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
26 * Summary table for x86 floatingpoint compares:
33 * pnc_Leg => NP (ordered)
55 #include "iredges_t.h"
59 #include "raw_bitset.h"
63 #include "../besched.h"
64 #include "../benode.h"
66 #include "../be_dbgout.h"
67 #include "../beemitter.h"
68 #include "../begnuas.h"
70 #include "../be_dbgout.h"
72 #include "ia32_emitter.h"
73 #include "gen_ia32_emitter.h"
74 #include "gen_ia32_regalloc_if.h"
75 #include "ia32_nodes_attr.h"
76 #include "ia32_new_nodes.h"
77 #include "ia32_map_regs.h"
78 #include "ia32_architecture.h"
79 #include "bearch_ia32_t.h"
81 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
83 #define SNPRINTF_BUF_LEN 128
85 static const ia32_isa_t *isa;
86 static ia32_code_gen_t *cg;
87 static char pic_base_label[128];
88 static ir_label_t exc_label_id;
89 static int mark_spill_reload = 0;
92 /** Return the next block in Block schedule */
93 static ir_node *get_prev_block_sched(const ir_node *block)
95 return get_irn_link(block);
98 /** Checks if the current block is a fall-through target. */
99 static int is_fallthrough(const ir_node *cfgpred)
103 if (!is_Proj(cfgpred))
105 pred = get_Proj_pred(cfgpred);
106 if (is_ia32_SwitchJmp(pred))
113 * returns non-zero if the given block needs a label
114 * because of being a jump-target (and not a fall-through)
116 static int block_needs_label(const ir_node *block)
119 int n_cfgpreds = get_Block_n_cfgpreds(block);
121 if (has_Block_entity(block))
124 if (n_cfgpreds == 0) {
126 } else if (n_cfgpreds == 1) {
127 ir_node *cfgpred = get_Block_cfgpred(block, 0);
128 ir_node *cfgpred_block = get_nodes_block(cfgpred);
130 if (get_prev_block_sched(block) == cfgpred_block
131 && is_fallthrough(cfgpred)) {
140 * Returns the register at in position pos.
142 static const arch_register_t *get_in_reg(const ir_node *irn, int pos)
145 const arch_register_t *reg = NULL;
147 assert(get_irn_arity(irn) > pos && "Invalid IN position");
149 /* The out register of the operator at position pos is the
150 in register we need. */
151 op = get_irn_n(irn, pos);
153 reg = arch_get_irn_register(op);
155 assert(reg && "no in register found");
157 if (reg == &ia32_gp_regs[REG_GP_NOREG])
158 panic("trying to emit noreg for %+F input %d", irn, pos);
160 /* in case of unknown register: just return a valid register */
161 if (reg == &ia32_gp_regs[REG_GP_UKNWN]) {
162 const arch_register_req_t *req = arch_get_register_req(irn, pos);
164 if (arch_register_req_is(req, limited)) {
165 /* in case of limited requirements: get the first allowed register */
166 unsigned idx = rbitset_next(req->limited, 0, 1);
167 reg = arch_register_for_index(req->cls, idx);
169 /* otherwise get first register in class */
170 reg = arch_register_for_index(req->cls, 0);
178 * Returns the register at out position pos.
180 static const arch_register_t *get_out_reg(const ir_node *irn, int pos)
183 const arch_register_t *reg = NULL;
185 /* 1st case: irn is not of mode_T, so it has only */
186 /* one OUT register -> good */
187 /* 2nd case: irn is of mode_T -> collect all Projs and ask the */
188 /* Proj with the corresponding projnum for the register */
190 if (get_irn_mode(irn) != mode_T) {
192 reg = arch_get_irn_register(irn);
193 } else if (is_ia32_irn(irn)) {
194 reg = arch_irn_get_register(irn, pos);
196 const ir_edge_t *edge;
198 foreach_out_edge(irn, edge) {
199 proj = get_edge_src_irn(edge);
200 assert(is_Proj(proj) && "non-Proj from mode_T node");
201 if (get_Proj_proj(proj) == pos) {
202 reg = arch_get_irn_register(proj);
208 assert(reg && "no out register found");
213 * Add a number to a prefix. This number will not be used a second time.
215 static char *get_unique_label(char *buf, size_t buflen, const char *prefix)
217 static unsigned long id = 0;
218 snprintf(buf, buflen, "%s%lu", prefix, ++id);
222 /*************************************************************
224 * (_) | | / _| | | | |
225 * _ __ _ __ _ _ __ | |_| |_ | |__ ___| |_ __ ___ _ __
226 * | '_ \| '__| | '_ \| __| _| | '_ \ / _ \ | '_ \ / _ \ '__|
227 * | |_) | | | | | | | |_| | | | | | __/ | |_) | __/ |
228 * | .__/|_| |_|_| |_|\__|_| |_| |_|\___|_| .__/ \___|_|
231 *************************************************************/
234 * Emit the name of the 8bit low register
236 static void emit_8bit_register(const arch_register_t *reg)
238 const char *reg_name = arch_register_get_name(reg);
241 be_emit_char(reg_name[1]);
246 * Emit the name of the 8bit high register
248 static void emit_8bit_register_high(const arch_register_t *reg)
250 const char *reg_name = arch_register_get_name(reg);
253 be_emit_char(reg_name[1]);
257 static void emit_16bit_register(const arch_register_t *reg)
259 const char *reg_name = ia32_get_mapped_reg_name(isa->regs_16bit, reg);
262 be_emit_string(reg_name);
266 * emit a register, possible shortened by a mode
268 * @param reg the register
269 * @param mode the mode of the register or NULL for full register
271 static void emit_register(const arch_register_t *reg, const ir_mode *mode)
273 const char *reg_name;
276 int size = get_mode_size_bits(mode);
278 case 8: emit_8bit_register(reg); return;
279 case 16: emit_16bit_register(reg); return;
281 assert(mode_is_float(mode) || size == 32);
284 reg_name = arch_register_get_name(reg);
287 be_emit_string(reg_name);
290 void ia32_emit_source_register(const ir_node *node, int pos)
292 const arch_register_t *reg = get_in_reg(node, pos);
294 emit_register(reg, NULL);
297 static void ia32_emit_entity(ir_entity *entity, int no_pic_adjust)
299 be_gas_emit_entity(entity);
301 if (get_entity_owner(entity) == get_tls_type()) {
302 if (get_entity_visibility(entity) == ir_visibility_external) {
303 be_emit_cstring("@INDNTPOFF");
305 be_emit_cstring("@NTPOFF");
309 if (do_pic && !no_pic_adjust) {
311 be_emit_string(pic_base_label);
315 static void emit_ia32_Immediate_no_prefix(const ir_node *node)
317 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node);
319 if (attr->symconst != NULL) {
322 ia32_emit_entity(attr->symconst, attr->no_pic_adjust);
324 if (attr->symconst == NULL || attr->offset != 0) {
325 if (attr->symconst != NULL) {
326 be_emit_irprintf("%+d", attr->offset);
328 be_emit_irprintf("0x%X", attr->offset);
333 static void emit_ia32_Immediate(const ir_node *node)
336 emit_ia32_Immediate_no_prefix(node);
339 void ia32_emit_8bit_source_register_or_immediate(const ir_node *node, int pos)
341 const arch_register_t *reg;
342 const ir_node *in = get_irn_n(node, pos);
343 if (is_ia32_Immediate(in)) {
344 emit_ia32_Immediate(in);
348 reg = get_in_reg(node, pos);
349 emit_8bit_register(reg);
352 void ia32_emit_8bit_high_source_register(const ir_node *node, int pos)
354 const arch_register_t *reg = get_in_reg(node, pos);
355 emit_8bit_register_high(reg);
358 void ia32_emit_16bit_source_register_or_immediate(const ir_node *node, int pos)
360 const arch_register_t *reg;
361 const ir_node *in = get_irn_n(node, pos);
362 if (is_ia32_Immediate(in)) {
363 emit_ia32_Immediate(in);
367 reg = get_in_reg(node, pos);
368 emit_16bit_register(reg);
371 void ia32_emit_dest_register(const ir_node *node, int pos)
373 const arch_register_t *reg = get_out_reg(node, pos);
375 emit_register(reg, NULL);
378 void ia32_emit_dest_register_size(const ir_node *node, int pos)
380 const arch_register_t *reg = get_out_reg(node, pos);
382 emit_register(reg, get_ia32_ls_mode(node));
385 void ia32_emit_8bit_dest_register(const ir_node *node, int pos)
387 const arch_register_t *reg = get_out_reg(node, pos);
389 emit_register(reg, mode_Bu);
392 void ia32_emit_x87_register(const ir_node *node, int pos)
394 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
398 be_emit_string(attr->x87[pos]->name);
401 static void ia32_emit_mode_suffix_mode(const ir_mode *mode)
403 assert(mode_is_int(mode) || mode_is_reference(mode));
404 switch (get_mode_size_bits(mode)) {
405 case 8: be_emit_char('b'); return;
406 case 16: be_emit_char('w'); return;
407 case 32: be_emit_char('l'); return;
408 /* gas docu says q is the suffix but gcc, objdump and icc use ll
410 case 64: be_emit_cstring("ll"); return;
412 panic("Can't output mode_suffix for %+F", mode);
415 void ia32_emit_mode_suffix(const ir_node *node)
417 ir_mode *mode = get_ia32_ls_mode(node);
421 ia32_emit_mode_suffix_mode(mode);
424 void ia32_emit_x87_mode_suffix(const ir_node *node)
428 /* we only need to emit the mode on address mode */
429 if (get_ia32_op_type(node) == ia32_Normal)
432 mode = get_ia32_ls_mode(node);
433 assert(mode != NULL);
435 if (mode_is_float(mode)) {
436 switch (get_mode_size_bits(mode)) {
437 case 32: be_emit_char('s'); return;
438 case 64: be_emit_char('l'); return;
440 case 96: be_emit_char('t'); return;
443 assert(mode_is_int(mode));
444 switch (get_mode_size_bits(mode)) {
445 case 16: be_emit_char('s'); return;
446 case 32: be_emit_char('l'); return;
447 /* gas docu says q is the suffix but gcc, objdump and icc use ll
449 case 64: be_emit_cstring("ll"); return;
452 panic("Can't output mode_suffix for %+F", mode);
455 static char get_xmm_mode_suffix(ir_mode *mode)
457 assert(mode_is_float(mode));
458 switch(get_mode_size_bits(mode)) {
461 default: panic("Invalid XMM mode");
465 void ia32_emit_xmm_mode_suffix(const ir_node *node)
467 ir_mode *mode = get_ia32_ls_mode(node);
468 assert(mode != NULL);
470 be_emit_char(get_xmm_mode_suffix(mode));
473 void ia32_emit_xmm_mode_suffix_s(const ir_node *node)
475 ir_mode *mode = get_ia32_ls_mode(node);
476 assert(mode != NULL);
477 be_emit_char(get_xmm_mode_suffix(mode));
480 void ia32_emit_extend_suffix(const ir_node *node)
482 ir_mode *mode = get_ia32_ls_mode(node);
483 if (get_mode_size_bits(mode) == 32)
485 be_emit_char(mode_is_signed(mode) ? 's' : 'z');
486 ia32_emit_mode_suffix_mode(mode);
489 void ia32_emit_source_register_or_immediate(const ir_node *node, int pos)
491 ir_node *in = get_irn_n(node, pos);
492 if (is_ia32_Immediate(in)) {
493 emit_ia32_Immediate(in);
495 const ir_mode *mode = get_ia32_ls_mode(node);
496 const arch_register_t *reg = get_in_reg(node, pos);
497 emit_register(reg, mode);
502 * Returns the target block for a control flow node.
504 static ir_node *get_cfop_target_block(const ir_node *irn)
506 assert(get_irn_mode(irn) == mode_X);
507 return get_irn_link(irn);
511 * Emits the target label for a control flow node.
513 static void ia32_emit_cfop_target(const ir_node *node)
515 ir_node *block = get_cfop_target_block(node);
516 be_gas_emit_block_name(block);
520 * positive conditions for signed compares
522 static const char *const cmp2condition_s[] = {
523 NULL, /* always false */
530 NULL /* always true */
534 * positive conditions for unsigned compares
536 static const char *const cmp2condition_u[] = {
537 NULL, /* always false */
544 NULL /* always true */
548 * Emit the suffix for a compare instruction.
550 static void ia32_emit_cmp_suffix(int pnc)
554 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,
572 EMIT_HIGH_REG = 1U << 3,
573 EMIT_LOW_REG = 1U << 4
577 * Emits address mode.
579 void ia32_emit_am(const ir_node *node)
581 ir_entity *ent = get_ia32_am_sc(node);
582 int offs = get_ia32_am_offs_int(node);
583 ir_node *base = get_irn_n(node, n_ia32_base);
584 int has_base = !is_ia32_NoReg_GP(base);
585 ir_node *index = get_irn_n(node, n_ia32_index);
586 int has_index = !is_ia32_NoReg_GP(index);
588 /* just to be sure... */
589 assert(!is_ia32_use_frame(node) || get_ia32_frame_ent(node) != NULL);
593 const ia32_attr_t *attr = get_ia32_attr_const(node);
594 if (is_ia32_am_sc_sign(node))
596 ia32_emit_entity(ent, attr->data.am_sc_no_pic_adjust);
599 /* also handle special case if nothing is set */
600 if (offs != 0 || (ent == NULL && !has_base && !has_index)) {
602 be_emit_irprintf("%+d", offs);
604 be_emit_irprintf("%d", offs);
608 if (has_base || has_index) {
613 const arch_register_t *reg = get_in_reg(node, n_ia32_base);
614 emit_register(reg, NULL);
617 /* emit index + scale */
619 const arch_register_t *reg = get_in_reg(node, n_ia32_index);
622 emit_register(reg, NULL);
624 scale = get_ia32_am_scale(node);
626 be_emit_irprintf(",%d", 1 << scale);
634 * fmt parameter output
635 * ---- ---------------------- ---------------------------------------------
637 * %AM <node> address mode of the node
638 * %AR const arch_register_t* address mode of the node or register
639 * %ASx <node> address mode of the node or source register x
640 * %Dx <node> destination register x
641 * %I <node> immediate of the node
642 * %L <node> control flow target of the node
643 * %M <node> mode suffix of the node
644 * %P int condition code
645 * %R const arch_register_t* register
646 * %Sx <node> source register x
647 * %s const char* string
648 * %u unsigned int unsigned int
649 * %d signed int signed int
652 * # modifier for %ASx, %D, %R, and %S uses ls mode of node to alter register width
653 * * modifier does not prefix immediates with $, but AM with *
654 * l modifier for %lu and %ld
655 * > modifier to output high 8bit register (ah, bh)
656 * < modifier to output low 8bit register (al, bl)
658 static void ia32_emitf(const ir_node *node, const char *fmt, ...)
