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 #include "tcg-has.h"
60 #ifdef CONFIG_USER_ONLY
61 #include "user/guest-base.h"
62 #endif
63
64 /* Forward declarations for functions declared in tcg-target.c.inc and
65 used here. */
66 static void tcg_target_init(TCGContext *s);
67 static void tcg_target_qemu_prologue(TCGContext *s);
68 static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
69 intptr_t value, intptr_t addend);
70 static void tcg_out_nop_fill(tcg_insn_unit *p, int count);
71
72 typedef struct TCGLabelQemuLdst TCGLabelQemuLdst;
73 static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l);
74 static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l);
75
76 /* The CIE and FDE header definitions will be common to all hosts. */
77 typedef struct {
78 uint32_t len __attribute__((aligned((sizeof(void *)))));
79 uint32_t id;
80 uint8_t version;
81 char augmentation[1];
82 uint8_t code_align;
83 uint8_t data_align;
84 uint8_t return_column;
85 } DebugFrameCIE;
86
87 typedef struct QEMU_PACKED {
88 uint32_t len __attribute__((aligned((sizeof(void *)))));
89 uint32_t cie_offset;
90 uintptr_t func_start;
91 uintptr_t func_len;
92 } DebugFrameFDEHeader;
93
94 typedef struct QEMU_PACKED {
95 DebugFrameCIE cie;
96 DebugFrameFDEHeader fde;
97 } DebugFrameHeader;
98
99 struct TCGLabelQemuLdst {
100 bool is_ld; /* qemu_ld: true, qemu_st: false */
101 MemOpIdx oi;
102 TCGType type; /* result type of a load */
103 TCGReg addr_reg; /* reg index for guest virtual addr */
104 TCGReg datalo_reg; /* reg index for low word to be loaded or stored */
105 TCGReg datahi_reg; /* reg index for high word to be loaded or stored */
106 const tcg_insn_unit *raddr; /* addr of the next IR of qemu_ld/st IR */
107 tcg_insn_unit *label_ptr[2]; /* label pointers to be updated */
108 QSIMPLEQ_ENTRY(TCGLabelQemuLdst) next;
109 };
110
111 static void tcg_register_jit_int(const void *buf, size_t size,
112 const void *debug_frame,
113 size_t debug_frame_size)
114 __attribute__((unused));
115
116 /* Forward declarations for functions declared and used in tcg-target.c.inc. */
117 static void tcg_out_tb_start(TCGContext *s);
118 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1,
119 intptr_t arg2);
120 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg);
121 static void tcg_out_movi(TCGContext *s, TCGType type,
122 TCGReg ret, tcg_target_long arg);
123 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg);
124 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg);
125 static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg);
126 static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg);
127 static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg);
128 static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg);
129 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg);
130 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg);
131 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg);
132 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long);
133 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2);
134 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg);
135 static void tcg_out_goto_tb(TCGContext *s, int which);
136 static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
137 const TCGArg args[TCG_MAX_OP_ARGS],
138 const int const_args[TCG_MAX_OP_ARGS]);
139 #if TCG_TARGET_MAYBE_vec
140 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
141 TCGReg dst, TCGReg src);
142 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
143 TCGReg dst, TCGReg base, intptr_t offset);
144 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
145 TCGReg dst, int64_t arg);
146 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
147 unsigned vecl, unsigned vece,
148 const TCGArg args[TCG_MAX_OP_ARGS],
149 const int const_args[TCG_MAX_OP_ARGS]);
150 #else
tcg_out_dup_vec(TCGContext * s,TCGType type,unsigned vece,TCGReg dst,TCGReg src)151 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
152 TCGReg dst, TCGReg src)
153 {
154 g_assert_not_reached();
155 }
tcg_out_dupm_vec(TCGContext * s,TCGType type,unsigned vece,TCGReg dst,TCGReg base,intptr_t offset)156 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
157 TCGReg dst, TCGReg base, intptr_t offset)
158 {
159 g_assert_not_reached();
160 }
tcg_out_dupi_vec(TCGContext * s,TCGType type,unsigned vece,TCGReg dst,int64_t arg)161 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
162 TCGReg dst, int64_t arg)
163 {
164 g_assert_not_reached();
165 }
tcg_out_vec_op(TCGContext * s,TCGOpcode opc,unsigned vecl,unsigned vece,const TCGArg args[TCG_MAX_OP_ARGS],const int const_args[TCG_MAX_OP_ARGS])166 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
167 unsigned vecl, unsigned vece,
168 const TCGArg args[TCG_MAX_OP_ARGS],
169 const int const_args[TCG_MAX_OP_ARGS])
170 {
171 g_assert_not_reached();
172 }
tcg_can_emit_vec_op(TCGOpcode o,TCGType t,unsigned ve)173 int tcg_can_emit_vec_op(TCGOpcode o, TCGType t, unsigned ve)
174 {
175 return 0;
176 }
177 #endif
178 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1,
179 intptr_t arg2);
180 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
181 TCGReg base, intptr_t ofs);
182 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target,
183 const TCGHelperInfo *info);
184 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot);
185 static bool tcg_target_const_match(int64_t val, int ct,
186 TCGType type, TCGCond cond, int vece);
187
188 #ifndef CONFIG_USER_ONLY
189 #define guest_base ({ qemu_build_not_reached(); (uintptr_t)0; })
190 #endif
191
192 typedef struct TCGLdstHelperParam {
193 TCGReg (*ra_gen)(TCGContext *s, const TCGLabelQemuLdst *l, int arg_reg);
194 unsigned ntmp;
195 int tmp[3];
196 } TCGLdstHelperParam;
197
198 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *l,
199 const TCGLdstHelperParam *p)
200 __attribute__((unused));
201 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *l,
202 bool load_sign, const TCGLdstHelperParam *p)
203 __attribute__((unused));
204 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *l,
205 const TCGLdstHelperParam *p)
206 __attribute__((unused));
207
208 static void * const qemu_ld_helpers[MO_SSIZE + 1] __attribute__((unused)) = {
209 [MO_UB] = helper_ldub_mmu,
210 [MO_SB] = helper_ldsb_mmu,
211 [MO_UW] = helper_lduw_mmu,
212 [MO_SW] = helper_ldsw_mmu,
213 [MO_UL] = helper_ldul_mmu,
214 [MO_UQ] = helper_ldq_mmu,
215 #if TCG_TARGET_REG_BITS == 64
216 [MO_SL] = helper_ldsl_mmu,
217 [MO_128] = helper_ld16_mmu,
218 #endif
219 };
220
221 static void * const qemu_st_helpers[MO_SIZE + 1] __attribute__((unused)) = {
222 [MO_8] = helper_stb_mmu,
223 [MO_16] = helper_stw_mmu,
224 [MO_32] = helper_stl_mmu,
225 [MO_64] = helper_stq_mmu,
226 #if TCG_TARGET_REG_BITS == 64
227 [MO_128] = helper_st16_mmu,
228 #endif
229 };
230
231 typedef struct {
232 MemOp atom; /* lg2 bits of atomicity required */
233 MemOp align; /* lg2 bits of alignment to use */
234 } TCGAtomAlign;
235
236 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc,
237 MemOp host_atom, bool allow_two_ops)
238 __attribute__((unused));
239
240 #ifdef CONFIG_USER_ONLY
241 bool tcg_use_softmmu;
242 #endif
243
244 TCGContext tcg_init_ctx;
245 __thread TCGContext *tcg_ctx;
246
247 TCGContext **tcg_ctxs;
248 unsigned int tcg_cur_ctxs;
249 unsigned int tcg_max_ctxs;
250 TCGv_env tcg_env;
251 const void *tcg_code_gen_epilogue;
252 uintptr_t tcg_splitwx_diff;
253
254 #ifndef CONFIG_TCG_INTERPRETER
255 tcg_prologue_fn *tcg_qemu_tb_exec;
256 #endif
257
258 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT];
259 static TCGRegSet tcg_target_call_clobber_regs;
260
261 #if TCG_TARGET_INSN_UNIT_SIZE == 1
tcg_out8(TCGContext * s,uint8_t v)262 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v)
263 {
264 *s->code_ptr++ = v;
265 }
266
tcg_patch8(tcg_insn_unit * p,uint8_t v)267 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p,
268 uint8_t v)
269 {
270 *p = v;
271 }
272 #endif
273
274 #if TCG_TARGET_INSN_UNIT_SIZE <= 2
tcg_out16(TCGContext * s,uint16_t v)275 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v)
276 {
277 if (TCG_TARGET_INSN_UNIT_SIZE == 2) {
278 *s->code_ptr++ = v;
279 } else {
280 tcg_insn_unit *p = s->code_ptr;
281 memcpy(p, &v, sizeof(v));
282 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE);
283 }
284 }
285
tcg_patch16(tcg_insn_unit * p,uint16_t v)286 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p,
287 uint16_t v)
288 {
289 if (TCG_TARGET_INSN_UNIT_SIZE == 2) {
290 *p = v;
291 } else {
292 memcpy(p, &v, sizeof(v));
293 }
294 }
295 #endif
296
297 #if TCG_TARGET_INSN_UNIT_SIZE <= 4
tcg_out32(TCGContext * s,uint32_t v)298 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v)
299 {
300 if (TCG_TARGET_INSN_UNIT_SIZE == 4) {
301 *s->code_ptr++ = v;
302 } else {
303 tcg_insn_unit *p = s->code_ptr;
304 memcpy(p, &v, sizeof(v));
305 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE);
306 }
307 }
308
tcg_patch32(tcg_insn_unit * p,uint32_t v)309 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p,
310 uint32_t v)
311 {
312 if (TCG_TARGET_INSN_UNIT_SIZE == 4) {
313 *p = v;
314 } else {
315 memcpy(p, &v, sizeof(v));
316 }
317 }
318 #endif
319
320 #if TCG_TARGET_INSN_UNIT_SIZE <= 8
tcg_out64(TCGContext * s,uint64_t v)321 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v)
322 {
323 if (TCG_TARGET_INSN_UNIT_SIZE == 8) {
324 *s->code_ptr++ = v;
325 } else {
326 tcg_insn_unit *p = s->code_ptr;
327 memcpy(p, &v, sizeof(v));
328 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE);
329 }
330 }
331
tcg_patch64(tcg_insn_unit * p,uint64_t v)332 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p,
333 uint64_t v)
334 {
335 if (TCG_TARGET_INSN_UNIT_SIZE == 8) {
336 *p = v;
337 } else {
338 memcpy(p, &v, sizeof(v));
339 }
340 }
341 #endif
342
343 /* label relocation processing */
344
tcg_out_reloc(TCGContext * s,tcg_insn_unit * code_ptr,int type,TCGLabel * l,intptr_t addend)345 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type,
346 TCGLabel *l, intptr_t addend)
347 {
348 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation));
349
350 r->type = type;
351 r->ptr = code_ptr;
352 r->addend = addend;
353 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next);
354 }
355
tcg_out_label(TCGContext * s,TCGLabel * l)356 static void tcg_out_label(TCGContext *s, TCGLabel *l)
357 {
358 tcg_debug_assert(!l->has_value);
359 l->has_value = 1;
360 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr);
361 }
362
gen_new_label(void)363 TCGLabel *gen_new_label(void)
364 {
365 TCGContext *s = tcg_ctx;
366 TCGLabel *l = tcg_malloc(sizeof(TCGLabel));
367
368 memset(l, 0, sizeof(TCGLabel));
369 l->id = s->nb_labels++;
370 QSIMPLEQ_INIT(&l->branches);
371 QSIMPLEQ_INIT(&l->relocs);
372
373 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next);
374
375 return l;
376 }
377
tcg_resolve_relocs(TCGContext * s)378 static bool tcg_resolve_relocs(TCGContext *s)
379 {
380 TCGLabel *l;
381
382 QSIMPLEQ_FOREACH(l, &s->labels, next) {
383 TCGRelocation *r;
384 uintptr_t value = l->u.value;
385
386 QSIMPLEQ_FOREACH(r, &l->relocs, next) {
387 if (!patch_reloc(r->ptr, r->type, value, r->addend)) {
388 return false;
389 }
390 }
391 }
392 return true;
393 }
394
set_jmp_reset_offset(TCGContext * s,int which)395 static void set_jmp_reset_offset(TCGContext *s, int which)
396 {
397 /*
398 * We will check for overflow at the end of the opcode loop in
399 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX.
400 */
401 s->gen_tb->jmp_reset_offset[which] = tcg_current_code_size(s);
402 }
403
set_jmp_insn_offset(TCGContext * s,int which)404 static void G_GNUC_UNUSED set_jmp_insn_offset(TCGContext *s, int which)
405 {
406 /*
407 * We will check for overflow at the end of the opcode loop in
408 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX.
409 */
410 s->gen_tb->jmp_insn_offset[which] = tcg_current_code_size(s);
411 }
412
get_jmp_target_addr(TCGContext * s,int which)413 static uintptr_t G_GNUC_UNUSED get_jmp_target_addr(TCGContext *s, int which)
414 {
415 /*
416 * Return the read-execute version of the pointer, for the benefit
417 * of any pc-relative addressing mode.
418 */
419 return (uintptr_t)tcg_splitwx_to_rx(&s->gen_tb->jmp_target_addr[which]);
420 }
421
422 static int __attribute__((unused))
tlb_mask_table_ofs(TCGContext * s,int which)423 tlb_mask_table_ofs(TCGContext *s, int which)
424 {
425 return (offsetof(CPUNegativeOffsetState, tlb.f[which]) -
426 sizeof(CPUNegativeOffsetState));
427 }
428
429 /* Signal overflow, starting over with fewer guest insns. */
430 static G_NORETURN
tcg_raise_tb_overflow(TCGContext * s)431 void tcg_raise_tb_overflow(TCGContext *s)
432 {
433 siglongjmp(s->jmp_trans, -2);
434 }
435
436 /*
437 * Used by tcg_out_movext{1,2} to hold the arguments for tcg_out_movext.
438 * By the time we arrive at tcg_out_movext1, @dst is always a TCGReg.
439 *
440 * However, tcg_out_helper_load_slots reuses this field to hold an
441 * argument slot number (which may designate a argument register or an
442 * argument stack slot), converting to TCGReg once all arguments that
443 * are destined for the stack are processed.
444 */
445 typedef struct TCGMovExtend {
446 unsigned dst;
447 TCGReg src;
448 TCGType dst_type;
449 TCGType src_type;
450 MemOp src_ext;
451 } TCGMovExtend;
452
453 /**
454 * tcg_out_movext -- move and extend
455 * @s: tcg context
456 * @dst_type: integral type for destination
457 * @dst: destination register
458 * @src_type: integral type for source
459 * @src_ext: extension to apply to source
460 * @src: source register
461 *
462 * Move or extend @src into @dst, depending on @src_ext and the types.
463 */
tcg_out_movext(TCGContext * s,TCGType dst_type,TCGReg dst,TCGType src_type,MemOp src_ext,TCGReg src)464 static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst,
465 TCGType src_type, MemOp src_ext, TCGReg src)
466 {
467 switch (src_ext) {
468 case MO_UB:
469 tcg_out_ext8u(s, dst, src);
470 break;
471 case MO_SB:
472 tcg_out_ext8s(s, dst_type, dst, src);
473 break;
474 case MO_UW:
475 tcg_out_ext16u(s, dst, src);
476 break;
477 case MO_SW:
478 tcg_out_ext16s(s, dst_type, dst, src);
479 break;
480 case MO_UL:
481 case MO_SL:
482 if (dst_type == TCG_TYPE_I32) {
483 if (src_type == TCG_TYPE_I32) {
484 tcg_out_mov(s, TCG_TYPE_I32, dst, src);
485 } else {
486 tcg_out_extrl_i64_i32(s, dst, src);
487 }
488 } else if (src_type == TCG_TYPE_I32) {
489 if (src_ext & MO_SIGN) {
490 tcg_out_exts_i32_i64(s, dst, src);
491 } else {
492 tcg_out_extu_i32_i64(s, dst, src);
493 }
494 } else {
495 if (src_ext & MO_SIGN) {
496 tcg_out_ext32s(s, dst, src);
497 } else {
498 tcg_out_ext32u(s, dst, src);
499 }
500 }
501 break;
502 case MO_UQ:
503 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
504 if (dst_type == TCG_TYPE_I32) {
505 tcg_out_extrl_i64_i32(s, dst, src);
506 } else {
507 tcg_out_mov(s, TCG_TYPE_I64, dst, src);
508 }
509 break;
510 default:
511 g_assert_not_reached();
512 }
513 }
514
515 /* Minor variations on a theme, using a structure. */
tcg_out_movext1_new_src(TCGContext * s,const TCGMovExtend * i,TCGReg src)516 static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i,
517 TCGReg src)
518 {
519 tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src);
520 }
521
tcg_out_movext1(TCGContext * s,const TCGMovExtend * i)522 static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i)
523 {
524 tcg_out_movext1_new_src(s, i, i->src);
525 }
526
527 /**
528 * tcg_out_movext2 -- move and extend two pair
529 * @s: tcg context
530 * @i1: first move description
531 * @i2: second move description
532 * @scratch: temporary register, or -1 for none
533 *
534 * As tcg_out_movext, for both @i1 and @i2, caring for overlap
535 * between the sources and destinations.
536 */
537
tcg_out_movext2(TCGContext * s,const TCGMovExtend * i1,const TCGMovExtend * i2,int scratch)538 static void tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1,
539 const TCGMovExtend *i2, int scratch)
540 {
541 TCGReg src1 = i1->src;
542 TCGReg src2 = i2->src;
543
544 if (i1->dst != src2) {
545 tcg_out_movext1(s, i1);
546 tcg_out_movext1(s, i2);
547 return;
548 }
549 if (i2->dst == src1) {
550 TCGType src1_type = i1->src_type;
551 TCGType src2_type = i2->src_type;
552
553 if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) {
554 /* The data is now in the correct registers, now extend. */
555 src1 = i2->src;
556 src2 = i1->src;
557 } else {
558 tcg_debug_assert(scratch >= 0);
559 tcg_out_mov(s, src1_type, scratch, src1);
560 src1 = scratch;
561 }
562 }
563 tcg_out_movext1_new_src(s, i2, src2);
564 tcg_out_movext1_new_src(s, i1, src1);
565 }
566
567 /**
568 * tcg_out_movext3 -- move and extend three pair
569 * @s: tcg context
570 * @i1: first move description
571 * @i2: second move description
572 * @i3: third move description
573 * @scratch: temporary register, or -1 for none
574 *
575 * As tcg_out_movext, for all of @i1, @i2 and @i3, caring for overlap
576 * between the sources and destinations.
577 */
578
tcg_out_movext3(TCGContext * s,const TCGMovExtend * i1,const TCGMovExtend * i2,const TCGMovExtend * i3,int scratch)579 static void tcg_out_movext3(TCGContext *s, const TCGMovExtend *i1,
580 const TCGMovExtend *i2, const TCGMovExtend *i3,
581 int scratch)
582 {
583 TCGReg src1 = i1->src;
584 TCGReg src2 = i2->src;
585 TCGReg src3 = i3->src;
586
587 if (i1->dst != src2 && i1->dst != src3) {
588 tcg_out_movext1(s, i1);
589 tcg_out_movext2(s, i2, i3, scratch);
590 return;
591 }
592 if (i2->dst != src1 && i2->dst != src3) {
593 tcg_out_movext1(s, i2);
594 tcg_out_movext2(s, i1, i3, scratch);
595 return;
596 }
597 if (i3->dst != src1 && i3->dst != src2) {
598 tcg_out_movext1(s, i3);
599 tcg_out_movext2(s, i1, i2, scratch);
600 return;
601 }
602
603 /*
604 * There is a cycle. Since there are only 3 nodes, the cycle is
605 * either "clockwise" or "anti-clockwise", and can be solved with
606 * a single scratch or two xchg.
607 */
608 if (i1->dst == src2 && i2->dst == src3 && i3->dst == src1) {
609 /* "Clockwise" */
610 if (tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2)) {
611 tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3);
612 /* The data is now in the correct registers, now extend. */
613 tcg_out_movext1_new_src(s, i1, i1->dst);
614 tcg_out_movext1_new_src(s, i2, i2->dst);
615 tcg_out_movext1_new_src(s, i3, i3->dst);
616 } else {
617 tcg_debug_assert(scratch >= 0);
618 tcg_out_mov(s, i1->src_type, scratch, src1);
619 tcg_out_movext1(s, i3);
620 tcg_out_movext1(s, i2);
621 tcg_out_movext1_new_src(s, i1, scratch);
622 }
623 } else if (i1->dst == src3 && i2->dst == src1 && i3->dst == src2) {
624 /* "Anti-clockwise" */
625 if (tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3)) {
626 tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2);
627 /* The data is now in the correct registers, now extend. */
628 tcg_out_movext1_new_src(s, i1, i1->dst);
629 tcg_out_movext1_new_src(s, i2, i2->dst);
630 tcg_out_movext1_new_src(s, i3, i3->dst);
631 } else {
632 tcg_debug_assert(scratch >= 0);
633 tcg_out_mov(s, i1->src_type, scratch, src1);
634 tcg_out_movext1(s, i2);
635 tcg_out_movext1(s, i3);
636 tcg_out_movext1_new_src(s, i1, scratch);
637 }
638 } else {
639 g_assert_not_reached();
640 }
641 }
642
643 /*
644 * Allocate a new TCGLabelQemuLdst entry.
