?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
migration: simplify exec migration functions
[qemu/armbru.git] / tcg / tcg.c
blobd6670237fb7c69f7f3ee3dfc7b47abc85cc3d173
1 /*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25 #include "qemu/osdep.h"
26
27 /* Define to jump the ELF file used to communicate with GDB. */
28 #undef DEBUG_JIT
29
30 #include "qemu/error-report.h"
31 #include "qemu/cutils.h"
32 #include "qemu/host-utils.h"
33 #include "qemu/qemu-print.h"
34 #include "qemu/cacheflush.h"
35 #include "qemu/cacheinfo.h"
36 #include "qemu/timer.h"
37 #include "exec/translation-block.h"
38 #include "exec/tlb-common.h"
39 #include "tcg/startup.h"
40 #include "tcg/tcg-op-common.h"
41
42 #if UINTPTR_MAX == UINT32_MAX
43 # define ELF_CLASS ELFCLASS32
44 #else
45 # define ELF_CLASS ELFCLASS64
46 #endif
47 #if HOST_BIG_ENDIAN
48 # define ELF_DATA ELFDATA2MSB
49 #else
50 # define ELF_DATA ELFDATA2LSB
51 #endif
52
53 #include "elf.h"
54 #include "exec/log.h"
55 #include "tcg/tcg-ldst.h"
56 #include "tcg/tcg-temp-internal.h"
57 #include "tcg-internal.h"
58 #include "tcg/perf.h"
59 #ifdef CONFIG_USER_ONLY
60 #include "exec/user/guest-base.h"
61 #endif
62
63 /* Forward declarations for functions declared in tcg-target.c.inc and
64 used here. */
65 static void tcg_target_init(TCGContext *s);
66 static void tcg_target_qemu_prologue(TCGContext *s);
67 static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
68 intptr_t value, intptr_t addend);
69
70 /* The CIE and FDE header definitions will be common to all hosts. */
71 typedef struct {
72 uint32_t len __attribute__((aligned((sizeof(void *)))));
73 uint32_t id;
74 uint8_t version;
75 char augmentation[1];
76 uint8_t code_align;
77 uint8_t data_align;
78 uint8_t return_column;
79 } DebugFrameCIE;
80
81 typedef struct QEMU_PACKED {
82 uint32_t len __attribute__((aligned((sizeof(void *)))));
83 uint32_t cie_offset;
84 uintptr_t func_start;
85 uintptr_t func_len;
86 } DebugFrameFDEHeader;
87
88 typedef struct QEMU_PACKED {
89 DebugFrameCIE cie;
90 DebugFrameFDEHeader fde;
91 } DebugFrameHeader;
92
93 typedef struct TCGLabelQemuLdst {
94 bool is_ld; /* qemu_ld: true, qemu_st: false */
95 MemOpIdx oi;
96 TCGType type; /* result type of a load */
97 TCGReg addrlo_reg; /* reg index for low word of guest virtual addr */
98 TCGReg addrhi_reg; /* reg index for high word of guest virtual addr */
99 TCGReg datalo_reg; /* reg index for low word to be loaded or stored */
100 TCGReg datahi_reg; /* reg index for high word to be loaded or stored */
101 const tcg_insn_unit *raddr; /* addr of the next IR of qemu_ld/st IR */
102 tcg_insn_unit *label_ptr[2]; /* label pointers to be updated */
103 QSIMPLEQ_ENTRY(TCGLabelQemuLdst) next;
104 } TCGLabelQemuLdst;
105
106 static void tcg_register_jit_int(const void *buf, size_t size,
107 const void *debug_frame,
108 size_t debug_frame_size)
109 __attribute__((unused));
110
111 /* Forward declarations for functions declared and used in tcg-target.c.inc. */
112 static void tcg_out_tb_start(TCGContext *s);
113 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1,
114 intptr_t arg2);
115 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg);
116 static void tcg_out_movi(TCGContext *s, TCGType type,
117 TCGReg ret, tcg_target_long arg);
118 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg);
119 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg);
120 static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg);
121 static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg);
122 static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg);
123 static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg);
124 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg);
125 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg);
126 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg);
127 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long);
128 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2);
129 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg);
130 static void tcg_out_goto_tb(TCGContext *s, int which);
131 static void tcg_out_op(TCGContext *s, TCGOpcode opc,
132 const TCGArg args[TCG_MAX_OP_ARGS],
133 const int const_args[TCG_MAX_OP_ARGS]);
134 #if TCG_TARGET_MAYBE_vec
135 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
136 TCGReg dst, TCGReg src);
137 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
138 TCGReg dst, TCGReg base, intptr_t offset);
139 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
140 TCGReg dst, int64_t arg);
141 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
142 unsigned vecl, unsigned vece,
143 const TCGArg args[TCG_MAX_OP_ARGS],
144 const int const_args[TCG_MAX_OP_ARGS]);
145 #else
146 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
147 TCGReg dst, TCGReg src)
148 {
149 g_assert_not_reached();
150 }
151 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
152 TCGReg dst, TCGReg base, intptr_t offset)
153 {
154 g_assert_not_reached();
155 }
156 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
157 TCGReg dst, int64_t arg)
158 {
159 g_assert_not_reached();
160 }
161 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
162 unsigned vecl, unsigned vece,
163 const TCGArg args[TCG_MAX_OP_ARGS],
164 const int const_args[TCG_MAX_OP_ARGS])
165 {
166 g_assert_not_reached();
167 }
168 #endif
169 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1,
170 intptr_t arg2);
171 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
172 TCGReg base, intptr_t ofs);
173 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target,
174 const TCGHelperInfo *info);
175 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot);
176 static bool tcg_target_const_match(int64_t val, int ct,
177 TCGType type, TCGCond cond, int vece);
178 #ifdef TCG_TARGET_NEED_LDST_LABELS
179 static int tcg_out_ldst_finalize(TCGContext *s);
180 #endif
181
182 #ifndef CONFIG_USER_ONLY
183 #define guest_base ({ qemu_build_not_reached(); (uintptr_t)0; })
184 #endif
185
186 typedef struct TCGLdstHelperParam {
187 TCGReg (*ra_gen)(TCGContext *s, const TCGLabelQemuLdst *l, int arg_reg);
188 unsigned ntmp;
189 int tmp[3];
190 } TCGLdstHelperParam;
191
192 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *l,
193 const TCGLdstHelperParam *p)
194 __attribute__((unused));
195 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *l,
196 bool load_sign, const TCGLdstHelperParam *p)
197 __attribute__((unused));
198 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *l,
199 const TCGLdstHelperParam *p)
200 __attribute__((unused));
201
202 static void * const qemu_ld_helpers[MO_SSIZE + 1] __attribute__((unused)) = {
203 [MO_UB] = helper_ldub_mmu,
204 [MO_SB] = helper_ldsb_mmu,
205 [MO_UW] = helper_lduw_mmu,
206 [MO_SW] = helper_ldsw_mmu,
207 [MO_UL] = helper_ldul_mmu,
208 [MO_UQ] = helper_ldq_mmu,
209 #if TCG_TARGET_REG_BITS == 64
210 [MO_SL] = helper_ldsl_mmu,
211 [MO_128] = helper_ld16_mmu,
212 #endif
213 };
214
215 static void * const qemu_st_helpers[MO_SIZE + 1] __attribute__((unused)) = {
216 [MO_8] = helper_stb_mmu,
217 [MO_16] = helper_stw_mmu,
218 [MO_32] = helper_stl_mmu,
219 [MO_64] = helper_stq_mmu,
220 #if TCG_TARGET_REG_BITS == 64
221 [MO_128] = helper_st16_mmu,
222 #endif
223 };
224
225 typedef struct {
226 MemOp atom; /* lg2 bits of atomicity required */
227 MemOp align; /* lg2 bits of alignment to use */
228 } TCGAtomAlign;
229
230 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc,
231 MemOp host_atom, bool allow_two_ops)
232 __attribute__((unused));
233
234 #ifdef CONFIG_USER_ONLY
235 bool tcg_use_softmmu;
236 #endif
237
238 TCGContext tcg_init_ctx;
239 __thread TCGContext *tcg_ctx;
240
241 TCGContext **tcg_ctxs;
242 unsigned int tcg_cur_ctxs;
243 unsigned int tcg_max_ctxs;
244 TCGv_env tcg_env;
245 const void *tcg_code_gen_epilogue;
246 uintptr_t tcg_splitwx_diff;
247
248 #ifndef CONFIG_TCG_INTERPRETER
249 tcg_prologue_fn *tcg_qemu_tb_exec;
250 #endif
251
252 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT];
253 static TCGRegSet tcg_target_call_clobber_regs;
254
255 #if TCG_TARGET_INSN_UNIT_SIZE == 1
256 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v)
257 {
258 *s->code_ptr++ = v;
259 }
260
261 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p,
262 uint8_t v)
263 {
264 *p = v;
265 }
266 #endif
267
268 #if TCG_TARGET_INSN_UNIT_SIZE <= 2
269 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v)
270 {
271 if (TCG_TARGET_INSN_UNIT_SIZE == 2) {
272 *s->code_ptr++ = v;
273 } else {
274 tcg_insn_unit *p = s->code_ptr;
275 memcpy(p, &v, sizeof(v));
276 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE);
277 }
278 }
279
280 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p,
281 uint16_t v)
282 {
283 if (TCG_TARGET_INSN_UNIT_SIZE == 2) {
284 *p = v;
285 } else {
286 memcpy(p, &v, sizeof(v));
287 }
288 }
289 #endif
290
291 #if TCG_TARGET_INSN_UNIT_SIZE <= 4
292 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v)
293 {
294 if (TCG_TARGET_INSN_UNIT_SIZE == 4) {
295 *s->code_ptr++ = v;
296 } else {
297 tcg_insn_unit *p = s->code_ptr;
298 memcpy(p, &v, sizeof(v));
299 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE);
300 }
301 }
302
303 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p,
304 uint32_t v)
305 {
306 if (TCG_TARGET_INSN_UNIT_SIZE == 4) {
307 *p = v;
308 } else {
309 memcpy(p, &v, sizeof(v));
310 }
311 }
312 #endif
313
314 #if TCG_TARGET_INSN_UNIT_SIZE <= 8
315 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v)
316 {
317 if (TCG_TARGET_INSN_UNIT_SIZE == 8) {
318 *s->code_ptr++ = v;
319 } else {
320 tcg_insn_unit *p = s->code_ptr;
321 memcpy(p, &v, sizeof(v));
322 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE);
323 }
324 }
325
326 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p,
327 uint64_t v)
328 {
329 if (TCG_TARGET_INSN_UNIT_SIZE == 8) {
330 *p = v;
331 } else {
332 memcpy(p, &v, sizeof(v));
333 }
334 }
335 #endif
336
337 /* label relocation processing */
338
339 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type,
340 TCGLabel *l, intptr_t addend)
341 {
342 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation));
343
344 r->type = type;
345 r->ptr = code_ptr;
346 r->addend = addend;
347 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next);
348 }
349
350 static void tcg_out_label(TCGContext *s, TCGLabel *l)
351 {
352 tcg_debug_assert(!l->has_value);
353 l->has_value = 1;
354 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr);
355 }
356
357 TCGLabel *gen_new_label(void)
358 {
359 TCGContext *s = tcg_ctx;
360 TCGLabel *l = tcg_malloc(sizeof(TCGLabel));
361
362 memset(l, 0, sizeof(TCGLabel));
363 l->id = s->nb_labels++;
364 QSIMPLEQ_INIT(&l->branches);
365 QSIMPLEQ_INIT(&l->relocs);
366
367 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next);
368
369 return l;
370 }
371
372 static bool tcg_resolve_relocs(TCGContext *s)
373 {
374 TCGLabel *l;
375
376 QSIMPLEQ_FOREACH(l, &s->labels, next) {
377 TCGRelocation *r;
378 uintptr_t value = l->u.value;
379
380 QSIMPLEQ_FOREACH(r, &l->relocs, next) {
381 if (!patch_reloc(r->ptr, r->type, value, r->addend)) {
382 return false;
383 }
384 }
385 }
386 return true;
387 }
388
389 static void set_jmp_reset_offset(TCGContext *s, int which)
390 {
391 /*
392 * We will check for overflow at the end of the opcode loop in
393 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX.
394 */
395 s->gen_tb->jmp_reset_offset[which] = tcg_current_code_size(s);
396 }
397
398 static void G_GNUC_UNUSED set_jmp_insn_offset(TCGContext *s, int which)
399 {
400 /*
401 * We will check for overflow at the end of the opcode loop in
402 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX.
403 */
404 s->gen_tb->jmp_insn_offset[which] = tcg_current_code_size(s);
405 }
406
407 static uintptr_t G_GNUC_UNUSED get_jmp_target_addr(TCGContext *s, int which)
408 {
409 /*
410 * Return the read-execute version of the pointer, for the benefit
411 * of any pc-relative addressing mode.
412 */
413 return (uintptr_t)tcg_splitwx_to_rx(&s->gen_tb->jmp_target_addr[which]);
414 }
415
416 static int __attribute__((unused))
417 tlb_mask_table_ofs(TCGContext *s, int which)
418 {
419 return (offsetof(CPUNegativeOffsetState, tlb.f[which]) -
420 sizeof(CPUNegativeOffsetState));
421 }
422
423 /* Signal overflow, starting over with fewer guest insns. */
424 static G_NORETURN
425 void tcg_raise_tb_overflow(TCGContext *s)
426 {
427 siglongjmp(s->jmp_trans, -2);
428 }
429
430 /*
431 * Used by tcg_out_movext{1,2} to hold the arguments for tcg_out_movext.
432 * By the time we arrive at tcg_out_movext1, @dst is always a TCGReg.
433 *
434 * However, tcg_out_helper_load_slots reuses this field to hold an
435 * argument slot number (which may designate a argument register or an
436 * argument stack slot), converting to TCGReg once all arguments that
437 * are destined for the stack are processed.
438 */
439 typedef struct TCGMovExtend {
440 unsigned dst;
441 TCGReg src;
442 TCGType dst_type;
443 TCGType src_type;
444 MemOp src_ext;
445 } TCGMovExtend;
446
447 /**
448 * tcg_out_movext -- move and extend
449 * @s: tcg context
450 * @dst_type: integral type for destination
451 * @dst: destination register
452 * @src_type: integral type for source
453 * @src_ext: extension to apply to source
454 * @src: source register
455 *
456 * Move or extend @src into @dst, depending on @src_ext and the types.
457 */
458 static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst,
459 TCGType src_type, MemOp src_ext, TCGReg src)
460 {
461 switch (src_ext) {
462 case MO_UB:
463 tcg_out_ext8u(s, dst, src);
464 break;
465 case MO_SB:
466 tcg_out_ext8s(s, dst_type, dst, src);
467 break;
468 case MO_UW:
469 tcg_out_ext16u(s, dst, src);
470 break;
471 case MO_SW:
472 tcg_out_ext16s(s, dst_type, dst, src);
473 break;
474 case MO_UL:
475 case MO_SL:
476 if (dst_type == TCG_TYPE_I32) {
477 if (src_type == TCG_TYPE_I32) {
478 tcg_out_mov(s, TCG_TYPE_I32, dst, src);
479 } else {
480 tcg_out_extrl_i64_i32(s, dst, src);
481 }
482 } else if (src_type == TCG_TYPE_I32) {
483 if (src_ext & MO_SIGN) {
484 tcg_out_exts_i32_i64(s, dst, src);
485 } else {
486 tcg_out_extu_i32_i64(s, dst, src);
487 }
488 } else {
489 if (src_ext & MO_SIGN) {
490 tcg_out_ext32s(s, dst, src);
491 } else {
492 tcg_out_ext32u(s, dst, src);
493 }
494 }
495 break;
496 case MO_UQ:
497 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
498 if (dst_type == TCG_TYPE_I32) {
499 tcg_out_extrl_i64_i32(s, dst, src);
500 } else {
501 tcg_out_mov(s, TCG_TYPE_I64, dst, src);
502 }
503 break;
504 default:
505 g_assert_not_reached();
506 }
507 }
508
509 /* Minor variations on a theme, using a structure. */
510 static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i,
511 TCGReg src)
512 {
513 tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src);
514 }
515
516 static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i)
517 {
518 tcg_out_movext1_new_src(s, i, i->src);
519 }
520
521 /**
522 * tcg_out_movext2 -- move and extend two pair
523 * @s: tcg context
524 * @i1: first move description
525 * @i2: second move description
526 * @scratch: temporary register, or -1 for none
527 *
528 * As tcg_out_movext, for both @i1 and @i2, caring for overlap
529 * between the sources and destinations.
530 */
531
532 static void tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1,
533 const TCGMovExtend *i2, int scratch)
534 {
535 TCGReg src1 = i1->src;
536 TCGReg src2 = i2->src;
537
538 if (i1->dst != src2) {
539 tcg_out_movext1(s, i1);
540 tcg_out_movext1(s, i2);
541 return;
542 }
543 if (i2->dst == src1) {
544 TCGType src1_type = i1->src_type;
545 TCGType src2_type = i2->src_type;
546
547 if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) {
548 /* The data is now in the correct registers, now extend. */
549 src1 = i2->src;
550 src2 = i1->src;
551 } else {
552 tcg_debug_assert(scratch >= 0);
553 tcg_out_mov(s, src1_type, scratch, src1);
554 src1 = scratch;
555 }
556 }
557 tcg_out_movext1_new_src(s, i2, src2);
558 tcg_out_movext1_new_src(s, i1, src1);
559 }
560
561 /**
562 * tcg_out_movext3 -- move and extend three pair
563 * @s: tcg context
564 * @i1: first move description
565 * @i2: second move description
566 * @i3: third move description
567 * @scratch: temporary register, or -1 for none
568 *
569 * As tcg_out_movext, for all of @i1, @i2 and @i3, caring for overlap
570 * between the sources and destinations.
571 */
572
573 static void tcg_out_movext3(TCGContext *s, const TCGMovExtend *i1,
574 const TCGMovExtend *i2, const TCGMovExtend *i3,
575 int scratch)
576 {
577 TCGReg src1 = i1->src;
578 TCGReg src2 = i2->src;
579 TCGReg src3 = i3->src;
580
581 if (i1->dst != src2 && i1->dst != src3) {
582 tcg_out_movext1(s, i1);
583 tcg_out_movext2(s, i2, i3, scratch);
584 return;
585 }
586 if (i2->dst != src1 && i2->dst != src3) {
587 tcg_out_movext1(s, i2);
588 tcg_out_movext2(s, i1, i3, scratch);
589 return;
590 }
591 if (i3->dst != src1 && i3->dst != src2) {
592 tcg_out_movext1(s, i3);
593 tcg_out_movext2(s, i1, i2, scratch);
594 return;
595 }
596
597 /*
598 * There is a cycle. Since there are only 3 nodes, the cycle is
599 * either "clockwise" or "anti-clockwise", and can be solved with
600 * a single scratch or two xchg.
601 */
602 if (i1->dst == src2 && i2->dst == src3 && i3->dst == src1) {
603 /* "Clockwise" */
604 if (tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2)) {
605 tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3);
606 /* The data is now in the correct registers, now extend. */
607 tcg_out_movext1_new_src(s, i1, i1->dst);
608 tcg_out_movext1_new_src(s, i2, i2->dst);
609 tcg_out_movext1_new_src(s, i3, i3->dst);
610 } else {
611 tcg_debug_assert(scratch >= 0);
612 tcg_out_mov(s, i1->src_type, scratch, src1);
613 tcg_out_movext1(s, i3);
614 tcg_out_movext1(s, i2);
615 tcg_out_movext1_new_src(s, i1, scratch);
616 }
617 } else if (i1->dst == src3 && i2->dst == src1 && i3->dst == src2) {
618 /* "Anti-clockwise" */
619 if (tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3)) {
620 tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2);
621 /* The data is now in the correct registers, now extend. */
622 tcg_out_movext1_new_src(s, i1, i1->dst);
623 tcg_out_movext1_new_src(s, i2, i2->dst);
624 tcg_out_movext1_new_src(s, i3, i3->dst);
625 } else {
626 tcg_debug_assert(scratch >= 0);
627 tcg_out_mov(s, i1->src_type, scratch, src1);
628 tcg_out_movext1(s, i2);
629 tcg_out_movext1(s, i3);
630 tcg_out_movext1_new_src(s, i1, scratch);
631 }
632 } else {
633 g_assert_not_reached();
634 }
635 }
636
637 #define C_PFX1(P, A) P##A
638 #define C_PFX2(P, A, B) P##A##_##B
639 #define C_PFX3(P, A, B, C) P##A##_##B##_##C
640 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D
641 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E
642 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F
643
644 /* Define an enumeration for the various combinations. */
645
646 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1),
647 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2),
648 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3),
649 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4),
650
651 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1),
652 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2),
653 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3),
654 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4),
655
656 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2),
657 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1),
658 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1),
659
660 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1),
661 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2),
662 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3),
663 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4),
664 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4),
665
666 typedef enum {
667 #include "tcg-target-con-set.h"
668 } TCGConstraintSetIndex;
669
670 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode);
671
672 #undef C_O0_I1
673 #undef C_O0_I2
674 #undef C_O0_I3
675 #undef C_O0_I4
676 #undef C_O1_I1
677 #undef C_O1_I2
678 #undef C_O1_I3
679 #undef C_O1_I4
680 #undef C_N1_I2
681 #undef C_N1O1_I1
682 #undef C_N2_I1
683 #undef C_O2_I1
684 #undef C_O2_I2
685 #undef C_O2_I3
686 #undef C_O2_I4
687 #undef C_N1_O1_I4
688
689 /* Put all of the constraint sets into an array, indexed by the enum. */
690
691 #define C_O0_I1(I1) { .args_ct_str = { #I1 } },
692 #define C_O0_I2(I1, I2) { .args_ct_str = { #I1, #I2 } },
693 #define C_O0_I3(I1, I2, I3) { .args_ct_str = { #I1, #I2, #I3 } },
694 #define C_O0_I4(I1, I2, I3, I4) { .args_ct_str = { #I1, #I2, #I3, #I4 } },
695
696 #define C_O1_I1(O1, I1) { .args_ct_str = { #O1, #I1 } },
697 #define C_O1_I2(O1, I1, I2) { .args_ct_str = { #O1, #I1, #I2 } },
698 #define C_O1_I3(O1, I1, I2, I3) { .args_ct_str = { #O1, #I1, #I2, #I3 } },
699 #define C_O1_I4(O1, I1, I2, I3, I4) { .args_ct_str = { #O1, #I1, #I2, #I3, #I4 } },
700
701 #define C_N1_I2(O1, I1, I2) { .args_ct_str = { "&" #O1, #I1, #I2 } },
702 #define C_N1O1_I1(O1, O2, I1) { .args_ct_str = { "&" #O1, #O2, #I1 } },
703 #define C_N2_I1(O1, O2, I1) { .args_ct_str = { "&" #O1, "&" #O2, #I1 } },
704
705 #define C_O2_I1(O1, O2, I1) { .args_ct_str = { #O1, #O2, #I1 } },
706 #define C_O2_I2(O1, O2, I1, I2) { .args_ct_str = { #O1, #O2, #I1, #I2 } },
707 #define C_O2_I3(O1, O2, I1, I2, I3) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3 } },
708 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3, #I4 } },
709 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { "&" #O1, #O2, #I1, #I2, #I3, #I4 } },
710
711 static const TCGTargetOpDef constraint_sets[] = {
712 #include "tcg-target-con-set.h"
713 };
714
715
716 #undef C_O0_I1
717 #undef C_O0_I2
718 #undef C_O0_I3
719 #undef C_O0_I4
720 #undef C_O1_I1
721 #undef C_O1_I2
722 #undef C_O1_I3
723 #undef C_O1_I4
724 #undef C_N1_I2
725 #undef C_N1O1_I1
726 #undef C_N2_I1
727 #undef C_O2_I1
728 #undef C_O2_I2
729 #undef C_O2_I3
730 #undef C_O2_I4
731 #undef C_N1_O1_I4
732
733 /* Expand the enumerator to be returned from tcg_target_op_def(). */
734
735 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1)
736 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2)
737 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3)
738 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4)
739
740 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1)
741 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2)
742 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3)
743 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4)
744
745 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2)
746 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1)
747 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1)
748
749 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1)
750 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2)
751 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3)
752 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4)
753 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4)
754
755 #include "tcg-target.c.inc"
756
757 #ifndef CONFIG_TCG_INTERPRETER
758 /* Validate CPUTLBDescFast placement. */
759 QEMU_BUILD_BUG_ON((int)(offsetof(CPUNegativeOffsetState, tlb.f[0]) -
760 sizeof(CPUNegativeOffsetState))
761 < MIN_TLB_MASK_TABLE_OFS);
762 #endif
763
764 static void alloc_tcg_plugin_context(TCGContext *s)
765 {
766 #ifdef CONFIG_PLUGIN
767 s->plugin_tb = g_new0(struct qemu_plugin_tb, 1);
768 s->plugin_tb->insns =
769 g_ptr_array_new_with_free_func(qemu_plugin_insn_cleanup_fn);
770 #endif
771 }
772
773 /*
774 * All TCG threads except the parent (i.e. the one that called tcg_context_init
775 * and registered the target's TCG globals) must register with this function
776 * before initiating translation.
777 *
778 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation
779 * of tcg_region_init() for the reasoning behind this.
780 *
781 * In system-mode each caller registers its context in tcg_ctxs[]. Note that in
782 * system-mode tcg_ctxs[] does not track tcg_ctx_init, since the initial context
783 * is not used anymore for translation once this function is called.
784 *
785 * Not tracking tcg_init_ctx in tcg_ctxs[] in system-mode keeps code that
786 * iterates over the array (e.g. tcg_code_size() the same for both system/user
787 * modes.
788 */
789 #ifdef CONFIG_USER_ONLY
790 void tcg_register_thread(void)
791 {
792 tcg_ctx = &tcg_init_ctx;
793 }
794 #else
795 void tcg_register_thread(void)
796 {
797 TCGContext *s = g_malloc(sizeof(*s));
798 unsigned int i, n;
799
800 *s = tcg_init_ctx;
801
802 /* Relink mem_base. */
803 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) {
804 if (tcg_init_ctx.temps[i].mem_base) {
805 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps;
806 tcg_debug_assert(b >= 0 && b < n);
807 s->temps[i].mem_base = &s->temps[b];
808 }
809 }
810
811 /* Claim an entry in tcg_ctxs */
812 n = qatomic_fetch_inc(&tcg_cur_ctxs);
813 g_assert(n < tcg_max_ctxs);
814 qatomic_set(&tcg_ctxs[n], s);
815
816 if (n > 0) {
817 alloc_tcg_plugin_context(s);
818 tcg_region_initial_alloc(s);
819 }
820
821 tcg_ctx = s;
822 }
823 #endif /* !CONFIG_USER_ONLY */
824
825 /* pool based memory allocation */
826 void *tcg_malloc_internal(TCGContext *s, int size)
827 {
828 TCGPool *p;
829 int pool_size;
830
831 if (size > TCG_POOL_CHUNK_SIZE) {
832 /* big malloc: insert a new pool (XXX: could optimize) */
833 p = g_malloc(sizeof(TCGPool) + size);
834 p->size = size;
835 p->next = s->pool_first_large;
836 s->pool_first_large = p;
837 return p->data;
838 } else {
839 p = s->pool_current;
840 if (!p) {
841 p = s->pool_first;
842 if (!p)
843 goto new_pool;
844 } else {
845 if (!p->next) {
846 new_pool:
847 pool_size = TCG_POOL_CHUNK_SIZE;
848 p = g_malloc(sizeof(TCGPool) + pool_size);
849 p->size = pool_size;
850 p->next = NULL;
851 if (s->pool_current) {
852 s->pool_current->next = p;
853 } else {
854 s->pool_first = p;
855 }
856 } else {
857 p = p->next;
858 }
859 }
860 }
861 s->pool_current = p;
862 s->pool_cur = p->data + size;
863 s->pool_end = p->data + p->size;
864 return p->data;
865 }
866
867 void tcg_pool_reset(TCGContext *s)
868 {
869 TCGPool *p, *t;
870 for (p = s->pool_first_large; p; p = t) {
871 t = p->next;
872 g_free(p);
873 }
874 s->pool_first_large = NULL;
875 s->pool_cur = s->pool_end = NULL;
876 s->pool_current = NULL;
877 }
878
879 /*
880 * Create TCGHelperInfo structures for "tcg/tcg-ldst.h" functions,
881 * akin to what "exec/helper-tcg.h" does with DEF_HELPER_FLAGS_N.
882 * We only use these for layout in tcg_out_ld_helper_ret and
883 * tcg_out_st_helper_args, and share them between several of
884 * the helpers, with the end result that it's easier to build manually.
