 mark_irn_visited(node);
919 if (be_is_Copy(node)) {
920 return find_original_value(be_get_Copy_op(node));
921 } else if (be_is_CopyKeep(node)) {
922 return find_original_value(be_get_CopyKeep_op(node));
923 } else if (is_Proj(node)) {
924 ir_node *pred = get_Proj_pred(node);
925 if (be_is_Perm(pred)) {
926 return find_original_value(get_irn_n(pred, get_Proj_proj(node)));
927 } else if (be_is_MemPerm(pred)) {
928 return find_original_value(get_irn_n(pred, get_Proj_proj(node) + 1));
929 } else if (is_ia32_Load(pred)) {
930 return find_original_value(get_irn_n(pred, n_ia32_Load_mem));
934 } else if (is_ia32_Store(node)) {
935 return find_original_value(get_irn_n(node, n_ia32_Store_val));
936 } else if (is_Phi(node)) {
938 arity = get_irn_arity(node);
939 for (i = 0; i < arity; ++i) {
940 ir_node *in = get_irn_n(node, i);
941 ir_node *res = find_original_value(in);
952 static int determine_final_pnc(const ir_node *node, int flags_pos,
955 ir_node *flags = get_irn_n(node, flags_pos);
956 const ia32_attr_t *flags_attr;
957 flags = skip_Proj(flags);
959 if (is_ia32_Sahf(flags)) {
960 ir_node *cmp = get_irn_n(flags, n_ia32_Sahf_val);
961 if (!(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp)
962 || is_ia32_FucomppFnstsw(cmp) || is_ia32_FtstFnstsw(cmp))) {
963 inc_irg_visited(current_ir_graph);
964 cmp = find_original_value(cmp);
966 assert(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp)
967 || is_ia32_FucomppFnstsw(cmp) || is_ia32_FtstFnstsw(cmp));
970 flags_attr = get_ia32_attr_const(cmp);
971 if (flags_attr->data.ins_permuted)
972 pnc = get_mirrored_pnc(pnc);
973 pnc |= ia32_pn_Cmp_float;
974 } else if (is_ia32_Ucomi(flags) || is_ia32_Fucomi(flags)
975 || is_ia32_Fucompi(flags)) {
976 flags_attr = get_ia32_attr_const(flags);
978 if (flags_attr->data.ins_permuted)
979 pnc = get_mirrored_pnc(pnc);
980 pnc |= ia32_pn_Cmp_float;
982 flags_attr = get_ia32_attr_const(flags);
984 if (flags_attr->data.ins_permuted)
985 pnc = get_mirrored_pnc(pnc);
986 if (flags_attr->data.cmp_unsigned)
987 pnc |= ia32_pn_Cmp_unsigned;
993 static pn_Cmp ia32_get_negated_pnc(pn_Cmp pnc)
995 ir_mode *mode = pnc & ia32_pn_Cmp_float ? mode_F : mode_Iu;
996 return get_negated_pnc(pnc, mode);
999 void ia32_emit_cmp_suffix_node(const ir_node *node,
1002 const ia32_attr_t *attr = get_ia32_attr_const(node);
1004 pn_Cmp pnc = get_ia32_condcode(node);
1006 pnc = determine_final_pnc(node, flags_pos, pnc);
1007 if (attr->data.ins_permuted)
1008 pnc = ia32_get_negated_pnc(pnc);
1010 ia32_emit_cmp_suffix(pnc);
1014 * Emits an exception label for a given node.
1016 static void ia32_emit_exc_label(const ir_node *node)
1018 be_emit_string(be_gas_insn_label_prefix());
1019 be_emit_irprintf("%lu", get_ia32_exc_label_id(node));
1023 * Returns the Proj with projection number proj and NOT mode_M
1025 static ir_node *get_proj(const ir_node *node, long proj)
1027 const ir_edge_t *edge;
1030 assert(get_irn_mode(node) == mode_T && "expected mode_T node");
1032 foreach_out_edge(node, edge) {
1033 src = get_edge_src_irn(edge);
1035 assert(is_Proj(src) && "Proj expected");
1036 if (get_irn_mode(src) == mode_M)
1039 if (get_Proj_proj(src) == proj)
1045 static int can_be_fallthrough(const ir_node *node)
1047 ir_node *target_block = get_cfop_target_block(node);
1048 ir_node *block = get_nodes_block(node);
1049 return get_prev_block_sched(target_block) == block;
1053 * Emits the jump sequence for a conditional jump (cmp + jmp_true + jmp_false)
1055 static void emit_ia32_Jcc(const ir_node *node)
1057 int need_parity_label = 0;
1058 const ir_node *proj_true;
1059 const ir_node *proj_false;
1060 const ir_node *block;
1061 pn_Cmp pnc = get_ia32_condcode(node);
1063 pnc = determine_final_pnc(node, 0, pnc);
1065 /* get both Projs */
1066 proj_true = get_proj(node, pn_ia32_Jcc_true);
1067 assert(proj_true && "Jcc without true Proj");
1069 proj_false = get_proj(node, pn_ia32_Jcc_false);
1070 assert(proj_false && "Jcc without false Proj");
1072 block = get_nodes_block(node);
1074 if (can_be_fallthrough(proj_true)) {
1075 /* exchange both proj's so the second one can be omitted */
1076 const ir_node *t = proj_true;
1078 proj_true = proj_false;
1080 pnc = ia32_get_negated_pnc(pnc);
1083 if (pnc & ia32_pn_Cmp_float) {
1084 /* Some floating point comparisons require a test of the parity flag,
1085 * which indicates that the result is unordered */
1088 ia32_emitf(proj_true, "\tjp %L\n");
1093 ia32_emitf(proj_true, "\tjnp %L\n");
1099 /* we need a local label if the false proj is a fallthrough
1100 * as the falseblock might have no label emitted then */
1101 if (can_be_fallthrough(proj_false)) {
1102 need_parity_label = 1;
1103 ia32_emitf(proj_false, "\tjp 1f\n");
1105 ia32_emitf(proj_false, "\tjp %L\n");
1112 ia32_emitf(proj_true, "\tjp %L\n");
1120 ia32_emitf(proj_true, "\tj%P %L\n", pnc);
1123 if (need_parity_label) {
1124 ia32_emitf(NULL, "1:\n");
1127 /* the second Proj might be a fallthrough */
1128 if (can_be_fallthrough(proj_false)) {
1129 ia32_emitf(proj_false, "\t/* fallthrough to %L */\n");
1131 ia32_emitf(proj_false, "\tjmp %L\n");
1135 static void emit_ia32_CMov(const ir_node *node)
1137 const ia32_attr_t *attr = get_ia32_attr_const(node);
1138 int ins_permuted = attr->data.ins_permuted;
1139 const arch_register_t *out = arch_irn_get_register(node, pn_ia32_res);
1140 pn_Cmp pnc = get_ia32_condcode(node);
1141 const arch_register_t *in_true;
1142 const arch_register_t *in_false;
1144 pnc = determine_final_pnc(node, n_ia32_CMov_eflags, pnc);
1146 in_true = arch_get_irn_register(get_irn_n(node, n_ia32_CMov_val_true));
1147 in_false = arch_get_irn_register(get_irn_n(node, n_ia32_CMov_val_false));
1149 /* should be same constraint fullfilled? */
1150 if (out == in_false) {
1151 /* yes -> nothing to do */
1152 } else if (out == in_true) {
1153 const arch_register_t *tmp;
1155 assert(get_ia32_op_type(node) == ia32_Normal);
1157 ins_permuted = !ins_permuted;
1164 ia32_emitf(node, "\tmovl %R, %R\n", in_false, out);
1168 pnc = ia32_get_negated_pnc(pnc);
1170 /* TODO: handling of Nans isn't correct yet */
1172 ia32_emitf(node, "\tcmov%P %#AR, %#R\n", pnc, in_true, out);
1175 /*********************************************************
1178 * ___ _ __ ___ _| |_ _ _ _ _ __ ___ _ __ ___
1179 * / _ \ '_ ` _ \| | __| | | | | | '_ ` _ \| '_ \/ __|
1180 * | __/ | | | | | | |_ | | |_| | | | | | | |_) \__ \
1181 * \___|_| |_| |_|_|\__| | |\__,_|_| |_| |_| .__/|___/
1184 *********************************************************/
1186 /* jump table entry (target and corresponding number) */
1187 typedef struct _branch_t {
1192 /* jump table for switch generation */
1193 typedef struct _jmp_tbl_t {
1194 ir_node *defProj; /**< default target */
1195 long min_value; /**< smallest switch case */
1196 long max_value; /**< largest switch case */
1197 long num_branches; /**< number of jumps */