645 */
646
647 __attribute__((unused))
new_ldst_label(TCGContext * s)648 static TCGLabelQemuLdst *new_ldst_label(TCGContext *s)
649 {
650 TCGLabelQemuLdst *l = tcg_malloc(sizeof(*l));
651
652 memset(l, 0, sizeof(*l));
653 QSIMPLEQ_INSERT_TAIL(&s->ldst_labels, l, next);
654
655 return l;
656 }
657
658 /*
659 * Allocate new constant pool entries.
660 */
661
662 typedef struct TCGLabelPoolData {
663 struct TCGLabelPoolData *next;
664 tcg_insn_unit *label;
665 intptr_t addend;
666 int rtype;
667 unsigned nlong;
668 tcg_target_ulong data[];
669 } TCGLabelPoolData;
670
new_pool_alloc(TCGContext * s,int nlong,int rtype,tcg_insn_unit * label,intptr_t addend)671 static TCGLabelPoolData *new_pool_alloc(TCGContext *s, int nlong, int rtype,
672 tcg_insn_unit *label, intptr_t addend)
673 {
674 TCGLabelPoolData *n = tcg_malloc(sizeof(TCGLabelPoolData)
675 + sizeof(tcg_target_ulong) * nlong);
676
677 n->label = label;
678 n->addend = addend;
679 n->rtype = rtype;
680 n->nlong = nlong;
681 return n;
682 }
683
new_pool_insert(TCGContext * s,TCGLabelPoolData * n)684 static void new_pool_insert(TCGContext *s, TCGLabelPoolData *n)
685 {
686 TCGLabelPoolData *i, **pp;
687 int nlong = n->nlong;
688
689 /* Insertion sort on the pool. */
690 for (pp = &s->pool_labels; (i = *pp) != NULL; pp = &i->next) {
691 if (nlong > i->nlong) {
692 break;
693 }
694 if (nlong < i->nlong) {
695 continue;
696 }
697 if (memcmp(n->data, i->data, sizeof(tcg_target_ulong) * nlong) >= 0) {
698 break;
699 }
700 }
701 n->next = *pp;
702 *pp = n;
703 }
704
705 /* The "usual" for generic integer code. */
706 __attribute__((unused))
new_pool_label(TCGContext * s,tcg_target_ulong d,int rtype,tcg_insn_unit * label,intptr_t addend)707 static void new_pool_label(TCGContext *s, tcg_target_ulong d, int rtype,
708 tcg_insn_unit *label, intptr_t addend)
709 {
710 TCGLabelPoolData *n = new_pool_alloc(s, 1, rtype, label, addend);
711 n->data[0] = d;
712 new_pool_insert(s, n);
713 }
714
715 /* For v64 or v128, depending on the host. */
716 __attribute__((unused))
new_pool_l2(TCGContext * s,int rtype,tcg_insn_unit * label,intptr_t addend,tcg_target_ulong d0,tcg_target_ulong d1)717 static void new_pool_l2(TCGContext *s, int rtype, tcg_insn_unit *label,
718 intptr_t addend, tcg_target_ulong d0,
719 tcg_target_ulong d1)
720 {
721 TCGLabelPoolData *n = new_pool_alloc(s, 2, rtype, label, addend);
722 n->data[0] = d0;
723 n->data[1] = d1;
724 new_pool_insert(s, n);
725 }
726
727 /* For v128 or v256, depending on the host. */
728 __attribute__((unused))
new_pool_l4(TCGContext * s,int rtype,tcg_insn_unit * label,intptr_t addend,tcg_target_ulong d0,tcg_target_ulong d1,tcg_target_ulong d2,tcg_target_ulong d3)729 static void new_pool_l4(TCGContext *s, int rtype, tcg_insn_unit *label,
730 intptr_t addend, tcg_target_ulong d0,
731 tcg_target_ulong d1, tcg_target_ulong d2,
732 tcg_target_ulong d3)
733 {
734 TCGLabelPoolData *n = new_pool_alloc(s, 4, rtype, label, addend);
735 n->data[0] = d0;
736 n->data[1] = d1;
737 n->data[2] = d2;
738 n->data[3] = d3;
739 new_pool_insert(s, n);
740 }
741
742 /* For v256, for 32-bit host. */
743 __attribute__((unused))
new_pool_l8(TCGContext * s,int rtype,tcg_insn_unit * label,intptr_t addend,tcg_target_ulong d0,tcg_target_ulong d1,tcg_target_ulong d2,tcg_target_ulong d3,tcg_target_ulong d4,tcg_target_ulong d5,tcg_target_ulong d6,tcg_target_ulong d7)744 static void new_pool_l8(TCGContext *s, int rtype, tcg_insn_unit *label,
745 intptr_t addend, tcg_target_ulong d0,
746 tcg_target_ulong d1, tcg_target_ulong d2,
747 tcg_target_ulong d3, tcg_target_ulong d4,
748 tcg_target_ulong d5, tcg_target_ulong d6,
749 tcg_target_ulong d7)
750 {
751 TCGLabelPoolData *n = new_pool_alloc(s, 8, rtype, label, addend);
752 n->data[0] = d0;
753 n->data[1] = d1;
754 n->data[2] = d2;
755 n->data[3] = d3;
756 n->data[4] = d4;
757 n->data[5] = d5;
758 n->data[6] = d6;
759 n->data[7] = d7;
760 new_pool_insert(s, n);
761 }
762
763 /*
764 * Generate TB finalization at the end of block
765 */
766
tcg_out_ldst_finalize(TCGContext * s)767 static int tcg_out_ldst_finalize(TCGContext *s)
768 {
769 TCGLabelQemuLdst *lb;
770
771 /* qemu_ld/st slow paths */
772 QSIMPLEQ_FOREACH(lb, &s->ldst_labels, next) {
773 if (lb->is_ld
774 ? !tcg_out_qemu_ld_slow_path(s, lb)
775 : !tcg_out_qemu_st_slow_path(s, lb)) {
776 return -2;
777 }
778
779 /*
780 * Test for (pending) buffer overflow. The assumption is that any
781 * one operation beginning below the high water mark cannot overrun
782 * the buffer completely. Thus we can test for overflow after
783 * generating code without having to check during generation.
784 */
785 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) {
786 return -1;
787 }
788 }
789 return 0;
790 }
791
tcg_out_pool_finalize(TCGContext * s)792 static int tcg_out_pool_finalize(TCGContext *s)
793 {
794 TCGLabelPoolData *p = s->pool_labels;
795 TCGLabelPoolData *l = NULL;
796 void *a;
797
798 if (p == NULL) {
799 return 0;
800 }
801
802 /*
803 * ??? Round up to qemu_icache_linesize, but then do not round
804 * again when allocating the next TranslationBlock structure.
805 */
806 a = (void *)ROUND_UP((uintptr_t)s->code_ptr,
807 sizeof(tcg_target_ulong) * p->nlong);
808 tcg_out_nop_fill(s->code_ptr, (tcg_insn_unit *)a - s->code_ptr);
809 s->data_gen_ptr = a;
810
811 for (; p != NULL; p = p->next) {
812 size_t size = sizeof(tcg_target_ulong) * p->nlong;
813 uintptr_t value;
814
815 if (!l || l->nlong != p->nlong || memcmp(l->data, p->data, size)) {
816 if (unlikely(a > s->code_gen_highwater)) {
817 return -1;
818 }
819 memcpy(a, p->data, size);
820 a += size;
821 l = p;
822 }
823
824 value = (uintptr_t)tcg_splitwx_to_rx(a) - size;
825 if (!patch_reloc(p->label, p->rtype, value, p->addend)) {
826 return -2;
827 }
828 }
829
830 s->code_ptr = a;
831 return 0;
832 }
833
834 #define C_PFX1(P, A) P##A
835 #define C_PFX2(P, A, B) P##A##_##B
836 #define C_PFX3(P, A, B, C) P##A##_##B##_##C
837 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D
838 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E
839 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F
840
841 /* Define an enumeration for the various combinations. */
842
843 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1),
844 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2),
845 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3),
846 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4),
847
848 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1),
849 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2),
850 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3),
851 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4),
852
853 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2),
854 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1),
855 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1),
856
857 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1),
858 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2),
859 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3),
860 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4),
861 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4),
862
863 typedef enum {
864 C_NotImplemented = -1,
865 #include "tcg-target-con-set.h"
866 } TCGConstraintSetIndex;
867
868 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode, TCGType, unsigned);
869
870 #undef C_O0_I1
871 #undef C_O0_I2
872 #undef C_O0_I3
873 #undef C_O0_I4
874 #undef C_O1_I1
875 #undef C_O1_I2
876 #undef C_O1_I3
877 #undef C_O1_I4
878 #undef C_N1_I2
879 #undef C_N1O1_I1
880 #undef C_N2_I1
881 #undef C_O2_I1
882 #undef C_O2_I2
883 #undef C_O2_I3
884 #undef C_O2_I4
885 #undef C_N1_O1_I4
886
887 /* Put all of the constraint sets into an array, indexed by the enum. */
888
889 typedef struct TCGConstraintSet {
890 uint8_t nb_oargs, nb_iargs;
891 const char *args_ct_str[TCG_MAX_OP_ARGS];
892 } TCGConstraintSet;
893
894 #define C_O0_I1(I1) { 0, 1, { #I1 } },
895 #define C_O0_I2(I1, I2) { 0, 2, { #I1, #I2 } },
896 #define C_O0_I3(I1, I2, I3) { 0, 3, { #I1, #I2, #I3 } },
897 #define C_O0_I4(I1, I2, I3, I4) { 0, 4, { #I1, #I2, #I3, #I4 } },
898
899 #define C_O1_I1(O1, I1) { 1, 1, { #O1, #I1 } },
900 #define C_O1_I2(O1, I1, I2) { 1, 2, { #O1, #I1, #I2 } },
901 #define C_O1_I3(O1, I1, I2, I3) { 1, 3, { #O1, #I1, #I2, #I3 } },
902 #define C_O1_I4(O1, I1, I2, I3, I4) { 1, 4, { #O1, #I1, #I2, #I3, #I4 } },
903
904 #define C_N1_I2(O1, I1, I2) { 1, 2, { "&" #O1, #I1, #I2 } },
905 #define C_N1O1_I1(O1, O2, I1) { 2, 1, { "&" #O1, #O2, #I1 } },
906 #define C_N2_I1(O1, O2, I1) { 2, 1, { "&" #O1, "&" #O2, #I1 } },
907
908 #define C_O2_I1(O1, O2, I1) { 2, 1, { #O1, #O2, #I1 } },
909 #define C_O2_I2(O1, O2, I1, I2) { 2, 2, { #O1, #O2, #I1, #I2 } },
910 #define C_O2_I3(O1, O2, I1, I2, I3) { 2, 3, { #O1, #O2, #I1, #I2, #I3 } },
911 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { 2, 4, { #O1, #O2, #I1, #I2, #I3, #I4 } },
912 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) { 2, 4, { "&" #O1, #O2, #I1, #I2, #I3, #I4 } },
913
914 static const TCGConstraintSet constraint_sets[] = {
915 #include "tcg-target-con-set.h"
916 };
917
918 #undef C_O0_I1
919 #undef C_O0_I2
920 #undef C_O0_I3
921 #undef C_O0_I4
922 #undef C_O1_I1
923 #undef C_O1_I2
924 #undef C_O1_I3
925 #undef C_O1_I4
926 #undef C_N1_I2
927 #undef C_N1O1_I1
928 #undef C_N2_I1
929 #undef C_O2_I1
930 #undef C_O2_I2
931 #undef C_O2_I3
932 #undef C_O2_I4
933 #undef C_N1_O1_I4
934
935 /* Expand the enumerator to be returned from tcg_target_op_def(). */
936
937 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1)
938 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2)
939 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3)
940 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4)
941
942 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1)
943 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2)
944 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3)
945 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4)
946
947 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2)
948 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1)
949 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1)
950
951 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1)
952 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2)
953 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3)
954 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4)
955 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4)
956
957 #include "tcg-target.c.inc"
958
959 #ifndef CONFIG_TCG_INTERPRETER
960 /* Validate CPUTLBDescFast placement. */
961 QEMU_BUILD_BUG_ON((int)(offsetof(CPUNegativeOffsetState, tlb.f[0]) -
962 sizeof(CPUNegativeOffsetState))
963 < MIN_TLB_MASK_TABLE_OFS);
964 #endif
965
966 /*
967 * All TCG threads except the parent (i.e. the one that called tcg_context_init
968 * and registered the target's TCG globals) must register with this function
969 * before initiating translation.
970 *
971 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation
972 * of tcg_region_init() for the reasoning behind this.
973 *
974 * In system-mode each caller registers its context in tcg_ctxs[]. Note that in
975 * system-mode tcg_ctxs[] does not track tcg_ctx_init, since the initial context
976 * is not used anymore for translation once this function is called.
977 *
978 * Not tracking tcg_init_ctx in tcg_ctxs[] in system-mode keeps code that
979 * iterates over the array (e.g. tcg_code_size() the same for both system/user
980 * modes.
981 */
982 #ifdef CONFIG_USER_ONLY
tcg_register_thread(void)983 void tcg_register_thread(void)
984 {
985 tcg_ctx = &tcg_init_ctx;
986 }
987 #else
tcg_register_thread(void)988 void tcg_register_thread(void)
989 {
990 TCGContext *s = g_malloc(sizeof(*s));
991 unsigned int i, n;
992
993 *s = tcg_init_ctx;
994
995 /* Relink mem_base. */
996 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) {
997 if (tcg_init_ctx.temps[i].mem_base) {
998 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps;
999 tcg_debug_assert(b >= 0 && b < n);
1000 s->temps[i].mem_base = &s->temps[b];
1001 }
1002 }
1003
1004 /* Claim an entry in tcg_ctxs */
1005 n = qatomic_fetch_inc(&tcg_cur_ctxs);
1006 g_assert(n < tcg_max_ctxs);
1007 qatomic_set(&tcg_ctxs[n], s);
1008
1009 if (n > 0) {
1010 tcg_region_initial_alloc(s);
1011 }
1012
1013 tcg_ctx = s;
1014 }
1015 #endif /* !CONFIG_USER_ONLY */
1016
1017 /* pool based memory allocation */
tcg_malloc_internal(TCGContext * s,int size)1018 void *tcg_malloc_internal(TCGContext *s, int size)
1019 {
1020 TCGPool *p;
1021 int pool_size;
1022
1023 if (size > TCG_POOL_CHUNK_SIZE) {
1024 /* big malloc: insert a new pool (XXX: could optimize) */
1025 p = g_malloc(sizeof(TCGPool) + size);
1026 p->size = size;
1027 p->next = s->pool_first_large;
1028 s->pool_first_large = p;
1029 return p->data;
1030 } else {
1031 p = s->pool_current;
1032 if (!p) {
1033 p = s->pool_first;
1034 if (!p)
1035 goto new_pool;
1036 } else {
1037 if (!p->next) {
1038 new_pool:
1039 pool_size = TCG_POOL_CHUNK_SIZE;
1040 p = g_malloc(sizeof(TCGPool) + pool_size);
1041 p->size = pool_size;
1042 p->next = NULL;
1043 if (s->pool_current) {
1044 s->pool_current->next = p;
1045 } else {
1046 s->pool_first = p;
1047 }
1048 } else {
1049 p = p->next;
1050 }
1051 }
1052 }
1053 s->pool_current = p;
1054 s->pool_cur = p->data + size;
1055 s->pool_end = p->data + p->size;
1056 return p->data;
1057 }
1058
tcg_pool_reset(TCGContext * s)1059 void tcg_pool_reset(TCGContext *s)
1060 {
1061 TCGPool *p, *t;
1062 for (p = s->pool_first_large; p; p = t) {
1063 t = p->next;
1064 g_free(p);
1065 }
1066 s->pool_first_large = NULL;
1067 s->pool_cur = s->pool_end = NULL;
1068 s->pool_current = NULL;
1069 }
1070
1071 /*
1072 * Create TCGHelperInfo structures for "tcg/tcg-ldst.h" functions,
1073 * akin to what "exec/helper-tcg.h" does with DEF_HELPER_FLAGS_N.
1074 * We only use these for layout in tcg_out_ld_helper_ret and
1075 * tcg_out_st_helper_args, and share them between several of
1076 * the helpers, with the end result that it's easier to build manually.
1077 */
1078
1079 #if TCG_TARGET_REG_BITS == 32
1080 # define dh_typecode_ttl dh_typecode_i32
1081 #else
1082 # define dh_typecode_ttl dh_typecode_i64
1083 #endif
1084
1085 static TCGHelperInfo info_helper_ld32_mmu = {
1086 .flags = TCG_CALL_NO_WG,
1087 .typemask = dh_typemask(ttl, 0) /* return tcg_target_ulong */
1088 | dh_typemask(env, 1)
1089 | dh_typemask(i64, 2) /* uint64_t addr */
1090 | dh_typemask(i32, 3) /* unsigned oi */
1091 | dh_typemask(ptr, 4) /* uintptr_t ra */
1092 };
1093
1094 static TCGHelperInfo info_helper_ld64_mmu = {
1095 .flags = TCG_CALL_NO_WG,
1096 .typemask = dh_typemask(i64, 0) /* return uint64_t */
1097 | dh_typemask(env, 1)
1098 | dh_typemask(i64, 2) /* uint64_t addr */
1099 | dh_typemask(i32, 3) /* unsigned oi */
1100 | dh_typemask(ptr, 4) /* uintptr_t ra */
1101 };
1102
1103 static TCGHelperInfo info_helper_ld128_mmu = {
1104 .flags = TCG_CALL_NO_WG,
1105 .typemask = dh_typemask(i128, 0) /* return Int128 */
1106 | dh_typemask(env, 1)
1107 | dh_typemask(i64, 2) /* uint64_t addr */
1108 | dh_typemask(i32, 3) /* unsigned oi */
1109 | dh_typemask(ptr, 4) /* uintptr_t ra */
1110 };
1111
1112 static TCGHelperInfo info_helper_st32_mmu = {
1113 .flags = TCG_CALL_NO_WG,
1114 .typemask = dh_typemask(void, 0)
1115 | dh_typemask(env, 1)
1116 | dh_typemask(i64, 2) /* uint64_t addr */
1117 | dh_typemask(i32, 3) /* uint32_t data */
1118 | dh_typemask(i32, 4) /* unsigned oi */
1119 | dh_typemask(ptr, 5) /* uintptr_t ra */
1120 };
1121
1122 static TCGHelperInfo info_helper_st64_mmu = {
1123 .flags = TCG_CALL_NO_WG,
1124 .typemask = dh_typemask(void, 0)
1125 | dh_typemask(env, 1)
1126 | dh_typemask(i64, 2) /* uint64_t addr */
1127 | dh_typemask(i64, 3) /* uint64_t data */
1128 | dh_typemask(i32, 4) /* unsigned oi */
1129 | dh_typemask(ptr, 5) /* uintptr_t ra */
1130 };
1131
1132 static TCGHelperInfo info_helper_st128_mmu = {
1133 .flags = TCG_CALL_NO_WG,
1134 .typemask = dh_typemask(void, 0)
1135 | dh_typemask(env, 1)
1136 | dh_typemask(i64, 2) /* uint64_t addr */
1137 | dh_typemask(i128, 3) /* Int128 data */
1138 | dh_typemask(i32, 4) /* unsigned oi */
1139 | dh_typemask(ptr, 5) /* uintptr_t ra */
1140 };
1141
1142 #ifdef CONFIG_TCG_INTERPRETER
typecode_to_ffi(int argmask)1143 static ffi_type *typecode_to_ffi(int argmask)
1144 {
1145 /*
1146 * libffi does not support __int128_t, so we have forced Int128
1147 * to use the structure definition instead of the builtin type.
1148 */
1149 static ffi_type *ffi_type_i128_elements[3] = {
1150 &ffi_type_uint64,
1151 &ffi_type_uint64,
1152 NULL
1153 };
1154 static ffi_type ffi_type_i128 = {
1155 .size = 16,
1156 .alignment = __alignof__(Int128),
1157 .type = FFI_TYPE_STRUCT,
1158 .elements = ffi_type_i128_elements,
1159 };
1160
1161 switch (argmask) {
1162 case dh_typecode_void:
1163 return &ffi_type_void;
1164 case dh_typecode_i32:
1165 return &ffi_type_uint32;
1166 case dh_typecode_s32:
1167 return &ffi_type_sint32;
1168 case dh_typecode_i64:
1169 return &ffi_type_uint64;
1170 case dh_typecode_s64:
1171 return &ffi_type_sint64;
1172 case dh_typecode_ptr:
1173 return &ffi_type_pointer;
1174 case dh_typecode_i128:
1175 return &ffi_type_i128;
1176 }
1177 g_assert_not_reached();
1178 }
1179
init_ffi_layout(TCGHelperInfo * info)1180 static ffi_cif *init_ffi_layout(TCGHelperInfo *info)
1181 {
1182 unsigned typemask = info->typemask;
1183 struct {
1184 ffi_cif cif;
1185 ffi_type *args[];
1186 } *ca;
1187 ffi_status status;
1188 int nargs;
1189
1190 /* Ignoring the return type, find the last non-zero field. */
1191 nargs = 32 - clz32(typemask >> 3);
1192 nargs = DIV_ROUND_UP(nargs, 3);
1193 assert(nargs <= MAX_CALL_IARGS);
1194
1195 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *));
1196 ca->cif.rtype = typecode_to_ffi(typemask & 7);
1197 ca->cif.nargs = nargs;
1198
1199 if (nargs != 0) {
1200 ca->cif.arg_types = ca->args;
1201 for (int j = 0; j < nargs; ++j) {
1202 int typecode = extract32(typemask, (j + 1) * 3, 3);
1203 ca->args[j] = typecode_to_ffi(typecode);
1204 }
1205 }
1206
1207 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs,
1208 ca->cif.rtype, ca->cif.arg_types);
1209 assert(status == FFI_OK);
1210
1211 return &ca->cif;
1212 }
1213
1214 #define HELPER_INFO_INIT(I) (&(I)->cif)
1215 #define HELPER_INFO_INIT_VAL(I) init_ffi_layout(I)
1216 #else
1217 #define HELPER_INFO_INIT(I) (&(I)->init)
1218 #define HELPER_INFO_INIT_VAL(I) 1
1219 #endif /* CONFIG_TCG_INTERPRETER */
1220
arg_slot_reg_p(unsigned arg_slot)1221 static inline bool arg_slot_reg_p(unsigned arg_slot)
1222 {
1223 /*
1224 * Split the sizeof away from the comparison to avoid Werror from
1225 * "unsigned < 0 is always false", when iarg_regs is empty.