885 */
886
887 #if TCG_TARGET_REG_BITS == 32
888 # define dh_typecode_ttl dh_typecode_i32
889 #else
890 # define dh_typecode_ttl dh_typecode_i64
891 #endif
892
893 static TCGHelperInfo info_helper_ld32_mmu = {
894 .flags = TCG_CALL_NO_WG,
895 .typemask = dh_typemask(ttl, 0) /* return tcg_target_ulong */
896 | dh_typemask(env, 1)
897 | dh_typemask(i64, 2) /* uint64_t addr */
898 | dh_typemask(i32, 3) /* unsigned oi */
899 | dh_typemask(ptr, 4) /* uintptr_t ra */
900 };
901
902 static TCGHelperInfo info_helper_ld64_mmu = {
903 .flags = TCG_CALL_NO_WG,
904 .typemask = dh_typemask(i64, 0) /* return uint64_t */
905 | dh_typemask(env, 1)
906 | dh_typemask(i64, 2) /* uint64_t addr */
907 | dh_typemask(i32, 3) /* unsigned oi */
908 | dh_typemask(ptr, 4) /* uintptr_t ra */
909 };
910
911 static TCGHelperInfo info_helper_ld128_mmu = {
912 .flags = TCG_CALL_NO_WG,
913 .typemask = dh_typemask(i128, 0) /* return Int128 */
914 | dh_typemask(env, 1)
915 | dh_typemask(i64, 2) /* uint64_t addr */
916 | dh_typemask(i32, 3) /* unsigned oi */
917 | dh_typemask(ptr, 4) /* uintptr_t ra */
918 };
919
920 static TCGHelperInfo info_helper_st32_mmu = {
921 .flags = TCG_CALL_NO_WG,
922 .typemask = dh_typemask(void, 0)
923 | dh_typemask(env, 1)
924 | dh_typemask(i64, 2) /* uint64_t addr */
925 | dh_typemask(i32, 3) /* uint32_t data */
926 | dh_typemask(i32, 4) /* unsigned oi */
927 | dh_typemask(ptr, 5) /* uintptr_t ra */
928 };
929
930 static TCGHelperInfo info_helper_st64_mmu = {
931 .flags = TCG_CALL_NO_WG,
932 .typemask = dh_typemask(void, 0)
933 | dh_typemask(env, 1)
934 | dh_typemask(i64, 2) /* uint64_t addr */
935 | dh_typemask(i64, 3) /* uint64_t data */
936 | dh_typemask(i32, 4) /* unsigned oi */
937 | dh_typemask(ptr, 5) /* uintptr_t ra */
938 };
939
940 static TCGHelperInfo info_helper_st128_mmu = {
941 .flags = TCG_CALL_NO_WG,
942 .typemask = dh_typemask(void, 0)
943 | dh_typemask(env, 1)
944 | dh_typemask(i64, 2) /* uint64_t addr */
945 | dh_typemask(i128, 3) /* Int128 data */
946 | dh_typemask(i32, 4) /* unsigned oi */
947 | dh_typemask(ptr, 5) /* uintptr_t ra */
948 };
949
950 #ifdef CONFIG_TCG_INTERPRETER
951 static ffi_type *typecode_to_ffi(int argmask)
952 {
953 /*
954 * libffi does not support __int128_t, so we have forced Int128
955 * to use the structure definition instead of the builtin type.
956 */
957 static ffi_type *ffi_type_i128_elements[3] = {
958 &ffi_type_uint64,
959 &ffi_type_uint64,
960 NULL
961 };
962 static ffi_type ffi_type_i128 = {
963 .size = 16,
964 .alignment = __alignof__(Int128),
965 .type = FFI_TYPE_STRUCT,
966 .elements = ffi_type_i128_elements,
967 };
968
969 switch (argmask) {
970 case dh_typecode_void:
971 return &ffi_type_void;
972 case dh_typecode_i32:
973 return &ffi_type_uint32;
974 case dh_typecode_s32:
975 return &ffi_type_sint32;
976 case dh_typecode_i64:
977 return &ffi_type_uint64;
978 case dh_typecode_s64:
979 return &ffi_type_sint64;
980 case dh_typecode_ptr:
981 return &ffi_type_pointer;
982 case dh_typecode_i128:
983 return &ffi_type_i128;
984 }
985 g_assert_not_reached();
986 }
987
988 static ffi_cif *init_ffi_layout(TCGHelperInfo *info)
989 {
990 unsigned typemask = info->typemask;
991 struct {
992 ffi_cif cif;
993 ffi_type *args[];
994 } *ca;
995 ffi_status status;
996 int nargs;
997
998 /* Ignoring the return type, find the last non-zero field. */
999 nargs = 32 - clz32(typemask >> 3);
1000 nargs = DIV_ROUND_UP(nargs, 3);
1001 assert(nargs <= MAX_CALL_IARGS);
1002
1003 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *));
1004 ca->cif.rtype = typecode_to_ffi(typemask & 7);
1005 ca->cif.nargs = nargs;
1006
1007 if (nargs != 0) {
1008 ca->cif.arg_types = ca->args;
1009 for (int j = 0; j < nargs; ++j) {
1010 int typecode = extract32(typemask, (j + 1) * 3, 3);
1011 ca->args[j] = typecode_to_ffi(typecode);
1012 }
1013 }
1014
1015 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs,
1016 ca->cif.rtype, ca->cif.arg_types);
1017 assert(status == FFI_OK);
1018
1019 return &ca->cif;
1020 }
1021
1022 #define HELPER_INFO_INIT(I) (&(I)->cif)
1023 #define HELPER_INFO_INIT_VAL(I) init_ffi_layout(I)
1024 #else
1025 #define HELPER_INFO_INIT(I) (&(I)->init)
1026 #define HELPER_INFO_INIT_VAL(I) 1
1027 #endif /* CONFIG_TCG_INTERPRETER */
1028
1029 static inline bool arg_slot_reg_p(unsigned arg_slot)
1030 {
1031 /*
1032 * Split the sizeof away from the comparison to avoid Werror from
1033 * "unsigned < 0 is always false", when iarg_regs is empty.
1034 */
1035 unsigned nreg = ARRAY_SIZE(tcg_target_call_iarg_regs);
1036 return arg_slot < nreg;
1037 }
1038
1039 static inline int arg_slot_stk_ofs(unsigned arg_slot)
1040 {
1041 unsigned max = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long);
1042 unsigned stk_slot = arg_slot - ARRAY_SIZE(tcg_target_call_iarg_regs);
1043
1044 tcg_debug_assert(stk_slot < max);
1045 return TCG_TARGET_CALL_STACK_OFFSET + stk_slot * sizeof(tcg_target_long);
1046 }
1047
1048 typedef struct TCGCumulativeArgs {
1049 int arg_idx; /* tcg_gen_callN args[] */
1050 int info_in_idx; /* TCGHelperInfo in[] */
1051 int arg_slot; /* regs+stack slot */
1052 int ref_slot; /* stack slots for references */
1053 } TCGCumulativeArgs;
1054
1055 static void layout_arg_even(TCGCumulativeArgs *cum)
1056 {
1057 cum->arg_slot += cum->arg_slot & 1;
1058 }
1059
1060 static void layout_arg_1(TCGCumulativeArgs *cum, TCGHelperInfo *info,
1061 TCGCallArgumentKind kind)
1062 {
1063 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx];
1064
1065 *loc = (TCGCallArgumentLoc){
1066 .kind = kind,
1067 .arg_idx = cum->arg_idx,
1068 .arg_slot = cum->arg_slot,
1069 };
1070 cum->info_in_idx++;
1071 cum->arg_slot++;
1072 }
1073
1074 static void layout_arg_normal_n(TCGCumulativeArgs *cum,
1075 TCGHelperInfo *info, int n)
1076 {
1077 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx];
1078
1079 for (int i = 0; i < n; ++i) {
1080 /* Layout all using the same arg_idx, adjusting the subindex. */
1081 loc[i] = (TCGCallArgumentLoc){
1082 .kind = TCG_CALL_ARG_NORMAL,
1083 .arg_idx = cum->arg_idx,
1084 .tmp_subindex = i,
1085 .arg_slot = cum->arg_slot + i,
1086 };
1087 }
1088 cum->info_in_idx += n;
1089 cum->arg_slot += n;
1090 }
1091
1092 static void layout_arg_by_ref(TCGCumulativeArgs *cum, TCGHelperInfo *info)
1093 {
1094 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx];
1095 int n = 128 / TCG_TARGET_REG_BITS;
1096
1097 /* The first subindex carries the pointer. */
1098 layout_arg_1(cum, info, TCG_CALL_ARG_BY_REF);
1099
1100 /*
1101 * The callee is allowed to clobber memory associated with
1102 * structure pass by-reference. Therefore we must make copies.
1103 * Allocate space from "ref_slot", which will be adjusted to
1104 * follow the parameters on the stack.
1105 */
1106 loc[0].ref_slot = cum->ref_slot;
1107
1108 /*
1109 * Subsequent words also go into the reference slot, but
1110 * do not accumulate into the regular arguments.
1111 */
1112 for (int i = 1; i < n; ++i) {
1113 loc[i] = (TCGCallArgumentLoc){
1114 .kind = TCG_CALL_ARG_BY_REF_N,
1115 .arg_idx = cum->arg_idx,
1116 .tmp_subindex = i,
1117 .ref_slot = cum->ref_slot + i,
1118 };
1119 }
1120 cum->info_in_idx += n - 1; /* i=0 accounted for in layout_arg_1 */
1121 cum->ref_slot += n;
1122 }
1123
1124 static void init_call_layout(TCGHelperInfo *info)
1125 {
1126 int max_reg_slots = ARRAY_SIZE(tcg_target_call_iarg_regs);
1127 int max_stk_slots = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long);
1128 unsigned typemask = info->typemask;
1129 unsigned typecode;
1130 TCGCumulativeArgs cum = { };
1131
1132 /*
1133 * Parse and place any function return value.
1134 */
1135 typecode = typemask & 7;
1136 switch (typecode) {
1137 case dh_typecode_void:
1138 info->nr_out = 0;
1139 break;
1140 case dh_typecode_i32:
1141 case dh_typecode_s32:
1142 case dh_typecode_ptr:
1143 info->nr_out = 1;
1144 info->out_kind = TCG_CALL_RET_NORMAL;
1145 break;
1146 case dh_typecode_i64:
1147 case dh_typecode_s64:
1148 info->nr_out = 64 / TCG_TARGET_REG_BITS;
1149 info->out_kind = TCG_CALL_RET_NORMAL;
1150 /* Query the last register now to trigger any assert early. */
1151 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1);
1152 break;
1153 case dh_typecode_i128:
1154 info->nr_out = 128 / TCG_TARGET_REG_BITS;
1155 info->out_kind = TCG_TARGET_CALL_RET_I128;
1156 switch (TCG_TARGET_CALL_RET_I128) {
1157 case TCG_CALL_RET_NORMAL:
1158 /* Query the last register now to trigger any assert early. */
1159 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1);
1160 break;
1161 case TCG_CALL_RET_BY_VEC:
1162 /* Query the single register now to trigger any assert early. */
1163 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0);
1164 break;
1165 case TCG_CALL_RET_BY_REF:
1166 /*
1167 * Allocate the first argument to the output.
1168 * We don't need to store this anywhere, just make it
1169 * unavailable for use in the input loop below.
1170 */
1171 cum.arg_slot = 1;
1172 break;
1173 default:
1174 qemu_build_not_reached();
1175 }
1176 break;
1177 default:
1178 g_assert_not_reached();
1179 }
1180
1181 /*
1182 * Parse and place function arguments.
1183 */
1184 for (typemask >>= 3; typemask; typemask >>= 3, cum.arg_idx++) {
1185 TCGCallArgumentKind kind;
1186 TCGType type;
1187
1188 typecode = typemask & 7;
1189 switch (typecode) {
1190 case dh_typecode_i32:
1191 case dh_typecode_s32:
1192 type = TCG_TYPE_I32;
1193 break;
1194 case dh_typecode_i64:
1195 case dh_typecode_s64:
1196 type = TCG_TYPE_I64;
1197 break;
1198 case dh_typecode_ptr:
1199 type = TCG_TYPE_PTR;
1200 break;
1201 case dh_typecode_i128:
1202 type = TCG_TYPE_I128;
1203 break;
1204 default:
1205 g_assert_not_reached();
1206 }
1207
1208 switch (type) {
1209 case TCG_TYPE_I32:
1210 switch (TCG_TARGET_CALL_ARG_I32) {
1211 case TCG_CALL_ARG_EVEN:
1212 layout_arg_even(&cum);
1213 /* fall through */
1214 case TCG_CALL_ARG_NORMAL:
1215 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL);
1216 break;
1217 case TCG_CALL_ARG_EXTEND:
1218 kind = TCG_CALL_ARG_EXTEND_U + (typecode & 1);
1219 layout_arg_1(&cum, info, kind);
1220 break;
1221 default:
1222 qemu_build_not_reached();
1223 }
1224 break;
1225
1226 case TCG_TYPE_I64:
1227 switch (TCG_TARGET_CALL_ARG_I64) {
1228 case TCG_CALL_ARG_EVEN:
1229 layout_arg_even(&cum);
1230 /* fall through */
1231 case TCG_CALL_ARG_NORMAL:
1232 if (TCG_TARGET_REG_BITS == 32) {
1233 layout_arg_normal_n(&cum, info, 2);
1234 } else {
1235 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL);
1236 }
1237 break;
1238 default:
1239 qemu_build_not_reached();
1240 }
1241 break;
1242
1243 case TCG_TYPE_I128:
1244 switch (TCG_TARGET_CALL_ARG_I128) {
1245 case TCG_CALL_ARG_EVEN:
1246 layout_arg_even(&cum);
1247 /* fall through */
1248 case TCG_CALL_ARG_NORMAL:
1249 layout_arg_normal_n(&cum, info, 128 / TCG_TARGET_REG_BITS);
1250 break;
1251 case TCG_CALL_ARG_BY_REF:
1252 layout_arg_by_ref(&cum, info);
1253 break;
1254 default:
1255 qemu_build_not_reached();
1256 }
1257 break;
1258
1259 default:
1260 g_assert_not_reached();
1261 }
1262 }
1263 info->nr_in = cum.info_in_idx;
1264
1265 /* Validate that we didn't overrun the input array. */
1266 assert(cum.info_in_idx <= ARRAY_SIZE(info->in));
1267 /* Validate the backend has enough argument space. */
1268 assert(cum.arg_slot <= max_reg_slots + max_stk_slots);
1269
1270 /*
1271 * Relocate the "ref_slot" area to the end of the parameters.
1272 * Minimizing this stack offset helps code size for x86,
1273 * which has a signed 8-bit offset encoding.
1274 */
1275 if (cum.ref_slot != 0) {
1276 int ref_base = 0;
1277
1278 if (cum.arg_slot > max_reg_slots) {
1279 int align = __alignof(Int128) / sizeof(tcg_target_long);
1280
1281 ref_base = cum.arg_slot - max_reg_slots;
1282 if (align > 1) {
1283 ref_base = ROUND_UP(ref_base, align);
1284 }
1285 }
1286 assert(ref_base + cum.ref_slot <= max_stk_slots);
1287 ref_base += max_reg_slots;
1288
1289 if (ref_base != 0) {
1290 for (int i = cum.info_in_idx - 1; i >= 0; --i) {
1291 TCGCallArgumentLoc *loc = &info->in[i];
1292 switch (loc->kind) {
1293 case TCG_CALL_ARG_BY_REF:
1294 case TCG_CALL_ARG_BY_REF_N:
1295 loc->ref_slot += ref_base;
1296 break;
1297 default:
1298 break;
1299 }
1300 }
1301 }
1302 }
1303 }
1304
1305 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)];
1306 static void process_op_defs(TCGContext *s);
1307 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type,
1308 TCGReg reg, const char *name);
1309
1310 static void tcg_context_init(unsigned max_cpus)
1311 {
1312 TCGContext *s = &tcg_init_ctx;
1313 int op, total_args, n, i;
1314 TCGOpDef *def;
1315 TCGArgConstraint *args_ct;
1316 TCGTemp *ts;
1317
1318 memset(s, 0, sizeof(*s));
1319 s->nb_globals = 0;
1320
1321 /* Count total number of arguments and allocate the corresponding
1322 space */
1323 total_args = 0;
1324 for(op = 0; op < NB_OPS; op++) {
1325 def = &tcg_op_defs[op];
1326 n = def->nb_iargs + def->nb_oargs;
1327 total_args += n;
1328 }
1329
1330 args_ct = g_new0(TCGArgConstraint, total_args);
1331
1332 for(op = 0; op < NB_OPS; op++) {
1333 def = &tcg_op_defs[op];
1334 def->args_ct = args_ct;
1335 n = def->nb_iargs + def->nb_oargs;
1336 args_ct += n;
1337 }
1338
1339 init_call_layout(&info_helper_ld32_mmu);
1340 init_call_layout(&info_helper_ld64_mmu);
1341 init_call_layout(&info_helper_ld128_mmu);
1342 init_call_layout(&info_helper_st32_mmu);
1343 init_call_layout(&info_helper_st64_mmu);
1344 init_call_layout(&info_helper_st128_mmu);
1345
1346 tcg_target_init(s);
1347 process_op_defs(s);
1348
1349 /* Reverse the order of the saved registers, assuming they're all at
1350 the start of tcg_target_reg_alloc_order. */
1351 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) {
1352 int r = tcg_target_reg_alloc_order[n];
1353 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) {
1354 break;
1355 }
1356 }
1357 for (i = 0; i < n; ++i) {
1358 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i];
1359 }
1360 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) {
1361 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i];
1362 }
1363
1364 alloc_tcg_plugin_context(s);
1365
1366 tcg_ctx = s;
1367 /*
1368 * In user-mode we simply share the init context among threads, since we
1369 * use a single region. See the documentation tcg_region_init() for the
1370 * reasoning behind this.
1371 * In system-mode we will have at most max_cpus TCG threads.
1372 */
1373 #ifdef CONFIG_USER_ONLY
1374 tcg_ctxs = &tcg_ctx;
1375 tcg_cur_ctxs = 1;
1376 tcg_max_ctxs = 1;
1377 #else
1378 tcg_max_ctxs = max_cpus;
1379 tcg_ctxs = g_new0(TCGContext *, max_cpus);
1380 #endif
1381
1382 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0));
1383 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env");
1384 tcg_env = temp_tcgv_ptr(ts);
1385 }
1386
1387 void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus)
1388 {
1389 tcg_context_init(max_cpus);
1390 tcg_region_init(tb_size, splitwx, max_cpus);
1391 }
1392
1393 /*
1394 * Allocate TBs right before their corresponding translated code, making
1395 * sure that TBs and code are on different cache lines.
1396 */
1397 TranslationBlock *tcg_tb_alloc(TCGContext *s)
1398 {
1399 uintptr_t align = qemu_icache_linesize;
1400 TranslationBlock *tb;
1401 void *next;
1402
1403 retry:
1404 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align);
1405 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align);
1406
1407 if (unlikely(next > s->code_gen_highwater)) {
1408 if (tcg_region_alloc(s)) {
1409 return NULL;
1410 }
1411 goto retry;
1412 }
1413 qatomic_set(&s->code_gen_ptr, next);
1414 s->data_gen_ptr = NULL;
1415 return tb;
1416 }
1417
1418 void tcg_prologue_init(void)
1419 {
1420 TCGContext *s = tcg_ctx;
1421 size_t prologue_size;
1422
1423 s->code_ptr = s->code_gen_ptr;
1424 s->code_buf = s->code_gen_ptr;
1425 s->data_gen_ptr = NULL;
1426
1427 #ifndef CONFIG_TCG_INTERPRETER
1428 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr);
1429 #endif
1430
1431 #ifdef TCG_TARGET_NEED_POOL_LABELS
1432 s->pool_labels = NULL;
1433 #endif
1434
1435 qemu_thread_jit_write();
1436 /* Generate the prologue. */
1437 tcg_target_qemu_prologue(s);
1438
1439 #ifdef TCG_TARGET_NEED_POOL_LABELS
1440 /* Allow the prologue to put e.g. guest_base into a pool entry. */
1441 {
1442 int result = tcg_out_pool_finalize(s);
1443 tcg_debug_assert(result == 0);
1444 }
1445 #endif
1446
1447 prologue_size = tcg_current_code_size(s);
1448 perf_report_prologue(s->code_gen_ptr, prologue_size);
1449
1450 #ifndef CONFIG_TCG_INTERPRETER
1451 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf),
1452 (uintptr_t)s->code_buf, prologue_size);
1453 #endif
1454
1455 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) {
1456 FILE *logfile = qemu_log_trylock();
1457 if (logfile) {
1458 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size);
1459 if (s->data_gen_ptr) {
1460 size_t code_size = s->data_gen_ptr - s->code_gen_ptr;
1461 size_t data_size = prologue_size - code_size;
1462 size_t i;
1463
1464 disas(logfile, s->code_gen_ptr, code_size);
1465
1466 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) {
1467 if (sizeof(tcg_target_ulong) == 8) {
1468 fprintf(logfile,
1469 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n",
1470 (uintptr_t)s->data_gen_ptr + i,
1471 *(uint64_t *)(s->data_gen_ptr + i));
1472 } else {
1473 fprintf(logfile,
1474 "0x%08" PRIxPTR ": .long 0x%08x\n",
1475 (uintptr_t)s->data_gen_ptr + i,
1476 *(uint32_t *)(s->data_gen_ptr + i));
1477 }
1478 }
1479 } else {
1480 disas(logfile, s->code_gen_ptr, prologue_size);
1481 }
1482 fprintf(logfile, "\n");
1483 qemu_log_unlock(logfile);
1484 }
1485 }
1486
1487 #ifndef CONFIG_TCG_INTERPRETER
1488 /*
1489 * Assert that goto_ptr is implemented completely, setting an epilogue.
1490 * For tci, we use NULL as the signal to return from the interpreter,
1491 * so skip this check.
1492 */
1493 tcg_debug_assert(tcg_code_gen_epilogue != NULL);
1494 #endif
1495
1496 tcg_region_prologue_set(s);
1497 }
1498
1499 void tcg_func_start(TCGContext *s)
1500 {
1501 tcg_pool_reset(s);
1502 s->nb_temps = s->nb_globals;
1503
1504 /* No temps have been previously allocated for size or locality. */
1505 memset(s->free_temps, 0, sizeof(s->free_temps));
1506
1507 /* No constant temps have been previously allocated. */
1508 for (int i = 0; i < TCG_TYPE_COUNT; ++i) {
1509 if (s->const_table[i]) {
1510 g_hash_table_remove_all(s->const_table[i]);
1511 }
1512 }
1513
1514 s->nb_ops = 0;
1515 s->nb_labels = 0;
1516 s->current_frame_offset = s->frame_start;
1517
1518 #ifdef CONFIG_DEBUG_TCG
1519 s->goto_tb_issue_mask = 0;
1520 #endif
1521
1522 QTAILQ_INIT(&s->ops);
1523 QTAILQ_INIT(&s->free_ops);
1524 QSIMPLEQ_INIT(&s->labels);
1525
1526 tcg_debug_assert(s->addr_type == TCG_TYPE_I32 ||
1527 s->addr_type == TCG_TYPE_I64);
1528
1529 tcg_debug_assert(s->insn_start_words > 0);
1530 }
1531
1532 static TCGTemp *tcg_temp_alloc(TCGContext *s)
1533 {
1534 int n = s->nb_temps++;
1535
1536 if (n >= TCG_MAX_TEMPS) {
1537 tcg_raise_tb_overflow(s);
1538 }
1539 return memset(&s->temps[n], 0, sizeof(TCGTemp));
1540 }
1541
1542 static TCGTemp *tcg_global_alloc(TCGContext *s)
1543 {
1544 TCGTemp *ts;
1545
1546 tcg_debug_assert(s->nb_globals == s->nb_temps);
1547 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS);
1548 s->nb_globals++;
1549 ts = tcg_temp_alloc(s);
1550 ts->kind = TEMP_GLOBAL;
1551
1552 return ts;
1553 }
1554
1555 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type,
1556 TCGReg reg, const char *name)
1557 {
1558 TCGTemp *ts;
1559
1560 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32);
1561
1562 ts = tcg_global_alloc(s);
1563 ts->base_type = type;
1564 ts->type = type;
1565 ts->kind = TEMP_FIXED;
1566 ts->reg = reg;
1567 ts->name = name;
1568 tcg_regset_set_reg(s->reserved_regs, reg);
1569
1570 return ts;
1571 }
1572
1573 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size)
1574 {
1575 s->frame_start = start;
1576 s->frame_end = start + size;
1577 s->frame_temp
1578 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame");
1579 }
1580
1581 static TCGTemp *tcg_global_mem_new_internal(TCGv_ptr base, intptr_t offset,
1582 const char *name, TCGType type)
1583 {
1584 TCGContext *s = tcg_ctx;
1585 TCGTemp *base_ts = tcgv_ptr_temp(base);
1586 TCGTemp *ts = tcg_global_alloc(s);
1587 int indirect_reg = 0;
1588
1589 switch (base_ts->kind) {
1590 case TEMP_FIXED:
1591 break;
1592 case TEMP_GLOBAL:
1593 /* We do not support double-indirect registers. */
1594 tcg_debug_assert(!base_ts->indirect_reg);
1595 base_ts->indirect_base = 1;
1596 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64
1597 ? 2 : 1);
1598 indirect_reg = 1;
1599 break;
1600 default:
1601 g_assert_not_reached();
1602 }
1603
1604 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) {
1605 TCGTemp *ts2 = tcg_global_alloc(s);
1606 char buf[64];
1607
1608 ts->base_type = TCG_TYPE_I64;
1609 ts->type = TCG_TYPE_I32;
1610 ts->indirect_reg = indirect_reg;
1611 ts->mem_allocated = 1;
1612 ts->mem_base = base_ts;
1613 ts->mem_offset = offset;
1614 pstrcpy(buf, sizeof(buf), name);
1615 pstrcat(buf, sizeof(buf), "_0");
1616 ts->name = strdup(buf);
1617
1618 tcg_debug_assert(ts2 == ts + 1);
1619 ts2->base_type = TCG_TYPE_I64;
1620 ts2->type = TCG_TYPE_I32;
1621 ts2->indirect_reg = indirect_reg;
1622 ts2->mem_allocated = 1;
1623 ts2->mem_base = base_ts;
1624 ts2->mem_offset = offset + 4;
1625 ts2->temp_subindex = 1;
1626 pstrcpy(buf, sizeof(buf), name);
1627 pstrcat(buf, sizeof(buf), "_1");
1628 ts2->name = strdup(buf);
1629 } else {
1630 ts->base_type = type;
1631 ts->type = type;
1632 ts->indirect_reg = indirect_reg;
1633 ts->mem_allocated = 1;
1634 ts->mem_base = base_ts;
1635 ts->mem_offset = offset;
1636 ts->name = name;
1637 }
1638 return ts;
1639 }
1640
1641 TCGv_i32 tcg_global_mem_new_i32(TCGv_ptr reg, intptr_t off, const char *name)
1642 {
1643 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I32);
1644 return temp_tcgv_i32(ts);
1645 }
1646
1647 TCGv_i64 tcg_global_mem_new_i64(TCGv_ptr reg, intptr_t off, const char *name)
1648 {
1649 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I64);
1650 return temp_tcgv_i64(ts);
1651 }
1652
1653 TCGv_ptr tcg_global_mem_new_ptr(TCGv_ptr reg, intptr_t off, const char *name)
1654 {
1655 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_PTR);
1656 return temp_tcgv_ptr(ts);
1657 }
1658
1659 TCGTemp *tcg_temp_new_internal(TCGType type, TCGTempKind kind)
1660 {
1661 TCGContext *s = tcg_ctx;
1662 TCGTemp *ts;
1663 int n;
1664
1665 if (kind == TEMP_EBB) {
1666 int idx = find_first_bit(s->free_temps[type].l, TCG_MAX_TEMPS);
1667
1668 if (idx < TCG_MAX_TEMPS) {
1669 /* There is already an available temp with the right type. */
1670 clear_bit(idx, s->free_temps[type].l);
1671
1672 ts = &s->temps[idx];
1673 ts->temp_allocated = 1;
1674 tcg_debug_assert(ts->base_type == type);
1675 tcg_debug_assert(ts->kind == kind);
1676 return ts;
1677 }
1678 } else {
1679 tcg_debug_assert(kind == TEMP_TB);
1680 }
1681
1682 switch (type) {
1683 case TCG_TYPE_I32:
1684 case TCG_TYPE_V64:
1685 case TCG_TYPE_V128:
1686 case TCG_TYPE_V256:
1687 n = 1;
1688 break;
1689 case TCG_TYPE_I64:
1690 n = 64 / TCG_TARGET_REG_BITS;
1691 break;
1692 case TCG_TYPE_I128:
1693 n = 128 / TCG_TARGET_REG_BITS;
1694 break;
1695 default:
1696 g_assert_not_reached();
1697 }
1698
1699 ts = tcg_temp_alloc(s);
1700 ts->base_type = type;
1701 ts->temp_allocated = 1;
1702 ts->kind = kind;
1703
1704 if (n == 1) {
1705 ts->type = type;
1706 } else {
1707 ts->type = TCG_TYPE_REG;
1708
1709 for (int i = 1; i < n; ++i) {
1710 TCGTemp *ts2 = tcg_temp_alloc(s);
1711
1712 tcg_debug_assert(ts2 == ts + i);
1713 ts2->base_type = type;
1714 ts2->type = TCG_TYPE_REG;
1715 ts2->temp_allocated = 1;
1716 ts2->temp_subindex = i;
1717 ts2->kind = kind;
1718 }
1719 }
1720 return ts;
1721 }
1722
1723 TCGv_i32 tcg_temp_new_i32(void)
1724 {
1725 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_TB));
1726 }
1727
1728 TCGv_i32 tcg_temp_ebb_new_i32(void)
1729 {
1730 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_EBB));
1731 }
1732
1733 TCGv_i64 tcg_temp_new_i64(void)
1734 {
1735 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_TB));
1736 }
1737
1738 TCGv_i64 tcg_temp_ebb_new_i64(void)
1739 {
1740 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_EBB));
1741 }
1742
1743 TCGv_ptr tcg_temp_new_ptr(void)
1744 {
1745 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_TB));
1746 }
1747
1748 TCGv_ptr tcg_temp_ebb_new_ptr(void)
1749 {
1750 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_EBB));
1751 }
1752
1753 TCGv_i128 tcg_temp_new_i128(void)
1754 {
1755 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_TB));
1756 }
1757
1758 TCGv_i128 tcg_temp_ebb_new_i128(void)
1759 {
1760 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_EBB));
1761 }
1762
1763 TCGv_vec tcg_temp_new_vec(TCGType type)
1764 {
1765 TCGTemp *t;
1766
1767 #ifdef CONFIG_DEBUG_TCG
1768 switch (type) {
1769 case TCG_TYPE_V64:
1770 assert(TCG_TARGET_HAS_v64);
1771 break;
1772 case TCG_TYPE_V128:
1773 assert(TCG_TARGET_HAS_v128);
1774 break;
1775 case TCG_TYPE_V256:
1776 assert(TCG_TARGET_HAS_v256);
1777 break;
1778 default:
1779 g_assert_not_reached();
1780 }
1781 #endif
1782
1783 t = tcg_temp_new_internal(type, TEMP_EBB);
1784 return temp_tcgv_vec(t);
1785 }
1786
1787 /* Create a new temp of the same type as an existing temp. */
1788 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match)
1789 {
1790 TCGTemp *t = tcgv_vec_temp(match);
1791
1792 tcg_debug_assert(t->temp_allocated != 0);
1793
1794 t = tcg_temp_new_internal(t->base_type, TEMP_EBB);
1795 return temp_tcgv_vec(t);
1796 }
1797
1798 void tcg_temp_free_internal(TCGTemp *ts)
1799 {
1800 TCGContext *s = tcg_ctx;
1801
1802 switch (ts->kind) {
1803 case TEMP_CONST:
1804 case TEMP_TB:
1805 /* Silently ignore free. */
1806 break;
1807 case TEMP_EBB:
1808 tcg_debug_assert(ts->temp_allocated != 0);
1809 ts->temp_allocated = 0;
1810 set_bit(temp_idx(ts), s->free_temps[ts->base_type].l);
1811 break;
1812 default:
1813 /* It never made sense to free TEMP_FIXED or TEMP_GLOBAL. */
1814 g_assert_not_reached();
1815 }
1816 }
1817
1818 void tcg_temp_free_i32(TCGv_i32 arg)
1819 {
1820 tcg_temp_free_internal(tcgv_i32_temp(arg));
1821 }
1822
1823 void tcg_temp_free_i64(TCGv_i64 arg)
1824 {
1825 tcg_temp_free_internal(tcgv_i64_temp(arg));
1826 }
1827
1828 void tcg_temp_free_i128(TCGv_i128 arg)
1829 {
1830 tcg_temp_free_internal(tcgv_i128_temp(arg));
1831 }
1832
1833 void tcg_temp_free_ptr(TCGv_ptr arg)
1834 {
1835 tcg_temp_free_internal(tcgv_ptr_temp(arg));
1836 }
1837
1838 void tcg_temp_free_vec(TCGv_vec arg)
1839 {
1840 tcg_temp_free_internal(tcgv_vec_temp(arg));
1841 }
1842
1843 TCGTemp *tcg_constant_internal(TCGType type, int64_t val)
1844 {
1845 TCGContext *s = tcg_ctx;
1846 GHashTable *h = s->const_table[type];
1847 TCGTemp *ts;
1848
1849 if (h == NULL) {
1850 h = g_hash_table_new(g_int64_hash, g_int64_equal);
1851 s->const_table[type] = h;
1852 }
1853
1854 ts = g_hash_table_lookup(h, &val);
1855 if (ts == NULL) {
1856 int64_t *val_ptr;
1857
1858 ts = tcg_temp_alloc(s);
1859
1860 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) {
1861 TCGTemp *ts2 = tcg_temp_alloc(s);
1862
1863 tcg_debug_assert(ts2 == ts + 1);
1864
1865 ts->base_type = TCG_TYPE_I64;
1866 ts->type = TCG_TYPE_I32;
1867 ts->kind = TEMP_CONST;
1868 ts->temp_allocated = 1;
1869
1870 ts2->base_type = TCG_TYPE_I64;
1871 ts2->type = TCG_TYPE_I32;
1872 ts2->kind = TEMP_CONST;
1873 ts2->temp_allocated = 1;
1874 ts2->temp_subindex = 1;
1875
1876 /*
1877 * Retain the full value of the 64-bit constant in the low
1878 * part, so that the hash table works. Actual uses will
1879 * truncate the value to the low part.
