?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
x86: fix a bug that we misgenerated the the 8-bit imul with a constant
[ajla.git] / codegen.c
blob1b942fd0c9797772dd738154a8ec11f768c19438
1 /*
2 * Copyright (C) 2024 Mikulas Patocka
3 *
4 * This file is part of Ajla.
5 *
6 * Ajla is free software: you can redistribute it and/or modify it under the
7 * terms of the GNU General Public License as published by the Free Software
8 * Foundation, either version 3 of the License, or (at your option) any later
9 * version.
10 *
11 * Ajla is distributed in the hope that it will be useful, but WITHOUT ANY
12 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
13 * A PARTICULAR PURPOSE. See the GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License along with
16 * Ajla. If not, see <https://www.gnu.org/licenses/>.
17 */
18
19 #include "ajla.h"
20
21 #ifndef FILE_OMIT
22
23 #include "data.h"
24 #include "os.h"
25 #include "os_util.h"
26 #include "util.h"
27 #include "ipfn.h"
28 #include "ipret.h"
29 #include "funct.h"
30 #include "thread.h"
31 #include "task.h"
32 #include "save.h"
33
34 #include "codegen.h"
35
36 #ifdef HAVE_CODEGEN
37
38 #define flag_cache_chicken 0
39 #define must_be_flat_chicken 0
40 #define ra_chicken 0
41
42 #define INLINE_BITMAP_SLOTS 16
43 #define INLINE_COPY_SIZE 64
44
45 /*#define DEBUG_INSNS*/
46 /*#define DEBUG_GARBAGE*/
47
48 #if (defined(ARCH_X86_64) || defined(ARCH_X86_X32)) && !defined(ARCH_X86_WIN_ABI)
49 #if defined(HAVE_SYSCALL) && defined(HAVE_ASM_PRCTL_H) && defined(HAVE_SYS_SYSCALL_H)
50 #include <asm/prctl.h>
51 #include <sys/syscall.h>
52 #include <unistd.h>
53 #endif
54 #if (defined(HAVE_AMD64_SET_GSBASE) || defined(HAVE_SYSARCH)) && defined(HAVE_X86_SYSARCH_H)
55 #include <x86/sysarch.h>
56 #endif
57 #endif
58
59 code_return_t (*codegen_entry)(frame_s *, struct cg_upcall_vector_s *, tick_stamp_t, void *);
60 static void *codegen_ptr;
61 static size_t codegen_size;
62
63 #ifdef DEBUG_ENV
64 static mutex_t dump_mutex;
65 static uint64_t dump_seq = 0;
66 #endif
67
68 /*
69 * insn:
70 * opcode - 16 bits
71 * op_size - 3 bits
72 * aux - 7 bits
73 * writes flags - 2 bit
74 * jmp size - 2 bits
75 */
76
77 #define INSN_OPCODE 0x0000ffffUL
78 #define INSN_OP_SIZE 0x00070000UL
79 #define INSN_AUX 0x03f80000UL
80 #define INSN_WRITES_FLAGS 0x0c000000UL
81 #define INSN_JUMP_SIZE 0x30000000UL
82
83 #define INSN_OPCODE_SHIFT 0
84 #define INSN_OP_SIZE_SHIFT 16
85 #define INSN_AUX_SHIFT 19
86 #define INSN_WRITES_FLAGS_SHIFT 26
87 #define INSN_JUMP_SIZE_SHIFT 28
88
89 #define insn_opcode(insn) (((insn) >> INSN_OPCODE_SHIFT) & (INSN_OPCODE >> INSN_OPCODE_SHIFT))
90 #define insn_op_size(insn) (((insn) >> INSN_OP_SIZE_SHIFT) & (INSN_OP_SIZE >> INSN_OP_SIZE_SHIFT))
91 #define insn_aux(insn) (((insn) >> INSN_AUX_SHIFT) & (INSN_AUX >> INSN_AUX_SHIFT))
92 #define insn_writes_flags(insn) (((insn) >> INSN_WRITES_FLAGS_SHIFT) & (INSN_WRITES_FLAGS >> INSN_WRITES_FLAGS_SHIFT))
93 #define insn_jump_size(insn) (((insn) >> INSN_JUMP_SIZE_SHIFT) & (INSN_JUMP_SIZE >> INSN_JUMP_SIZE_SHIFT))
94
95 #define ALU_ADD 0x00
96 #define ALU_OR 0x01
97 #define ALU_ADC 0x02
98 #define ALU_SBB 0x03
99 #define ALU_AND 0x04
100 #define ALU_SUB 0x05
101 #define ALU_XOR 0x06
102 #define ALU_ORN 0x08
103 #define ALU_ANDN 0x09
104 #define ALU_XORN 0x0a
105 #define ALU_MUL 0x10
106 #define ALU_UMULH 0x11
107 #define ALU_SMULH 0x12
108 #define ALU_UDIV 0x13
109 #define ALU_SDIV 0x14
110 #define ALU_UREM 0x15
111 #define ALU_SREM 0x16
112 #define ALU_SAVE 0x17
113 #define ALU_EXTBL 0x18
114 #define ALU_EXTWL 0x19
115 #define ALU_EXTLL 0x1a
116 #define ALU_EXTLH 0x1b
117 #define ALU_INSBL 0x1c
118 #define ALU_MSKBL 0x1d
119 #define ALU_ZAP 0x20
120 #define ALU_ZAPNOT 0x21
121
122 #define ALU1_NOT 0x00
123 #define ALU1_NEG 0x01
124 #define ALU1_NGC 0x02
125 #define ALU1_INC 0x03
126 #define ALU1_DEC 0x04
127 #define ALU1_BSWAP 0x05
128 #define ALU1_BSWAP16 0x06
129 #define ALU1_BREV 0x07
130 #define ALU1_BSF 0x08
131 #define ALU1_BSR 0x09
132 #define ALU1_LZCNT 0x0a
133 #define ALU1_POPCNT 0x0b
134
135 #define FP_ALU_ADD 0
136 #define FP_ALU_SUB 1
137 #define FP_ALU_MUL 2
138 #define FP_ALU_DIV 3
139 #define FP_ALU1_NEG 0
140 #define FP_ALU1_SQRT 1
141 #define FP_ALU1_ROUND 2
142 #define FP_ALU1_FLOOR 3
143 #define FP_ALU1_CEIL 4
144 #define FP_ALU1_TRUNC 5
145 #define FP_ALU1_VCNT8 6
146 #define FP_ALU1_VPADDL 7
147 #define FP_ALU1_ADDV 8
148
149 #define COND_O 0x0
150 #define COND_NO 0x1
151 #define COND_B 0x2
152 #define COND_AE 0x3
153 #define COND_E 0x4
154 #define COND_NE 0x5
155 #define COND_BE 0x6
156 #define COND_A 0x7
157 #define COND_S 0x8
158 #define COND_NS 0x9
159 #define COND_P 0xa
160 #define COND_NP 0xb
161 #define COND_L 0xc
162 #define COND_GE 0xd
163 #define COND_LE 0xe
164 #define COND_G 0xf
165 #define COND_BLBC 0x10
166 #define COND_BLBS 0x11
167 #define COND_ALWAYS 0x12
168
169 #define COND_FP 0x20
170 #define FP_COND_P (COND_FP | COND_P)
171 #define FP_COND_NP (COND_FP | COND_NP)
172 #define FP_COND_E (COND_FP | COND_E)
173 #define FP_COND_NE (COND_FP | COND_NE)
174 #define FP_COND_A (COND_FP | COND_A)
175 #define FP_COND_BE (COND_FP | COND_BE)
176 #define FP_COND_B (COND_FP | COND_B)
177 #define FP_COND_AE (COND_FP | COND_AE)
178
179 #define ROT_ROL 0x0
180 #define ROT_ROR 0x1
181 #define ROT_RCL 0x2
182 #define ROT_RCR 0x3
183 #define ROT_SHL 0x4
184 #define ROT_SHR 0x5
185 #define ROT_SAR 0x7
186 #define ROT_SAL 0x8
187
188 #define BTX_BT 0x0
189 #define BTX_BTS 0x1
190 #define BTX_BTR 0x2
191 #define BTX_BTC 0x3
192 #define BTX_BTEXT 0x4
193
194 #define OP_SIZE_1 0
195 #define OP_SIZE_2 1
196 #define OP_SIZE_4 2
197 #define OP_SIZE_8 3
198 #define OP_SIZE_16 4
199 #define OP_SIZE_10 7
200
201 #define MOV_MASK_0_16 0x0
202 #define MOV_MASK_16_32 0x1
203 #define MOV_MASK_32_48 0x2
204 #define MOV_MASK_48_64 0x3
205 #define MOV_MASK_0_8 0x4
206 #define MOV_MASK_32_64 0x5
207 #define MOV_MASK_52_64 0x6
208
209 #define JMP_SHORTEST 0x0000
210 #define JMP_SHORT 0x0001
211 #define JMP_LONG 0x0002
212 #define JMP_EXTRA_LONG 0x0003
213
214 enum {
215 INSN_ENTRY,
216 INSN_LABEL,
217 INSN_RET,
218 INSN_RET_IMM,
219 INSN_ARM_PUSH,
220 INSN_ARM_POP,
221 INSN_S390_PUSH,
222 INSN_S390_POP,
223 INSN_IA64_ALLOC,
224 INSN_IA64_DEALLOC,
225 INSN_PUSH,
226 INSN_POP,
227 INSN_CALL,
228 INSN_CALL_INDIRECT,
229 INSN_MOV,
230 INSN_MOVSX,
231 INSN_MOV_U,
232 INSN_MOV_LR,
233 INSN_CMP,
234 INSN_CMP_DEST_REG,
235 INSN_CMN,
236 INSN_TEST,
237 INSN_TEST_DEST_REG,
238 INSN_ALU,
239 INSN_ALU_PARTIAL,
240 INSN_ALU_FLAGS,
241 INSN_ALU_FLAGS_PARTIAL,
242 INSN_ALU_TRAP,
243 INSN_ALU_FLAGS_TRAP,
244 INSN_ALU1,
245 INSN_ALU1_PARTIAL,
246 INSN_ALU1_FLAGS,
247 INSN_ALU1_FLAGS_PARTIAL,
248 INSN_ALU1_TRAP,
249 INSN_LEA3,
250 INSN_ROT,
251 INSN_ROT_PARTIAL,
252 INSN_BT,
253 INSN_BTX,
254 INSN_MUL_L,
255 INSN_DIV_L,
256 INSN_MADD,
257 INSN_CBW,
258 INSN_CBW_PARTIAL,
259 INSN_CWD,
260 INSN_CWD_PARTIAL,
261 INSN_SET_COND,
262 INSN_SET_COND_PARTIAL,
263 INSN_CMOV,
264 INSN_CMOV_XCC,
265 INSN_CMP_CMOV,
266 INSN_MOVR,
267 INSN_CSEL_SEL,
268 INSN_CSEL_INC,
269 INSN_CSEL_INV,
270 INSN_CSEL_NEG,
271 INSN_STP,
272 INSN_LDP,
273 INSN_MOV_MASK,
274 INSN_MEMCPY,
275 INSN_MEMSET,
276 INSN_FP_CMP,
277 INSN_FP_CMP_DEST_REG,
278 INSN_FP_CMP_DEST_REG_TRAP,
279 INSN_FP_CMP_UNORDERED_DEST_REG,
280 INSN_FP_CMP_COND,
281 INSN_FP_TEST_REG,
282 INSN_FP_TO_INT_FLAGS,
283 INSN_FP_ALU,
284 INSN_FP_ALU1,
285 INSN_FP_TO_INT32,
286 INSN_FP_TO_INT64,
287 INSN_FP_TO_INT64_TRAP,
288 INSN_FP_FROM_INT32,
289 INSN_FP_FROM_INT64,
290 INSN_FP_INT64_TO_INT32_TRAP,
291 INSN_FP_CVT,
292 INSN_X87_FLD,
293 INSN_X87_FILD,
294 INSN_X87_FSTP,
295 INSN_X87_FISTP,
296 INSN_X87_FISTTP,
297 INSN_X87_FCOMP,
298 INSN_X87_FCOMPP,
299 INSN_X87_FCOMIP,
300 INSN_X87_ALU,
301 INSN_X87_ALUP,
302 INSN_X87_FCHS,
303 INSN_X87_FSQRT,
304 INSN_X87_FRNDINT,
305 INSN_X87_FNSTSW,
306 INSN_X87_FLDCW,
307 INSN_JMP,
308 INSN_JMP_COND,
309 INSN_JMP_COND_LOGICAL,
310 INSN_JMP_REG,
311 INSN_JMP_REG_BIT,
312 INSN_JMP_2REGS,
313 INSN_JMP_FP_TEST,
314 INSN_JMP_INDIRECT,
315 INSN_MB,
316 INSN_CALL_MILLICODE,
317 };
318
319 #define ARG_REGS_MAX 0xc0
320 #define ARG_SHIFTED_REGISTER 0xc0
321 #define ARG_SHIFT_AMOUNT 0x3f
322 #define ARG_SHIFT_MODE 0xc0
323 #define ARG_SHIFT_LSL 0x00
324 #define ARG_SHIFT_LSR 0x40
325 #define ARG_SHIFT_ASR 0x80
326 #define ARG_SHIFT_ROR 0xc0
327 #define ARG_EXTENDED_REGISTER 0xc1
328 #define ARG_EXTEND_SHIFT 0x07
329 #define ARG_EXTEND_MODE 0x38
330 #define ARG_EXTEND_UXTB 0x00
331 #define ARG_EXTEND_UXTH 0x08
332 #define ARG_EXTEND_UXTW 0x10
333 #define ARG_EXTEND_UXTX 0x18
334 #define ARG_EXTEND_SXTB 0x20
335 #define ARG_EXTEND_SXTH 0x28
336 #define ARG_EXTEND_SXTW 0x30
337 #define ARG_EXTEND_SXTX 0x38
338 #define ARG_ADDRESS_0 0xd0
339 #define ARG_ADDRESS_1 0xd1
340 #define ARG_ADDRESS_1_2 0xd2
341 #define ARG_ADDRESS_1_4 0xd3
342 #define ARG_ADDRESS_1_8 0xd4
343 #define ARG_ADDRESS_1_PRE_I 0xd5
344 #define ARG_ADDRESS_1_POST_I 0xd6
345 #define ARG_ADDRESS_2 0xd7
346 #define ARG_ADDRESS_2_2 0xd8
347 #define ARG_ADDRESS_2_4 0xd9
348 #define ARG_ADDRESS_2_8 0xda
349 #define ARG_ADDRESS_2_UXTW 0xdb
350 #define ARG_ADDRESS_2_SXTW 0xdc
351 #define ARG_IMM 0xe0
352
353 #define ARG_IS_ADDRESS(a) ((a) >= ARG_ADDRESS_0 && (a) <= ARG_ADDRESS_2_SXTW)
354
355 #ifdef POINTER_COMPRESSION
356 #define OP_SIZE_SLOT OP_SIZE_4
357 #else
358 #define OP_SIZE_SLOT OP_SIZE_ADDRESS
359 #endif
360
361 #define OP_SIZE_BITMAP (bitmap_64bit ? OP_SIZE_8 : OP_SIZE_4)
362
363 #define OP_SIZE_INT log_2(sizeof(int_default_t))
364
365 #define check_insn(insn) \
366 do { \
367 /*if ((insn_opcode(insn) == INSN_ALU || insn_opcode(insn) == INSN_ALU1) && insn_op_size(insn) != OP_SIZE_NATIVE) internal(file_line, "invalid insn %08x", (unsigned)(insn));*/\
368 /*if (insn == 0x001a000e) internal(file_line, "invalid insn %08x", insn);*/\
369 } while (0)
370
371 #ifdef DEBUG_INSNS
372 #define gen_line() gen_four(__LINE__ + (insn_file << 24))
373 #else
374 #define gen_line() do { } while (0)
375 #endif
376
377 #ifdef ARCH_IA64
378 #define ARCH_CONTEXT struct { \
379 uint64_t insns[3]; \