1226 */
1227 unsigned nreg = ARRAY_SIZE(tcg_target_call_iarg_regs);
1228 return arg_slot < nreg;
1229 }
1230
arg_slot_stk_ofs(unsigned arg_slot)1231 static inline int arg_slot_stk_ofs(unsigned arg_slot)
1232 {
1233 unsigned max = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long);
1234 unsigned stk_slot = arg_slot - ARRAY_SIZE(tcg_target_call_iarg_regs);
1235
1236 tcg_debug_assert(stk_slot < max);
1237 return TCG_TARGET_CALL_STACK_OFFSET + stk_slot * sizeof(tcg_target_long);
1238 }
1239
1240 typedef struct TCGCumulativeArgs {
1241 int arg_idx; /* tcg_gen_callN args[] */
1242 int info_in_idx; /* TCGHelperInfo in[] */
1243 int arg_slot; /* regs+stack slot */
1244 int ref_slot; /* stack slots for references */
1245 } TCGCumulativeArgs;
1246
layout_arg_even(TCGCumulativeArgs * cum)1247 static void layout_arg_even(TCGCumulativeArgs *cum)
1248 {
1249 cum->arg_slot += cum->arg_slot & 1;
1250 }
1251
layout_arg_1(TCGCumulativeArgs * cum,TCGHelperInfo * info,TCGCallArgumentKind kind)1252 static void layout_arg_1(TCGCumulativeArgs *cum, TCGHelperInfo *info,
1253 TCGCallArgumentKind kind)
1254 {
1255 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx];
1256
1257 *loc = (TCGCallArgumentLoc){
1258 .kind = kind,
1259 .arg_idx = cum->arg_idx,
1260 .arg_slot = cum->arg_slot,
1261 };
1262 cum->info_in_idx++;
1263 cum->arg_slot++;
1264 }
1265
layout_arg_normal_n(TCGCumulativeArgs * cum,TCGHelperInfo * info,int n)1266 static void layout_arg_normal_n(TCGCumulativeArgs *cum,
1267 TCGHelperInfo *info, int n)
1268 {
1269 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx];
1270
1271 for (int i = 0; i < n; ++i) {
1272 /* Layout all using the same arg_idx, adjusting the subindex. */
1273 loc[i] = (TCGCallArgumentLoc){
1274 .kind = TCG_CALL_ARG_NORMAL,
1275 .arg_idx = cum->arg_idx,
1276 .tmp_subindex = i,
1277 .arg_slot = cum->arg_slot + i,
1278 };
1279 }
1280 cum->info_in_idx += n;
1281 cum->arg_slot += n;
1282 }
1283
layout_arg_by_ref(TCGCumulativeArgs * cum,TCGHelperInfo * info)1284 static void layout_arg_by_ref(TCGCumulativeArgs *cum, TCGHelperInfo *info)
1285 {
1286 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx];
1287 int n = 128 / TCG_TARGET_REG_BITS;
1288
1289 /* The first subindex carries the pointer. */
1290 layout_arg_1(cum, info, TCG_CALL_ARG_BY_REF);
1291
1292 /*
1293 * The callee is allowed to clobber memory associated with
1294 * structure pass by-reference. Therefore we must make copies.
1295 * Allocate space from "ref_slot", which will be adjusted to
1296 * follow the parameters on the stack.
1297 */
1298 loc[0].ref_slot = cum->ref_slot;
1299
1300 /*
1301 * Subsequent words also go into the reference slot, but
1302 * do not accumulate into the regular arguments.
1303 */
1304 for (int i = 1; i < n; ++i) {
1305 loc[i] = (TCGCallArgumentLoc){
1306 .kind = TCG_CALL_ARG_BY_REF_N,
1307 .arg_idx = cum->arg_idx,
1308 .tmp_subindex = i,
1309 .ref_slot = cum->ref_slot + i,
1310 };
1311 }
1312 cum->info_in_idx += n - 1; /* i=0 accounted for in layout_arg_1 */
1313 cum->ref_slot += n;
1314 }
1315
init_call_layout(TCGHelperInfo * info)1316 static void init_call_layout(TCGHelperInfo *info)
1317 {
1318 int max_reg_slots = ARRAY_SIZE(tcg_target_call_iarg_regs);
1319 int max_stk_slots = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long);
1320 unsigned typemask = info->typemask;
1321 unsigned typecode;
1322 TCGCumulativeArgs cum = { };
1323
1324 /*
1325 * Parse and place any function return value.
1326 */
1327 typecode = typemask & 7;
1328 switch (typecode) {
1329 case dh_typecode_void:
1330 info->nr_out = 0;
1331 break;
1332 case dh_typecode_i32:
1333 case dh_typecode_s32:
1334 case dh_typecode_ptr:
1335 info->nr_out = 1;
1336 info->out_kind = TCG_CALL_RET_NORMAL;
1337 break;
1338 case dh_typecode_i64:
1339 case dh_typecode_s64:
1340 info->nr_out = 64 / TCG_TARGET_REG_BITS;
1341 info->out_kind = TCG_CALL_RET_NORMAL;
1342 /* Query the last register now to trigger any assert early. */
1343 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1);
1344 break;
1345 case dh_typecode_i128:
1346 info->nr_out = 128 / TCG_TARGET_REG_BITS;
1347 info->out_kind = TCG_TARGET_CALL_RET_I128;
1348 switch (TCG_TARGET_CALL_RET_I128) {
1349 case TCG_CALL_RET_NORMAL:
1350 /* Query the last register now to trigger any assert early. */
1351 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1);
1352 break;
1353 case TCG_CALL_RET_BY_VEC:
1354 /* Query the single register now to trigger any assert early. */
1355 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0);
1356 break;
1357 case TCG_CALL_RET_BY_REF:
1358 /*
1359 * Allocate the first argument to the output.
1360 * We don't need to store this anywhere, just make it
1361 * unavailable for use in the input loop below.
1362 */
1363 cum.arg_slot = 1;
1364 break;
1365 default:
1366 qemu_build_not_reached();
1367 }
1368 break;
1369 default:
1370 g_assert_not_reached();
1371 }
1372
1373 /*
1374 * Parse and place function arguments.
1375 */
1376 for (typemask >>= 3; typemask; typemask >>= 3, cum.arg_idx++) {
1377 TCGCallArgumentKind kind;
1378 TCGType type;
1379
1380 typecode = typemask & 7;
1381 switch (typecode) {
1382 case dh_typecode_i32:
1383 case dh_typecode_s32:
1384 type = TCG_TYPE_I32;
1385 break;
1386 case dh_typecode_i64:
1387 case dh_typecode_s64:
1388 type = TCG_TYPE_I64;
1389 break;
1390 case dh_typecode_ptr:
1391 type = TCG_TYPE_PTR;
1392 break;
1393 case dh_typecode_i128:
1394 type = TCG_TYPE_I128;
1395 break;
1396 default:
1397 g_assert_not_reached();
1398 }
1399
1400 switch (type) {
1401 case TCG_TYPE_I32:
1402 switch (TCG_TARGET_CALL_ARG_I32) {
1403 case TCG_CALL_ARG_EVEN:
1404 layout_arg_even(&cum);
1405 /* fall through */
1406 case TCG_CALL_ARG_NORMAL:
1407 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL);
1408 break;
1409 case TCG_CALL_ARG_EXTEND:
1410 kind = TCG_CALL_ARG_EXTEND_U + (typecode & 1);
1411 layout_arg_1(&cum, info, kind);
1412 break;
1413 default:
1414 qemu_build_not_reached();
1415 }
1416 break;
1417
1418 case TCG_TYPE_I64:
1419 switch (TCG_TARGET_CALL_ARG_I64) {
1420 case TCG_CALL_ARG_EVEN:
1421 layout_arg_even(&cum);
1422 /* fall through */
1423 case TCG_CALL_ARG_NORMAL:
1424 if (TCG_TARGET_REG_BITS == 32) {
1425 layout_arg_normal_n(&cum, info, 2);
1426 } else {
1427 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL);
1428 }
1429 break;
1430 default:
1431 qemu_build_not_reached();
1432 }
1433 break;
1434
1435 case TCG_TYPE_I128:
1436 switch (TCG_TARGET_CALL_ARG_I128) {
1437 case TCG_CALL_ARG_EVEN:
1438 layout_arg_even(&cum);
1439 /* fall through */
1440 case TCG_CALL_ARG_NORMAL:
1441 layout_arg_normal_n(&cum, info, 128 / TCG_TARGET_REG_BITS);
1442 break;
1443 case TCG_CALL_ARG_BY_REF:
1444 layout_arg_by_ref(&cum, info);
1445 break;
1446 default:
1447 qemu_build_not_reached();
1448 }
1449 break;
1450
1451 default:
1452 g_assert_not_reached();
1453 }
1454 }
1455 info->nr_in = cum.info_in_idx;
1456
1457 /* Validate that we didn't overrun the input array. */
1458 assert(cum.info_in_idx <= ARRAY_SIZE(info->in));
1459 /* Validate the backend has enough argument space. */
1460 assert(cum.arg_slot <= max_reg_slots + max_stk_slots);
1461
1462 /*
1463 * Relocate the "ref_slot" area to the end of the parameters.
1464 * Minimizing this stack offset helps code size for x86,
1465 * which has a signed 8-bit offset encoding.
1466 */
1467 if (cum.ref_slot != 0) {
1468 int ref_base = 0;
1469
1470 if (cum.arg_slot > max_reg_slots) {
1471 int align = __alignof(Int128) / sizeof(tcg_target_long);
1472
1473 ref_base = cum.arg_slot - max_reg_slots;
1474 if (align > 1) {
1475 ref_base = ROUND_UP(ref_base, align);
1476 }
1477 }
1478 assert(ref_base + cum.ref_slot <= max_stk_slots);
1479 ref_base += max_reg_slots;
1480
1481 if (ref_base != 0) {
1482 for (int i = cum.info_in_idx - 1; i >= 0; --i) {
1483 TCGCallArgumentLoc *loc = &info->in[i];
1484 switch (loc->kind) {
1485 case TCG_CALL_ARG_BY_REF:
1486 case TCG_CALL_ARG_BY_REF_N:
1487 loc->ref_slot += ref_base;
1488 break;
1489 default:
1490 break;
1491 }
1492 }
1493 }
1494 }
1495 }
1496
1497 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)];
1498 static void process_constraint_sets(void);
1499 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type,
1500 TCGReg reg, const char *name);
1501
tcg_context_init(unsigned max_cpus)1502 static void tcg_context_init(unsigned max_cpus)
1503 {
1504 TCGContext *s = &tcg_init_ctx;
1505 int n, i;
1506 TCGTemp *ts;
1507
1508 memset(s, 0, sizeof(*s));
1509 s->nb_globals = 0;
1510
1511 init_call_layout(&info_helper_ld32_mmu);
1512 init_call_layout(&info_helper_ld64_mmu);
1513 init_call_layout(&info_helper_ld128_mmu);
1514 init_call_layout(&info_helper_st32_mmu);
1515 init_call_layout(&info_helper_st64_mmu);
1516 init_call_layout(&info_helper_st128_mmu);
1517
1518 tcg_target_init(s);
1519 process_constraint_sets();
1520
1521 /* Reverse the order of the saved registers, assuming they're all at
1522 the start of tcg_target_reg_alloc_order. */
1523 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) {
1524 int r = tcg_target_reg_alloc_order[n];
1525 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) {
1526 break;
1527 }
1528 }
1529 for (i = 0; i < n; ++i) {
1530 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i];
1531 }
1532 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) {
1533 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i];
1534 }
1535
1536 tcg_ctx = s;
1537 /*
1538 * In user-mode we simply share the init context among threads, since we
1539 * use a single region. See the documentation tcg_region_init() for the
1540 * reasoning behind this.
1541 * In system-mode we will have at most max_cpus TCG threads.
1542 */
1543 #ifdef CONFIG_USER_ONLY
1544 tcg_ctxs = &tcg_ctx;
1545 tcg_cur_ctxs = 1;
1546 tcg_max_ctxs = 1;
1547 #else
1548 tcg_max_ctxs = max_cpus;
1549 tcg_ctxs = g_new0(TCGContext *, max_cpus);
1550 #endif
1551
1552 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0));
1553 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env");
1554 tcg_env = temp_tcgv_ptr(ts);
1555 }
1556
tcg_init(size_t tb_size,int splitwx,unsigned max_cpus)1557 void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus)
1558 {
1559 tcg_context_init(max_cpus);
1560 tcg_region_init(tb_size, splitwx, max_cpus);
1561 }
1562
1563 /*
1564 * Allocate TBs right before their corresponding translated code, making
1565 * sure that TBs and code are on different cache lines.
1566 */
tcg_tb_alloc(TCGContext * s)1567 TranslationBlock *tcg_tb_alloc(TCGContext *s)
1568 {
1569 uintptr_t align = qemu_icache_linesize;
1570 TranslationBlock *tb;
1571 void *next;
1572
1573 retry:
1574 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align);
1575 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align);
1576
1577 if (unlikely(next > s->code_gen_highwater)) {
1578 if (tcg_region_alloc(s)) {
1579 return NULL;
1580 }
1581 goto retry;
1582 }
1583 qatomic_set(&s->code_gen_ptr, next);
1584 return tb;
1585 }
1586
tcg_prologue_init(void)1587 void tcg_prologue_init(void)
1588 {
1589 TCGContext *s = tcg_ctx;
1590 size_t prologue_size;
1591
1592 s->code_ptr = s->code_gen_ptr;
1593 s->code_buf = s->code_gen_ptr;
1594 s->data_gen_ptr = NULL;
1595
1596 #ifndef CONFIG_TCG_INTERPRETER
1597 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr);
1598 #endif
1599
1600 s->pool_labels = NULL;
1601
1602 qemu_thread_jit_write();
1603 /* Generate the prologue. */
1604 tcg_target_qemu_prologue(s);
1605
1606 /* Allow the prologue to put e.g. guest_base into a pool entry. */
1607 {
1608 int result = tcg_out_pool_finalize(s);
1609 tcg_debug_assert(result == 0);
1610 }
1611
1612 prologue_size = tcg_current_code_size(s);
1613 perf_report_prologue(s->code_gen_ptr, prologue_size);
1614
1615 #ifndef CONFIG_TCG_INTERPRETER
1616 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf),
1617 (uintptr_t)s->code_buf, prologue_size);
1618 #endif
1619
1620 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) {
1621 FILE *logfile = qemu_log_trylock();
1622 if (logfile) {
1623 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size);
1624 if (s->data_gen_ptr) {
1625 size_t code_size = s->data_gen_ptr - s->code_gen_ptr;
1626 size_t data_size = prologue_size - code_size;
1627 size_t i;
1628
1629 disas(logfile, s->code_gen_ptr, code_size);
1630
1631 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) {
1632 if (sizeof(tcg_target_ulong) == 8) {
1633 fprintf(logfile,
1634 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n",
1635 (uintptr_t)s->data_gen_ptr + i,
1636 *(uint64_t *)(s->data_gen_ptr + i));
1637 } else {
1638 fprintf(logfile,
1639 "0x%08" PRIxPTR ": .long 0x%08x\n",
1640 (uintptr_t)s->data_gen_ptr + i,
1641 *(uint32_t *)(s->data_gen_ptr + i));
1642 }
1643 }
1644 } else {
1645 disas(logfile, s->code_gen_ptr, prologue_size);
1646 }
1647 fprintf(logfile, "\n");
1648 qemu_log_unlock(logfile);
1649 }
1650 }
1651
1652 #ifndef CONFIG_TCG_INTERPRETER
1653 /*
1654 * Assert that goto_ptr is implemented completely, setting an epilogue.
1655 * For tci, we use NULL as the signal to return from the interpreter,
1656 * so skip this check.