1198 char *label; /**< label of the jump table */
1199 branch_t *branches; /**< jump array */
1203 * Compare two variables of type branch_t. Used to sort all switch cases
1205 static int ia32_cmp_branch_t(const void *a, const void *b)
1207 branch_t *b1 = (branch_t *)a;
1208 branch_t *b2 = (branch_t *)b;
1210 if (b1->value <= b2->value)
1217 * Emits code for a SwitchJmp (creates a jump table if
1218 * possible otherwise a cmp-jmp cascade). Port from
1221 static void emit_ia32_SwitchJmp(const ir_node *node)
1223 unsigned long interval;
1229 const ir_edge_t *edge;
1231 /* fill the table structure */
1232 tbl.label = XMALLOCN(char, SNPRINTF_BUF_LEN);
1233 tbl.label = get_unique_label(tbl.label, SNPRINTF_BUF_LEN, ".TBL_");
1235 tbl.num_branches = get_irn_n_edges(node) - 1;
1236 tbl.branches = XMALLOCNZ(branch_t, tbl.num_branches);
1237 tbl.min_value = INT_MAX;
1238 tbl.max_value = INT_MIN;
1240 default_pn = get_ia32_condcode(node);
1242 /* go over all proj's and collect them */
1243 foreach_out_edge(node, edge) {
1244 proj = get_edge_src_irn(edge);
1245 assert(is_Proj(proj) && "Only proj allowed at SwitchJmp");
1247 pnc = get_Proj_proj(proj);
1249 /* check for default proj */
1250 if (pnc == default_pn) {
1251 assert(tbl.defProj == NULL && "found two default Projs at SwitchJmp");
1254 tbl.min_value = pnc < tbl.min_value ? pnc : tbl.min_value;
1255 tbl.max_value = pnc > tbl.max_value ? pnc : tbl.max_value;
1257 /* create branch entry */
1258 tbl.branches[i].target = proj;
1259 tbl.branches[i].value = pnc;
1264 assert(i == tbl.num_branches);
1266 /* sort the branches by their number */
1267 qsort(tbl.branches, tbl.num_branches, sizeof(tbl.branches[0]), ia32_cmp_branch_t);
1269 /* two-complement's magic make this work without overflow */
1270 interval = tbl.max_value - tbl.min_value;
1272 /* emit the table */
1273 ia32_emitf(node, "\tcmpl $%u, %S0\n", interval);
1274 ia32_emitf(tbl.defProj, "\tja %L\n");
1276 if (tbl.num_branches > 1) {
1278 ia32_emitf(node, "\tjmp *%s(,%S0,4)\n", tbl.label);
1280 be_gas_emit_switch_section(GAS_SECTION_RODATA);
1281 ia32_emitf(NULL, "\t.align 4\n");
1282 ia32_emitf(NULL, "%s:\n", tbl.label);
1284 last_value = tbl.branches[0].value;
1285 for (i = 0; i != tbl.num_branches; ++i) {
1286 while (last_value != tbl.branches[i].value) {
1287 ia32_emitf(tbl.defProj, ".long %L\n");
1290 ia32_emitf(tbl.branches[i].target, ".long %L\n");
1293 be_gas_emit_switch_section(GAS_SECTION_TEXT);
1295 /* one jump is enough */
1296 ia32_emitf(tbl.branches[0].target, "\tjmp %L\n");
1306 * Emits code for a unconditional jump.
1308 static void emit_ia32_Jmp(const ir_node *node)
1312 /* for now, the code works for scheduled and non-schedules blocks */
1313 block = get_nodes_block(node);
1315 /* we have a block schedule */
1316 if (can_be_fallthrough(node)) {
1317 ia32_emitf(node, "\t/* fallthrough to %L */\n");
1319 ia32_emitf(node, "\tjmp %L\n");
1324 * Emit an inline assembler operand.
1326 * @param node the ia32_ASM node
1327 * @param s points to the operand (a %c)
1329 * @return pointer to the first char in s NOT in the current operand
1331 static const char* emit_asm_operand(const ir_node *node, const char *s)
1333 const ia32_attr_t *ia32_attr = get_ia32_attr_const(node);
1334 const ia32_asm_attr_t *attr = CONST_CAST_IA32_ATTR(ia32_asm_attr_t,
1336 const arch_register_t *reg;
1337 const ia32_asm_reg_t *asm_regs = attr->register_map;
1338 const ia32_asm_reg_t *asm_reg;
1339 const char *reg_name;
1348 /* parse modifiers */
1351 ir_fprintf(stderr, "Warning: asm text (%+F) ends with %%\n", node);
1376 "Warning: asm text (%+F) contains unknown modifier '%c' for asm op\n",
1383 sscanf(s, "%d%n", &num, &p);
1385 ir_fprintf(stderr, "Warning: Couldn't parse assembler operand (%+F)\n",
1392 if (num < 0 || ARR_LEN(asm_regs) <= num) {
1394 "Error: Custom assembler references invalid input/output (%+F)\n",
1398 asm_reg = & asm_regs[num];
1399 assert(asm_reg->valid);
1402 if (asm_reg->use_input == 0) {
1403 reg = get_out_reg(node, asm_reg->inout_pos);
1405 ir_node *pred = get_irn_n(node, asm_reg->inout_pos);
1407 /* might be an immediate value */
1408 if (is_ia32_Immediate(pred)) {
1409 emit_ia32_Immediate(pred);
1412 reg = get_in_reg(node, asm_reg->inout_pos);
1416 "Warning: no register assigned for %d asm op (%+F)\n",
1421 if (asm_reg->memory) {
1426 if (modifier != 0) {
1430 reg_name = ia32_get_mapped_reg_name(isa->regs_8bit, reg);
1433 reg_name = ia32_get_mapped_reg_name(isa->regs_8bit_high, reg);
1436 reg_name = ia32_get_mapped_reg_name(isa->regs_16bit, reg);
1439 panic("Invalid asm op modifier");
1441 be_emit_string(reg_name);
1443 emit_register(reg, asm_reg->mode);
1446 if (asm_reg->memory) {
1454 * Emits code for an ASM pseudo op.
1456 static void emit_ia32_Asm(const ir_node *node)
1458 const void *gen_attr = get_irn_generic_attr_const(node);
1459 const ia32_asm_attr_t *attr
1460 = CONST_CAST_IA32_ATTR(ia32_asm_attr_t, gen_attr);
1461 ident *asm_text = attr->asm_text;
1462 const char *s = get_id_str(asm_text);
1464 ia32_emitf(node, "#APP\t\n");
1471 s = emit_asm_operand(node, s);
1477 ia32_emitf(NULL, "\n#NO_APP\n");
1480 /**********************************
1483 * | | ___ _ __ _ _| |_) |
1484 * | | / _ \| '_ \| | | | _ <
1485 * | |___| (_) | |_) | |_| | |_) |
1486 * \_____\___/| .__/ \__, |____/
1489 **********************************/
1492 * Emit movsb/w instructions to make mov count divideable by 4
1494 static void emit_CopyB_prolog(unsigned size)
1497 ia32_emitf(NULL, "\tmovsb\n");
1499 ia32_emitf(NULL, "\tmovsw\n");
1503 * Emit rep movsd instruction for memcopy.
1505 static void emit_ia32_CopyB(const ir_node *node)
1507 unsigned size = get_ia32_copyb_size(node);
1509 emit_CopyB_prolog(size);
1510 ia32_emitf(node, "\trep movsd\n");
1514 * Emits unrolled memcopy.
1516 static void emit_ia32_CopyB_i(const ir_node *node)
1518 unsigned size = get_ia32_copyb_size(node);
1520 emit_CopyB_prolog(size);
1524 ia32_emitf(NULL, "\tmovsd\n");
1530 /***************************
1534 * | | / _ \| '_ \ \ / /
1535 * | |___| (_) | | | \ V /
1536 * \_____\___/|_| |_|\_/
1538 ***************************/
1541 * Emit code for conversions (I, FP), (FP, I) and (FP, FP).
1543 static void emit_ia32_Conv_with_FP(const ir_node *node, const char* conv_f,
1546 ir_mode *ls_mode = get_ia32_ls_mode(node);
1547 int ls_bits = get_mode_size_bits(ls_mode);
1548 const char *conv = ls_bits == 32 ? conv_f : conv_d;
1550 ia32_emitf(node, "\tcvt%s %AS3, %D0\n", conv);
1553 static void emit_ia32_Conv_I2FP(const ir_node *node)
1555 emit_ia32_Conv_with_FP(node, "si2ss", "si2sd");
1558 static void emit_ia32_Conv_FP2I(const ir_node *node)
1560 emit_ia32_Conv_with_FP(node, "ss2si", "sd2si");
1563 static void emit_ia32_Conv_FP2FP(const ir_node *node)
1565 emit_ia32_Conv_with_FP(node, "sd2ss", "ss2sd");
1569 * Emits code for an Int conversion.