380 uint8_t insn_units[3]; \
381 bool insn_stops[3]; \
382 uint64_t wr_mask[4]; \
383 }
384 #endif
385
386 #define gen_insn(opcode, op_size, aux, writes_flags) \
387 do { \
388 uint32_t dword = \
389 (uint32_t)(opcode) << INSN_OPCODE_SHIFT | \
390 (uint32_t)(op_size) << INSN_OP_SIZE_SHIFT | \
391 (uint32_t)(aux) << INSN_AUX_SHIFT | \
392 (uint32_t)(writes_flags) << INSN_WRITES_FLAGS_SHIFT; \
393 check_insn(dword); \
394 gen_line(); \
395 gen_four(dword); \
396 } while (0)
397
398 static size_t arg_size(uint8_t arg)
399 {
400 if (arg < ARG_REGS_MAX)
401 return 1;
402 if (arg >= ARG_SHIFTED_REGISTER && arg <= ARG_EXTENDED_REGISTER)
403 return 3;
404 if (arg == ARG_ADDRESS_0)
405 return 9;
406 if (arg >= ARG_ADDRESS_1 && arg <= ARG_ADDRESS_1_POST_I)
407 return 10;
408 if (arg >= ARG_ADDRESS_2 && arg <= ARG_ADDRESS_2_SXTW)
409 return 11;
410 if (arg == ARG_IMM)
411 return 9;
412 internal(file_line, "arg_size: invalid argument %02x", arg);
413 return 0;
414 }
415
416 struct relocation {
417 uint32_t label_id;
418 uint8_t length;
419 size_t position;
420 size_t jmp_instr;
421 };
422
423 struct code_arg {
424 frame_t slot;
425 frame_t flags;
426 frame_t type;
427 };
428
429 struct cg_entry {
430 size_t entry_to_pos;
431 frame_t *variables;
432 size_t n_variables;
433 uint32_t entry_label;
434 uint32_t nonflat_label;
435 };
436
437 struct cg_exit {
438 uint32_t undo_label;
439 uint8_t undo_opcode;
440 uint8_t undo_op_size;
441 uint8_t undo_aux;
442 uint8_t undo_writes_flags;
443 uint8_t undo_parameters[35];
444 uint8_t undo_parameters_len;
445 uint32_t escape_label;
446 };
447
448 #define FLAG_CACHE_IS_FLAT 0x01
449 #define FLAG_CACHE_IS_NOT_FLAT 0x02
450 #define FLAG_CACHE_IS_NOT_THUNK 0x04
451
452 struct codegen_context {
453 struct data *fn;
454 struct data **local_directory;
455
456 const code_t *instr_start;
457 const code_t *current_position;
458 uchar_efficient_t arg_mode;
459
460 uint32_t label_id;
461 struct cg_entry *entries;
462 frame_t n_entries;
463
464 uint8_t *code;
465 size_t code_size;
466
467 uint8_t *code_position;
468
469 uint32_t *code_labels;
470 struct cg_exit **code_exits;
471 uint32_t escape_nospill_label;
472 uint32_t call_label;
473 uint32_t reload_label;
474
475 uint8_t *mcode;
476 size_t mcode_size;
477
478 size_t *label_to_pos;
479 struct relocation *reloc;
480 size_t reloc_size;
481
482 struct trap_record *trap_records;
483 size_t trap_records_size;
484
485 struct code_arg *args;
486 size_t args_l;
487 const code_t *return_values;
488
489 uint8_t *flag_cache;
490 short *registers;
491 frame_t *need_spill;
492 size_t need_spill_l;
493
494 unsigned base_reg;
495 bool offset_reg;
496 int64_t offset_imm;
497
498 bool const_reg;
499 int64_t const_imm;
500
501 struct data *codegen;
502
503 int upcall_args;
504 frame_t *var_aux;
505
506 ajla_error_t err;
507
508 #ifdef ARCH_CONTEXT
509 ARCH_CONTEXT a;
510 #endif
511 };
512
513 static void init_ctx(struct codegen_context *ctx)
514 {
515 ctx->local_directory = NULL;
516 ctx->label_id = 0;
517 ctx->entries = NULL;
518 ctx->n_entries = 0;
519 ctx->code = NULL;
520 ctx->code_labels = NULL;
521 ctx->code_exits = NULL;
522 ctx->escape_nospill_label = 0;
523 ctx->call_label = 0;
524 ctx->reload_label = 0;
525 ctx->mcode = NULL;
526 ctx->label_to_pos = NULL;
527 ctx->reloc = NULL;
528 ctx->trap_records = NULL;
529 ctx->args = NULL;
530 ctx->flag_cache = NULL;
531 ctx->registers = NULL;
532 ctx->need_spill = NULL;
533 ctx->codegen = NULL;
534 ctx->upcall_args = -1;
535 ctx->var_aux = NULL;
536 }
537
538 static void done_ctx(struct codegen_context *ctx)
539 {
540 if (ctx->local_directory)
541 mem_free(ctx->local_directory);
542 if (ctx->entries) {
543 size_t i;
544 for (i = 0; i < ctx->n_entries; i++) {
545 struct cg_entry *ce = &ctx->entries[i];
546 if (ce->variables)
547 mem_free(ce->variables);
548 }
549 mem_free(ctx->entries);
550 }
551 if (ctx->code)
552 mem_free(ctx->code);
553 if (ctx->code_labels)
554 mem_free(ctx->code_labels);
555 if (ctx->code_exits) {
556 ip_t ip;
557 ip_t cs = da(ctx->fn,function)->code_size;
558 for (ip = 0; ip < cs; ip++) {
559 if (ctx->code_exits[ip])
560 mem_free(ctx->code_exits[ip]);
561 }
562 mem_free(ctx->code_exits);
563 }
564 if (ctx->mcode)
565 mem_free(ctx->mcode);
566 if (ctx->label_to_pos)
567 mem_free(ctx->label_to_pos);
568 if (ctx->reloc)
569 mem_free(ctx->reloc);
570 if (ctx->trap_records)
571 mem_free(ctx->trap_records);
572 if (ctx->args)
573 mem_free(ctx->args);
574 if (ctx->flag_cache)
575 mem_free(ctx->flag_cache);
576 if (ctx->registers)
577 mem_free(ctx->registers);
578 if (ctx->need_spill)
579 mem_free(ctx->need_spill);
580 if (ctx->codegen)
581 data_free(ctx->codegen);
582 if (ctx->var_aux)
583 mem_free(ctx->var_aux);
584 }
585
586
587 static inline code_t get_code(struct codegen_context *ctx)
588 {
589 ajla_assert(ctx->current_position < da(ctx->fn,function)->code + da(ctx->fn,function)->code_size, (file_line, "get_code: ran out of code"));
590 return *ctx->current_position++;
591 }
592
593 static inline uint32_t get_uint32(struct codegen_context *ctx)
594 {
595 uint32_t a1 = get_code(ctx);
596 uint32_t a2 = get_code(ctx);
597 #if !CODE_ENDIAN
598 return a1 + (a2 << 16);
599 #else
600 return a2 + (a1 << 16);
601 #endif
602 }
603
604 static int32_t get_jump_offset(struct codegen_context *ctx)
605 {
606 if (SIZEOF_IP_T == 2) {
607 return (int32_t)(int16_t)get_code(ctx);
608 } else if (SIZEOF_IP_T == 4) {
609 return (int32_t)get_uint32(ctx);
610 } else {
611 not_reached();
612 return -1;
613 }
614 }
615
616 static inline void get_one(struct codegen_context *ctx, frame_t *v)
617 {
618 if (!ctx->arg_mode) {
619 code_t c = get_code(ctx);
620 ajla_assert(!(c & ~0xff), (file_line, "get_one: high byte is not cleared: %u", (unsigned)c));
621 *v = c & 0xff;
622 } else if (ctx->arg_mode == 1) {
623 *v = get_code(ctx);
624 #if ARG_MODE_N >= 2
625 } else if (ctx->arg_mode == 2) {
626 *v = get_uint32(ctx);
627 #endif
628 } else {
629 internal(file_line, "get_one: invalid arg mode %u", ctx->arg_mode);
630 }
631 }
632
633 static inline void get_two(struct codegen_context *ctx, frame_t *v1, frame_t *v2)
634 {
635 if (!ctx->arg_mode) {
636 code_t c = get_code(ctx);
637 *v1 = c & 0xff;
638 *v2 = c >> 8;
639 } else if (ctx->arg_mode == 1) {
640 *v1 = get_code(ctx);
641 *v2 = get_code(ctx);
642 #if ARG_MODE_N >= 2
643 } else if (ctx->arg_mode == 2) {
644 *v1 = get_uint32(ctx);
645 *v2 = get_uint32(ctx);
646 #endif
647 } else {
648 internal(file_line, "get_two: invalid arg mode %u", ctx->arg_mode);
649 }
650 }
651
652
653 static uint32_t alloc_label(struct codegen_context *ctx)
654 {
655 return ++ctx->label_id;
656 }
657
658 static struct cg_exit *alloc_cg_exit_for_ip(struct codegen_context *ctx, const code_t *code)
659 {
660 ip_t ip = code - da(ctx->fn,function)->code;
661 struct cg_exit *ce = ctx->code_exits[ip];
662 if (!ce) {
663 ce = mem_calloc_mayfail(struct cg_exit *, sizeof(struct cg_exit), &ctx->err);
664 if (unlikely(!ce))
665 return NULL;
666 ctx->code_exits[ip] = ce;
667 }
668 return ce;
669 }
670
671 static struct cg_exit *alloc_undo_label(struct codegen_context *ctx)
672 {
673 struct cg_exit *ce = alloc_cg_exit_for_ip(ctx, ctx->instr_start);
674 if (unlikely(!ce))
675 return NULL;
676 if (unlikely(ce->undo_label != 0))
677 internal(file_line, "alloc_cg_exit: undo label already allocated");
678 ce->undo_label = alloc_label(ctx);
679 if (unlikely(!ce->undo_label))
680 return NULL;
681 return ce;
682 }
683
684 static uint32_t alloc_escape_label_for_ip(struct codegen_context *ctx, const code_t *code)
685 {
686 struct cg_exit *ce = alloc_cg_exit_for_ip(ctx, code);
687 if (!ce)
688 return 0;
689 if (!ce->escape_label)
690 ce->escape_label = alloc_label(ctx);
691 return ce->escape_label;
692 }
693
694 static uint32_t alloc_escape_label(struct codegen_context *ctx)
695 {
696 return alloc_escape_label_for_ip(ctx, ctx->instr_start);
697 }
698
699 static uint32_t attr_unused alloc_call_label(struct codegen_context *ctx)
700 {
701 if (!ctx->call_label) {
702 ctx->call_label = alloc_label(ctx);
703 }
704 return ctx->call_label;
705 }
706
707 static uint32_t alloc_reload_label(struct codegen_context *ctx)
708 {
709 if (!ctx->reload_label) {
710 ctx->reload_label = alloc_label(ctx);
711 }
712 return ctx->reload_label;
713 }
714
715 static size_t attr_unused mark_params(struct codegen_context *ctx)
716 {
717 return ctx->code_size;
718 }
719
720 static void attr_unused copy_params(struct codegen_context *ctx, struct cg_exit *ce, size_t mark)
721 {
722 if (ctx->code_size - mark > n_array_elements(ce->undo_parameters))
723 internal(file_line, "undo_parameters is too small: %"PRIuMAX" > %"PRIuMAX"", (uintmax_t)(ctx->code_size - mark), (uintmax_t)n_array_elements(ce->undo_parameters));
724 memcpy(ce->undo_parameters, ctx->code + mark, ctx->code_size - mark);
725 ce->undo_parameters_len = ctx->code_size - mark;
726 ctx->code_size = mark;
727 }
728
729 #define g(call) \
730 do { \
731 if (unlikely(!call)) \
732 return false; \
733 } while (0)
734
735 #define gen_one(byte) \
736 do { \
737 /*debug("gen %d: %02x", __LINE__, (uint8_t)(byte))*/; \
738 if (unlikely(!array_add_mayfail(uint8_t, &ctx->code, &ctx->code_size, byte, NULL, &ctx->err)))\
739 return false; \
740 } while (0)
741
742 #if defined(C_LITTLE_ENDIAN)
743 #define gen_two(word) \
744 do { \
745 uint16_t word_ = (word); \
746 /*debug("gen %d: %04x", __LINE__, (uint16_t)(word_));*/ \
747 if (unlikely(!array_add_multiple_mayfail(uint8_t, &ctx->code, &ctx->code_size, cast_ptr(uint8_t *, &word_), 2, NULL, &ctx->err)))\
748 return false; \
749 } while (0)
750 #define gen_four(dword) \
751 do { \
752 uint32_t dword_ = (dword); \
753 /*debug("gen %d: %08x", __LINE__, (uint32_t)(dword_));*/ \
754 if (unlikely(!array_add_multiple_mayfail(uint8_t, &ctx->code, &ctx->code_size, cast_ptr(uint8_t *, &dword_), 4, NULL, &ctx->err)))\
755 return false; \
756 } while (0)
757 #define gen_eight(qword) \
758 do { \
759 uint64_t qword_ = (qword); \
760 /*debug("gen %d: %016lx", __LINE__, (uint64_t)(qword_));*/ \