1657 */
1658 tcg_debug_assert(tcg_code_gen_epilogue != NULL);
1659 #endif
1660
1661 tcg_region_prologue_set(s);
1662 }
1663
tcg_func_start(TCGContext * s)1664 void tcg_func_start(TCGContext *s)
1665 {
1666 tcg_pool_reset(s);
1667 s->nb_temps = s->nb_globals;
1668
1669 /* No temps have been previously allocated for size or locality. */
1670 tcg_temp_ebb_reset_freed(s);
1671
1672 /* No constant temps have been previously allocated. */
1673 for (int i = 0; i < TCG_TYPE_COUNT; ++i) {
1674 if (s->const_table[i]) {
1675 g_hash_table_remove_all(s->const_table[i]);
1676 }
1677 }
1678
1679 s->nb_ops = 0;
1680 s->nb_labels = 0;
1681 s->current_frame_offset = s->frame_start;
1682
1683 #ifdef CONFIG_DEBUG_TCG
1684 s->goto_tb_issue_mask = 0;
1685 #endif
1686
1687 QTAILQ_INIT(&s->ops);
1688 QTAILQ_INIT(&s->free_ops);
1689 s->emit_before_op = NULL;
1690 QSIMPLEQ_INIT(&s->labels);
1691
1692 tcg_debug_assert(s->addr_type <= TCG_TYPE_REG);
1693 tcg_debug_assert(s->insn_start_words > 0);
1694 }
1695
tcg_temp_alloc(TCGContext * s)1696 static TCGTemp *tcg_temp_alloc(TCGContext *s)
1697 {
1698 int n = s->nb_temps++;
1699
1700 if (n >= TCG_MAX_TEMPS) {
1701 tcg_raise_tb_overflow(s);
1702 }
1703 return memset(&s->temps[n], 0, sizeof(TCGTemp));
1704 }
1705
tcg_global_alloc(TCGContext * s)1706 static TCGTemp *tcg_global_alloc(TCGContext *s)
1707 {
1708 TCGTemp *ts;
1709
1710 tcg_debug_assert(s->nb_globals == s->nb_temps);
1711 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS);
1712 s->nb_globals++;
1713 ts = tcg_temp_alloc(s);
1714 ts->kind = TEMP_GLOBAL;
1715
1716 return ts;
1717 }
1718
tcg_global_reg_new_internal(TCGContext * s,TCGType type,TCGReg reg,const char * name)1719 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type,
1720 TCGReg reg, const char *name)
1721 {
1722 TCGTemp *ts;
1723
1724 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32);
1725
1726 ts = tcg_global_alloc(s);
1727 ts->base_type = type;
1728 ts->type = type;
1729 ts->kind = TEMP_FIXED;
1730 ts->reg = reg;
1731 ts->name = name;
1732 tcg_regset_set_reg(s->reserved_regs, reg);
1733
1734 return ts;
1735 }
1736
tcg_set_frame(TCGContext * s,TCGReg reg,intptr_t start,intptr_t size)1737 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size)
1738 {
1739 s->frame_start = start;
1740 s->frame_end = start + size;
1741 s->frame_temp
1742 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame");
1743 }
1744
tcg_global_mem_new_internal(TCGv_ptr base,intptr_t offset,const char * name,TCGType type)1745 static TCGTemp *tcg_global_mem_new_internal(TCGv_ptr base, intptr_t offset,
1746 const char *name, TCGType type)
1747 {
1748 TCGContext *s = tcg_ctx;
1749 TCGTemp *base_ts = tcgv_ptr_temp(base);
1750 TCGTemp *ts = tcg_global_alloc(s);
1751 int indirect_reg = 0;
1752
1753 switch (base_ts->kind) {
1754 case TEMP_FIXED:
1755 break;
1756 case TEMP_GLOBAL:
1757 /* We do not support double-indirect registers. */
1758 tcg_debug_assert(!base_ts->indirect_reg);
1759 base_ts->indirect_base = 1;
1760 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64
1761 ? 2 : 1);
1762 indirect_reg = 1;
1763 break;
1764 default:
1765 g_assert_not_reached();
1766 }
1767
1768 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) {
1769 TCGTemp *ts2 = tcg_global_alloc(s);
1770 char buf[64];
1771
1772 ts->base_type = TCG_TYPE_I64;
1773 ts->type = TCG_TYPE_I32;
1774 ts->indirect_reg = indirect_reg;
1775 ts->mem_allocated = 1;
1776 ts->mem_base = base_ts;
1777 ts->mem_offset = offset;
1778 pstrcpy(buf, sizeof(buf), name);
1779 pstrcat(buf, sizeof(buf), "_0");
1780 ts->name = strdup(buf);
1781
1782 tcg_debug_assert(ts2 == ts + 1);
1783 ts2->base_type = TCG_TYPE_I64;
1784 ts2->type = TCG_TYPE_I32;
1785 ts2->indirect_reg = indirect_reg;
1786 ts2->mem_allocated = 1;
1787 ts2->mem_base = base_ts;
1788 ts2->mem_offset = offset + 4;
1789 ts2->temp_subindex = 1;
1790 pstrcpy(buf, sizeof(buf), name);
1791 pstrcat(buf, sizeof(buf), "_1");
1792 ts2->name = strdup(buf);
1793 } else {
1794 ts->base_type = type;
1795 ts->type = type;
1796 ts->indirect_reg = indirect_reg;
1797 ts->mem_allocated = 1;
1798 ts->mem_base = base_ts;
1799 ts->mem_offset = offset;
1800 ts->name = name;
1801 }
1802 return ts;
1803 }
1804
tcg_global_mem_new_i32(TCGv_ptr reg,intptr_t off,const char * name)1805 TCGv_i32 tcg_global_mem_new_i32(TCGv_ptr reg, intptr_t off, const char *name)
1806 {
1807 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I32);
1808 return temp_tcgv_i32(ts);
1809 }
1810
tcg_global_mem_new_i64(TCGv_ptr reg,intptr_t off,const char * name)1811 TCGv_i64 tcg_global_mem_new_i64(TCGv_ptr reg, intptr_t off, const char *name)
1812 {
1813 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I64);
1814 return temp_tcgv_i64(ts);
1815 }
1816
tcg_global_mem_new_ptr(TCGv_ptr reg,intptr_t off,const char * name)1817 TCGv_ptr tcg_global_mem_new_ptr(TCGv_ptr reg, intptr_t off, const char *name)
1818 {
1819 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_PTR);
1820 return temp_tcgv_ptr(ts);
1821 }
1822
tcg_temp_new_internal(TCGType type,TCGTempKind kind)1823 TCGTemp *tcg_temp_new_internal(TCGType type, TCGTempKind kind)
1824 {
1825 TCGContext *s = tcg_ctx;
1826 TCGTemp *ts;
1827 int n;
1828
1829 if (kind == TEMP_EBB) {
1830 int idx = find_first_bit(s->free_temps[type].l, TCG_MAX_TEMPS);
1831
1832 if (idx < TCG_MAX_TEMPS) {
1833 /* There is already an available temp with the right type. */
1834 clear_bit(idx, s->free_temps[type].l);
1835
1836 ts = &s->temps[idx];
1837 ts->temp_allocated = 1;
1838 tcg_debug_assert(ts->base_type == type);
1839 tcg_debug_assert(ts->kind == kind);
1840 return ts;
1841 }
1842 } else {
1843 tcg_debug_assert(kind == TEMP_TB);
1844 }
1845
1846 switch (type) {
1847 case TCG_TYPE_I32:
1848 case TCG_TYPE_V64:
1849 case TCG_TYPE_V128:
1850 case TCG_TYPE_V256:
1851 n = 1;
1852 break;
1853 case TCG_TYPE_I64:
1854 n = 64 / TCG_TARGET_REG_BITS;
1855 break;
1856 case TCG_TYPE_I128:
1857 n = 128 / TCG_TARGET_REG_BITS;
1858 break;
1859 default:
1860 g_assert_not_reached();
1861 }
1862
1863 ts = tcg_temp_alloc(s);
1864 ts->base_type = type;
1865 ts->temp_allocated = 1;
1866 ts->kind = kind;
1867
1868 if (n == 1) {
1869 ts->type = type;
1870 } else {
1871 ts->type = TCG_TYPE_REG;
1872
1873 for (int i = 1; i < n; ++i) {
1874 TCGTemp *ts2 = tcg_temp_alloc(s);
1875
1876 tcg_debug_assert(ts2 == ts + i);
1877 ts2->base_type = type;
1878 ts2->type = TCG_TYPE_REG;
1879 ts2->temp_allocated = 1;
1880 ts2->temp_subindex = i;
1881 ts2->kind = kind;
1882 }
1883 }
1884 return ts;
1885 }
1886
tcg_temp_new_i32(void)1887 TCGv_i32 tcg_temp_new_i32(void)
1888 {
1889 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_TB));
1890 }
1891
tcg_temp_ebb_new_i32(void)1892 TCGv_i32 tcg_temp_ebb_new_i32(void)
1893 {
1894 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_EBB));
1895 }
1896
tcg_temp_new_i64(void)1897 TCGv_i64 tcg_temp_new_i64(void)
1898 {
1899 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_TB));
1900 }
1901
tcg_temp_ebb_new_i64(void)1902 TCGv_i64 tcg_temp_ebb_new_i64(void)
1903 {
1904 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_EBB));
1905 }
1906
tcg_temp_new_ptr(void)1907 TCGv_ptr tcg_temp_new_ptr(void)
1908 {
1909 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_TB));
1910 }
1911
tcg_temp_ebb_new_ptr(void)1912 TCGv_ptr tcg_temp_ebb_new_ptr(void)
1913 {
1914 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_EBB));
1915 }
1916
tcg_temp_new_i128(void)1917 TCGv_i128 tcg_temp_new_i128(void)
1918 {
1919 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_TB));
1920 }
1921
tcg_temp_ebb_new_i128(void)1922 TCGv_i128 tcg_temp_ebb_new_i128(void)
1923 {
1924 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_EBB));
1925 }
1926
tcg_temp_new_vec(TCGType type)1927 TCGv_vec tcg_temp_new_vec(TCGType type)
1928 {
1929 TCGTemp *t;
1930
1931 #ifdef CONFIG_DEBUG_TCG
1932 switch (type) {
1933 case TCG_TYPE_V64:
1934 assert(TCG_TARGET_HAS_v64);
1935 break;
1936 case TCG_TYPE_V128:
1937 assert(TCG_TARGET_HAS_v128);
1938 break;
1939 case TCG_TYPE_V256:
1940 assert(TCG_TARGET_HAS_v256);
1941 break;
1942 default:
1943 g_assert_not_reached();
1944 }
1945 #endif
1946
1947 t = tcg_temp_new_internal(type, TEMP_EBB);
1948 return temp_tcgv_vec(t);
1949 }
1950
1951 /* Create a new temp of the same type as an existing temp. */
tcg_temp_new_vec_matching(TCGv_vec match)1952 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match)
1953 {
1954 TCGTemp *t = tcgv_vec_temp(match);
1955
1956 tcg_debug_assert(t->temp_allocated != 0);
1957
1958 t = tcg_temp_new_internal(t->base_type, TEMP_EBB);
1959 return temp_tcgv_vec(t);
1960 }
1961
tcg_temp_free_internal(TCGTemp * ts)1962 void tcg_temp_free_internal(TCGTemp *ts)
1963 {
1964 TCGContext *s = tcg_ctx;
1965
1966 switch (ts->kind) {
1967 case TEMP_CONST:
1968 case TEMP_TB:
1969 /* Silently ignore free. */
1970 break;
1971 case TEMP_EBB:
1972 tcg_debug_assert(ts->temp_allocated != 0);
1973 ts->temp_allocated = 0;
1974 set_bit(temp_idx(ts), s->free_temps[ts->base_type].l);
1975 break;
1976 default:
1977 /* It never made sense to free TEMP_FIXED or TEMP_GLOBAL. */
1978 g_assert_not_reached();
1979 }
1980 }
1981
tcg_temp_free_i32(TCGv_i32 arg)1982 void tcg_temp_free_i32(TCGv_i32 arg)
1983 {
1984 tcg_temp_free_internal(tcgv_i32_temp(arg));
1985 }
1986
tcg_temp_free_i64(TCGv_i64 arg)1987 void tcg_temp_free_i64(TCGv_i64 arg)
1988 {
1989 tcg_temp_free_internal(tcgv_i64_temp(arg));
1990 }
1991
tcg_temp_free_i128(TCGv_i128 arg)1992 void tcg_temp_free_i128(TCGv_i128 arg)
1993 {
1994 tcg_temp_free_internal(tcgv_i128_temp(arg));
1995 }
1996
tcg_temp_free_ptr(TCGv_ptr arg)1997 void tcg_temp_free_ptr(TCGv_ptr arg)
1998 {
1999 tcg_temp_free_internal(tcgv_ptr_temp(arg));
2000 }
2001
tcg_temp_free_vec(TCGv_vec arg)2002 void tcg_temp_free_vec(TCGv_vec arg)
2003 {
2004 tcg_temp_free_internal(tcgv_vec_temp(arg));
2005 }
2006
tcg_constant_internal(TCGType type,int64_t val)2007 TCGTemp *tcg_constant_internal(TCGType type, int64_t val)
2008 {
2009 TCGContext *s = tcg_ctx;
2010 GHashTable *h = s->const_table[type];
2011 TCGTemp *ts;
2012
2013 if (h == NULL) {
2014 h = g_hash_table_new(g_int64_hash, g_int64_equal);
2015 s->const_table[type] = h;
2016 }
2017
2018 ts = g_hash_table_lookup(h, &val);
2019 if (ts == NULL) {
2020 int64_t *val_ptr;
2021
2022 ts = tcg_temp_alloc(s);
2023
2024 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) {
2025 TCGTemp *ts2 = tcg_temp_alloc(s);
2026
2027 tcg_debug_assert(ts2 == ts + 1);
2028
2029 ts->base_type = TCG_TYPE_I64;
2030 ts->type = TCG_TYPE_I32;
2031 ts->kind = TEMP_CONST;
2032 ts->temp_allocated = 1;
2033
2034 ts2->base_type = TCG_TYPE_I64;
2035 ts2->type = TCG_TYPE_I32;
2036 ts2->kind = TEMP_CONST;
2037 ts2->temp_allocated = 1;
2038 ts2->temp_subindex = 1;
2039
2040 /*
2041 * Retain the full value of the 64-bit constant in the low
2042 * part, so that the hash table works. Actual uses will
2043 * truncate the value to the low part.
2044 */
2045 ts[HOST_BIG_ENDIAN].val = val;
2046 ts[!HOST_BIG_ENDIAN].val = val >> 32;
2047 val_ptr = &ts[HOST_BIG_ENDIAN].val;
2048 } else {
2049 ts->base_type = type;
2050 ts->type = type;
2051 ts->kind = TEMP_CONST;
2052 ts->temp_allocated = 1;
2053 ts->val = val;
2054 val_ptr = &ts->val;
2055 }
2056 g_hash_table_insert(h, val_ptr, ts);
2057 }
2058
2059 return ts;
2060 }
2061
tcg_constant_i32(int32_t val)2062 TCGv_i32 tcg_constant_i32(int32_t val)
2063 {
2064 return temp_tcgv_i32(tcg_constant_internal(TCG_TYPE_I32, val));
2065 }
2066
tcg_constant_i64(int64_t val)2067 TCGv_i64 tcg_constant_i64(int64_t val)
2068 {
2069 return temp_tcgv_i64(tcg_constant_internal(TCG_TYPE_I64, val));
2070 }
2071
tcg_constant_ptr_int(intptr_t val)2072 TCGv_ptr tcg_constant_ptr_int(intptr_t val)
2073 {
2074 return temp_tcgv_ptr(tcg_constant_internal(TCG_TYPE_PTR, val));
2075 }
2076
tcg_constant_vec(TCGType type,unsigned vece,int64_t val)2077 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val)
2078 {
2079 val = dup_const(vece, val);
2080 return temp_tcgv_vec(tcg_constant_internal(type, val));
2081 }
2082
tcg_constant_vec_matching(TCGv_vec match,unsigned vece,int64_t val)2083 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val)
2084 {
2085 TCGTemp *t = tcgv_vec_temp(match);
2086
2087 tcg_debug_assert(t->temp_allocated != 0);
2088 return tcg_constant_vec(t->base_type, vece, val);
2089 }
2090
2091 #ifdef CONFIG_DEBUG_TCG
temp_idx(TCGTemp * ts)2092 size_t temp_idx(TCGTemp *ts)
2093 {
2094 ptrdiff_t n = ts - tcg_ctx->temps;
2095 assert(n >= 0 && n < tcg_ctx->nb_temps);
2096 return n;
2097 }
2098
tcgv_i32_temp(TCGv_i32 v)2099 TCGTemp *tcgv_i32_temp(TCGv_i32 v)
2100 {
2101 uintptr_t o = (uintptr_t)v - offsetof(TCGContext, temps);
2102
2103 assert(o < sizeof(TCGTemp) * tcg_ctx->nb_temps);
2104 assert(o % sizeof(TCGTemp) == 0);
2105
2106 return (void *)tcg_ctx + (uintptr_t)v;
2107 }
2108 #endif /* CONFIG_DEBUG_TCG */
2109
2110 /*
2111 * Return true if OP may appear in the opcode stream with TYPE.
2112 * Test the runtime variable that controls each opcode.
2113 */
tcg_op_supported(TCGOpcode op,TCGType type,unsigned flags)2114 bool tcg_op_supported(TCGOpcode op, TCGType type, unsigned flags)
2115 {
2116 bool has_type;
2117
2118 switch (type) {
2119 case TCG_TYPE_I32:
2120 has_type = true;
2121 break;
2122 case TCG_TYPE_I64:
2123 has_type = TCG_TARGET_REG_BITS == 64;
2124 break;
2125 case TCG_TYPE_V64:
2126 has_type = TCG_TARGET_HAS_v64;
2127 break;
2128 case TCG_TYPE_V128:
2129 has_type = TCG_TARGET_HAS_v128;
2130 break;
2131 case TCG_TYPE_V256:
2132 has_type = TCG_TARGET_HAS_v256;
2133 break;
2134 default:
2135 has_type = false;
2136 break;
2137 }
2138
2139 switch (op) {
2140 case INDEX_op_discard:
2141 case INDEX_op_set_label:
2142 case INDEX_op_call:
2143 case INDEX_op_br:
2144 case INDEX_op_mb:
2145 case INDEX_op_insn_start:
2146 case INDEX_op_exit_tb:
2147 case INDEX_op_goto_tb:
2148 case INDEX_op_goto_ptr:
2149 case INDEX_op_qemu_ld_i32:
2150 case INDEX_op_qemu_st_i32:
2151 case INDEX_op_qemu_ld_i64:
2152 case INDEX_op_qemu_st_i64:
2153 return true;
2154
2155 case INDEX_op_qemu_st8_i32:
2156 return TCG_TARGET_HAS_qemu_st8_i32;
2157
2158 case INDEX_op_qemu_ld_i128:
2159 case INDEX_op_qemu_st_i128:
2160 return TCG_TARGET_HAS_qemu_ldst_i128;
2161
2162 case INDEX_op_mov_i32:
2163 case INDEX_op_setcond_i32:
2164 case INDEX_op_brcond_i32:
2165 case INDEX_op_movcond_i32:
2166 case INDEX_op_ld8u_i32:
2167 case INDEX_op_ld8s_i32:
2168 case INDEX_op_ld16u_i32:
2169 case INDEX_op_ld16s_i32:
2170 case INDEX_op_ld_i32:
2171 case INDEX_op_st8_i32:
2172 case INDEX_op_st16_i32:
2173 case INDEX_op_st_i32:
2174 case INDEX_op_add_i32:
2175 case INDEX_op_sub_i32:
2176 case INDEX_op_neg_i32:
2177 case INDEX_op_mul_i32:
2178 case INDEX_op_and_i32:
2179 case INDEX_op_or_i32:
2180 case INDEX_op_xor_i32:
2181 case INDEX_op_shl_i32:
2182 case INDEX_op_shr_i32:
2183 case INDEX_op_sar_i32:
2184 case INDEX_op_extract_i32:
2185 case INDEX_op_sextract_i32:
2186 case INDEX_op_deposit_i32:
2187 return true;
2188
2189 case INDEX_op_negsetcond_i32:
2190 return TCG_TARGET_HAS_negsetcond_i32;
2191 case INDEX_op_div_i32:
2192 case INDEX_op_divu_i32:
2193 return TCG_TARGET_HAS_div_i32;
2194 case INDEX_op_rem_i32:
2195 case INDEX_op_remu_i32:
2196 return TCG_TARGET_HAS_rem_i32;
2197 case INDEX_op_div2_i32:
2198 case INDEX_op_divu2_i32:
2199 return TCG_TARGET_HAS_div2_i32;
2200 case INDEX_op_rotl_i32:
2201 case INDEX_op_rotr_i32:
2202 return TCG_TARGET_HAS_rot_i32;
2203 case INDEX_op_extract2_i32:
2204 return TCG_TARGET_HAS_extract2_i32;
2205 case INDEX_op_add2_i32:
2206 return TCG_TARGET_HAS_add2_i32;
2207 case INDEX_op_sub2_i32:
2208 return TCG_TARGET_HAS_sub2_i32;
2209 case INDEX_op_mulu2_i32:
2210 return TCG_TARGET_HAS_mulu2_i32;
2211 case INDEX_op_muls2_i32:
2212 return TCG_TARGET_HAS_muls2_i32;
2213 case INDEX_op_muluh_i32:
2214 return TCG_TARGET_HAS_muluh_i32;
2215 case INDEX_op_mulsh_i32:
2216 return TCG_TARGET_HAS_mulsh_i32;
2217 case INDEX_op_ext8s_i32:
2218 return TCG_TARGET_HAS_ext8s_i32;
2219 case INDEX_op_ext16s_i32:
2220 return TCG_TARGET_HAS_ext16s_i32;
2221 case INDEX_op_ext8u_i32:
2222 return TCG_TARGET_HAS_ext8u_i32;
2223 case INDEX_op_ext16u_i32:
2224 return TCG_TARGET_HAS_ext16u_i32;
2225 case INDEX_op_bswap16_i32:
2226 return TCG_TARGET_HAS_bswap16_i32;
2227 case INDEX_op_bswap32_i32:
2228 return TCG_TARGET_HAS_bswap32_i32;
2229 case INDEX_op_not_i32:
2230 return TCG_TARGET_HAS_not_i32;
2231 case INDEX_op_andc_i32:
2232 return TCG_TARGET_HAS_andc_i32;
2233 case INDEX_op_orc_i32:
2234 return TCG_TARGET_HAS_orc_i32;
2235 case INDEX_op_eqv_i32:
2236 return TCG_TARGET_HAS_eqv_i32;
2237 case INDEX_op_nand_i32:
2238 return TCG_TARGET_HAS_nand_i32;
2239 case INDEX_op_nor_i32:
2240 return TCG_TARGET_HAS_nor_i32;
2241 case INDEX_op_clz_i32:
2242 return TCG_TARGET_HAS_clz_i32;
2243 case INDEX_op_ctz_i32:
2244 return TCG_TARGET_HAS_ctz_i32;
2245 case INDEX_op_ctpop_i32:
2246 return TCG_TARGET_HAS_ctpop_i32;
2247
2248 case INDEX_op_brcond2_i32:
2249 case INDEX_op_setcond2_i32:
2250 return TCG_TARGET_REG_BITS == 32;
2251
2252 case INDEX_op_mov_i64:
2253 case INDEX_op_setcond_i64:
2254 case INDEX_op_brcond_i64:
2255 case INDEX_op_movcond_i64:
2256 case INDEX_op_ld8u_i64:
2257 case INDEX_op_ld8s_i64:
2258 case INDEX_op_ld16u_i64:
2259 case INDEX_op_ld16s_i64:
2260 case INDEX_op_ld32u_i64:
2261 case INDEX_op_ld32s_i64:
2262 case INDEX_op_ld_i64:
2263 case INDEX_op_st8_i64:
2264 case INDEX_op_st16_i64:
2265 case INDEX_op_st32_i64:
2266 case INDEX_op_st_i64:
2267 case INDEX_op_add_i64:
2268 case INDEX_op_sub_i64:
2269 case INDEX_op_neg_i64:
2270 case INDEX_op_mul_i64:
2271 case INDEX_op_and_i64:
2272 case INDEX_op_or_i64:
2273 case INDEX_op_xor_i64:
2274 case INDEX_op_shl_i64:
2275 case INDEX_op_shr_i64:
2276 case INDEX_op_sar_i64:
2277 case INDEX_op_ext_i32_i64:
2278 case INDEX_op_extu_i32_i64:
2279 case INDEX_op_extract_i64:
2280 case INDEX_op_sextract_i64:
2281 case INDEX_op_deposit_i64:
2282 return TCG_TARGET_REG_BITS == 64;
2283
2284 case INDEX_op_negsetcond_i64:
2285 return TCG_TARGET_HAS_negsetcond_i64;
2286 case INDEX_op_div_i64:
2287 case INDEX_op_divu_i64:
2288 return TCG_TARGET_HAS_div_i64;
2289 case INDEX_op_rem_i64:
2290 case INDEX_op_remu_i64:
2291 return TCG_TARGET_HAS_rem_i64;
2292 case INDEX_op_div2_i64:
2293 case INDEX_op_divu2_i64:
2294 return TCG_TARGET_HAS_div2_i64;
2295 case INDEX_op_rotl_i64:
2296 case INDEX_op_rotr_i64:
2297 return TCG_TARGET_HAS_rot_i64;
2298 case INDEX_op_extract2_i64:
2299 return TCG_TARGET_HAS_extract2_i64;
2300 case INDEX_op_extrl_i64_i32:
2301 case INDEX_op_extrh_i64_i32:
2302 return TCG_TARGET_HAS_extr_i64_i32;
2303 case INDEX_op_ext8s_i64:
2304 return TCG_TARGET_HAS_ext8s_i64;
2305 case INDEX_op_ext16s_i64:
2306 return TCG_TARGET_HAS_ext16s_i64;
2307 case INDEX_op_ext32s_i64:
2308 return TCG_TARGET_HAS_ext32s_i64;
2309 case INDEX_op_ext8u_i64:
2310 return TCG_TARGET_HAS_ext8u_i64;
2311 case INDEX_op_ext16u_i64:
2312 return TCG_TARGET_HAS_ext16u_i64;
2313 case INDEX_op_ext32u_i64:
2314 return TCG_TARGET_HAS_ext32u_i64;
2315 case INDEX_op_bswap16_i64:
2316 return TCG_TARGET_HAS_bswap16_i64;
2317 case INDEX_op_bswap32_i64:
2318 return TCG_TARGET_HAS_bswap32_i64;
2319 case INDEX_op_bswap64_i64:
2320 return TCG_TARGET_HAS_bswap64_i64;
2321 case INDEX_op_not_i64:
2322 return TCG_TARGET_HAS_not_i64;
2323 case INDEX_op_andc_i64:
2324 return TCG_TARGET_HAS_andc_i64;
2325 case INDEX_op_orc_i64:
2326 return TCG_TARGET_HAS_orc_i64;
2327 case INDEX_op_eqv_i64:
2328 return TCG_TARGET_HAS_eqv_i64;
2329 case INDEX_op_nand_i64:
2330 return TCG_TARGET_HAS_nand_i64;
2331 case INDEX_op_nor_i64:
2332 return TCG_TARGET_HAS_nor_i64;
2333 case INDEX_op_clz_i64:
2334 return TCG_TARGET_HAS_clz_i64;
2335 case INDEX_op_ctz_i64:
2336 return TCG_TARGET_HAS_ctz_i64;
2337 case INDEX_op_ctpop_i64:
2338 return TCG_TARGET_HAS_ctpop_i64;
2339 case INDEX_op_add2_i64:
2340 return TCG_TARGET_HAS_add2_i64;
2341 case INDEX_op_sub2_i64:
2342 return TCG_TARGET_HAS_sub2_i64;
2343 case INDEX_op_mulu2_i64:
2344 return TCG_TARGET_HAS_mulu2_i64;
2345 case INDEX_op_muls2_i64:
2346 return TCG_TARGET_HAS_muls2_i64;
2347 case INDEX_op_muluh_i64:
2348 return TCG_TARGET_HAS_muluh_i64;
2349 case INDEX_op_mulsh_i64:
2350 return TCG_TARGET_HAS_mulsh_i64;
2351
2352 case INDEX_op_mov_vec:
2353 case INDEX_op_dup_vec:
2354 case INDEX_op_dupm_vec:
2355 case INDEX_op_ld_vec:
2356 case INDEX_op_st_vec:
2357 case INDEX_op_add_vec:
2358 case INDEX_op_sub_vec:
2359 case INDEX_op_and_vec:
2360 case INDEX_op_or_vec:
2361 case INDEX_op_xor_vec:
2362 case INDEX_op_cmp_vec:
2363 return has_type;
2364 case INDEX_op_dup2_vec:
2365 return has_type && TCG_TARGET_REG_BITS == 32;
2366 case INDEX_op_not_vec:
2367 return has_type && TCG_TARGET_HAS_not_vec;
2368 case INDEX_op_neg_vec:
2369 return has_type && TCG_TARGET_HAS_neg_vec;
2370 case INDEX_op_abs_vec:
2371 return has_type && TCG_TARGET_HAS_abs_vec;
2372 case INDEX_op_andc_vec:
2373 return has_type && TCG_TARGET_HAS_andc_vec;
2374 case INDEX_op_orc_vec:
2375 return has_type && TCG_TARGET_HAS_orc_vec;
2376 case INDEX_op_nand_vec:
2377 return has_type && TCG_TARGET_HAS_nand_vec;
2378 case INDEX_op_nor_vec:
2379 return has_type && TCG_TARGET_HAS_nor_vec;
2380 case INDEX_op_eqv_vec:
2381 return has_type && TCG_TARGET_HAS_eqv_vec;
2382 case INDEX_op_mul_vec:
2383 return has_type && TCG_TARGET_HAS_mul_vec;
2384 case INDEX_op_shli_vec:
2385 case INDEX_op_shri_vec:
2386 case INDEX_op_sari_vec:
2387 return has_type && TCG_TARGET_HAS_shi_vec;
2388 case INDEX_op_shls_vec:
2389 case INDEX_op_shrs_vec:
2390 case INDEX_op_sars_vec:
2391 return has_type && TCG_TARGET_HAS_shs_vec;
2392 case INDEX_op_shlv_vec:
2393 case INDEX_op_shrv_vec:
2394 case INDEX_op_sarv_vec:
2395 return has_type && TCG_TARGET_HAS_shv_vec;
2396 case INDEX_op_rotli_vec:
2397 return has_type && TCG_TARGET_HAS_roti_vec;
2398 case INDEX_op_rotls_vec:
2399 return has_type && TCG_TARGET_HAS_rots_vec;
2400 case INDEX_op_rotlv_vec:
2401 case INDEX_op_rotrv_vec:
2402 return has_type && TCG_TARGET_HAS_rotv_vec;
2403 case INDEX_op_ssadd_vec:
2404 case INDEX_op_usadd_vec:
2405 case INDEX_op_sssub_vec:
2406 case INDEX_op_ussub_vec:
2407 return has_type && TCG_TARGET_HAS_sat_vec;
2408 case INDEX_op_smin_vec:
2409 case INDEX_op_umin_vec:
2410 case INDEX_op_smax_vec:
2411 case INDEX_op_umax_vec:
2412 return has_type && TCG_TARGET_HAS_minmax_vec;
2413 case INDEX_op_bitsel_vec:
2414 return has_type && TCG_TARGET_HAS_bitsel_vec;
2415 case INDEX_op_cmpsel_vec:
2416 return has_type && TCG_TARGET_HAS_cmpsel_vec;
2417
2418 default:
2419 tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS);
2420 return true;
2421 }
2422 }
2423
tcg_op_deposit_valid(TCGType type,unsigned ofs,unsigned len)2424 bool tcg_op_deposit_valid(TCGType type, unsigned ofs, unsigned len)
2425 {
2426 unsigned width;
2427
2428 tcg_debug_assert(type == TCG_TYPE_I32 || type == TCG_TYPE_I64);
2429 width = (type == TCG_TYPE_I32 ? 32 : 64);
2430
2431 tcg_debug_assert(ofs < width);
2432 tcg_debug_assert(len > 0);
2433 tcg_debug_assert(len <= width - ofs);
2434
2435 return TCG_TARGET_deposit_valid(type, ofs, len);
2436 }
2437
2438 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs);
2439
tcg_gen_callN(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp ** args)2440 static void tcg_gen_callN(void *func, TCGHelperInfo *info,
2441 TCGTemp *ret, TCGTemp **args)
2442 {
2443 TCGv_i64 extend_free[MAX_CALL_IARGS];
2444 int n_extend = 0;
2445 TCGOp *op;
2446 int i, n, pi = 0, total_args;
2447
2448 if (unlikely(g_once_init_enter(HELPER_INFO_INIT(info)))) {
2449 init_call_layout(info);
2450 g_once_init_leave(HELPER_INFO_INIT(info), HELPER_INFO_INIT_VAL(info));
2451 }
2452
2453 total_args = info->nr_out + info->nr_in + 2;
2454 op = tcg_op_alloc(INDEX_op_call, total_args);
2455
2456 #ifdef CONFIG_PLUGIN
2457 /* Flag helpers that may affect guest state */
2458 if (tcg_ctx->plugin_insn && !(info->flags & TCG_CALL_NO_SIDE_EFFECTS)) {
2459 tcg_ctx->plugin_insn->calls_helpers = true;
2460 }
2461 #endif
2462
2463 TCGOP_CALLO(op) = n = info->nr_out;
2464 switch (n) {
2465 case 0:
2466 tcg_debug_assert(ret == NULL);
2467 break;
2468 case 1:
2469 tcg_debug_assert(ret != NULL);
2470 op->args[pi++] = temp_arg(ret);
2471 break;
2472 case 2:
2473 case 4:
2474 tcg_debug_assert(ret != NULL);
2475 tcg_debug_assert(ret->base_type == ret->type + ctz32(n));
2476 tcg_debug_assert(ret->temp_subindex == 0);
2477 for (i = 0; i < n; ++i) {
2478 op->args[pi++] = temp_arg(ret + i);
2479 }
2480 break;
2481 default:
2482 g_assert_not_reached();
2483 }
2484
2485 TCGOP_CALLI(op) = n = info->nr_in;
2486 for (i = 0; i < n; i++) {
2487 const TCGCallArgumentLoc *loc = &info->in[i];
2488 TCGTemp *ts = args[loc->arg_idx] + loc->tmp_subindex;
2489
2490 switch (loc->kind) {
2491 case TCG_CALL_ARG_NORMAL:
2492 case TCG_CALL_ARG_BY_REF:
2493 case TCG_CALL_ARG_BY_REF_N:
2494 op->args[pi++] = temp_arg(ts);
2495 break;
2496
2497 case TCG_CALL_ARG_EXTEND_U:
2498 case TCG_CALL_ARG_EXTEND_S:
2499 {
2500 TCGv_i64 temp = tcg_temp_ebb_new_i64();
2501 TCGv_i32 orig = temp_tcgv_i32(ts);
2502
2503 if (loc->kind == TCG_CALL_ARG_EXTEND_S) {
2504 tcg_gen_ext_i32_i64(temp, orig);
2505 } else {
2506 tcg_gen_extu_i32_i64(temp, orig);
2507 }
2508 op->args[pi++] = tcgv_i64_arg(temp);
2509 extend_free[n_extend++] = temp;
2510 }
2511 break;
2512
2513 default:
2514 g_assert_not_reached();
2515 }
2516 }
2517 op->args[pi++] = (uintptr_t)func;
2518 op->args[pi++] = (uintptr_t)info;
2519 tcg_debug_assert(pi == total_args);
2520
2521 if (tcg_ctx->emit_before_op) {
2522 QTAILQ_INSERT_BEFORE(tcg_ctx->emit_before_op, op, link);
2523 } else {
2524 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link);
2525 }
2526
2527 tcg_debug_assert(n_extend < ARRAY_SIZE(extend_free));
2528 for (i = 0; i < n_extend; ++i) {
2529 tcg_temp_free_i64(extend_free[i]);
2530 }
2531 }
2532
tcg_gen_call0(void * func,TCGHelperInfo * info,TCGTemp * ret)2533 void tcg_gen_call0(void *func, TCGHelperInfo *info, TCGTemp *ret)
2534 {
2535 tcg_gen_callN(func, info, ret, NULL);
2536 }
2537
tcg_gen_call1(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1)2538 void tcg_gen_call1(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1)
2539 {
2540 tcg_gen_callN(func, info, ret, &t1);
2541 }
2542
tcg_gen_call2(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1,TCGTemp * t2)2543 void tcg_gen_call2(void *func, TCGHelperInfo *info, TCGTemp *ret,
2544 TCGTemp *t1, TCGTemp *t2)
2545 {
2546 TCGTemp *args[2] = { t1, t2 };
2547 tcg_gen_callN(func, info, ret, args);
2548 }
2549
tcg_gen_call3(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1,TCGTemp * t2,TCGTemp * t3)2550 void tcg_gen_call3(void *func, TCGHelperInfo *info, TCGTemp *ret,
2551 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3)
2552 {
2553 TCGTemp *args[3] = { t1, t2, t3 };
2554 tcg_gen_callN(func, info, ret, args);
2555 }
2556
tcg_gen_call4(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1,TCGTemp * t2,TCGTemp * t3,TCGTemp * t4)2557 void tcg_gen_call4(void *func, TCGHelperInfo *info, TCGTemp *ret,
2558 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3, TCGTemp *t4)
2559 {
2560 TCGTemp *args[4] = { t1, t2, t3, t4 };
2561 tcg_gen_callN(func, info, ret, args);
2562 }
2563
tcg_gen_call5(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1,TCGTemp * t2,TCGTemp * t3,TCGTemp * t4,TCGTemp * t5)2564 void tcg_gen_call5(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1,
2565 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, TCGTemp *t5)
2566 {
2567 TCGTemp *args[5] = { t1, t2, t3, t4, t5 };
2568 tcg_gen_callN(func, info, ret, args);
2569 }
2570
tcg_gen_call6(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1,TCGTemp * t2,TCGTemp * t3,TCGTemp * t4,TCGTemp * t5,TCGTemp * t6)2571 void tcg_gen_call6(void *func, TCGHelperInfo *info, TCGTemp *ret,
2572 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3,
2573 TCGTemp *t4, TCGTemp *t5, TCGTemp *t6)
2574 {
2575 TCGTemp *args[6] = { t1, t2, t3, t4, t5, t6 };
2576 tcg_gen_callN(func, info, ret, args);
2577 }
2578
tcg_gen_call7(void * func,TCGHelperInfo * info,TCGTemp * ret,TCGTemp * t1,TCGTemp * t2,TCGTemp * t3,TCGTemp * t4,TCGTemp * t5,TCGTemp * t6,TCGTemp * t7)2579 void tcg_gen_call7(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1,
2580 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4,
2581 TCGTemp *t5, TCGTemp *t6, TCGTemp *t7)
2582 {
2583 TCGTemp *args[7] = { t1, t2, t3, t4, t5, t6, t7 };
2584 tcg_gen_callN(func, info, ret, args);
2585 }
2586
tcg_reg_alloc_start(TCGContext * s)2587 static void tcg_reg_alloc_start(TCGContext *s)
2588 {
2589 int i, n;
2590
2591 for (i = 0, n = s->nb_temps; i < n; i++) {
2592 TCGTemp *ts = &s->temps[i];
2593 TCGTempVal val = TEMP_VAL_MEM;
2594
2595 switch (ts->kind) {
2596 case TEMP_CONST:
2597 val = TEMP_VAL_CONST;
2598 break;
2599 case TEMP_FIXED:
2600 val = TEMP_VAL_REG;
2601 break;
2602 case TEMP_GLOBAL:
2603 break;
2604 case TEMP_EBB:
2605 val = TEMP_VAL_DEAD;
2606 /* fall through */
2607 case TEMP_TB:
2608 ts->mem_allocated = 0;
2609 break;
2610 default:
2611 g_assert_not_reached();
2612 }
2613 ts->val_type = val;
2614 }
2615
2616 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp));
2617 }
2618
tcg_get_arg_str_ptr(TCGContext * s,char * buf,int buf_size,TCGTemp * ts)2619 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size,
2620 TCGTemp *ts)
2621 {
2622 int idx = temp_idx(ts);
2623
2624 switch (ts->kind) {
2625 case TEMP_FIXED:
2626 case TEMP_GLOBAL:
2627 pstrcpy(buf, buf_size, ts->name);
2628 break;
2629 case TEMP_TB:
2630 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals);
2631 break;
2632 case TEMP_EBB:
2633 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals);
2634 break;
2635 case TEMP_CONST:
2636 switch (ts->type) {
2637 case TCG_TYPE_I32:
2638 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val);
2639 break;
2640 #if TCG_TARGET_REG_BITS > 32
2641 case TCG_TYPE_I64:
2642 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val);
2643 break;
2644 #endif
2645 case TCG_TYPE_V64:
2646 case TCG_TYPE_V128:
2647 case TCG_TYPE_V256:
2648 snprintf(buf, buf_size, "v%d$0x%" PRIx64,
2649 64 << (ts->type - TCG_TYPE_V64), ts->val);
2650 break;
2651 default:
2652 g_assert_not_reached();
2653 }
2654 break;
2655 }
2656 return buf;
2657 }
2658
tcg_get_arg_str(TCGContext * s,char * buf,int buf_size,TCGArg arg)2659 static char *tcg_get_arg_str(TCGContext *s, char *buf,
2660 int buf_size, TCGArg arg)
2661 {
2662 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg));
2663 }
2664
2665 static const char * const cond_name[] =
2666 {
2667 [TCG_COND_NEVER] = "never",
2668 [TCG_COND_ALWAYS] = "always",
2669 [TCG_COND_EQ] = "eq",
2670 [TCG_COND_NE] = "ne",
2671 [TCG_COND_LT] = "lt",
2672 [TCG_COND_GE] = "ge",
2673 [TCG_COND_LE] = "le",
2674 [TCG_COND_GT] = "gt",
2675 [TCG_COND_LTU] = "ltu",
2676 [TCG_COND_GEU] = "geu",
2677 [TCG_COND_LEU] = "leu",
2678 [TCG_COND_GTU] = "gtu",
2679 [TCG_COND_TSTEQ] = "tsteq",
2680 [TCG_COND_TSTNE] = "tstne",
2681 };
2682
2683 static const char * const ldst_name[(MO_BSWAP | MO_SSIZE) + 1] =
2684 {
2685 [MO_UB] = "ub",
2686 [MO_SB] = "sb",
2687 [MO_LEUW] = "leuw",
2688 [MO_LESW] = "lesw",
2689 [MO_LEUL] = "leul",
2690 [MO_LESL] = "lesl",
2691 [MO_LEUQ] = "leq",
2692 [MO_BEUW] = "beuw",
2693 [MO_BESW] = "besw",
2694 [MO_BEUL] = "beul",
2695 [MO_BESL] = "besl",
2696 [MO_BEUQ] = "beq",
2697 [MO_128 + MO_BE] = "beo",
2698 [MO_128 + MO_LE] = "leo",
2699 };
2700
2701 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = {
2702 [MO_UNALN >> MO_ASHIFT] = "un+",
2703 [MO_ALIGN >> MO_ASHIFT] = "al+",
2704 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+",
2705 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+",
2706 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+",
2707 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+",
2708 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+",
2709 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+",
2710 };
2711
2712 static const char * const atom_name[(MO_ATOM_MASK >> MO_ATOM_SHIFT) + 1] = {
2713 [MO_ATOM_IFALIGN >> MO_ATOM_SHIFT] = "",
2714 [MO_ATOM_IFALIGN_PAIR >> MO_ATOM_SHIFT] = "pair+",
2715 [MO_ATOM_WITHIN16 >> MO_ATOM_SHIFT] = "w16+",
2716 [MO_ATOM_WITHIN16_PAIR >> MO_ATOM_SHIFT] = "w16p+",
2717 [MO_ATOM_SUBALIGN >> MO_ATOM_SHIFT] = "sub+",
2718 [MO_ATOM_NONE >> MO_ATOM_SHIFT] = "noat+",
2719 };
2720
2721 static const char bswap_flag_name[][6] = {
2722 [TCG_BSWAP_IZ] = "iz",
2723 [TCG_BSWAP_OZ] = "oz",
2724 [TCG_BSWAP_OS] = "os",
2725 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz",
2726 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os",
2727 };
2728
2729 #ifdef CONFIG_PLUGIN
2730 static const char * const plugin_from_name[] = {
2731 "from-tb",
2732 "from-insn",
2733 "after-insn",
2734 "after-tb",
2735 };
2736 #endif
2737
tcg_regset_single(TCGRegSet d)2738 static inline bool tcg_regset_single(TCGRegSet d)
2739 {
2740 return (d & (d - 1)) == 0;
2741 }
2742
tcg_regset_first(TCGRegSet d)2743 static inline TCGReg tcg_regset_first(TCGRegSet d)
2744 {
2745 if (TCG_TARGET_NB_REGS <= 32) {
2746 return ctz32(d);
2747 } else {
2748 return ctz64(d);
2749 }
2750 }
2751
2752 /* Return only the number of characters output -- no error return. */
2753 #define ne_fprintf(...) \
2754 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; })
2755
tcg_dump_ops(TCGContext * s,FILE * f,bool have_prefs)2756 void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs)
2757 {
2758 char buf[128];
2759 TCGOp *op;
2760
2761 QTAILQ_FOREACH(op, &s->ops, link) {
2762 int i, k, nb_oargs, nb_iargs, nb_cargs;
2763 const TCGOpDef *def;
2764 TCGOpcode c;
2765 int col = 0;
2766
2767 c = op->opc;
2768 def = &tcg_op_defs[c];
2769
2770 if (c == INDEX_op_insn_start) {
2771 nb_oargs = 0;
2772 col += ne_fprintf(f, "\n ----");
2773
2774 for (i = 0, k = s->insn_start_words; i < k; ++i) {
2775 col += ne_fprintf(f, " %016" PRIx64,
2776 tcg_get_insn_start_param(op, i));
2777 }
2778 } else if (c == INDEX_op_call) {
2779 const TCGHelperInfo *info = tcg_call_info(op);
2780 void *func = tcg_call_func(op);
2781
2782 /* variable number of arguments */
2783 nb_oargs = TCGOP_CALLO(op);
2784 nb_iargs = TCGOP_CALLI(op);
2785 nb_cargs = def->nb_cargs;
2786
2787 col += ne_fprintf(f, " %s ", def->name);
2788
2789 /*
2790 * Print the function name from TCGHelperInfo, if available.
2791 * Note that plugins have a template function for the info,
2792 * but the actual function pointer comes from the plugin.