1571 static void emit_ia32_Conv_I2I(const ir_node *node)
1573 ir_mode *smaller_mode = get_ia32_ls_mode(node);
1574 int signed_mode = mode_is_signed(smaller_mode);
1575 const char *sign_suffix;
1577 assert(!mode_is_float(smaller_mode));
1579 sign_suffix = signed_mode ? "s" : "z";
1580 ia32_emitf(node, "\tmov%s%Ml %#AS3, %D0\n", sign_suffix);
1586 static void emit_ia32_Call(const ir_node *node)
1588 /* Special case: Call must not have its immediates prefixed by $, instead
1589 * address mode is prefixed by *. */
1590 ia32_emitf(node, "\tcall %*AS3\n");
1594 /*******************************************
1597 * | |__ ___ _ __ ___ __| | ___ ___
1598 * | '_ \ / _ \ '_ \ / _ \ / _` |/ _ \/ __|
1599 * | |_) | __/ | | | (_) | (_| | __/\__ \
1600 * |_.__/ \___|_| |_|\___/ \__,_|\___||___/
1602 *******************************************/
1605 * Emits code to increase stack pointer.
1607 static void emit_be_IncSP(const ir_node *node)
1609 int offs = be_get_IncSP_offset(node);
1615 ia32_emitf(node, "\tsubl $%u, %D0\n", offs);
1617 ia32_emitf(node, "\taddl $%u, %D0\n", -offs);
1621 static inline bool is_unknown_reg(const arch_register_t *reg)
1623 if(reg == &ia32_gp_regs[REG_GP_UKNWN]
1624 || reg == &ia32_xmm_regs[REG_XMM_UKNWN]
1625 || reg == &ia32_vfp_regs[REG_VFP_UKNWN])
1632 * Emits code for Copy/CopyKeep.
1634 static void Copy_emitter(const ir_node *node, const ir_node *op)
1636 const arch_register_t *in = arch_get_irn_register(op);
1637 const arch_register_t *out = arch_get_irn_register(node);
1642 if (is_unknown_reg(in))
1644 /* copies of vf nodes aren't real... */
1645 if (arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp])
1648 if (get_irn_mode(node) == mode_E) {
1649 ia32_emitf(node, "\tmovsd %R, %R\n", in, out);
1651 ia32_emitf(node, "\tmovl %R, %R\n", in, out);
1655 static void emit_be_Copy(const ir_node *node)
1657 Copy_emitter(node, be_get_Copy_op(node));
1660 static void emit_be_CopyKeep(const ir_node *node)
1662 Copy_emitter(node, be_get_CopyKeep_op(node));
1666 * Emits code for exchange.
1668 static void emit_be_Perm(const ir_node *node)
1670 const arch_register_t *in0, *in1;
1671 const arch_register_class_t *cls0, *cls1;
1673 in0 = arch_get_irn_register(get_irn_n(node, 0));
1674 in1 = arch_get_irn_register(get_irn_n(node, 1));
1676 cls0 = arch_register_get_class(in0);
1677 cls1 = arch_register_get_class(in1);
1679 assert(cls0 == cls1 && "Register class mismatch at Perm");
1681 if (cls0 == &ia32_reg_classes[CLASS_ia32_gp]) {
1682 ia32_emitf(node, "\txchg %R, %R\n", in1, in0);
1683 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_xmm]) {
1684 ia32_emitf(NULL, "\txorpd %R, %R\n", in1, in0);
1685 ia32_emitf(NULL, "\txorpd %R, %R\n", in0, in1);
1686 ia32_emitf(node, "\txorpd %R, %R\n", in1, in0);
1687 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_vfp]) {
1689 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_st]) {
1692 panic("unexpected register class in be_Perm (%+F)", node);
1697 * Emits code for Constant loading.
1699 static void emit_ia32_Const(const ir_node *node)
1701 ia32_emitf(node, "\tmovl %I, %D0\n");
1705 * Emits code to load the TLS base
1707 static void emit_ia32_LdTls(const ir_node *node)
1709 ia32_emitf(node, "\tmovl %%gs:0, %D0\n");
1712 /* helper function for emit_ia32_Minus64Bit */
1713 static void emit_mov(const ir_node* node, const arch_register_t *src, const arch_register_t *dst)
1715 ia32_emitf(node, "\tmovl %R, %R\n", src, dst);
1718 /* helper function for emit_ia32_Minus64Bit */
1719 static void emit_neg(const ir_node* node, const arch_register_t *reg)
1721 ia32_emitf(node, "\tnegl %R\n", reg);
1724 /* helper function for emit_ia32_Minus64Bit */
1725 static void emit_sbb0(const ir_node* node, const arch_register_t *reg)
1727 ia32_emitf(node, "\tsbbl $0, %R\n", reg);
1730 /* helper function for emit_ia32_Minus64Bit */
1731 static void emit_sbb(const ir_node* node, const arch_register_t *src, const arch_register_t *dst)
1733 ia32_emitf(node, "\tsbbl %R, %R\n", src, dst);
1736 /* helper function for emit_ia32_Minus64Bit */
1737 static void emit_xchg(const ir_node* node, const arch_register_t *src, const arch_register_t *dst)
1739 ia32_emitf(node, "\txchgl %R, %R\n", src, dst);
1742 /* helper function for emit_ia32_Minus64Bit */
1743 static void emit_zero(const ir_node* node, const arch_register_t *reg)
1745 ia32_emitf(node, "\txorl %R, %R\n", reg, reg);
1748 static void emit_ia32_Minus64Bit(const ir_node *node)
1750 const arch_register_t *in_lo = get_in_reg(node, 0);
1751 const arch_register_t *in_hi = get_in_reg(node, 1);
1752 const arch_register_t *out_lo = get_out_reg(node, 0);
1753 const arch_register_t *out_hi = get_out_reg(node, 1);
1755 if (out_lo == in_lo) {
1756 if (out_hi != in_hi) {
1757 /* a -> a, b -> d */
1760 /* a -> a, b -> b */
1763 } else if (out_lo == in_hi) {
1764 if (out_hi == in_lo) {
1765 /* a -> b, b -> a */
1766 emit_xchg(node, in_lo, in_hi);
1769 /* a -> b, b -> d */
1770 emit_mov(node, in_hi, out_hi);
1771 emit_mov(node, in_lo, out_lo);
1775 if (out_hi == in_lo) {
1776 /* a -> c, b -> a */
1777 emit_mov(node, in_lo, out_lo);
1779 } else if (out_hi == in_hi) {
1780 /* a -> c, b -> b */
1781 emit_mov(node, in_lo, out_lo);
1784 /* a -> c, b -> d */
1785 emit_mov(node, in_lo, out_lo);
1791 emit_neg( node, out_hi);
1792 emit_neg( node, out_lo);
1793 emit_sbb0(node, out_hi);
1797 emit_zero(node, out_hi);
1798 emit_neg( node, out_lo);
1799 emit_sbb( node, in_hi, out_hi);
1802 static void emit_ia32_GetEIP(const ir_node *node)
1804 ia32_emitf(node, "\tcall %s\n", pic_base_label);
1805 ia32_emitf(NULL, "%s:\n", pic_base_label);
1806 ia32_emitf(node, "\tpopl %D0\n");
1809 static void emit_ia32_ClimbFrame(const ir_node *node)
1811 const ia32_climbframe_attr_t *attr = get_ia32_climbframe_attr_const(node);
1813 ia32_emitf(node, "\tmovl %S0, %D0\n");
1814 ia32_emitf(node, "\tmovl $%u, %S1\n", attr->count);
1815 ia32_emitf(NULL, BLOCK_PREFIX "%ld:\n", get_irn_node_nr(node));
1816 ia32_emitf(node, "\tmovl (%D0), %D0\n");
1817 ia32_emitf(node, "\tdec %S1\n");
1818 ia32_emitf(node, "\tjnz " BLOCK_PREFIX "%ld\n", get_irn_node_nr(node));
1821 static void emit_be_Return(const ir_node *node)
1823 unsigned pop = be_Return_get_pop(node);
1825 if (pop > 0 || be_Return_get_emit_pop(node)) {
1826 ia32_emitf(node, "\tret $%u\n", pop);
1828 ia32_emitf(node, "\tret\n");
1832 static void emit_Nothing(const ir_node *node)
1838 /***********************************************************************************
1841 * _ __ ___ __ _ _ _ __ | |_ _ __ __ _ _ __ ___ _____ _____ _ __| | __
1842 * | '_ ` _ \ / _` | | '_ \ | _| '__/ _` | '_ ` _ \ / _ \ \ /\ / / _ \| '__| |/ /
1843 * | | | | | | (_| | | | | | | | | | | (_| | | | | | | __/\ V V / (_) | | | <
1844 * |_| |_| |_|\__,_|_|_| |_| |_| |_| \__,_|_| |_| |_|\___| \_/\_/ \___/|_| |_|\_\
1846 ***********************************************************************************/
1849 * Enters the emitter functions for handled nodes into the generic
1850 * pointer of an opcode.