1880 */
1881 ts[HOST_BIG_ENDIAN].val = val;
1882 ts[!HOST_BIG_ENDIAN].val = val >> 32;
1883 val_ptr = &ts[HOST_BIG_ENDIAN].val;
1884 } else {
1885 ts->base_type = type;
1886 ts->type = type;
1887 ts->kind = TEMP_CONST;
1888 ts->temp_allocated = 1;
1889 ts->val = val;
1890 val_ptr = &ts->val;
1891 }
1892 g_hash_table_insert(h, val_ptr, ts);
1893 }
1894
1895 return ts;
1896 }
1897
1898 TCGv_i32 tcg_constant_i32(int32_t val)
1899 {
1900 return temp_tcgv_i32(tcg_constant_internal(TCG_TYPE_I32, val));
1901 }
1902
1903 TCGv_i64 tcg_constant_i64(int64_t val)
1904 {
1905 return temp_tcgv_i64(tcg_constant_internal(TCG_TYPE_I64, val));
1906 }
1907
1908 TCGv_ptr tcg_constant_ptr_int(intptr_t val)
1909 {
1910 return temp_tcgv_ptr(tcg_constant_internal(TCG_TYPE_PTR, val));
1911 }
1912
1913 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val)
1914 {
1915 val = dup_const(vece, val);
1916 return temp_tcgv_vec(tcg_constant_internal(type, val));
1917 }
1918
1919 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val)
1920 {
1921 TCGTemp *t = tcgv_vec_temp(match);
1922
1923 tcg_debug_assert(t->temp_allocated != 0);
1924 return tcg_constant_vec(t->base_type, vece, val);
1925 }
1926
1927 #ifdef CONFIG_DEBUG_TCG
1928 size_t temp_idx(TCGTemp *ts)
1929 {
1930 ptrdiff_t n = ts - tcg_ctx->temps;
1931 assert(n >= 0 && n < tcg_ctx->nb_temps);
1932 return n;
1933 }
1934
1935 TCGTemp *tcgv_i32_temp(TCGv_i32 v)
1936 {
1937 uintptr_t o = (uintptr_t)v - offsetof(TCGContext, temps);
1938
1939 assert(o < sizeof(TCGTemp) * tcg_ctx->nb_temps);
1940 assert(o % sizeof(TCGTemp) == 0);
1941
1942 return (void *)tcg_ctx + (uintptr_t)v;
1943 }
1944 #endif /* CONFIG_DEBUG_TCG */
1945
1946 /* Return true if OP may appear in the opcode stream.
1947 Test the runtime variable that controls each opcode. */
1948 bool tcg_op_supported(TCGOpcode op)
1949 {
1950 const bool have_vec
1951 = TCG_TARGET_HAS_v64 | TCG_TARGET_HAS_v128 | TCG_TARGET_HAS_v256;
1952
1953 switch (op) {
1954 case INDEX_op_discard:
1955 case INDEX_op_set_label:
1956 case INDEX_op_call:
1957 case INDEX_op_br:
1958 case INDEX_op_mb:
1959 case INDEX_op_insn_start:
1960 case INDEX_op_exit_tb:
1961 case INDEX_op_goto_tb:
1962 case INDEX_op_goto_ptr:
1963 case INDEX_op_qemu_ld_a32_i32:
1964 case INDEX_op_qemu_ld_a64_i32:
1965 case INDEX_op_qemu_st_a32_i32:
1966 case INDEX_op_qemu_st_a64_i32:
1967 case INDEX_op_qemu_ld_a32_i64:
1968 case INDEX_op_qemu_ld_a64_i64:
1969 case INDEX_op_qemu_st_a32_i64:
1970 case INDEX_op_qemu_st_a64_i64:
1971 return true;
1972
1973 case INDEX_op_qemu_st8_a32_i32:
1974 case INDEX_op_qemu_st8_a64_i32:
1975 return TCG_TARGET_HAS_qemu_st8_i32;
1976
1977 case INDEX_op_qemu_ld_a32_i128:
1978 case INDEX_op_qemu_ld_a64_i128:
1979 case INDEX_op_qemu_st_a32_i128:
1980 case INDEX_op_qemu_st_a64_i128:
1981 return TCG_TARGET_HAS_qemu_ldst_i128;
1982
1983 case INDEX_op_mov_i32:
1984 case INDEX_op_setcond_i32:
1985 case INDEX_op_brcond_i32:
1986 case INDEX_op_movcond_i32:
1987 case INDEX_op_ld8u_i32:
1988 case INDEX_op_ld8s_i32:
1989 case INDEX_op_ld16u_i32:
1990 case INDEX_op_ld16s_i32:
1991 case INDEX_op_ld_i32:
1992 case INDEX_op_st8_i32:
1993 case INDEX_op_st16_i32:
1994 case INDEX_op_st_i32:
1995 case INDEX_op_add_i32:
1996 case INDEX_op_sub_i32:
1997 case INDEX_op_neg_i32:
1998 case INDEX_op_mul_i32:
1999 case INDEX_op_and_i32:
2000 case INDEX_op_or_i32:
2001 case INDEX_op_xor_i32:
2002 case INDEX_op_shl_i32:
2003 case INDEX_op_shr_i32:
2004 case INDEX_op_sar_i32:
2005 return true;
2006
2007 case INDEX_op_negsetcond_i32:
2008 return TCG_TARGET_HAS_negsetcond_i32;
2009 case INDEX_op_div_i32:
2010 case INDEX_op_divu_i32:
2011 return TCG_TARGET_HAS_div_i32;
2012 case INDEX_op_rem_i32:
2013 case INDEX_op_remu_i32:
2014 return TCG_TARGET_HAS_rem_i32;
2015 case INDEX_op_div2_i32:
2016 case INDEX_op_divu2_i32:
2017 return TCG_TARGET_HAS_div2_i32;
2018 case INDEX_op_rotl_i32:
2019 case INDEX_op_rotr_i32:
2020 return TCG_TARGET_HAS_rot_i32;
2021 case INDEX_op_deposit_i32:
2022 return TCG_TARGET_HAS_deposit_i32;
2023 case INDEX_op_extract_i32:
2024 return TCG_TARGET_HAS_extract_i32;
2025 case INDEX_op_sextract_i32:
2026 return TCG_TARGET_HAS_sextract_i32;
2027 case INDEX_op_extract2_i32:
2028 return TCG_TARGET_HAS_extract2_i32;
2029 case INDEX_op_add2_i32:
2030 return TCG_TARGET_HAS_add2_i32;
2031 case INDEX_op_sub2_i32:
2032 return TCG_TARGET_HAS_sub2_i32;
2033 case INDEX_op_mulu2_i32:
2034 return TCG_TARGET_HAS_mulu2_i32;
2035 case INDEX_op_muls2_i32:
2036 return TCG_TARGET_HAS_muls2_i32;
2037 case INDEX_op_muluh_i32:
2038 return TCG_TARGET_HAS_muluh_i32;
2039 case INDEX_op_mulsh_i32:
2040 return TCG_TARGET_HAS_mulsh_i32;
2041 case INDEX_op_ext8s_i32:
2042 return TCG_TARGET_HAS_ext8s_i32;
2043 case INDEX_op_ext16s_i32:
2044 return TCG_TARGET_HAS_ext16s_i32;
2045 case INDEX_op_ext8u_i32:
2046 return TCG_TARGET_HAS_ext8u_i32;
2047 case INDEX_op_ext16u_i32:
2048 return TCG_TARGET_HAS_ext16u_i32;
2049 case INDEX_op_bswap16_i32:
2050 return TCG_TARGET_HAS_bswap16_i32;
2051 case INDEX_op_bswap32_i32:
2052 return TCG_TARGET_HAS_bswap32_i32;
2053 case INDEX_op_not_i32:
2054 return TCG_TARGET_HAS_not_i32;
2055 case INDEX_op_andc_i32:
2056 return TCG_TARGET_HAS_andc_i32;
2057 case INDEX_op_orc_i32:
2058 return TCG_TARGET_HAS_orc_i32;
2059 case INDEX_op_eqv_i32:
2060 return TCG_TARGET_HAS_eqv_i32;
2061 case INDEX_op_nand_i32:
2062 return TCG_TARGET_HAS_nand_i32;
2063 case INDEX_op_nor_i32:
2064 return TCG_TARGET_HAS_nor_i32;
2065 case INDEX_op_clz_i32:
2066 return TCG_TARGET_HAS_clz_i32;
2067 case INDEX_op_ctz_i32:
2068 return TCG_TARGET_HAS_ctz_i32;
2069 case INDEX_op_ctpop_i32:
2070 return TCG_TARGET_HAS_ctpop_i32;
2071
2072 case INDEX_op_brcond2_i32:
2073 case INDEX_op_setcond2_i32:
2074 return TCG_TARGET_REG_BITS == 32;
2075
2076 case INDEX_op_mov_i64:
2077 case INDEX_op_setcond_i64:
2078 case INDEX_op_brcond_i64:
2079 case INDEX_op_movcond_i64:
2080 case INDEX_op_ld8u_i64:
2081 case INDEX_op_ld8s_i64:
2082 case INDEX_op_ld16u_i64:
2083 case INDEX_op_ld16s_i64:
2084 case INDEX_op_ld32u_i64:
2085 case INDEX_op_ld32s_i64:
2086 case INDEX_op_ld_i64:
2087 case INDEX_op_st8_i64:
2088 case INDEX_op_st16_i64:
2089 case INDEX_op_st32_i64:
2090 case INDEX_op_st_i64:
2091 case INDEX_op_add_i64:
2092 case INDEX_op_sub_i64:
2093 case INDEX_op_neg_i64:
2094 case INDEX_op_mul_i64:
2095 case INDEX_op_and_i64:
2096 case INDEX_op_or_i64:
2097 case INDEX_op_xor_i64:
2098 case INDEX_op_shl_i64:
2099 case INDEX_op_shr_i64:
2100 case INDEX_op_sar_i64:
2101 case INDEX_op_ext_i32_i64:
2102 case INDEX_op_extu_i32_i64:
2103 return TCG_TARGET_REG_BITS == 64;
2104
2105 case INDEX_op_negsetcond_i64:
2106 return TCG_TARGET_HAS_negsetcond_i64;
2107 case INDEX_op_div_i64:
2108 case INDEX_op_divu_i64:
2109 return TCG_TARGET_HAS_div_i64;
2110 case INDEX_op_rem_i64:
2111 case INDEX_op_remu_i64:
2112 return TCG_TARGET_HAS_rem_i64;
2113 case INDEX_op_div2_i64:
2114 case INDEX_op_divu2_i64:
2115 return TCG_TARGET_HAS_div2_i64;
2116 case INDEX_op_rotl_i64:
2117 case INDEX_op_rotr_i64:
2118 return TCG_TARGET_HAS_rot_i64;
2119 case INDEX_op_deposit_i64:
2120 return TCG_TARGET_HAS_deposit_i64;
2121 case INDEX_op_extract_i64:
2122 return TCG_TARGET_HAS_extract_i64;
2123 case INDEX_op_sextract_i64:
2124 return TCG_TARGET_HAS_sextract_i64;
2125 case INDEX_op_extract2_i64:
2126 return TCG_TARGET_HAS_extract2_i64;
2127 case INDEX_op_extrl_i64_i32:
2128 case INDEX_op_extrh_i64_i32:
2129 return TCG_TARGET_HAS_extr_i64_i32;
2130 case INDEX_op_ext8s_i64:
2131 return TCG_TARGET_HAS_ext8s_i64;
2132 case INDEX_op_ext16s_i64:
2133 return TCG_TARGET_HAS_ext16s_i64;
2134 case INDEX_op_ext32s_i64:
2135 return TCG_TARGET_HAS_ext32s_i64;
2136 case INDEX_op_ext8u_i64:
2137 return TCG_TARGET_HAS_ext8u_i64;
2138 case INDEX_op_ext16u_i64:
2139 return TCG_TARGET_HAS_ext16u_i64;
2140 case INDEX_op_ext32u_i64:
2141 return TCG_TARGET_HAS_ext32u_i64;
2142 case INDEX_op_bswap16_i64:
2143 return TCG_TARGET_HAS_bswap16_i64;
2144 case INDEX_op_bswap32_i64:
2145 return TCG_TARGET_HAS_bswap32_i64;
2146 case INDEX_op_bswap64_i64:
2147 return TCG_TARGET_HAS_bswap64_i64;
2148 case INDEX_op_not_i64:
2149 return TCG_TARGET_HAS_not_i64;
2150 case INDEX_op_andc_i64:
2151 return TCG_TARGET_HAS_andc_i64;
2152 case INDEX_op_orc_i64:
2153 return TCG_TARGET_HAS_orc_i64;
2154 case INDEX_op_eqv_i64:
2155 return TCG_TARGET_HAS_eqv_i64;
2156 case INDEX_op_nand_i64:
2157 return TCG_TARGET_HAS_nand_i64;
2158 case INDEX_op_nor_i64:
2159 return TCG_TARGET_HAS_nor_i64;
2160 case INDEX_op_clz_i64:
2161 return TCG_TARGET_HAS_clz_i64;
2162 case INDEX_op_ctz_i64:
2163 return TCG_TARGET_HAS_ctz_i64;
2164 case INDEX_op_ctpop_i64:
2165 return TCG_TARGET_HAS_ctpop_i64;
2166 case INDEX_op_add2_i64:
2167 return TCG_TARGET_HAS_add2_i64;
2168 case INDEX_op_sub2_i64:
2169 return TCG_TARGET_HAS_sub2_i64;
2170 case INDEX_op_mulu2_i64:
2171 return TCG_TARGET_HAS_mulu2_i64;
2172 case INDEX_op_muls2_i64:
2173 return TCG_TARGET_HAS_muls2_i64;
2174 case INDEX_op_muluh_i64:
2175 return TCG_TARGET_HAS_muluh_i64;
2176 case INDEX_op_mulsh_i64:
2177 return TCG_TARGET_HAS_mulsh_i64;
2178
2179 case INDEX_op_mov_vec:
2180 case INDEX_op_dup_vec:
2181 case INDEX_op_dupm_vec:
2182 case INDEX_op_ld_vec:
2183 case INDEX_op_st_vec:
2184 case INDEX_op_add_vec:
2185 case INDEX_op_sub_vec:
2186 case INDEX_op_and_vec:
2187 case INDEX_op_or_vec:
2188 case INDEX_op_xor_vec:
2189 case INDEX_op_cmp_vec:
2190 return have_vec;
2191 case INDEX_op_dup2_vec:
2192 return have_vec && TCG_TARGET_REG_BITS == 32;
2193 case INDEX_op_not_vec:
2194 return have_vec && TCG_TARGET_HAS_not_vec;
2195 case INDEX_op_neg_vec:
2196 return have_vec && TCG_TARGET_HAS_neg_vec;
2197 case INDEX_op_abs_vec:
2198 return have_vec && TCG_TARGET_HAS_abs_vec;
2199 case INDEX_op_andc_vec:
2200 return have_vec && TCG_TARGET_HAS_andc_vec;
2201 case INDEX_op_orc_vec:
2202 return have_vec && TCG_TARGET_HAS_orc_vec;
2203 case INDEX_op_nand_vec:
2204 return have_vec && TCG_TARGET_HAS_nand_vec;
2205 case INDEX_op_nor_vec:
2206 return have_vec && TCG_TARGET_HAS_nor_vec;
2207 case INDEX_op_eqv_vec:
2208 return have_vec && TCG_TARGET_HAS_eqv_vec;
2209 case INDEX_op_mul_vec:
2210 return have_vec && TCG_TARGET_HAS_mul_vec;
2211 case INDEX_op_shli_vec:
2212 case INDEX_op_shri_vec:
2213 case INDEX_op_sari_vec:
2214 return have_vec && TCG_TARGET_HAS_shi_vec;
2215 case INDEX_op_shls_vec:
2216 case INDEX_op_shrs_vec:
2217 case INDEX_op_sars_vec:
2218 return have_vec && TCG_TARGET_HAS_shs_vec;
2219 case INDEX_op_shlv_vec:
2220 case INDEX_op_shrv_vec:
2221 case INDEX_op_sarv_vec:
2222 return have_vec && TCG_TARGET_HAS_shv_vec;
2223 case INDEX_op_rotli_vec:
2224 return have_vec && TCG_TARGET_HAS_roti_vec;
2225 case INDEX_op_rotls_vec:
2226 return have_vec && TCG_TARGET_HAS_rots_vec;
2227 case INDEX_op_rotlv_vec:
2228 case INDEX_op_rotrv_vec:
2229 return have_vec && TCG_TARGET_HAS_rotv_vec;
2230 case INDEX_op_ssadd_vec:
2231 case INDEX_op_usadd_vec:
2232 case INDEX_op_sssub_vec:
2233 case INDEX_op_ussub_vec:
2234 return have_vec && TCG_TARGET_HAS_sat_vec;
2235 case INDEX_op_smin_vec:
2236 case INDEX_op_umin_vec:
2237 case INDEX_op_smax_vec:
2238 case INDEX_op_umax_vec:
2239 return have_vec && TCG_TARGET_HAS_minmax_vec;
2240 case INDEX_op_bitsel_vec:
2241 return have_vec && TCG_TARGET_HAS_bitsel_vec;
2242 case INDEX_op_cmpsel_vec:
2243 return have_vec && TCG_TARGET_HAS_cmpsel_vec;
2244
2245 default:
2246 tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS);
2247 return true;
2248 }
2249 }
2250
2251 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs);
2252
2253 static void tcg_gen_callN(TCGHelperInfo *info, TCGTemp *ret, TCGTemp **args)
2254 {
2255 TCGv_i64 extend_free[MAX_CALL_IARGS];
2256 int n_extend = 0;
2257 TCGOp *op;
2258 int i, n, pi = 0, total_args;
2259
2260 if (unlikely(g_once_init_enter(HELPER_INFO_INIT(info)))) {
2261 init_call_layout(info);
2262 g_once_init_leave(HELPER_INFO_INIT(info), HELPER_INFO_INIT_VAL(info));
2263 }
2264
2265 total_args = info->nr_out + info->nr_in + 2;
2266 op = tcg_op_alloc(INDEX_op_call, total_args);
2267
2268 #ifdef CONFIG_PLUGIN
2269 /* Flag helpers that may affect guest state */
2270 if (tcg_ctx->plugin_insn &&
2271 !(info->flags & TCG_CALL_PLUGIN) &&
2272 !(info->flags & TCG_CALL_NO_SIDE_EFFECTS)) {
2273 tcg_ctx->plugin_insn->calls_helpers = true;
2274 }
2275 #endif
2276
2277 TCGOP_CALLO(op) = n = info->nr_out;
2278 switch (n) {
2279 case 0:
2280 tcg_debug_assert(ret == NULL);
2281 break;
2282 case 1:
2283 tcg_debug_assert(ret != NULL);
2284 op->args[pi++] = temp_arg(ret);
2285 break;
2286 case 2:
2287 case 4:
2288 tcg_debug_assert(ret != NULL);
2289 tcg_debug_assert(ret->base_type == ret->type + ctz32(n));
2290 tcg_debug_assert(ret->temp_subindex == 0);
2291 for (i = 0; i < n; ++i) {
2292 op->args[pi++] = temp_arg(ret + i);
2293 }
2294 break;
2295 default:
2296 g_assert_not_reached();
2297 }
2298
2299 TCGOP_CALLI(op) = n = info->nr_in;
2300 for (i = 0; i < n; i++) {
2301 const TCGCallArgumentLoc *loc = &info->in[i];
2302 TCGTemp *ts = args[loc->arg_idx] + loc->tmp_subindex;
2303
2304 switch (loc->kind) {
2305 case TCG_CALL_ARG_NORMAL:
2306 case TCG_CALL_ARG_BY_REF:
2307 case TCG_CALL_ARG_BY_REF_N:
2308 op->args[pi++] = temp_arg(ts);
2309 break;
2310
2311 case TCG_CALL_ARG_EXTEND_U:
2312 case TCG_CALL_ARG_EXTEND_S:
2313 {
2314 TCGv_i64 temp = tcg_temp_ebb_new_i64();
2315 TCGv_i32 orig = temp_tcgv_i32(ts);
2316
2317 if (loc->kind == TCG_CALL_ARG_EXTEND_S) {
2318 tcg_gen_ext_i32_i64(temp, orig);
2319 } else {
2320 tcg_gen_extu_i32_i64(temp, orig);
2321 }
2322 op->args[pi++] = tcgv_i64_arg(temp);
2323 extend_free[n_extend++] = temp;
2324 }
2325 break;
2326
2327 default:
2328 g_assert_not_reached();
2329 }
2330 }
2331 op->args[pi++] = (uintptr_t)info->func;
2332 op->args[pi++] = (uintptr_t)info;
2333 tcg_debug_assert(pi == total_args);
2334
2335 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link);
2336
2337 tcg_debug_assert(n_extend < ARRAY_SIZE(extend_free));
2338 for (i = 0; i < n_extend; ++i) {
2339 tcg_temp_free_i64(extend_free[i]);
2340 }
2341 }
2342
2343 void tcg_gen_call0(TCGHelperInfo *info, TCGTemp *ret)
2344 {
2345 tcg_gen_callN(info, ret, NULL);
2346 }
2347
2348 void tcg_gen_call1(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1)
2349 {
2350 tcg_gen_callN(info, ret, &t1);
2351 }
2352
2353 void tcg_gen_call2(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, TCGTemp *t2)
2354 {
2355 TCGTemp *args[2] = { t1, t2 };
2356 tcg_gen_callN(info, ret, args);
2357 }
2358
2359 void tcg_gen_call3(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1,
2360 TCGTemp *t2, TCGTemp *t3)
2361 {
2362 TCGTemp *args[3] = { t1, t2, t3 };
2363 tcg_gen_callN(info, ret, args);
2364 }
2365
2366 void tcg_gen_call4(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1,
2367 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4)
2368 {
2369 TCGTemp *args[4] = { t1, t2, t3, t4 };
2370 tcg_gen_callN(info, ret, args);
2371 }
2372
2373 void tcg_gen_call5(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1,
2374 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, TCGTemp *t5)
2375 {
2376 TCGTemp *args[5] = { t1, t2, t3, t4, t5 };
2377 tcg_gen_callN(info, ret, args);
2378 }
2379
2380 void tcg_gen_call6(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, TCGTemp *t2,
2381 TCGTemp *t3, TCGTemp *t4, TCGTemp *t5, TCGTemp *t6)
2382 {
2383 TCGTemp *args[6] = { t1, t2, t3, t4, t5, t6 };
2384 tcg_gen_callN(info, ret, args);
2385 }
2386
2387 void tcg_gen_call7(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1,
2388 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4,
2389 TCGTemp *t5, TCGTemp *t6, TCGTemp *t7)
2390 {
2391 TCGTemp *args[7] = { t1, t2, t3, t4, t5, t6, t7 };
2392 tcg_gen_callN(info, ret, args);
2393 }
2394
2395 static void tcg_reg_alloc_start(TCGContext *s)
2396 {
2397 int i, n;
2398
2399 for (i = 0, n = s->nb_temps; i < n; i++) {
2400 TCGTemp *ts = &s->temps[i];
2401 TCGTempVal val = TEMP_VAL_MEM;
2402
2403 switch (ts->kind) {
2404 case TEMP_CONST:
2405 val = TEMP_VAL_CONST;
2406 break;
2407 case TEMP_FIXED:
2408 val = TEMP_VAL_REG;
2409 break;
2410 case TEMP_GLOBAL:
2411 break;
2412 case TEMP_EBB:
2413 val = TEMP_VAL_DEAD;
2414 /* fall through */
2415 case TEMP_TB:
2416 ts->mem_allocated = 0;
2417 break;
2418 default:
2419 g_assert_not_reached();
2420 }
2421 ts->val_type = val;
2422 }
2423
2424 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp));
2425 }
2426
2427 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size,
2428 TCGTemp *ts)
2429 {
2430 int idx = temp_idx(ts);
2431
2432 switch (ts->kind) {
2433 case TEMP_FIXED:
2434 case TEMP_GLOBAL:
2435 pstrcpy(buf, buf_size, ts->name);
2436 break;
2437 case TEMP_TB:
2438 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals);
2439 break;
2440 case TEMP_EBB:
2441 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals);
2442 break;
2443 case TEMP_CONST:
2444 switch (ts->type) {
2445 case TCG_TYPE_I32:
2446 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val);
2447 break;
2448 #if TCG_TARGET_REG_BITS > 32
2449 case TCG_TYPE_I64:
2450 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val);
2451 break;
2452 #endif
2453 case TCG_TYPE_V64:
2454 case TCG_TYPE_V128:
2455 case TCG_TYPE_V256:
2456 snprintf(buf, buf_size, "v%d$0x%" PRIx64,
2457 64 << (ts->type - TCG_TYPE_V64), ts->val);
2458 break;
2459 default:
2460 g_assert_not_reached();
2461 }
2462 break;
2463 }
2464 return buf;
2465 }
2466
2467 static char *tcg_get_arg_str(TCGContext *s, char *buf,
2468 int buf_size, TCGArg arg)
2469 {
2470 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg));
2471 }
2472
2473 static const char * const cond_name[] =
2474 {
2475 [TCG_COND_NEVER] = "never",
2476 [TCG_COND_ALWAYS] = "always",
2477 [TCG_COND_EQ] = "eq",
2478 [TCG_COND_NE] = "ne",
2479 [TCG_COND_LT] = "lt",
2480 [TCG_COND_GE] = "ge",
2481 [TCG_COND_LE] = "le",
2482 [TCG_COND_GT] = "gt",
2483 [TCG_COND_LTU] = "ltu",
2484 [TCG_COND_GEU] = "geu",
2485 [TCG_COND_LEU] = "leu",
2486 [TCG_COND_GTU] = "gtu",
2487 [TCG_COND_TSTEQ] = "tsteq",
2488 [TCG_COND_TSTNE] = "tstne",
2489 };
2490
2491 static const char * const ldst_name[(MO_BSWAP | MO_SSIZE) + 1] =
2492 {
2493 [MO_UB] = "ub",
2494 [MO_SB] = "sb",
2495 [MO_LEUW] = "leuw",
2496 [MO_LESW] = "lesw",
2497 [MO_LEUL] = "leul",
2498 [MO_LESL] = "lesl",
2499 [MO_LEUQ] = "leq",
2500 [MO_BEUW] = "beuw",
2501 [MO_BESW] = "besw",
2502 [MO_BEUL] = "beul",
2503 [MO_BESL] = "besl",
2504 [MO_BEUQ] = "beq",
2505 [MO_128 + MO_BE] = "beo",
2506 [MO_128 + MO_LE] = "leo",
2507 };
2508
2509 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = {
2510 [MO_UNALN >> MO_ASHIFT] = "un+",
2511 [MO_ALIGN >> MO_ASHIFT] = "al+",
2512 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+",
2513 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+",
2514 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+",
2515 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+",
2516 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+",
2517 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+",
2518 };
2519
2520 static const char * const atom_name[(MO_ATOM_MASK >> MO_ATOM_SHIFT) + 1] = {
2521 [MO_ATOM_IFALIGN >> MO_ATOM_SHIFT] = "",
2522 [MO_ATOM_IFALIGN_PAIR >> MO_ATOM_SHIFT] = "pair+",
2523 [MO_ATOM_WITHIN16 >> MO_ATOM_SHIFT] = "w16+",
2524 [MO_ATOM_WITHIN16_PAIR >> MO_ATOM_SHIFT] = "w16p+",
2525 [MO_ATOM_SUBALIGN >> MO_ATOM_SHIFT] = "sub+",
2526 [MO_ATOM_NONE >> MO_ATOM_SHIFT] = "noat+",
2527 };
2528
2529 static const char bswap_flag_name[][6] = {
2530 [TCG_BSWAP_IZ] = "iz",
2531 [TCG_BSWAP_OZ] = "oz",
2532 [TCG_BSWAP_OS] = "os",
2533 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz",
2534 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os",
2535 };
2536
2537 static inline bool tcg_regset_single(TCGRegSet d)
2538 {
2539 return (d & (d - 1)) == 0;
2540 }
2541
2542 static inline TCGReg tcg_regset_first(TCGRegSet d)
2543 {
2544 if (TCG_TARGET_NB_REGS <= 32) {
2545 return ctz32(d);
2546 } else {
2547 return ctz64(d);
2548 }
2549 }
2550
2551 /* Return only the number of characters output -- no error return. */
2552 #define ne_fprintf(...) \
2553 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; })
2554
2555 static void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs)
2556 {
2557 char buf[128];
2558 TCGOp *op;
2559
2560 QTAILQ_FOREACH(op, &s->ops, link) {
2561 int i, k, nb_oargs, nb_iargs, nb_cargs;
2562 const TCGOpDef *def;
2563 TCGOpcode c;
2564 int col = 0;
2565
2566 c = op->opc;
2567 def = &tcg_op_defs[c];
2568
2569 if (c == INDEX_op_insn_start) {
2570 nb_oargs = 0;
2571 col += ne_fprintf(f, "\n ----");
2572
2573 for (i = 0, k = s->insn_start_words; i < k; ++i) {
2574 col += ne_fprintf(f, " %016" PRIx64,
2575 tcg_get_insn_start_param(op, i));
2576 }
2577 } else if (c == INDEX_op_call) {
2578 const TCGHelperInfo *info = tcg_call_info(op);
2579 void *func = tcg_call_func(op);
2580
2581 /* variable number of arguments */
2582 nb_oargs = TCGOP_CALLO(op);
2583 nb_iargs = TCGOP_CALLI(op);
2584 nb_cargs = def->nb_cargs;
2585
2586 col += ne_fprintf(f, " %s ", def->name);
2587
2588 /*
2589 * Print the function name from TCGHelperInfo, if available.