761 if (unlikely(!array_add_multiple_mayfail(uint8_t, &ctx->code, &ctx->code_size, cast_ptr(uint8_t *, &qword_), 8, NULL, &ctx->err)))\
762 return false; \
763 } while (0)
764 #else
765 #define gen_two(word) \
766 do { \
767 uint16_t word_ = (word); \
768 gen_one(word_ & 0xffU); \
769 gen_one(word_ >> 8); \
770 } while (0)
771 #define gen_four(dword) \
772 do { \
773 uint32_t dword_ = (dword); \
774 gen_two(dword_ & 0xffffU); \
775 gen_two(dword_ >> 15 >> 1); \
776 } while (0)
777 #define gen_eight(qword) \
778 do { \
779 uint64_t qword_ = (qword); \
780 gen_four(qword_ & 0xffffffffUL); \
781 gen_four(qword_ >> 15 >> 15 >> 2); \
782 } while (0)
783 #endif
784
785 #define gen_label(label_id) \
786 do { \
787 gen_insn(INSN_LABEL, 0, 0, 0); \
788 gen_four(label_id); \
789 } while (0)
790
791
792 static uint8_t attr_unused cget_one(struct codegen_context *ctx)
793 {
794 ajla_assert(ctx->code_position < ctx->code + ctx->code_size, (file_line, "cget_one: ran out of code"));
795 return *ctx->code_position++;
796 }
797
798 static uint16_t attr_unused cget_two(struct codegen_context *ctx)
799 {
800 #if defined(C_LITTLE_ENDIAN)
801 uint16_t r;
802 ajla_assert(ctx->code_position < ctx->code + ctx->code_size, (file_line, "cget_two: ran out of code"));
803 memcpy(&r, ctx->code_position, 2);
804 ctx->code_position += 2;
805 return r;
806 #else
807 uint16_t r = cget_one(ctx);
808 r |= cget_one(ctx) << 8;
809 return r;
810 #endif
811 }
812
813 static uint32_t cget_four(struct codegen_context *ctx)
814 {
815 #if defined(C_LITTLE_ENDIAN)
816 uint32_t r;
817 ajla_assert(ctx->code_position < ctx->code + ctx->code_size, (file_line, "cget_four: ran out of code"));
818 memcpy(&r, ctx->code_position, 4);
819 ctx->code_position += 4;
820 return r;
821 #else
822 uint32_t r = cget_two(ctx);
823 r |= (uint32_t)cget_two(ctx) << 16;
824 return r;
825 #endif
826 }
827
828 static uint64_t attr_unused cget_eight(struct codegen_context *ctx)
829 {
830 #if defined(C_LITTLE_ENDIAN)
831 uint64_t r;
832 ajla_assert(ctx->code_position < ctx->code + ctx->code_size, (file_line, "cget_eight: ran out of code"));
833 memcpy(&r, ctx->code_position, 8);
834 ctx->code_position += 8;
835 return r;
836 #else
837 uint64_t r = cget_four(ctx);
838 r |= (uint64_t)cget_four(ctx) << 32;
839 return r;
840 #endif
841 }
842
843 static int64_t get_imm(uint8_t *ptr)
844 {
845 #if defined(C_LITTLE_ENDIAN)
846 int64_t r;
847 memcpy(&r, ptr, 8);
848 return r;
849 #else
850 int64_t r;
851 r = (uint64_t)ptr[0] |
852 ((uint64_t)ptr[1] << 8) |
853 ((uint64_t)ptr[2] << 16) |
854 ((uint64_t)ptr[3] << 24) |
855 ((uint64_t)ptr[4] << 32) |
856 ((uint64_t)ptr[5] << 40) |
857 ((uint64_t)ptr[6] << 48) |
858 ((uint64_t)ptr[7] << 56);
859 return r;
860 #endif
861 }
862
863 #define cgen_one(byte) \
864 do { \
865 if (unlikely(!array_add_mayfail(uint8_t, &ctx->mcode, &ctx->mcode_size, byte, NULL, &ctx->err)))\
866 return false; \
867 } while (0)
868
869 #if defined(C_LITTLE_ENDIAN) || 1
870 #define cgen_two(word) \
871 do { \
872 uint16_t word_ = (word); \
873 if (unlikely(!array_add_multiple_mayfail(uint8_t, &ctx->mcode, &ctx->mcode_size, cast_ptr(uint8_t *, &word_), 2, NULL, &ctx->err)))\
874 return false; \
875 } while (0)
876 #define cgen_four(dword) \
877 do { \
878 uint32_t dword_ = (dword); \
879 /*if (dword_ == 0x1ee02000) internal(file_line, "invalid instruction");*/\
880 if (unlikely(!array_add_multiple_mayfail(uint8_t, &ctx->mcode, &ctx->mcode_size, cast_ptr(uint8_t *, &dword_), 4, NULL, &ctx->err)))\
881 return false; \
882 } while (0)
883 #define cgen_eight(qword) \
884 do { \
885 uint64_t qword_ = (qword); \
886 if (unlikely(!array_add_multiple_mayfail(uint8_t, &ctx->mcode, &ctx->mcode_size, cast_ptr(uint8_t *, &qword_), 8, NULL, &ctx->err)))\
887 return false; \
888 } while (0)
889 #else
890 #define cgen_two(word) \
891 do { \
892 cgen_one((word) & 0xff); \
893 cgen_one((word) >> 8); \
894 } while (0)
895 #define cgen_four(dword) \
896 do { \
897 cgen_two((dword) & 0xffff); \
898 cgen_two((dword) >> 15 >> 1); \
899 } while (0)
900 #define cgen_eight(qword) \
901 do { \
902 cgen_four((qword) & 0xffffffff); \
903 cgen_four((qword) >> 15 >> 15 >> 2); \
904 } while (0)
905 #endif
906
907
908 #define IMM_PURPOSE_LDR_OFFSET 1
909 #define IMM_PURPOSE_LDR_SX_OFFSET 2
910 #define IMM_PURPOSE_STR_OFFSET 3
911 #define IMM_PURPOSE_LDP_STP_OFFSET 4
912 #define IMM_PURPOSE_VLDR_VSTR_OFFSET 5
913 #define IMM_PURPOSE_MVI_CLI_OFFSET 6
914 #define IMM_PURPOSE_STORE_VALUE 7
915 #define IMM_PURPOSE_ADD 8
916 #define IMM_PURPOSE_SUB 9
917 #define IMM_PURPOSE_CMP 10
918 #define IMM_PURPOSE_CMP_LOGICAL 11
919 #define IMM_PURPOSE_AND 12
920 #define IMM_PURPOSE_OR 13
921 #define IMM_PURPOSE_XOR 14
922 #define IMM_PURPOSE_ANDN 15
923 #define IMM_PURPOSE_TEST 16
924 #define IMM_PURPOSE_JMP_2REGS 17
925 #define IMM_PURPOSE_MUL 18
926 #define IMM_PURPOSE_CMOV 19
927 #define IMM_PURPOSE_MOVR 20
928 #define IMM_PURPOSE_BITWISE 21
929 #define IMM_PURPOSE_ADD_TRAP 22
930 #define IMM_PURPOSE_SUB_TRAP 23
931
932
933 static unsigned alu_purpose(unsigned alu)
934 {
935 unsigned purpose =
936 alu == ALU_ADD ? IMM_PURPOSE_ADD :
937 alu == ALU_ADC ? IMM_PURPOSE_ADD :
938 alu == ALU_SUB ? IMM_PURPOSE_SUB :
939 alu == ALU_SBB ? IMM_PURPOSE_SUB :
940 alu == ALU_MUL ? IMM_PURPOSE_MUL :
941 alu == ALU_UMULH ? IMM_PURPOSE_MUL :
942 alu == ALU_SMULH ? IMM_PURPOSE_MUL :
943 alu == ALU_ANDN ? IMM_PURPOSE_ANDN :
944 alu == ALU_AND ? IMM_PURPOSE_AND :
945 alu == ALU_OR ? IMM_PURPOSE_OR :
946 alu == ALU_XOR ? IMM_PURPOSE_XOR :
947 alu == ALU_EXTBL ? IMM_PURPOSE_OR :
948 alu == ALU_EXTWL ? IMM_PURPOSE_OR :
949 alu == ALU_EXTLL ? IMM_PURPOSE_OR :
950 alu == ALU_EXTLH ? IMM_PURPOSE_OR :
951 alu == ALU_INSBL ? IMM_PURPOSE_OR :
952 alu == ALU_MSKBL ? IMM_PURPOSE_OR :
953 alu == ALU_ZAP ? IMM_PURPOSE_ANDN :
954 alu == ALU_ZAPNOT ? IMM_PURPOSE_AND :
955 -1U;
956 if (unlikely(purpose == -1U))
957 internal(file_line, "alu_purpose: invalid alu %u", alu);
958 return purpose;
959 }
960
961 static unsigned alu_trap_purpose(unsigned alu)
962 {
963 unsigned purpose =
964 alu == ALU_ADD ? IMM_PURPOSE_ADD_TRAP :
965 alu == ALU_SUB ? IMM_PURPOSE_SUB_TRAP :
966 -1U;
967 if (unlikely(purpose == -1U))
968 internal(file_line, "alu_trap_purpose: invalid alu %u", alu);
969 return purpose;
970 }
971
972
973 static bool attr_w gen_upcall_end(struct codegen_context *ctx, unsigned args);
974
975 #define gen_address_offset() \
976 do { \
977 if (likely(!ctx->offset_reg)) { \
978 gen_one(ARG_ADDRESS_1); \
979 gen_one(ctx->base_reg); \
980 gen_eight(ctx->offset_imm); \
981 } else { \
982 gen_one(ARG_ADDRESS_2); \
983 gen_one(ctx->base_reg); \
984 gen_one(R_OFFSET_IMM); \
985 gen_eight(0); \
986 } \
987 } while (0)
988
989 #define gen_imm_offset() \
990 do { \
991 if (likely(!ctx->const_reg)) { \
992 gen_one(ARG_IMM); \
993 gen_eight(ctx->const_imm); \
994 } else { \
995 gen_one(R_CONST_IMM); \
996 } \
997 } while (0)
998
999 #define is_imm() (!ctx->const_reg)
1000
1001
1002 static inline bool slot_is_register(struct codegen_context *ctx, frame_t slot)
1003 {
1004 if (frame_t_is_const(slot))
1005 return false;
1006 if (unlikely(slot >= function_n_variables(ctx->fn)))
1007 internal(file_line, "slot_is_register: invalid slot %lu", (unsigned long)slot);
1008 return ctx->registers[slot] >= 0;
1009 }
1010
1011
1012 #define insn_file 1
1013 #if defined(ARCH_ALPHA)
1014 #include "c1-alpha.inc"
1015 #elif defined(ARCH_ARM32)
1016 #include "c1-arm.inc"
1017 #elif defined(ARCH_ARM64)
1018 #include "c1-arm64.inc"
1019 #elif defined(ARCH_IA64)
1020 #include "c1-ia64.inc"
1021 #elif defined(ARCH_LOONGARCH64)
1022 #include "c1-loong.inc"
1023 #elif defined(ARCH_MIPS)
1024 #include "c1-mips.inc"
1025 #elif defined(ARCH_PARISC)
1026 #include "c1-hppa.inc"
1027 #elif defined(ARCH_POWER)
1028 #include "c1-power.inc"
1029 #elif defined(ARCH_S390)
1030 #include "c1-s390.inc"
1031 #elif defined(ARCH_SPARC)
1032 #include "c1-sparc.inc"
1033 #elif defined(ARCH_RISCV64)
1034 #include "c1-riscv.inc"
1035 #elif defined(ARCH_X86)
1036 #include "c1-x86.inc"
1037 #endif
1038 #undef insn_file
1039
1040
1041 #ifndef ARCH_SUPPORTS_TRAPS
1042 #define ARCH_SUPPORTS_TRAPS 0
1043 #define ARCH_TRAP_BEFORE 0
1044 #endif
1045
1046
1047 #define insn_file 2
1048 #include "cg-util.inc"
1049 #undef insn_file
1050
1051 #define insn_file 3
1052 #include "cg-frame.inc"
1053 #undef insn_file
1054
1055 #define insn_file 4
1056 #include "cg-flags.inc"
1057 #undef insn_file
1058
1059 #define insn_file 5
1060 #include "cg-flcch.inc"
1061 #undef insn_file
1062
1063 #define insn_file 6
1064 #include "cg-ptr.inc"
1065 #undef insn_file
1066
1067 #define insn_file 7
1068 #include "cg-alu.inc"
1069 #undef insn_file
1070
1071 #define insn_file 8
1072 #include "cg-ops.inc"
1073 #undef insn_file
1074
1075 #define insn_file 0
1076
1077
1078 #ifndef n_regs_saved
1079 #define n_regs_saved n_array_elements(regs_saved)
1080 #endif
1081
1082 #ifndef n_regs_volatile
1083 #define n_regs_volatile n_array_elements(regs_volatile)
1084 #endif
1085
1086 #ifndef n_fp_saved
1087 #define n_fp_saved n_array_elements(fp_saved)
1088 #endif
1089
1090 #ifndef n_fp_volatile
1091 #define n_fp_volatile n_array_elements(fp_volatile)
1092 #endif
1093
1094 #ifndef n_vector_volatile
1095 #define n_vector_volatile n_array_elements(vector_volatile)
1096 #endif
1097
1098 static bool attr_w gen_registers(struct codegen_context *ctx)
1099 {
1100 frame_t v;
1101 size_t index_saved = 0;
1102 size_t index_volatile = 0;
1103 size_t index_fp_saved = 0;
1104 size_t index_fp_volatile = 0;
1105 size_t attr_unused index_vector_volatile = 0;
1106 #ifdef ARCH_S390
1107 bool uses_x = false;
1108 for (v = MIN_USEABLE_SLOT; v < function_n_variables(ctx->fn); v++) {
1109 const struct type *t = get_type_of_local(ctx, v);
1110 if (t && TYPE_TAG_IS_REAL(t->tag) && TYPE_TAG_IDX_REAL(t->tag) == 4) {
1111 uses_x = true;
1112 break;
1113 }
1114 }
1115 #endif
1116 /*for (v = function_n_variables(ctx->fn) - 1; v >= MIN_USEABLE_SLOT; v--)*/
1117 for (v = MIN_USEABLE_SLOT; v < function_n_variables(ctx->fn); v++) {
1118 const struct type *t;
1119 ctx->registers[v] = -1;