2793 */
2794 if (func == info->func) {
2795 col += ne_fprintf(f, "%s", info->name);
2796 } else {
2797 col += ne_fprintf(f, "plugin(%p)", func);
2798 }
2799
2800 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs);
2801 for (i = 0; i < nb_oargs; i++) {
2802 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf),
2803 op->args[i]));
2804 }
2805 for (i = 0; i < nb_iargs; i++) {
2806 TCGArg arg = op->args[nb_oargs + i];
2807 const char *t = tcg_get_arg_str(s, buf, sizeof(buf), arg);
2808 col += ne_fprintf(f, ",%s", t);
2809 }
2810 } else {
2811 col += ne_fprintf(f, " %s ", def->name);
2812
2813 nb_oargs = def->nb_oargs;
2814 nb_iargs = def->nb_iargs;
2815 nb_cargs = def->nb_cargs;
2816
2817 if (def->flags & TCG_OPF_VECTOR) {
2818 col += ne_fprintf(f, "v%d,e%d,",
2819 8 * tcg_type_size(TCGOP_TYPE(op)),
2820 8 << TCGOP_VECE(op));
2821 }
2822
2823 k = 0;
2824 for (i = 0; i < nb_oargs; i++) {
2825 const char *sep = k ? "," : "";
2826 col += ne_fprintf(f, "%s%s", sep,
2827 tcg_get_arg_str(s, buf, sizeof(buf),
2828 op->args[k++]));
2829 }
2830 for (i = 0; i < nb_iargs; i++) {
2831 const char *sep = k ? "," : "";
2832 col += ne_fprintf(f, "%s%s", sep,
2833 tcg_get_arg_str(s, buf, sizeof(buf),
2834 op->args[k++]));
2835 }
2836 switch (c) {
2837 case INDEX_op_brcond_i32:
2838 case INDEX_op_setcond_i32:
2839 case INDEX_op_negsetcond_i32:
2840 case INDEX_op_movcond_i32:
2841 case INDEX_op_brcond2_i32:
2842 case INDEX_op_setcond2_i32:
2843 case INDEX_op_brcond_i64:
2844 case INDEX_op_setcond_i64:
2845 case INDEX_op_negsetcond_i64:
2846 case INDEX_op_movcond_i64:
2847 case INDEX_op_cmp_vec:
2848 case INDEX_op_cmpsel_vec:
2849 if (op->args[k] < ARRAY_SIZE(cond_name)
2850 && cond_name[op->args[k]]) {
2851 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]);
2852 } else {
2853 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]);
2854 }
2855 i = 1;
2856 break;
2857 case INDEX_op_qemu_ld_i32:
2858 case INDEX_op_qemu_st_i32:
2859 case INDEX_op_qemu_st8_i32:
2860 case INDEX_op_qemu_ld_i64:
2861 case INDEX_op_qemu_st_i64:
2862 case INDEX_op_qemu_ld_i128:
2863 case INDEX_op_qemu_st_i128:
2864 {
2865 const char *s_al, *s_op, *s_at;
2866 MemOpIdx oi = op->args[k++];
2867 MemOp mop = get_memop(oi);
2868 unsigned ix = get_mmuidx(oi);
2869
2870 s_al = alignment_name[(mop & MO_AMASK) >> MO_ASHIFT];
2871 s_op = ldst_name[mop & (MO_BSWAP | MO_SSIZE)];
2872 s_at = atom_name[(mop & MO_ATOM_MASK) >> MO_ATOM_SHIFT];
2873 mop &= ~(MO_AMASK | MO_BSWAP | MO_SSIZE | MO_ATOM_MASK);
2874
2875 /* If all fields are accounted for, print symbolically. */
2876 if (!mop && s_al && s_op && s_at) {
2877 col += ne_fprintf(f, ",%s%s%s,%u",
2878 s_at, s_al, s_op, ix);
2879 } else {
2880 mop = get_memop(oi);
2881 col += ne_fprintf(f, ",$0x%x,%u", mop, ix);
2882 }
2883 i = 1;
2884 }
2885 break;
2886 case INDEX_op_bswap16_i32:
2887 case INDEX_op_bswap16_i64:
2888 case INDEX_op_bswap32_i32:
2889 case INDEX_op_bswap32_i64:
2890 case INDEX_op_bswap64_i64:
2891 {
2892 TCGArg flags = op->args[k];
2893 const char *name = NULL;
2894
2895 if (flags < ARRAY_SIZE(bswap_flag_name)) {
2896 name = bswap_flag_name[flags];
2897 }
2898 if (name) {
2899 col += ne_fprintf(f, ",%s", name);
2900 } else {
2901 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags);
2902 }
2903 i = k = 1;
2904 }
2905 break;
2906 #ifdef CONFIG_PLUGIN
2907 case INDEX_op_plugin_cb:
2908 {
2909 TCGArg from = op->args[k++];
2910 const char *name = NULL;
2911
2912 if (from < ARRAY_SIZE(plugin_from_name)) {
2913 name = plugin_from_name[from];
2914 }
2915 if (name) {
2916 col += ne_fprintf(f, "%s", name);
2917 } else {
2918 col += ne_fprintf(f, "$0x%" TCG_PRIlx, from);
2919 }
2920 i = 1;
2921 }
2922 break;
2923 #endif
2924 default:
2925 i = 0;
2926 break;
2927 }
2928 switch (c) {
2929 case INDEX_op_set_label:
2930 case INDEX_op_br:
2931 case INDEX_op_brcond_i32:
2932 case INDEX_op_brcond_i64:
2933 case INDEX_op_brcond2_i32:
2934 col += ne_fprintf(f, "%s$L%d", k ? "," : "",
2935 arg_label(op->args[k])->id);
2936 i++, k++;
2937 break;
2938 case INDEX_op_mb:
2939 {
2940 TCGBar membar = op->args[k];
2941 const char *b_op, *m_op;
2942
2943 switch (membar & TCG_BAR_SC) {
2944 case 0:
2945 b_op = "none";
2946 break;
2947 case TCG_BAR_LDAQ:
2948 b_op = "acq";
2949 break;
2950 case TCG_BAR_STRL:
2951 b_op = "rel";
2952 break;
2953 case TCG_BAR_SC:
2954 b_op = "seq";
2955 break;
2956 default:
2957 g_assert_not_reached();
2958 }
2959
2960 switch (membar & TCG_MO_ALL) {
2961 case 0:
2962 m_op = "none";
2963 break;
2964 case TCG_MO_LD_LD:
2965 m_op = "rr";
2966 break;
2967 case TCG_MO_LD_ST:
2968 m_op = "rw";
2969 break;
2970 case TCG_MO_ST_LD:
2971 m_op = "wr";
2972 break;
2973 case TCG_MO_ST_ST:
2974 m_op = "ww";
2975 break;
2976 case TCG_MO_LD_LD | TCG_MO_LD_ST:
2977 m_op = "rr+rw";
2978 break;
2979 case TCG_MO_LD_LD | TCG_MO_ST_LD:
2980 m_op = "rr+wr";
2981 break;
2982 case TCG_MO_LD_LD | TCG_MO_ST_ST:
2983 m_op = "rr+ww";
2984 break;
2985 case TCG_MO_LD_ST | TCG_MO_ST_LD:
2986 m_op = "rw+wr";
2987 break;
2988 case TCG_MO_LD_ST | TCG_MO_ST_ST:
2989 m_op = "rw+ww";
2990 break;
2991 case TCG_MO_ST_LD | TCG_MO_ST_ST:
2992 m_op = "wr+ww";
2993 break;
2994 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_LD:
2995 m_op = "rr+rw+wr";
2996 break;
2997 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_ST:
2998 m_op = "rr+rw+ww";
2999 break;
3000 case TCG_MO_LD_LD | TCG_MO_ST_LD | TCG_MO_ST_ST:
3001 m_op = "rr+wr+ww";
3002 break;
3003 case TCG_MO_LD_ST | TCG_MO_ST_LD | TCG_MO_ST_ST:
3004 m_op = "rw+wr+ww";
3005 break;
3006 case TCG_MO_ALL:
3007 m_op = "all";
3008 break;
3009 default:
3010 g_assert_not_reached();
3011 }
3012
3013 col += ne_fprintf(f, "%s%s:%s", (k ? "," : ""), b_op, m_op);
3014 i++, k++;
3015 }
3016 break;
3017 default:
3018 break;
3019 }
3020 for (; i < nb_cargs; i++, k++) {
3021 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "",
3022 op->args[k]);
3023 }
3024 }
3025
3026 if (have_prefs || op->life) {
3027 for (; col < 40; ++col) {
3028 putc(' ', f);
3029 }
3030 }
3031
3032 if (op->life) {
3033 unsigned life = op->life;
3034
3035 if (life & (SYNC_ARG * 3)) {
3036 ne_fprintf(f, " sync:");
3037 for (i = 0; i < 2; ++i) {
3038 if (life & (SYNC_ARG << i)) {
3039 ne_fprintf(f, " %d", i);
3040 }
3041 }
3042 }
3043 life /= DEAD_ARG;
3044 if (life) {
3045 ne_fprintf(f, " dead:");
3046 for (i = 0; life; ++i, life >>= 1) {
3047 if (life & 1) {
3048 ne_fprintf(f, " %d", i);
3049 }
3050 }
3051 }
3052 }
3053
3054 if (have_prefs) {
3055 for (i = 0; i < nb_oargs; ++i) {
3056 TCGRegSet set = output_pref(op, i);
3057
3058 if (i == 0) {
3059 ne_fprintf(f, " pref=");
3060 } else {
3061 ne_fprintf(f, ",");
3062 }
3063 if (set == 0) {
3064 ne_fprintf(f, "none");
3065 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) {
3066 ne_fprintf(f, "all");
3067 #ifdef CONFIG_DEBUG_TCG
3068 } else if (tcg_regset_single(set)) {
3069 TCGReg reg = tcg_regset_first(set);
3070 ne_fprintf(f, "%s", tcg_target_reg_names[reg]);
3071 #endif
3072 } else if (TCG_TARGET_NB_REGS <= 32) {
3073 ne_fprintf(f, "0x%x", (uint32_t)set);
3074 } else {
3075 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set);
3076 }
3077 }
3078 }
3079
3080 putc('\n', f);
3081 }
3082 }
3083
3084 /* we give more priority to constraints with less registers */
get_constraint_priority(const TCGArgConstraint * arg_ct,int k)3085 static int get_constraint_priority(const TCGArgConstraint *arg_ct, int k)
3086 {
3087 int n;
3088
3089 arg_ct += k;
3090 n = ctpop64(arg_ct->regs);
3091
3092 /*
3093 * Sort constraints of a single register first, which includes output
3094 * aliases (which must exactly match the input already allocated).
3095 */
3096 if (n == 1 || arg_ct->oalias) {
3097 return INT_MAX;
3098 }
3099
3100 /*
3101 * Sort register pairs next, first then second immediately after.
3102 * Arbitrarily sort multiple pairs by the index of the first reg;
3103 * there shouldn't be many pairs.
3104 */
3105 switch (arg_ct->pair) {
3106 case 1:
3107 case 3:
3108 return (k + 1) * 2;
3109 case 2:
3110 return (arg_ct->pair_index + 1) * 2 - 1;
3111 }
3112
3113 /* Finally, sort by decreasing register count. */
3114 assert(n > 1);
3115 return -n;
3116 }
3117
3118 /* sort from highest priority to lowest */
sort_constraints(TCGArgConstraint * a,int start,int n)3119 static void sort_constraints(TCGArgConstraint *a, int start, int n)
3120 {
3121 int i, j;
3122
3123 for (i = 0; i < n; i++) {
3124 a[start + i].sort_index = start + i;
3125 }
3126 if (n <= 1) {
3127 return;
3128 }
3129 for (i = 0; i < n - 1; i++) {
3130 for (j = i + 1; j < n; j++) {
3131 int p1 = get_constraint_priority(a, a[start + i].sort_index);
3132 int p2 = get_constraint_priority(a, a[start + j].sort_index);
3133 if (p1 < p2) {
3134 int tmp = a[start + i].sort_index;
3135 a[start + i].sort_index = a[start + j].sort_index;
3136 a[start + j].sort_index = tmp;
3137 }
3138 }
3139 }
3140 }
3141
3142 static const TCGArgConstraint empty_cts[TCG_MAX_OP_ARGS];
3143 static TCGArgConstraint all_cts[ARRAY_SIZE(constraint_sets)][TCG_MAX_OP_ARGS];
3144
process_constraint_sets(void)3145 static void process_constraint_sets(void)
3146 {
3147 for (size_t c = 0; c < ARRAY_SIZE(constraint_sets); ++c) {
3148 const TCGConstraintSet *tdefs = &constraint_sets[c];
3149 TCGArgConstraint *args_ct = all_cts[c];
3150 int nb_oargs = tdefs->nb_oargs;
3151 int nb_iargs = tdefs->nb_iargs;
3152 int nb_args = nb_oargs + nb_iargs;
3153 bool saw_alias_pair = false;
3154
3155 for (int i = 0; i < nb_args; i++) {
3156 const char *ct_str = tdefs->args_ct_str[i];
3157 bool input_p = i >= nb_oargs;
3158 int o;
3159
3160 switch (*ct_str) {
3161 case '0' ... '9':
3162 o = *ct_str - '0';
3163 tcg_debug_assert(input_p);
3164 tcg_debug_assert(o < nb_oargs);
3165 tcg_debug_assert(args_ct[o].regs != 0);
3166 tcg_debug_assert(!args_ct[o].oalias);
3167 args_ct[i] = args_ct[o];
3168 /* The output sets oalias. */
3169 args_ct[o].oalias = 1;
3170 args_ct[o].alias_index = i;
3171 /* The input sets ialias. */
3172 args_ct[i].ialias = 1;
3173 args_ct[i].alias_index = o;
3174 if (args_ct[i].pair) {
3175 saw_alias_pair = true;
3176 }
3177 tcg_debug_assert(ct_str[1] == '\0');
3178 continue;
3179
3180 case '&':
3181 tcg_debug_assert(!input_p);
3182 args_ct[i].newreg = true;
3183 ct_str++;
3184 break;
3185
3186 case 'p': /* plus */
3187 /* Allocate to the register after the previous. */
3188 tcg_debug_assert(i > (input_p ? nb_oargs : 0));
3189 o = i - 1;
3190 tcg_debug_assert(!args_ct[o].pair);
3191 tcg_debug_assert(!args_ct[o].ct);
3192 args_ct[i] = (TCGArgConstraint){
3193 .pair = 2,
3194 .pair_index = o,
3195 .regs = args_ct[o].regs << 1,
3196 .newreg = args_ct[o].newreg,
3197 };
3198 args_ct[o].pair = 1;
3199 args_ct[o].pair_index = i;
3200 tcg_debug_assert(ct_str[1] == '\0');
3201 continue;
3202
3203 case 'm': /* minus */
3204 /* Allocate to the register before the previous. */
3205 tcg_debug_assert(i > (input_p ? nb_oargs : 0));
3206 o = i - 1;
3207 tcg_debug_assert(!args_ct[o].pair);
3208 tcg_debug_assert(!args_ct[o].ct);
3209 args_ct[i] = (TCGArgConstraint){
3210 .pair = 1,
3211 .pair_index = o,
3212 .regs = args_ct[o].regs >> 1,
3213 .newreg = args_ct[o].newreg,
3214 };
3215 args_ct[o].pair = 2;
3216 args_ct[o].pair_index = i;
3217 tcg_debug_assert(ct_str[1] == '\0');
3218 continue;
3219 }
3220
3221 do {
3222 switch (*ct_str) {
3223 case 'i':
3224 args_ct[i].ct |= TCG_CT_CONST;
3225 break;
3226 #ifdef TCG_REG_ZERO
3227 case 'z':
3228 args_ct[i].ct |= TCG_CT_REG_ZERO;
3229 break;
3230 #endif
3231
3232 /* Include all of the target-specific constraints. */
3233
3234 #undef CONST
3235 #define CONST(CASE, MASK) \
3236 case CASE: args_ct[i].ct |= MASK; break;
3237 #define REGS(CASE, MASK) \
3238 case CASE: args_ct[i].regs |= MASK; break;
3239
3240 #include "tcg-target-con-str.h"
3241
3242 #undef REGS
3243 #undef CONST
3244 default:
3245 case '0' ... '9':
3246 case '&':
3247 case 'p':
3248 case 'm':
3249 /* Typo in TCGConstraintSet constraint. */
3250 g_assert_not_reached();
3251 }
3252 } while (*++ct_str != '\0');
3253 }
3254
3255 /*
3256 * Fix up output pairs that are aliased with inputs.
3257 * When we created the alias, we copied pair from the output.
3258 * There are three cases:
3259 * (1a) Pairs of inputs alias pairs of outputs.
3260 * (1b) One input aliases the first of a pair of outputs.
3261 * (2) One input aliases the second of a pair of outputs.
3262 *
3263 * Case 1a is handled by making sure that the pair_index'es are
3264 * properly updated so that they appear the same as a pair of inputs.
3265 *
3266 * Case 1b is handled by setting the pair_index of the input to
3267 * itself, simply so it doesn't point to an unrelated argument.
3268 * Since we don't encounter the "second" during the input allocation
3269 * phase, nothing happens with the second half of the input pair.
3270 *
3271 * Case 2 is handled by setting the second input to pair=3, the
3272 * first output to pair=3, and the pair_index'es to match.
3273 */
3274 if (saw_alias_pair) {
3275 for (int i = nb_oargs; i < nb_args; i++) {
3276 int o, o2, i2;
3277
3278 /*
3279 * Since [0-9pm] must be alone in the constraint string,
3280 * the only way they can both be set is if the pair comes
3281 * from the output alias.
3282 */
3283 if (!args_ct[i].ialias) {
3284 continue;
3285 }
3286 switch (args_ct[i].pair) {
3287 case 0:
3288 break;
3289 case 1:
3290 o = args_ct[i].alias_index;
3291 o2 = args_ct[o].pair_index;
3292 tcg_debug_assert(args_ct[o].pair == 1);
3293 tcg_debug_assert(args_ct[o2].pair == 2);
3294 if (args_ct[o2].oalias) {
3295 /* Case 1a */
3296 i2 = args_ct[o2].alias_index;
3297 tcg_debug_assert(args_ct[i2].pair == 2);
3298 args_ct[i2].pair_index = i;
3299 args_ct[i].pair_index = i2;
3300 } else {
3301 /* Case 1b */
3302 args_ct[i].pair_index = i;
3303 }
3304 break;
3305 case 2:
3306 o = args_ct[i].alias_index;
3307 o2 = args_ct[o].pair_index;
3308 tcg_debug_assert(args_ct[o].pair == 2);
3309 tcg_debug_assert(args_ct[o2].pair == 1);
3310 if (args_ct[o2].oalias) {
3311 /* Case 1a */
3312 i2 = args_ct[o2].alias_index;
3313 tcg_debug_assert(args_ct[i2].pair == 1);
3314 args_ct[i2].pair_index = i;
3315 args_ct[i].pair_index = i2;
3316 } else {
3317 /* Case 2 */
3318 args_ct[i].pair = 3;
3319 args_ct[o2].pair = 3;
3320 args_ct[i].pair_index = o2;
3321 args_ct[o2].pair_index = i;
3322 }
3323 break;
3324 default:
3325 g_assert_not_reached();
3326 }
3327 }
3328 }
3329
3330 /* sort the constraints (XXX: this is just an heuristic) */
3331 sort_constraints(args_ct, 0, nb_oargs);
3332 sort_constraints(args_ct, nb_oargs, nb_iargs);
3333 }
3334 }
3335
opcode_args_ct(const TCGOp * op)3336 static const TCGArgConstraint *opcode_args_ct(const TCGOp *op)
3337 {
3338 const TCGOpDef *def = &tcg_op_defs[op->opc];
3339 TCGConstraintSetIndex con_set;
3340
3341 #ifdef CONFIG_DEBUG_TCG
3342 assert(tcg_op_supported(op->opc, TCGOP_TYPE(op), TCGOP_FLAGS(op)));
3343 #endif
3344
3345 if (def->flags & TCG_OPF_NOT_PRESENT) {
3346 return empty_cts;
3347 }
3348
3349 con_set = tcg_target_op_def(op->opc, TCGOP_TYPE(op), TCGOP_FLAGS(op));
3350 tcg_debug_assert(con_set >= 0 && con_set < ARRAY_SIZE(constraint_sets));
3351
3352 /* The constraint arguments must match TCGOpcode arguments. */
3353 tcg_debug_assert(constraint_sets[con_set].nb_oargs == def->nb_oargs);
3354 tcg_debug_assert(constraint_sets[con_set].nb_iargs == def->nb_iargs);
3355
3356 return all_cts[con_set];
3357 }
3358
remove_label_use(TCGOp * op,int idx)3359 static void remove_label_use(TCGOp *op, int idx)
3360 {
3361 TCGLabel *label = arg_label(op->args[idx]);
3362 TCGLabelUse *use;
3363
3364 QSIMPLEQ_FOREACH(use, &label->branches, next) {
3365 if (use->op == op) {
3366 QSIMPLEQ_REMOVE(&label->branches, use, TCGLabelUse, next);
3367 return;
3368 }
3369 }
3370 g_assert_not_reached();
3371 }
3372
tcg_op_remove(TCGContext * s,TCGOp * op)3373 void tcg_op_remove(TCGContext *s, TCGOp *op)
3374 {
3375 switch (op->opc) {
3376 case INDEX_op_br:
3377 remove_label_use(op, 0);
3378 break;
3379 case INDEX_op_brcond_i32:
3380 case INDEX_op_brcond_i64:
3381 remove_label_use(op, 3);
3382 break;
3383 case INDEX_op_brcond2_i32:
3384 remove_label_use(op, 5);
3385 break;
3386 default:
3387 break;
3388 }
3389
3390 QTAILQ_REMOVE(&s->ops, op, link);
3391 QTAILQ_INSERT_TAIL(&s->free_ops, op, link);
3392 s->nb_ops--;
3393 }
3394
tcg_remove_ops_after(TCGOp * op)3395 void tcg_remove_ops_after(TCGOp *op)
3396 {
3397 TCGContext *s = tcg_ctx;
3398
3399 while (true) {
3400 TCGOp *last = tcg_last_op();
3401 if (last == op) {
3402 return;
3403 }
3404 tcg_op_remove(s, last);
3405 }
3406 }
3407
tcg_op_alloc(TCGOpcode opc,unsigned nargs)3408 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs)
3409 {
3410 TCGContext *s = tcg_ctx;
3411 TCGOp *op = NULL;
3412
3413 if (unlikely(!QTAILQ_EMPTY(&s->free_ops))) {
3414 QTAILQ_FOREACH(op, &s->free_ops, link) {
3415 if (nargs <= op->nargs) {
3416 QTAILQ_REMOVE(&s->free_ops, op, link);
3417 nargs = op->nargs;
3418 goto found;
3419 }
3420 }
3421 }
3422
3423 /* Most opcodes have 3 or 4 operands: reduce fragmentation. */
3424 nargs = MAX(4, nargs);
3425 op = tcg_malloc(sizeof(TCGOp) + sizeof(TCGArg) * nargs);
3426
3427 found:
3428 memset(op, 0, offsetof(TCGOp, link));
3429 op->opc = opc;
3430 op->nargs = nargs;
3431
3432 /* Check for bitfield overflow. */
3433 tcg_debug_assert(op->nargs == nargs);
3434
3435 s->nb_ops++;
3436 return op;
3437 }
3438
tcg_emit_op(TCGOpcode opc,unsigned nargs)3439 TCGOp *tcg_emit_op(TCGOpcode opc, unsigned nargs)
3440 {
3441 TCGOp *op = tcg_op_alloc(opc, nargs);
3442
3443 if (tcg_ctx->emit_before_op) {
3444 QTAILQ_INSERT_BEFORE(tcg_ctx->emit_before_op, op, link);
3445 } else {
3446 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link);
3447 }
3448 return op;
3449 }
3450
tcg_op_insert_before(TCGContext * s,TCGOp * old_op,TCGOpcode opc,unsigned nargs)3451 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op,
3452 TCGOpcode opc, unsigned nargs)
3453 {
3454 TCGOp *new_op = tcg_op_alloc(opc, nargs);
3455
3456 TCGOP_TYPE(new_op) = TCGOP_TYPE(old_op);
3457 QTAILQ_INSERT_BEFORE(old_op, new_op, link);
3458 return new_op;
3459 }
3460
tcg_op_insert_after(TCGContext * s,TCGOp * old_op,TCGOpcode opc,unsigned nargs)3461 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op,
3462 TCGOpcode opc, unsigned nargs)
3463 {
3464 TCGOp *new_op = tcg_op_alloc(opc, nargs);
3465
3466 TCGOP_TYPE(new_op) = TCGOP_TYPE(old_op);
3467 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link);
3468 return new_op;
3469 }
3470
move_label_uses(TCGLabel * to,TCGLabel * from)3471 static void move_label_uses(TCGLabel *to, TCGLabel *from)
3472 {
3473 TCGLabelUse *u;
3474
3475 QSIMPLEQ_FOREACH(u, &from->branches, next) {
3476 TCGOp *op = u->op;
3477 switch (op->opc) {
3478 case INDEX_op_br:
3479 op->args[0] = label_arg(to);
3480 break;
3481 case INDEX_op_brcond_i32:
3482 case INDEX_op_brcond_i64:
3483 op->args[3] = label_arg(to);
3484 break;
3485 case INDEX_op_brcond2_i32:
3486 op->args[5] = label_arg(to);
3487 break;
3488 default:
3489 g_assert_not_reached();
3490 }
3491 }
3492
3493 QSIMPLEQ_CONCAT(&to->branches, &from->branches);
3494 }
3495
3496 /* Reachable analysis : remove unreachable code. */
3497 static void __attribute__((noinline))
reachable_code_pass(TCGContext * s)3498 reachable_code_pass(TCGContext *s)
3499 {
3500 TCGOp *op, *op_next, *op_prev;
3501 bool dead = false;
3502
3503 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
3504 bool remove = dead;
3505 TCGLabel *label;
3506
3507 switch (op->opc) {
3508 case INDEX_op_set_label:
3509 label = arg_label(op->args[0]);
3510
3511 /*
3512 * Note that the first op in the TB is always a load,
3513 * so there is always something before a label.