2590 * Note that plugins have a template function for the info,
2591 * but the actual function pointer comes from the plugin.
2592 */
2593 if (func == info->func) {
2594 col += ne_fprintf(f, "%s", info->name);
2595 } else {
2596 col += ne_fprintf(f, "plugin(%p)", func);
2597 }
2598
2599 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs);
2600 for (i = 0; i < nb_oargs; i++) {
2601 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf),
2602 op->args[i]));
2603 }
2604 for (i = 0; i < nb_iargs; i++) {
2605 TCGArg arg = op->args[nb_oargs + i];
2606 const char *t = tcg_get_arg_str(s, buf, sizeof(buf), arg);
2607 col += ne_fprintf(f, ",%s", t);
2608 }
2609 } else {
2610 col += ne_fprintf(f, " %s ", def->name);
2611
2612 nb_oargs = def->nb_oargs;
2613 nb_iargs = def->nb_iargs;
2614 nb_cargs = def->nb_cargs;
2615
2616 if (def->flags & TCG_OPF_VECTOR) {
2617 col += ne_fprintf(f, "v%d,e%d,", 64 << TCGOP_VECL(op),
2618 8 << TCGOP_VECE(op));
2619 }
2620
2621 k = 0;
2622 for (i = 0; i < nb_oargs; i++) {
2623 const char *sep = k ? "," : "";
2624 col += ne_fprintf(f, "%s%s", sep,
2625 tcg_get_arg_str(s, buf, sizeof(buf),
2626 op->args[k++]));
2627 }
2628 for (i = 0; i < nb_iargs; i++) {
2629 const char *sep = k ? "," : "";
2630 col += ne_fprintf(f, "%s%s", sep,
2631 tcg_get_arg_str(s, buf, sizeof(buf),
2632 op->args[k++]));
2633 }
2634 switch (c) {
2635 case INDEX_op_brcond_i32:
2636 case INDEX_op_setcond_i32:
2637 case INDEX_op_negsetcond_i32:
2638 case INDEX_op_movcond_i32:
2639 case INDEX_op_brcond2_i32:
2640 case INDEX_op_setcond2_i32:
2641 case INDEX_op_brcond_i64:
2642 case INDEX_op_setcond_i64:
2643 case INDEX_op_negsetcond_i64:
2644 case INDEX_op_movcond_i64:
2645 case INDEX_op_cmp_vec:
2646 case INDEX_op_cmpsel_vec:
2647 if (op->args[k] < ARRAY_SIZE(cond_name)
2648 && cond_name[op->args[k]]) {
2649 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]);
2650 } else {
2651 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]);
2652 }
2653 i = 1;
2654 break;
2655 case INDEX_op_qemu_ld_a32_i32:
2656 case INDEX_op_qemu_ld_a64_i32:
2657 case INDEX_op_qemu_st_a32_i32:
2658 case INDEX_op_qemu_st_a64_i32:
2659 case INDEX_op_qemu_st8_a32_i32:
2660 case INDEX_op_qemu_st8_a64_i32:
2661 case INDEX_op_qemu_ld_a32_i64:
2662 case INDEX_op_qemu_ld_a64_i64:
2663 case INDEX_op_qemu_st_a32_i64:
2664 case INDEX_op_qemu_st_a64_i64:
2665 case INDEX_op_qemu_ld_a32_i128:
2666 case INDEX_op_qemu_ld_a64_i128:
2667 case INDEX_op_qemu_st_a32_i128:
2668 case INDEX_op_qemu_st_a64_i128:
2669 {
2670 const char *s_al, *s_op, *s_at;
2671 MemOpIdx oi = op->args[k++];
2672 MemOp mop = get_memop(oi);
2673 unsigned ix = get_mmuidx(oi);
2674
2675 s_al = alignment_name[(mop & MO_AMASK) >> MO_ASHIFT];
2676 s_op = ldst_name[mop & (MO_BSWAP | MO_SSIZE)];
2677 s_at = atom_name[(mop & MO_ATOM_MASK) >> MO_ATOM_SHIFT];
2678 mop &= ~(MO_AMASK | MO_BSWAP | MO_SSIZE | MO_ATOM_MASK);
2679
2680 /* If all fields are accounted for, print symbolically. */
2681 if (!mop && s_al && s_op && s_at) {
2682 col += ne_fprintf(f, ",%s%s%s,%u",
2683 s_at, s_al, s_op, ix);
2684 } else {
2685 mop = get_memop(oi);
2686 col += ne_fprintf(f, ",$0x%x,%u", mop, ix);
2687 }
2688 i = 1;
2689 }
2690 break;
2691 case INDEX_op_bswap16_i32:
2692 case INDEX_op_bswap16_i64:
2693 case INDEX_op_bswap32_i32:
2694 case INDEX_op_bswap32_i64:
2695 case INDEX_op_bswap64_i64:
2696 {
2697 TCGArg flags = op->args[k];
2698 const char *name = NULL;
2699
2700 if (flags < ARRAY_SIZE(bswap_flag_name)) {
2701 name = bswap_flag_name[flags];
2702 }
2703 if (name) {
2704 col += ne_fprintf(f, ",%s", name);
2705 } else {
2706 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags);
2707 }
2708 i = k = 1;
2709 }
2710 break;
2711 default:
2712 i = 0;
2713 break;
2714 }
2715 switch (c) {
2716 case INDEX_op_set_label:
2717 case INDEX_op_br:
2718 case INDEX_op_brcond_i32:
2719 case INDEX_op_brcond_i64:
2720 case INDEX_op_brcond2_i32:
2721 col += ne_fprintf(f, "%s$L%d", k ? "," : "",
2722 arg_label(op->args[k])->id);
2723 i++, k++;
2724 break;
2725 case INDEX_op_mb:
2726 {
2727 TCGBar membar = op->args[k];
2728 const char *b_op, *m_op;
2729
2730 switch (membar & TCG_BAR_SC) {
2731 case 0:
2732 b_op = "none";
2733 break;
2734 case TCG_BAR_LDAQ:
2735 b_op = "acq";
2736 break;
2737 case TCG_BAR_STRL:
2738 b_op = "rel";
2739 break;
2740 case TCG_BAR_SC:
2741 b_op = "seq";
2742 break;
2743 default:
2744 g_assert_not_reached();
2745 }
2746
2747 switch (membar & TCG_MO_ALL) {
2748 case 0:
2749 m_op = "none";
2750 break;
2751 case TCG_MO_LD_LD:
2752 m_op = "rr";
2753 break;
2754 case TCG_MO_LD_ST:
2755 m_op = "rw";
2756 break;
2757 case TCG_MO_ST_LD:
2758 m_op = "wr";
2759 break;
2760 case TCG_MO_ST_ST:
2761 m_op = "ww";
2762 break;
2763 case TCG_MO_LD_LD | TCG_MO_LD_ST:
2764 m_op = "rr+rw";
2765 break;
2766 case TCG_MO_LD_LD | TCG_MO_ST_LD:
2767 m_op = "rr+wr";
2768 break;
2769 case TCG_MO_LD_LD | TCG_MO_ST_ST:
2770 m_op = "rr+ww";
2771 break;
2772 case TCG_MO_LD_ST | TCG_MO_ST_LD:
2773 m_op = "rw+wr";
2774 break;
2775 case TCG_MO_LD_ST | TCG_MO_ST_ST:
2776 m_op = "rw+ww";
2777 break;
2778 case TCG_MO_ST_LD | TCG_MO_ST_ST:
2779 m_op = "wr+ww";
2780 break;
2781 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_LD:
2782 m_op = "rr+rw+wr";
2783 break;
2784 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_ST:
2785 m_op = "rr+rw+ww";
2786 break;
2787 case TCG_MO_LD_LD | TCG_MO_ST_LD | TCG_MO_ST_ST:
2788 m_op = "rr+wr+ww";
2789 break;
2790 case TCG_MO_LD_ST | TCG_MO_ST_LD | TCG_MO_ST_ST:
2791 m_op = "rw+wr+ww";
2792 break;
2793 case TCG_MO_ALL:
2794 m_op = "all";
2795 break;
2796 default:
2797 g_assert_not_reached();
2798 }
2799
2800 col += ne_fprintf(f, "%s%s:%s", (k ? "," : ""), b_op, m_op);
2801 i++, k++;
2802 }
2803 break;
2804 default:
2805 break;
2806 }
2807 for (; i < nb_cargs; i++, k++) {
2808 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "",
2809 op->args[k]);
2810 }
2811 }
2812
2813 if (have_prefs || op->life) {
2814 for (; col < 40; ++col) {
2815 putc(' ', f);
2816 }
2817 }
2818
2819 if (op->life) {
2820 unsigned life = op->life;
2821
2822 if (life & (SYNC_ARG * 3)) {
2823 ne_fprintf(f, " sync:");
2824 for (i = 0; i < 2; ++i) {
2825 if (life & (SYNC_ARG << i)) {
2826 ne_fprintf(f, " %d", i);
2827 }
2828 }
2829 }
2830 life /= DEAD_ARG;
2831 if (life) {
2832 ne_fprintf(f, " dead:");
2833 for (i = 0; life; ++i, life >>= 1) {
2834 if (life & 1) {
2835 ne_fprintf(f, " %d", i);
2836 }
2837 }
2838 }
2839 }
2840
2841 if (have_prefs) {
2842 for (i = 0; i < nb_oargs; ++i) {
2843 TCGRegSet set = output_pref(op, i);
2844
2845 if (i == 0) {
2846 ne_fprintf(f, " pref=");
2847 } else {
2848 ne_fprintf(f, ",");
2849 }
2850 if (set == 0) {
2851 ne_fprintf(f, "none");
2852 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) {
2853 ne_fprintf(f, "all");
2854 #ifdef CONFIG_DEBUG_TCG
2855 } else if (tcg_regset_single(set)) {
2856 TCGReg reg = tcg_regset_first(set);
2857 ne_fprintf(f, "%s", tcg_target_reg_names[reg]);
2858 #endif
2859 } else if (TCG_TARGET_NB_REGS <= 32) {
2860 ne_fprintf(f, "0x%x", (uint32_t)set);
2861 } else {
2862 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set);
2863 }
2864 }
2865 }
2866
2867 putc('\n', f);
2868 }
2869 }
2870
2871 /* we give more priority to constraints with less registers */
2872 static int get_constraint_priority(const TCGOpDef *def, int k)
2873 {
2874 const TCGArgConstraint *arg_ct = &def->args_ct[k];
2875 int n = ctpop64(arg_ct->regs);
2876
2877 /*
2878 * Sort constraints of a single register first, which includes output
2879 * aliases (which must exactly match the input already allocated).
2880 */
2881 if (n == 1 || arg_ct->oalias) {
2882 return INT_MAX;
2883 }
2884
2885 /*
2886 * Sort register pairs next, first then second immediately after.
2887 * Arbitrarily sort multiple pairs by the index of the first reg;
2888 * there shouldn't be many pairs.
2889 */
2890 switch (arg_ct->pair) {
2891 case 1:
2892 case 3:
2893 return (k + 1) * 2;
2894 case 2:
2895 return (arg_ct->pair_index + 1) * 2 - 1;
2896 }
2897
2898 /* Finally, sort by decreasing register count. */
2899 assert(n > 1);
2900 return -n;
2901 }
2902
2903 /* sort from highest priority to lowest */
2904 static void sort_constraints(TCGOpDef *def, int start, int n)
2905 {
2906 int i, j;
2907 TCGArgConstraint *a = def->args_ct;
2908
2909 for (i = 0; i < n; i++) {
2910 a[start + i].sort_index = start + i;
2911 }
2912 if (n <= 1) {
2913 return;
2914 }
2915 for (i = 0; i < n - 1; i++) {
2916 for (j = i + 1; j < n; j++) {
2917 int p1 = get_constraint_priority(def, a[start + i].sort_index);
2918 int p2 = get_constraint_priority(def, a[start + j].sort_index);
2919 if (p1 < p2) {
2920 int tmp = a[start + i].sort_index;
2921 a[start + i].sort_index = a[start + j].sort_index;
2922 a[start + j].sort_index = tmp;
2923 }
2924 }
2925 }
2926 }
2927
2928 static void process_op_defs(TCGContext *s)
2929 {
2930 TCGOpcode op;
2931
2932 for (op = 0; op < NB_OPS; op++) {
2933 TCGOpDef *def = &tcg_op_defs[op];
2934 const TCGTargetOpDef *tdefs;
2935 bool saw_alias_pair = false;
2936 int i, o, i2, o2, nb_args;
2937
2938 if (def->flags & TCG_OPF_NOT_PRESENT) {
2939 continue;
2940 }
2941
2942 nb_args = def->nb_iargs + def->nb_oargs;
2943 if (nb_args == 0) {
2944 continue;
2945 }
2946
2947 /*
2948 * Macro magic should make it impossible, but double-check that
2949 * the array index is in range. Since the signness of an enum
2950 * is implementation defined, force the result to unsigned.
2951 */
2952 unsigned con_set = tcg_target_op_def(op);
2953 tcg_debug_assert(con_set < ARRAY_SIZE(constraint_sets));
2954 tdefs = &constraint_sets[con_set];
2955
2956 for (i = 0; i < nb_args; i++) {
2957 const char *ct_str = tdefs->args_ct_str[i];
2958 bool input_p = i >= def->nb_oargs;
2959
2960 /* Incomplete TCGTargetOpDef entry. */
2961 tcg_debug_assert(ct_str != NULL);
2962
2963 switch (*ct_str) {
2964 case '0' ... '9':
2965 o = *ct_str - '0';
2966 tcg_debug_assert(input_p);
2967 tcg_debug_assert(o < def->nb_oargs);
2968 tcg_debug_assert(def->args_ct[o].regs != 0);
2969 tcg_debug_assert(!def->args_ct[o].oalias);
2970 def->args_ct[i] = def->args_ct[o];
2971 /* The output sets oalias. */
2972 def->args_ct[o].oalias = 1;
2973 def->args_ct[o].alias_index = i;
2974 /* The input sets ialias. */
2975 def->args_ct[i].ialias = 1;
2976 def->args_ct[i].alias_index = o;
2977 if (def->args_ct[i].pair) {
2978 saw_alias_pair = true;
2979 }
2980 tcg_debug_assert(ct_str[1] == '\0');
2981 continue;
2982
2983 case '&':
2984 tcg_debug_assert(!input_p);
2985 def->args_ct[i].newreg = true;
2986 ct_str++;
2987 break;
2988
2989 case 'p': /* plus */
2990 /* Allocate to the register after the previous. */
2991 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0));
2992 o = i - 1;
2993 tcg_debug_assert(!def->args_ct[o].pair);
2994 tcg_debug_assert(!def->args_ct[o].ct);
2995 def->args_ct[i] = (TCGArgConstraint){
2996 .pair = 2,
2997 .pair_index = o,
2998 .regs = def->args_ct[o].regs << 1,
2999 .newreg = def->args_ct[o].newreg,
3000 };
3001 def->args_ct[o].pair = 1;
3002 def->args_ct[o].pair_index = i;
3003 tcg_debug_assert(ct_str[1] == '\0');
3004 continue;
3005
3006 case 'm': /* minus */
3007 /* Allocate to the register before the previous. */
3008 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0));
3009 o = i - 1;
3010 tcg_debug_assert(!def->args_ct[o].pair);
3011 tcg_debug_assert(!def->args_ct[o].ct);
3012 def->args_ct[i] = (TCGArgConstraint){
3013 .pair = 1,
3014 .pair_index = o,
3015 .regs = def->args_ct[o].regs >> 1,
3016 .newreg = def->args_ct[o].newreg,
3017 };
3018 def->args_ct[o].pair = 2;
3019 def->args_ct[o].pair_index = i;
3020 tcg_debug_assert(ct_str[1] == '\0');
3021 continue;
3022 }
3023
3024 do {
3025 switch (*ct_str) {
3026 case 'i':
3027 def->args_ct[i].ct |= TCG_CT_CONST;
3028 break;
3029
3030 /* Include all of the target-specific constraints. */
3031
3032 #undef CONST
3033 #define CONST(CASE, MASK) \
3034 case CASE: def->args_ct[i].ct |= MASK; break;
3035 #define REGS(CASE, MASK) \
3036 case CASE: def->args_ct[i].regs |= MASK; break;
3037
3038 #include "tcg-target-con-str.h"
3039
3040 #undef REGS
3041 #undef CONST
3042 default:
3043 case '0' ... '9':
3044 case '&':
3045 case 'p':
3046 case 'm':
3047 /* Typo in TCGTargetOpDef constraint. */
3048 g_assert_not_reached();
3049 }
3050 } while (*++ct_str != '\0');
3051 }
3052
3053 /* TCGTargetOpDef entry with too much information? */
3054 tcg_debug_assert(i == TCG_MAX_OP_ARGS || tdefs->args_ct_str[i] == NULL);
3055
3056 /*
3057 * Fix up output pairs that are aliased with inputs.
3058 * When we created the alias, we copied pair from the output.
3059 * There are three cases:
3060 * (1a) Pairs of inputs alias pairs of outputs.
3061 * (1b) One input aliases the first of a pair of outputs.
3062 * (2) One input aliases the second of a pair of outputs.
3063 *
3064 * Case 1a is handled by making sure that the pair_index'es are
3065 * properly updated so that they appear the same as a pair of inputs.
3066 *
3067 * Case 1b is handled by setting the pair_index of the input to
3068 * itself, simply so it doesn't point to an unrelated argument.
3069 * Since we don't encounter the "second" during the input allocation
3070 * phase, nothing happens with the second half of the input pair.
3071 *
3072 * Case 2 is handled by setting the second input to pair=3, the
3073 * first output to pair=3, and the pair_index'es to match.
3074 */
3075 if (saw_alias_pair) {
3076 for (i = def->nb_oargs; i < nb_args; i++) {
3077 /*
3078 * Since [0-9pm] must be alone in the constraint string,
3079 * the only way they can both be set is if the pair comes
3080 * from the output alias.
3081 */
3082 if (!def->args_ct[i].ialias) {
3083 continue;
3084 }
3085 switch (def->args_ct[i].pair) {
3086 case 0:
3087 break;
3088 case 1:
3089 o = def->args_ct[i].alias_index;
3090 o2 = def->args_ct[o].pair_index;
3091 tcg_debug_assert(def->args_ct[o].pair == 1);
3092 tcg_debug_assert(def->args_ct[o2].pair == 2);
3093 if (def->args_ct[o2].oalias) {
3094 /* Case 1a */
3095 i2 = def->args_ct[o2].alias_index;
3096 tcg_debug_assert(def->args_ct[i2].pair == 2);
3097 def->args_ct[i2].pair_index = i;
3098 def->args_ct[i].pair_index = i2;
3099 } else {
3100 /* Case 1b */
3101 def->args_ct[i].pair_index = i;
3102 }
3103 break;
3104 case 2:
3105 o = def->args_ct[i].alias_index;
3106 o2 = def->args_ct[o].pair_index;
3107 tcg_debug_assert(def->args_ct[o].pair == 2);
3108 tcg_debug_assert(def->args_ct[o2].pair == 1);
3109 if (def->args_ct[o2].oalias) {
3110 /* Case 1a */
3111 i2 = def->args_ct[o2].alias_index;
3112 tcg_debug_assert(def->args_ct[i2].pair == 1);
3113 def->args_ct[i2].pair_index = i;
3114 def->args_ct[i].pair_index = i2;
3115 } else {
3116 /* Case 2 */
3117 def->args_ct[i].pair = 3;
3118 def->args_ct[o2].pair = 3;
3119 def->args_ct[i].pair_index = o2;
3120 def->args_ct[o2].pair_index = i;
3121 }
3122 break;
3123 default:
3124 g_assert_not_reached();
3125 }
3126 }
3127 }
3128
3129 /* sort the constraints (XXX: this is just an heuristic) */
3130 sort_constraints(def, 0, def->nb_oargs);
3131 sort_constraints(def, def->nb_oargs, def->nb_iargs);
3132 }
3133 }
3134
3135 static void remove_label_use(TCGOp *op, int idx)
3136 {
3137 TCGLabel *label = arg_label(op->args[idx]);
3138 TCGLabelUse *use;
3139
3140 QSIMPLEQ_FOREACH(use, &label->branches, next) {
3141 if (use->op == op) {
3142 QSIMPLEQ_REMOVE(&label->branches, use, TCGLabelUse, next);
3143 return;
3144 }
3145 }
3146 g_assert_not_reached();
3147 }
3148
3149 void tcg_op_remove(TCGContext *s, TCGOp *op)
3150 {
3151 switch (op->opc) {
3152 case INDEX_op_br:
3153 remove_label_use(op, 0);
3154 break;
3155 case INDEX_op_brcond_i32:
3156 case INDEX_op_brcond_i64:
3157 remove_label_use(op, 3);
3158 break;
3159 case INDEX_op_brcond2_i32:
3160 remove_label_use(op, 5);
3161 break;
3162 default:
3163 break;
3164 }
3165
3166 QTAILQ_REMOVE(&s->ops, op, link);
3167 QTAILQ_INSERT_TAIL(&s->free_ops, op, link);
3168 s->nb_ops--;
3169 }
3170
3171 void tcg_remove_ops_after(TCGOp *op)
3172 {
3173 TCGContext *s = tcg_ctx;
3174
3175 while (true) {
3176 TCGOp *last = tcg_last_op();
3177 if (last == op) {
3178 return;
3179 }
3180 tcg_op_remove(s, last);
3181 }
3182 }
3183
3184 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs)
3185 {
3186 TCGContext *s = tcg_ctx;
3187 TCGOp *op = NULL;
3188
3189 if (unlikely(!QTAILQ_EMPTY(&s->free_ops))) {
3190 QTAILQ_FOREACH(op, &s->free_ops, link) {
3191 if (nargs <= op->nargs) {
3192 QTAILQ_REMOVE(&s->free_ops, op, link);
3193 nargs = op->nargs;
3194 goto found;
3195 }
3196 }
3197 }
3198
3199 /* Most opcodes have 3 or 4 operands: reduce fragmentation. */
3200 nargs = MAX(4, nargs);
3201 op = tcg_malloc(sizeof(TCGOp) + sizeof(TCGArg) * nargs);
3202
3203 found:
3204 memset(op, 0, offsetof(TCGOp, link));
3205 op->opc = opc;
3206 op->nargs = nargs;
3207
3208 /* Check for bitfield overflow. */
3209 tcg_debug_assert(op->nargs == nargs);
3210
3211 s->nb_ops++;
3212 return op;
3213 }
3214
3215 TCGOp *tcg_emit_op(TCGOpcode opc, unsigned nargs)
3216 {
3217 TCGOp *op = tcg_op_alloc(opc, nargs);
3218 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link);
3219 return op;
3220 }
3221
3222 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op,
3223 TCGOpcode opc, unsigned nargs)
3224 {
3225 TCGOp *new_op = tcg_op_alloc(opc, nargs);
3226 QTAILQ_INSERT_BEFORE(old_op, new_op, link);
3227 return new_op;
3228 }
3229
3230 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op,
3231 TCGOpcode opc, unsigned nargs)
3232 {
3233 TCGOp *new_op = tcg_op_alloc(opc, nargs);
3234 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link);
3235 return new_op;
3236 }
3237
3238 static void move_label_uses(TCGLabel *to, TCGLabel *from)
3239 {
3240 TCGLabelUse *u;
3241
3242 QSIMPLEQ_FOREACH(u, &from->branches, next) {
3243 TCGOp *op = u->op;
3244 switch (op->opc) {
3245 case INDEX_op_br:
3246 op->args[0] = label_arg(to);
3247 break;
3248 case INDEX_op_brcond_i32:
3249 case INDEX_op_brcond_i64:
3250 op->args[3] = label_arg(to);
3251 break;
3252 case INDEX_op_brcond2_i32:
3253 op->args[5] = label_arg(to);
3254 break;
3255 default:
3256 g_assert_not_reached();
3257 }
3258 }
3259
3260 QSIMPLEQ_CONCAT(&to->branches, &from->branches);
3261 }
3262
3263 /* Reachable analysis : remove unreachable code. */
3264 static void __attribute__((noinline))
3265 reachable_code_pass(TCGContext *s)
3266 {
3267 TCGOp *op, *op_next, *op_prev;
3268 bool dead = false;
3269
3270 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
3271 bool remove = dead;
3272 TCGLabel *label;
3273
3274 switch (op->opc) {
3275 case INDEX_op_set_label:
3276 label = arg_label(op->args[0]);
3277
3278 /*
3279 * Note that the first op in the TB is always a load,
3280 * so there is always something before a label.
3281 */
3282 op_prev = QTAILQ_PREV(op, link);
3283
3284 /*
3285 * If we find two sequential labels, move all branches to
3286 * reference the second label and remove the first label.
3287 * Do this before branch to next optimization, so that the
3288 * middle label is out of the way.
3289 */
3290 if (op_prev->opc == INDEX_op_set_label) {
3291 move_label_uses(label, arg_label(op_prev->args[0]));
3292 tcg_op_remove(s, op_prev);
3293 op_prev = QTAILQ_PREV(op, link);
3294 }
3295
3296 /*
3297 * Optimization can fold conditional branches to unconditional.