1120 if (ra_chicken)
1121 continue;
1122 t = get_type_of_local(ctx, v);
1123 if (unlikely(!t))
1124 continue;
1125 if (!da(ctx->fn,function)->local_variables_flags[v].must_be_flat &&
1126 !da(ctx->fn,function)->local_variables_flags[v].must_be_data)
1127 continue;
1128 if (!ARCH_HAS_BWX && t->size < 1U << OP_SIZE_4)
1129 continue;
1130 if (TYPE_TAG_IS_FIXED(t->tag) || TYPE_TAG_IS_INT(t->tag) || t->tag == TYPE_TAG_flat_option || t->tag == TYPE_TAG_unknown || t->tag == TYPE_TAG_record) {
1131 if (TYPE_TAG_IS_BUILTIN(t->tag)) {
1132 if (!is_power_of_2(t->size) || t->size > 1U << OP_SIZE_NATIVE)
1133 continue;
1134 }
1135 if (index_saved < n_regs_saved + zero
1136 #if defined(ARCH_PARISC) || defined(ARCH_SPARC)
1137 && t->size <= 1U << OP_SIZE_ADDRESS
1138 #endif
1139 ) {
1140 ctx->registers[v] = regs_saved[index_saved++];
1141 } else if (index_volatile < n_regs_volatile + zero) {
1142 ctx->registers[v] = regs_volatile[index_volatile++];
1143 } else {
1144 continue;
1145 }
1146 } else if (TYPE_TAG_IS_REAL(t->tag)) {
1147 unsigned real_type = TYPE_TAG_IDX_REAL(t->tag);
1148 if ((SUPPORTED_FP >> real_type) & 1) {
1149 #ifdef ARCH_POWER
1150 if (real_type == 4) {
1151 if (index_vector_volatile < n_vector_volatile + zero) {
1152 ctx->registers[v] = vector_volatile[index_vector_volatile++];
1153 goto success;
1154 }
1155 continue;
1156 }
1157 #endif
1158 #ifdef ARCH_S390
1159 if (real_type == 4) {
1160 if (!(index_fp_saved & 1) && index_fp_saved + 1 < n_fp_saved + zero) {
1161 ctx->registers[v] = fp_saved[index_fp_saved++];
1162 index_fp_saved++;
1163 goto success;
1164 }
1165 if (index_fp_saved & 1 && index_fp_saved + 2 < n_fp_saved + zero) {
1166 index_fp_saved++;
1167 ctx->registers[v] = fp_saved[index_fp_saved++];
1168 index_fp_saved++;
1169 goto success;
1170 }
1171 if (!(index_fp_volatile & 1) && index_fp_volatile + 1 < n_fp_volatile + zero) {
1172 ctx->registers[v] = fp_volatile[index_fp_volatile++];
1173 index_fp_volatile++;
1174 goto success;
1175 }
1176 if (index_fp_volatile & 1 && index_fp_volatile + 2 < n_fp_volatile + zero) {
1177 index_fp_volatile++;
1178 ctx->registers[v] = fp_volatile[index_fp_volatile++];
1179 index_fp_volatile++;
1180 goto success;
1181 }
1182 continue;
1183 }
1184 #endif
1185 if (index_fp_saved < n_fp_saved + zero) {
1186 ctx->registers[v] = fp_saved[index_fp_saved++];
1187 } else if (index_fp_volatile < n_fp_volatile + zero) {
1188 ctx->registers[v] = fp_volatile[index_fp_volatile++];
1189 } else {
1190 continue;
1191 }
1192 } else {
1193 continue;
1194 }
1195 } else {
1196 continue;
1197 }
1198 goto success;
1199 success:
1200 if (!reg_is_saved(ctx->registers[v])) {
1201 if (unlikely(!array_add_mayfail(frame_t, &ctx->need_spill, &ctx->need_spill_l, v, NULL, &ctx->err)))
1202 return false;
1203 }
1204 }
1205
1206 return true;
1207 }
1208
1209 static bool attr_w gen_fused_binary(struct codegen_context *ctx, unsigned mode, unsigned op_size, unsigned op, uint32_t escape_label, frame_t slot_1, frame_t slot_2, frame_t slot_r, bool *failed)
1210 {
1211 const code_t *backup = ctx->current_position;
1212 code_t code;
1213 frame_t slot_dr, slot_test;
1214 int32_t offs_false;
1215
1216 *failed = false;
1217
1218 next_code:
1219 code = get_code(ctx);
1220 ctx->arg_mode = code / OPCODE_MODE_MULT;
1221 code %= OPCODE_MODE_MULT;
1222 ajla_assert_lo(ctx->arg_mode < ARG_MODE_N, (file_line, "gen_fused_binary: invalid opcode %04x", (unsigned)*ctx->instr_start));
1223
1224 if (code == OPCODE_DEREFERENCE) {
1225 get_one(ctx, &slot_dr);
1226 const struct type *t = get_type_of_local(ctx, slot_dr);
1227 if (!TYPE_TAG_IS_BUILTIN(t->tag)) {
1228 *failed = true;
1229 goto fail;
1230 }
1231 if (unlikely(!flag_is_clear(ctx, slot_dr))) {
1232 *failed = true;
1233 goto fail;
1234 }
1235 goto next_code;
1236 }
1237 if (code == OPCODE_DEREFERENCE_CLEAR) {
1238 *failed = true;
1239 goto fail;
1240 }
1241 if (unlikely(code != OPCODE_JMP_FALSE))
1242 internal(file_line, "gen_fused_binary: binary operation is not followed by jmp false: %x, %s", code, decode_opcode(code, true));
1243 get_one(ctx, &slot_test);
1244 if (unlikely(slot_test != slot_r))
1245 internal(file_line, "gen_fused_binary: the result of the binary operation and the tested variable do not match");
1246 offs_false = get_jump_offset(ctx);
1247 get_jump_offset(ctx);
1248
1249 if (mode == MODE_ARRAY_LEN_GT) {
1250 g(gen_array_len(ctx, slot_1, slot_2, slot_r, true, offs_false));
1251 } else if (mode == MODE_REAL) {
1252 g(gen_fp_alu_jmp(ctx, op_size, op, escape_label, slot_1, slot_2, offs_false, failed));
1253 } else {
1254 g(gen_alu_jmp(ctx, mode, op_size, op, slot_1, slot_2, offs_false, failed));
1255 }
1256
1257 fail:
1258 if (*failed)
1259 ctx->current_position = backup;
1260
1261 return true;
1262 }
1263
1264 static bool attr_w gen_function(struct codegen_context *ctx)
1265 {
1266 ctx->current_position = da(ctx->fn,function)->code;
1267
1268 ctx->escape_nospill_label = alloc_label(ctx);
1269 if (unlikely(!ctx->escape_nospill_label))
1270 return false;
1271
1272 while (ctx->current_position != da(ctx->fn,function)->code + da(ctx->fn,function)->code_size) {
1273 ip_t ip;
1274 code_t code;
1275 unsigned op, type;
1276 frame_t slot_1, slot_2, slot_3, slot_r, flags, fn_idx, opt;
1277 arg_t n_args, n_ret, i_arg;
1278 uint32_t label_id;
1279 uint32_t escape_label;
1280 bool failed;
1281
1282 ajla_assert_lo(ctx->current_position < da(ctx->fn,function)->code + da(ctx->fn,function)->code_size, (file_line, "gen_function: ran out of code in %s", da(ctx->fn,function)->function_name));
1283
1284 ctx->instr_start = ctx->current_position;
1285
1286 /*debug("%s: %04x, %s", da(ctx->fn,function)->function_name, *ctx->instr_start, decode_opcode(*ctx->instr_start, true));*/
1287
1288 ip = ctx->instr_start - da(ctx->fn,function)->code;
1289 if (likely(!ctx->code_labels[ip])) {
1290 ctx->code_labels[ip] = alloc_label(ctx);
1291 if (unlikely(!ctx->code_labels[ip]))
1292 return false;
1293 }
1294 gen_label(ctx->code_labels[ip]);
1295
1296 code = get_code(ctx);
1297 ctx->arg_mode = code / OPCODE_MODE_MULT;
1298 code %= OPCODE_MODE_MULT;
1299 ajla_assert_lo(ctx->arg_mode < ARG_MODE_N, (file_line, "gen_function: invalid opcode %04x", (unsigned)*ctx->instr_start));
1300
1301 if (code >= OPCODE_FIXED_OP + uzero && code < OPCODE_INT_OP) {
1302 code -= OPCODE_FIXED_OP;
1303 op = (code / OPCODE_FIXED_OP_MULT) % OPCODE_FIXED_TYPE_MULT;
1304 type = code / OPCODE_FIXED_TYPE_MULT;
1305 if (op < OPCODE_FIXED_OP_UNARY) {
1306 get_two(ctx, &slot_1, &slot_2);
1307 get_two(ctx, &slot_r, &flags);
1308 escape_label = alloc_escape_label(ctx);
1309 if (unlikely(!escape_label))
1310 return false;
1311 g(gen_test_2_cached(ctx, slot_1, slot_2, escape_label));
1312 flag_set(ctx, slot_1, false);
1313 flag_set(ctx, slot_2, false);
1314 flag_set(ctx, slot_r, false);
1315 if (flags & OPCODE_FLAG_FUSED) {
1316 g(gen_fused_binary(ctx, MODE_FIXED, type, op, escape_label, slot_1, slot_2, slot_r, &failed));
1317 if (unlikely(!failed))
1318 continue;
1319 }
1320 g(gen_alu(ctx, MODE_FIXED, type, op, escape_label, slot_1, slot_2, slot_r));
1321 continue;
1322 } else if (op < OPCODE_FIXED_OP_N) {
1323 get_two(ctx, &slot_1, &slot_r);
1324 get_one(ctx, &flags);
1325 escape_label = alloc_escape_label(ctx);
1326 if (unlikely(!escape_label))
1327 return false;
1328 g(gen_test_1_cached(ctx, slot_1, escape_label));
1329 flag_set(ctx, slot_1, false);
1330 flag_set(ctx, slot_r, false);
1331 g(gen_alu1(ctx, MODE_FIXED, type, op, escape_label, slot_1, slot_r));
1332 continue;
1333 } else if (op == OPCODE_FIXED_OP_ldc) {
1334 unsigned i;
1335 get_one(ctx, &slot_r);
1336 g(gen_constant(ctx, false, type, false, slot_r));
1337 for (i = 0; i < 1U << type; i += 2)
1338 get_code(ctx);
1339 flag_set(ctx, slot_r, false);
1340 continue;
1341 } else if (op == OPCODE_FIXED_OP_ldc16) {
1342 get_one(ctx, &slot_r);
1343 g(gen_constant(ctx, false, type, true, slot_r));
1344 get_code(ctx);
1345 flag_set(ctx, slot_r, false);
1346 continue;
1347 } else if (op == OPCODE_FIXED_OP_move || op == OPCODE_FIXED_OP_copy) {
1348 get_two(ctx, &slot_1, &slot_r);
1349 escape_label = alloc_escape_label(ctx);
1350 if (unlikely(!escape_label))
1351 return false;
1352 g(gen_test_1_cached(ctx, slot_1, escape_label));
1353 flag_set(ctx, slot_1, false);
1354 flag_set(ctx, slot_r, false);
1355 g(gen_copy(ctx, type, slot_1, slot_r));
1356 continue;
1357 } else {
1358 internal(file_line, "gen_function: bad fixed code %04x", *ctx->instr_start);
1359 }
1360 } else if (code >= OPCODE_INT_OP && code < OPCODE_REAL_OP) {
1361 code -= OPCODE_INT_OP;
1362 op = (code / OPCODE_INT_OP_MULT) % OPCODE_INT_TYPE_MULT;
1363 type = code / OPCODE_INT_TYPE_MULT;
1364 if (op < OPCODE_INT_OP_C) {
1365 get_two(ctx, &slot_1, &slot_2);
1366 get_two(ctx, &slot_r, &flags);
1367 escape_label = alloc_escape_label(ctx);
1368 if (unlikely(!escape_label))
1369 return false;
1370 g(gen_test_2_cached(ctx, slot_1, slot_2, escape_label));
1371 flag_set(ctx, slot_1, false);
1372 flag_set(ctx, slot_2, false);
1373 flag_set(ctx, slot_r, false);
1374 if (flags & OPCODE_FLAG_FUSED) {
1375 g(gen_fused_binary(ctx, MODE_INT, type, op, escape_label, slot_1, slot_2, slot_r, &failed));
1376 if (unlikely(!failed))
1377 continue;
1378 }
1379 g(gen_alu(ctx, MODE_INT, type, op, escape_label, slot_1, slot_2, slot_r));
1380 continue;
1381 } else if (op < OPCODE_INT_OP_UNARY) {
1382 op -= OPCODE_INT_OP_C;
1383 get_two(ctx, &slot_1, &slot_2);
1384 get_two(ctx, &slot_r, &flags);
1385 escape_label = alloc_escape_label(ctx);
1386 if (unlikely(!escape_label))
1387 return false;
1388 g(gen_test_1_cached(ctx, slot_1, escape_label));
1389 flag_set(ctx, slot_1, false);
1390 flag_set(ctx, slot_r, false);
1391 slot_2 = frame_t_from_const((int32_t)slot_2);
1392 if (flags & OPCODE_FLAG_FUSED) {
1393 g(gen_fused_binary(ctx, MODE_INT, type, op, escape_label, slot_1, slot_2, slot_r, &failed));
1394 if (unlikely(!failed))
1395 continue;
1396 }
1397 g(gen_alu(ctx, MODE_INT, type, op, escape_label, slot_1, slot_2, slot_r));
1398 continue;
1399 } else if (op < OPCODE_INT_OP_N) {