3514 */
3515 op_prev = QTAILQ_PREV(op, link);
3516
3517 /*
3518 * If we find two sequential labels, move all branches to
3519 * reference the second label and remove the first label.
3520 * Do this before branch to next optimization, so that the
3521 * middle label is out of the way.
3522 */
3523 if (op_prev->opc == INDEX_op_set_label) {
3524 move_label_uses(label, arg_label(op_prev->args[0]));
3525 tcg_op_remove(s, op_prev);
3526 op_prev = QTAILQ_PREV(op, link);
3527 }
3528
3529 /*
3530 * Optimization can fold conditional branches to unconditional.
3531 * If we find a label which is preceded by an unconditional
3532 * branch to next, remove the branch. We couldn't do this when
3533 * processing the branch because any dead code between the branch
3534 * and label had not yet been removed.
3535 */
3536 if (op_prev->opc == INDEX_op_br &&
3537 label == arg_label(op_prev->args[0])) {
3538 tcg_op_remove(s, op_prev);
3539 /* Fall through means insns become live again. */
3540 dead = false;
3541 }
3542
3543 if (QSIMPLEQ_EMPTY(&label->branches)) {
3544 /*
3545 * While there is an occasional backward branch, virtually
3546 * all branches generated by the translators are forward.
3547 * Which means that generally we will have already removed
3548 * all references to the label that will be, and there is
3549 * little to be gained by iterating.
3550 */
3551 remove = true;
3552 } else {
3553 /* Once we see a label, insns become live again. */
3554 dead = false;
3555 remove = false;
3556 }
3557 break;
3558
3559 case INDEX_op_br:
3560 case INDEX_op_exit_tb:
3561 case INDEX_op_goto_ptr:
3562 /* Unconditional branches; everything following is dead. */
3563 dead = true;
3564 break;
3565
3566 case INDEX_op_call:
3567 /* Notice noreturn helper calls, raising exceptions. */
3568 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) {
3569 dead = true;
3570 }
3571 break;
3572
3573 case INDEX_op_insn_start:
3574 /* Never remove -- we need to keep these for unwind. */
3575 remove = false;
3576 break;
3577
3578 default:
3579 break;
3580 }
3581
3582 if (remove) {
3583 tcg_op_remove(s, op);
3584 }
3585 }
3586 }
3587
3588 #define TS_DEAD 1
3589 #define TS_MEM 2
3590
3591 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n)))
3592 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n)))
3593
3594 /* For liveness_pass_1, the register preferences for a given temp. */
la_temp_pref(TCGTemp * ts)3595 static inline TCGRegSet *la_temp_pref(TCGTemp *ts)
3596 {
3597 return ts->state_ptr;
3598 }
3599
3600 /* For liveness_pass_1, reset the preferences for a given temp to the
3601 * maximal regset for its type.
3602 */
la_reset_pref(TCGTemp * ts)3603 static inline void la_reset_pref(TCGTemp *ts)
3604 {
3605 *la_temp_pref(ts)
3606 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]);
3607 }
3608
3609 /* liveness analysis: end of function: all temps are dead, and globals
3610 should be in memory. */
la_func_end(TCGContext * s,int ng,int nt)3611 static void la_func_end(TCGContext *s, int ng, int nt)
3612 {
3613 int i;
3614
3615 for (i = 0; i < ng; ++i) {
3616 s->temps[i].state = TS_DEAD | TS_MEM;
3617 la_reset_pref(&s->temps[i]);
3618 }
3619 for (i = ng; i < nt; ++i) {
3620 s->temps[i].state = TS_DEAD;
3621 la_reset_pref(&s->temps[i]);
3622 }
3623 }
3624
3625 /* liveness analysis: end of basic block: all temps are dead, globals
3626 and local temps should be in memory. */
la_bb_end(TCGContext * s,int ng,int nt)3627 static void la_bb_end(TCGContext *s, int ng, int nt)
3628 {
3629 int i;
3630
3631 for (i = 0; i < nt; ++i) {
3632 TCGTemp *ts = &s->temps[i];
3633 int state;
3634
3635 switch (ts->kind) {
3636 case TEMP_FIXED:
3637 case TEMP_GLOBAL:
3638 case TEMP_TB:
3639 state = TS_DEAD | TS_MEM;
3640 break;
3641 case TEMP_EBB:
3642 case TEMP_CONST:
3643 state = TS_DEAD;
3644 break;
3645 default:
3646 g_assert_not_reached();
3647 }
3648 ts->state = state;
3649 la_reset_pref(ts);
3650 }
3651 }
3652
3653 /* liveness analysis: sync globals back to memory. */
la_global_sync(TCGContext * s,int ng)3654 static void la_global_sync(TCGContext *s, int ng)
3655 {
3656 int i;
3657
3658 for (i = 0; i < ng; ++i) {
3659 int state = s->temps[i].state;
3660 s->temps[i].state = state | TS_MEM;
3661 if (state == TS_DEAD) {
3662 /* If the global was previously dead, reset prefs. */
3663 la_reset_pref(&s->temps[i]);
3664 }
3665 }
3666 }
3667
3668 /*
3669 * liveness analysis: conditional branch: all temps are dead unless
3670 * explicitly live-across-conditional-branch, globals and local temps
3671 * should be synced.
3672 */
la_bb_sync(TCGContext * s,int ng,int nt)3673 static void la_bb_sync(TCGContext *s, int ng, int nt)
3674 {
3675 la_global_sync(s, ng);
3676
3677 for (int i = ng; i < nt; ++i) {
3678 TCGTemp *ts = &s->temps[i];
3679 int state;
3680
3681 switch (ts->kind) {
3682 case TEMP_TB:
3683 state = ts->state;
3684 ts->state = state | TS_MEM;
3685 if (state != TS_DEAD) {
3686 continue;
3687 }
3688 break;
3689 case TEMP_EBB:
3690 case TEMP_CONST:
3691 continue;
3692 default:
3693 g_assert_not_reached();
3694 }
3695 la_reset_pref(&s->temps[i]);
3696 }
3697 }
3698
3699 /* liveness analysis: sync globals back to memory and kill. */
la_global_kill(TCGContext * s,int ng)3700 static void la_global_kill(TCGContext *s, int ng)
3701 {
3702 int i;
3703
3704 for (i = 0; i < ng; i++) {
3705 s->temps[i].state = TS_DEAD | TS_MEM;
3706 la_reset_pref(&s->temps[i]);
3707 }
3708 }
3709
3710 /* liveness analysis: note live globals crossing calls. */
la_cross_call(TCGContext * s,int nt)3711 static void la_cross_call(TCGContext *s, int nt)
3712 {
3713 TCGRegSet mask = ~tcg_target_call_clobber_regs;
3714 int i;
3715
3716 for (i = 0; i < nt; i++) {
3717 TCGTemp *ts = &s->temps[i];
3718 if (!(ts->state & TS_DEAD)) {
3719 TCGRegSet *pset = la_temp_pref(ts);
3720 TCGRegSet set = *pset;
3721
3722 set &= mask;
3723 /* If the combination is not possible, restart. */
3724 if (set == 0) {
3725 set = tcg_target_available_regs[ts->type] & mask;
3726 }
3727 *pset = set;
3728 }
3729 }
3730 }
3731
3732 /*
3733 * Liveness analysis: Verify the lifetime of TEMP_TB, and reduce
3734 * to TEMP_EBB, if possible.
3735 */
3736 static void __attribute__((noinline))
liveness_pass_0(TCGContext * s)3737 liveness_pass_0(TCGContext *s)
3738 {
3739 void * const multiple_ebb = (void *)(uintptr_t)-1;
3740 int nb_temps = s->nb_temps;
3741 TCGOp *op, *ebb;
3742
3743 for (int i = s->nb_globals; i < nb_temps; ++i) {
3744 s->temps[i].state_ptr = NULL;
3745 }
3746
3747 /*
3748 * Represent each EBB by the op at which it begins. In the case of
3749 * the first EBB, this is the first op, otherwise it is a label.
3750 * Collect the uses of each TEMP_TB: NULL for unused, EBB for use
3751 * within a single EBB, else MULTIPLE_EBB.
3752 */
3753 ebb = QTAILQ_FIRST(&s->ops);
3754 QTAILQ_FOREACH(op, &s->ops, link) {
3755 const TCGOpDef *def;
3756 int nb_oargs, nb_iargs;
3757
3758 switch (op->opc) {
3759 case INDEX_op_set_label:
3760 ebb = op;
3761 continue;
3762 case INDEX_op_discard:
3763 continue;
3764 case INDEX_op_call:
3765 nb_oargs = TCGOP_CALLO(op);
3766 nb_iargs = TCGOP_CALLI(op);
3767 break;
3768 default:
3769 def = &tcg_op_defs[op->opc];
3770 nb_oargs = def->nb_oargs;
3771 nb_iargs = def->nb_iargs;
3772 break;
3773 }
3774
3775 for (int i = 0; i < nb_oargs + nb_iargs; ++i) {
3776 TCGTemp *ts = arg_temp(op->args[i]);
3777
3778 if (ts->kind != TEMP_TB) {
3779 continue;
3780 }
3781 if (ts->state_ptr == NULL) {
3782 ts->state_ptr = ebb;
3783 } else if (ts->state_ptr != ebb) {
3784 ts->state_ptr = multiple_ebb;
3785 }
3786 }
3787 }
3788
3789 /*
3790 * For TEMP_TB that turned out not to be used beyond one EBB,
3791 * reduce the liveness to TEMP_EBB.
3792 */
3793 for (int i = s->nb_globals; i < nb_temps; ++i) {
3794 TCGTemp *ts = &s->temps[i];
3795 if (ts->kind == TEMP_TB && ts->state_ptr != multiple_ebb) {
3796 ts->kind = TEMP_EBB;
3797 }
3798 }
3799 }
3800
3801 /* Liveness analysis : update the opc_arg_life array to tell if a
3802 given input arguments is dead. Instructions updating dead
3803 temporaries are removed. */
3804 static void __attribute__((noinline))
liveness_pass_1(TCGContext * s)3805 liveness_pass_1(TCGContext *s)
3806 {
3807 int nb_globals = s->nb_globals;
3808 int nb_temps = s->nb_temps;
3809 TCGOp *op, *op_prev;
3810 TCGRegSet *prefs;
3811 int i;
3812
3813 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps);
3814 for (i = 0; i < nb_temps; ++i) {
3815 s->temps[i].state_ptr = prefs + i;
3816 }
3817
3818 /* ??? Should be redundant with the exit_tb that ends the TB. */
3819 la_func_end(s, nb_globals, nb_temps);
3820
3821 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) {
3822 int nb_iargs, nb_oargs;
3823 TCGOpcode opc_new, opc_new2;
3824 bool have_opc_new2;
3825 TCGLifeData arg_life = 0;
3826 TCGTemp *ts;
3827 TCGOpcode opc = op->opc;
3828 const TCGOpDef *def = &tcg_op_defs[opc];
3829 const TCGArgConstraint *args_ct;
3830
3831 switch (opc) {
3832 case INDEX_op_call:
3833 {
3834 const TCGHelperInfo *info = tcg_call_info(op);
3835 int call_flags = tcg_call_flags(op);
3836
3837 nb_oargs = TCGOP_CALLO(op);
3838 nb_iargs = TCGOP_CALLI(op);
3839
3840 /* pure functions can be removed if their result is unused */
3841 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) {
3842 for (i = 0; i < nb_oargs; i++) {
3843 ts = arg_temp(op->args[i]);
3844 if (ts->state != TS_DEAD) {
3845 goto do_not_remove_call;
3846 }
3847 }
3848 goto do_remove;
3849 }
3850 do_not_remove_call:
3851
3852 /* Output args are dead. */
3853 for (i = 0; i < nb_oargs; i++) {
3854 ts = arg_temp(op->args[i]);
3855 if (ts->state & TS_DEAD) {
3856 arg_life |= DEAD_ARG << i;
3857 }
3858 if (ts->state & TS_MEM) {
3859 arg_life |= SYNC_ARG << i;
3860 }
3861 ts->state = TS_DEAD;
3862 la_reset_pref(ts);
3863 }
3864
3865 /* Not used -- it will be tcg_target_call_oarg_reg(). */
3866 memset(op->output_pref, 0, sizeof(op->output_pref));
3867
3868 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS |
3869 TCG_CALL_NO_READ_GLOBALS))) {
3870 la_global_kill(s, nb_globals);
3871 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) {
3872 la_global_sync(s, nb_globals);
3873 }
3874
3875 /* Record arguments that die in this helper. */
3876 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
3877 ts = arg_temp(op->args[i]);
3878 if (ts->state & TS_DEAD) {
3879 arg_life |= DEAD_ARG << i;
3880 }
3881 }
3882
3883 /* For all live registers, remove call-clobbered prefs. */
3884 la_cross_call(s, nb_temps);
3885
3886 /*
3887 * Input arguments are live for preceding opcodes.
3888 *
3889 * For those arguments that die, and will be allocated in
3890 * registers, clear the register set for that arg, to be
3891 * filled in below. For args that will be on the stack,
3892 * reset to any available reg. Process arguments in reverse
3893 * order so that if a temp is used more than once, the stack
3894 * reset to max happens before the register reset to 0.
3895 */
3896 for (i = nb_iargs - 1; i >= 0; i--) {
3897 const TCGCallArgumentLoc *loc = &info->in[i];
3898 ts = arg_temp(op->args[nb_oargs + i]);
3899
3900 if (ts->state & TS_DEAD) {
3901 switch (loc->kind) {
3902 case TCG_CALL_ARG_NORMAL:
3903 case TCG_CALL_ARG_EXTEND_U:
3904 case TCG_CALL_ARG_EXTEND_S:
3905 if (arg_slot_reg_p(loc->arg_slot)) {
3906 *la_temp_pref(ts) = 0;
3907 break;
3908 }
3909 /* fall through */
3910 default:
3911 *la_temp_pref(ts) =
3912 tcg_target_available_regs[ts->type];
3913 break;
3914 }
3915 ts->state &= ~TS_DEAD;
3916 }
3917 }
3918
3919 /*
3920 * For each input argument, add its input register to prefs.
3921 * If a temp is used once, this produces a single set bit;
3922 * if a temp is used multiple times, this produces a set.