664 const char *start = fmt;
665 ia32_emit_mod_t mod = 0;
667 while (*fmt != '%' && *fmt != '\n' && *fmt != '\0')
670 be_emit_string_len(start, fmt - start);
674 be_emit_finish_line_gas(node);
687 case '*': mod |= EMIT_ALTERNATE_AM; break;
688 case '#': mod |= EMIT_RESPECT_LS; break;
689 case 'l': mod |= EMIT_LONG; break;
690 case '>': mod |= EMIT_HIGH_REG; break;
691 case '<': mod |= EMIT_LOW_REG; break;
708 if (mod & EMIT_ALTERNATE_AM)
714 const arch_register_t *reg = va_arg(ap, const arch_register_t*);
715 if (get_ia32_op_type(node) == ia32_AddrModeS) {
718 if (mod & EMIT_ALTERNATE_AM)
720 emit_register(reg, NULL);
726 if (get_ia32_op_type(node) == ia32_AddrModeS) {
730 assert(get_ia32_op_type(node) == ia32_Normal);
735 default: goto unknown;
742 const arch_register_t *reg;
744 if (*fmt < '0' || '9' <= *fmt)
748 reg = get_out_reg(node, pos);
749 emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL);
754 if (!(mod & EMIT_ALTERNATE_AM))
756 emit_ia32_Immediate_no_prefix(node);
760 ia32_emit_cfop_target(node);
764 ia32_emit_mode_suffix_mode(get_ia32_ls_mode(node));
769 int pnc = va_arg(ap, int);
770 ia32_emit_cmp_suffix(pnc);
775 const arch_register_t *reg = va_arg(ap, const arch_register_t*);
776 if (mod & EMIT_HIGH_REG) {
777 emit_8bit_register_high(reg);
778 } else if (mod & EMIT_LOW_REG) {
779 emit_8bit_register(reg);
781 emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL);
791 if (*fmt < '0' || '9' <= *fmt)
795 in = get_irn_n(node, pos);
796 if (is_ia32_Immediate(in)) {
797 if (!(mod & EMIT_ALTERNATE_AM))
799 emit_ia32_Immediate_no_prefix(in);
801 const arch_register_t *reg;
803 if (mod & EMIT_ALTERNATE_AM)
805 reg = get_in_reg(node, pos);
806 emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL);
812 const char *str = va_arg(ap, const char*);
818 if (mod & EMIT_LONG) {
819 unsigned long num = va_arg(ap, unsigned long);
820 be_emit_irprintf("%lu", num);
822 unsigned num = va_arg(ap, unsigned);
823 be_emit_irprintf("%u", num);
828 if (mod & EMIT_LONG) {
829 long num = va_arg(ap, long);
830 be_emit_irprintf("%ld", num);
832 int num = va_arg(ap, int);
833 be_emit_irprintf("%d", num);
839 panic("unknown format conversion in ia32_emitf()");
847 * Emits registers and/or address mode of a binary operation.
849 void ia32_emit_binop(const ir_node *node)
851 if (is_ia32_Immediate(get_irn_n(node, n_ia32_binary_right))) {
852 ia32_emitf(node, "%#S4, %#AS3");
854 ia32_emitf(node, "%#AS4, %#S3");
859 * Emits registers and/or address mode of a binary operation.
861 void ia32_emit_x87_binop(const ir_node *node)
863 switch(get_ia32_op_type(node)) {
866 const ia32_x87_attr_t *x87_attr = get_ia32_x87_attr_const(node);
867 const arch_register_t *in1 = x87_attr->x87[0];
868 const arch_register_t *in = x87_attr->x87[1];
869 const arch_register_t *out = x87_attr->x87[2];
873 } else if (out == in) {
878 be_emit_string(arch_register_get_name(in));
879 be_emit_cstring(", %");
880 be_emit_string(arch_register_get_name(out));
888 assert(0 && "unsupported op type");
893 * Emits registers and/or address mode of a unary operation.
895 void ia32_emit_unop(const ir_node *node, int pos)
899 ia32_emitf(node, fmt);
902 static void emit_ia32_IMul(const ir_node *node)
904 ir_node *left = get_irn_n(node, n_ia32_IMul_left);
905 const arch_register_t *out_reg = get_out_reg(node, pn_ia32_IMul_res);
907 /* do we need the 3-address form? */
908 if (is_ia32_NoReg_GP(left) ||
909 get_in_reg(node, n_ia32_IMul_left) != out_reg) {
910 ia32_emitf(node, "\timul%M %#S4, %#AS3, %#D0\n");
912 ia32_emitf(node, "\timul%M %#AS4, %#S3\n");
917 * walks up a tree of copies/perms/spills/reloads to find the original value
918 * that is moved around
920 static ir_node *find_original_value(ir_node *node)
922 if (irn_visited(node))
925 mark_irn_visited(node);
926 if (be_is_Copy(node)) {
927 return find_original_value(be_get_Copy_op(node));
928 } else if (be_is_CopyKeep(node)) {
929 return find_original_value(be_get_CopyKeep_op(node));
930 } else if (is_Proj(node)) {
931 ir_node *pred = get_Proj_pred(node);
932 if (be_is_Perm(pred)) {
933 return find_original_value(get_irn_n(pred, get_Proj_proj(node)));
934 } else if (be_is_MemPerm(pred)) {
935 return find_original_value(get_irn_n(pred, get_Proj_proj(node) + 1));
936 } else if (is_ia32_Load(pred)) {
937 return find_original_value(get_irn_n(pred, n_ia32_Load_mem));
941 } else if (is_ia32_Store(node)) {
942 return find_original_value(get_irn_n(node, n_ia32_Store_val));
943 } else if (is_Phi(node)) {
945 arity = get_irn_arity(node);
946 for (i = 0; i < arity; ++i) {
947 ir_node *in = get_irn_n(node, i);
948 ir_node *res = find_original_value(in);
959 static int determine_final_pnc(const ir_node *node, int flags_pos, int pnc)
961 ir_node *flags = get_irn_n(node, flags_pos);
962 const ia32_attr_t *flags_attr;
963 flags = skip_Proj(flags);
965 if (is_ia32_Sahf(flags)) {
966 ir_node *cmp = get_irn_n(flags, n_ia32_Sahf_val);
967 if (!(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp)
968 || is_ia32_FucomppFnstsw(cmp) || is_ia32_FtstFnstsw(cmp))) {
969 inc_irg_visited(current_ir_graph);
970 cmp = find_original_value(cmp);
972 assert(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp)
973 || is_ia32_FucomppFnstsw(cmp) || is_ia32_FtstFnstsw(cmp));
976 flags_attr = get_ia32_attr_const(cmp);
977 if (flags_attr->data.ins_permuted)
978 pnc = get_mirrored_pnc(pnc);
979 pnc |= ia32_pn_Cmp_float;
980 } else if (is_ia32_Ucomi(flags) || is_ia32_Fucomi(flags)
981 || is_ia32_Fucompi(flags)) {
982 flags_attr = get_ia32_attr_const(flags);
984 if (flags_attr->data.ins_permuted)
985 pnc = get_mirrored_pnc(pnc);
986 pnc |= ia32_pn_Cmp_float;
988 flags_attr = get_ia32_attr_const(flags);
990 if (flags_attr->data.ins_permuted)
991 pnc = get_mirrored_pnc(pnc);
992 if (flags_attr->data.cmp_unsigned)
993 pnc |= ia32_pn_Cmp_unsigned;
999 static pn_Cmp ia32_get_negated_pnc(pn_Cmp pnc)
1001 ir_mode *mode = pnc & ia32_pn_Cmp_float ? mode_F : mode_Iu;
1002 return get_negated_pnc(pnc, mode);
1005 void ia32_emit_cmp_suffix_node(const ir_node *node, int flags_pos)
1007 pn_Cmp pnc = get_ia32_condcode(node);
1008 pnc = determine_final_pnc(node, flags_pos, 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 */
1086 switch (pnc & 0x0f) {
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");
1136 * Emits an ia32 Setcc. This is mostly easy but some floating point compares
1139 static void emit_ia32_Setcc(const ir_node *node)
1141 const arch_register_t *dreg = get_out_reg(node, pn_ia32_Setcc_res);
1143 pn_Cmp pnc = get_ia32_condcode(node);
1144 pnc = determine_final_pnc(node, n_ia32_Setcc_eflags, pnc);
1145 if (pnc & ia32_pn_Cmp_float) {
1146 switch (pnc & 0x0f) {
1148 ia32_emitf(node, "\tsetp %#R\n", dreg);
1152 ia32_emitf(node, "\tsetnp %#R\n", dreg);
1158 ia32_emitf(node, "\tset%P %<R\n", pnc, dreg);
1159 ia32_emitf(node, "\tsetnp %>R\n", dreg);
1160 ia32_emitf(node, "\tandb %>R, %<R\n", dreg, dreg);
1166 ia32_emitf(node, "\tset%P %<R\n", pnc, dreg);
1167 ia32_emitf(node, "\tsetp %>R\n", dreg);
1168 ia32_emitf(node, "\torb %>R, %<R\n", dreg, dreg);
1175 ia32_emitf(node, "\tset%P %#R\n", pnc, dreg);
1178 static void emit_ia32_CMovcc(const ir_node *node)
1180 const ia32_attr_t *attr = get_ia32_attr_const(node);
1181 const arch_register_t *out = arch_irn_get_register(node, pn_ia32_res);
1182 pn_Cmp pnc = get_ia32_condcode(node);
1183 const arch_register_t *in_true;
1184 const arch_register_t *in_false;
1186 pnc = determine_final_pnc(node, n_ia32_CMovcc_eflags, pnc);
1187 /* although you can't set ins_permuted in the constructor it might still
1188 * be set by memory operand folding
1189 * Permuting inputs of a cmov means the condition is negated!
1191 if (attr->data.ins_permuted)
1192 pnc = ia32_get_negated_pnc(pnc);
1194 in_true = arch_get_irn_register(get_irn_n(node, n_ia32_CMovcc_val_true));
1195 in_false = arch_get_irn_register(get_irn_n(node, n_ia32_CMovcc_val_false));
1197 /* should be same constraint fullfilled? */
1198 if (out == in_false) {
1199 /* yes -> nothing to do */
1200 } else if (out == in_true) {
1201 const arch_register_t *tmp;
1203 assert(get_ia32_op_type(node) == ia32_Normal);
1205 pnc = ia32_get_negated_pnc(pnc);
1212 ia32_emitf(node, "\tmovl %R, %R\n", in_false, out);
1215 /* TODO: handling of Nans isn't correct yet */
1216 if (pnc & ia32_pn_Cmp_float) {
1217 switch (pnc & 0x0f) {
1226 panic("CMov with floatingpoint compare/parity not supported yet");
1230 ia32_emitf(node, "\tcmov%P %#AR, %#R\n", pnc, in_true, out);
1233 /*********************************************************
1236 * ___ _ __ ___ _| |_ _ _ _ _ __ ___ _ __ ___
1237 * / _ \ '_ ` _ \| | __| | | | | | '_ ` _ \| '_ \/ __|
1238 * | __/ | | | | | | |_ | | |_| | | | | | | |_) \__ \
1239 * \___|_| |_| |_|_|\__| | |\__,_|_| |_| |_| .__/|___/
1242 *********************************************************/
1244 /* jump table entry (target and corresponding number) */
1245 typedef struct _branch_t {
1250 /* jump table for switch generation */
1251 typedef struct _jmp_tbl_t {
1252 ir_node *defProj; /**< default target */
1253 long min_value; /**< smallest switch case */
1254 long max_value; /**< largest switch case */
1255 long num_branches; /**< number of jumps */
1256 char label[SNPRINTF_BUF_LEN]; /**< label of the jump table */
1257 branch_t *branches; /**< jump array */
1261 * Compare two variables of type branch_t. Used to sort all switch cases
1263 static int ia32_cmp_branch_t(const void *a, const void *b)
1265 branch_t *b1 = (branch_t *)a;
1266 branch_t *b2 = (branch_t *)b;
1268 if (b1->value <= b2->value)
1274 static void generate_jump_table(jmp_tbl_t *tbl, const ir_node *node)
1280 const ir_edge_t *edge;
1282 /* fill the table structure */
1283 get_unique_label(tbl->label, SNPRINTF_BUF_LEN, ".TBL_");
1284 tbl->defProj = NULL;
1285 tbl->num_branches = get_irn_n_edges(node) - 1;
1286 tbl->branches = XMALLOCNZ(branch_t, tbl->num_branches);
1287 tbl->min_value = LONG_MAX;
1288 tbl->max_value = LONG_MIN;
1290 default_pn = get_ia32_condcode(node);
1292 /* go over all proj's and collect them */
1293 foreach_out_edge(node, edge) {
1294 proj = get_edge_src_irn(edge);
1295 assert(is_Proj(proj) && "Only proj allowed at SwitchJmp");
1297 pnc = get_Proj_proj(proj);
1299 /* check for default proj */
1300 if (pnc == default_pn) {
1301 assert(tbl->defProj == NULL && "found two default Projs at SwitchJmp");
1302 tbl->defProj = proj;
1304 tbl->min_value = pnc < tbl->min_value ? pnc : tbl->min_value;
1305 tbl->max_value = pnc > tbl->max_value ? pnc : tbl->max_value;
1307 /* create branch entry */
1308 tbl->branches[i].target = proj;
1309 tbl->branches[i].value = pnc;
1314 assert(i == tbl->num_branches);
1316 /* sort the branches by their number */
1317 qsort(tbl->branches, tbl->num_branches, sizeof(tbl->branches[0]), ia32_cmp_branch_t);
1321 * Emits code for a SwitchJmp (creates a jump table if
1322 * possible otherwise a cmp-jmp cascade). Port from
1325 static void emit_ia32_SwitchJmp(const ir_node *node)
1327 unsigned long interval;
1331 /* fill the table structure */
1332 generate_jump_table(&tbl, node);
1334 /* two-complement's magic make this work without overflow */
1335 interval = tbl.max_value - tbl.min_value;
1337 /* emit the table */
1338 ia32_emitf(node, "\tcmpl $%u, %S0\n", interval);
1339 ia32_emitf(tbl.defProj, "\tja %L\n");
1341 if (tbl.num_branches > 1) {
1343 ia32_emitf(node, "\tjmp *%s(,%S0,4)\n", tbl.label);
1345 be_gas_emit_switch_section(GAS_SECTION_RODATA);
1346 ia32_emitf(NULL, "\t.align 4\n");
1347 ia32_emitf(NULL, "%s:\n", tbl.label);
1349 last_value = tbl.branches[0].value;
1350 for (i = 0; i != tbl.num_branches; ++i) {
1351 while (last_value != tbl.branches[i].value) {
1352 ia32_emitf(tbl.defProj, ".long %L\n");
1355 ia32_emitf(tbl.branches[i].target, ".long %L\n");
1358 be_gas_emit_switch_section(GAS_SECTION_TEXT);
1360 /* one jump is enough */
1361 ia32_emitf(tbl.branches[0].target, "\tjmp %L\n");
1368 * Emits code for a unconditional jump.