1400 get_two(ctx, &slot_1, &slot_r);
1401 get_one(ctx, &flags);
1402 if ((op == OPCODE_INT_OP_to_int || op == OPCODE_INT_OP_from_int) && slot_1 == slot_r)
1403 continue;
1404 escape_label = alloc_escape_label(ctx);
1405 if (unlikely(!escape_label))
1406 return false;
1407 g(gen_test_1_cached(ctx, slot_1, escape_label));
1408 flag_set(ctx, slot_1, false);
1409 flag_set(ctx, slot_r, false);
1410 g(gen_alu1(ctx, MODE_INT, type, op, escape_label, slot_1, slot_r));
1411 continue;
1412 } else if (op == OPCODE_INT_OP_ldc) {
1413 unsigned i;
1414 get_one(ctx, &slot_r);
1415 g(gen_constant(ctx, false, type, false, slot_r));
1416 for (i = 0; i < 1U << type; i += 2)
1417 get_code(ctx);
1418 flag_set(ctx, slot_r, false);
1419 continue;
1420 } else if (op == OPCODE_INT_OP_ldc16) {
1421 get_one(ctx, &slot_r);
1422 g(gen_constant(ctx, false, type, true, slot_r));
1423 get_code(ctx);
1424 flag_set(ctx, slot_r, false);
1425 continue;
1426 } else if (op == OPCODE_INT_OP_move || op == OPCODE_INT_OP_copy) {
1427 get_two(ctx, &slot_1, &slot_r);
1428 escape_label = alloc_escape_label(ctx);
1429 if (unlikely(!escape_label))
1430 return false;
1431 g(gen_test_1_cached(ctx, slot_1, escape_label));
1432 flag_set(ctx, slot_1, false);
1433 flag_set(ctx, slot_r, false);
1434 g(gen_copy(ctx, type, slot_1, slot_r));
1435 continue;
1436 } else {
1437 internal(file_line, "gen_function: bad integer code %04x", *ctx->instr_start);
1438 }
1439 } else if (code >= OPCODE_REAL_OP && code < OPCODE_BOOL_OP) {
1440 code -= OPCODE_REAL_OP;
1441 op = (code / OPCODE_REAL_OP_MULT) % OPCODE_REAL_TYPE_MULT;
1442 type = code / OPCODE_REAL_TYPE_MULT;
1443 if (op < OPCODE_REAL_OP_UNARY) {
1444 get_two(ctx, &slot_1, &slot_2);
1445 get_two(ctx, &slot_r, &flags);
1446 escape_label = alloc_escape_label(ctx);
1447 if (unlikely(!escape_label))
1448 return false;
1449 g(gen_test_2_cached(ctx, slot_1, slot_2, escape_label));
1450 flag_set(ctx, slot_1, false);
1451 flag_set(ctx, slot_2, false);
1452 flag_set(ctx, slot_r, false);
1453 if (flags & OPCODE_FLAG_FUSED) {
1454 g(gen_fused_binary(ctx, MODE_REAL, type, op, escape_label, slot_1, slot_2, slot_r, &failed));
1455 if (unlikely(!failed))
1456 continue;
1457 }
1458 g(gen_fp_alu(ctx, type, op, escape_label, slot_1, slot_2, slot_r));
1459 continue;
1460 } else if (op < OPCODE_REAL_OP_N) {
1461 get_two(ctx, &slot_1, &slot_r);
1462 get_one(ctx, &flags);
1463 escape_label = alloc_escape_label(ctx);
1464 if (unlikely(!escape_label))
1465 return false;
1466 g(gen_test_1_cached(ctx, slot_1, escape_label));
1467 flag_set(ctx, slot_1, false);
1468 flag_set(ctx, slot_r, false);
1469 g(gen_fp_alu1(ctx, type, op, escape_label, slot_1, slot_r));
1470 continue;
1471 } else if (op == OPCODE_REAL_OP_ldc) {
1472 const struct type *t;
1473 unsigned i;
1474 get_one(ctx, &slot_r);
1475 t = type_get_real(type);
1476 g(gen_real_constant(ctx, t, slot_r));
1477 for (i = 0; i < t->size; i += 2)
1478 get_code(ctx);
1479 flag_set(ctx, slot_r, false);
1480 continue;
1481 } else if (op == OPCODE_REAL_OP_move || op == OPCODE_REAL_OP_copy) {
1482 get_two(ctx, &slot_1, &slot_r);
1483 escape_label = alloc_escape_label(ctx);
1484 if (unlikely(!escape_label))
1485 return false;
1486 g(gen_test_1_cached(ctx, slot_1, escape_label));
1487 flag_set(ctx, slot_1, false);
1488 flag_set(ctx, slot_r, false);
1489 g(gen_memcpy_slots(ctx, slot_r, slot_1));
1490 continue;
1491 } else {
1492 internal(file_line, "gen_function: bad real code %04x", *ctx->instr_start);
1493 }
1494 } else if (code >= OPCODE_BOOL_OP && code < OPCODE_EXTRA) {
1495 code -= OPCODE_BOOL_OP;
1496 op = (code / OPCODE_BOOL_OP_MULT) % OPCODE_BOOL_TYPE_MULT;
1497 type = log_2(sizeof(ajla_flat_option_t));
1498 if (op < OPCODE_BOOL_OP_UNARY) {
1499 get_two(ctx, &slot_1, &slot_2);
1500 get_two(ctx, &slot_r, &flags);
1501 escape_label = alloc_escape_label(ctx);
1502 if (unlikely(!escape_label))
1503 return false;
1504 g(gen_test_2_cached(ctx, slot_1, slot_2, escape_label));
1505 flag_set(ctx, slot_1, false);
1506 flag_set(ctx, slot_2, false);
1507 flag_set(ctx, slot_r, false);
1508 if (flags & OPCODE_FLAG_FUSED) {
1509 g(gen_fused_binary(ctx, MODE_BOOL, type, op, escape_label, slot_1, slot_2, slot_r, &failed));
1510 if (unlikely(!failed))
1511 continue;
1512 }
1513 g(gen_alu(ctx, MODE_BOOL, type, op, escape_label, slot_1, slot_2, slot_r));
1514 continue;
1515 } else if (op < OPCODE_BOOL_OP_N) {
1516 get_two(ctx, &slot_1, &slot_r);
1517 get_one(ctx, &flags);
1518 escape_label = alloc_escape_label(ctx);
1519 if (unlikely(!escape_label))
1520 return false;
1521 g(gen_test_1_cached(ctx, slot_1, escape_label));
1522 flag_set(ctx, slot_1, false);
1523 flag_set(ctx, slot_r, false);
1524 g(gen_alu1(ctx, MODE_BOOL, type, op, escape_label, slot_1, slot_r));
1525 continue;
1526 } else if (op == OPCODE_BOOL_OP_move || op == OPCODE_BOOL_OP_copy) {
1527 get_two(ctx, &slot_1, &slot_r);
1528 escape_label = alloc_escape_label(ctx);
1529 if (unlikely(!escape_label))
1530 return false;
1531 g(gen_test_1_cached(ctx, slot_1, escape_label));
1532 flag_set(ctx, slot_1, false);
1533 flag_set(ctx, slot_r, false);
1534 g(gen_copy(ctx, type, slot_1, slot_r));
1535 continue;
1536 } else {
1537 internal(file_line, "gen_function: bad boolean code %04x", *ctx->instr_start);
1538 }
1539 } else switch (code) {
1540 case OPCODE_INT_LDC_LONG: {
1541 uint32_t words, w;
1542 get_one(ctx, &slot_r);
1543 words = get_uint32(ctx);
1544 for (w = 0; w < words; w++)
1545 get_code(ctx);
1546 unconditional_escape:
1547 escape_label = alloc_escape_label(ctx);
1548 if (unlikely(!escape_label))
1549 return false;
1550 gen_insn(INSN_JMP, 0, 0, 0);
1551 gen_four(escape_label);
1552 continue;
1553 }
1554 case OPCODE_IS_EXCEPTION: {
1555 get_two(ctx, &slot_1, &slot_r);
1556 get_one(ctx, &flags);
1557 g(gen_is_exception(ctx, slot_1, slot_r));
1558 continue;
1559 }
1560 case OPCODE_EXCEPTION_CLASS:
1561 case OPCODE_EXCEPTION_TYPE:
1562 case OPCODE_EXCEPTION_AUX: {
1563 get_two(ctx, &slot_1, &slot_r);
1564 get_one(ctx, &flags);
1565 goto unconditional_escape;
1566 }
1567 case OPCODE_SYSTEM_PROPERTY: {
1568 get_two(ctx, &slot_1, &slot_r);
1569 get_one(ctx, &flags);
1570 g(gen_system_property(ctx, slot_1, slot_r));
1571 continue;
1572 }
1573 case OPCODE_FLAT_MOVE:
1574 case OPCODE_FLAT_COPY: {
1575 get_two(ctx, &slot_1, &slot_r);
1576 g(gen_flat_move_copy(ctx, slot_1, slot_r));
1577 continue;
1578 }
1579 case OPCODE_REF_MOVE:
1580 case OPCODE_REF_MOVE_CLEAR:
1581 case OPCODE_REF_COPY: {
1582 get_two(ctx, &slot_1, &slot_r);
1583 g(gen_ref_move_copy(ctx, code, slot_1, slot_r));
1584 continue;
1585 }
1586 case OPCODE_BOX_MOVE_CLEAR:
1587 case OPCODE_BOX_COPY: {
1588 get_two(ctx, &slot_1, &slot_r);
1589 g(gen_box_move_copy(ctx, code, slot_1, slot_r));
1590 continue;
1591 }
1592 case OPCODE_TAKE_BORROWED:
1593 get_one(ctx, &slot_1);
1594 if (!da(ctx->fn,function)->local_variables_flags[slot_1].may_be_borrowed)
1595 continue;
1596 if (unlikely(!(label_id = alloc_label(ctx))))
1597 return false;
1598 if (flag_is_set(ctx, slot_1))
1599 goto take_borrowed_done;
1600 if (flag_is_clear(ctx, slot_1)) {
1601 g(gen_set_1(ctx, R_FRAME, slot_1, 0, true));
1602 goto do_take_borrowed;
1603 }
1604 g(gen_test_1(ctx, R_FRAME, slot_1, 0, label_id, false, TEST_SET));
1605 do_take_borrowed:
1606 g(gen_upcall_start(ctx, 1));
1607 g(gen_frame_load(ctx, OP_SIZE_SLOT, garbage, slot_1, 0, false, R_ARG0));
1608 g(gen_upcall_argument(ctx, 0));
1609 g(gen_upcall(ctx, offsetof(struct cg_upcall_vector_s, cg_upcall_pointer_reference_owned), 1));
1610 flag_set(ctx, slot_1, true);
1611 take_borrowed_done:
1612 gen_label(label_id);
1613 continue;
1614 case OPCODE_DEREFERENCE:
1615 case OPCODE_DEREFERENCE_CLEAR: {
1616 bool need_bit_test;
1617 /*const struct type *type;*/
1618 get_one(ctx, &slot_1);
1619 if (flag_is_clear(ctx, slot_1))
1620 goto skip_dereference;
1621 /*type = get_type_of_local(ctx, slot_1);*/
1622 /*need_bit_test = 1 || TYPE_IS_FLAT(type) || da(ctx->fn,function)->local_variables[slot_1].may_be_borrowed;*/
1623 need_bit_test = !flag_is_set(ctx, slot_1);
1624 if (need_bit_test) {
1625 if (unlikely(!(label_id = alloc_label(ctx))))
1626 return false;
1627 g(gen_test_1(ctx, R_FRAME, slot_1, 0, label_id, true, TEST_CLEAR));
1628 } else {
1629 g(gen_set_1(ctx, R_FRAME, slot_1, 0, false));
1630 label_id = 0; /* avoid warning */
1631 }
1632 g(gen_upcall_start(ctx, 1));
1633 g(gen_frame_load(ctx, OP_SIZE_SLOT, garbage, slot_1, 0, false, R_ARG0));
1634 g(gen_upcall_argument(ctx, 0));
1635 g(gen_upcall(ctx, offsetof(struct cg_upcall_vector_s, cg_upcall_pointer_dereference), 1));
1636 if (need_bit_test)
1637 gen_label(label_id);
1638 skip_dereference:
1639 if (code == OPCODE_DEREFERENCE_CLEAR)
1640 g(gen_frame_clear(ctx, OP_SIZE_SLOT, slot_1));
1641 flag_set_unknown(ctx, slot_1);
1642 flag_set(ctx, slot_1, false);
1643 continue;
1644 }
1645 case OPCODE_EVAL: {
1646 get_one(ctx, &slot_1);
1647 g(gen_eval(ctx, slot_1));
1648 continue;
1649 }
1650 case OPCODE_ESCAPE_NONFLAT: {
1651 frame_t n, i;
1652 frame_t *vars;
1653
1654 get_one(ctx, &n);
1655 vars = mem_alloc_array_mayfail(mem_alloc_mayfail, frame_t *, 0, 0, n, sizeof(frame_t), &ctx->err);
1656 if (unlikely(!vars))
1657 return false;
1658 for (i = 0; i < n; i++) {
1659 get_one(ctx, &vars[i]);
1660 }
1661
1662 escape_label = alloc_escape_label(ctx);
1663 if (unlikely(!escape_label)) {
1664 mem_free(vars);
1665 return false;
1666 }
1667
1668 if (unlikely(!gen_test_variables(ctx, vars, n, escape_label))) {
1669 mem_free(vars);
1670 return false;
1671 }
1672 mem_free(vars);
1673
1674 continue;
1675 }
1676 case OPCODE_CHECKPOINT: {
1677 frame_t n_vars;
1678
1679 g(clear_flag_cache(ctx));
1680
1681 if (SIZEOF_IP_T == 2) {
1682 slot_1 = get_code(ctx);
1683 } else if (SIZEOF_IP_T == 4) {
1684 slot_1 = get_uint32(ctx);
1685 } else {
1686 not_reached();
1687 continue;
1688 }
1689
1690 if (unlikely(!(slot_1 + 1)))
1691 return false;
1692 while (slot_1 >= ctx->n_entries) {
1693 void *err_entries;
1694 struct cg_entry e;
1695 if (unlikely(!ctx->entries)) {
1696 if (unlikely(!array_init_mayfail(struct cg_entry, &ctx->entries, &ctx->n_entries, &ctx->err)))