3298 * If we find a label which is preceded by an unconditional
3299 * branch to next, remove the branch. We couldn't do this when
3300 * processing the branch because any dead code between the branch
3301 * and label had not yet been removed.
3302 */
3303 if (op_prev->opc == INDEX_op_br &&
3304 label == arg_label(op_prev->args[0])) {
3305 tcg_op_remove(s, op_prev);
3306 /* Fall through means insns become live again. */
3307 dead = false;
3308 }
3309
3310 if (QSIMPLEQ_EMPTY(&label->branches)) {
3311 /*
3312 * While there is an occasional backward branch, virtually
3313 * all branches generated by the translators are forward.
3314 * Which means that generally we will have already removed
3315 * all references to the label that will be, and there is
3316 * little to be gained by iterating.
3317 */
3318 remove = true;
3319 } else {
3320 /* Once we see a label, insns become live again. */
3321 dead = false;
3322 remove = false;
3323 }
3324 break;
3325
3326 case INDEX_op_br:
3327 case INDEX_op_exit_tb:
3328 case INDEX_op_goto_ptr:
3329 /* Unconditional branches; everything following is dead. */
3330 dead = true;
3331 break;
3332
3333 case INDEX_op_call:
3334 /* Notice noreturn helper calls, raising exceptions. */
3335 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) {
3336 dead = true;
3337 }
3338 break;
3339
3340 case INDEX_op_insn_start:
3341 /* Never remove -- we need to keep these for unwind. */
3342 remove = false;
3343 break;
3344
3345 default:
3346 break;
3347 }
3348
3349 if (remove) {
3350 tcg_op_remove(s, op);
3351 }
3352 }
3353 }
3354
3355 #define TS_DEAD 1
3356 #define TS_MEM 2
3357
3358 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n)))
3359 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n)))
3360
3361 /* For liveness_pass_1, the register preferences for a given temp. */
3362 static inline TCGRegSet *la_temp_pref(TCGTemp *ts)
3363 {
3364 return ts->state_ptr;
3365 }
3366
3367 /* For liveness_pass_1, reset the preferences for a given temp to the
3368 * maximal regset for its type.
3369 */
3370 static inline void la_reset_pref(TCGTemp *ts)
3371 {
3372 *la_temp_pref(ts)
3373 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]);
3374 }
3375
3376 /* liveness analysis: end of function: all temps are dead, and globals
3377 should be in memory. */
3378 static void la_func_end(TCGContext *s, int ng, int nt)
3379 {
3380 int i;
3381
3382 for (i = 0; i < ng; ++i) {
3383 s->temps[i].state = TS_DEAD | TS_MEM;
3384 la_reset_pref(&s->temps[i]);
3385 }
3386 for (i = ng; i < nt; ++i) {
3387 s->temps[i].state = TS_DEAD;
3388 la_reset_pref(&s->temps[i]);
3389 }
3390 }
3391
3392 /* liveness analysis: end of basic block: all temps are dead, globals
3393 and local temps should be in memory. */
3394 static void la_bb_end(TCGContext *s, int ng, int nt)
3395 {
3396 int i;
3397
3398 for (i = 0; i < nt; ++i) {
3399 TCGTemp *ts = &s->temps[i];
3400 int state;
3401
3402 switch (ts->kind) {
3403 case TEMP_FIXED:
3404 case TEMP_GLOBAL:
3405 case TEMP_TB:
3406 state = TS_DEAD | TS_MEM;
3407 break;
3408 case TEMP_EBB:
3409 case TEMP_CONST:
3410 state = TS_DEAD;
3411 break;
3412 default:
3413 g_assert_not_reached();
3414 }
3415 ts->state = state;
3416 la_reset_pref(ts);
3417 }
3418 }
3419
3420 /* liveness analysis: sync globals back to memory. */
3421 static void la_global_sync(TCGContext *s, int ng)
3422 {
3423 int i;
3424
3425 for (i = 0; i < ng; ++i) {
3426 int state = s->temps[i].state;
3427 s->temps[i].state = state | TS_MEM;
3428 if (state == TS_DEAD) {
3429 /* If the global was previously dead, reset prefs. */
3430 la_reset_pref(&s->temps[i]);
3431 }
3432 }
3433 }
3434
3435 /*
3436 * liveness analysis: conditional branch: all temps are dead unless
3437 * explicitly live-across-conditional-branch, globals and local temps
3438 * should be synced.
3439 */
3440 static void la_bb_sync(TCGContext *s, int ng, int nt)
3441 {
3442 la_global_sync(s, ng);
3443
3444 for (int i = ng; i < nt; ++i) {
3445 TCGTemp *ts = &s->temps[i];
3446 int state;
3447
3448 switch (ts->kind) {
3449 case TEMP_TB:
3450 state = ts->state;
3451 ts->state = state | TS_MEM;
3452 if (state != TS_DEAD) {
3453 continue;
3454 }
3455 break;
3456 case TEMP_EBB:
3457 case TEMP_CONST:
3458 continue;
3459 default:
3460 g_assert_not_reached();
3461 }
3462 la_reset_pref(&s->temps[i]);
3463 }
3464 }
3465
3466 /* liveness analysis: sync globals back to memory and kill. */
3467 static void la_global_kill(TCGContext *s, int ng)
3468 {
3469 int i;
3470
3471 for (i = 0; i < ng; i++) {
3472 s->temps[i].state = TS_DEAD | TS_MEM;
3473 la_reset_pref(&s->temps[i]);
3474 }
3475 }
3476
3477 /* liveness analysis: note live globals crossing calls. */
3478 static void la_cross_call(TCGContext *s, int nt)
3479 {
3480 TCGRegSet mask = ~tcg_target_call_clobber_regs;
3481 int i;
3482
3483 for (i = 0; i < nt; i++) {
3484 TCGTemp *ts = &s->temps[i];
3485 if (!(ts->state & TS_DEAD)) {
3486 TCGRegSet *pset = la_temp_pref(ts);
3487 TCGRegSet set = *pset;
3488
3489 set &= mask;
3490 /* If the combination is not possible, restart. */
3491 if (set == 0) {
3492 set = tcg_target_available_regs[ts->type] & mask;
3493 }
3494 *pset = set;
3495 }
3496 }
3497 }
3498
3499 /*
3500 * Liveness analysis: Verify the lifetime of TEMP_TB, and reduce
3501 * to TEMP_EBB, if possible.
3502 */
3503 static void __attribute__((noinline))
3504 liveness_pass_0(TCGContext *s)
3505 {
3506 void * const multiple_ebb = (void *)(uintptr_t)-1;
3507 int nb_temps = s->nb_temps;
3508 TCGOp *op, *ebb;
3509
3510 for (int i = s->nb_globals; i < nb_temps; ++i) {
3511 s->temps[i].state_ptr = NULL;
3512 }
3513
3514 /*
3515 * Represent each EBB by the op at which it begins. In the case of
3516 * the first EBB, this is the first op, otherwise it is a label.
3517 * Collect the uses of each TEMP_TB: NULL for unused, EBB for use
3518 * within a single EBB, else MULTIPLE_EBB.
3519 */
3520 ebb = QTAILQ_FIRST(&s->ops);
3521 QTAILQ_FOREACH(op, &s->ops, link) {
3522 const TCGOpDef *def;
3523 int nb_oargs, nb_iargs;
3524
3525 switch (op->opc) {
3526 case INDEX_op_set_label:
3527 ebb = op;
3528 continue;
3529 case INDEX_op_discard:
3530 continue;
3531 case INDEX_op_call:
3532 nb_oargs = TCGOP_CALLO(op);
3533 nb_iargs = TCGOP_CALLI(op);
3534 break;
3535 default:
3536 def = &tcg_op_defs[op->opc];
3537 nb_oargs = def->nb_oargs;
3538 nb_iargs = def->nb_iargs;
3539 break;
3540 }
3541
3542 for (int i = 0; i < nb_oargs + nb_iargs; ++i) {
3543 TCGTemp *ts = arg_temp(op->args[i]);
3544
3545 if (ts->kind != TEMP_TB) {
3546 continue;
3547 }
3548 if (ts->state_ptr == NULL) {
3549 ts->state_ptr = ebb;
3550 } else if (ts->state_ptr != ebb) {
3551 ts->state_ptr = multiple_ebb;
3552 }
3553 }
3554 }
3555
3556 /*
3557 * For TEMP_TB that turned out not to be used beyond one EBB,
3558 * reduce the liveness to TEMP_EBB.
3559 */
3560 for (int i = s->nb_globals; i < nb_temps; ++i) {
3561 TCGTemp *ts = &s->temps[i];
3562 if (ts->kind == TEMP_TB && ts->state_ptr != multiple_ebb) {
3563 ts->kind = TEMP_EBB;
3564 }
3565 }
3566 }
3567
3568 /* Liveness analysis : update the opc_arg_life array to tell if a
3569 given input arguments is dead. Instructions updating dead
3570 temporaries are removed. */
3571 static void __attribute__((noinline))
3572 liveness_pass_1(TCGContext *s)
3573 {
3574 int nb_globals = s->nb_globals;
3575 int nb_temps = s->nb_temps;
3576 TCGOp *op, *op_prev;
3577 TCGRegSet *prefs;
3578 int i;
3579
3580 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps);
3581 for (i = 0; i < nb_temps; ++i) {
3582 s->temps[i].state_ptr = prefs + i;
3583 }
3584
3585 /* ??? Should be redundant with the exit_tb that ends the TB. */
3586 la_func_end(s, nb_globals, nb_temps);
3587
3588 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) {
3589 int nb_iargs, nb_oargs;
3590 TCGOpcode opc_new, opc_new2;
3591 bool have_opc_new2;
3592 TCGLifeData arg_life = 0;
3593 TCGTemp *ts;
3594 TCGOpcode opc = op->opc;
3595 const TCGOpDef *def = &tcg_op_defs[opc];
3596
3597 switch (opc) {
3598 case INDEX_op_call:
3599 {
3600 const TCGHelperInfo *info = tcg_call_info(op);
3601 int call_flags = tcg_call_flags(op);
3602
3603 nb_oargs = TCGOP_CALLO(op);
3604 nb_iargs = TCGOP_CALLI(op);
3605
3606 /* pure functions can be removed if their result is unused */
3607 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) {
3608 for (i = 0; i < nb_oargs; i++) {
3609 ts = arg_temp(op->args[i]);
3610 if (ts->state != TS_DEAD) {
3611 goto do_not_remove_call;
3612 }
3613 }
3614 goto do_remove;
3615 }
3616 do_not_remove_call:
3617
3618 /* Output args are dead. */
3619 for (i = 0; i < nb_oargs; i++) {
3620 ts = arg_temp(op->args[i]);
3621 if (ts->state & TS_DEAD) {
3622 arg_life |= DEAD_ARG << i;
3623 }
3624 if (ts->state & TS_MEM) {
3625 arg_life |= SYNC_ARG << i;
3626 }
3627 ts->state = TS_DEAD;
3628 la_reset_pref(ts);
3629 }
3630
3631 /* Not used -- it will be tcg_target_call_oarg_reg(). */
3632 memset(op->output_pref, 0, sizeof(op->output_pref));
3633
3634 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS |
3635 TCG_CALL_NO_READ_GLOBALS))) {
3636 la_global_kill(s, nb_globals);
3637 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) {
3638 la_global_sync(s, nb_globals);
3639 }
3640
3641 /* Record arguments that die in this helper. */
3642 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
3643 ts = arg_temp(op->args[i]);
3644 if (ts->state & TS_DEAD) {
3645 arg_life |= DEAD_ARG << i;
3646 }
3647 }
3648
3649 /* For all live registers, remove call-clobbered prefs. */
3650 la_cross_call(s, nb_temps);
3651
3652 /*
3653 * Input arguments are live for preceding opcodes.
3654 *
3655 * For those arguments that die, and will be allocated in
3656 * registers, clear the register set for that arg, to be
3657 * filled in below. For args that will be on the stack,
3658 * reset to any available reg. Process arguments in reverse
3659 * order so that if a temp is used more than once, the stack
3660 * reset to max happens before the register reset to 0.
3661 */
3662 for (i = nb_iargs - 1; i >= 0; i--) {
3663 const TCGCallArgumentLoc *loc = &info->in[i];
3664 ts = arg_temp(op->args[nb_oargs + i]);
3665
3666 if (ts->state & TS_DEAD) {
3667 switch (loc->kind) {
3668 case TCG_CALL_ARG_NORMAL:
3669 case TCG_CALL_ARG_EXTEND_U:
3670 case TCG_CALL_ARG_EXTEND_S:
3671 if (arg_slot_reg_p(loc->arg_slot)) {
3672 *la_temp_pref(ts) = 0;
3673 break;
3674 }
3675 /* fall through */
3676 default:
3677 *la_temp_pref(ts) =
3678 tcg_target_available_regs[ts->type];
3679 break;
3680 }
3681 ts->state &= ~TS_DEAD;
3682 }
3683 }
3684
3685 /*
3686 * For each input argument, add its input register to prefs.
3687 * If a temp is used once, this produces a single set bit;
3688 * if a temp is used multiple times, this produces a set.
3689 */
3690 for (i = 0; i < nb_iargs; i++) {
3691 const TCGCallArgumentLoc *loc = &info->in[i];
3692 ts = arg_temp(op->args[nb_oargs + i]);
3693
3694 switch (loc->kind) {
3695 case TCG_CALL_ARG_NORMAL:
3696 case TCG_CALL_ARG_EXTEND_U:
3697 case TCG_CALL_ARG_EXTEND_S:
3698 if (arg_slot_reg_p(loc->arg_slot)) {
3699 tcg_regset_set_reg(*la_temp_pref(ts),
3700 tcg_target_call_iarg_regs[loc->arg_slot]);
3701 }
3702 break;
3703 default:
3704 break;
3705 }
3706 }
3707 }
3708 break;
3709 case INDEX_op_insn_start:
3710 break;
3711 case INDEX_op_discard:
3712 /* mark the temporary as dead */
3713 ts = arg_temp(op->args[0]);
3714 ts->state = TS_DEAD;
3715 la_reset_pref(ts);
3716 break;
3717
3718 case INDEX_op_add2_i32:
3719 opc_new = INDEX_op_add_i32;
3720 goto do_addsub2;
3721 case INDEX_op_sub2_i32:
3722 opc_new = INDEX_op_sub_i32;
3723 goto do_addsub2;
3724 case INDEX_op_add2_i64:
3725 opc_new = INDEX_op_add_i64;
3726 goto do_addsub2;
3727 case INDEX_op_sub2_i64:
3728 opc_new = INDEX_op_sub_i64;
3729 do_addsub2:
3730 nb_iargs = 4;
3731 nb_oargs = 2;
3732 /* Test if the high part of the operation is dead, but not
3733 the low part. The result can be optimized to a simple
3734 add or sub. This happens often for x86_64 guest when the
3735 cpu mode is set to 32 bit. */
3736 if (arg_temp(op->args[1])->state == TS_DEAD) {
3737 if (arg_temp(op->args[0])->state == TS_DEAD) {
3738 goto do_remove;
3739 }
3740 /* Replace the opcode and adjust the args in place,
3741 leaving 3 unused args at the end. */
3742 op->opc = opc = opc_new;
3743 op->args[1] = op->args[2];
3744 op->args[2] = op->args[4];
3745 /* Fall through and mark the single-word operation live. */
3746 nb_iargs = 2;
3747 nb_oargs = 1;
3748 }
3749 goto do_not_remove;
3750
3751 case INDEX_op_mulu2_i32:
3752 opc_new = INDEX_op_mul_i32;
3753 opc_new2 = INDEX_op_muluh_i32;
3754 have_opc_new2 = TCG_TARGET_HAS_muluh_i32;
3755 goto do_mul2;
3756 case INDEX_op_muls2_i32:
3757 opc_new = INDEX_op_mul_i32;
3758 opc_new2 = INDEX_op_mulsh_i32;
3759 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32;
3760 goto do_mul2;
3761 case INDEX_op_mulu2_i64:
3762 opc_new = INDEX_op_mul_i64;
3763 opc_new2 = INDEX_op_muluh_i64;
3764 have_opc_new2 = TCG_TARGET_HAS_muluh_i64;
3765 goto do_mul2;
3766 case INDEX_op_muls2_i64:
3767 opc_new = INDEX_op_mul_i64;
3768 opc_new2 = INDEX_op_mulsh_i64;
3769 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64;
3770 goto do_mul2;
3771 do_mul2:
3772 nb_iargs = 2;
3773 nb_oargs = 2;
3774 if (arg_temp(op->args[1])->state == TS_DEAD) {
3775 if (arg_temp(op->args[0])->state == TS_DEAD) {
3776 /* Both parts of the operation are dead. */
3777 goto do_remove;
3778 }
3779 /* The high part of the operation is dead; generate the low. */
3780 op->opc = opc = opc_new;
3781 op->args[1] = op->args[2];
3782 op->args[2] = op->args[3];
3783 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) {
3784 /* The low part of the operation is dead; generate the high. */
3785 op->opc = opc = opc_new2;
3786 op->args[0] = op->args[1];
3787 op->args[1] = op->args[2];
3788 op->args[2] = op->args[3];
3789 } else {
3790 goto do_not_remove;
3791 }
3792 /* Mark the single-word operation live. */
3793 nb_oargs = 1;
3794 goto do_not_remove;
3795
3796 default:
3797 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */
3798 nb_iargs = def->nb_iargs;
3799 nb_oargs = def->nb_oargs;
3800
3801 /* Test if the operation can be removed because all
3802 its outputs are dead. We assume that nb_oargs == 0
3803 implies side effects */
3804 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) {
3805 for (i = 0; i < nb_oargs; i++) {
3806 if (arg_temp(op->args[i])->state != TS_DEAD) {
3807 goto do_not_remove;
3808 }
3809 }
3810 goto do_remove;
3811 }
3812 goto do_not_remove;
3813
3814 do_remove:
3815 tcg_op_remove(s, op);
3816 break;
3817
3818 do_not_remove:
3819 for (i = 0; i < nb_oargs; i++) {
3820 ts = arg_temp(op->args[i]);
3821
3822 /* Remember the preference of the uses that followed. */
3823 if (i < ARRAY_SIZE(op->output_pref)) {
3824 op->output_pref[i] = *la_temp_pref(ts);
3825 }
3826
3827 /* Output args are dead. */
3828 if (ts->state & TS_DEAD) {
3829 arg_life |= DEAD_ARG << i;
3830 }
3831 if (ts->state & TS_MEM) {
3832 arg_life |= SYNC_ARG << i;
3833 }
3834 ts->state = TS_DEAD;
3835 la_reset_pref(ts);
3836 }
3837
3838 /* If end of basic block, update. */
3839 if (def->flags & TCG_OPF_BB_EXIT) {
3840 la_func_end(s, nb_globals, nb_temps);
3841 } else if (def->flags & TCG_OPF_COND_BRANCH) {
3842 la_bb_sync(s, nb_globals, nb_temps);
3843 } else if (def->flags & TCG_OPF_BB_END) {
3844 la_bb_end(s, nb_globals, nb_temps);
3845 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) {
3846 la_global_sync(s, nb_globals);
3847 if (def->flags & TCG_OPF_CALL_CLOBBER) {
3848 la_cross_call(s, nb_temps);
3849 }
3850 }
3851
3852 /* Record arguments that die in this opcode. */
3853 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
3854 ts = arg_temp(op->args[i]);
3855 if (ts->state & TS_DEAD) {
3856 arg_life |= DEAD_ARG << i;
3857 }
3858 }
3859
3860 /* Input arguments are live for preceding opcodes. */
3861 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
3862 ts = arg_temp(op->args[i]);
3863 if (ts->state & TS_DEAD) {
3864 /* For operands that were dead, initially allow
3865 all regs for the type. */
3866 *la_temp_pref(ts) = tcg_target_available_regs[ts->type];
3867 ts->state &= ~TS_DEAD;
3868 }
3869 }
3870
3871 /* Incorporate constraints for this operand. */
3872 switch (opc) {
3873 case INDEX_op_mov_i32:
3874 case INDEX_op_mov_i64:
3875 /* Note that these are TCG_OPF_NOT_PRESENT and do not
3876 have proper constraints. That said, special case
3877 moves to propagate preferences backward. */
3878 if (IS_DEAD_ARG(1)) {
3879 *la_temp_pref(arg_temp(op->args[0]))
3880 = *la_temp_pref(arg_temp(op->args[1]));
3881 }
3882 break;
3883
3884 default:
3885 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
3886 const TCGArgConstraint *ct = &def->args_ct[i];
3887 TCGRegSet set, *pset;
3888
3889 ts = arg_temp(op->args[i]);
3890 pset = la_temp_pref(ts);
3891 set = *pset;
3892
3893 set &= ct->regs;
3894 if (ct->ialias) {
3895 set &= output_pref(op, ct->alias_index);
3896 }
3897 /* If the combination is not possible, restart. */
3898 if (set == 0) {
3899 set = ct->regs;
3900 }
3901 *pset = set;
3902 }
3903 break;
3904 }
3905 break;
3906 }
3907 op->life = arg_life;
3908 }
3909 }
3910
3911 /* Liveness analysis: Convert indirect regs to direct temporaries. */
3912 static bool __attribute__((noinline))
3913 liveness_pass_2(TCGContext *s)
3914 {
3915 int nb_globals = s->nb_globals;
3916 int nb_temps, i;
3917 bool changes = false;
3918 TCGOp *op, *op_next;
3919
3920 /* Create a temporary for each indirect global. */
3921 for (i = 0; i < nb_globals; ++i) {
3922 TCGTemp *its = &s->temps[i];
3923 if (its->indirect_reg) {
3924 TCGTemp *dts = tcg_temp_alloc(s);
3925 dts->type = its->type;
3926 dts->base_type = its->base_type;
3927 dts->temp_subindex = its->temp_subindex;
3928 dts->kind = TEMP_EBB;
3929 its->state_ptr = dts;
3930 } else {
3931 its->state_ptr = NULL;
3932 }
3933 /* All globals begin dead. */
3934 its->state = TS_DEAD;
3935 }
3936 for (nb_temps = s->nb_temps; i < nb_temps; ++i) {
3937 TCGTemp *its = &s->temps[i];
3938 its->state_ptr = NULL;
3939 its->state = TS_DEAD;
3940 }
3941
3942 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
3943 TCGOpcode opc = op->opc;
3944 const TCGOpDef *def = &tcg_op_defs[opc];
3945 TCGLifeData arg_life = op->life;
3946 int nb_iargs, nb_oargs, call_flags;
3947 TCGTemp *arg_ts, *dir_ts;
3948
3949 if (opc == INDEX_op_call) {
3950 nb_oargs = TCGOP_CALLO(op);
3951 nb_iargs = TCGOP_CALLI(op);
3952 call_flags = tcg_call_flags(op);
3953 } else {
3954 nb_iargs = def->nb_iargs;
3955 nb_oargs = def->nb_oargs;
3956
3957 /* Set flags similar to how calls require. */
3958 if (def->flags & TCG_OPF_COND_BRANCH) {
3959 /* Like reading globals: sync_globals */
3960 call_flags = TCG_CALL_NO_WRITE_GLOBALS;
3961 } else if (def->flags & TCG_OPF_BB_END) {
3962 /* Like writing globals: save_globals */
3963 call_flags = 0;
3964 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) {
3965 /* Like reading globals: sync_globals */
3966 call_flags = TCG_CALL_NO_WRITE_GLOBALS;
3967 } else {
3968 /* No effect on globals. */
3969 call_flags = (TCG_CALL_NO_READ_GLOBALS |
3970 TCG_CALL_NO_WRITE_GLOBALS);
3971 }
3972 }
3973
3974 /* Make sure that input arguments are available. */
3975 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
3976 arg_ts = arg_temp(op->args[i]);
3977 dir_ts = arg_ts->state_ptr;
3978 if (dir_ts && arg_ts->state == TS_DEAD) {
3979 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32
3980 ? INDEX_op_ld_i32
3981 : INDEX_op_ld_i64);
3982 TCGOp *lop = tcg_op_insert_before(s, op, lopc, 3);
3983
3984 lop->args[0] = temp_arg(dir_ts);
3985 lop->args[1] = temp_arg(arg_ts->mem_base);
3986 lop->args[2] = arg_ts->mem_offset;
3987
3988 /* Loaded, but synced with memory. */
3989 arg_ts->state = TS_MEM;
3990 }
3991 }
3992
3993 /* Perform input replacement, and mark inputs that became dead.
3994 No action is required except keeping temp_state up to date
3995 so that we reload when needed. */
3996 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
3997 arg_ts = arg_temp(op->args[i]);
3998 dir_ts = arg_ts->state_ptr;
3999 if (dir_ts) {
4000 op->args[i] = temp_arg(dir_ts);
4001 changes = true;
4002 if (IS_DEAD_ARG(i)) {
4003 arg_ts->state = TS_DEAD;
4004 }
4005 }
4006 }
4007
4008 /* Liveness analysis should ensure that the following are
4009 all correct, for call sites and basic block end points. */
4010 if (call_flags & TCG_CALL_NO_READ_GLOBALS) {
4011 /* Nothing to do */
4012 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) {
4013 for (i = 0; i < nb_globals; ++i) {
4014 /* Liveness should see that globals are synced back,
4015 that is, either TS_DEAD or TS_MEM. */
4016 arg_ts = &s->temps[i];
4017 tcg_debug_assert(arg_ts->state_ptr == 0
4018 || arg_ts->state != 0);
4019 }
4020 } else {
4021 for (i = 0; i < nb_globals; ++i) {
4022 /* Liveness should see that globals are saved back,
4023 that is, TS_DEAD, waiting to be reloaded. */
4024 arg_ts = &s->temps[i];
4025 tcg_debug_assert(arg_ts->state_ptr == 0
4026 || arg_ts->state == TS_DEAD);
4027 }
4028 }
4029
4030 /* Outputs become available. */
4031 if (opc == INDEX_op_mov_i32 || opc == INDEX_op_mov_i64) {
4032 arg_ts = arg_temp(op->args[0]);
4033 dir_ts = arg_ts->state_ptr;
4034 if (dir_ts) {
4035 op->args[0] = temp_arg(dir_ts);
4036 changes = true;
4037
4038 /* The output is now live and modified. */
4039 arg_ts->state = 0;
4040
4041 if (NEED_SYNC_ARG(0)) {
4042 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32
4043 ? INDEX_op_st_i32
4044 : INDEX_op_st_i64);
4045 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3);
4046 TCGTemp *out_ts = dir_ts;
4047
4048 if (IS_DEAD_ARG(0)) {
4049 out_ts = arg_temp(op->args[1]);
4050 arg_ts->state = TS_DEAD;
4051 tcg_op_remove(s, op);
4052 } else {
4053 arg_ts->state = TS_MEM;
4054 }
4055
4056 sop->args[0] = temp_arg(out_ts);
4057 sop->args[1] = temp_arg(arg_ts->mem_base);
4058 sop->args[2] = arg_ts->mem_offset;
4059 } else {
4060 tcg_debug_assert(!IS_DEAD_ARG(0));
4061 }
4062 }
4063 } else {
4064 for (i = 0; i < nb_oargs; i++) {
4065 arg_ts = arg_temp(op->args[i]);
4066 dir_ts = arg_ts->state_ptr;
4067 if (!dir_ts) {
4068 continue;
4069 }
4070 op->args[i] = temp_arg(dir_ts);
4071 changes = true;
4072
4073 /* The output is now live and modified. */
4074 arg_ts->state = 0;
4075
4076 /* Sync outputs upon their last write. */
4077 if (NEED_SYNC_ARG(i)) {
4078 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32
4079 ? INDEX_op_st_i32
4080 : INDEX_op_st_i64);
4081 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3);
4082
4083 sop->args[0] = temp_arg(dir_ts);
4084 sop->args[1] = temp_arg(arg_ts->mem_base);
4085 sop->args[2] = arg_ts->mem_offset;
4086
4087 arg_ts->state = TS_MEM;
4088 }
4089 /* Drop outputs that are dead. */
4090 if (IS_DEAD_ARG(i)) {
4091 arg_ts->state = TS_DEAD;
4092 }
4093 }
4094 }
4095 }
4096
4097 return changes;
4098 }
4099
4100 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts)
4101 {
4102 intptr_t off;
4103 int size, align;
4104
4105 /* When allocating an object, look at the full type. */
4106 size = tcg_type_size(ts->base_type);
4107 switch (ts->base_type) {
4108 case TCG_TYPE_I32:
4109 align = 4;
4110 break;
4111 case TCG_TYPE_I64:
4112 case TCG_TYPE_V64:
4113 align = 8;
4114 break;
4115 case TCG_TYPE_I128:
4116 case TCG_TYPE_V128:
4117 case TCG_TYPE_V256:
4118 /*
4119 * Note that we do not require aligned storage for V256,
4120 * and that we provide alignment for I128 to match V128,
4121 * even if that's above what the host ABI requires.
4122 */
4123 align = 16;
4124 break;
4125 default:
4126 g_assert_not_reached();
4127 }
4128
4129 /*
4130 * Assume the stack is sufficiently aligned.
4131 * This affects e.g. ARM NEON, where we have 8 byte stack alignment
4132 * and do not require 16 byte vector alignment. This seems slightly
4133 * easier than fully parameterizing the above switch statement.