1852 static void ia32_register_emitters(void)
1854 #define IA32_EMIT2(a,b) op_ia32_##a->ops.generic = (op_func)emit_ia32_##b
1855 #define IA32_EMIT(a) IA32_EMIT2(a,a)
1856 #define EMIT(a) op_##a->ops.generic = (op_func)emit_##a
1857 #define IGN(a) op_##a->ops.generic = (op_func)emit_Nothing
1858 #define BE_EMIT(a) op_be_##a->ops.generic = (op_func)emit_be_##a
1859 #define BE_IGN(a) op_be_##a->ops.generic = (op_func)emit_Nothing
1861 /* first clear the generic function pointer for all ops */
1862 clear_irp_opcodes_generic_func();
1864 /* register all emitter functions defined in spec */
1865 ia32_register_spec_emitters();
1867 /* other ia32 emitter functions */
1868 IA32_EMIT2(Conv_I2I8Bit, Conv_I2I);
1873 IA32_EMIT(Conv_FP2FP);
1874 IA32_EMIT(Conv_FP2I);
1875 IA32_EMIT(Conv_I2FP);
1876 IA32_EMIT(Conv_I2I);
1883 IA32_EMIT(Minus64Bit);
1884 IA32_EMIT(SwitchJmp);
1885 IA32_EMIT(ClimbFrame);
1888 /* benode emitter */
1909 typedef void (*emit_func_ptr) (const ir_node *);
1912 * Assign and emit an exception label if the current instruction can fail.
1914 static void ia32_assign_exc_label(ir_node *node)
1916 /* assign a new ID to the instruction */
1917 set_ia32_exc_label_id(node, ++exc_label_id);
1919 ia32_emit_exc_label(node);
1921 be_emit_pad_comment();
1922 be_emit_cstring("/* exception to Block ");
1923 ia32_emit_cfop_target(node);
1924 be_emit_cstring(" */\n");
1925 be_emit_write_line();
1929 * Emits code for a node.
1931 static void ia32_emit_node(ir_node *node)
1933 ir_op *op = get_irn_op(node);
1935 DBG((dbg, LEVEL_1, "emitting code for %+F\n", node));
1937 if (is_ia32_irn(node)) {
1938 if (get_ia32_exc_label(node)) {
1939 /* emit the exception label of this instruction */
1940 ia32_assign_exc_label(node);
1942 if (mark_spill_reload) {
1943 if (is_ia32_is_spill(node)) {
1944 ia32_emitf(NULL, "\txchg %ebx, %ebx /* spill mark */\n");
1946 if (is_ia32_is_reload(node)) {
1947 ia32_emitf(NULL, "\txchg %edx, %edx /* reload mark */\n");
1949 if (is_ia32_is_remat(node)) {
1950 ia32_emitf(NULL, "\txchg %ecx, %ecx /* remat mark */\n");
1954 if (op->ops.generic) {
1955 emit_func_ptr func = (emit_func_ptr) op->ops.generic;
1957 be_dbg_set_dbg_info(get_irn_dbg_info(node));
1962 ir_fprintf(stderr, "Error: No emit handler for node %+F (%+G, graph %+F)\n", node, node, current_ir_graph);
1968 * Emits gas alignment directives
1970 static void ia32_emit_alignment(unsigned align, unsigned skip)
1972 ia32_emitf(NULL, "\t.p2align %u,,%u\n", align, skip);
1976 * Emits gas alignment directives for Labels depended on cpu architecture.
1978 static void ia32_emit_align_label(void)
1980 unsigned align = ia32_cg_config.label_alignment;
1981 unsigned maximum_skip = ia32_cg_config.label_alignment_max_skip;
1982 ia32_emit_alignment(align, maximum_skip);
1986 * Test whether a block should be aligned.
1987 * For cpus in the P4/Athlon class it is useful to align jump labels to
1988 * 16 bytes. However we should only do that if the alignment nops before the
1989 * label aren't executed more often than we have jumps to the label.
1991 static int should_align_block(const ir_node *block)
1993 static const double DELTA = .0001;
1994 ir_exec_freq *exec_freq = cg->birg->exec_freq;
1995 ir_node *prev = get_prev_block_sched(block);
1997 double prev_freq = 0; /**< execfreq of the fallthrough block */
1998 double jmp_freq = 0; /**< execfreq of all non-fallthrough blocks */
2001 if (exec_freq == NULL)
2003 if (ia32_cg_config.label_alignment_factor <= 0)
2006 block_freq = get_block_execfreq(exec_freq, block);
2007 if (block_freq < DELTA)
2010 n_cfgpreds = get_Block_n_cfgpreds(block);
2011 for(i = 0; i < n_cfgpreds; ++i) {
2012 const ir_node *pred = get_Block_cfgpred_block(block, i);
2013 double pred_freq = get_block_execfreq(exec_freq, pred);
2016 prev_freq += pred_freq;
2018 jmp_freq += pred_freq;
2022 if (prev_freq < DELTA && !(jmp_freq < DELTA))
2025 jmp_freq /= prev_freq;
2027 return jmp_freq > ia32_cg_config.label_alignment_factor;
2031 * Emit the block header for a block.
2033 * @param block the block
2034 * @param prev_block the previous block
2036 static void ia32_emit_block_header(ir_node *block)
2038 ir_graph *irg = current_ir_graph;
2039 int need_label = block_needs_label(block);
2041 ir_exec_freq *exec_freq = cg->birg->exec_freq;
2043 if (block == get_irg_end_block(irg))
2046 if (ia32_cg_config.label_alignment > 0) {
2047 /* align the current block if:
2048 * a) if should be aligned due to its execution frequency
2049 * b) there is no fall-through here
2051 if (should_align_block(block)) {
2052 ia32_emit_align_label();
2054 /* if the predecessor block has no fall-through,
2055 we can always align the label. */
2057 int has_fallthrough = 0;
2059 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
2060 ir_node *cfg_pred = get_Block_cfgpred(block, i);
2061 if (can_be_fallthrough(cfg_pred)) {
2062 has_fallthrough = 1;
2067 if (!has_fallthrough)
2068 ia32_emit_align_label();
2073 ia32_emit_block_name(block);
2076 be_emit_pad_comment();
2077 be_emit_cstring(" /* ");
2079 be_emit_cstring("\t/* ");
2080 ia32_emit_block_name(block);
2081 be_emit_cstring(": ");
2084 be_emit_cstring("preds:");
2086 /* emit list of pred blocks in comment */
2087 arity = get_irn_arity(block);
2089 be_emit_cstring(" none");
2091 for (i = 0; i < arity; ++i) {
2092 ir_node *predblock = get_Block_cfgpred_block(block, i);
2093 be_emit_irprintf(" %d", get_irn_node_nr(predblock));
2096 if (exec_freq != NULL) {
2097 be_emit_irprintf(", freq: %f",
2098 get_block_execfreq(exec_freq, block));
2100 be_emit_cstring(" */\n");
2101 be_emit_write_line();
2105 * Walks over the nodes in a block connected by scheduling edges
2106 * and emits code for each node.
2108 static void ia32_gen_block(ir_node *block)
2112 ia32_emit_block_header(block);
2114 /* emit the contents of the block */
2115 be_dbg_set_dbg_info(get_irn_dbg_info(block));
2116 sched_foreach(block, node) {
2117 ia32_emit_node(node);
2121 typedef struct exc_entry {
2122 ir_node *exc_instr; /** The instruction that can issue an exception. */
2123 ir_node *block; /** The block to call then. */
2128 * Sets labels for control flow nodes (jump target).
2129 * Links control predecessors to there destination blocks.
2131 static void ia32_gen_labels(ir_node *block, void *data)
2133 exc_entry **exc_list = data;
2137 for (n = get_Block_n_cfgpreds(block) - 1; n >= 0; --n) {
2138 pred = get_Block_cfgpred(block, n);
2139 set_irn_link(pred, block);
2141 pred = skip_Proj(pred);
2142 if (is_ia32_irn(pred) && get_ia32_exc_label(pred)) {
2147 ARR_APP1(exc_entry, *exc_list, e);
2148 set_irn_link(pred, block);
2154 * Compare two exception_entries.