1697 return false;
1698 }
1699 memset(&e, 0, sizeof(struct cg_entry));
1700 if (unlikely(!array_add_mayfail(struct cg_entry, &ctx->entries, &ctx->n_entries, e, &err_entries, &ctx->err))) {
1701 ctx->entries = err_entries;
1702 return false;
1703 }
1704 }
1705
1706 get_one(ctx, &n_vars);
1707
1708 escape_label = 0; /* avoid warning */
1709 if (likely(slot_1 != 0)) {
1710 escape_label = alloc_escape_label(ctx);
1711 if (unlikely(!escape_label))
1712 return false;
1713 }
1714
1715 if (n_vars || !slot_1) {
1716 frame_t i;
1717 uint32_t entry_label, nonflat_label;
1718 struct cg_entry *ce = &ctx->entries[slot_1];
1719
1720 if (unlikely(!array_init_mayfail(frame_t, &ce->variables, &ce->n_variables, &ctx->err)))
1721 return false;
1722 for (i = 0; i < n_vars; i++) {
1723 frame_t v;
1724 get_one(ctx, &v);
1725 if (unlikely(!array_add_mayfail(frame_t, &ce->variables, &ce->n_variables, v, NULL, &ctx->err)))
1726 return false;
1727 }
1728 if (!slot_1) {
1729 g(gen_test_variables(ctx, ce->variables, ce->n_variables, ctx->escape_nospill_label));
1730 }
1731 entry_label = alloc_label(ctx);
1732 if (unlikely(!entry_label))
1733 return false;
1734 gen_label(entry_label);
1735 ce->entry_label = entry_label;
1736
1737 nonflat_label = alloc_escape_label_for_ip(ctx, ctx->current_position);
1738 if (unlikely(!nonflat_label))
1739 return false;
1740 ce->nonflat_label = nonflat_label;
1741
1742 if (unlikely(!slot_1))
1743 g(gen_timestamp_test(ctx, ctx->escape_nospill_label));
1744 else
1745 g(gen_timestamp_test(ctx, escape_label));
1746 } else {
1747 g(gen_timestamp_test(ctx, escape_label));
1748
1749 gen_insn(INSN_ENTRY, 0, 0, 0);
1750 gen_four(slot_1);
1751 }
1752 continue;
1753 }
1754 case OPCODE_JMP: {
1755 int32_t x = get_jump_offset(ctx);
1756 g(gen_jump(ctx, x, OP_SIZE_NATIVE, COND_ALWAYS, -1U, -1U));
1757 continue;
1758 }
1759 case OPCODE_JMP_BACK_16: {
1760 int32_t x = get_code(ctx);
1761 g(gen_jump(ctx, -x - (int)(2 * sizeof(code_t)), OP_SIZE_NATIVE, COND_ALWAYS, -1U, -1U));
1762 continue;
1763 }
1764 case OPCODE_JMP_FALSE: {
1765 int32_t offs_false;
1766 get_one(ctx, &slot_1);
1767 offs_false = get_jump_offset(ctx);
1768 get_jump_offset(ctx);
1769 escape_label = alloc_escape_label(ctx);
1770 if (unlikely(!escape_label))
1771 return false;
1772 g(gen_test_1_cached(ctx, slot_1, escape_label));
1773 flag_set(ctx, slot_1, false);
1774 g(gen_cond_jump(ctx, slot_1, offs_false));
1775 continue;
1776 }
1777 case OPCODE_LABEL: {
1778 g(clear_flag_cache(ctx));
1779 continue;
1780 }
1781 #define init_args \
1782 do { \
1783 if (ctx->args != NULL) \
1784 mem_free(ctx->args); \
1785 g(array_init_mayfail(struct code_arg, &ctx->args, &ctx->args_l, &ctx->err));\
1786 } while (0)
1787 #define load_args \
1788 do { \
1789 init_args; \
1790 for (i_arg = 0; i_arg < n_args; i_arg++) { \
1791 struct code_arg a; \
1792 get_two(ctx, &a.slot, &a.flags); \
1793 a.type = 0; \
1794 g(array_add_mayfail(struct code_arg, &ctx->args, &ctx->args_l, a, NULL, &ctx->err));\
1795 } \
1796 } while (0)
1797 case OPCODE_LOAD_FN:
1798 get_two(ctx, &n_args, &slot_r);
1799 get_one(ctx, &fn_idx);
1800 load_args;
1801 g(gen_load_fn_or_curry(ctx, fn_idx, NO_FRAME_T, slot_r, 0));
1802 continue;
1803 case OPCODE_CURRY:
1804 get_two(ctx, &n_args, &slot_r);
1805 get_two(ctx, &slot_1, &flags);
1806 load_args;
1807 g(gen_load_fn_or_curry(ctx, NO_FRAME_T, slot_1, slot_r, flags));
1808 continue;
1809 case OPCODE_CALL:
1810 case OPCODE_CALL_STRICT:
1811 case OPCODE_CALL_SPARK:
1812 case OPCODE_CALL_LAZY:
1813 case OPCODE_CALL_CACHE:
1814 case OPCODE_CALL_SAVE: {
1815 get_two(ctx, &n_args, &n_ret);
1816 get_one(ctx, &fn_idx);
1817 jump_over_arguments_and_return:
1818 load_args;
1819 ctx->return_values = ctx->current_position;
1820 for (i_arg = 0; i_arg < n_ret; i_arg++) {
1821 #if ARG_MODE_N >= 3
1822 get_uint32(ctx);
1823 #else
1824 get_code(ctx);
1825 #endif
1826 get_code(ctx);
1827 }
1828 if (unlikely(profiling))
1829 goto unconditional_escape;
1830 if (code == OPCODE_CALL || code == OPCODE_CALL_STRICT) {
1831 g(gen_call(ctx, code, fn_idx));
1832 continue;
1833 }
1834 /*if (code == OPCODE_CALL_INDIRECT || code == OPCODE_CALL_INDIRECT_STRICT) {
1835 if (unlikely(!gen_call_indirect(ctx, code, slot_1, flags)))
1836 return false;
1837 continue;
1838 }*/
1839 goto unconditional_escape;
1840 }
1841 case OPCODE_CALL_INDIRECT:
1842 case OPCODE_CALL_INDIRECT_STRICT:
1843 case OPCODE_CALL_INDIRECT_SPARK:
1844 case OPCODE_CALL_INDIRECT_LAZY:
1845 case OPCODE_CALL_INDIRECT_CACHE:
1846 case OPCODE_CALL_INDIRECT_SAVE: {
1847 fn_idx = 0; /* avoid warning */
1848 get_two(ctx, &n_args, &n_ret);
1849 get_two(ctx, &slot_1, &flags);
1850 goto jump_over_arguments_and_return;
1851 }
1852 case OPCODE_RETURN: {
1853 n_args = da(ctx->fn,function)->n_return_values;
1854 load_args;
1855 if (unlikely(profiling))
1856 goto unconditional_escape;
1857 g(gen_return(ctx));
1858 continue;
1859 }
1860 case OPCODE_STRUCTURED: {
1861 init_args;
1862 get_two(ctx, &slot_1, &slot_2);
1863 do {
1864 struct code_arg a;
1865 get_two(ctx, &flags, &slot_r);
1866 get_one(ctx, &opt);
1867 a.slot = slot_r;
1868 a.flags = flags;
1869 a.type = opt;
1870 g(array_add_mayfail(struct code_arg, &ctx->args, &ctx->args_l, a, NULL, &ctx->err));
1871 } while (!(flags & OPCODE_STRUCTURED_FLAG_END));
1872 g(gen_structured(ctx, slot_1, slot_2));
1873 continue;
1874 }
1875 case OPCODE_RECORD_CREATE: {
1876 init_args;
1877 get_two(ctx, &slot_r, &n_args);
1878 for (i_arg = 0; i_arg < n_args; i_arg++) {
1879 struct code_arg a;
1880 get_two(ctx, &slot_1, &flags);
1881 a.slot = slot_1;
1882 a.flags = flags;
1883 a.type = 0;
1884 g(array_add_mayfail(struct code_arg, &ctx->args, &ctx->args_l, a, NULL, &ctx->err));
1885 }
1886 g(gen_record_create(ctx, slot_r));
1887 continue;
1888 }
1889 case OPCODE_RECORD_LOAD: {
1890 get_two(ctx, &slot_1, &opt);
1891 get_two(ctx, &slot_r, &flags);
1892 g(gen_record_load(ctx, slot_1, slot_r, opt, flags));
1893 continue;
1894 }
1895 case OPCODE_OPTION_CREATE_EMPTY_FLAT: {
1896 get_two(ctx, &slot_r, &opt);
1897 g(gen_option_create_empty_flat(ctx, opt, slot_r));
1898 continue;
1899 }
1900 case OPCODE_OPTION_CREATE_EMPTY: {
1901 get_two(ctx, &slot_r, &opt);
1902 g(gen_option_create_empty(ctx, opt, slot_r));
1903 continue;
1904 }
1905 case OPCODE_OPTION_CREATE: {
1906 get_two(ctx, &slot_r, &opt);
1907 get_two(ctx, &slot_1, &flags);
1908 g(gen_option_create(ctx, opt, slot_1, slot_r, flags));
1909 continue;
1910 }
1911 case OPCODE_OPTION_LOAD: {
1912 get_two(ctx, &slot_1, &opt);
1913 get_two(ctx, &slot_r, &flags);
1914 g(gen_option_load(ctx, slot_1, slot_r, opt, flags));
1915 continue;
1916 }
1917 case OPCODE_OPTION_TEST_FLAT: {
1918 get_two(ctx, &slot_1, &opt);
1919 get_one(ctx, &slot_r);
1920 g(gen_option_test_flat(ctx, slot_1, opt, slot_r));
1921 continue;
1922 }
1923 case OPCODE_OPTION_TEST: {
1924 get_two(ctx, &slot_1, &opt);
1925 get_one(ctx, &slot_r);
1926 g(gen_option_test(ctx, slot_1, opt, slot_r));
1927 continue;
1928 }
1929 case OPCODE_OPTION_ORD_FLAT: {
1930 get_two(ctx, &slot_1, &slot_r);
1931 g(gen_option_ord(ctx, slot_1, slot_r, true));
1932 continue;
1933 }
1934 case OPCODE_OPTION_ORD: {
1935 get_two(ctx, &slot_1, &slot_r);
1936 g(gen_option_ord(ctx, slot_1, slot_r, false));
1937 continue;
1938 }
1939 case OPCODE_ARRAY_CREATE: {
1940 init_args;
1941 get_two(ctx, &slot_r, &n_args);
1942 for (i_arg = 0; i_arg < n_args; i_arg++) {
1943 struct code_arg a;
1944 get_two(ctx, &slot_1, &flags);
1945 a.slot = slot_1;
1946 a.flags = flags;
1947 a.type = 0;
1948 g(array_add_mayfail(struct code_arg, &ctx->args, &ctx->args_l, a, NULL, &ctx->err));
1949 }
1950 g(gen_array_create(ctx, slot_r));
1951 continue;
1952 }
1953 case OPCODE_ARRAY_CREATE_EMPTY_FLAT: {
1954 get_two(ctx, &slot_r, &flags);
1955 g(gen_array_create_empty_flat(ctx, slot_r, flags));
1956 continue;
1957 }
1958 case OPCODE_ARRAY_CREATE_EMPTY: {
1959 get_one(ctx, &slot_r);
1960 g(gen_array_create_empty(ctx, slot_r));
1961 continue;
1962 }
1963 case OPCODE_ARRAY_FILL: {
1964 get_two(ctx, &slot_1, &flags);
1965 get_two(ctx, &slot_2, &slot_r);
1966 g(gen_array_fill(ctx, slot_1, flags, slot_2, slot_r));
1967 continue;
1968 }
1969 case OPCODE_ARRAY_STRING: {
1970 frame_t i;
1971 get_two(ctx, &slot_r, &i);
1972 g(gen_array_string(ctx, type_get_fixed(0, true)->tag, cast_ptr(uint8_t *, ctx->current_position), i, slot_r));
1973 ctx->current_position += (i + 1) >> 1;
1974 continue;
1975 }
1976 case OPCODE_ARRAY_UNICODE: {
1977 frame_t i;
1978 get_two(ctx, &slot_r, &i);
1979 g(gen_array_string(ctx, type_get_int(2)->tag, cast_ptr(uint8_t *, ctx->current_position), i, slot_r));
1980 ctx->current_position += i * 2;
1981 continue;
1982 }
1983 case OPCODE_ARRAY_LOAD: {
1984 get_two(ctx, &slot_1, &slot_2);
1985 get_two(ctx, &slot_r, &flags);
1986 g(gen_array_load(ctx, slot_1, slot_2, slot_r, flags));
1987 continue;
1988 }
1989 case OPCODE_ARRAY_LEN: {
1990 get_two(ctx, &slot_1, &slot_r);
1991 get_one(ctx, &flags);
1992 g(gen_array_len(ctx, slot_1, NO_FRAME_T, slot_r, false, 0));
1993 continue;
1994 }
1995 case OPCODE_ARRAY_LEN_GREATER_THAN: {
1996 get_two(ctx, &slot_1, &slot_2);
1997 get_two(ctx, &slot_r, &flags);
1998 escape_label = alloc_escape_label(ctx);
1999 if (unlikely(!escape_label))
2000 return false;
2001 if (flags & OPCODE_FLAG_FUSED) {
2002 g(gen_fused_binary(ctx, MODE_ARRAY_LEN_GT, 0, 0, escape_label, slot_1, slot_2, slot_r, &failed));
2003 if (unlikely(!failed))
2004 continue;
2005 }
2006 g(gen_array_len(ctx, slot_1, slot_2, slot_r, false, 0));
2007 continue;
2008 }
2009 case OPCODE_ARRAY_SUB: {
2010 get_two(ctx, &slot_1, &slot_2);
2011 get_two(ctx, &slot_3, &slot_r);
2012 get_one(ctx, &flags);
2013 g(gen_array_sub(ctx, slot_1, slot_2, slot_3, slot_r, flags));
2014 continue;
2015 }
2016 case OPCODE_ARRAY_SKIP: {
2017 get_two(ctx, &slot_1, &slot_2);
2018 get_two(ctx, &slot_r, &flags);
2019 g(gen_array_skip(ctx, slot_1, slot_2, slot_r, flags));
2020 continue;
2021 }
2022 case OPCODE_ARRAY_APPEND: {
2023 get_two(ctx, &slot_r, &flags);
2024 get_two(ctx, &slot_1, &slot_2);