4134 */
4135 align = MIN(TCG_TARGET_STACK_ALIGN, align);
4136 off = ROUND_UP(s->current_frame_offset, align);
4137
4138 /* If we've exhausted the stack frame, restart with a smaller TB. */
4139 if (off + size > s->frame_end) {
4140 tcg_raise_tb_overflow(s);
4141 }
4142 s->current_frame_offset = off + size;
4143 #if defined(__sparc__)
4144 off += TCG_TARGET_STACK_BIAS;
4145 #endif
4146
4147 /* If the object was subdivided, assign memory to all the parts. */
4148 if (ts->base_type != ts->type) {
4149 int part_size = tcg_type_size(ts->type);
4150 int part_count = size / part_size;
4151
4152 /*
4153 * Each part is allocated sequentially in tcg_temp_new_internal.
4154 * Jump back to the first part by subtracting the current index.
4155 */
4156 ts -= ts->temp_subindex;
4157 for (int i = 0; i < part_count; ++i) {
4158 ts[i].mem_offset = off + i * part_size;
4159 ts[i].mem_base = s->frame_temp;
4160 ts[i].mem_allocated = 1;
4161 }
4162 } else {
4163 ts->mem_offset = off;
4164 ts->mem_base = s->frame_temp;
4165 ts->mem_allocated = 1;
4166 }
4167 }
4168
4169 /* Assign @reg to @ts, and update reg_to_temp[]. */
4170 static void set_temp_val_reg(TCGContext *s, TCGTemp *ts, TCGReg reg)
4171 {
4172 if (ts->val_type == TEMP_VAL_REG) {
4173 TCGReg old = ts->reg;
4174 tcg_debug_assert(s->reg_to_temp[old] == ts);
4175 if (old == reg) {
4176 return;
4177 }
4178 s->reg_to_temp[old] = NULL;
4179 }
4180 tcg_debug_assert(s->reg_to_temp[reg] == NULL);
4181 s->reg_to_temp[reg] = ts;
4182 ts->val_type = TEMP_VAL_REG;
4183 ts->reg = reg;
4184 }
4185
4186 /* Assign a non-register value type to @ts, and update reg_to_temp[]. */
4187 static void set_temp_val_nonreg(TCGContext *s, TCGTemp *ts, TCGTempVal type)
4188 {
4189 tcg_debug_assert(type != TEMP_VAL_REG);
4190 if (ts->val_type == TEMP_VAL_REG) {
4191 TCGReg reg = ts->reg;
4192 tcg_debug_assert(s->reg_to_temp[reg] == ts);
4193 s->reg_to_temp[reg] = NULL;
4194 }
4195 ts->val_type = type;
4196 }
4197
4198 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet);
4199
4200 /* Mark a temporary as free or dead. If 'free_or_dead' is negative,
4201 mark it free; otherwise mark it dead. */
4202 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead)
4203 {
4204 TCGTempVal new_type;
4205
4206 switch (ts->kind) {
4207 case TEMP_FIXED:
4208 return;
4209 case TEMP_GLOBAL:
4210 case TEMP_TB:
4211 new_type = TEMP_VAL_MEM;
4212 break;
4213 case TEMP_EBB:
4214 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD;
4215 break;
4216 case TEMP_CONST:
4217 new_type = TEMP_VAL_CONST;
4218 break;
4219 default:
4220 g_assert_not_reached();
4221 }
4222 set_temp_val_nonreg(s, ts, new_type);
4223 }
4224
4225 /* Mark a temporary as dead. */
4226 static inline void temp_dead(TCGContext *s, TCGTemp *ts)
4227 {
4228 temp_free_or_dead(s, ts, 1);
4229 }
4230
4231 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary
4232 registers needs to be allocated to store a constant. If 'free_or_dead'
4233 is non-zero, subsequently release the temporary; if it is positive, the
4234 temp is dead; if it is negative, the temp is free. */
4235 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs,
4236 TCGRegSet preferred_regs, int free_or_dead)
4237 {
4238 if (!temp_readonly(ts) && !ts->mem_coherent) {
4239 if (!ts->mem_allocated) {
4240 temp_allocate_frame(s, ts);
4241 }
4242 switch (ts->val_type) {
4243 case TEMP_VAL_CONST:
4244 /* If we're going to free the temp immediately, then we won't
4245 require it later in a register, so attempt to store the
4246 constant to memory directly. */
4247 if (free_or_dead
4248 && tcg_out_sti(s, ts->type, ts->val,
4249 ts->mem_base->reg, ts->mem_offset)) {
4250 break;
4251 }
4252 temp_load(s, ts, tcg_target_available_regs[ts->type],
4253 allocated_regs, preferred_regs);
4254 /* fallthrough */
4255
4256 case TEMP_VAL_REG:
4257 tcg_out_st(s, ts->type, ts->reg,
4258 ts->mem_base->reg, ts->mem_offset);
4259 break;
4260
4261 case TEMP_VAL_MEM:
4262 break;
4263
4264 case TEMP_VAL_DEAD:
4265 default:
4266 g_assert_not_reached();
4267 }
4268 ts->mem_coherent = 1;
4269 }
4270 if (free_or_dead) {
4271 temp_free_or_dead(s, ts, free_or_dead);
4272 }
4273 }
4274
4275 /* free register 'reg' by spilling the corresponding temporary if necessary */
4276 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs)
4277 {
4278 TCGTemp *ts = s->reg_to_temp[reg];
4279 if (ts != NULL) {
4280 temp_sync(s, ts, allocated_regs, 0, -1);
4281 }
4282 }
4283
4284 /**
4285 * tcg_reg_alloc:
4286 * @required_regs: Set of registers in which we must allocate.
4287 * @allocated_regs: Set of registers which must be avoided.
4288 * @preferred_regs: Set of registers we should prefer.
4289 * @rev: True if we search the registers in "indirect" order.
4290 *
4291 * The allocated register must be in @required_regs & ~@allocated_regs,
4292 * but if we can put it in @preferred_regs we may save a move later.
4293 */
4294 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs,
4295 TCGRegSet allocated_regs,
4296 TCGRegSet preferred_regs, bool rev)
4297 {
4298 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order);
4299 TCGRegSet reg_ct[2];
4300 const int *order;
4301
4302 reg_ct[1] = required_regs & ~allocated_regs;
4303 tcg_debug_assert(reg_ct[1] != 0);
4304 reg_ct[0] = reg_ct[1] & preferred_regs;
4305
4306 /* Skip the preferred_regs option if it cannot be satisfied,
4307 or if the preference made no difference. */
4308 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1];
4309
4310 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order;
4311
4312 /* Try free registers, preferences first. */
4313 for (j = f; j < 2; j++) {
4314 TCGRegSet set = reg_ct[j];
4315
4316 if (tcg_regset_single(set)) {
4317 /* One register in the set. */
4318 TCGReg reg = tcg_regset_first(set);
4319 if (s->reg_to_temp[reg] == NULL) {
4320 return reg;
4321 }
4322 } else {
4323 for (i = 0; i < n; i++) {
4324 TCGReg reg = order[i];
4325 if (s->reg_to_temp[reg] == NULL &&
4326 tcg_regset_test_reg(set, reg)) {
4327 return reg;
4328 }
4329 }
4330 }
4331 }
4332
4333 /* We must spill something. */
4334 for (j = f; j < 2; j++) {
4335 TCGRegSet set = reg_ct[j];
4336
4337 if (tcg_regset_single(set)) {
4338 /* One register in the set. */
4339 TCGReg reg = tcg_regset_first(set);
4340 tcg_reg_free(s, reg, allocated_regs);
4341 return reg;
4342 } else {
4343 for (i = 0; i < n; i++) {
4344 TCGReg reg = order[i];
4345 if (tcg_regset_test_reg(set, reg)) {
4346 tcg_reg_free(s, reg, allocated_regs);
4347 return reg;
4348 }
4349 }
4350 }
4351 }
4352
4353 g_assert_not_reached();
4354 }
4355
4356 static TCGReg tcg_reg_alloc_pair(TCGContext *s, TCGRegSet required_regs,
4357 TCGRegSet allocated_regs,
4358 TCGRegSet preferred_regs, bool rev)
4359 {
4360 int i, j, k, fmin, n = ARRAY_SIZE(tcg_target_reg_alloc_order);
4361 TCGRegSet reg_ct[2];
4362 const int *order;
4363
4364 /* Ensure that if I is not in allocated_regs, I+1 is not either. */
4365 reg_ct[1] = required_regs & ~(allocated_regs | (allocated_regs >> 1));
4366 tcg_debug_assert(reg_ct[1] != 0);
4367 reg_ct[0] = reg_ct[1] & preferred_regs;
4368
4369 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order;
4370
4371 /*
4372 * Skip the preferred_regs option if it cannot be satisfied,
4373 * or if the preference made no difference.
4374 */
4375 k = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1];
4376
4377 /*
4378 * Minimize the number of flushes by looking for 2 free registers first,
4379 * then a single flush, then two flushes.
4380 */
4381 for (fmin = 2; fmin >= 0; fmin--) {
4382 for (j = k; j < 2; j++) {
4383 TCGRegSet set = reg_ct[j];
4384
4385 for (i = 0; i < n; i++) {
4386 TCGReg reg = order[i];
4387
4388 if (tcg_regset_test_reg(set, reg)) {
4389 int f = !s->reg_to_temp[reg] + !s->reg_to_temp[reg + 1];
4390 if (f >= fmin) {
4391 tcg_reg_free(s, reg, allocated_regs);
4392 tcg_reg_free(s, reg + 1, allocated_regs);
4393 return reg;
4394 }
4395 }
4396 }
4397 }
4398 }
4399 g_assert_not_reached();
4400 }
4401
4402 /* Make sure the temporary is in a register. If needed, allocate the register
4403 from DESIRED while avoiding ALLOCATED. */
4404 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs,
4405 TCGRegSet allocated_regs, TCGRegSet preferred_regs)
4406 {
4407 TCGReg reg;
4408
4409 switch (ts->val_type) {
4410 case TEMP_VAL_REG:
4411 return;
4412 case TEMP_VAL_CONST:
4413 reg = tcg_reg_alloc(s, desired_regs, allocated_regs,
4414 preferred_regs, ts->indirect_base);
4415 if (ts->type <= TCG_TYPE_I64) {
4416 tcg_out_movi(s, ts->type, reg, ts->val);
4417 } else {
4418 uint64_t val = ts->val;
4419 MemOp vece = MO_64;
4420
4421 /*
4422 * Find the minimal vector element that matches the constant.
4423 * The targets will, in general, have to do this search anyway,
4424 * do this generically.
4425 */
4426 if (val == dup_const(MO_8, val)) {
4427 vece = MO_8;
4428 } else if (val == dup_const(MO_16, val)) {
4429 vece = MO_16;
4430 } else if (val == dup_const(MO_32, val)) {
4431 vece = MO_32;
4432 }
4433
4434 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val);
4435 }
4436 ts->mem_coherent = 0;
4437 break;
4438 case TEMP_VAL_MEM:
4439 reg = tcg_reg_alloc(s, desired_regs, allocated_regs,
4440 preferred_regs, ts->indirect_base);
4441 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset);
4442 ts->mem_coherent = 1;
4443 break;
4444 case TEMP_VAL_DEAD:
4445 default:
4446 g_assert_not_reached();
4447 }
4448 set_temp_val_reg(s, ts, reg);
4449 }
4450
4451 /* Save a temporary to memory. 'allocated_regs' is used in case a
4452 temporary registers needs to be allocated to store a constant. */
4453 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs)
4454 {
4455 /* The liveness analysis already ensures that globals are back
4456 in memory. Keep an tcg_debug_assert for safety. */
4457 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts));
4458 }
4459
4460 /* save globals to their canonical location and assume they can be
4461 modified be the following code. 'allocated_regs' is used in case a
4462 temporary registers needs to be allocated to store a constant. */
4463 static void save_globals(TCGContext *s, TCGRegSet allocated_regs)
4464 {
4465 int i, n;
4466
4467 for (i = 0, n = s->nb_globals; i < n; i++) {
4468 temp_save(s, &s->temps[i], allocated_regs);
4469 }
4470 }
4471
4472 /* sync globals to their canonical location and assume they can be
4473 read by the following code. 'allocated_regs' is used in case a
4474 temporary registers needs to be allocated to store a constant. */
4475 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs)
4476 {
4477 int i, n;
4478
4479 for (i = 0, n = s->nb_globals; i < n; i++) {
4480 TCGTemp *ts = &s->temps[i];
4481 tcg_debug_assert(ts->val_type != TEMP_VAL_REG
4482 || ts->kind == TEMP_FIXED
4483 || ts->mem_coherent);
4484 }
4485 }
4486
4487 /* at the end of a basic block, we assume all temporaries are dead and
4488 all globals are stored at their canonical location. */
4489 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs)
4490 {
4491 int i;
4492
4493 for (i = s->nb_globals; i < s->nb_temps; i++) {
4494 TCGTemp *ts = &s->temps[i];
4495
4496 switch (ts->kind) {
4497 case TEMP_TB:
4498 temp_save(s, ts, allocated_regs);
4499 break;
4500 case TEMP_EBB:
4501 /* The liveness analysis already ensures that temps are dead.
4502 Keep an tcg_debug_assert for safety. */
4503 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD);
4504 break;
4505 case TEMP_CONST:
4506 /* Similarly, we should have freed any allocated register. */
4507 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST);
4508 break;
4509 default:
4510 g_assert_not_reached();
4511 }
4512 }
4513
4514 save_globals(s, allocated_regs);
4515 }
4516
4517 /*
4518 * At a conditional branch, we assume all temporaries are dead unless
4519 * explicitly live-across-conditional-branch; all globals and local
4520 * temps are synced to their location.
4521 */
4522 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs)
4523 {
4524 sync_globals(s, allocated_regs);
4525
4526 for (int i = s->nb_globals; i < s->nb_temps; i++) {
4527 TCGTemp *ts = &s->temps[i];
4528 /*
4529 * The liveness analysis already ensures that temps are dead.
4530 * Keep tcg_debug_asserts for safety.
4531 */
4532 switch (ts->kind) {
4533 case TEMP_TB:
4534 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent);
4535 break;
4536 case TEMP_EBB:
4537 case TEMP_CONST:
4538 break;
4539 default:
4540 g_assert_not_reached();
4541 }
4542 }
4543 }
4544
4545 /*
4546 * Specialized code generation for INDEX_op_mov_* with a constant.
4547 */
4548 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots,
4549 tcg_target_ulong val, TCGLifeData arg_life,
4550 TCGRegSet preferred_regs)
4551 {
4552 /* ENV should not be modified. */
4553 tcg_debug_assert(!temp_readonly(ots));
4554
4555 /* The movi is not explicitly generated here. */
4556 set_temp_val_nonreg(s, ots, TEMP_VAL_CONST);
4557 ots->val = val;
4558 ots->mem_coherent = 0;
4559 if (NEED_SYNC_ARG(0)) {
4560 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0));
4561 } else if (IS_DEAD_ARG(0)) {
4562 temp_dead(s, ots);
4563 }
4564 }
4565
4566 /*
4567 * Specialized code generation for INDEX_op_mov_*.
4568 */
4569 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op)
4570 {
4571 const TCGLifeData arg_life = op->life;
4572 TCGRegSet allocated_regs, preferred_regs;
4573 TCGTemp *ts, *ots;
4574 TCGType otype, itype;
4575 TCGReg oreg, ireg;
4576
4577 allocated_regs = s->reserved_regs;
4578 preferred_regs = output_pref(op, 0);
4579 ots = arg_temp(op->args[0]);
4580 ts = arg_temp(op->args[1]);
4581
4582 /* ENV should not be modified. */
4583 tcg_debug_assert(!temp_readonly(ots));
4584
4585 /* Note that otype != itype for no-op truncation. */
4586 otype = ots->type;
4587 itype = ts->type;
4588
4589 if (ts->val_type == TEMP_VAL_CONST) {
4590 /* propagate constant or generate sti */
4591 tcg_target_ulong val = ts->val;
4592 if (IS_DEAD_ARG(1)) {
4593 temp_dead(s, ts);
4594 }
4595 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs);
4596 return;
4597 }
4598
4599 /* If the source value is in memory we're going to be forced
4600 to have it in a register in order to perform the copy. Copy
4601 the SOURCE value into its own register first, that way we
4602 don't have to reload SOURCE the next time it is used. */
4603 if (ts->val_type == TEMP_VAL_MEM) {
4604 temp_load(s, ts, tcg_target_available_regs[itype],
4605 allocated_regs, preferred_regs);
4606 }
4607 tcg_debug_assert(ts->val_type == TEMP_VAL_REG);
4608 ireg = ts->reg;
4609
4610 if (IS_DEAD_ARG(0)) {
4611 /* mov to a non-saved dead register makes no sense (even with
4612 liveness analysis disabled). */
4613 tcg_debug_assert(NEED_SYNC_ARG(0));
4614 if (!ots->mem_allocated) {
4615 temp_allocate_frame(s, ots);
4616 }
4617 tcg_out_st(s, otype, ireg, ots->mem_base->reg, ots->mem_offset);
4618 if (IS_DEAD_ARG(1)) {
4619 temp_dead(s, ts);
4620 }
4621 temp_dead(s, ots);
4622 return;
4623 }
4624
4625 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) {
4626 /*
4627 * The mov can be suppressed. Kill input first, so that it
4628 * is unlinked from reg_to_temp, then set the output to the
4629 * reg that we saved from the input.
4630 */
4631 temp_dead(s, ts);
4632 oreg = ireg;
4633 } else {
4634 if (ots->val_type == TEMP_VAL_REG) {
4635 oreg = ots->reg;
4636 } else {
4637 /* Make sure to not spill the input register during allocation. */
4638 oreg = tcg_reg_alloc(s, tcg_target_available_regs[otype],
4639 allocated_regs | ((TCGRegSet)1 << ireg),
4640 preferred_regs, ots->indirect_base);
4641 }
4642 if (!tcg_out_mov(s, otype, oreg, ireg)) {
4643 /*
4644 * Cross register class move not supported.
4645 * Store the source register into the destination slot
4646 * and leave the destination temp as TEMP_VAL_MEM.
4647 */
4648 assert(!temp_readonly(ots));
4649 if (!ts->mem_allocated) {
4650 temp_allocate_frame(s, ots);
4651 }
4652 tcg_out_st(s, ts->type, ireg, ots->mem_base->reg, ots->mem_offset);
4653 set_temp_val_nonreg(s, ts, TEMP_VAL_MEM);
4654 ots->mem_coherent = 1;
4655 return;
4656 }
4657 }
4658 set_temp_val_reg(s, ots, oreg);
4659 ots->mem_coherent = 0;
4660
4661 if (NEED_SYNC_ARG(0)) {
4662 temp_sync(s, ots, allocated_regs, 0, 0);
4663 }
4664 }
4665
4666 /*
4667 * Specialized code generation for INDEX_op_dup_vec.
4668 */
4669 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op)
4670 {
4671 const TCGLifeData arg_life = op->life;
4672 TCGRegSet dup_out_regs, dup_in_regs;
4673 TCGTemp *its, *ots;
4674 TCGType itype, vtype;
4675 unsigned vece;
4676 int lowpart_ofs;
4677 bool ok;
4678
4679 ots = arg_temp(op->args[0]);
4680 its = arg_temp(op->args[1]);
4681
4682 /* ENV should not be modified. */
4683 tcg_debug_assert(!temp_readonly(ots));
4684
4685 itype = its->type;
4686 vece = TCGOP_VECE(op);
4687 vtype = TCGOP_VECL(op) + TCG_TYPE_V64;
4688
4689 if (its->val_type == TEMP_VAL_CONST) {
4690 /* Propagate constant via movi -> dupi. */
4691 tcg_target_ulong val = its->val;
4692 if (IS_DEAD_ARG(1)) {
4693 temp_dead(s, its);
4694 }
4695 tcg_reg_alloc_do_movi(s, ots, val, arg_life, output_pref(op, 0));
4696 return;
4697 }
4698
4699 dup_out_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs;
4700 dup_in_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[1].regs;
4701
4702 /* Allocate the output register now. */
4703 if (ots->val_type != TEMP_VAL_REG) {
4704 TCGRegSet allocated_regs = s->reserved_regs;
4705 TCGReg oreg;
4706
4707 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) {
4708 /* Make sure to not spill the input register. */
4709 tcg_regset_set_reg(allocated_regs, its->reg);
4710 }
4711 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs,
4712 output_pref(op, 0), ots->indirect_base);
4713 set_temp_val_reg(s, ots, oreg);
4714 }
4715
4716 switch (its->val_type) {
4717 case TEMP_VAL_REG:
4718 /*
4719 * The dup constriaints must be broad, covering all possible VECE.
4720 * However, tcg_op_dup_vec() gets to see the VECE and we allow it
4721 * to fail, indicating that extra moves are required for that case.
4722 */
4723 if (tcg_regset_test_reg(dup_in_regs, its->reg)) {
4724 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) {
4725 goto done;
4726 }
4727 /* Try again from memory or a vector input register. */
4728 }
4729 if (!its->mem_coherent) {
4730 /*
4731 * The input register is not synced, and so an extra store
4732 * would be required to use memory. Attempt an integer-vector
4733 * register move first. We do not have a TCGRegSet for this.
4734 */
4735 if (tcg_out_mov(s, itype, ots->reg, its->reg)) {
4736 break;
4737 }
4738 /* Sync the temp back to its slot and load from there. */
4739 temp_sync(s, its, s->reserved_regs, 0, 0);
4740 }
4741 /* fall through */
4742
4743 case TEMP_VAL_MEM:
4744 lowpart_ofs = 0;
4745 if (HOST_BIG_ENDIAN) {
4746 lowpart_ofs = tcg_type_size(itype) - (1 << vece);
4747 }
4748 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg,
4749 its->mem_offset + lowpart_ofs)) {
4750 goto done;
4751 }
4752 /* Load the input into the destination vector register. */
4753 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset);
4754 break;
4755
4756 default:
4757 g_assert_not_reached();
4758 }
4759
4760 /* We now have a vector input register, so dup must succeed. */
4761 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg);
4762 tcg_debug_assert(ok);
4763
4764 done:
4765 ots->mem_coherent = 0;
4766 if (IS_DEAD_ARG(1)) {
4767 temp_dead(s, its);
4768 }
4769 if (NEED_SYNC_ARG(0)) {
4770 temp_sync(s, ots, s->reserved_regs, 0, 0);
4771 }
4772 if (IS_DEAD_ARG(0)) {
4773 temp_dead(s, ots);
4774 }
4775 }
4776
4777 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op)
4778 {
4779 const TCGLifeData arg_life = op->life;
4780 const TCGOpDef * const def = &tcg_op_defs[op->opc];
4781 TCGRegSet i_allocated_regs;
4782 TCGRegSet o_allocated_regs;
4783 int i, k, nb_iargs, nb_oargs;
4784 TCGReg reg;
4785 TCGArg arg;
4786 const TCGArgConstraint *arg_ct;
4787 TCGTemp *ts;
4788 TCGArg new_args[TCG_MAX_OP_ARGS];
4789 int const_args[TCG_MAX_OP_ARGS];
4790 TCGCond op_cond;
4791
4792 nb_oargs = def->nb_oargs;
4793 nb_iargs = def->nb_iargs;
4794
4795 /* copy constants */
4796 memcpy(new_args + nb_oargs + nb_iargs,
4797 op->args + nb_oargs + nb_iargs,
4798 sizeof(TCGArg) * def->nb_cargs);
4799
4800 i_allocated_regs = s->reserved_regs;
4801 o_allocated_regs = s->reserved_regs;
4802
4803 switch (op->opc) {
4804 case INDEX_op_brcond_i32:
4805 case INDEX_op_brcond_i64:
4806 op_cond = op->args[2];
4807 break;
4808 case INDEX_op_setcond_i32:
4809 case INDEX_op_setcond_i64:
4810 case INDEX_op_negsetcond_i32:
4811 case INDEX_op_negsetcond_i64:
4812 case INDEX_op_cmp_vec:
4813 op_cond = op->args[3];
4814 break;
4815 case INDEX_op_brcond2_i32:
4816 op_cond = op->args[4];
4817 break;
4818 case INDEX_op_movcond_i32:
4819 case INDEX_op_movcond_i64:
4820 case INDEX_op_setcond2_i32:
4821 case INDEX_op_cmpsel_vec:
4822 op_cond = op->args[5];
4823 break;
4824 default:
4825 /* No condition within opcode. */
4826 op_cond = TCG_COND_ALWAYS;
4827 break;
4828 }
4829
4830 /* satisfy input constraints */
4831 for (k = 0; k < nb_iargs; k++) {
4832 TCGRegSet i_preferred_regs, i_required_regs;
4833 bool allocate_new_reg, copyto_new_reg;
4834 TCGTemp *ts2;
4835 int i1, i2;
4836
4837 i = def->args_ct[nb_oargs + k].sort_index;
4838 arg = op->args[i];
4839 arg_ct = &def->args_ct[i];
4840 ts = arg_temp(arg);
4841
4842 if (ts->val_type == TEMP_VAL_CONST
4843 && tcg_target_const_match(ts->val, arg_ct->ct, ts->type,
4844 op_cond, TCGOP_VECE(op))) {
4845 /* constant is OK for instruction */
4846 const_args[i] = 1;
4847 new_args[i] = ts->val;
4848 continue;
4849 }
4850
4851 reg = ts->reg;
4852 i_preferred_regs = 0;
4853 i_required_regs = arg_ct->regs;
4854 allocate_new_reg = false;
4855 copyto_new_reg = false;
4856
4857 switch (arg_ct->pair) {
4858 case 0: /* not paired */
4859 if (arg_ct->ialias) {
4860 i_preferred_regs = output_pref(op, arg_ct->alias_index);
4861
4862 /*
4863 * If the input is readonly, then it cannot also be an
4864 * output and aliased to itself. If the input is not
4865 * dead after the instruction, we must allocate a new
4866 * register and move it.
4867 */
4868 if (temp_readonly(ts) || !IS_DEAD_ARG(i)
4869 || def->args_ct[arg_ct->alias_index].newreg) {
4870 allocate_new_reg = true;
4871 } else if (ts->val_type == TEMP_VAL_REG) {
4872 /*
4873 * Check if the current register has already been
4874 * allocated for another input.
4875 */
4876 allocate_new_reg =
4877 tcg_regset_test_reg(i_allocated_regs, reg);
4878 }
4879 }
4880 if (!allocate_new_reg) {
4881 temp_load(s, ts, i_required_regs, i_allocated_regs,
4882 i_preferred_regs);
4883 reg = ts->reg;
4884 allocate_new_reg = !tcg_regset_test_reg(i_required_regs, reg);
4885 }
4886 if (allocate_new_reg) {
4887 /*
4888 * Allocate a new register matching the constraint
4889 * and move the temporary register into it.
4890 */
4891 temp_load(s, ts, tcg_target_available_regs[ts->type],
4892 i_allocated_regs, 0);
4893 reg = tcg_reg_alloc(s, i_required_regs, i_allocated_regs,
4894 i_preferred_regs, ts->indirect_base);
4895 copyto_new_reg = true;
4896 }
4897 break;
4898
4899 case 1:
4900 /* First of an input pair; if i1 == i2, the second is an output. */
4901 i1 = i;
4902 i2 = arg_ct->pair_index;
4903 ts2 = i1 != i2 ? arg_temp(op->args[i2]) : NULL;
4904
4905 /*
4906 * It is easier to default to allocating a new pair
4907 * and to identify a few cases where it's not required.
4908 */
4909 if (arg_ct->ialias) {
4910 i_preferred_regs = output_pref(op, arg_ct->alias_index);
4911 if (IS_DEAD_ARG(i1) &&
4912 IS_DEAD_ARG(i2) &&
4913 !temp_readonly(ts) &&
4914 ts->val_type == TEMP_VAL_REG &&
4915 ts->reg < TCG_TARGET_NB_REGS - 1 &&
4916 tcg_regset_test_reg(i_required_regs, reg) &&
4917 !tcg_regset_test_reg(i_allocated_regs, reg) &&
4918 !tcg_regset_test_reg(i_allocated_regs, reg + 1) &&
4919 (ts2
4920 ? ts2->val_type == TEMP_VAL_REG &&
4921 ts2->reg == reg + 1 &&
4922 !temp_readonly(ts2)
4923 : s->reg_to_temp[reg + 1] == NULL)) {
4924 break;
4925 }
4926 } else {
4927 /* Without aliasing, the pair must also be an input. */
4928 tcg_debug_assert(ts2);
4929 if (ts->val_type == TEMP_VAL_REG &&
4930 ts2->val_type == TEMP_VAL_REG &&
4931 ts2->reg == reg + 1 &&
4932 tcg_regset_test_reg(i_required_regs, reg)) {
4933 break;
4934 }
4935 }
4936 reg = tcg_reg_alloc_pair(s, i_required_regs, i_allocated_regs,
4937 0, ts->indirect_base);
4938 goto do_pair;
4939
4940 case 2: /* pair second */
4941 reg = new_args[arg_ct->pair_index] + 1;
4942 goto do_pair;
4943
4944 case 3: /* ialias with second output, no first input */
4945 tcg_debug_assert(arg_ct->ialias);
4946 i_preferred_regs = output_pref(op, arg_ct->alias_index);
4947
4948 if (IS_DEAD_ARG(i) &&
4949 !temp_readonly(ts) &&
4950 ts->val_type == TEMP_VAL_REG &&
4951 reg > 0 &&
4952 s->reg_to_temp[reg - 1] == NULL &&
4953 tcg_regset_test_reg(i_required_regs, reg) &&
4954 !tcg_regset_test_reg(i_allocated_regs, reg) &&
4955 !tcg_regset_test_reg(i_allocated_regs, reg - 1)) {
4956 tcg_regset_set_reg(i_allocated_regs, reg - 1);
4957 break;
4958 }
4959 reg = tcg_reg_alloc_pair(s, i_required_regs >> 1,
4960 i_allocated_regs, 0,
4961 ts->indirect_base);
4962 tcg_regset_set_reg(i_allocated_regs, reg);
4963 reg += 1;
4964 goto do_pair;
4965
4966 do_pair:
4967 /*
4968 * If an aliased input is not dead after the instruction,
4969 * we must allocate a new register and move it.