2156 static int cmp_exc_entry(const void *a, const void *b)
2158 const exc_entry *ea = a;
2159 const exc_entry *eb = b;
2161 if (get_ia32_exc_label_id(ea->exc_instr) < get_ia32_exc_label_id(eb->exc_instr))
2167 * Main driver. Emits the code for one routine.
2169 void ia32_gen_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg)
2171 ir_entity *entity = get_irg_entity(irg);
2172 exc_entry *exc_list = NEW_ARR_F(exc_entry, 0);
2177 do_pic = cg->birg->main_env->options->pic;
2179 ia32_register_emitters();
2181 get_unique_label(pic_base_label, sizeof(pic_base_label), ".PIC_BASE");
2183 be_dbg_method_begin(entity, be_abi_get_stack_layout(cg->birg->abi));
2184 be_gas_emit_function_prolog(entity, ia32_cg_config.function_alignment);
2186 /* we use links to point to target blocks */
2187 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
2188 irg_block_walk_graph(irg, ia32_gen_labels, NULL, &exc_list);
2190 /* initialize next block links */
2191 n = ARR_LEN(cg->blk_sched);
2192 for (i = 0; i < n; ++i) {
2193 ir_node *block = cg->blk_sched[i];
2194 ir_node *prev = i > 0 ? cg->blk_sched[i-1] : NULL;
2196 set_irn_link(block, prev);
2199 for (i = 0; i < n; ++i) {
2200 ir_node *block = cg->blk_sched[i];
2202 ia32_gen_block(block);
2205 be_gas_emit_function_epilog(entity);
2206 be_dbg_method_end();
2208 be_emit_write_line();
2210 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2212 /* Sort the exception table using the exception label id's.
2213 Those are ascending with ascending addresses. */
2214 qsort(exc_list, ARR_LEN(exc_list), sizeof(exc_list[0]), cmp_exc_entry);
2218 for (i = 0; i < ARR_LEN(exc_list); ++i) {
2219 be_emit_cstring("\t.long ");
2220 ia32_emit_exc_label(exc_list[i].exc_instr);
2222 be_emit_cstring("\t.long ");
2223 ia32_emit_block_name(exc_list[i].block);
2227 DEL_ARR_F(exc_list);
2230 static const lc_opt_table_entry_t ia32_emitter_options[] = {
2231 LC_OPT_ENT_BOOL("mark_spill_reload", "mark spills and reloads with ud opcodes", &mark_spill_reload),
2235 /* ==== Experimental binary emitter ==== */
2237 static unsigned char reg_gp_map[N_ia32_gp_REGS];
2238 static unsigned char reg_mmx_map[N_ia32_mmx_REGS];
2239 static unsigned char reg_sse_map[N_ia32_xmm_REGS];
2240 static unsigned char pnc_map_signed[8];
2241 static unsigned char pnc_map_unsigned[8];
2243 static void build_reg_map(void)
2245 reg_gp_map[REG_EAX] = 0x0;
2246 reg_gp_map[REG_ECX] = 0x1;
2247 reg_gp_map[REG_EDX] = 0x2;
2248 reg_gp_map[REG_EBX] = 0x3;
2249 reg_gp_map[REG_ESP] = 0x4;
2250 reg_gp_map[REG_EBP] = 0x5;
2251 reg_gp_map[REG_ESI] = 0x6;
2252 reg_gp_map[REG_EDI] = 0x7;
2254 pnc_map_signed[pn_Cmp_Eq] = 0x04;
2255 pnc_map_signed[pn_Cmp_Lt] = 0x0C;
2256 pnc_map_signed[pn_Cmp_Le] = 0x0E;
2257 pnc_map_signed[pn_Cmp_Gt] = 0x0F;
2258 pnc_map_signed[pn_Cmp_Ge] = 0x0D;
2259 pnc_map_signed[pn_Cmp_Lg] = 0x05;
2261 pnc_map_unsigned[pn_Cmp_Eq] = 0x04;
2262 pnc_map_unsigned[pn_Cmp_Lt] = 0x02;
2263 pnc_map_unsigned[pn_Cmp_Le] = 0x06;
2264 pnc_map_unsigned[pn_Cmp_Gt] = 0x07;
2265 pnc_map_unsigned[pn_Cmp_Ge] = 0x03;
2266 pnc_map_unsigned[pn_Cmp_Lg] = 0x05;
2269 static unsigned char pnc2cc(int pnc)
2272 if (pnc == ia32_pn_Cmp_parity) {
2274 } else if (pnc & ia32_pn_Cmp_float || pnc & ia32_pn_Cmp_unsigned) {
2275 cc = pnc_map_unsigned[pnc & 0x07];
2277 cc = pnc_map_signed[pnc & 0x07];
2283 /** Sign extension bit values for binops */
2285 UNSIGNED_IMM = 0, /**< unsigned immediate */
2286 SIGNEXT_IMM = 2, /**< sign extended immediate */
2289 /** The mod encoding of the ModR/M */
2291 MOD_IND = 0x00, /**< [reg1] */
2292 MOD_IND_BYTE_OFS = 0x40, /**< [reg1 + byte ofs] */
2293 MOD_IND_WORD_OFS = 0x80, /**< [reg1 + word ofs] */
2294 MOD_REG = 0xC0 /**< reg1 */
2297 /** create R/M encoding for ModR/M */
2298 #define ENC_RM(x) (x)
2299 /** create REG encoding for ModR/M */
2300 #define ENC_REG(x) ((x) << 3)
2302 /** create encoding for a SIB byte */
2303 #define ENC_SIB(scale, index, base) ((scale) << 6 | (index) << 3 | (base))
2305 /* Node: The following routines are supposed to append bytes, words, dwords
2306 to the output stream.
2307 Currently the implementation is stupid in that it still creates output
2308 for an "assembler" in the form of .byte, .long
2309 We will change this when enough infrastructure is there to create complete
2310 machine code in memory/object files */
2312 static void bemit8(const unsigned char byte)
2314 be_emit_irprintf("\t.byte 0x%x\n", byte);
2315 be_emit_write_line();
2318 static void bemit16(const unsigned u16)
2320 be_emit_irprintf("\t.word 0x%x\n", u16);
2321 be_emit_write_line();
2324 static void bemit32(const unsigned u32)
2326 be_emit_irprintf("\t.long 0x%x\n", u32);
2327 be_emit_write_line();
2331 * Emit address of an entity. If @p is_relative is true then a relative
2332 * offset from behind the address to the entity is created.
2334 static void bemit_entity(ir_entity *entity, bool entity_sign, int offset,
2337 if (entity == NULL) {
2346 /* the final version should remember the position in the bytestream
2347 and patch it with the correct address at linktime... */
2348 be_emit_cstring("\t.long ");
2351 set_entity_backend_marked(entity, 1);
2352 be_gas_emit_entity(entity);
2355 be_emit_cstring("-.");
2359 be_emit_irprintf("%+d", offset);
2362 be_emit_write_line();
2365 static void bemit_jmp_destination(const ir_node *dest_block)
2367 be_emit_cstring("\t.long ");
2368 ia32_emit_block_name(dest_block);
2369 be_emit_cstring(" - . - 4\n");
2370 be_emit_write_line();
2373 /* end emit routines, all emitters following here should only use the functions
2376 /** Create a ModR/M byte for src1,src2 registers */
2377 static void bemit_modrr(const arch_register_t *src1,
2378 const arch_register_t *src2)
2380 unsigned char modrm = MOD_REG;
2381 modrm |= ENC_RM(reg_gp_map[src1->index]);
2382 modrm |= ENC_REG(reg_gp_map[src2->index]);
2386 /** Create a ModR/M byte for one register and extension */
2387 static void bemit_modru(const arch_register_t *reg, unsigned ext)
2389 unsigned char modrm = MOD_REG;
2391 modrm |= ENC_RM(reg_gp_map[reg->index]);
2392 modrm |= ENC_REG(ext);
2397 * Calculate the size of an (unsigned) immediate in bytes.
2399 * @param offset an offset
2401 static unsigned get_unsigned_imm_size(unsigned offset)
2405 } else if (offset < 65536) {
2413 * Calculate the size of an signed immediate in bytes.
2415 * @param offset an offset
2417 static unsigned get_signed_imm_size(int offset)
2419 if (-128 <= offset && offset < 128) {
2421 } else if (-32768 <= offset && offset < 32768) {
2429 * Emit an address mode.