1370 static void emit_ia32_Jmp(const ir_node *node)
1374 /* for now, the code works for scheduled and non-schedules blocks */
1375 block = get_nodes_block(node);
1377 /* we have a block schedule */
1378 if (can_be_fallthrough(node)) {
1379 ia32_emitf(node, "\t/* fallthrough to %L */\n");
1381 ia32_emitf(node, "\tjmp %L\n");
1386 * Emit an inline assembler operand.
1388 * @param node the ia32_ASM node
1389 * @param s points to the operand (a %c)
1391 * @return pointer to the first char in s NOT in the current operand
1393 static const char* emit_asm_operand(const ir_node *node, const char *s)
1395 const ia32_attr_t *ia32_attr = get_ia32_attr_const(node);
1396 const ia32_asm_attr_t *attr = CONST_CAST_IA32_ATTR(ia32_asm_attr_t,
1398 const arch_register_t *reg;
1399 const ia32_asm_reg_t *asm_regs = attr->register_map;
1400 const ia32_asm_reg_t *asm_reg;
1401 const char *reg_name;
1410 /* parse modifiers */
1413 ir_fprintf(stderr, "Warning: asm text (%+F) ends with %%\n", node);
1438 "Warning: asm text (%+F) contains unknown modifier '%c' for asm op\n",
1445 sscanf(s, "%d%n", &num, &p);
1447 ir_fprintf(stderr, "Warning: Couldn't parse assembler operand (%+F)\n",
1454 if (num < 0 || ARR_LEN(asm_regs) <= num) {
1456 "Error: Custom assembler references invalid input/output (%+F)\n",
1460 asm_reg = & asm_regs[num];
1461 assert(asm_reg->valid);
1464 if (asm_reg->use_input == 0) {
1465 reg = get_out_reg(node, asm_reg->inout_pos);
1467 ir_node *pred = get_irn_n(node, asm_reg->inout_pos);
1469 /* might be an immediate value */
1470 if (is_ia32_Immediate(pred)) {
1471 emit_ia32_Immediate(pred);
1474 reg = get_in_reg(node, asm_reg->inout_pos);
1478 "Warning: no register assigned for %d asm op (%+F)\n",
1483 if (asm_reg->memory) {
1488 if (modifier != 0) {
1492 reg_name = ia32_get_mapped_reg_name(isa->regs_8bit, reg);
1495 reg_name = ia32_get_mapped_reg_name(isa->regs_8bit_high, reg);
1498 reg_name = ia32_get_mapped_reg_name(isa->regs_16bit, reg);
1501 panic("Invalid asm op modifier");
1503 be_emit_string(reg_name);
1505 emit_register(reg, asm_reg->mode);
1508 if (asm_reg->memory) {
1516 * Emits code for an ASM pseudo op.
1518 static void emit_ia32_Asm(const ir_node *node)
1520 const void *gen_attr = get_irn_generic_attr_const(node);
1521 const ia32_asm_attr_t *attr
1522 = CONST_CAST_IA32_ATTR(ia32_asm_attr_t, gen_attr);
1523 ident *asm_text = attr->asm_text;
1524 const char *s = get_id_str(asm_text);
1526 ia32_emitf(node, "#APP\t\n");
1533 s = emit_asm_operand(node, s);
1539 ia32_emitf(NULL, "\n#NO_APP\n");
1542 /**********************************
1545 * | | ___ _ __ _ _| |_) |
1546 * | | / _ \| '_ \| | | | _ <
1547 * | |___| (_) | |_) | |_| | |_) |
1548 * \_____\___/| .__/ \__, |____/
1551 **********************************/
1554 * Emit movsb/w instructions to make mov count divideable by 4
1556 static void emit_CopyB_prolog(unsigned size)
1559 ia32_emitf(NULL, "\tmovsb\n");
1561 ia32_emitf(NULL, "\tmovsw\n");
1565 * Emit rep movsd instruction for memcopy.
1567 static void emit_ia32_CopyB(const ir_node *node)
1569 unsigned size = get_ia32_copyb_size(node);
1571 emit_CopyB_prolog(size);
1572 ia32_emitf(node, "\trep movsd\n");
1576 * Emits unrolled memcopy.
1578 static void emit_ia32_CopyB_i(const ir_node *node)
1580 unsigned size = get_ia32_copyb_size(node);
1582 emit_CopyB_prolog(size);
1586 ia32_emitf(NULL, "\tmovsd\n");
1592 /***************************
1596 * | | / _ \| '_ \ \ / /
1597 * | |___| (_) | | | \ V /
1598 * \_____\___/|_| |_|\_/
1600 ***************************/
1603 * Emit code for conversions (I, FP), (FP, I) and (FP, FP).
1605 static void emit_ia32_Conv_with_FP(const ir_node *node, const char* conv_f,
1608 ir_mode *ls_mode = get_ia32_ls_mode(node);
1609 int ls_bits = get_mode_size_bits(ls_mode);
1610 const char *conv = ls_bits == 32 ? conv_f : conv_d;
1612 ia32_emitf(node, "\tcvt%s %AS3, %D0\n", conv);
1615 static void emit_ia32_Conv_I2FP(const ir_node *node)
1617 emit_ia32_Conv_with_FP(node, "si2ss", "si2sd");
1620 static void emit_ia32_Conv_FP2I(const ir_node *node)
1622 emit_ia32_Conv_with_FP(node, "ss2si", "sd2si");
1625 static void emit_ia32_Conv_FP2FP(const ir_node *node)
1627 emit_ia32_Conv_with_FP(node, "sd2ss", "ss2sd");
1631 * Emits code for an Int conversion.
1633 static void emit_ia32_Conv_I2I(const ir_node *node)
1635 ir_mode *smaller_mode = get_ia32_ls_mode(node);
1636 int signed_mode = mode_is_signed(smaller_mode);
1637 const char *sign_suffix;
1639 assert(!mode_is_float(smaller_mode));
1641 sign_suffix = signed_mode ? "s" : "z";
1642 ia32_emitf(node, "\tmov%s%Ml %#AS3, %D0\n", sign_suffix);
1648 static void emit_ia32_Call(const ir_node *node)
1650 /* Special case: Call must not have its immediates prefixed by $, instead
1651 * address mode is prefixed by *. */
1652 ia32_emitf(node, "\tcall %*AS3\n");
1656 /*******************************************
1659 * | |__ ___ _ __ ___ __| | ___ ___
1660 * | '_ \ / _ \ '_ \ / _ \ / _` |/ _ \/ __|
1661 * | |_) | __/ | | | (_) | (_| | __/\__ \
1662 * |_.__/ \___|_| |_|\___/ \__,_|\___||___/
1664 *******************************************/
1667 * Emits code to increase stack pointer.
1669 static void emit_be_IncSP(const ir_node *node)
1671 int offs = be_get_IncSP_offset(node);
1677 ia32_emitf(node, "\tsubl $%u, %D0\n", offs);
1679 ia32_emitf(node, "\taddl $%u, %D0\n", -offs);
1683 static inline bool is_unknown_reg(const arch_register_t *reg)
1685 if(reg == &ia32_gp_regs[REG_GP_UKNWN]
1686 || reg == &ia32_xmm_regs[REG_XMM_UKNWN]
1687 || reg == &ia32_vfp_regs[REG_VFP_UKNWN])
1694 * Emits code for Copy/CopyKeep.
1696 static void Copy_emitter(const ir_node *node, const ir_node *op)
1698 const arch_register_t *in = arch_get_irn_register(op);
1699 const arch_register_t *out = arch_get_irn_register(node);
1704 if (is_unknown_reg(in))
1706 /* copies of vf nodes aren't real... */
1707 if (arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp])
1710 if (get_irn_mode(node) == mode_E) {
1711 ia32_emitf(node, "\tmovsd %R, %R\n", in, out);
1713 ia32_emitf(node, "\tmovl %R, %R\n", in, out);
1717 static void emit_be_Copy(const ir_node *node)
1719 Copy_emitter(node, be_get_Copy_op(node));
1722 static void emit_be_CopyKeep(const ir_node *node)
1724 Copy_emitter(node, be_get_CopyKeep_op(node));
1728 * Emits code for exchange.
1730 static void emit_be_Perm(const ir_node *node)
1732 const arch_register_t *in0, *in1;
1733 const arch_register_class_t *cls0, *cls1;
1735 in0 = arch_get_irn_register(get_irn_n(node, 0));
1736 in1 = arch_get_irn_register(get_irn_n(node, 1));
1738 cls0 = arch_register_get_class(in0);
1739 cls1 = arch_register_get_class(in1);
1741 assert(cls0 == cls1 && "Register class mismatch at Perm");
1743 if (cls0 == &ia32_reg_classes[CLASS_ia32_gp]) {
1744 ia32_emitf(node, "\txchg %R, %R\n", in1, in0);
1745 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_xmm]) {
1746 ia32_emitf(NULL, "\txorpd %R, %R\n", in1, in0);
1747 ia32_emitf(NULL, "\txorpd %R, %R\n", in0, in1);
1748 ia32_emitf(node, "\txorpd %R, %R\n", in1, in0);
1749 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_vfp]) {
1751 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_st]) {
1754 panic("unexpected register class in be_Perm (%+F)", node);
1759 * Emits code for Constant loading.
1761 static void emit_ia32_Const(const ir_node *node)
1763 ia32_emitf(node, "\tmovl %I, %D0\n");
1767 * Emits code to load the TLS base
1769 static void emit_ia32_LdTls(const ir_node *node)
1771 ia32_emitf(node, "\tmovl %%gs:0, %D0\n");
1774 /* helper function for emit_ia32_Minus64Bit */
1775 static void emit_mov(const ir_node* node, const arch_register_t *src, const arch_register_t *dst)
1777 ia32_emitf(node, "\tmovl %R, %R\n", src, dst);
1780 /* helper function for emit_ia32_Minus64Bit */
1781 static void emit_neg(const ir_node* node, const arch_register_t *reg)
1783 ia32_emitf(node, "\tnegl %R\n", reg);
1786 /* helper function for emit_ia32_Minus64Bit */
1787 static void emit_sbb0(const ir_node* node, const arch_register_t *reg)
1789 ia32_emitf(node, "\tsbbl $0, %R\n", reg);
1792 /* helper function for emit_ia32_Minus64Bit */
1793 static void emit_sbb(const ir_node* node, const arch_register_t *src, const arch_register_t *dst)
1795 ia32_emitf(node, "\tsbbl %R, %R\n", src, dst);
1798 /* helper function for emit_ia32_Minus64Bit */
1799 static void emit_xchg(const ir_node* node, const arch_register_t *src, const arch_register_t *dst)
1801 ia32_emitf(node, "\txchgl %R, %R\n", src, dst);
1804 /* helper function for emit_ia32_Minus64Bit */
1805 static void emit_zero(const ir_node* node, const arch_register_t *reg)
1807 ia32_emitf(node, "\txorl %R, %R\n", reg, reg);
1810 static void emit_ia32_Minus64Bit(const ir_node *node)
1812 const arch_register_t *in_lo = get_in_reg(node, 0);
1813 const arch_register_t *in_hi = get_in_reg(node, 1);
1814 const arch_register_t *out_lo = get_out_reg(node, 0);
1815 const arch_register_t *out_hi = get_out_reg(node, 1);
1817 if (out_lo == in_lo) {
1818 if (out_hi != in_hi) {
1819 /* a -> a, b -> d */
1822 /* a -> a, b -> b */
1825 } else if (out_lo == in_hi) {
1826 if (out_hi == in_lo) {
1827 /* a -> b, b -> a */
1828 emit_xchg(node, in_lo, in_hi);
1831 /* a -> b, b -> d */
1832 emit_mov(node, in_hi, out_hi);
1833 emit_mov(node, in_lo, out_lo);
1837 if (out_hi == in_lo) {
1838 /* a -> c, b -> a */
1839 emit_mov(node, in_lo, out_lo);
1841 } else if (out_hi == in_hi) {
1842 /* a -> c, b -> b */
1843 emit_mov(node, in_lo, out_lo);
1846 /* a -> c, b -> d */
1847 emit_mov(node, in_lo, out_lo);
1853 emit_neg( node, out_hi);
1854 emit_neg( node, out_lo);
1855 emit_sbb0(node, out_hi);
1859 emit_zero(node, out_hi);
1860 emit_neg( node, out_lo);
1861 emit_sbb( node, in_hi, out_hi);
1864 static void emit_ia32_GetEIP(const ir_node *node)
1866 ia32_emitf(node, "\tcall %s\n", pic_base_label);
1867 ia32_emitf(NULL, "%s:\n", pic_base_label);
1868 ia32_emitf(node, "\tpopl %D0\n");
1871 static void emit_ia32_ClimbFrame(const ir_node *node)
1873 const ia32_climbframe_attr_t *attr = get_ia32_climbframe_attr_const(node);
1875 ia32_emitf(node, "\tmovl %S0, %D0\n");
1876 ia32_emitf(node, "\tmovl $%u, %S1\n", attr->count);
1877 be_gas_emit_block_name(node);
1878 be_emit_cstring(":\n");
1879 be_emit_write_line();
1880 ia32_emitf(node, "\tmovl (%D0), %D0\n");
1881 ia32_emitf(node, "\tdec %S1\n");
1882 be_emit_cstring("\tjnz ");
1883 be_gas_emit_block_name(node);
1884 be_emit_finish_line_gas(node);
1887 static void emit_be_Return(const ir_node *node)
1889 unsigned pop = be_Return_get_pop(node);
1891 if (pop > 0 || be_Return_get_emit_pop(node)) {
1892 ia32_emitf(node, "\tret $%u\n", pop);
1894 ia32_emitf(node, "\tret\n");
1898 static void emit_Nothing(const ir_node *node)
1904 /***********************************************************************************
1907 * _ __ ___ __ _ _ _ __ | |_ _ __ __ _ _ __ ___ _____ _____ _ __| | __
1908 * | '_ ` _ \ / _` | | '_ \ | _| '__/ _` | '_ ` _ \ / _ \ \ /\ / / _ \| '__| |/ /
1909 * | | | | | | (_| | | | | | | | | | | (_| | | | | | | __/\ V V / (_) | | | <
1910 * |_| |_| |_|\__,_|_|_| |_| |_| |_| \__,_|_| |_| |_|\___| \_/\_/ \___/|_| |_|\_\
1912 ***********************************************************************************/
1915 * Enters the emitter functions for handled nodes into the generic
1916 * pointer of an opcode.