2025 g(gen_array_append(ctx, slot_1, slot_2, slot_r, flags));
2026 continue;
2027 }
2028 case OPCODE_ARRAY_APPEND_ONE_FLAT: {
2029 get_two(ctx, &slot_r, &flags);
2030 get_two(ctx, &slot_1, &slot_2);
2031 g(gen_array_append_one_flat(ctx, slot_1, slot_2, slot_r, flags));
2032 continue;
2033 }
2034 case OPCODE_ARRAY_APPEND_ONE: {
2035 get_two(ctx, &slot_r, &flags);
2036 get_two(ctx, &slot_1, &slot_2);
2037 g(gen_array_append_one(ctx, slot_1, slot_2, slot_r, flags));
2038 continue;
2039 }
2040 case OPCODE_ARRAY_FLATTEN: {
2041 get_two(ctx, &slot_r, &flags);
2042 get_one(ctx, &slot_1);
2043 goto unconditional_escape;
2044 }
2045 case OPCODE_IO: {
2046 get_two(ctx, &flags, &slot_1);
2047 get_two(ctx, &slot_2, &slot_3);
2048 g(gen_io(ctx, flags, slot_1, slot_2, slot_3));
2049 continue;
2050 }
2051 case OPCODE_INTERNAL_FUNCTION:
2052 case OPCODE_EXIT_THREAD:
2053 case OPCODE_UNREACHABLE: {
2054 goto unconditional_escape;
2055 }
2056 default: {
2057 #if 1
2058 /*if (getenv("DUMP") && !strcmp(da(ctx->fn,function)->function_name, getenv("DUMP")))*/
2059 warning("gen_function: %s: unknown opcode %04x, %s", da(ctx->fn,function)->function_name, *ctx->instr_start, decode_opcode(*ctx->instr_start, false));
2060 #endif
2061 return false;
2062 }
2063 }
2064 }
2065
2066 return true;
2067 }
2068
2069 static bool attr_w gen_entries(struct codegen_context *ctx)
2070 {
2071 size_t i;
2072 for (i = 0; i < ctx->n_entries; i++) {
2073 struct cg_entry *ce = &ctx->entries[i];
2074 if (ce->entry_label) {
2075 gen_insn(INSN_ENTRY, 0, 0, 0);
2076 gen_four(i);
2077
2078 g(gen_test_variables(ctx, ce->variables, ce->n_variables, ce->nonflat_label));
2079
2080 gen_insn(INSN_JMP, 0, 0, 0);
2081 gen_four(ce->entry_label);
2082 }
2083 }
2084 return true;
2085 }
2086
2087 static bool attr_w gen_epilogues(struct codegen_context *ctx)
2088 {
2089 frame_t v;
2090 ip_t ip;
2091 uint32_t escape_label, nospill_label;
2092 escape_label = alloc_label(ctx);
2093 if (unlikely(!escape_label))
2094 return false;
2095 nospill_label = alloc_label(ctx);
2096 if (unlikely(!nospill_label))
2097 return false;
2098 #if defined(ARCH_PARISC)
2099 if (ctx->call_label) {
2100 gen_label(ctx->call_label);
2101 g(gen_call_millicode(ctx));
2102 }
2103 #endif
2104 if (ctx->reload_label) {
2105 gen_label(ctx->reload_label);
2106 g(gen_mov(ctx, i_size(OP_SIZE_ADDRESS), R_FRAME, R_RET0));
2107 g(gen_escape_arg(ctx, (ip_t)-1, nospill_label));
2108 }
2109 gen_label(ctx->escape_nospill_label);
2110 g(gen_escape_arg(ctx, 0, nospill_label));
2111 for (ip = 0; ip < da(ctx->fn,function)->code_size; ip++) {
2112 struct cg_exit *ce = ctx->code_exits[ip];
2113 if (ce && (ce->undo_label || ce->escape_label)) {
2114 if (ce->undo_label) {
2115 size_t i;
2116 gen_label(ce->undo_label);
2117 gen_insn(ce->undo_opcode, ce->undo_op_size, ce->undo_aux, ce->undo_writes_flags);
2118 for (i = 0; i < ce->undo_parameters_len; i++)
2119 gen_one(ce->undo_parameters[i]);
2120 }
2121 if (ce->escape_label) {
2122 gen_label(ce->escape_label);
2123 }
2124 g(gen_escape_arg(ctx, ip, escape_label));
2125 }
2126 }
2127 gen_label(escape_label);
2128 for (v = MIN_USEABLE_SLOT; v < function_n_variables(ctx->fn); v++) {
2129 if (slot_is_register(ctx, v)) {
2130 g(spill(ctx, v));
2131 }
2132 }
2133 gen_label(nospill_label);
2134 g(gen_escape(ctx));
2135 return true;
2136 }
2137
2138 static bool attr_w cgen_entry(struct codegen_context *ctx)
2139 {
2140 uint32_t entry_id = cget_four(ctx);
2141 ajla_assert_lo(entry_id < ctx->n_entries, (file_line, "cgen_entry: invalid entry %lx", (unsigned long)entry_id));
2142 ctx->entries[entry_id].entry_to_pos = ctx->mcode_size;
2143 return true;
2144 }
2145
2146 static bool attr_w cgen_label(struct codegen_context *ctx)
2147 {
2148 uint32_t label_id = cget_four(ctx);
2149 ctx->label_to_pos[label_id] = ctx->mcode_size;
2150 return true;
2151 }
2152
2153 static bool attr_w attr_unused cgen_trap(struct codegen_context *ctx, uint32_t label)
2154 {
2155 struct trap_record tr;
2156 tr.source_ip = ctx->mcode_size;
2157 tr.destination_ip = label;
2158 if (unlikely(!array_add_mayfail(struct trap_record, &ctx->trap_records, &ctx->trap_records_size, tr, NULL, &ctx->err)))
2159 return false;
2160 return true;
2161 }
2162
2163 static bool attr_w add_relocation(struct codegen_context *ctx, unsigned length, int offset, bool *known)
2164 {
2165 struct relocation rel;
2166 rel.label_id = cget_four(ctx);
2167 rel.length = length;
2168 rel.position = ctx->mcode_size;
2169 rel.jmp_instr = ctx->code_position - 8 - offset - ctx->code;
2170 if (unlikely(!array_add_mayfail(struct relocation, &ctx->reloc, &ctx->reloc_size, rel, NULL, &ctx->err)))
2171 return false;
2172 if (known)
2173 *known = ctx->label_to_pos[rel.label_id] != (size_t)-1;
2174 return true;
2175 }
2176
2177
2178 #if defined(ARCH_ALPHA)
2179 #include "c2-alpha.inc"
2180 #elif defined(ARCH_ARM32)
2181 #include "c2-arm.inc"
2182 #elif defined(ARCH_ARM64)
2183 #include "c2-arm64.inc"
2184 #elif defined(ARCH_IA64)
2185 #include "c2-ia64.inc"
2186 #elif defined(ARCH_LOONGARCH64)
2187 #include "c2-loong.inc"
2188 #elif defined(ARCH_MIPS)
2189 #include "c2-mips.inc"
2190 #elif defined(ARCH_PARISC)
2191 #include "c2-hppa.inc"
2192 #elif defined(ARCH_POWER)
2193 #include "c2-power.inc"
2194 #elif defined(ARCH_S390)
2195 #include "c2-s390.inc"
2196 #elif defined(ARCH_SPARC)
2197 #include "c2-sparc.inc"
2198 #elif defined(ARCH_RISCV64)
2199 #include "c2-riscv.inc"
2200 #elif defined(ARCH_X86)
2201 #include "c2-x86.inc"
2202 #endif
2203
2204
2205 static bool attr_w gen_mcode(struct codegen_context *ctx)
2206 {
2207 ctx->code_position = ctx->code;
2208
2209 while (ctx->code_position != ctx->code + ctx->code_size) {
2210 uint32_t insn;
2211 ajla_assert_lo(ctx->code_position < ctx->code + ctx->code_size, (file_line, "gen_mcode: ran out of code"));
2212 #ifdef DEBUG_INSNS
2213 insn = cget_four(ctx);
2214 debug("line: %u/%u", insn >> 24, insn & 0x00FFFFFFU);
2215 #endif
2216 insn = cget_four(ctx);
2217 g(cgen_insn(ctx, insn));
2218 }
2219
2220 return true;
2221 }
2222
2223 #define RELOCS_RETRY -1
2224 #define RELOCS_FAIL 0
2225 #define RELOCS_OK 1
2226
2227 static int8_t resolve_relocs(struct codegen_context *ctx)
2228 {
2229 size_t i;
2230 int8_t status = RELOCS_OK;
2231 for (i = 0; i < ctx->reloc_size; i++) {
2232 struct relocation *reloc = &ctx->reloc[i];
2233 if (!resolve_relocation(ctx, reloc)) {
2234 uint8_t *jmp_instr;
2235 uint32_t insn;
2236 uint32_t new_length;
2237 status = RELOCS_RETRY;
2238 if (unlikely(reloc->length + zero >= JMP_LIMIT))
2239 return RELOCS_FAIL;
2240 new_length = reloc->length + 1;
2241 jmp_instr = ctx->code + reloc->jmp_instr;
2242 insn = (uint32_t)jmp_instr[0] +
2243 ((uint32_t)jmp_instr[1] << 8) +
2244 ((uint32_t)jmp_instr[2] << 16) +
2245 ((uint32_t)jmp_instr[3] << 24);
2246 insn &= ~INSN_JUMP_SIZE;
2247 insn |= (uint32_t)new_length << INSN_JUMP_SIZE_SHIFT;
2248 jmp_instr[0] = insn;
2249 jmp_instr[1] = insn >> 8;
2250 jmp_instr[2] = insn >> 16;
2251 jmp_instr[3] = insn >> 24;
2252 }
2253 }
2254 return status;
2255 }
2256
2257 static void resolve_traps(struct codegen_context *ctx)
2258 {
2259 size_t i;
2260 for (i = 0; i < ctx->trap_records_size; i++) {
2261 struct trap_record *tr = &ctx->trap_records[i];
2262 tr->destination_ip = ctx->label_to_pos[tr->destination_ip];
2263 }
2264 }
2265
2266
2267 static bool attr_w codegen_map(struct codegen_context *ctx)
2268 {
2269 void *ptr;
2270 frame_t i;
2271 array_finish(uint8_t, &ctx->mcode, &ctx->mcode_size);
2272 ptr = os_code_map(ctx->mcode, ctx->mcode_size, &ctx->err);
2273 ctx->mcode = NULL;
2274 if (unlikely(!ptr)) {
2275 return false;
2276 }
2277 for (i = 0; i < ctx->n_entries; i++) {
2278 char *entry = cast_ptr(char *, ptr) + ctx->entries[i].entry_to_pos;
2279 da(ctx->codegen,codegen)->unoptimized_code[i] = entry;
2280 da(ctx->codegen,codegen)->n_entries++;
2281 }
2282 da(ctx->codegen,codegen)->unoptimized_code_base = ptr;
2283 da(ctx->codegen,codegen)->unoptimized_code_size = ctx->mcode_size;
2284
2285 return true;
2286 }
2287
2288
2289 void *codegen_fn(frame_s *fp, const code_t *ip, union internal_arg ia[])
2290 {
2291 struct codegen_context ctx_;
2292 struct codegen_context *ctx = &ctx_;
2293 frame_t i;
2294 int8_t rr;
2295 struct data *codegen;
2296 uint32_t l;
2297
2298 init_ctx(ctx);
2299 ctx->fn = ia[0].ptr;
2300
2301 #ifdef DEBUG_ENV
2302 if (getenv("CG") && strcmp(da(ctx->fn,function)->function_name, getenv("CG")))
2303 goto fail;
2304 #endif
2305
2306 ctx->local_directory = mem_alloc_array_mayfail(mem_calloc_mayfail, struct data **, 0, 0, da(ctx->fn,function)->local_directory_size, sizeof(struct data *), &ctx->err);
2307 if (unlikely(!ctx->local_directory))
2308 goto fail;
2309
2310 if (0) for (i = 0; i < da(ctx->fn,function)->local_directory_size; i++) {
2311 struct data *callee;
2312 pointer_t *ptr;
2313 ptr = da(ctx->fn,function)->local_directory[i];
2314 pointer_follow(ptr, false, callee, PF_SPARK, NULL, 0,
2315 SUBMIT_EX(ex_);
2316 goto next_one,
2317 goto next_one;
2318 );
2319 ctx->local_directory[i] = callee;
2320 next_one:;
2321 }
2322 for (i = 0; i < da(ctx->fn,function)->local_directory_size; i++) {
2323 struct data *callee;
2324 pointer_t *ptr;
2325 if (ctx->local_directory[i])
2326 continue;
2327 ptr = da(ctx->fn,function)->local_directory[i];
2328 pointer_follow(ptr, false, callee, PF_WAIT, fp, ip,
2329 done_ctx(ctx);
2330 return ex_,
2331 goto fail
2332 );
2333 ctx->local_directory[i] = callee;
2334 /*debug("processing call: %s -> %s", da(ctx->fn,function)->function_name, da(callee,function)->function_name);*/
2335 }
2336
2337 if (da(ctx->fn,function)->module_designator) {
2338 struct function_descriptor *sfd = save_find_function_descriptor(da(ctx->fn,function)->module_designator, da(ctx->fn,function)->function_designator);
2339 if (sfd && sfd->unoptimized_code_size) {
2340 codegen = data_alloc_flexible(codegen, unoptimized_code, sfd->n_entries, &ctx->err);
2341 if (unlikely(!codegen))
2342 goto fail;
2343 da(codegen,codegen)->unoptimized_code_base = sfd->unoptimized_code_base;
2344 da(codegen,codegen)->unoptimized_code_size = sfd->unoptimized_code_size;