2431 * @param reg content of the reg field: either a register index or an opcode extension
2432 * @param node the node
2434 static void bemit_mod_am(unsigned reg, const ir_node *node)
2436 ir_entity *ent = get_ia32_am_sc(node);
2437 int offs = get_ia32_am_offs_int(node);
2438 ir_node *base = get_irn_n(node, n_ia32_base);
2439 int has_base = !is_ia32_NoReg_GP(base);
2440 ir_node *index = get_irn_n(node, n_ia32_index);
2441 int has_index = !is_ia32_NoReg_GP(index);
2444 unsigned emitoffs = 0;
2445 bool emitsib = false;
2448 /* set the mod part depending on displacement */
2450 modrm |= MOD_IND_WORD_OFS;
2452 } else if (offs == 0) {
2455 } else if (-128 <= offs && offs < 128) {
2456 modrm |= MOD_IND_BYTE_OFS;
2459 modrm |= MOD_IND_WORD_OFS;
2464 const arch_register_t *base_reg = arch_get_irn_register(base);
2465 base_enc = reg_gp_map[base_reg->index];
2467 /* Use the EBP encoding + MOD_IND if NO base register. There is
2468 * always a 32bit offset present in this case. */
2474 /* Determine if we need a SIB byte. */
2476 const arch_register_t *reg_index = arch_get_irn_register(index);
2477 int scale = get_ia32_am_scale(node);
2479 /* R/M set to ESP means SIB in 32bit mode. */
2480 modrm |= ENC_RM(0x04);
2481 sib = ENC_SIB(scale, reg_gp_map[reg_index->index], base_enc);
2483 } else if (base_enc == 0x04) {
2484 /* for the above reason we are forced to emit a SIB when base is ESP.
2485 * Only the base is used, index must be ESP too, which means no index.
2487 modrm |= ENC_RM(0x04);
2488 sib = ENC_SIB(0, 0x04, 0x04);
2491 modrm |= ENC_RM(base_enc);
2494 /* We are forced to emit an 8bit offset as EBP base without offset is a
2495 * special case for SIB without base register. */
2496 if (base_enc == 0x05 && emitoffs == 0) {
2497 modrm |= MOD_IND_BYTE_OFS;
2501 modrm |= ENC_REG(reg);
2507 /* emit displacement */
2508 if (emitoffs == 8) {
2509 bemit8((unsigned) offs);
2510 } else if (emitoffs == 32) {
2511 bemit_entity(ent, is_ia32_am_sc_sign(node), offs, false);
2516 * Emit a binop with a immediate operand.
2518 * @param node the node to emit
2519 * @param opcode_eax the opcode for the op eax, imm variant
2520 * @param opcode the opcode for the reg, imm variant
2521 * @param ruval the opcode extension for opcode
2523 static void bemit_binop_with_imm(
2524 const ir_node *node,
2525 unsigned char opcode_ax,
2526 unsigned char opcode, unsigned char ruval)
2528 /* Use in-reg, because some instructions (cmp, test) have no out-reg. */
2529 const ir_node *op = get_irn_n(node, n_ia32_binary_right);
2530 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(op);
2533 /* Some instructions (test) have no short form with 32bit value + 8bit
2535 if (attr->symconst != NULL || opcode & SIGNEXT_IMM) {
2538 /* check for sign extension */
2539 size = get_signed_imm_size(attr->offset);
2544 bemit8(opcode | SIGNEXT_IMM);
2545 /* cmp has this special mode */
2546 if (get_ia32_op_type(node) == ia32_AddrModeS) {
2547 bemit_mod_am(ruval, node);
2549 const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left);
2550 bemit_modru(reg, ruval);
2552 bemit8((unsigned char)attr->offset);
2556 /* check for eax variant: this variant is shorter for 32bit immediates only */
2557 if (get_ia32_op_type(node) == ia32_AddrModeS) {
2559 bemit_mod_am(ruval, node);
2561 const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left);
2562 if (reg->index == REG_EAX) {
2566 bemit_modru(reg, ruval);
2569 bemit_entity(attr->symconst, attr->sc_sign, attr->offset, false);
2572 panic("invalid imm size?!?");
2578 static void bemit_binop_2(const ir_node *node, unsigned code)
2580 const arch_register_t *out = get_in_reg(node, n_ia32_binary_left);
2581 ia32_op_type_t am_type = get_ia32_op_type(node);
2582 unsigned char d = 0;
2583 const arch_register_t *op2;
2586 case ia32_AddrModeS:
2589 case ia32_AddrModeD:
2591 bemit_mod_am(reg_gp_map[out->index], node);
2595 op2 = get_in_reg(node, n_ia32_binary_right);
2596 bemit_modrr(out, op2);
2599 panic("invalid address mode");
2605 static void bemit_binop(const ir_node *node, const unsigned char opcodes[4])
2607 ir_node *right = get_irn_n(node, n_ia32_binary_right);
2608 if (is_ia32_Immediate(right)) {
2609 bemit_binop_with_imm(node, opcodes[1], opcodes[2], opcodes[3]);
2611 bemit_binop_2(node, opcodes[0]);
2618 static void bemit_unop(const ir_node *node, unsigned char code, unsigned char ext, int input)
2621 if (get_ia32_op_type(node) == ia32_Normal) {
2622 const arch_register_t *in = get_in_reg(node, input);
2623 bemit_modru(in, ext);
2625 bemit_mod_am(ext, node);
2629 static void bemit_unop_reg(const ir_node *node, unsigned char code, int input)
2631 const arch_register_t *out = get_out_reg(node, 0);
2632 bemit_unop(node, code, reg_gp_map[out->index], input);
2635 static void bemit_immediate(const ir_node *node, bool relative)
2637 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node);
2638 bemit_entity(attr->symconst, attr->sc_sign, attr->offset, relative);
2641 static void bemit_copy(const ir_node *copy)
2643 const arch_register_t *in = get_in_reg(copy, 0);
2644 const arch_register_t *out = get_out_reg(copy, 0);
2646 if (in == out || is_unknown_reg(in))
2648 /* copies of vf nodes aren't real... */
2649 if (arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp])
2652 if (get_irn_mode(copy) == mode_E) {
2655 assert(arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_gp]);
2657 bemit_modrr(out, in);
2661 static void bemit_perm(const ir_node *node)
2663 const arch_register_t *in0 = arch_get_irn_register(get_irn_n(node, 0));
2664 const arch_register_t *in1 = arch_get_irn_register(get_irn_n(node, 1));
2665 const arch_register_class_t *cls0 = arch_register_get_class(in0);
2667 assert(cls0 == arch_register_get_class(in1) && "Register class mismatch at Perm");
2669 if (cls0 == &ia32_reg_classes[CLASS_ia32_gp]) {
2670 if (in0->index == REG_EAX) {
2671 bemit8(0x90 + reg_gp_map[in1->index]);
2672 } else if (in1->index == REG_EAX) {
2673 bemit8(0x90 + reg_gp_map[in0->index]);
2676 bemit_modrr(in0, in1);
2678 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_xmm]) {
2679 panic("unimplemented"); // TODO implement
2680 //ia32_emitf(NULL, "\txorpd %R, %R\n", in1, in0);
2681 //ia32_emitf(NULL, "\txorpd %R, %R\n", in0, in1);
2682 //ia32_emitf(node, "\txorpd %R, %R\n", in1, in0);
2683 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_vfp]) {
2685 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_st]) {
2688 panic("unexpected register class in be_Perm (%+F)", node);
2692 static void bemit_xor0(const ir_node *node)
2694 const arch_register_t *out = get_out_reg(node, 0);
2696 bemit_modrr(out, out);
2699 static void bemit_mov_const(const ir_node *node)
2701 const arch_register_t *out = get_out_reg(node, 0);
2702 bemit8(0xB8 + reg_gp_map[out->index]);
2703 bemit_immediate(node, false);
2707 * Creates a function for a Binop with 3 possible encodings.
2709 #define BINOP(op, op0, op1, op2, op2_ext) \
2710 static void bemit_ ## op(const ir_node *node) { \
2711 static const unsigned char op ## _codes[] = {op0, op1, op2, op2_ext}; \
2712 bemit_binop(node, op ## _codes); \
2715 /* insn def eax,imm imm */
2716 BINOP(add, 0x01, 0x05, 0x81, 0)
2717 BINOP(or, 0x09, 0x0D, 0x81, 1)
2718 BINOP(adc, 0x11, 0x15, 0x81, 2)
2719 BINOP(sbb, 0x19, 0x1D, 0x81, 3)
2720 BINOP(and, 0x21, 0x25, 0x81, 4)
2721 BINOP(sub, 0x29, 0x2D, 0x81, 5)
2722 BINOP(xor, 0x31, 0x35, 0x81, 6)
2723 BINOP(cmp, 0x39, 0x3D, 0x81, 7)
2724 BINOP(test, 0x85, 0xA9, 0xF7, 0)
2727 * Creates a function for an Unop with code /ext encoding.