1918 static void ia32_register_emitters(void)
1920 #define IA32_EMIT2(a,b) op_ia32_##a->ops.generic = (op_func)emit_ia32_##b
1921 #define IA32_EMIT(a) IA32_EMIT2(a,a)
1922 #define EMIT(a) op_##a->ops.generic = (op_func)emit_##a
1923 #define IGN(a) op_##a->ops.generic = (op_func)emit_Nothing
1924 #define BE_EMIT(a) op_be_##a->ops.generic = (op_func)emit_be_##a
1925 #define BE_IGN(a) op_be_##a->ops.generic = (op_func)emit_Nothing
1927 /* first clear the generic function pointer for all ops */
1928 clear_irp_opcodes_generic_func();
1930 /* register all emitter functions defined in spec */
1931 ia32_register_spec_emitters();
1933 /* other ia32 emitter functions */
1934 IA32_EMIT2(Conv_I2I8Bit, Conv_I2I);
1939 IA32_EMIT(Conv_FP2FP);
1940 IA32_EMIT(Conv_FP2I);
1941 IA32_EMIT(Conv_I2FP);
1942 IA32_EMIT(Conv_I2I);
1950 IA32_EMIT(Minus64Bit);
1951 IA32_EMIT(SwitchJmp);
1952 IA32_EMIT(ClimbFrame);
1955 /* benode emitter */
1976 typedef void (*emit_func_ptr) (const ir_node *);
1979 * Assign and emit an exception label if the current instruction can fail.
1981 static void ia32_assign_exc_label(ir_node *node)
1983 /* assign a new ID to the instruction */
1984 set_ia32_exc_label_id(node, ++exc_label_id);
1986 ia32_emit_exc_label(node);
1988 be_emit_pad_comment();
1989 be_emit_cstring("/* exception to Block ");
1990 ia32_emit_cfop_target(node);
1991 be_emit_cstring(" */\n");
1992 be_emit_write_line();
1996 * Emits code for a node.
1998 static void ia32_emit_node(ir_node *node)
2000 ir_op *op = get_irn_op(node);
2002 DBG((dbg, LEVEL_1, "emitting code for %+F\n", node));
2004 if (is_ia32_irn(node)) {
2005 if (get_ia32_exc_label(node)) {
2006 /* emit the exception label of this instruction */
2007 ia32_assign_exc_label(node);
2009 if (mark_spill_reload) {
2010 if (is_ia32_is_spill(node)) {
2011 ia32_emitf(NULL, "\txchg %ebx, %ebx /* spill mark */\n");
2013 if (is_ia32_is_reload(node)) {
2014 ia32_emitf(NULL, "\txchg %edx, %edx /* reload mark */\n");
2016 if (is_ia32_is_remat(node)) {
2017 ia32_emitf(NULL, "\txchg %ecx, %ecx /* remat mark */\n");
2021 if (op->ops.generic) {
2022 emit_func_ptr func = (emit_func_ptr) op->ops.generic;
2024 be_dbg_set_dbg_info(get_irn_dbg_info(node));
2029 ir_fprintf(stderr, "Error: No emit handler for node %+F (%+G, graph %+F)\n", node, node, current_ir_graph);
2035 * Emits gas alignment directives
2037 static void ia32_emit_alignment(unsigned align, unsigned skip)
2039 ia32_emitf(NULL, "\t.p2align %u,,%u\n", align, skip);
2043 * Emits gas alignment directives for Labels depended on cpu architecture.
2045 static void ia32_emit_align_label(void)
2047 unsigned align = ia32_cg_config.label_alignment;
2048 unsigned maximum_skip = ia32_cg_config.label_alignment_max_skip;
2049 ia32_emit_alignment(align, maximum_skip);
2053 * Test whether a block should be aligned.
2054 * For cpus in the P4/Athlon class it is useful to align jump labels to
2055 * 16 bytes. However we should only do that if the alignment nops before the
2056 * label aren't executed more often than we have jumps to the label.
2058 static int should_align_block(const ir_node *block)
2060 static const double DELTA = .0001;
2061 ir_exec_freq *exec_freq = cg->birg->exec_freq;
2062 ir_node *prev = get_prev_block_sched(block);
2064 double prev_freq = 0; /**< execfreq of the fallthrough block */
2065 double jmp_freq = 0; /**< execfreq of all non-fallthrough blocks */
2068 if (exec_freq == NULL)
2070 if (ia32_cg_config.label_alignment_factor <= 0)
2073 block_freq = get_block_execfreq(exec_freq, block);
2074 if (block_freq < DELTA)
2077 n_cfgpreds = get_Block_n_cfgpreds(block);
2078 for(i = 0; i < n_cfgpreds; ++i) {
2079 const ir_node *pred = get_Block_cfgpred_block(block, i);
2080 double pred_freq = get_block_execfreq(exec_freq, pred);
2083 prev_freq += pred_freq;
2085 jmp_freq += pred_freq;
2089 if (prev_freq < DELTA && !(jmp_freq < DELTA))
2092 jmp_freq /= prev_freq;
2094 return jmp_freq > ia32_cg_config.label_alignment_factor;
2098 * Emit the block header for a block.
2100 * @param block the block
2101 * @param prev_block the previous block
2103 static void ia32_emit_block_header(ir_node *block)
2105 ir_graph *irg = current_ir_graph;
2106 int need_label = block_needs_label(block);
2108 ir_exec_freq *exec_freq = cg->birg->exec_freq;
2110 if (block == get_irg_end_block(irg))
2113 if (ia32_cg_config.label_alignment > 0) {
2114 /* align the current block if:
2115 * a) if should be aligned due to its execution frequency
2116 * b) there is no fall-through here
2118 if (should_align_block(block)) {
2119 ia32_emit_align_label();
2121 /* if the predecessor block has no fall-through,
2122 we can always align the label. */
2124 int has_fallthrough = 0;
2126 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
2127 ir_node *cfg_pred = get_Block_cfgpred(block, i);
2128 if (can_be_fallthrough(cfg_pred)) {
2129 has_fallthrough = 1;
2134 if (!has_fallthrough)
2135 ia32_emit_align_label();
2140 be_gas_emit_block_name(block);
2143 be_emit_pad_comment();
2144 be_emit_cstring(" /* ");
2146 be_emit_cstring("\t/* ");
2147 be_gas_emit_block_name(block);
2148 be_emit_cstring(": ");
2151 be_emit_cstring("preds:");
2153 /* emit list of pred blocks in comment */
2154 arity = get_irn_arity(block);
2156 be_emit_cstring(" none");
2158 for (i = 0; i < arity; ++i) {
2159 ir_node *predblock = get_Block_cfgpred_block(block, i);
2160 be_emit_irprintf(" %d", get_irn_node_nr(predblock));
2163 if (exec_freq != NULL) {
2164 be_emit_irprintf(", freq: %f",
2165 get_block_execfreq(exec_freq, block));
2167 be_emit_cstring(" */\n");
2168 be_emit_write_line();
2172 * Walks over the nodes in a block connected by scheduling edges
2173 * and emits code for each node.
2175 static void ia32_gen_block(ir_node *block)
2179 ia32_emit_block_header(block);
2181 /* emit the contents of the block */
2182 be_dbg_set_dbg_info(get_irn_dbg_info(block));
2183 sched_foreach(block, node) {
2184 ia32_emit_node(node);
2188 typedef struct exc_entry {
2189 ir_node *exc_instr; /** The instruction that can issue an exception. */
2190 ir_node *block; /** The block to call then. */
2195 * Sets labels for control flow nodes (jump target).
2196 * Links control predecessors to there destination blocks.
2198 static void ia32_gen_labels(ir_node *block, void *data)
2200 exc_entry **exc_list = data;
2204 for (n = get_Block_n_cfgpreds(block) - 1; n >= 0; --n) {
2205 pred = get_Block_cfgpred(block, n);
2206 set_irn_link(pred, block);
2208 pred = skip_Proj(pred);
2209 if (is_ia32_irn(pred) && get_ia32_exc_label(pred)) {
2214 ARR_APP1(exc_entry, *exc_list, e);
2215 set_irn_link(pred, block);
2221 * Compare two exception_entries.
2223 static int cmp_exc_entry(const void *a, const void *b)
2225 const exc_entry *ea = a;
2226 const exc_entry *eb = b;
2228 if (get_ia32_exc_label_id(ea->exc_instr) < get_ia32_exc_label_id(eb->exc_instr))
2234 * Main driver. Emits the code for one routine.
2236 void ia32_gen_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg)
2238 ir_entity *entity = get_irg_entity(irg);
2239 exc_entry *exc_list = NEW_ARR_F(exc_entry, 0);
2244 do_pic = cg->birg->main_env->options->pic;
2246 be_gas_elf_type_char = '@';
2248 ia32_register_emitters();
2250 get_unique_label(pic_base_label, sizeof(pic_base_label), ".PIC_BASE");
2252 be_dbg_method_begin(entity, be_abi_get_stack_layout(cg->birg->abi));
2253 be_gas_emit_function_prolog(entity, ia32_cg_config.function_alignment);
2255 /* we use links to point to target blocks */
2256 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
2257 irg_block_walk_graph(irg, ia32_gen_labels, NULL, &exc_list);
2259 /* initialize next block links */
2260 n = ARR_LEN(cg->blk_sched);
2261 for (i = 0; i < n; ++i) {
2262 ir_node *block = cg->blk_sched[i];
2263 ir_node *prev = i > 0 ? cg->blk_sched[i-1] : NULL;
2265 set_irn_link(block, prev);
2268 for (i = 0; i < n; ++i) {
2269 ir_node *block = cg->blk_sched[i];
2271 ia32_gen_block(block);
2274 be_gas_emit_function_epilog(entity);
2275 be_dbg_method_end();
2277 be_emit_write_line();
2279 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2281 /* Sort the exception table using the exception label id's.
2282 Those are ascending with ascending addresses. */
2283 qsort(exc_list, ARR_LEN(exc_list), sizeof(exc_list[0]), cmp_exc_entry);
2287 for (i = 0; i < ARR_LEN(exc_list); ++i) {
2288 be_emit_cstring("\t.long ");
2289 ia32_emit_exc_label(exc_list[i].exc_instr);
2291 be_emit_cstring("\t.long ");
2292 be_gas_emit_block_name(exc_list[i].block);
2296 DEL_ARR_F(exc_list);
2299 static const lc_opt_table_entry_t ia32_emitter_options[] = {
2300 LC_OPT_ENT_BOOL("mark_spill_reload", "mark spills and reloads with ud opcodes", &mark_spill_reload),
2304 /* ==== Experimental binary emitter ==== */
2306 static unsigned char reg_gp_map[N_ia32_gp_REGS];
2307 //static unsigned char reg_mmx_map[N_ia32_mmx_REGS];
2308 //static unsigned char reg_sse_map[N_ia32_xmm_REGS];
2309 static unsigned char pnc_map_signed[8];
2310 static unsigned char pnc_map_unsigned[8];
2312 static void build_reg_map(void)
2314 reg_gp_map[REG_EAX] = 0x0;
2315 reg_gp_map[REG_ECX] = 0x1;
2316 reg_gp_map[REG_EDX] = 0x2;
2317 reg_gp_map[REG_EBX] = 0x3;
2318 reg_gp_map[REG_ESP] = 0x4;
2319 reg_gp_map[REG_EBP] = 0x5;
2320 reg_gp_map[REG_ESI] = 0x6;
2321 reg_gp_map[REG_EDI] = 0x7;
2323 pnc_map_signed[pn_Cmp_Eq] = 0x04;
2324 pnc_map_signed[pn_Cmp_Lt] = 0x0C;
2325 pnc_map_signed[pn_Cmp_Le] = 0x0E;
2326 pnc_map_signed[pn_Cmp_Gt] = 0x0F;
2327 pnc_map_signed[pn_Cmp_Ge] = 0x0D;
2328 pnc_map_signed[pn_Cmp_Lg] = 0x05;
2330 pnc_map_unsigned[pn_Cmp_Eq] = 0x04;
2331 pnc_map_unsigned[pn_Cmp_Lt] = 0x02;
2332 pnc_map_unsigned[pn_Cmp_Le] = 0x06;
2333 pnc_map_unsigned[pn_Cmp_Gt] = 0x07;
2334 pnc_map_unsigned[pn_Cmp_Ge] = 0x03;
2335 pnc_map_unsigned[pn_Cmp_Lg] = 0x05;
2338 /** Returns the encoding for a pnc field. */
2339 static unsigned char pnc2cc(int pnc)
2342 if (pnc == ia32_pn_Cmp_parity) {
2344 } else if (pnc & ia32_pn_Cmp_float || pnc & ia32_pn_Cmp_unsigned) {
2345 cc = pnc_map_unsigned[pnc & 0x07];
2347 cc = pnc_map_signed[pnc & 0x07];
2353 /** Sign extension bit values for binops */
2355 UNSIGNED_IMM = 0, /**< unsigned immediate */
2356 SIGNEXT_IMM = 2, /**< sign extended immediate */
2359 /** The mod encoding of the ModR/M */
2361 MOD_IND = 0x00, /**< [reg1] */
2362 MOD_IND_BYTE_OFS = 0x40, /**< [reg1 + byte ofs] */
2363 MOD_IND_WORD_OFS = 0x80, /**< [reg1 + word ofs] */
2364 MOD_REG = 0xC0 /**< reg1 */
2367 /** create R/M encoding for ModR/M */
2368 #define ENC_RM(x) (x)
2369 /** create REG encoding for ModR/M */
2370 #define ENC_REG(x) ((x) << 3)
2372 /** create encoding for a SIB byte */
2373 #define ENC_SIB(scale, index, base) ((scale) << 6 | (index) << 3 | (base))
2375 /* Node: The following routines are supposed to append bytes, words, dwords
2376 to the output stream.