4970 */
4971 if (arg_ct->ialias && (!IS_DEAD_ARG(i) || temp_readonly(ts))) {
4972 TCGRegSet t_allocated_regs = i_allocated_regs;
4973
4974 /*
4975 * Because of the alias, and the continued life, make sure
4976 * that the temp is somewhere *other* than the reg pair,
4977 * and we get a copy in reg.
4978 */
4979 tcg_regset_set_reg(t_allocated_regs, reg);
4980 tcg_regset_set_reg(t_allocated_regs, reg + 1);
4981 if (ts->val_type == TEMP_VAL_REG && ts->reg == reg) {
4982 /* If ts was already in reg, copy it somewhere else. */
4983 TCGReg nr;
4984 bool ok;
4985
4986 tcg_debug_assert(ts->kind != TEMP_FIXED);
4987 nr = tcg_reg_alloc(s, tcg_target_available_regs[ts->type],
4988 t_allocated_regs, 0, ts->indirect_base);
4989 ok = tcg_out_mov(s, ts->type, nr, reg);
4990 tcg_debug_assert(ok);
4991
4992 set_temp_val_reg(s, ts, nr);
4993 } else {
4994 temp_load(s, ts, tcg_target_available_regs[ts->type],
4995 t_allocated_regs, 0);
4996 copyto_new_reg = true;
4997 }
4998 } else {
4999 /* Preferably allocate to reg, otherwise copy. */
5000 i_required_regs = (TCGRegSet)1 << reg;
5001 temp_load(s, ts, i_required_regs, i_allocated_regs,
5002 i_preferred_regs);
5003 copyto_new_reg = ts->reg != reg;
5004 }
5005 break;
5006
5007 default:
5008 g_assert_not_reached();
5009 }
5010
5011 if (copyto_new_reg) {
5012 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) {
5013 /*
5014 * Cross register class move not supported. Sync the
5015 * temp back to its slot and load from there.
5016 */
5017 temp_sync(s, ts, i_allocated_regs, 0, 0);
5018 tcg_out_ld(s, ts->type, reg,
5019 ts->mem_base->reg, ts->mem_offset);
5020 }
5021 }
5022 new_args[i] = reg;
5023 const_args[i] = 0;
5024 tcg_regset_set_reg(i_allocated_regs, reg);
5025 }
5026
5027 /* mark dead temporaries and free the associated registers */
5028 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
5029 if (IS_DEAD_ARG(i)) {
5030 temp_dead(s, arg_temp(op->args[i]));
5031 }
5032 }
5033
5034 if (def->flags & TCG_OPF_COND_BRANCH) {
5035 tcg_reg_alloc_cbranch(s, i_allocated_regs);
5036 } else if (def->flags & TCG_OPF_BB_END) {
5037 tcg_reg_alloc_bb_end(s, i_allocated_regs);
5038 } else {
5039 if (def->flags & TCG_OPF_CALL_CLOBBER) {
5040 /* XXX: permit generic clobber register list ? */
5041 for (i = 0; i < TCG_TARGET_NB_REGS; i++) {
5042 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) {
5043 tcg_reg_free(s, i, i_allocated_regs);
5044 }
5045 }
5046 }
5047 if (def->flags & TCG_OPF_SIDE_EFFECTS) {
5048 /* sync globals if the op has side effects and might trigger
5049 an exception. */
5050 sync_globals(s, i_allocated_regs);
5051 }
5052
5053 /* satisfy the output constraints */
5054 for(k = 0; k < nb_oargs; k++) {
5055 i = def->args_ct[k].sort_index;
5056 arg = op->args[i];
5057 arg_ct = &def->args_ct[i];
5058 ts = arg_temp(arg);
5059
5060 /* ENV should not be modified. */
5061 tcg_debug_assert(!temp_readonly(ts));
5062
5063 switch (arg_ct->pair) {
5064 case 0: /* not paired */
5065 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) {
5066 reg = new_args[arg_ct->alias_index];
5067 } else if (arg_ct->newreg) {
5068 reg = tcg_reg_alloc(s, arg_ct->regs,
5069 i_allocated_regs | o_allocated_regs,
5070 output_pref(op, k), ts->indirect_base);
5071 } else {
5072 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs,
5073 output_pref(op, k), ts->indirect_base);
5074 }
5075 break;
5076
5077 case 1: /* first of pair */
5078 if (arg_ct->oalias) {
5079 reg = new_args[arg_ct->alias_index];
5080 } else if (arg_ct->newreg) {
5081 reg = tcg_reg_alloc_pair(s, arg_ct->regs,
5082 i_allocated_regs | o_allocated_regs,
5083 output_pref(op, k),
5084 ts->indirect_base);
5085 } else {
5086 reg = tcg_reg_alloc_pair(s, arg_ct->regs, o_allocated_regs,
5087 output_pref(op, k),
5088 ts->indirect_base);
5089 }
5090 break;
5091
5092 case 2: /* second of pair */
5093 if (arg_ct->oalias) {
5094 reg = new_args[arg_ct->alias_index];
5095 } else {
5096 reg = new_args[arg_ct->pair_index] + 1;
5097 }
5098 break;
5099
5100 case 3: /* first of pair, aliasing with a second input */
5101 tcg_debug_assert(!arg_ct->newreg);
5102 reg = new_args[arg_ct->pair_index] - 1;
5103 break;
5104
5105 default:
5106 g_assert_not_reached();
5107 }
5108 tcg_regset_set_reg(o_allocated_regs, reg);
5109 set_temp_val_reg(s, ts, reg);
5110 ts->mem_coherent = 0;
5111 new_args[i] = reg;
5112 }
5113 }
5114
5115 /* emit instruction */
5116 switch (op->opc) {
5117 case INDEX_op_ext8s_i32:
5118 tcg_out_ext8s(s, TCG_TYPE_I32, new_args[0], new_args[1]);
5119 break;
5120 case INDEX_op_ext8s_i64:
5121 tcg_out_ext8s(s, TCG_TYPE_I64, new_args[0], new_args[1]);
5122 break;
5123 case INDEX_op_ext8u_i32:
5124 case INDEX_op_ext8u_i64:
5125 tcg_out_ext8u(s, new_args[0], new_args[1]);
5126 break;
5127 case INDEX_op_ext16s_i32:
5128 tcg_out_ext16s(s, TCG_TYPE_I32, new_args[0], new_args[1]);
5129 break;
5130 case INDEX_op_ext16s_i64:
5131 tcg_out_ext16s(s, TCG_TYPE_I64, new_args[0], new_args[1]);
5132 break;
5133 case INDEX_op_ext16u_i32:
5134 case INDEX_op_ext16u_i64:
5135 tcg_out_ext16u(s, new_args[0], new_args[1]);
5136 break;
5137 case INDEX_op_ext32s_i64:
5138 tcg_out_ext32s(s, new_args[0], new_args[1]);
5139 break;
5140 case INDEX_op_ext32u_i64:
5141 tcg_out_ext32u(s, new_args[0], new_args[1]);
5142 break;
5143 case INDEX_op_ext_i32_i64:
5144 tcg_out_exts_i32_i64(s, new_args[0], new_args[1]);
5145 break;
5146 case INDEX_op_extu_i32_i64:
5147 tcg_out_extu_i32_i64(s, new_args[0], new_args[1]);
5148 break;
5149 case INDEX_op_extrl_i64_i32:
5150 tcg_out_extrl_i64_i32(s, new_args[0], new_args[1]);
5151 break;
5152 default:
5153 if (def->flags & TCG_OPF_VECTOR) {
5154 tcg_out_vec_op(s, op->opc, TCGOP_VECL(op), TCGOP_VECE(op),
5155 new_args, const_args);
5156 } else {
5157 tcg_out_op(s, op->opc, new_args, const_args);
5158 }
5159 break;
5160 }
5161
5162 /* move the outputs in the correct register if needed */
5163 for(i = 0; i < nb_oargs; i++) {
5164 ts = arg_temp(op->args[i]);
5165
5166 /* ENV should not be modified. */
5167 tcg_debug_assert(!temp_readonly(ts));
5168
5169 if (NEED_SYNC_ARG(i)) {
5170 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i));
5171 } else if (IS_DEAD_ARG(i)) {
5172 temp_dead(s, ts);
5173 }
5174 }
5175 }
5176
5177 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op)
5178 {
5179 const TCGLifeData arg_life = op->life;
5180 TCGTemp *ots, *itsl, *itsh;
5181 TCGType vtype = TCGOP_VECL(op) + TCG_TYPE_V64;
5182
5183 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */
5184 tcg_debug_assert(TCG_TARGET_REG_BITS == 32);
5185 tcg_debug_assert(TCGOP_VECE(op) == MO_64);
5186
5187 ots = arg_temp(op->args[0]);
5188 itsl = arg_temp(op->args[1]);
5189 itsh = arg_temp(op->args[2]);
5190
5191 /* ENV should not be modified. */
5192 tcg_debug_assert(!temp_readonly(ots));
5193
5194 /* Allocate the output register now. */
5195 if (ots->val_type != TEMP_VAL_REG) {
5196 TCGRegSet allocated_regs = s->reserved_regs;
5197 TCGRegSet dup_out_regs =
5198 tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs;
5199 TCGReg oreg;
5200
5201 /* Make sure to not spill the input registers. */
5202 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) {
5203 tcg_regset_set_reg(allocated_regs, itsl->reg);
5204 }
5205 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) {
5206 tcg_regset_set_reg(allocated_regs, itsh->reg);
5207 }
5208
5209 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs,
5210 output_pref(op, 0), ots->indirect_base);
5211 set_temp_val_reg(s, ots, oreg);
5212 }
5213
5214 /* Promote dup2 of immediates to dupi_vec. */
5215 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) {
5216 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val);
5217 MemOp vece = MO_64;
5218
5219 if (val == dup_const(MO_8, val)) {
5220 vece = MO_8;
5221 } else if (val == dup_const(MO_16, val)) {
5222 vece = MO_16;
5223 } else if (val == dup_const(MO_32, val)) {
5224 vece = MO_32;
5225 }
5226
5227 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val);
5228 goto done;
5229 }
5230
5231 /* If the two inputs form one 64-bit value, try dupm_vec. */
5232 if (itsl->temp_subindex == HOST_BIG_ENDIAN &&
5233 itsh->temp_subindex == !HOST_BIG_ENDIAN &&
5234 itsl == itsh + (HOST_BIG_ENDIAN ? 1 : -1)) {
5235 TCGTemp *its = itsl - HOST_BIG_ENDIAN;
5236
5237 temp_sync(s, its + 0, s->reserved_regs, 0, 0);
5238 temp_sync(s, its + 1, s->reserved_regs, 0, 0);
5239
5240 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg,
5241 its->mem_base->reg, its->mem_offset)) {
5242 goto done;
5243 }
5244 }
5245
5246 /* Fall back to generic expansion. */
5247 return false;
5248
5249 done:
5250 ots->mem_coherent = 0;
5251 if (IS_DEAD_ARG(1)) {
5252 temp_dead(s, itsl);
5253 }
5254 if (IS_DEAD_ARG(2)) {
5255 temp_dead(s, itsh);
5256 }
5257 if (NEED_SYNC_ARG(0)) {
5258 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0));
5259 } else if (IS_DEAD_ARG(0)) {
5260 temp_dead(s, ots);
5261 }
5262 return true;
5263 }
5264
5265 static void load_arg_reg(TCGContext *s, TCGReg reg, TCGTemp *ts,
5266 TCGRegSet allocated_regs)
5267 {
5268 if (ts->val_type == TEMP_VAL_REG) {
5269 if (ts->reg != reg) {
5270 tcg_reg_free(s, reg, allocated_regs);
5271 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) {
5272 /*
5273 * Cross register class move not supported. Sync the
5274 * temp back to its slot and load from there.
5275 */
5276 temp_sync(s, ts, allocated_regs, 0, 0);
5277 tcg_out_ld(s, ts->type, reg,
5278 ts->mem_base->reg, ts->mem_offset);
5279 }
5280 }
5281 } else {
5282 TCGRegSet arg_set = 0;
5283
5284 tcg_reg_free(s, reg, allocated_regs);
5285 tcg_regset_set_reg(arg_set, reg);
5286 temp_load(s, ts, arg_set, allocated_regs, 0);
5287 }
5288 }
5289
5290 static void load_arg_stk(TCGContext *s, unsigned arg_slot, TCGTemp *ts,
5291 TCGRegSet allocated_regs)
5292 {
5293 /*
5294 * When the destination is on the stack, load up the temp and store.
5295 * If there are many call-saved registers, the temp might live to
5296 * see another use; otherwise it'll be discarded.
5297 */
5298 temp_load(s, ts, tcg_target_available_regs[ts->type], allocated_regs, 0);
5299 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK,
5300 arg_slot_stk_ofs(arg_slot));
5301 }
5302
5303 static void load_arg_normal(TCGContext *s, const TCGCallArgumentLoc *l,
5304 TCGTemp *ts, TCGRegSet *allocated_regs)
5305 {
5306 if (arg_slot_reg_p(l->arg_slot)) {
5307 TCGReg reg = tcg_target_call_iarg_regs[l->arg_slot];
5308 load_arg_reg(s, reg, ts, *allocated_regs);
5309 tcg_regset_set_reg(*allocated_regs, reg);
5310 } else {
5311 load_arg_stk(s, l->arg_slot, ts, *allocated_regs);
5312 }
5313 }
5314
5315 static void load_arg_ref(TCGContext *s, unsigned arg_slot, TCGReg ref_base,
5316 intptr_t ref_off, TCGRegSet *allocated_regs)
5317 {
5318 TCGReg reg;
5319
5320 if (arg_slot_reg_p(arg_slot)) {
5321 reg = tcg_target_call_iarg_regs[arg_slot];
5322 tcg_reg_free(s, reg, *allocated_regs);
5323 tcg_out_addi_ptr(s, reg, ref_base, ref_off);
5324 tcg_regset_set_reg(*allocated_regs, reg);
5325 } else {
5326 reg = tcg_reg_alloc(s, tcg_target_available_regs[TCG_TYPE_PTR],
5327 *allocated_regs, 0, false);
5328 tcg_out_addi_ptr(s, reg, ref_base, ref_off);
5329 tcg_out_st(s, TCG_TYPE_PTR, reg, TCG_REG_CALL_STACK,
5330 arg_slot_stk_ofs(arg_slot));
5331 }
5332 }
5333
5334 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op)
5335 {
5336 const int nb_oargs = TCGOP_CALLO(op);
5337 const int nb_iargs = TCGOP_CALLI(op);
5338 const TCGLifeData arg_life = op->life;
5339 const TCGHelperInfo *info = tcg_call_info(op);
5340 TCGRegSet allocated_regs = s->reserved_regs;
5341 int i;
5342
5343 /*
5344 * Move inputs into place in reverse order,
5345 * so that we place stacked arguments first.
5346 */
5347 for (i = nb_iargs - 1; i >= 0; --i) {
5348 const TCGCallArgumentLoc *loc = &info->in[i];
5349 TCGTemp *ts = arg_temp(op->args[nb_oargs + i]);
5350
5351 switch (loc->kind) {
5352 case TCG_CALL_ARG_NORMAL:
5353 case TCG_CALL_ARG_EXTEND_U:
5354 case TCG_CALL_ARG_EXTEND_S:
5355 load_arg_normal(s, loc, ts, &allocated_regs);
5356 break;
5357 case TCG_CALL_ARG_BY_REF:
5358 load_arg_stk(s, loc->ref_slot, ts, allocated_regs);
5359 load_arg_ref(s, loc->arg_slot, TCG_REG_CALL_STACK,
5360 arg_slot_stk_ofs(loc->ref_slot),
5361 &allocated_regs);
5362 break;
5363 case TCG_CALL_ARG_BY_REF_N:
5364 load_arg_stk(s, loc->ref_slot, ts, allocated_regs);
5365 break;
5366 default:
5367 g_assert_not_reached();
5368 }
5369 }
5370
5371 /* Mark dead temporaries and free the associated registers. */
5372 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
5373 if (IS_DEAD_ARG(i)) {
5374 temp_dead(s, arg_temp(op->args[i]));
5375 }
5376 }
5377
5378 /* Clobber call registers. */
5379 for (i = 0; i < TCG_TARGET_NB_REGS; i++) {
5380 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) {
5381 tcg_reg_free(s, i, allocated_regs);
5382 }
5383 }
5384
5385 /*
5386 * Save globals if they might be written by the helper,
5387 * sync them if they might be read.
5388 */
5389 if (info->flags & TCG_CALL_NO_READ_GLOBALS) {
5390 /* Nothing to do */
5391 } else if (info->flags & TCG_CALL_NO_WRITE_GLOBALS) {
5392 sync_globals(s, allocated_regs);
5393 } else {
5394 save_globals(s, allocated_regs);
5395 }
5396
5397 /*
5398 * If the ABI passes a pointer to the returned struct as the first
5399 * argument, load that now. Pass a pointer to the output home slot.
5400 */
5401 if (info->out_kind == TCG_CALL_RET_BY_REF) {
5402 TCGTemp *ts = arg_temp(op->args[0]);
5403
5404 if (!ts->mem_allocated) {
5405 temp_allocate_frame(s, ts);
5406 }
5407 load_arg_ref(s, 0, ts->mem_base->reg, ts->mem_offset, &allocated_regs);
5408 }
5409
5410 tcg_out_call(s, tcg_call_func(op), info);
5411
5412 /* Assign output registers and emit moves if needed. */
5413 switch (info->out_kind) {
5414 case TCG_CALL_RET_NORMAL:
5415 for (i = 0; i < nb_oargs; i++) {
5416 TCGTemp *ts = arg_temp(op->args[i]);
5417 TCGReg reg = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, i);
5418
5419 /* ENV should not be modified. */
5420 tcg_debug_assert(!temp_readonly(ts));
5421
5422 set_temp_val_reg(s, ts, reg);
5423 ts->mem_coherent = 0;
5424 }
5425 break;
5426
5427 case TCG_CALL_RET_BY_VEC:
5428 {
5429 TCGTemp *ts = arg_temp(op->args[0]);
5430
5431 tcg_debug_assert(ts->base_type == TCG_TYPE_I128);
5432 tcg_debug_assert(ts->temp_subindex == 0);
5433 if (!ts->mem_allocated) {
5434 temp_allocate_frame(s, ts);
5435 }
5436 tcg_out_st(s, TCG_TYPE_V128,
5437 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0),
5438 ts->mem_base->reg, ts->mem_offset);
5439 }
5440 /* fall through to mark all parts in memory */
5441
5442 case TCG_CALL_RET_BY_REF:
5443 /* The callee has performed a write through the reference. */
5444 for (i = 0; i < nb_oargs; i++) {
5445 TCGTemp *ts = arg_temp(op->args[i]);
5446 ts->val_type = TEMP_VAL_MEM;
5447 }
5448 break;
5449
5450 default:
5451 g_assert_not_reached();
5452 }
5453
5454 /* Flush or discard output registers as needed. */
5455 for (i = 0; i < nb_oargs; i++) {
5456 TCGTemp *ts = arg_temp(op->args[i]);
5457 if (NEED_SYNC_ARG(i)) {
5458 temp_sync(s, ts, s->reserved_regs, 0, IS_DEAD_ARG(i));
5459 } else if (IS_DEAD_ARG(i)) {
5460 temp_dead(s, ts);
5461 }
5462 }
5463 }
5464
5465 /**
5466 * atom_and_align_for_opc:
5467 * @s: tcg context
5468 * @opc: memory operation code
5469 * @host_atom: MO_ATOM_{IFALIGN,WITHIN16,SUBALIGN} for host operations
5470 * @allow_two_ops: true if we are prepared to issue two operations
5471 *
5472 * Return the alignment and atomicity to use for the inline fast path
5473 * for the given memory operation. The alignment may be larger than
5474 * that specified in @opc, and the correct alignment will be diagnosed
5475 * by the slow path helper.
5476 *
5477 * If @allow_two_ops, the host is prepared to test for 2x alignment,
5478 * and issue two loads or stores for subalignment.
5479 */
5480 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc,
5481 MemOp host_atom, bool allow_two_ops)
5482 {
5483 MemOp align = get_alignment_bits(opc);
5484 MemOp size = opc & MO_SIZE;
5485 MemOp half = size ? size - 1 : 0;
5486 MemOp atom = opc & MO_ATOM_MASK;
5487 MemOp atmax;
5488
5489 switch (atom) {
5490 case MO_ATOM_NONE:
5491 /* The operation requires no specific atomicity. */
5492 atmax = MO_8;
5493 break;
5494
5495 case MO_ATOM_IFALIGN:
5496 atmax = size;
5497 break;
5498
5499 case MO_ATOM_IFALIGN_PAIR:
5500 atmax = half;
5501 break;
5502
5503 case MO_ATOM_WITHIN16:
5504 atmax = size;
5505 if (size == MO_128) {
5506 /* Misalignment implies !within16, and therefore no atomicity. */
5507 } else if (host_atom != MO_ATOM_WITHIN16) {
5508 /* The host does not implement within16, so require alignment. */
5509 align = MAX(align, size);
5510 }
5511 break;
5512
5513 case MO_ATOM_WITHIN16_PAIR:
5514 atmax = size;
5515 /*
5516 * Misalignment implies !within16, and therefore half atomicity.
5517 * Any host prepared for two operations can implement this with
5518 * half alignment.
5519 */
5520 if (host_atom != MO_ATOM_WITHIN16 && allow_two_ops) {
5521 align = MAX(align, half);
5522 }
5523 break;
5524
5525 case MO_ATOM_SUBALIGN:
5526 atmax = size;
5527 if (host_atom != MO_ATOM_SUBALIGN) {
5528 /* If unaligned but not odd, there are subobjects up to half. */
5529 if (allow_two_ops) {
5530 align = MAX(align, half);
5531 } else {
5532 align = MAX(align, size);
5533 }
5534 }
5535 break;
5536
5537 default:
5538 g_assert_not_reached();
5539 }
5540
5541 return (TCGAtomAlign){ .atom = atmax, .align = align };
5542 }
5543
5544 /*
5545 * Similarly for qemu_ld/st slow path helpers.
5546 * We must re-implement tcg_gen_callN and tcg_reg_alloc_call simultaneously,
5547 * using only the provided backend tcg_out_* functions.
5548 */
5549
5550 static int tcg_out_helper_stk_ofs(TCGType type, unsigned slot)
5551 {
5552 int ofs = arg_slot_stk_ofs(slot);
5553
5554 /*
5555 * Each stack slot is TCG_TARGET_LONG_BITS. If the host does not
5556 * require extension to uint64_t, adjust the address for uint32_t.
5557 */
5558 if (HOST_BIG_ENDIAN &&
5559 TCG_TARGET_REG_BITS == 64 &&
5560 type == TCG_TYPE_I32) {
5561 ofs += 4;
5562 }
5563 return ofs;
5564 }
5565
5566 static void tcg_out_helper_load_slots(TCGContext *s,
5567 unsigned nmov, TCGMovExtend *mov,
5568 const TCGLdstHelperParam *parm)
5569 {
5570 unsigned i;
5571 TCGReg dst3;
5572
5573 /*
5574 * Start from the end, storing to the stack first.
5575 * This frees those registers, so we need not consider overlap.
5576 */
5577 for (i = nmov; i-- > 0; ) {
5578 unsigned slot = mov[i].dst;
5579
5580 if (arg_slot_reg_p(slot)) {
5581 goto found_reg;
5582 }
5583
5584 TCGReg src = mov[i].src;
5585 TCGType dst_type = mov[i].dst_type;
5586 MemOp dst_mo = dst_type == TCG_TYPE_I32 ? MO_32 : MO_64;
5587
5588 /* The argument is going onto the stack; extend into scratch. */
5589 if ((mov[i].src_ext & MO_SIZE) != dst_mo) {
5590 tcg_debug_assert(parm->ntmp != 0);
5591 mov[i].dst = src = parm->tmp[0];
5592 tcg_out_movext1(s, &mov[i]);
5593 }
5594
5595 tcg_out_st(s, dst_type, src, TCG_REG_CALL_STACK,
5596 tcg_out_helper_stk_ofs(dst_type, slot));
5597 }
5598 return;
5599
5600 found_reg:
5601 /*
5602 * The remaining arguments are in registers.
5603 * Convert slot numbers to argument registers.
5604 */
5605 nmov = i + 1;
5606 for (i = 0; i < nmov; ++i) {
5607 mov[i].dst = tcg_target_call_iarg_regs[mov[i].dst];
5608 }
5609
5610 switch (nmov) {
5611 case 4:
5612 /* The backend must have provided enough temps for the worst case. */
5613 tcg_debug_assert(parm->ntmp >= 2);
5614
5615 dst3 = mov[3].dst;
5616 for (unsigned j = 0; j < 3; ++j) {
5617 if (dst3 == mov[j].src) {
5618 /*
5619 * Conflict. Copy the source to a temporary, perform the
5620 * remaining moves, then the extension from our scratch
5621 * on the way out.
5622 */
5623 TCGReg scratch = parm->tmp[1];
5624
5625 tcg_out_mov(s, mov[3].src_type, scratch, mov[3].src);
5626 tcg_out_movext3(s, mov, mov + 1, mov + 2, parm->tmp[0]);
5627 tcg_out_movext1_new_src(s, &mov[3], scratch);
5628 break;
5629 }
5630 }
5631
5632 /* No conflicts: perform this move and continue. */
5633 tcg_out_movext1(s, &mov[3]);
5634 /* fall through */
5635
5636 case 3:
5637 tcg_out_movext3(s, mov, mov + 1, mov + 2,
5638 parm->ntmp ? parm->tmp[0] : -1);
5639 break;
5640 case 2:
5641 tcg_out_movext2(s, mov, mov + 1,
5642 parm->ntmp ? parm->tmp[0] : -1);
5643 break;
5644 case 1:
5645 tcg_out_movext1(s, mov);
5646 break;
5647 default:
5648 g_assert_not_reached();
5649 }
5650 }
5651
5652 static void tcg_out_helper_load_imm(TCGContext *s, unsigned slot,
5653 TCGType type, tcg_target_long imm,
5654 const TCGLdstHelperParam *parm)
5655 {
5656 if (arg_slot_reg_p(slot)) {
5657 tcg_out_movi(s, type, tcg_target_call_iarg_regs[slot], imm);
5658 } else {
5659 int ofs = tcg_out_helper_stk_ofs(type, slot);
5660 if (!tcg_out_sti(s, type, imm, TCG_REG_CALL_STACK, ofs)) {
5661 tcg_debug_assert(parm->ntmp != 0);
5662 tcg_out_movi(s, type, parm->tmp[0], imm);
5663 tcg_out_st(s, type, parm->tmp[0], TCG_REG_CALL_STACK, ofs);
5664 }
5665 }
5666 }
5667
5668 static void tcg_out_helper_load_common_args(TCGContext *s,
5669 const TCGLabelQemuLdst *ldst,
5670 const TCGLdstHelperParam *parm,
5671 const TCGHelperInfo *info,
5672 unsigned next_arg)
5673 {
5674 TCGMovExtend ptr_mov = {
5675 .dst_type = TCG_TYPE_PTR,
5676 .src_type = TCG_TYPE_PTR,
5677 .src_ext = sizeof(void *) == 4 ? MO_32 : MO_64
5678 };
5679 const TCGCallArgumentLoc *loc = &info->in[0];
5680 TCGType type;
5681 unsigned slot;
5682 tcg_target_ulong imm;
5683
5684 /*
5685 * Handle env, which is always first.
5686 */
5687 ptr_mov.dst = loc->arg_slot;
5688 ptr_mov.src = TCG_AREG0;
5689 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm);
5690
5691 /*
5692 * Handle oi.
5693 */
5694 imm = ldst->oi;
5695 loc = &info->in[next_arg];
5696 type = TCG_TYPE_I32;
5697 switch (loc->kind) {
5698 case TCG_CALL_ARG_NORMAL:
5699 break;
5700 case TCG_CALL_ARG_EXTEND_U:
5701 case TCG_CALL_ARG_EXTEND_S:
5702 /* No extension required for MemOpIdx. */
5703 tcg_debug_assert(imm <= INT32_MAX);
5704 type = TCG_TYPE_REG;
5705 break;
5706 default:
5707 g_assert_not_reached();
5708 }
5709 tcg_out_helper_load_imm(s, loc->arg_slot, type, imm, parm);
5710 next_arg++;
5711
5712 /*
5713 * Handle ra.