2729 #define UNOP(op, code, ext, input) \
2730 static void bemit_ ## op(const ir_node *node) { \
2731 bemit_unop(node, code, ext, input); \
2734 UNOP(not, 0xF7, 2, n_ia32_Not_val)
2735 UNOP(neg, 0xF7, 3, n_ia32_Neg_val)
2736 UNOP(mul, 0xF7, 4, n_ia32_Mul_right)
2737 UNOP(imul1op, 0xF7, 5, n_ia32_IMul1OP_right)
2738 UNOP(div, 0xF7, 6, n_ia32_Div_divisor)
2739 UNOP(idiv, 0xF7, 7, n_ia32_IDiv_divisor)
2741 /* TODO: am support for IJmp */
2742 UNOP(ijmp, 0xFF, 4, n_ia32_IJmp_target)
2744 #define SHIFT(op, ext) \
2745 static void bemit_##op(const ir_node *node) \
2747 const arch_register_t *out = get_out_reg(node, 0); \
2748 ir_node *count = get_irn_n(node, 1); \
2749 if (is_ia32_Immediate(count)) { \
2750 int offset = get_ia32_immediate_attr_const(count)->offset; \
2751 if (offset == 1) { \
2753 bemit_modru(out, ext); \
2756 bemit_modru(out, ext); \
2761 bemit_modru(out, ext); \
2771 static void bemit_imul(const ir_node *node)
2773 ir_node *right = get_irn_n(node, n_ia32_IMul_right);
2774 /* Do we need the immediate form? */
2775 if (is_ia32_Immediate(right)) {
2776 int imm = get_ia32_immediate_attr_const(right)->offset;
2777 if (get_signed_imm_size(imm) == 1) {
2778 bemit_unop_reg(node, 0x6B, n_ia32_IMul_left);
2781 bemit_unop_reg(node, 0x69, n_ia32_IMul_left);
2786 bemit_unop_reg(node, 0xAF, n_ia32_IMul_right);
2790 static void bemit_dec(const ir_node *node)
2792 const arch_register_t *out = get_out_reg(node, pn_ia32_Dec_res);
2793 bemit8(0x48 + reg_gp_map[out->index]);
2796 static void bemit_inc(const ir_node *node)
2798 const arch_register_t *out = get_out_reg(node, pn_ia32_Inc_res);
2799 bemit8(0x40 + reg_gp_map[out->index]);
2802 static void bemit_set(const ir_node *node)
2808 pnc = get_ia32_condcode(node);
2809 pnc = determine_final_pnc(node, n_ia32_Set_eflags, pnc);
2810 if (get_ia32_attr_const(node)->data.ins_permuted)
2811 pnc = ia32_get_negated_pnc(pnc);
2813 bemit8(0x90 + pnc2cc(pnc));
2814 bemit_modru(get_out_reg(node, pn_ia32_Set_res), 2);
2820 static void bemit_lea(const ir_node *node)
2822 const arch_register_t *out = get_out_reg(node, 0);
2824 bemit_mod_am(reg_gp_map[out->index], node);
2828 * Emit a single opcode.
2830 #define EMIT_SINGLEOP(op, code) \
2831 static void bemit_ ## op(const ir_node *node) { \
2836 //EMIT_SINGLEOP(daa, 0x27)
2837 //EMIT_SINGLEOP(das, 0x2F)
2838 //EMIT_SINGLEOP(aaa, 0x37)
2839 //EMIT_SINGLEOP(aas, 0x3F)
2840 //EMIT_SINGLEOP(nop, 0x90)
2841 EMIT_SINGLEOP(cwtl, 0x98)
2842 EMIT_SINGLEOP(cltd, 0x99)
2843 //EMIT_SINGLEOP(fwait, 0x9B)
2844 EMIT_SINGLEOP(sahf, 0x9E)
2845 //EMIT_SINGLEOP(popf, 0x9D)
2846 EMIT_SINGLEOP(int3, 0xCC)
2847 //EMIT_SINGLEOP(iret, 0xCF)
2848 //EMIT_SINGLEOP(xlat, 0xD7)
2849 //EMIT_SINGLEOP(lock, 0xF0)
2850 EMIT_SINGLEOP(rep, 0xF3)
2851 //EMIT_SINGLEOP(halt, 0xF4)
2852 EMIT_SINGLEOP(cmc, 0xF5)
2853 EMIT_SINGLEOP(stc, 0xF9)
2854 //EMIT_SINGLEOP(cli, 0xFA)
2855 //EMIT_SINGLEOP(sti, 0xFB)
2856 //EMIT_SINGLEOP(std, 0xFD)
2859 * Emits a MOV out, [MEM].
2861 static void bemit_load(const ir_node *node)
2863 const arch_register_t *out = get_out_reg(node, 0);
2865 if (out->index == REG_EAX) {
2866 ir_entity *ent = get_ia32_am_sc(node);
2867 int offs = get_ia32_am_offs_int(node);
2868 ir_node *base = get_irn_n(node, n_ia32_base);
2869 int has_base = !is_ia32_NoReg_GP(base);
2870 ir_node *index = get_irn_n(node, n_ia32_index);
2871 int has_index = !is_ia32_NoReg_GP(index);
2873 if (ent == NULL && !has_base && !has_index) {
2874 /* load from constant address to EAX can be encoded
2877 bemit_entity(NULL, 0, offs, false);
2882 bemit_mod_am(reg_gp_map[out->index], node);
2886 * Emits a MOV [mem], in.
2888 static void bemit_store(const ir_node *node)
2890 const ir_node *value = get_irn_n(node, n_ia32_Store_val);
2892 if (is_ia32_Immediate(value)) {
2894 bemit_mod_am(0, node);
2895 bemit_immediate(value, false);
2897 const arch_register_t *in = get_in_reg(node, n_ia32_Store_val);
2899 if (in->index == REG_EAX) {
2900 ir_entity *ent = get_ia32_am_sc(node);
2901 int offs = get_ia32_am_offs_int(node);
2902 ir_node *base = get_irn_n(node, n_ia32_base);
2903 int has_base = !is_ia32_NoReg_GP(base);
2904 ir_node *index = get_irn_n(node, n_ia32_index);
2905 int has_index = !is_ia32_NoReg_GP(index);
2907 if (ent == NULL && !has_base && !has_index) {
2908 /* store to constant address from EAX can be encoded as
2911 bemit_entity(NULL, 0, offs, false);
2916 bemit_mod_am(reg_gp_map[in->index], node);
2920 static void bemit_conv_i2i(const ir_node *node)
2922 ir_mode *smaller_mode = get_ia32_ls_mode(node);
2931 if (mode_is_signed(smaller_mode)) opcode |= 0x08;
2932 if (get_mode_size_bits(smaller_mode) == 16) opcode |= 0x01;
2933 bemit_unop_reg(node, opcode, n_ia32_Conv_I2I_val);
2939 static void bemit_push(const ir_node *node)
2941 const ir_node *value = get_irn_n(node, n_ia32_Push_val);
2943 if (is_ia32_Immediate(value)) {
2944 const ia32_immediate_attr_t *attr
2945 = get_ia32_immediate_attr_const(value);
2946 unsigned size = get_unsigned_imm_size(attr->offset);
2952 bemit8((unsigned char)attr->offset);
2957 bemit_immediate(value, false);
2960 } else if (is_ia32_NoReg_GP(value)) {
2962 bemit_mod_am(6, node);
2964 const arch_register_t *reg = get_in_reg(node, n_ia32_Push_val);
2965 bemit8(0x50 + reg_gp_map[reg->index]);
2972 static void bemit_pop(const ir_node *node)
2974 const arch_register_t *reg = get_out_reg(node, pn_ia32_Pop_res);
2975 bemit8(0x58 + reg_gp_map[reg->index]);
2978 static void bemit_popmem(const ir_node *node)
2981 bemit_mod_am(0, node);
2984 static void bemit_call(const ir_node *node)
2986 ir_node *proc = get_irn_n(node, n_ia32_Call_addr);
2988 if (is_ia32_Immediate(proc)) {
2990 bemit_immediate(proc, true);
2992 bemit_unop(node, 0xFF, 2, n_ia32_Call_addr);
2996 static void bemit_jmp(const ir_node *dest_block)
2999 bemit_jmp_destination(dest_block);
3002 static void bemit_jump(const ir_node *node)
3004 if (can_be_fallthrough(node))
3007 bemit_jmp(get_cfop_target_block(node));
3010 static void bemit_jcc(int pnc, const ir_node *dest_block)
3012 unsigned char cc = pnc2cc(pnc);
3015 bemit_jmp_destination(dest_block);
3018 static void bemit_ia32_jcc(const ir_node *node)
3020 int pnc = get_ia32_condcode(node);
3021 int need_parity_label = 0;
3022 const ir_node *proj_true;
3023 const ir_node *proj_false;
3024 const ir_node *dest_true;
3025 const ir_node *dest_false;
3026 const ir_node *block;
3028 pnc = determine_final_pnc(node, 0, pnc);
3030 /* get both Projs */
3031 proj_true = get_proj(node, pn_ia32_Jcc_true);