2377 Currently the implementation is stupid in that it still creates output
2378 for an "assembler" in the form of .byte, .long
2379 We will change this when enough infrastructure is there to create complete
2380 machine code in memory/object files */
2382 static void bemit8(const unsigned char byte)
2384 be_emit_irprintf("\t.byte 0x%x\n", byte);
2385 be_emit_write_line();
2388 static void bemit16(const unsigned short u16)
2390 be_emit_irprintf("\t.word 0x%x\n", u16);
2391 be_emit_write_line();
2394 static void bemit32(const unsigned u32)
2396 be_emit_irprintf("\t.long 0x%x\n", u32);
2397 be_emit_write_line();
2401 * Emit address of an entity. If @p is_relative is true then a relative
2402 * offset from behind the address to the entity is created.
2404 static void bemit_entity(ir_entity *entity, bool entity_sign, int offset,
2407 if (entity == NULL) {
2412 /* the final version should remember the position in the bytestream
2413 and patch it with the correct address at linktime... */
2414 be_emit_cstring("\t.long ");
2417 be_gas_emit_entity(entity);
2419 if (get_entity_owner(entity) == get_tls_type()) {
2420 if (get_entity_visibility(entity) == ir_visibility_external) {
2421 be_emit_cstring("@INDNTPOFF");
2423 be_emit_cstring("@NTPOFF");
2428 be_emit_cstring("-.");
2433 be_emit_irprintf("%+d", offset);
2436 be_emit_write_line();
2439 static void bemit_jmp_destination(const ir_node *dest_block)
2441 be_emit_cstring("\t.long ");
2442 be_gas_emit_block_name(dest_block);
2443 be_emit_cstring(" - . - 4\n");
2444 be_emit_write_line();
2447 /* end emit routines, all emitters following here should only use the functions
2450 typedef enum reg_modifier {
2455 /** Create a ModR/M byte for src1,src2 registers */
2456 static void bemit_modrr(const arch_register_t *src1,
2457 const arch_register_t *src2)
2459 unsigned char modrm = MOD_REG;
2460 modrm |= ENC_RM(reg_gp_map[src1->index]);
2461 modrm |= ENC_REG(reg_gp_map[src2->index]);
2465 /** Create a ModR/M8 byte for src1,src2 registers */
2466 static void bemit_modrr8(reg_modifier_t high_part1, const arch_register_t *src1,
2467 reg_modifier_t high_part2, const arch_register_t *src2)
2469 unsigned char modrm = MOD_REG;
2470 modrm |= ENC_RM(reg_gp_map[src1->index] + (high_part1 == REG_HIGH ? 4 : 0));
2471 modrm |= ENC_REG(reg_gp_map[src2->index] + (high_part2 == REG_HIGH ? 4 : 0));
2475 /** Create a ModR/M byte for one register and extension */
2476 static void bemit_modru(const arch_register_t *reg, unsigned ext)
2478 unsigned char modrm = MOD_REG;
2480 modrm |= ENC_RM(reg_gp_map[reg->index]);
2481 modrm |= ENC_REG(ext);
2485 /** Create a ModR/M8 byte for one register */
2486 static void bemit_modrm8(reg_modifier_t high_part, const arch_register_t *reg)
2488 unsigned char modrm = MOD_REG;
2489 assert(reg_gp_map[reg->index] < 4);
2490 modrm |= ENC_RM(reg_gp_map[reg->index] + (high_part == REG_HIGH ? 4 : 0));
2496 * Calculate the size of an signed immediate in bytes.
2498 * @param offset an offset
2500 static unsigned get_signed_imm_size(int offset)
2502 if (-128 <= offset && offset < 128) {
2504 } else if (-32768 <= offset && offset < 32768) {
2512 * Emit an address mode.
2514 * @param reg content of the reg field: either a register index or an opcode extension
2515 * @param node the node
2517 static void bemit_mod_am(unsigned reg, const ir_node *node)
2519 ir_entity *ent = get_ia32_am_sc(node);
2520 int offs = get_ia32_am_offs_int(node);
2521 ir_node *base = get_irn_n(node, n_ia32_base);
2522 int has_base = !is_ia32_NoReg_GP(base);
2523 ir_node *index = get_irn_n(node, n_ia32_index);
2524 int has_index = !is_ia32_NoReg_GP(index);
2527 unsigned emitoffs = 0;
2528 bool emitsib = false;
2531 /* set the mod part depending on displacement */
2533 modrm |= MOD_IND_WORD_OFS;
2535 } else if (offs == 0) {
2538 } else if (-128 <= offs && offs < 128) {
2539 modrm |= MOD_IND_BYTE_OFS;
2542 modrm |= MOD_IND_WORD_OFS;
2547 const arch_register_t *base_reg = arch_get_irn_register(base);
2548 base_enc = reg_gp_map[base_reg->index];
2550 /* Use the EBP encoding + MOD_IND if NO base register. There is
2551 * always a 32bit offset present in this case. */
2557 /* Determine if we need a SIB byte. */
2559 const arch_register_t *reg_index = arch_get_irn_register(index);
2560 int scale = get_ia32_am_scale(node);
2562 /* R/M set to ESP means SIB in 32bit mode. */
2563 modrm |= ENC_RM(0x04);
2564 sib = ENC_SIB(scale, reg_gp_map[reg_index->index], base_enc);
2566 } else if (base_enc == 0x04) {
2567 /* for the above reason we are forced to emit a SIB when base is ESP.
2568 * Only the base is used, index must be ESP too, which means no index.
2570 modrm |= ENC_RM(0x04);
2571 sib = ENC_SIB(0, 0x04, 0x04);
2574 modrm |= ENC_RM(base_enc);
2577 /* We are forced to emit an 8bit offset as EBP base without offset is a
2578 * special case for SIB without base register. */
2579 if (base_enc == 0x05 && emitoffs == 0) {
2580 modrm |= MOD_IND_BYTE_OFS;
2584 modrm |= ENC_REG(reg);
2590 /* emit displacement */
2591 if (emitoffs == 8) {
2592 bemit8((unsigned) offs);
2593 } else if (emitoffs == 32) {
2594 bemit_entity(ent, is_ia32_am_sc_sign(node), offs, false);
2599 * Emit a binop with a immediate operand.
2601 * @param node the node to emit
2602 * @param opcode_eax the opcode for the op eax, imm variant
2603 * @param opcode the opcode for the reg, imm variant
2604 * @param ruval the opcode extension for opcode
2606 static void bemit_binop_with_imm(
2607 const ir_node *node,
2608 unsigned char opcode_ax,
2609 unsigned char opcode, unsigned char ruval)
2611 /* Use in-reg, because some instructions (cmp, test) have no out-reg. */
2612 const ir_node *op = get_irn_n(node, n_ia32_binary_right);
2613 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(op);
2616 /* Some instructions (test) have no short form with 32bit value + 8bit
2618 if (attr->symconst != NULL || opcode & SIGNEXT_IMM) {
2621 /* check for sign extension */
2622 size = get_signed_imm_size(attr->offset);
2627 bemit8(opcode | SIGNEXT_IMM);
2628 /* cmp has this special mode */
2629 if (get_ia32_op_type(node) == ia32_AddrModeS) {
2630 bemit_mod_am(ruval, node);
2632 const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left);
2633 bemit_modru(reg, ruval);
2635 bemit8((unsigned char)attr->offset);
2639 /* check for eax variant: this variant is shorter for 32bit immediates only */
2640 if (get_ia32_op_type(node) == ia32_AddrModeS) {
2642 bemit_mod_am(ruval, node);
2644 const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left);
2645 if (reg->index == REG_EAX) {
2649 bemit_modru(reg, ruval);
2652 bemit_entity(attr->symconst, attr->sc_sign, attr->offset, false);
2655 panic("invalid imm size?!?");
2661 static void bemit_binop_2(const ir_node *node, unsigned code)
2663 const arch_register_t *out = get_in_reg(node, n_ia32_binary_left);
2665 if (get_ia32_op_type(node) == ia32_Normal) {
2666 const arch_register_t *op2 = get_in_reg(node, n_ia32_binary_right);
2667 bemit_modrr(op2, out);
2669 bemit_mod_am(reg_gp_map[out->index], node);
2676 static void bemit_binop(const ir_node *node, const unsigned char opcodes[4])
2678 ir_node *right = get_irn_n(node, n_ia32_binary_right);
2679 if (is_ia32_Immediate(right)) {
2680 bemit_binop_with_imm(node, opcodes[1], opcodes[2], opcodes[3]);
2682 bemit_binop_2(node, opcodes[0]);
2689 static void bemit_unop(const ir_node *node, unsigned char code, unsigned char ext, int input)
2692 if (get_ia32_op_type(node) == ia32_Normal) {
2693 const arch_register_t *in = get_in_reg(node, input);
2694 bemit_modru(in, ext);
2696 bemit_mod_am(ext, node);
2700 static void bemit_unop_reg(const ir_node *node, unsigned char code, int input)
2702 const arch_register_t *out = get_out_reg(node, 0);
2703 bemit_unop(node, code, reg_gp_map[out->index], input);
2706 static void bemit_unop_mem(const ir_node *node, unsigned char code, unsigned char ext)
2708 unsigned size = get_mode_size_bits(get_ia32_ls_mode(node));
2711 bemit8(size == 8 ? code : code + 1);
2712 bemit_mod_am(ext, node);
2715 static void bemit_immediate(const ir_node *node, bool relative)
2717 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node);
2718 bemit_entity(attr->symconst, attr->sc_sign, attr->offset, relative);
2721 static void bemit_copy(const ir_node *copy)
2723 const arch_register_t *in = get_in_reg(copy, 0);
2724 const arch_register_t *out = get_out_reg(copy, 0);
2726 if (in == out || is_unknown_reg(in))
2728 /* copies of vf nodes aren't real... */
2729 if (arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp])
2732 if (get_irn_mode(copy) == mode_E) {
2735 assert(arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_gp]);
2737 bemit_modrr(in, out);
2741 static void bemit_perm(const ir_node *node)
2743 const arch_register_t *in0 = arch_get_irn_register(get_irn_n(node, 0));
2744 const arch_register_t *in1 = arch_get_irn_register(get_irn_n(node, 1));
2745 const arch_register_class_t *cls0 = arch_register_get_class(in0);
2747 assert(cls0 == arch_register_get_class(in1) && "Register class mismatch at Perm");
2749 if (cls0 == &ia32_reg_classes[CLASS_ia32_gp]) {
2750 if (in0->index == REG_EAX) {
2751 bemit8(0x90 + reg_gp_map[in1->index]);
2752 } else if (in1->index == REG_EAX) {
2753 bemit8(0x90 + reg_gp_map[in0->index]);
2756 bemit_modrr(in0, in1);
2758 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_xmm]) {
2759 panic("unimplemented"); // TODO implement
2760 //ia32_emitf(NULL, "\txorpd %R, %R\n", in1, in0);
2761 //ia32_emitf(NULL, "\txorpd %R, %R\n", in0, in1);
2762 //ia32_emitf(node, "\txorpd %R, %R\n", in1, in0);
2763 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_vfp]) {
2765 } else if (cls0 == &ia32_reg_classes[CLASS_ia32_st]) {
2768 panic("unexpected register class in be_Perm (%+F)", node);
2772 static void bemit_xor0(const ir_node *node)
2774 const arch_register_t *out = get_out_reg(node, 0);
2776 bemit_modrr(out, out);
2779 static void bemit_mov_const(const ir_node *node)
2781 const arch_register_t *out = get_out_reg(node, 0);
2782 bemit8(0xB8 + reg_gp_map[out->index]);
2783 bemit_immediate(node, false);
2787 * Creates a function for a Binop with 3 possible encodings.
2789 #define BINOP(op, op0, op1, op2, op2_ext) \
2790 static void bemit_ ## op(const ir_node *node) { \
2791 static const unsigned char op ## _codes[] = {op0, op1, op2, op2_ext}; \
2792 bemit_binop(node, op ## _codes); \
2795 /* insn def eax,imm imm */
2796 BINOP(add, 0x03, 0x05, 0x81, 0)
2797 BINOP(or, 0x0B, 0x0D, 0x81, 1)
2798 BINOP(adc, 0x13, 0x15, 0x81, 2)
2799 BINOP(sbb, 0x1B, 0x1D, 0x81, 3)
2800 BINOP(and, 0x23, 0x25, 0x81, 4)
2801 BINOP(sub, 0x2B, 0x2D, 0x81, 5)
2802 BINOP(xor, 0x33, 0x35, 0x81, 6)
2803 BINOP(test, 0x85, 0xA9, 0xF7, 0)
2805 #define BINOPMEM(op, ext) \
2806 static void bemit_##op(const ir_node *node) \
2809 unsigned size = get_mode_size_bits(get_ia32_ls_mode(node)); \
2812 val = get_irn_n(node, n_ia32_unary_op); \
2813 if (is_ia32_Immediate(val)) { \
2814 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(val); \
2815 int offset = attr->offset; \
2816 if (attr->symconst == NULL && get_signed_imm_size(offset) == 1) { \
2818 bemit_mod_am(ext, node); \
2822 bemit_mod_am(ext, node); \
2826 bemit_entity(attr->symconst, attr->sc_sign, offset, false); \
2830 bemit8(ext << 3 | 1); \
2831 bemit_mod_am(reg_gp_map[get_out_reg(val, 0)->index], node); \
2835 static void bemit_##op##8bit(const ir_node *node) \
2837 ir_node *val = get_irn_n(node, n_ia32_unary_op); \
2838 if (is_ia32_Immediate(val)) { \
2840 bemit_mod_am(ext, node); \
2841 bemit8(get_ia32_immediate_attr_const(val)->offset); \
2844 bemit_mod_am(reg_gp_map[get_out_reg(val, 0)->index], node); \
2856 * Creates a function for an Unop with code /ext encoding.