5714 */
5715 loc = &info->in[next_arg];
5716 slot = loc->arg_slot;
5717 if (parm->ra_gen) {
5718 int arg_reg = -1;
5719 TCGReg ra_reg;
5720
5721 if (arg_slot_reg_p(slot)) {
5722 arg_reg = tcg_target_call_iarg_regs[slot];
5723 }
5724 ra_reg = parm->ra_gen(s, ldst, arg_reg);
5725
5726 ptr_mov.dst = slot;
5727 ptr_mov.src = ra_reg;
5728 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm);
5729 } else {
5730 imm = (uintptr_t)ldst->raddr;
5731 tcg_out_helper_load_imm(s, slot, TCG_TYPE_PTR, imm, parm);
5732 }
5733 }
5734
5735 static unsigned tcg_out_helper_add_mov(TCGMovExtend *mov,
5736 const TCGCallArgumentLoc *loc,
5737 TCGType dst_type, TCGType src_type,
5738 TCGReg lo, TCGReg hi)
5739 {
5740 MemOp reg_mo;
5741
5742 if (dst_type <= TCG_TYPE_REG) {
5743 MemOp src_ext;
5744
5745 switch (loc->kind) {
5746 case TCG_CALL_ARG_NORMAL:
5747 src_ext = src_type == TCG_TYPE_I32 ? MO_32 : MO_64;
5748 break;
5749 case TCG_CALL_ARG_EXTEND_U:
5750 dst_type = TCG_TYPE_REG;
5751 src_ext = MO_UL;
5752 break;
5753 case TCG_CALL_ARG_EXTEND_S:
5754 dst_type = TCG_TYPE_REG;
5755 src_ext = MO_SL;
5756 break;
5757 default:
5758 g_assert_not_reached();
5759 }
5760
5761 mov[0].dst = loc->arg_slot;
5762 mov[0].dst_type = dst_type;
5763 mov[0].src = lo;
5764 mov[0].src_type = src_type;
5765 mov[0].src_ext = src_ext;
5766 return 1;
5767 }
5768
5769 if (TCG_TARGET_REG_BITS == 32) {
5770 assert(dst_type == TCG_TYPE_I64);
5771 reg_mo = MO_32;
5772 } else {
5773 assert(dst_type == TCG_TYPE_I128);
5774 reg_mo = MO_64;
5775 }
5776
5777 mov[0].dst = loc[HOST_BIG_ENDIAN].arg_slot;
5778 mov[0].src = lo;
5779 mov[0].dst_type = TCG_TYPE_REG;
5780 mov[0].src_type = TCG_TYPE_REG;
5781 mov[0].src_ext = reg_mo;
5782
5783 mov[1].dst = loc[!HOST_BIG_ENDIAN].arg_slot;
5784 mov[1].src = hi;
5785 mov[1].dst_type = TCG_TYPE_REG;
5786 mov[1].src_type = TCG_TYPE_REG;
5787 mov[1].src_ext = reg_mo;
5788
5789 return 2;
5790 }
5791
5792 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst,
5793 const TCGLdstHelperParam *parm)
5794 {
5795 const TCGHelperInfo *info;
5796 const TCGCallArgumentLoc *loc;
5797 TCGMovExtend mov[2];
5798 unsigned next_arg, nmov;
5799 MemOp mop = get_memop(ldst->oi);
5800
5801 switch (mop & MO_SIZE) {
5802 case MO_8:
5803 case MO_16:
5804 case MO_32:
5805 info = &info_helper_ld32_mmu;
5806 break;
5807 case MO_64:
5808 info = &info_helper_ld64_mmu;
5809 break;
5810 case MO_128:
5811 info = &info_helper_ld128_mmu;
5812 break;
5813 default:
5814 g_assert_not_reached();
5815 }
5816
5817 /* Defer env argument. */
5818 next_arg = 1;
5819
5820 loc = &info->in[next_arg];
5821 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) {
5822 /*
5823 * 32-bit host with 32-bit guest: zero-extend the guest address
5824 * to 64-bits for the helper by storing the low part, then
5825 * load a zero for the high part.
5826 */
5827 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN,
5828 TCG_TYPE_I32, TCG_TYPE_I32,
5829 ldst->addrlo_reg, -1);
5830 tcg_out_helper_load_slots(s, 1, mov, parm);
5831
5832 tcg_out_helper_load_imm(s, loc[!HOST_BIG_ENDIAN].arg_slot,
5833 TCG_TYPE_I32, 0, parm);
5834 next_arg += 2;
5835 } else {
5836 nmov = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type,
5837 ldst->addrlo_reg, ldst->addrhi_reg);
5838 tcg_out_helper_load_slots(s, nmov, mov, parm);
5839 next_arg += nmov;
5840 }
5841
5842 switch (info->out_kind) {
5843 case TCG_CALL_RET_NORMAL:
5844 case TCG_CALL_RET_BY_VEC:
5845 break;
5846 case TCG_CALL_RET_BY_REF:
5847 /*
5848 * The return reference is in the first argument slot.
5849 * We need memory in which to return: re-use the top of stack.
5850 */
5851 {
5852 int ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET;
5853
5854 if (arg_slot_reg_p(0)) {
5855 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[0],
5856 TCG_REG_CALL_STACK, ofs_slot0);
5857 } else {
5858 tcg_debug_assert(parm->ntmp != 0);
5859 tcg_out_addi_ptr(s, parm->tmp[0],
5860 TCG_REG_CALL_STACK, ofs_slot0);
5861 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0],
5862 TCG_REG_CALL_STACK, ofs_slot0);
5863 }
5864 }
5865 break;
5866 default:
5867 g_assert_not_reached();
5868 }
5869
5870 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg);
5871 }
5872
5873 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *ldst,
5874 bool load_sign,
5875 const TCGLdstHelperParam *parm)
5876 {
5877 MemOp mop = get_memop(ldst->oi);
5878 TCGMovExtend mov[2];
5879 int ofs_slot0;
5880
5881 switch (ldst->type) {
5882 case TCG_TYPE_I64:
5883 if (TCG_TARGET_REG_BITS == 32) {
5884 break;
5885 }
5886 /* fall through */
5887
5888 case TCG_TYPE_I32:
5889 mov[0].dst = ldst->datalo_reg;
5890 mov[0].src = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, 0);
5891 mov[0].dst_type = ldst->type;
5892 mov[0].src_type = TCG_TYPE_REG;
5893
5894 /*
5895 * If load_sign, then we allowed the helper to perform the
5896 * appropriate sign extension to tcg_target_ulong, and all
5897 * we need now is a plain move.
5898 *
5899 * If they do not, then we expect the relevant extension
5900 * instruction to be no more expensive than a move, and
5901 * we thus save the icache etc by only using one of two
5902 * helper functions.
5903 */
5904 if (load_sign || !(mop & MO_SIGN)) {
5905 if (TCG_TARGET_REG_BITS == 32 || ldst->type == TCG_TYPE_I32) {
5906 mov[0].src_ext = MO_32;
5907 } else {
5908 mov[0].src_ext = MO_64;
5909 }
5910 } else {
5911 mov[0].src_ext = mop & MO_SSIZE;
5912 }
5913 tcg_out_movext1(s, mov);
5914 return;
5915
5916 case TCG_TYPE_I128:
5917 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
5918 ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET;
5919 switch (TCG_TARGET_CALL_RET_I128) {
5920 case TCG_CALL_RET_NORMAL:
5921 break;
5922 case TCG_CALL_RET_BY_VEC:
5923 tcg_out_st(s, TCG_TYPE_V128,
5924 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0),
5925 TCG_REG_CALL_STACK, ofs_slot0);
5926 /* fall through */
5927 case TCG_CALL_RET_BY_REF:
5928 tcg_out_ld(s, TCG_TYPE_I64, ldst->datalo_reg,
5929 TCG_REG_CALL_STACK, ofs_slot0 + 8 * HOST_BIG_ENDIAN);
5930 tcg_out_ld(s, TCG_TYPE_I64, ldst->datahi_reg,
5931 TCG_REG_CALL_STACK, ofs_slot0 + 8 * !HOST_BIG_ENDIAN);
5932 return;
5933 default:
5934 g_assert_not_reached();
5935 }
5936 break;
5937
5938 default:
5939 g_assert_not_reached();
5940 }
5941
5942 mov[0].dst = ldst->datalo_reg;
5943 mov[0].src =
5944 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, HOST_BIG_ENDIAN);
5945 mov[0].dst_type = TCG_TYPE_REG;
5946 mov[0].src_type = TCG_TYPE_REG;
5947 mov[0].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64;
5948
5949 mov[1].dst = ldst->datahi_reg;
5950 mov[1].src =
5951 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, !HOST_BIG_ENDIAN);
5952 mov[1].dst_type = TCG_TYPE_REG;
5953 mov[1].src_type = TCG_TYPE_REG;
5954 mov[1].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64;
5955
5956 tcg_out_movext2(s, mov, mov + 1, parm->ntmp ? parm->tmp[0] : -1);
5957 }
5958
5959 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst,
5960 const TCGLdstHelperParam *parm)
5961 {
5962 const TCGHelperInfo *info;
5963 const TCGCallArgumentLoc *loc;
5964 TCGMovExtend mov[4];
5965 TCGType data_type;
5966 unsigned next_arg, nmov, n;
5967 MemOp mop = get_memop(ldst->oi);
5968
5969 switch (mop & MO_SIZE) {
5970 case MO_8:
5971 case MO_16:
5972 case MO_32:
5973 info = &info_helper_st32_mmu;
5974 data_type = TCG_TYPE_I32;
5975 break;
5976 case MO_64:
5977 info = &info_helper_st64_mmu;
5978 data_type = TCG_TYPE_I64;
5979 break;
5980 case MO_128:
5981 info = &info_helper_st128_mmu;
5982 data_type = TCG_TYPE_I128;
5983 break;
5984 default:
5985 g_assert_not_reached();
5986 }
5987
5988 /* Defer env argument. */
5989 next_arg = 1;
5990 nmov = 0;
5991
5992 /* Handle addr argument. */
5993 loc = &info->in[next_arg];
5994 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) {
5995 /*
5996 * 32-bit host with 32-bit guest: zero-extend the guest address
5997 * to 64-bits for the helper by storing the low part. Later,
5998 * after we have processed the register inputs, we will load a
5999 * zero for the high part.
6000 */
6001 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN,
6002 TCG_TYPE_I32, TCG_TYPE_I32,
6003 ldst->addrlo_reg, -1);
6004 next_arg += 2;
6005 nmov += 1;
6006 } else {
6007 n = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type,
6008 ldst->addrlo_reg, ldst->addrhi_reg);
6009 next_arg += n;
6010 nmov += n;
6011 }
6012
6013 /* Handle data argument. */
6014 loc = &info->in[next_arg];
6015 switch (loc->kind) {
6016 case TCG_CALL_ARG_NORMAL:
6017 case TCG_CALL_ARG_EXTEND_U:
6018 case TCG_CALL_ARG_EXTEND_S:
6019 n = tcg_out_helper_add_mov(mov + nmov, loc, data_type, ldst->type,
6020 ldst->datalo_reg, ldst->datahi_reg);
6021 next_arg += n;
6022 nmov += n;
6023 tcg_out_helper_load_slots(s, nmov, mov, parm);
6024 break;
6025
6026 case TCG_CALL_ARG_BY_REF:
6027 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
6028 tcg_debug_assert(data_type == TCG_TYPE_I128);
6029 tcg_out_st(s, TCG_TYPE_I64,
6030 HOST_BIG_ENDIAN ? ldst->datahi_reg : ldst->datalo_reg,
6031 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[0].ref_slot));
6032 tcg_out_st(s, TCG_TYPE_I64,
6033 HOST_BIG_ENDIAN ? ldst->datalo_reg : ldst->datahi_reg,
6034 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[1].ref_slot));
6035
6036 tcg_out_helper_load_slots(s, nmov, mov, parm);
6037
6038 if (arg_slot_reg_p(loc->arg_slot)) {
6039 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[loc->arg_slot],
6040 TCG_REG_CALL_STACK,
6041 arg_slot_stk_ofs(loc->ref_slot));
6042 } else {
6043 tcg_debug_assert(parm->ntmp != 0);
6044 tcg_out_addi_ptr(s, parm->tmp[0], TCG_REG_CALL_STACK,
6045 arg_slot_stk_ofs(loc->ref_slot));
6046 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0],
6047 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc->arg_slot));
6048 }
6049 next_arg += 2;
6050 break;
6051
6052 default:
6053 g_assert_not_reached();
6054 }
6055
6056 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) {
6057 /* Zero extend the address by loading a zero for the high part. */
6058 loc = &info->in[1 + !HOST_BIG_ENDIAN];
6059 tcg_out_helper_load_imm(s, loc->arg_slot, TCG_TYPE_I32, 0, parm);
6060 }
6061
6062 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg);
6063 }
6064
6065 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start)
6066 {
6067 int i, start_words, num_insns;
6068 TCGOp *op;
6069
6070 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP)
6071 && qemu_log_in_addr_range(pc_start))) {
6072 FILE *logfile = qemu_log_trylock();
6073 if (logfile) {
6074 fprintf(logfile, "OP:\n");
6075 tcg_dump_ops(s, logfile, false);
6076 fprintf(logfile, "\n");
6077 qemu_log_unlock(logfile);
6078 }
6079 }
6080
6081 #ifdef CONFIG_DEBUG_TCG
6082 /* Ensure all labels referenced have been emitted. */
6083 {
6084 TCGLabel *l;
6085 bool error = false;
6086
6087 QSIMPLEQ_FOREACH(l, &s->labels, next) {
6088 if (unlikely(!l->present) && !QSIMPLEQ_EMPTY(&l->branches)) {
6089 qemu_log_mask(CPU_LOG_TB_OP,
6090 "$L%d referenced but not present.\n", l->id);
6091 error = true;
6092 }
6093 }
6094 assert(!error);
6095 }
6096 #endif
6097
6098 tcg_optimize(s);
6099
6100 reachable_code_pass(s);
6101 liveness_pass_0(s);
6102 liveness_pass_1(s);
6103
6104 if (s->nb_indirects > 0) {
6105 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND)
6106 && qemu_log_in_addr_range(pc_start))) {
6107 FILE *logfile = qemu_log_trylock();
6108 if (logfile) {
6109 fprintf(logfile, "OP before indirect lowering:\n");
6110 tcg_dump_ops(s, logfile, false);
6111 fprintf(logfile, "\n");
6112 qemu_log_unlock(logfile);
6113 }
6114 }
6115
6116 /* Replace indirect temps with direct temps. */
6117 if (liveness_pass_2(s)) {
6118 /* If changes were made, re-run liveness. */
6119 liveness_pass_1(s);
6120 }
6121 }
6122
6123 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT)
6124 && qemu_log_in_addr_range(pc_start))) {
6125 FILE *logfile = qemu_log_trylock();
6126 if (logfile) {
6127 fprintf(logfile, "OP after optimization and liveness analysis:\n");
6128 tcg_dump_ops(s, logfile, true);
6129 fprintf(logfile, "\n");
6130 qemu_log_unlock(logfile);
6131 }
6132 }
6133
6134 /* Initialize goto_tb jump offsets. */
6135 tb->jmp_reset_offset[0] = TB_JMP_OFFSET_INVALID;
6136 tb->jmp_reset_offset[1] = TB_JMP_OFFSET_INVALID;
6137 tb->jmp_insn_offset[0] = TB_JMP_OFFSET_INVALID;
6138 tb->jmp_insn_offset[1] = TB_JMP_OFFSET_INVALID;
6139
6140 tcg_reg_alloc_start(s);
6141
6142 /*
6143 * Reset the buffer pointers when restarting after overflow.
6144 * TODO: Move this into translate-all.c with the rest of the
6145 * buffer management. Having only this done here is confusing.
6146 */
6147 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr);
6148 s->code_ptr = s->code_buf;
6149
6150 #ifdef TCG_TARGET_NEED_LDST_LABELS
6151 QSIMPLEQ_INIT(&s->ldst_labels);
6152 #endif
6153 #ifdef TCG_TARGET_NEED_POOL_LABELS
6154 s->pool_labels = NULL;
6155 #endif
6156
6157 start_words = s->insn_start_words;
6158 s->gen_insn_data =
6159 tcg_malloc(sizeof(uint64_t) * s->gen_tb->icount * start_words);
6160
6161 tcg_out_tb_start(s);
6162
6163 num_insns = -1;
6164 QTAILQ_FOREACH(op, &s->ops, link) {
6165 TCGOpcode opc = op->opc;
6166
6167 switch (opc) {
6168 case INDEX_op_mov_i32:
6169 case INDEX_op_mov_i64:
6170 case INDEX_op_mov_vec:
6171 tcg_reg_alloc_mov(s, op);
6172 break;
6173 case INDEX_op_dup_vec:
6174 tcg_reg_alloc_dup(s, op);
6175 break;
6176 case INDEX_op_insn_start:
6177 if (num_insns >= 0) {
6178 size_t off = tcg_current_code_size(s);
6179 s->gen_insn_end_off[num_insns] = off;
6180 /* Assert that we do not overflow our stored offset. */
6181 assert(s->gen_insn_end_off[num_insns] == off);
6182 }
6183 num_insns++;
6184 for (i = 0; i < start_words; ++i) {
6185 s->gen_insn_data[num_insns * start_words + i] =
6186 tcg_get_insn_start_param(op, i);
6187 }
6188 break;
6189 case INDEX_op_discard:
6190 temp_dead(s, arg_temp(op->args[0]));
6191 break;
6192 case INDEX_op_set_label:
6193 tcg_reg_alloc_bb_end(s, s->reserved_regs);
6194 tcg_out_label(s, arg_label(op->args[0]));
6195 break;
6196 case INDEX_op_call:
6197 tcg_reg_alloc_call(s, op);
6198 break;
6199 case INDEX_op_exit_tb:
6200 tcg_out_exit_tb(s, op->args[0]);
6201 break;
6202 case INDEX_op_goto_tb:
6203 tcg_out_goto_tb(s, op->args[0]);
6204 break;
6205 case INDEX_op_dup2_vec:
6206 if (tcg_reg_alloc_dup2(s, op)) {
6207 break;
6208 }
6209 /* fall through */
6210 default:
6211 /* Sanity check that we've not introduced any unhandled opcodes. */
6212 tcg_debug_assert(tcg_op_supported(opc));
6213 /* Note: in order to speed up the code, it would be much
6214 faster to have specialized register allocator functions for
6215 some common argument patterns */
6216 tcg_reg_alloc_op(s, op);
6217 break;
6218 }
6219 /* Test for (pending) buffer overflow. The assumption is that any
6220 one operation beginning below the high water mark cannot overrun
6221 the buffer completely. Thus we can test for overflow after
6222 generating code without having to check during generation. */
6223 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) {
6224 return -1;
6225 }
6226 /* Test for TB overflow, as seen by gen_insn_end_off. */
6227 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) {
6228 return -2;
6229 }
6230 }
6231 tcg_debug_assert(num_insns + 1 == s->gen_tb->icount);
6232 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s);
6233
6234 /* Generate TB finalization at the end of block */
6235 #ifdef TCG_TARGET_NEED_LDST_LABELS
6236 i = tcg_out_ldst_finalize(s);
6237 if (i < 0) {
6238 return i;
6239 }
6240 #endif
6241 #ifdef TCG_TARGET_NEED_POOL_LABELS
6242 i = tcg_out_pool_finalize(s);
6243 if (i < 0) {
6244 return i;
6245 }
6246 #endif
6247 if (!tcg_resolve_relocs(s)) {
6248 return -2;
6249 }
6250
6251 #ifndef CONFIG_TCG_INTERPRETER
6252 /* flush instruction cache */
6253 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf),
6254 (uintptr_t)s->code_buf,
6255 tcg_ptr_byte_diff(s->code_ptr, s->code_buf));
6256 #endif
6257
6258 return tcg_current_code_size(s);
6259 }
6260
6261 #ifdef ELF_HOST_MACHINE
6262 /* In order to use this feature, the backend needs to do three things:
6263
6264 (1) Define ELF_HOST_MACHINE to indicate both what value to
6265 put into the ELF image and to indicate support for the feature.
6266
6267 (2) Define tcg_register_jit. This should create a buffer containing
6268 the contents of a .debug_frame section that describes the post-
6269 prologue unwind info for the tcg machine.
6270
6271 (3) Call tcg_register_jit_int, with the constructed .debug_frame.
6272 */
6273
6274 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */
6275 typedef enum {
6276 JIT_NOACTION = 0,
6277 JIT_REGISTER_FN,
6278 JIT_UNREGISTER_FN
6279 } jit_actions_t;
6280
6281 struct jit_code_entry {
6282 struct jit_code_entry *next_entry;
6283 struct jit_code_entry *prev_entry;
6284 const void *symfile_addr;
6285 uint64_t symfile_size;
6286 };
6287
6288 struct jit_descriptor {
6289 uint32_t version;
6290 uint32_t action_flag;
6291 struct jit_code_entry *relevant_entry;
6292 struct jit_code_entry *first_entry;
6293 };
6294
6295 void __jit_debug_register_code(void) __attribute__((noinline));
6296 void __jit_debug_register_code(void)
6297 {
6298 asm("");
6299 }
6300
6301 /* Must statically initialize the version, because GDB may check
6302 the version before we can set it. */
6303 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 };
6304
6305 /* End GDB interface. */
6306
6307 static int find_string(const char *strtab, const char *str)
6308 {
6309 const char *p = strtab + 1;
6310
6311 while (1) {
6312 if (strcmp(p, str) == 0) {
6313 return p - strtab;
6314 }
6315 p += strlen(p) + 1;
6316 }
6317 }
6318
6319 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size,
6320 const void *debug_frame,
6321 size_t debug_frame_size)
6322 {
6323 struct __attribute__((packed)) DebugInfo {
6324 uint32_t len;
6325 uint16_t version;
6326 uint32_t abbrev;
6327 uint8_t ptr_size;
6328 uint8_t cu_die;
6329 uint16_t cu_lang;
6330 uintptr_t cu_low_pc;
6331 uintptr_t cu_high_pc;
6332 uint8_t fn_die;
6333 char fn_name[16];
6334 uintptr_t fn_low_pc;
6335 uintptr_t fn_high_pc;
6336 uint8_t cu_eoc;
6337 };
6338
6339 struct ElfImage {
6340 ElfW(Ehdr) ehdr;
6341 ElfW(Phdr) phdr;
6342 ElfW(Shdr) shdr[7];
6343 ElfW(Sym) sym[2];
6344 struct DebugInfo di;
6345 uint8_t da[24];
6346 char str[80];
6347 };
6348
6349 struct ElfImage *img;
6350
6351 static const struct ElfImage img_template = {
6352 .ehdr = {
6353 .e_ident[EI_MAG0] = ELFMAG0,
6354 .e_ident[EI_MAG1] = ELFMAG1,
6355 .e_ident[EI_MAG2] = ELFMAG2,
6356 .e_ident[EI_MAG3] = ELFMAG3,
6357 .e_ident[EI_CLASS] = ELF_CLASS,
6358 .e_ident[EI_DATA] = ELF_DATA,
6359 .e_ident[EI_VERSION] = EV_CURRENT,
6360 .e_type = ET_EXEC,
6361 .e_machine = ELF_HOST_MACHINE,
6362 .e_version = EV_CURRENT,
6363 .e_phoff = offsetof(struct ElfImage, phdr),
6364 .e_shoff = offsetof(struct ElfImage, shdr),
6365 .e_ehsize = sizeof(ElfW(Shdr)),
6366 .e_phentsize = sizeof(ElfW(Phdr)),
6367 .e_phnum = 1,
6368 .e_shentsize = sizeof(ElfW(Shdr)),
6369 .e_shnum = ARRAY_SIZE(img->shdr),
6370 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1,
6371 #ifdef ELF_HOST_FLAGS
6372 .e_flags = ELF_HOST_FLAGS,
6373 #endif
6374 #ifdef ELF_OSABI
6375 .e_ident[EI_OSABI] = ELF_OSABI,
6376 #endif
6377 },
6378 .phdr = {
6379 .p_type = PT_LOAD,
6380 .p_flags = PF_X,
6381 },
6382 .shdr = {
6383 [0] = { .sh_type = SHT_NULL },
6384 /* Trick: The contents of code_gen_buffer are not present in
6385 this fake ELF file; that got allocated elsewhere. Therefore
6386 we mark .text as SHT_NOBITS (similar to .bss) so that readers
6387 will not look for contents. We can record any address. */
6388 [1] = { /* .text */
6389 .sh_type = SHT_NOBITS,
6390 .sh_flags = SHF_EXECINSTR | SHF_ALLOC,
6391 },
6392 [2] = { /* .debug_info */
6393 .sh_type = SHT_PROGBITS,
6394 .sh_offset = offsetof(struct ElfImage, di),
6395 .sh_size = sizeof(struct DebugInfo),
6396 },
6397 [3] = { /* .debug_abbrev */
6398 .sh_type = SHT_PROGBITS,
6399 .sh_offset = offsetof(struct ElfImage, da),
6400 .sh_size = sizeof(img->da),
6401 },
6402 [4] = { /* .debug_frame */
6403 .sh_type = SHT_PROGBITS,
6404 .sh_offset = sizeof(struct ElfImage),
6405 },
6406 [5] = { /* .symtab */
6407 .sh_type = SHT_SYMTAB,
6408 .sh_offset = offsetof(struct ElfImage, sym),
6409 .sh_size = sizeof(img->sym),
6410 .sh_info = 1,
6411 .sh_link = ARRAY_SIZE(img->shdr) - 1,
6412 .sh_entsize = sizeof(ElfW(Sym)),
6413 },
6414 [6] = { /* .strtab */
6415 .sh_type = SHT_STRTAB,
6416 .sh_offset = offsetof(struct ElfImage, str),
6417 .sh_size = sizeof(img->str),
6418 }
6419 },
6420 .sym = {
6421 [1] = { /* code_gen_buffer */
6422 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC),
6423 .st_shndx = 1,
6424 }
6425 },
6426 .di = {
6427 .len = sizeof(struct DebugInfo) - 4,
6428 .version = 2,
6429 .ptr_size = sizeof(void *),
6430 .cu_die = 1,
6431 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */
6432 .fn_die = 2,
6433 .fn_name = "code_gen_buffer"
6434 },
6435 .da = {
6436 1, /* abbrev number (the cu) */
6437 0x11, 1, /* DW_TAG_compile_unit, has children */
6438 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */
6439 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */
6440 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */
6441 0, 0, /* end of abbrev */
6442 2, /* abbrev number (the fn) */
6443 0x2e, 0, /* DW_TAG_subprogram, no children */
6444 0x3, 0x8, /* DW_AT_name, DW_FORM_string */
6445 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */
6446 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */
6447 0, 0, /* end of abbrev */
6448 0 /* no more abbrev */
6449 },
6450 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0"
6451 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer",
6452 };
6453
6454 /* We only need a single jit entry; statically allocate it. */
6455 static struct jit_code_entry one_entry;
6456
6457 uintptr_t buf = (uintptr_t)buf_ptr;
6458 size_t img_size = sizeof(struct ElfImage) + debug_frame_size;
6459 DebugFrameHeader *dfh;
6460
6461 img = g_malloc(img_size);
6462 *img = img_template;
6463
6464 img->phdr.p_vaddr = buf;
6465 img->phdr.p_paddr = buf;
6466 img->phdr.p_memsz = buf_size;
6467
6468 img->shdr[1].sh_name = find_string(img->str, ".text");
6469 img->shdr[1].sh_addr = buf;
6470 img->shdr[1].sh_size = buf_size;
6471
6472 img->shdr[2].sh_name = find_string(img->str, ".debug_info");
6473 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev");
6474
6475 img->shdr[4].sh_name = find_string(img->str, ".debug_frame");
6476 img->shdr[4].sh_size = debug_frame_size;
6477
6478 img->shdr[5].sh_name = find_string(img->str, ".symtab");
6479 img->shdr[6].sh_name = find_string(img->str, ".strtab");
6480
6481 img->sym[1].st_name = find_string(img->str, "code_gen_buffer");
6482 img->sym[1].st_value = buf;
6483 img->sym[1].st_size = buf_size;
6484
6485 img->di.cu_low_pc = buf;
6486 img->di.cu_high_pc = buf + buf_size;
6487 img->di.fn_low_pc = buf;
6488 img->di.fn_high_pc = buf + buf_size;
6489
6490 dfh = (DebugFrameHeader *)(img + 1);
6491 memcpy(dfh, debug_frame, debug_frame_size);
6492 dfh->fde.func_start = buf;
6493 dfh->fde.func_len = buf_size;
6494
6495 #ifdef DEBUG_JIT
6496 /* Enable this block to be able to debug the ELF image file creation.
6497 One can use readelf, objdump, or other inspection utilities. */
6498 {
6499 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir());
6500 FILE *f = fopen(jit, "w+b");
6501 if (f) {
6502 if (fwrite(img, img_size, 1, f) != img_size) {
6503 /* Avoid stupid unused return value warning for fwrite. */
6504 }
6505 fclose(f);
6506 }
6507 }
6508 #endif
6509
6510 one_entry.symfile_addr = img;
6511 one_entry.symfile_size = img_size;
6512
6513 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN;
6514 __jit_debug_descriptor.relevant_entry = &one_entry;
6515 __jit_debug_descriptor.first_entry = &one_entry;
6516 __jit_debug_register_code();
6517 }
6518 #else
6519 /* No support for the feature. Provide the entry point expected by exec.c,
6520 and implement the internal function we declared earlier. */
6521
6522 static void tcg_register_jit_int(const void *buf, size_t size,
6523 const void *debug_frame,
6524 size_t debug_frame_size)
6525 {
6526 }
6527
6528 void tcg_register_jit(const void *buf, size_t buf_size)
6529 {
6530 }
6531 #endif /* ELF_HOST_MACHINE */
6532
6533 #if !TCG_TARGET_MAYBE_vec
6534 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...)
6535 {
6536 g_assert_not_reached();
6537 }
6538 #endif
armbru's QEMU hackery