3032 assert(proj_true && "Jcc without true Proj");
3034 proj_false = get_proj(node, pn_ia32_Jcc_false);
3035 assert(proj_false && "Jcc without false Proj");
3037 block = get_nodes_block(node);
3039 if (can_be_fallthrough(proj_true)) {
3040 /* exchange both proj's so the second one can be omitted */
3041 const ir_node *t = proj_true;
3043 proj_true = proj_false;
3045 pnc = ia32_get_negated_pnc(pnc);
3048 dest_true = get_cfop_target_block(proj_true);
3049 dest_false = get_cfop_target_block(proj_false);
3051 if (pnc & ia32_pn_Cmp_float) {
3052 panic("Float jump NIY");
3053 /* Some floating point comparisons require a test of the parity flag,
3054 * which indicates that the result is unordered */
3057 ia32_emitf(proj_true, "\tjp %L\n");
3062 ia32_emitf(proj_true, "\tjnp %L\n");
3068 /* we need a local label if the false proj is a fallthrough
3069 * as the falseblock might have no label emitted then */
3070 if (can_be_fallthrough(proj_false)) {
3071 need_parity_label = 1;
3072 ia32_emitf(proj_false, "\tjp 1f\n");
3074 ia32_emitf(proj_false, "\tjp %L\n");
3081 ia32_emitf(proj_true, "\tjp %L\n");
3089 bemit_jcc(pnc, dest_true);
3092 if (need_parity_label) {
3093 panic("parity label NIY");
3096 /* the second Proj might be a fallthrough */
3097 if (can_be_fallthrough(proj_false)) {
3098 /* it's a fallthrough */
3100 bemit_jmp(dest_false);
3107 static void bemit_return(const ir_node *node)
3109 unsigned pop = be_Return_get_pop(node);
3110 if (pop > 0 || be_Return_get_emit_pop(node)) {
3112 assert(pop <= 0xffff);
3119 static void bemit_incsp(const ir_node *node)
3122 const arch_register_t *reg;
3126 offs = be_get_IncSP_offset(node);
3137 size = get_signed_imm_size(offs);
3138 bemit8(size == 1 ? 0x83 : 0x81);
3140 reg = get_out_reg(node, 0);
3141 bemit_modru(reg, ext);
3151 * The type of a emitter function.
3153 typedef void (*emit_func) (const ir_node *);
3156 * Set a node emitter. Make it a bit more type safe.
3158 static void register_emitter(ir_op *op, emit_func func)
3160 op->ops.generic = (op_func) func;
3163 static void ia32_register_binary_emitters(void)
3165 /* first clear the generic function pointer for all ops */
3166 clear_irp_opcodes_generic_func();
3168 /* benode emitter */
3169 register_emitter(op_be_Copy, bemit_copy);
3170 register_emitter(op_be_CopyKeep, bemit_copy);
3171 register_emitter(op_be_IncSP, bemit_incsp);
3172 register_emitter(op_be_Perm, bemit_perm);
3173 register_emitter(op_be_Return, bemit_return);
3174 register_emitter(op_ia32_Adc, bemit_adc);
3175 register_emitter(op_ia32_Add, bemit_add);
3176 register_emitter(op_ia32_And, bemit_and);
3177 register_emitter(op_ia32_Breakpoint, bemit_int3);
3178 register_emitter(op_ia32_Call, bemit_call);
3179 register_emitter(op_ia32_Cltd, bemit_cltd);
3180 register_emitter(op_ia32_Cmc, bemit_cmc);
3181 register_emitter(op_ia32_Cmp, bemit_cmp);
3182 register_emitter(op_ia32_Const, bemit_mov_const);
3183 register_emitter(op_ia32_Conv_I2I, bemit_conv_i2i);
3184 register_emitter(op_ia32_Conv_I2I8Bit, bemit_conv_i2i);
3185 register_emitter(op_ia32_Cwtl, bemit_cwtl);
3186 register_emitter(op_ia32_Dec, bemit_dec);
3187 register_emitter(op_ia32_Div, bemit_div);
3188 register_emitter(op_ia32_IDiv, bemit_idiv);
3189 register_emitter(op_ia32_IJmp, bemit_ijmp);
3190 register_emitter(op_ia32_IMul, bemit_imul);
3191 register_emitter(op_ia32_IMul1OP, bemit_imul1op);
3192 register_emitter(op_ia32_Inc, bemit_inc);
3193 register_emitter(op_ia32_Jcc, bemit_ia32_jcc);
3194 register_emitter(op_ia32_Jmp, bemit_jump);
3195 register_emitter(op_ia32_Lea, bemit_lea);
3196 register_emitter(op_ia32_Load, bemit_load);
3197 register_emitter(op_ia32_Mul, bemit_mul);
3198 register_emitter(op_ia32_Neg, bemit_neg);
3199 register_emitter(op_ia32_Not, bemit_not);
3200 register_emitter(op_ia32_Or, bemit_or);
3201 register_emitter(op_ia32_Pop, bemit_pop);
3202 register_emitter(op_ia32_PopEbp, bemit_pop);
3203 register_emitter(op_ia32_PopMem, bemit_popmem);
3204 register_emitter(op_ia32_Push, bemit_push);
3205 register_emitter(op_ia32_RepPrefix, bemit_rep);
3206 register_emitter(op_ia32_Rol, bemit_rol);
3207 register_emitter(op_ia32_Ror, bemit_ror);
3208 register_emitter(op_ia32_Sahf, bemit_sahf);
3209 register_emitter(op_ia32_Sar, bemit_sar);
3210 register_emitter(op_ia32_Sbb, bemit_sbb);
3211 register_emitter(op_ia32_Set, bemit_set);
3212 register_emitter(op_ia32_Shl, bemit_shl);
3213 register_emitter(op_ia32_Shr, bemit_shr);
3214 register_emitter(op_ia32_Stc, bemit_stc);
3215 register_emitter(op_ia32_Store, bemit_store);
3216 register_emitter(op_ia32_Store8Bit, bemit_store);
3217 register_emitter(op_ia32_Sub, bemit_sub);
3218 register_emitter(op_ia32_Test, bemit_test);
3219 register_emitter(op_ia32_Xor, bemit_xor);
3220 register_emitter(op_ia32_Xor0, bemit_xor0);
3222 /* ignore the following nodes */
3223 register_emitter(op_ia32_ProduceVal, emit_Nothing);
3224 register_emitter(op_be_Barrier, emit_Nothing);
3225 register_emitter(op_be_Keep, emit_Nothing);
3226 register_emitter(op_be_Start, emit_Nothing);
3227 register_emitter(op_Phi, emit_Nothing);
3228 register_emitter(op_Start, emit_Nothing);
3231 static void gen_binary_block(ir_node *block)
3235 ia32_emit_block_header(block);
3237 /* emit the contents of the block */
3238 sched_foreach(block, node) {
3239 ia32_emit_node(node);
3243 void ia32_gen_binary_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg)
3245 ir_entity *entity = get_irg_entity(irg);
3251 ia32_register_binary_emitters();
3253 be_gas_emit_function_prolog(entity, ia32_cg_config.function_alignment);
3255 /* we use links to point to target blocks */
3256 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
3257 irg_block_walk_graph(irg, ia32_gen_labels, NULL, NULL);
3259 /* initialize next block links */
3260 n = ARR_LEN(cg->blk_sched);
3261 for (i = 0; i < n; ++i) {
3262 ir_node *block = cg->blk_sched[i];
3263 ir_node *prev = i > 0 ? cg->blk_sched[i-1] : NULL;
3265 set_irn_link(block, prev);
3268 for (i = 0; i < n; ++i) {
3269 ir_node *block = cg->blk_sched[i];
3270 gen_binary_block(block);
3273 be_gas_emit_function_epilog(entity);
3274 be_dbg_method_end();
3276 be_emit_write_line();
3278 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
3284 void ia32_init_emitter(void)
3286 lc_opt_entry_t *be_grp;
3287 lc_opt_entry_t *ia32_grp;
3289 be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
3290 ia32_grp = lc_opt_get_grp(be_grp, "ia32");
3292 lc_opt_add_table(ia32_grp, ia32_emitter_options);
3296 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.emitter");