2858 #define UNOP(op, code, ext, input) \
2859 static void bemit_ ## op(const ir_node *node) { \
2860 bemit_unop(node, code, ext, input); \
2863 UNOP(not, 0xF7, 2, n_ia32_Not_val)
2864 UNOP(neg, 0xF7, 3, n_ia32_Neg_val)
2865 UNOP(mul, 0xF7, 4, n_ia32_Mul_right)
2866 UNOP(imul1op, 0xF7, 5, n_ia32_IMul1OP_right)
2867 UNOP(div, 0xF7, 6, n_ia32_Div_divisor)
2868 UNOP(idiv, 0xF7, 7, n_ia32_IDiv_divisor)
2870 /* TODO: am support for IJmp */
2871 UNOP(ijmp, 0xFF, 4, n_ia32_IJmp_target)
2873 #define SHIFT(op, ext) \
2874 static void bemit_##op(const ir_node *node) \
2876 const arch_register_t *out = get_out_reg(node, 0); \
2877 ir_node *count = get_irn_n(node, 1); \
2878 if (is_ia32_Immediate(count)) { \
2879 int offset = get_ia32_immediate_attr_const(count)->offset; \
2880 if (offset == 1) { \
2882 bemit_modru(out, ext); \
2885 bemit_modru(out, ext); \
2890 bemit_modru(out, ext); \
2894 static void bemit_##op##mem(const ir_node *node) \
2897 unsigned size = get_mode_size_bits(get_ia32_ls_mode(node)); \
2900 count = get_irn_n(node, 1); \
2901 if (is_ia32_Immediate(count)) { \
2902 int offset = get_ia32_immediate_attr_const(count)->offset; \
2903 if (offset == 1) { \
2904 bemit8(size == 8 ? 0xD0 : 0xD1); \
2905 bemit_mod_am(ext, node); \
2907 bemit8(size == 8 ? 0xC0 : 0xC1); \
2908 bemit_mod_am(ext, node); \
2912 bemit8(size == 8 ? 0xD2 : 0xD3); \
2913 bemit_mod_am(ext, node); \
2923 static void bemit_shld(const ir_node *node)
2925 const arch_register_t *in = get_in_reg(node, n_ia32_ShlD_val_low);
2926 const arch_register_t *out = get_out_reg(node, pn_ia32_ShlD_res);
2927 ir_node *count = get_irn_n(node, n_ia32_ShlD_count);
2929 if (is_ia32_Immediate(count)) {
2931 bemit_modrr(out, in);
2932 bemit8(get_ia32_immediate_attr_const(count)->offset);
2935 bemit_modrr(out, in);
2939 static void bemit_shrd(const ir_node *node)
2941 const arch_register_t *in = get_in_reg(node, n_ia32_ShrD_val_low);
2942 const arch_register_t *out = get_out_reg(node, pn_ia32_ShrD_res);
2943 ir_node *count = get_irn_n(node, n_ia32_ShrD_count);
2945 if (is_ia32_Immediate(count)) {
2947 bemit_modrr(out, in);
2948 bemit8(get_ia32_immediate_attr_const(count)->offset);
2951 bemit_modrr(out, in);
2956 * binary emitter for setcc.
2958 static void bemit_setcc(const ir_node *node)
2960 const arch_register_t *dreg = get_out_reg(node, pn_ia32_Setcc_res);
2962 pn_Cmp pnc = get_ia32_condcode(node);
2963 pnc = determine_final_pnc(node, n_ia32_Setcc_eflags, pnc);
2964 if (pnc & ia32_pn_Cmp_float) {
2965 switch (pnc & 0x0f) {
2970 bemit_modrm8(REG_LOW, dreg);
2977 bemit_modrm8(REG_LOW, dreg);
2985 bemit8(0x90 | pnc2cc(pnc));
2986 bemit_modrm8(REG_LOW, dreg);
2991 bemit_modrm8(REG_HIGH, dreg);
2993 /* andb %>dreg, %<dreg */
2995 bemit_modrr8(REG_LOW, dreg, REG_HIGH, dreg);
3003 bemit8(0x90 | pnc2cc(pnc));
3004 bemit_modrm8(REG_LOW, dreg);
3009 bemit_modrm8(REG_HIGH, dreg);
3011 /* orb %>dreg, %<dreg */
3013 bemit_modrr8(REG_LOW, dreg, REG_HIGH, dreg);
3022 bemit8(0x90 | pnc2cc(pnc));
3023 bemit_modrm8(REG_LOW, dreg);
3026 static void bemit_cmovcc(const ir_node *node)
3028 const ia32_attr_t *attr = get_ia32_attr_const(node);
3029 int ins_permuted = attr->data.ins_permuted;
3030 const arch_register_t *out = arch_irn_get_register(node, pn_ia32_res);
3031 pn_Cmp pnc = get_ia32_condcode(node);
3032 const arch_register_t *in_true;
3033 const arch_register_t *in_false;
3035 pnc = determine_final_pnc(node, n_ia32_CMovcc_eflags, pnc);
3037 in_true = arch_get_irn_register(get_irn_n(node, n_ia32_CMovcc_val_true));
3038 in_false = arch_get_irn_register(get_irn_n(node, n_ia32_CMovcc_val_false));
3040 /* should be same constraint fullfilled? */
3041 if (out == in_false) {
3042 /* yes -> nothing to do */
3043 } else if (out == in_true) {
3044 assert(get_ia32_op_type(node) == ia32_Normal);
3045 ins_permuted = !ins_permuted;
3049 bemit8(0x8B); // mov %in_false, %out
3050 bemit_modrr(in_false, out);
3054 pnc = ia32_get_negated_pnc(pnc);
3056 /* TODO: handling of Nans isn't correct yet */
3059 bemit8(0x40 | pnc2cc(pnc));
3060 if (get_ia32_op_type(node) == ia32_Normal) {
3061 bemit_modrr(in_true, out);
3063 bemit_mod_am(reg_gp_map[out->index], node);
3067 static void bemit_cmp(const ir_node *node)
3069 unsigned ls_size = get_mode_size_bits(get_ia32_ls_mode(node));
3075 right = get_irn_n(node, n_ia32_binary_right);
3076 if (is_ia32_Immediate(right)) {
3077 /* Use in-reg, because some instructions (cmp, test) have no out-reg. */
3078 const ir_node *op = get_irn_n(node, n_ia32_binary_right);
3079 const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(op);
3082 if (attr->symconst != NULL) {
3085 /* check for sign extension */
3086 size = get_signed_imm_size(attr->offset);
3091 bemit8(0x81 | SIGNEXT_IMM);
3092 /* cmp has this special mode */
3093 if (get_ia32_op_type(node) == ia32_AddrModeS) {
3094 bemit_mod_am(7, node);
3096 const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left);
3097 bemit_modru(reg, 7);
3099 bemit8((unsigned char)attr->offset);
3103 /* check for eax variant: this variant is shorter for 32bit immediates only */
3104 if (get_ia32_op_type(node) == ia32_AddrModeS) {
3106 bemit_mod_am(7, node);
3108 const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left);
3109 if (reg->index == REG_EAX) {
3113 bemit_modru(reg, 7);
3116 if (ls_size == 16) {
3117 bemit16(attr->offset);
3119 bemit_entity(attr->symconst, attr->sc_sign, attr->offset, false);
3123 panic("invalid imm size?!?");
3125 const arch_register_t *out = get_in_reg(node, n_ia32_binary_left);
3127 if (get_ia32_op_type(node) == ia32_Normal) {
3128 const arch_register_t *op2 = get_in_reg(node, n_ia32_binary_right);
3129 bemit_modrr(op2, out);
3131 bemit_mod_am(reg_gp_map[out->index], node);
3136 static void bemit_cmp8bit(const ir_node *node)
3138 ir_node *right = get_irn_n(node, n_ia32_binary_right);
3139 if (is_ia32_Immediate(right)) {
3140 if (get_ia32_op_type(node) == ia32_Normal) {
3141 const arch_register_t *out = get_in_reg(node, n_ia32_Cmp_left);
3142 if (out->index == REG_EAX) {
3146 bemit_modru(out, 7);
3150 bemit_mod_am(7, node);
3152 bemit8(get_ia32_immediate_attr_const(right)->offset);
3154 const arch_register_t *out = get_in_reg(node, n_ia32_Cmp_left);
3156 if (get_ia32_op_type(node) == ia32_Normal) {
3157 const arch_register_t *in = get_in_reg(node, n_ia32_Cmp_right);
3158 bemit_modrr(out, in);
3160 bemit_mod_am(reg_gp_map[out->index], node);
3165 static void bemit_test8bit(const ir_node *node)
3167 ir_node *right = get_irn_n(node, n_ia32_Test8Bit_right);
3168 if (is_ia32_Immediate(right)) {
3169 if (get_ia32_op_type(node) == ia32_Normal) {
3170 const arch_register_t *out = get_in_reg(node, n_ia32_Test8Bit_left);
3171 if (out->index == REG_EAX) {
3175 bemit_modru(out, 0);
3179 bemit_mod_am(0, node);
3181 bemit8(get_ia32_immediate_attr_const(right)->offset);
3183 const arch_register_t *out = get_in_reg(node, n_ia32_Test8Bit_left);
3185 if (get_ia32_op_type(node) == ia32_Normal) {
3186 const arch_register_t *in = get_in_reg(node, n_ia32_Test8Bit_right);
3187 bemit_modrr(out, in);
3189 bemit_mod_am(reg_gp_map[out->index], node);
3194 static void bemit_imul(const ir_node *node)
3196 ir_node *right = get_irn_n(node, n_ia32_IMul_right);
3197 /* Do we need the immediate form? */
3198 if (is_ia32_Immediate(right)) {
3199 int imm = get_ia32_immediate_attr_const(right)->offset;
3200 if (get_signed_imm_size(imm) == 1) {
3201 bemit_unop_reg(node, 0x6B, n_ia32_IMul_left);
3204 bemit_unop_reg(node, 0x69, n_ia32_IMul_left);
3209 bemit_unop_reg(node, 0xAF, n_ia32_IMul_right);
3213 static void bemit_dec(const ir_node *node)
3215 const arch_register_t *out = get_out_reg(node, pn_ia32_Dec_res);
3216 bemit8(0x48 + reg_gp_map[out->index]);
3219 static void bemit_inc(const ir_node *node)
3221 const arch_register_t *out = get_out_reg(node, pn_ia32_Inc_res);
3222 bemit8(0x40 + reg_gp_map[out->index]);
3225 #define UNOPMEM(op, code, ext) \
3226 static void bemit_##op(const ir_node *node) \
3228 bemit_unop_mem(node, code, ext); \
3231 UNOPMEM(notmem, 0xF6, 2)
3232 UNOPMEM(negmem, 0xF6, 3)
3233 UNOPMEM(incmem, 0xFE, 0)
3234 UNOPMEM(decmem, 0xFE, 1)
3236 static void bemit_ldtls(const ir_node *node)
3238 const arch_register_t *out = get_out_reg(node, 0);
3240 bemit8(0x65); // gs:
3241 if (out->index == REG_EAX) {
3242 bemit8(0xA1); // movl 0, %eax
3244 bemit8(0x8B); // movl 0, %reg
3245 bemit8(MOD_IND | ENC_REG(reg_gp_map[out->index]) | ENC_RM(0x05));
3253 static void bemit_lea(const ir_node *node)
3255 const arch_register_t *out = get_out_reg(node, 0);
3257 bemit_mod_am(reg_gp_map[out->index], node);
3260 /* helper function for bemit_minus64bit */
3261 static void bemit_helper_mov(const arch_register_t *src, const arch_register_t *dst)
3263 bemit8(0x8B); // movl %src, %dst
3264 bemit_modrr(src, dst);
3267 /* helper function for bemit_minus64bit */
3268 static void bemit_helper_neg(const arch_register_t *reg)
3270 bemit8(0xF7); // negl %reg
3271 bemit_modru(reg, 3);
3274 /* helper function for bemit_minus64bit */
3275 static void bemit_helper_sbb0(const arch_register_t *reg)
3277 bemit8(0x83); // sbbl $0, %reg
3278 bemit_modru(reg, 3);
3282 /* helper function for bemit_minus64bit */
3283 static void bemit_helper_sbb(const arch_register_t *src, const arch_register_t *dst)
3285 bemit8(0x1B); // sbbl %src, %dst
3286 bemit_modrr(src, dst);
3289 /* helper function for bemit_minus64bit */
3290 static void bemit_helper_xchg(const arch_register_t *src, const arch_register_t *dst)
3292 if (src->index == REG_EAX) {
3293 bemit8(0x90 + reg_gp_map[dst->index]); // xchgl %eax, %dst
3294 } else if (dst->index == REG_EAX) {
3295 bemit8(0x90 + reg_gp_map[src->index]); // xchgl %src, %eax
3297 bemit8(0x87); // xchgl %src, %dst
3298 bemit_modrr(src, dst);
3302 /* helper function for bemit_minus64bit */
3303 static void bemit_helper_zero(const arch_register_t *reg)
3305 bemit8(0x33); // xorl %reg, %reg
3306 bemit_modrr(reg, reg);
3309 static void bemit_minus64bit(const ir_node *node)
3311 const arch_register_t *in_lo = get_in_reg(node, 0);
3312 const arch_register_t *in_hi = get_in_reg(node, 1);
3313 const arch_register_t *out_lo = get_out_reg(node, 0);
3314 const arch_register_t *out_hi = get_out_reg(node, 1);
3316 if (out_lo == in_lo) {
3317 if (out_hi != in_hi) {
3318 /* a -> a, b -> d */
3321 /* a -> a, b -> b */
3324 } else if (out_lo == in_hi) {
3325 if (out_hi == in_lo) {
3326 /* a -> b, b -> a */
3327 bemit_helper_xchg(in_lo, in_hi);
3330 /* a -> b, b -> d */
3331 bemit_helper_mov(in_hi, out_hi);
3332 bemit_helper_mov(in_lo, out_lo);
3336 if (out_hi == in_lo) {
3337 /* a -> c, b -> a */
3338 bemit_helper_mov(in_lo, out_lo);
3340 } else if (out_hi == in_hi) {
3341 /* a -> c, b -> b */
3342 bemit_helper_mov(in_lo, out_lo);
3345 /* a -> c, b -> d */
3346 bemit_helper_mov(in_lo, out_lo);
3352 bemit_helper_neg( out_hi);
3353 bemit_helper_neg( out_lo);
3354 bemit_helper_sbb0(out_hi);
3358 bemit_helper_zero(out_hi);
3359 bemit_helper_neg( out_lo);
3360 bemit_helper_sbb( in_hi, out_hi);
3364 * Emit a single opcode.