2345 da(codegen,codegen)->function = ctx->fn;
2346 da(codegen,codegen)->is_saved = true;
2347 da(codegen,codegen)->n_entries = sfd->n_entries;
2348 da(codegen,codegen)->offsets = NULL;
2349 for (i = 0; i < sfd->n_entries; i++) {
2350 da(codegen,codegen)->unoptimized_code[i] = cast_ptr(char *, da(codegen,codegen)->unoptimized_code_base) + sfd->entries[i];
2351 /*debug("%s: %p + %lx -> %p", da(ctx->fn,function)->function_name, da(codegen,codegen)->unoptimized_code_base, sfd->entries[i], da(codegen,codegen)->unoptimized_code[i]);*/
2352 }
2353 #ifdef HAVE_CODEGEN_TRAPS
2354 da(codegen,codegen)->trap_records = sfd->trap_records;
2355 da(codegen,codegen)->trap_records_size = sfd->trap_records_size;
2356 data_trap_insert(codegen);
2357 #endif
2358 goto have_codegen;
2359 }
2360 }
2361
2362 /*debug("trying: %s", da(ctx->fn,function)->function_name);*/
2363 if (unlikely(!array_init_mayfail(uint8_t, &ctx->code, &ctx->code_size, &ctx->err)))
2364 goto fail;
2365
2366 ctx->code_labels = mem_alloc_array_mayfail(mem_calloc_mayfail, uint32_t *, 0, 0, da(ctx->fn,function)->code_size, sizeof(uint32_t), &ctx->err);
2367 if (unlikely(!ctx->code_labels))
2368 goto fail;
2369
2370 ctx->code_exits = mem_alloc_array_mayfail(mem_calloc_mayfail, struct cg_exit **, 0, 0, da(ctx->fn,function)->code_size, sizeof(struct cg_exit *), &ctx->err);
2371 if (unlikely(!ctx->code_exits))
2372 goto fail;
2373
2374 ctx->flag_cache = mem_alloc_array_mayfail(mem_calloc_mayfail, uint8_t *, 0, 0, function_n_variables(ctx->fn), sizeof(int8_t), &ctx->err);
2375 if (unlikely(!ctx->flag_cache))
2376 goto fail;
2377
2378 ctx->registers = mem_alloc_array_mayfail(mem_alloc_mayfail, short *, 0, 0, function_n_variables(ctx->fn), sizeof(short), &ctx->err);
2379 if (unlikely(!ctx->registers))
2380 goto fail;
2381
2382 if (unlikely(!array_init_mayfail(frame_t, &ctx->need_spill, &ctx->need_spill_l, &ctx->err)))
2383 goto fail;
2384
2385 if (unlikely(!gen_registers(ctx)))
2386 goto fail;
2387
2388 if (unlikely(!gen_function(ctx)))
2389 goto fail;
2390
2391 if (unlikely(!gen_entries(ctx)))
2392 goto fail;
2393
2394 if (unlikely(!gen_epilogues(ctx)))
2395 goto fail;
2396
2397 if (unlikely(!(ctx->label_id + 1)))
2398 goto fail;
2399 if (unlikely(!(ctx->label_to_pos = mem_alloc_array_mayfail(mem_alloc_mayfail, size_t *, 0, 0, ctx->label_id + 1, sizeof(size_t), &ctx->err))))
2400 goto fail;
2401
2402 again:
2403 for (l = 0; l < ctx->label_id + 1; l++)
2404 ctx->label_to_pos[l] = (size_t)-1;
2405
2406 if (unlikely(!array_init_mayfail(uint8_t, &ctx->mcode, &ctx->mcode_size, &ctx->err)))
2407 goto fail;
2408
2409 if (unlikely(!array_init_mayfail(struct relocation, &ctx->reloc, &ctx->reloc_size, &ctx->err)))
2410 goto fail;
2411
2412 if (unlikely(!array_init_mayfail(struct trap_record, &ctx->trap_records, &ctx->trap_records_size, &ctx->err)))
2413 goto fail;
2414
2415 #ifdef ARCH_CONTEXT
2416 init_arch_context(ctx);
2417 #endif
2418
2419 if (unlikely(!gen_mcode(ctx)))
2420 goto fail;
2421
2422 rr = resolve_relocs(ctx);
2423 if (unlikely(rr == RELOCS_FAIL)) {
2424 /*debug("relocation fail: %s", da(ctx->fn,function)->function_name);*/
2425 goto fail;
2426 }
2427 if (rr == RELOCS_RETRY) {
2428 mem_free(ctx->mcode);
2429 ctx->mcode = NULL;
2430 mem_free(ctx->reloc);
2431 ctx->reloc = NULL;
2432 mem_free(ctx->trap_records);
2433 ctx->trap_records = NULL;
2434 goto again;
2435 }
2436
2437 resolve_traps(ctx);
2438
2439 #ifdef DEBUG_ENV
2440 if ((getenv("DUMP") && !strcmp(getenv("DUMP"), da(ctx->fn,function)->function_name)) || getenv("DUMP_ALL")) {
2441 char *hex;
2442 size_t hexl;
2443 size_t i;
2444 handle_t h;
2445
2446 mutex_lock(&dump_mutex);
2447 str_init(&hex, &hexl);
2448 str_add_string(&hex, &hexl, "_");
2449 str_add_unsigned(&hex, &hexl, dump_seq++, 10);
2450 str_add_string(&hex, &hexl, "_");
2451 str_add_string(&hex, &hexl, da(ctx->fn,function)->function_name);
2452 str_add_string(&hex, &hexl, ":");
2453 for (i = 0; i < hexl; i++)
2454 if (hex[i] == '/')
2455 hex[i] = '_';
2456 for (i = 0; i < ctx->mcode_size; i++) {
2457 uint8_t a = ctx->mcode[i];
2458 if (!(i & 0xff))
2459 str_add_string(&hex, &hexl, "\n .byte 0x");
2460 else
2461 str_add_string(&hex, &hexl, ",0x");
2462 if (a < 16)
2463 str_add_char(&hex, &hexl, '0');
2464 str_add_unsigned(&hex, &hexl, a, 16);
2465 }
2466 str_add_string(&hex, &hexl, "\n");
2467 h = os_open(os_cwd, "dump.s", O_WRONLY | O_APPEND, 0600, NULL);
2468 os_write_all(h, hex, hexl, NULL);
2469 os_close(h);
2470 mem_free(hex);
2471 mutex_unlock(&dump_mutex);
2472 }
2473 #endif
2474
2475 ctx->codegen = data_alloc_flexible(codegen, unoptimized_code, ctx->n_entries, &ctx->err);
2476 if (unlikely(!ctx->codegen))
2477 goto fail;
2478 da(ctx->codegen,codegen)->function = ctx->fn;
2479 da(ctx->codegen,codegen)->is_saved = false;
2480 da(ctx->codegen,codegen)->n_entries = 0;
2481 da(ctx->codegen,codegen)->offsets = NULL;
2482
2483 if (unlikely(!codegen_map(ctx)))
2484 goto fail;
2485
2486 codegen = ctx->codegen;
2487 ctx->codegen = NULL;
2488
2489 #ifdef HAVE_CODEGEN_TRAPS
2490 da(codegen,codegen)->trap_records = ctx->trap_records;
2491 da(codegen,codegen)->trap_records_size = ctx->trap_records_size;
2492 ctx->trap_records = NULL;
2493 data_trap_insert(codegen);
2494 #endif
2495
2496 have_codegen:
2497 done_ctx(ctx);
2498 return function_return(fp, pointer_data(codegen));
2499
2500 fail:
2501 /*debug("FAILED: %s", da(ctx->fn,function)->function_name);*/
2502 done_ctx(ctx);
2503 return function_return(fp, pointer_thunk(thunk_alloc_exception_error(error_ajla(EC_SYNC, AJLA_ERROR_NOT_SUPPORTED), NULL, NULL, NULL pass_file_line)));
2504 }
2505
2506 void codegen_free(struct data *codegen)
2507 {
2508 if (unlikely(da(codegen,codegen)->offsets != NULL))
2509 mem_free(da(codegen,codegen)->offsets);
2510 if (likely(da(codegen,codegen)->is_saved))
2511 return;
2512 #ifdef HAVE_CODEGEN_TRAPS
2513 mem_free(da(codegen,codegen)->trap_records);
2514 #endif
2515 os_code_unmap(da(codegen,codegen)->unoptimized_code_base, da(codegen,codegen)->unoptimized_code_size);
2516 }
2517
2518 #if defined(ARCH_IA64)
2519 static uintptr_t ia64_stub[2];
2520 #endif
2521 #if defined(ARCH_PARISC32) && defined(ARCH_PARISC_USE_STUBS)
2522 static uintptr_t parisc_stub[2];
2523 #endif
2524 #if defined(ARCH_PARISC64) && defined(ARCH_PARISC_USE_STUBS)
2525 static uintptr_t parisc_stub[4];
2526 #endif
2527 #if defined(ARCH_POWER) && defined(AIX_CALL)
2528 static uintptr_t ppc_stub[3];
2529 #endif
2530
2531 void name(codegen_init)(void)
2532 {
2533 struct codegen_context ctx_;
2534 struct codegen_context *ctx = &ctx_;
2535 void *ptr;
2536
2537 #if (defined(ARCH_X86_64) || defined(ARCH_X86_X32)) && !defined(ARCH_X86_WIN_ABI)
2538 #if defined(HAVE_SYSCALL) && defined(HAVE_ASM_PRCTL_H) && defined(HAVE_SYS_SYSCALL_H)
2539 if (!dll) {
2540 int r;
2541 EINTR_LOOP(r, syscall(SYS_arch_prctl, ARCH_SET_GS, &cg_upcall_vector));
2542 if (!r)
2543 upcall_register = R_GS;
2544 }
2545 #elif defined(HAVE_AMD64_SET_GSBASE) && defined(HAVE_X86_SYSARCH_H)
2546 if (!dll) {
2547 int r;
2548 EINTR_LOOP(r, amd64_set_gsbase(&cg_upcall_vector));
2549 if (!r)
2550 upcall_register = R_GS;
2551 }
2552 #elif defined(HAVE_SYSARCH) && defined(HAVE_X86_SYSARCH_H) && defined(X86_64_SET_GSBASE)
2553 if (!dll) {
2554 int r;
2555 void *ptr = &cg_upcall_vector;
2556 EINTR_LOOP(r, sysarch(X86_64_SET_GSBASE, &ptr));
2557 if (!r)
2558 upcall_register = R_GS;
2559 }
2560 #endif
2561 #endif
2562
2563 init_ctx(ctx);
2564 ctx->fn = NULL;
2565
2566 array_init(uint8_t, &ctx->code, &ctx->code_size);
2567
2568 if (unlikely(!gen_entry(ctx)))
2569 goto fail;
2570
2571 array_init(uint8_t, &ctx->mcode, &ctx->mcode_size);
2572
2573 #ifdef ARCH_CONTEXT
2574 init_arch_context(ctx);
2575 #endif
2576
2577 if (unlikely(!gen_mcode(ctx)))
2578 goto fail;
2579
2580 array_finish(uint8_t, &ctx->mcode, &ctx->mcode_size);
2581 ptr = os_code_map(ctx->mcode, ctx->mcode_size, NULL);
2582 codegen_ptr = ptr;
2583 codegen_size = ctx->mcode_size;
2584 ctx->mcode = NULL;
2585 #if defined(ARCH_IA64)
2586 ia64_stub[0] = ptr_to_num(ptr);
2587 ia64_stub[1] = 0;
2588 codegen_entry = cast_ptr(codegen_type, ia64_stub);
2589 #elif defined(ARCH_PARISC32) && defined(ARCH_PARISC_USE_STUBS)
2590 parisc_stub[0] = ptr_to_num(ptr);
2591 parisc_stub[1] = 0;
2592 codegen_entry = cast_ptr(codegen_type, cast_ptr(char *, parisc_stub) + 2);
2593 #elif defined(ARCH_PARISC64) && defined(ARCH_PARISC_USE_STUBS)
2594 parisc_stub[0] = 0;
2595 parisc_stub[1] = 0;
2596 parisc_stub[2] = ptr_to_num(ptr);
2597 parisc_stub[3] = 0;
2598 codegen_entry = cast_ptr(codegen_type, parisc_stub);
2599 #elif defined(ARCH_POWER) && defined(AIX_CALL)
2600 ppc_stub[0] = ptr_to_num(ptr);
2601 ppc_stub[1] = 0;
2602 ppc_stub[2] = 0;
2603 codegen_entry = cast_ptr(codegen_type, ppc_stub);
2604 #else
2605 codegen_entry = ptr;
2606 #endif
2607 done_ctx(ctx);
2608
2609 #ifdef DEBUG_ENV
2610 mutex_init(&dump_mutex);
2611 if (getenv("DUMP") || getenv("DUMP_ALL")) {
2612 size_t i;
2613 char *hex;
2614 size_t hexl;
2615 str_init(&hex, &hexl);
2616 #if defined(ARCH_RISCV64)
2617 str_add_string(&hex, &hexl, " .attribute arch, \"rv64i2p1_m2p0_a2p1_f2p2_d2p2_c2p0_zicsr2p0_zifencei2p0_zba1p0_zbb1p0_zbc1p0_zbs1p0\"\n");
2618 #endif
2619 for (i = 0; i < codegen_size; i++) {
2620 uint8_t a = cast_ptr(uint8_t *, codegen_ptr)[i];
2621 str_add_string(&hex, &hexl, " .byte 0x");
2622 if (a < 16)
2623 str_add_char(&hex, &hexl, '0');
2624 str_add_unsigned(&hex, &hexl, a, 16);
2625 str_add_char(&hex, &hexl, '\n');
2626 }
2627 os_write_atomic(".", "dump.s", hex, hexl, NULL);
2628 mem_free(hex);
2629 }
2630 #endif
2631
2632 return;
2633
2634 fail:
2635 fatal("couldn't compile global entry");
2636 }
2637
2638 void name(codegen_done)(void)
2639 {
2640 os_code_unmap(codegen_ptr, codegen_size);
2641 #ifdef DEBUG_ENV
2642 mutex_done(&dump_mutex);
2643 #endif
2644 }
2645
2646 #else
2647
2648 void name(codegen_init)(void)
2649 {
2650 }
2651
2652 void name(codegen_done)(void)
2653 {
2654 }
2655
2656 #endif
2657
2658 #endif
Ajla - a purely functional programming language