3366 #define EMIT_SINGLEOP(op, code) \
3367 static void bemit_ ## op(const ir_node *node) { \
3372 //EMIT_SINGLEOP(daa, 0x27)
3373 //EMIT_SINGLEOP(das, 0x2F)
3374 //EMIT_SINGLEOP(aaa, 0x37)
3375 //EMIT_SINGLEOP(aas, 0x3F)
3376 //EMIT_SINGLEOP(nop, 0x90)
3377 EMIT_SINGLEOP(cwtl, 0x98)
3378 EMIT_SINGLEOP(cltd, 0x99)
3379 //EMIT_SINGLEOP(fwait, 0x9B)
3380 EMIT_SINGLEOP(sahf, 0x9E)
3381 //EMIT_SINGLEOP(popf, 0x9D)
3382 EMIT_SINGLEOP(leave, 0xC9)
3383 EMIT_SINGLEOP(int3, 0xCC)
3384 //EMIT_SINGLEOP(iret, 0xCF)
3385 //EMIT_SINGLEOP(xlat, 0xD7)
3386 //EMIT_SINGLEOP(lock, 0xF0)
3387 EMIT_SINGLEOP(rep, 0xF3)
3388 //EMIT_SINGLEOP(halt, 0xF4)
3389 EMIT_SINGLEOP(cmc, 0xF5)
3390 EMIT_SINGLEOP(stc, 0xF9)
3391 //EMIT_SINGLEOP(cli, 0xFA)
3392 //EMIT_SINGLEOP(sti, 0xFB)
3393 //EMIT_SINGLEOP(std, 0xFD)
3396 * Emits a MOV out, [MEM].
3398 static void bemit_load(const ir_node *node)
3400 const arch_register_t *out = get_out_reg(node, 0);
3402 if (out->index == REG_EAX) {
3403 ir_node *base = get_irn_n(node, n_ia32_base);
3404 int has_base = !is_ia32_NoReg_GP(base);
3405 ir_node *index = get_irn_n(node, n_ia32_index);
3406 int has_index = !is_ia32_NoReg_GP(index);
3407 if (!has_base && !has_index) {
3408 ir_entity *ent = get_ia32_am_sc(node);
3409 int offs = get_ia32_am_offs_int(node);
3410 /* load from constant address to EAX can be encoded
3413 bemit_entity(ent, 0, offs, false);
3418 bemit_mod_am(reg_gp_map[out->index], node);
3422 * Emits a MOV [mem], in.
3424 static void bemit_store(const ir_node *node)
3426 const ir_node *value = get_irn_n(node, n_ia32_Store_val);
3427 unsigned size = get_mode_size_bits(get_ia32_ls_mode(node));
3429 if (is_ia32_Immediate(value)) {
3432 bemit_mod_am(0, node);
3433 bemit8(get_ia32_immediate_attr_const(value)->offset);
3434 } else if (size == 16) {
3437 bemit_mod_am(0, node);
3438 bemit16(get_ia32_immediate_attr_const(value)->offset);
3441 bemit_mod_am(0, node);
3442 bemit_immediate(value, false);
3445 const arch_register_t *in = get_in_reg(node, n_ia32_Store_val);
3447 if (in->index == REG_EAX) {
3448 ir_node *base = get_irn_n(node, n_ia32_base);
3449 int has_base = !is_ia32_NoReg_GP(base);
3450 ir_node *index = get_irn_n(node, n_ia32_index);
3451 int has_index = !is_ia32_NoReg_GP(index);
3452 if (!has_base && !has_index) {
3453 ir_entity *ent = get_ia32_am_sc(node);
3454 int offs = get_ia32_am_offs_int(node);
3455 /* store to constant address from EAX can be encoded as
3456 * 0xA2/0xA3 [offset]*/
3464 bemit_entity(ent, 0, offs, false);
3476 bemit_mod_am(reg_gp_map[in->index], node);
3480 static void bemit_conv_i2i(const ir_node *node)
3482 ir_mode *smaller_mode = get_ia32_ls_mode(node);
3491 if (mode_is_signed(smaller_mode)) opcode |= 0x08;
3492 if (get_mode_size_bits(smaller_mode) == 16) opcode |= 0x01;
3493 bemit_unop_reg(node, opcode, n_ia32_Conv_I2I_val);
3499 static void bemit_push(const ir_node *node)
3501 const ir_node *value = get_irn_n(node, n_ia32_Push_val);
3503 if (is_ia32_Immediate(value)) {
3504 const ia32_immediate_attr_t *attr
3505 = get_ia32_immediate_attr_const(value);
3506 unsigned size = get_signed_imm_size(attr->offset);
3512 bemit8((unsigned char)attr->offset);
3517 bemit_immediate(value, false);
3520 } else if (is_ia32_NoReg_GP(value)) {
3522 bemit_mod_am(6, node);
3524 const arch_register_t *reg = get_in_reg(node, n_ia32_Push_val);
3525 bemit8(0x50 + reg_gp_map[reg->index]);
3532 static void bemit_pop(const ir_node *node)
3534 const arch_register_t *reg = get_out_reg(node, pn_ia32_Pop_res);
3535 bemit8(0x58 + reg_gp_map[reg->index]);
3538 static void bemit_popmem(const ir_node *node)
3541 bemit_mod_am(0, node);
3544 static void bemit_call(const ir_node *node)
3546 ir_node *proc = get_irn_n(node, n_ia32_Call_addr);
3548 if (is_ia32_Immediate(proc)) {
3550 bemit_immediate(proc, true);
3552 bemit_unop(node, 0xFF, 2, n_ia32_Call_addr);
3556 static void bemit_jmp(const ir_node *dest_block)
3559 bemit_jmp_destination(dest_block);
3562 static void bemit_jump(const ir_node *node)
3564 if (can_be_fallthrough(node))
3567 bemit_jmp(get_cfop_target_block(node));
3570 static void bemit_jcc(int pnc, const ir_node *dest_block)
3572 unsigned char cc = pnc2cc(pnc);
3575 bemit_jmp_destination(dest_block);
3578 static void bemit_jp(bool odd, const ir_node *dest_block)
3582 bemit_jmp_destination(dest_block);
3585 static void bemit_ia32_jcc(const ir_node *node)
3587 int pnc = get_ia32_condcode(node);
3588 const ir_node *proj_true;
3589 const ir_node *proj_false;
3590 const ir_node *dest_true;
3591 const ir_node *dest_false;
3592 const ir_node *block;
3594 pnc = determine_final_pnc(node, 0, pnc);
3596 /* get both Projs */
3597 proj_true = get_proj(node, pn_ia32_Jcc_true);
3598 assert(proj_true && "Jcc without true Proj");
3600 proj_false = get_proj(node, pn_ia32_Jcc_false);
3601 assert(proj_false && "Jcc without false Proj");
3603 block = get_nodes_block(node);
3605 if (can_be_fallthrough(proj_true)) {
3606 /* exchange both proj's so the second one can be omitted */
3607 const ir_node *t = proj_true;
3609 proj_true = proj_false;
3611 pnc = ia32_get_negated_pnc(pnc);
3614 dest_true = get_cfop_target_block(proj_true);
3615 dest_false = get_cfop_target_block(proj_false);
3617 if (pnc & ia32_pn_Cmp_float) {
3618 /* Some floating point comparisons require a test of the parity flag,
3619 * which indicates that the result is unordered */
3622 bemit_jp(false, dest_true);
3627 bemit_jp(true, dest_true);
3633 /* we need a local label if the false proj is a fallthrough
3634 * as the falseblock might have no label emitted then */
3635 if (can_be_fallthrough(proj_false)) {
3637 bemit8(0x06); // jp + 6
3639 bemit_jp(false, dest_false);
3646 bemit_jp(false, dest_true);
3654 bemit_jcc(pnc, dest_true);
3657 /* the second Proj might be a fallthrough */
3658 if (can_be_fallthrough(proj_false)) {
3659 /* it's a fallthrough */
3661 bemit_jmp(dest_false);
3665 static void bemit_switchjmp(const ir_node *node)
3667 unsigned long interval;
3671 const arch_register_t *in;
3673 /* fill the table structure */
3674 generate_jump_table(&tbl, node);
3676 /* two-complement's magic make this work without overflow */
3677 interval = tbl.max_value - tbl.min_value;
3679 in = get_in_reg(node, 0);
3680 /* emit the table */
3681 if (get_signed_imm_size(interval) == 1) {
3682 bemit8(0x83); // cmpl $imm8, %in
3686 bemit8(0x81); // cmpl $imm32, %in
3690 bemit8(0x0F); // ja tbl.defProj
3692 ia32_emitf(tbl.defProj, ".long %L - . - 4\n");
3694 if (tbl.num_branches > 1) {
3696 bemit8(0xFF); // jmp *tbl.label(,%in,4)
3697 bemit8(MOD_IND | ENC_REG(4) | ENC_RM(0x04));
3698 bemit8(ENC_SIB(2, reg_gp_map[in->index], 0x05));
3699 be_emit_irprintf("\t.long %s\n", tbl.label);
3701 be_gas_emit_switch_section(GAS_SECTION_RODATA);
3702 be_emit_cstring(".align 4\n");
3703 be_emit_irprintf("%s:\n", tbl.label);
3705 last_value = tbl.branches[0].value;
3706 for (i = 0; i != tbl.num_branches; ++i) {
3707 while (last_value != tbl.branches[i].value) {
3708 ia32_emitf(tbl.defProj, ".long %L\n");
3711 ia32_emitf(tbl.branches[i].target, ".long %L\n");
3714 be_gas_emit_switch_section(GAS_SECTION_TEXT);
3716 /* one jump is enough */
3717 panic("switch only has one case");
3718 //ia32_emitf(tbl.branches[0].target, "\tjmp %L\n");
3721 be_emit_write_line();
3729 static void bemit_return(const ir_node *node)
3731 unsigned pop = be_Return_get_pop(node);
3732 if (pop > 0 || be_Return_get_emit_pop(node)) {
3734 assert(pop <= 0xffff);
3741 static void bemit_subsp(const ir_node *node)
3743 const arch_register_t *out;
3746 /* mov %esp, %out */
3748 out = get_out_reg(node, 1);
3749 bemit8(MOD_REG | ENC_REG(reg_gp_map[out->index]) | ENC_RM(0x04));
3752 static void bemit_incsp(const ir_node *node)
3755 const arch_register_t *reg;
3759 offs = be_get_IncSP_offset(node);
3770 size = get_signed_imm_size(offs);
3771 bemit8(size == 1 ? 0x83 : 0x81);
3773 reg = get_out_reg(node, 0);
3774 bemit_modru(reg, ext);
3783 static void bemit_copybi(const ir_node *node)
3785 unsigned size = get_ia32_copyb_size(node);
3787 bemit8(0xA4); // movsb
3790 bemit8(0xA5); // movsw
3794 bemit8(0xA5); // movsl
3798 static void bemit_fbinop(const ir_node *node, unsigned code, unsigned code_to)
3800 if (get_ia32_op_type(node) == ia32_Normal) {
3801 const ia32_x87_attr_t *x87_attr = get_ia32_x87_attr_const(node);
3802 const arch_register_t *in1 = x87_attr->x87[0];
3803 const arch_register_t *in = x87_attr->x87[1];
3804 const arch_register_t *out = x87_attr->x87[2];
3808 } else if (out == in) {
3812 if (out->index == 0) {
3814 bemit8(MOD_REG | ENC_REG(code) | ENC_RM(in->index));
3817 bemit8(MOD_REG | ENC_REG(code_to) | ENC_RM(out->index));
3820 if (get_mode_size_bits(get_ia32_ls_mode(node)) == 32) {
3825 bemit_mod_am(code, node);
3829 static void bemit_fbinopp(const ir_node *node, unsigned const code)
3831 const ia32_x87_attr_t *x87_attr = get_ia32_x87_attr_const(node);
3832 const arch_register_t *out = x87_attr->x87[2];
3834 bemit8(code + out->index);
3837 static void bemit_fabs(const ir_node *node)
3845 static void bemit_fadd(const ir_node *node)
3847 bemit_fbinop(node, 0, 0);
3850 static void bemit_faddp(const ir_node *node)
3852 bemit_fbinopp(node, 0xC0);
3855 static void bemit_fchs(const ir_node *node)
3863 static void bemit_fdiv(const ir_node *node)
3865 bemit_fbinop(node, 6, 7);
3868 static void bemit_fdivp(const ir_node *node)
3870 bemit_fbinopp(node, 0xF8);
3873 static void bemit_fdivr(const ir_node *node)
3875 bemit_fbinop(node, 7, 6);
3878 static void bemit_fdivrp(const ir_node *node)
3880 bemit_fbinopp(node, 0xF0);
3883 static void bemit_fild(const ir_node *node)
3885 switch (get_mode_size_bits(get_ia32_ls_mode(node))) {
3887 bemit8(0xDF); // filds
3888 bemit_mod_am(0, node);
3892 bemit8(0xDB); // fildl
3893 bemit_mod_am(0, node);
3897 bemit8(0xDF); // fildll
3898 bemit_mod_am(5, node);
3902 panic("invalid mode size");
3906 static void bemit_fist(const ir_node *node)
3908 switch (get_mode_size_bits(get_ia32_ls_mode(node))) {
3910 bemit8(0xDF); // fists
3914 bemit8(0xDB); // fistl
3918 panic("invalid mode size");
3920 bemit_mod_am(2, node);
3923 static void bemit_fistp(const ir_node *node)
3925 switch (get_mode_size_bits(get_ia32_ls_mode(node))) {
3927 bemit8(0xDF); // fistps
3928 bemit_mod_am(3, node);
3932 bemit8(0xDB); // fistpl
3933 bemit_mod_am(3, node);
3937 bemit8(0xDF); // fistpll
3938 bemit_mod_am(7, node);
3942 panic("invalid mode size");
3946 static void bemit_fld(const ir_node *node)
3948 switch (get_mode_size_bits(get_ia32_ls_mode(node))) {
3950 bemit8(0xD9); // flds
3951 bemit_mod_am(0, node);
3955 bemit8(0xDD); // fldl
3956 bemit_mod_am(0, node);
3961 bemit8(0xDB); // fldt
3962 bemit_mod_am(5, node);
3966 panic("invalid mode size");
3970 static void bemit_fld1(const ir_node *node)
3974 bemit8(0xE8); // fld1
3977 static void bemit_fldcw(const ir_node *node)
3979 bemit8(0xD9); // fldcw
3980 bemit_mod_am(5, node);
3983 static void bemit_fldz(const ir_node *node)
3987 bemit8(0xEE); // fldz
3990 static void bemit_fmul(const ir_node *node)
3992 bemit_fbinop(node, 1, 1);
3995 static void bemit_fmulp(const ir_node *node)
3997 bemit_fbinopp(node, 0xC8);
4000 static void bemit_fpop(const ir_node *node)
4002 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4004 bemit8(0xD8 + attr->x87[0]->index);
4007 static void bemit_fpush(const ir_node *node)
4009 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4011 bemit8(0xC0 + attr->x87[0]->index);
4014 static void bemit_fpushcopy(const ir_node *node)
4016 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4018 bemit8(0xC0 + attr->x87[0]->index);
4021 static void bemit_fst(const ir_node *node)
4023 switch (get_mode_size_bits(get_ia32_ls_mode(node))) {
4025 bemit8(0xD9); // fsts
4029 bemit8(0xDD); // fstl
4033 panic("invalid mode size");
4035 bemit_mod_am(2, node);
4038 static void bemit_fstp(const ir_node *node)
4040 switch (get_mode_size_bits(get_ia32_ls_mode(node))) {
4042 bemit8(0xD9); // fstps
4043 bemit_mod_am(3, node);
4047 bemit8(0xDD); // fstpl
4048 bemit_mod_am(3, node);
4053 bemit8(0xDB); // fstpt
4054 bemit_mod_am(7, node);
4058 panic("invalid mode size");
4062 static void bemit_fsub(const ir_node *node)
4064 bemit_fbinop(node, 4, 5);
4067 static void bemit_fsubp(const ir_node *node)
4069 bemit_fbinopp(node, 0xE8);
4072 static void bemit_fsubr(const ir_node *node)
4074 bemit_fbinop(node, 5, 4);
4077 static void bemit_fsubrp(const ir_node *node)
4079 bemit_fbinopp(node, 0xE0);
4082 static void bemit_fnstcw(const ir_node *node)
4084 bemit8(0xD9); // fnstcw
4085 bemit_mod_am(7, node);
4088 static void bemit_fnstsw(void)
4090 bemit8(0xDF); // fnstsw %ax
4094 static void bemit_ftstfnstsw(const ir_node *node)
4098 bemit8(0xD9); // ftst
4103 static void bemit_fucomi(const ir_node *node)
4105 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4106 bemit8(0xDB); // fucomi
4107 bemit8(0xE8 + attr->x87[1]->index);
4110 static void bemit_fucomip(const ir_node *node)
4112 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4113 bemit8(0xDF); // fucomip
4114 bemit8(0xE8 + attr->x87[1]->index);
4117 static void bemit_fucomfnstsw(const ir_node *node)
4119 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4120 bemit8(0xDD); // fucom
4121 bemit8(0xE0 + attr->x87[1]->index);
4125 static void bemit_fucompfnstsw(const ir_node *node)
4127 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4128 bemit8(0xDD); // fucomp
4129 bemit8(0xE8 + attr->x87[1]->index);
4133 static void bemit_fucomppfnstsw(const ir_node *node)
4137 bemit8(0xDA); // fucompp
4142 static void bemit_fxch(const ir_node *node)
4144 const ia32_x87_attr_t *attr = get_ia32_x87_attr_const(node);
4146 bemit8(0xC8 + attr->x87[0]->index);
4150 * The type of a emitter function.
4152 typedef void (*emit_func) (const ir_node *);
4155 * Set a node emitter. Make it a bit more type safe.
4157 static void register_emitter(ir_op *op, emit_func func)
4159 op->ops.generic = (op_func) func;
4162 static void ia32_register_binary_emitters(void)
4164 /* first clear the generic function pointer for all ops */
4165 clear_irp_opcodes_generic_func();
4167 /* benode emitter */
4168 register_emitter(op_be_Copy, bemit_copy);
4169 register_emitter(op_be_CopyKeep, bemit_copy);
4170 register_emitter(op_be_IncSP, bemit_incsp);
4171 register_emitter(op_be_Perm, bemit_perm);
4172 register_emitter(op_be_Return, bemit_return);
4173 register_emitter(op_ia32_Adc, bemit_adc);
4174 register_emitter(op_ia32_Add, bemit_add);
4175 register_emitter(op_ia32_AddMem, bemit_addmem);
4176 register_emitter(op_ia32_AddMem8Bit, bemit_addmem8bit);
4177 register_emitter(op_ia32_And, bemit_and);
4178 register_emitter(op_ia32_AndMem, bemit_andmem);
4179 register_emitter(op_ia32_AndMem8Bit, bemit_andmem8bit);
4180 register_emitter(op_ia32_Breakpoint, bemit_int3);
4181 register_emitter(op_ia32_CMovcc, bemit_cmovcc);
4182 register_emitter(op_ia32_Call, bemit_call);
4183 register_emitter(op_ia32_Cltd, bemit_cltd);
4184 register_emitter(op_ia32_Cmc, bemit_cmc);
4185 register_emitter(op_ia32_Cmp, bemit_cmp);
4186 register_emitter(op_ia32_Cmp8Bit, bemit_cmp8bit);
4187 register_emitter(op_ia32_Const, bemit_mov_const);
4188 register_emitter(op_ia32_Conv_I2I, bemit_conv_i2i);
4189 register_emitter(op_ia32_Conv_I2I8Bit, bemit_conv_i2i);
4190 register_emitter(op_ia32_CopyB_i, bemit_copybi);
4191 register_emitter(op_ia32_Cwtl, bemit_cwtl);
4192 register_emitter(op_ia32_Dec, bemit_dec);
4193 register_emitter(op_ia32_DecMem, bemit_decmem);
4194 register_emitter(op_ia32_Div, bemit_div);
4195 register_emitter(op_ia32_FldCW, bemit_fldcw);
4196 register_emitter(op_ia32_FnstCW, bemit_fnstcw);
4197 register_emitter(op_ia32_FtstFnstsw, bemit_ftstfnstsw);
4198 register_emitter(op_ia32_FucomFnstsw, bemit_fucomfnstsw);
4199 register_emitter(op_ia32_Fucomi, bemit_fucomi);
4200 register_emitter(op_ia32_FucompFnstsw, bemit_fucompfnstsw);
4201 register_emitter(op_ia32_Fucompi, bemit_fucomip);
4202 register_emitter(op_ia32_FucomppFnstsw, bemit_fucomppfnstsw);
4203 register_emitter(op_ia32_IDiv, bemit_idiv);
4204 register_emitter(op_ia32_IJmp, bemit_ijmp);
4205 register_emitter(op_ia32_IMul, bemit_imul);
4206 register_emitter(op_ia32_IMul1OP, bemit_imul1op);
4207 register_emitter(op_ia32_Inc, bemit_inc);
4208 register_emitter(op_ia32_IncMem, bemit_incmem);
4209 register_emitter(op_ia32_Jcc, bemit_ia32_jcc);
4210 register_emitter(op_ia32_Jmp, bemit_jump);
4211 register_emitter(op_ia32_LdTls, bemit_ldtls);
4212 register_emitter(op_ia32_Lea, bemit_lea);
4213 register_emitter(op_ia32_Leave, bemit_leave);
4214 register_emitter(op_ia32_Load, bemit_load);
4215 register_emitter(op_ia32_Minus64Bit, bemit_minus64bit);
4216 register_emitter(op_ia32_Mul, bemit_mul);
4217 register_emitter(op_ia32_Neg, bemit_neg);
4218 register_emitter(op_ia32_NegMem, bemit_negmem);
4219 register_emitter(op_ia32_Not, bemit_not);
4220 register_emitter(op_ia32_NotMem, bemit_notmem);
4221 register_emitter(op_ia32_Or, bemit_or);
4222 register_emitter(op_ia32_OrMem, bemit_ormem);
4223 register_emitter(op_ia32_OrMem8Bit, bemit_ormem8bit);
4224 register_emitter(op_ia32_Pop, bemit_pop);
4225 register_emitter(op_ia32_PopEbp, bemit_pop);
4226 register_emitter(op_ia32_PopMem, bemit_popmem);
4227 register_emitter(op_ia32_Push, bemit_push);
4228 register_emitter(op_ia32_RepPrefix, bemit_rep);
4229 register_emitter(op_ia32_Rol, bemit_rol);
4230 register_emitter(op_ia32_RolMem, bemit_rolmem);
4231 register_emitter(op_ia32_Ror, bemit_ror);
4232 register_emitter(op_ia32_RorMem, bemit_rormem);
4233 register_emitter(op_ia32_Sahf, bemit_sahf);
4234 register_emitter(op_ia32_Sar, bemit_sar);
4235 register_emitter(op_ia32_SarMem, bemit_sarmem);
4236 register_emitter(op_ia32_Sbb, bemit_sbb);
4237 register_emitter(op_ia32_Setcc, bemit_setcc);
4238 register_emitter(op_ia32_Shl, bemit_shl);
4239 register_emitter(op_ia32_ShlD, bemit_shld);
4240 register_emitter(op_ia32_ShlMem, bemit_shlmem);
4241 register_emitter(op_ia32_Shr, bemit_shr);
4242 register_emitter(op_ia32_ShrD, bemit_shrd);
4243 register_emitter(op_ia32_ShrMem, bemit_shrmem);
4244 register_emitter(op_ia32_Stc, bemit_stc);
4245 register_emitter(op_ia32_Store, bemit_store);
4246 register_emitter(op_ia32_Store8Bit, bemit_store);
4247 register_emitter(op_ia32_Sub, bemit_sub);
4248 register_emitter(op_ia32_SubMem, bemit_submem);
4249 register_emitter(op_ia32_SubMem8Bit, bemit_submem8bit);
4250 register_emitter(op_ia32_SubSP, bemit_subsp);
4251 register_emitter(op_ia32_SwitchJmp, bemit_switchjmp);
4252 register_emitter(op_ia32_Test, bemit_test);
4253 register_emitter(op_ia32_Test8Bit, bemit_test8bit);
4254 register_emitter(op_ia32_Xor, bemit_xor);
4255 register_emitter(op_ia32_Xor0, bemit_xor0);
4256 register_emitter(op_ia32_XorMem, bemit_xormem);
4257 register_emitter(op_ia32_XorMem8Bit, bemit_xormem8bit);
4258 register_emitter(op_ia32_fabs, bemit_fabs);
4259 register_emitter(op_ia32_fadd, bemit_fadd);
4260 register_emitter(op_ia32_faddp, bemit_faddp);
4261 register_emitter(op_ia32_fchs, bemit_fchs);
4262 register_emitter(op_ia32_fdiv, bemit_fdiv);
4263 register_emitter(op_ia32_fdivp, bemit_fdivp);
4264 register_emitter(op_ia32_fdivr, bemit_fdivr);
4265 register_emitter(op_ia32_fdivrp, bemit_fdivrp);
4266 register_emitter(op_ia32_fild, bemit_fild);
4267 register_emitter(op_ia32_fist, bemit_fist);
4268 register_emitter(op_ia32_fistp, bemit_fistp);
4269 register_emitter(op_ia32_fld, bemit_fld);
4270 register_emitter(op_ia32_fld1, bemit_fld1);
4271 register_emitter(op_ia32_fldz, bemit_fldz);
4272 register_emitter(op_ia32_fmul, bemit_fmul);
4273 register_emitter(op_ia32_fmulp, bemit_fmulp);
4274 register_emitter(op_ia32_fpop, bemit_fpop);
4275 register_emitter(op_ia32_fpush, bemit_fpush);
4276 register_emitter(op_ia32_fpushCopy, bemit_fpushcopy);
4277 register_emitter(op_ia32_fst, bemit_fst);
4278 register_emitter(op_ia32_fstp, bemit_fstp);
4279 register_emitter(op_ia32_fsub, bemit_fsub);
4280 register_emitter(op_ia32_fsubp, bemit_fsubp);
4281 register_emitter(op_ia32_fsubr, bemit_fsubr);
4282 register_emitter(op_ia32_fsubrp, bemit_fsubrp);
4283 register_emitter(op_ia32_fxch, bemit_fxch);
4285 /* ignore the following nodes */
4286 register_emitter(op_ia32_ProduceVal, emit_Nothing);
4287 register_emitter(op_be_Barrier, emit_Nothing);
4288 register_emitter(op_be_Keep, emit_Nothing);
4289 register_emitter(op_be_Start, emit_Nothing);
4290 register_emitter(op_Phi, emit_Nothing);
4291 register_emitter(op_Start, emit_Nothing);
4294 static void gen_binary_block(ir_node *block)
4298 ia32_emit_block_header(block);
4300 /* emit the contents of the block */
4301 sched_foreach(block, node) {
4302 ia32_emit_node(node);
4306 void ia32_gen_binary_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg)
4308 ir_entity *entity = get_irg_entity(irg);
4314 ia32_register_binary_emitters();
4316 be_gas_emit_function_prolog(entity, ia32_cg_config.function_alignment);
4318 /* we use links to point to target blocks */
4319 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
4320 irg_block_walk_graph(irg, ia32_gen_labels, NULL, NULL);
4322 /* initialize next block links */
4323 n = ARR_LEN(cg->blk_sched);
4324 for (i = 0; i < n; ++i) {
4325 ir_node *block = cg->blk_sched[i];
4326 ir_node *prev = i > 0 ? cg->blk_sched[i-1] : NULL;
4328 set_irn_link(block, prev);
4331 for (i = 0; i < n; ++i) {
4332 ir_node *block = cg->blk_sched[i];
4333 gen_binary_block(block);
4336 be_gas_emit_function_epilog(entity);
4337 be_dbg_method_end();
4339 be_emit_write_line();
4341 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
4347 void ia32_init_emitter(void)
4349 lc_opt_entry_t *be_grp;
4350 lc_opt_entry_t *ia32_grp;
4352 be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4353 ia32_grp = lc_opt_get_grp(be_grp, "ia32");
4355 lc_opt_add_table(ia32_grp, ia32_emitter_options);
4359 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.emitter");