3923 */
3924 for (i = 0; i < nb_iargs; i++) {
3925 const TCGCallArgumentLoc *loc = &info->in[i];
3926 ts = arg_temp(op->args[nb_oargs + i]);
3927
3928 switch (loc->kind) {
3929 case TCG_CALL_ARG_NORMAL:
3930 case TCG_CALL_ARG_EXTEND_U:
3931 case TCG_CALL_ARG_EXTEND_S:
3932 if (arg_slot_reg_p(loc->arg_slot)) {
3933 tcg_regset_set_reg(*la_temp_pref(ts),
3934 tcg_target_call_iarg_regs[loc->arg_slot]);
3935 }
3936 break;
3937 default:
3938 break;
3939 }
3940 }
3941 }
3942 break;
3943 case INDEX_op_insn_start:
3944 break;
3945 case INDEX_op_discard:
3946 /* mark the temporary as dead */
3947 ts = arg_temp(op->args[0]);
3948 ts->state = TS_DEAD;
3949 la_reset_pref(ts);
3950 break;
3951
3952 case INDEX_op_add2_i32:
3953 opc_new = INDEX_op_add_i32;
3954 goto do_addsub2;
3955 case INDEX_op_sub2_i32:
3956 opc_new = INDEX_op_sub_i32;
3957 goto do_addsub2;
3958 case INDEX_op_add2_i64:
3959 opc_new = INDEX_op_add_i64;
3960 goto do_addsub2;
3961 case INDEX_op_sub2_i64:
3962 opc_new = INDEX_op_sub_i64;
3963 do_addsub2:
3964 nb_iargs = 4;
3965 nb_oargs = 2;
3966 /* Test if the high part of the operation is dead, but not
3967 the low part. The result can be optimized to a simple
3968 add or sub. This happens often for x86_64 guest when the
3969 cpu mode is set to 32 bit. */
3970 if (arg_temp(op->args[1])->state == TS_DEAD) {
3971 if (arg_temp(op->args[0])->state == TS_DEAD) {
3972 goto do_remove;
3973 }
3974 /* Replace the opcode and adjust the args in place,
3975 leaving 3 unused args at the end. */
3976 op->opc = opc = opc_new;
3977 op->args[1] = op->args[2];
3978 op->args[2] = op->args[4];
3979 /* Fall through and mark the single-word operation live. */
3980 nb_iargs = 2;
3981 nb_oargs = 1;
3982 }
3983 goto do_not_remove;
3984
3985 case INDEX_op_mulu2_i32:
3986 opc_new = INDEX_op_mul_i32;
3987 opc_new2 = INDEX_op_muluh_i32;
3988 have_opc_new2 = TCG_TARGET_HAS_muluh_i32;
3989 goto do_mul2;
3990 case INDEX_op_muls2_i32:
3991 opc_new = INDEX_op_mul_i32;
3992 opc_new2 = INDEX_op_mulsh_i32;
3993 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32;
3994 goto do_mul2;
3995 case INDEX_op_mulu2_i64:
3996 opc_new = INDEX_op_mul_i64;
3997 opc_new2 = INDEX_op_muluh_i64;
3998 have_opc_new2 = TCG_TARGET_HAS_muluh_i64;
3999 goto do_mul2;
4000 case INDEX_op_muls2_i64:
4001 opc_new = INDEX_op_mul_i64;
4002 opc_new2 = INDEX_op_mulsh_i64;
4003 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64;
4004 goto do_mul2;
4005 do_mul2:
4006 nb_iargs = 2;
4007 nb_oargs = 2;
4008 if (arg_temp(op->args[1])->state == TS_DEAD) {
4009 if (arg_temp(op->args[0])->state == TS_DEAD) {
4010 /* Both parts of the operation are dead. */
4011 goto do_remove;
4012 }
4013 /* The high part of the operation is dead; generate the low. */
4014 op->opc = opc = opc_new;
4015 op->args[1] = op->args[2];
4016 op->args[2] = op->args[3];
4017 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) {
4018 /* The low part of the operation is dead; generate the high. */
4019 op->opc = opc = opc_new2;
4020 op->args[0] = op->args[1];
4021 op->args[1] = op->args[2];
4022 op->args[2] = op->args[3];
4023 } else {
4024 goto do_not_remove;
4025 }
4026 /* Mark the single-word operation live. */
4027 nb_oargs = 1;
4028 goto do_not_remove;
4029
4030 default:
4031 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */
4032 nb_iargs = def->nb_iargs;
4033 nb_oargs = def->nb_oargs;
4034
4035 /* Test if the operation can be removed because all
4036 its outputs are dead. We assume that nb_oargs == 0
4037 implies side effects */
4038 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) {
4039 for (i = 0; i < nb_oargs; i++) {
4040 if (arg_temp(op->args[i])->state != TS_DEAD) {
4041 goto do_not_remove;
4042 }
4043 }
4044 goto do_remove;
4045 }
4046 goto do_not_remove;
4047
4048 do_remove:
4049 tcg_op_remove(s, op);
4050 break;
4051
4052 do_not_remove:
4053 for (i = 0; i < nb_oargs; i++) {
4054 ts = arg_temp(op->args[i]);
4055
4056 /* Remember the preference of the uses that followed. */
4057 if (i < ARRAY_SIZE(op->output_pref)) {
4058 op->output_pref[i] = *la_temp_pref(ts);
4059 }
4060
4061 /* Output args are dead. */
4062 if (ts->state & TS_DEAD) {
4063 arg_life |= DEAD_ARG << i;
4064 }
4065 if (ts->state & TS_MEM) {
4066 arg_life |= SYNC_ARG << i;
4067 }
4068 ts->state = TS_DEAD;
4069 la_reset_pref(ts);
4070 }
4071
4072 /* If end of basic block, update. */
4073 if (def->flags & TCG_OPF_BB_EXIT) {
4074 la_func_end(s, nb_globals, nb_temps);
4075 } else if (def->flags & TCG_OPF_COND_BRANCH) {
4076 la_bb_sync(s, nb_globals, nb_temps);
4077 } else if (def->flags & TCG_OPF_BB_END) {
4078 la_bb_end(s, nb_globals, nb_temps);
4079 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) {
4080 la_global_sync(s, nb_globals);
4081 if (def->flags & TCG_OPF_CALL_CLOBBER) {
4082 la_cross_call(s, nb_temps);
4083 }
4084 }
4085
4086 /* Record arguments that die in this opcode. */
4087 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
4088 ts = arg_temp(op->args[i]);
4089 if (ts->state & TS_DEAD) {
4090 arg_life |= DEAD_ARG << i;
4091 }
4092 }
4093
4094 /* Input arguments are live for preceding opcodes. */
4095 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
4096 ts = arg_temp(op->args[i]);
4097 if (ts->state & TS_DEAD) {
4098 /* For operands that were dead, initially allow
4099 all regs for the type. */
4100 *la_temp_pref(ts) = tcg_target_available_regs[ts->type];
4101 ts->state &= ~TS_DEAD;
4102 }
4103 }
4104
4105 /* Incorporate constraints for this operand. */
4106 switch (opc) {
4107 case INDEX_op_mov_i32:
4108 case INDEX_op_mov_i64:
4109 /* Note that these are TCG_OPF_NOT_PRESENT and do not
4110 have proper constraints. That said, special case
4111 moves to propagate preferences backward. */
4112 if (IS_DEAD_ARG(1)) {
4113 *la_temp_pref(arg_temp(op->args[0]))
4114 = *la_temp_pref(arg_temp(op->args[1]));
4115 }
4116 break;
4117
4118 default:
4119 args_ct = opcode_args_ct(op);
4120 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
4121 const TCGArgConstraint *ct = &args_ct[i];
4122 TCGRegSet set, *pset;
4123
4124 ts = arg_temp(op->args[i]);
4125 pset = la_temp_pref(ts);
4126 set = *pset;
4127
4128 set &= ct->regs;
4129 if (ct->ialias) {
4130 set &= output_pref(op, ct->alias_index);
4131 }
4132 /* If the combination is not possible, restart. */
4133 if (set == 0) {
4134 set = ct->regs;
4135 }
4136 *pset = set;
4137 }
4138 break;
4139 }
4140 break;
4141 }
4142 op->life = arg_life;
4143 }
4144 }
4145
4146 /* Liveness analysis: Convert indirect regs to direct temporaries. */
4147 static bool __attribute__((noinline))
liveness_pass_2(TCGContext * s)4148 liveness_pass_2(TCGContext *s)
4149 {
4150 int nb_globals = s->nb_globals;
4151 int nb_temps, i;
4152 bool changes = false;
4153 TCGOp *op, *op_next;
4154
4155 /* Create a temporary for each indirect global. */
4156 for (i = 0; i < nb_globals; ++i) {
4157 TCGTemp *its = &s->temps[i];
4158 if (its->indirect_reg) {
4159 TCGTemp *dts = tcg_temp_alloc(s);
4160 dts->type = its->type;
4161 dts->base_type = its->base_type;
4162 dts->temp_subindex = its->temp_subindex;
4163 dts->kind = TEMP_EBB;
4164 its->state_ptr = dts;
4165 } else {
4166 its->state_ptr = NULL;
4167 }
4168 /* All globals begin dead. */
4169 its->state = TS_DEAD;
4170 }
4171 for (nb_temps = s->nb_temps; i < nb_temps; ++i) {
4172 TCGTemp *its = &s->temps[i];
4173 its->state_ptr = NULL;
4174 its->state = TS_DEAD;
4175 }
4176
4177 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
4178 TCGOpcode opc = op->opc;
4179 const TCGOpDef *def = &tcg_op_defs[opc];
4180 TCGLifeData arg_life = op->life;
4181 int nb_iargs, nb_oargs, call_flags;
4182 TCGTemp *arg_ts, *dir_ts;
4183
4184 if (opc == INDEX_op_call) {
4185 nb_oargs = TCGOP_CALLO(op);
4186 nb_iargs = TCGOP_CALLI(op);
4187 call_flags = tcg_call_flags(op);
4188 } else {
4189 nb_iargs = def->nb_iargs;
4190 nb_oargs = def->nb_oargs;
4191
4192 /* Set flags similar to how calls require. */
4193 if (def->flags & TCG_OPF_COND_BRANCH) {
4194 /* Like reading globals: sync_globals */
4195 call_flags = TCG_CALL_NO_WRITE_GLOBALS;
4196 } else if (def->flags & TCG_OPF_BB_END) {
4197 /* Like writing globals: save_globals */
4198 call_flags = 0;
4199 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) {
4200 /* Like reading globals: sync_globals */
4201 call_flags = TCG_CALL_NO_WRITE_GLOBALS;
4202 } else {
4203 /* No effect on globals. */
4204 call_flags = (TCG_CALL_NO_READ_GLOBALS |
4205 TCG_CALL_NO_WRITE_GLOBALS);
4206 }
4207 }
4208
4209 /* Make sure that input arguments are available. */
4210 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
4211 arg_ts = arg_temp(op->args[i]);
4212 dir_ts = arg_ts->state_ptr;
4213 if (dir_ts && arg_ts->state == TS_DEAD) {
4214 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32
4215 ? INDEX_op_ld_i32
4216 : INDEX_op_ld_i64);
4217 TCGOp *lop = tcg_op_insert_before(s, op, lopc, 3);
4218
4219 lop->args[0] = temp_arg(dir_ts);
4220 lop->args[1] = temp_arg(arg_ts->mem_base);
4221 lop->args[2] = arg_ts->mem_offset;
4222
4223 /* Loaded, but synced with memory. */
4224 arg_ts->state = TS_MEM;
4225 }
4226 }
4227
4228 /* Perform input replacement, and mark inputs that became dead.
4229 No action is required except keeping temp_state up to date
4230 so that we reload when needed. */
4231 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
4232 arg_ts = arg_temp(op->args[i]);
4233 dir_ts = arg_ts->state_ptr;
4234 if (dir_ts) {
4235 op->args[i] = temp_arg(dir_ts);
4236 changes = true;
4237 if (IS_DEAD_ARG(i)) {
4238 arg_ts->state = TS_DEAD;
4239 }
4240 }
4241 }
4242
4243 /* Liveness analysis should ensure that the following are
4244 all correct, for call sites and basic block end points. */
4245 if (call_flags & TCG_CALL_NO_READ_GLOBALS) {
4246 /* Nothing to do */
4247 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) {
4248 for (i = 0; i < nb_globals; ++i) {
4249 /* Liveness should see that globals are synced back,
4250 that is, either TS_DEAD or TS_MEM. */
4251 arg_ts = &s->temps[i];
4252 tcg_debug_assert(arg_ts->state_ptr == 0
4253 || arg_ts->state != 0);
4254 }
4255 } else {
4256 for (i = 0; i < nb_globals; ++i) {
4257 /* Liveness should see that globals are saved back,
4258 that is, TS_DEAD, waiting to be reloaded. */
4259 arg_ts = &s->temps[i];
4260 tcg_debug_assert(arg_ts->state_ptr == 0
4261 || arg_ts->state == TS_DEAD);
4262 }
4263 }
4264
4265 /* Outputs become available. */
4266 if (opc == INDEX_op_mov_i32 || opc == INDEX_op_mov_i64) {
4267 arg_ts = arg_temp(op->args[0]);
4268 dir_ts = arg_ts->state_ptr;
4269 if (dir_ts) {
4270 op->args[0] = temp_arg(dir_ts);
4271 changes = true;
4272
4273 /* The output is now live and modified. */
4274 arg_ts->state = 0;
4275
4276 if (NEED_SYNC_ARG(0)) {
4277 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32
4278 ? INDEX_op_st_i32
4279 : INDEX_op_st_i64);
4280 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3);
4281 TCGTemp *out_ts = dir_ts;
4282
4283 if (IS_DEAD_ARG(0)) {
4284 out_ts = arg_temp(op->args[1]);
4285 arg_ts->state = TS_DEAD;
4286 tcg_op_remove(s, op);
4287 } else {
4288 arg_ts->state = TS_MEM;
4289 }
4290
4291 sop->args[0] = temp_arg(out_ts);
4292 sop->args[1] = temp_arg(arg_ts->mem_base);
4293 sop->args[2] = arg_ts->mem_offset;
4294 } else {
4295 tcg_debug_assert(!IS_DEAD_ARG(0));
4296 }
4297 }
4298 } else {
4299 for (i = 0; i < nb_oargs; i++) {
4300 arg_ts = arg_temp(op->args[i]);
4301 dir_ts = arg_ts->state_ptr;
4302 if (!dir_ts) {
4303 continue;
4304 }
4305 op->args[i] = temp_arg(dir_ts);
4306 changes = true;
4307
4308 /* The output is now live and modified. */
4309 arg_ts->state = 0;
4310
4311 /* Sync outputs upon their last write. */
4312 if (NEED_SYNC_ARG(i)) {
4313 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32
4314 ? INDEX_op_st_i32
4315 : INDEX_op_st_i64);
4316 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3);
4317
4318 sop->args[0] = temp_arg(dir_ts);
4319 sop->args[1] = temp_arg(arg_ts->mem_base);
4320 sop->args[2] = arg_ts->mem_offset;
4321
4322 arg_ts->state = TS_MEM;
4323 }
4324 /* Drop outputs that are dead. */
4325 if (IS_DEAD_ARG(i)) {
4326 arg_ts->state = TS_DEAD;
4327 }
4328 }
4329 }
4330 }
4331
4332 return changes;
4333 }
4334
temp_allocate_frame(TCGContext * s,TCGTemp * ts)4335 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts)
4336 {
4337 intptr_t off;
4338 int size, align;
4339
4340 /* When allocating an object, look at the full type. */
4341 size = tcg_type_size(ts->base_type);
4342 switch (ts->base_type) {
4343 case TCG_TYPE_I32:
4344 align = 4;
4345 break;
4346 case TCG_TYPE_I64:
4347 case TCG_TYPE_V64:
4348 align = 8;
4349 break;
4350 case TCG_TYPE_I128:
4351 case TCG_TYPE_V128:
4352 case TCG_TYPE_V256:
4353 /*
4354 * Note that we do not require aligned storage for V256,
4355 * and that we provide alignment for I128 to match V128,
4356 * even if that's above what the host ABI requires.
4357 */
4358 align = 16;
4359 break;
4360 default:
4361 g_assert_not_reached();
4362 }
4363
4364 /*
4365 * Assume the stack is sufficiently aligned.
4366 * This affects e.g. ARM NEON, where we have 8 byte stack alignment
4367 * and do not require 16 byte vector alignment. This seems slightly
4368 * easier than fully parameterizing the above switch statement.
4369 */
4370 align = MIN(TCG_TARGET_STACK_ALIGN, align);
4371 off = ROUND_UP(s->current_frame_offset, align);
4372
4373 /* If we've exhausted the stack frame, restart with a smaller TB. */
4374 if (off + size > s->frame_end) {
4375 tcg_raise_tb_overflow(s);
4376 }
4377 s->current_frame_offset = off + size;
4378 #if defined(__sparc__)
4379 off += TCG_TARGET_STACK_BIAS;
4380 #endif
4381
4382 /* If the object was subdivided, assign memory to all the parts. */
4383 if (ts->base_type != ts->type) {
4384 int part_size = tcg_type_size(ts->type);
4385 int part_count = size / part_size;
4386
4387 /*
4388 * Each part is allocated sequentially in tcg_temp_new_internal.
4389 * Jump back to the first part by subtracting the current index.
4390 */
4391 ts -= ts->temp_subindex;
4392 for (int i = 0; i < part_count; ++i) {
4393 ts[i].mem_offset = off + i * part_size;
4394 ts[i].mem_base = s->frame_temp;
4395 ts[i].mem_allocated = 1;
4396 }
4397 } else {
4398 ts->mem_offset = off;
4399 ts->mem_base = s->frame_temp;
4400 ts->mem_allocated = 1;
4401 }
4402 }
4403
4404 /* Assign @reg to @ts, and update reg_to_temp[]. */
set_temp_val_reg(TCGContext * s,TCGTemp * ts,TCGReg reg)4405 static void set_temp_val_reg(TCGContext *s, TCGTemp *ts, TCGReg reg)
4406 {
4407 if (ts->val_type == TEMP_VAL_REG) {
4408 TCGReg old = ts->reg;
4409 tcg_debug_assert(s->reg_to_temp[old] == ts);
4410 if (old == reg) {
4411 return;
4412 }
4413 s->reg_to_temp[old] = NULL;
4414 }
4415 tcg_debug_assert(s->reg_to_temp[reg] == NULL);
4416 s->reg_to_temp[reg] = ts;
4417 ts->val_type = TEMP_VAL_REG;
4418 ts->reg = reg;
4419 }
4420
4421 /* Assign a non-register value type to @ts, and update reg_to_temp[]. */
set_temp_val_nonreg(TCGContext * s,TCGTemp * ts,TCGTempVal type)4422 static void set_temp_val_nonreg(TCGContext *s, TCGTemp *ts, TCGTempVal type)
4423 {
4424 tcg_debug_assert(type != TEMP_VAL_REG);
4425 if (ts->val_type == TEMP_VAL_REG) {
4426 TCGReg reg = ts->reg;
4427 tcg_debug_assert(s->reg_to_temp[reg] == ts);
4428 s->reg_to_temp[reg] = NULL;
4429 }
4430 ts->val_type = type;
4431 }
4432
4433 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet);
4434
4435 /* Mark a temporary as free or dead. If 'free_or_dead' is negative,
4436 mark it free; otherwise mark it dead. */
temp_free_or_dead(TCGContext * s,TCGTemp * ts,int free_or_dead)4437 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead)
4438 {
4439 TCGTempVal new_type;
4440
4441 switch (ts->kind) {
4442 case TEMP_FIXED:
4443 return;
4444 case TEMP_GLOBAL:
4445 case TEMP_TB:
4446 new_type = TEMP_VAL_MEM;
4447 break;
4448 case TEMP_EBB:
4449 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD;
4450 break;
4451 case TEMP_CONST:
4452 new_type = TEMP_VAL_CONST;
4453 break;
4454 default:
4455 g_assert_not_reached();
4456 }
4457 set_temp_val_nonreg(s, ts, new_type);
4458 }
4459
4460 /* Mark a temporary as dead. */
temp_dead(TCGContext * s,TCGTemp * ts)4461 static inline void temp_dead(TCGContext *s, TCGTemp *ts)
4462 {
4463 temp_free_or_dead(s, ts, 1);
4464 }
4465
4466 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary
4467 registers needs to be allocated to store a constant. If 'free_or_dead'
4468 is non-zero, subsequently release the temporary; if it is positive, the
4469 temp is dead; if it is negative, the temp is free. */
temp_sync(TCGContext * s,TCGTemp * ts,TCGRegSet allocated_regs,TCGRegSet preferred_regs,int free_or_dead)4470 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs,
4471 TCGRegSet preferred_regs, int free_or_dead)
4472 {
4473 if (!temp_readonly(ts) && !ts->mem_coherent) {
4474 if (!ts->mem_allocated) {
4475 temp_allocate_frame(s, ts);
4476 }
4477 switch (ts->val_type) {
4478 case TEMP_VAL_CONST:
4479 /* If we're going to free the temp immediately, then we won't
4480 require it later in a register, so attempt to store the
4481 constant to memory directly. */
4482 if (free_or_dead
4483 && tcg_out_sti(s, ts->type, ts->val,
4484 ts->mem_base->reg, ts->mem_offset)) {
4485 break;
4486 }
4487 temp_load(s, ts, tcg_target_available_regs[ts->type],
4488 allocated_regs, preferred_regs);
4489 /* fallthrough */
4490
4491 case TEMP_VAL_REG:
4492 tcg_out_st(s, ts->type, ts->reg,
4493 ts->mem_base->reg, ts->mem_offset);
4494 break;
4495
4496 case TEMP_VAL_MEM:
4497 break;
4498
4499 case TEMP_VAL_DEAD:
4500 default:
4501 g_assert_not_reached();
4502 }
4503 ts->mem_coherent = 1;
4504 }
4505 if (free_or_dead) {
4506 temp_free_or_dead(s, ts, free_or_dead);
4507 }
4508 }
4509
4510 /* free register 'reg' by spilling the corresponding temporary if necessary */
tcg_reg_free(TCGContext * s,TCGReg reg,TCGRegSet allocated_regs)4511 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs)
4512 {
4513 TCGTemp *ts = s->reg_to_temp[reg];
4514 if (ts != NULL) {
4515 temp_sync(s, ts, allocated_regs, 0, -1);
4516 }
4517 }
4518
4519 /**
4520 * tcg_reg_alloc:
4521 * @required_regs: Set of registers in which we must allocate.
4522 * @allocated_regs: Set of registers which must be avoided.
4523 * @preferred_regs: Set of registers we should prefer.
4524 * @rev: True if we search the registers in "indirect" order.
4525 *
4526 * The allocated register must be in @required_regs & ~@allocated_regs,
4527 * but if we can put it in @preferred_regs we may save a move later.
4528 */
tcg_reg_alloc(TCGContext * s,TCGRegSet required_regs,TCGRegSet allocated_regs,TCGRegSet preferred_regs,bool rev)4529 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs,
4530 TCGRegSet allocated_regs,
4531 TCGRegSet preferred_regs, bool rev)
4532 {
4533 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order);
4534 TCGRegSet reg_ct[2];
4535 const int *order;
4536
4537 reg_ct[1] = required_regs & ~allocated_regs;
4538 tcg_debug_assert(reg_ct[1] != 0);
4539 reg_ct[0] = reg_ct[1] & preferred_regs;
4540
4541 /* Skip the preferred_regs option if it cannot be satisfied,
4542 or if the preference made no difference. */
4543 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1];
4544
4545 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order;
4546
4547 /* Try free registers, preferences first. */
4548 for (j = f; j < 2; j++) {
4549 TCGRegSet set = reg_ct[j];
4550
4551 if (tcg_regset_single(set)) {
4552 /* One register in the set. */
4553 TCGReg reg = tcg_regset_first(set);
4554 if (s->reg_to_temp[reg] == NULL) {
4555 return reg;
4556 }
4557 } else {
4558 for (i = 0; i < n; i++) {
4559 TCGReg reg = order[i];
4560 if (s->reg_to_temp[reg] == NULL &&
4561 tcg_regset_test_reg(set, reg)) {
4562 return reg;
4563 }
4564 }
4565 }
4566 }
4567
4568 /* We must spill something. */
4569 for (j = f; j < 2; j++) {
4570 TCGRegSet set = reg_ct[j];
4571
4572 if (tcg_regset_single(set)) {
4573 /* One register in the set. */
4574 TCGReg reg = tcg_regset_first(set);
4575 tcg_reg_free(s, reg, allocated_regs);
4576 return reg;
4577 } else {
4578 for (i = 0; i < n; i++) {
4579 TCGReg reg = order[i];
4580 if (tcg_regset_test_reg(set, reg)) {
4581 tcg_reg_free(s, reg, allocated_regs);
4582 return reg;
4583 }
4584 }
4585 }
4586 }
4587
4588 g_assert_not_reached();
4589 }
4590
tcg_reg_alloc_pair(TCGContext * s,TCGRegSet required_regs,TCGRegSet allocated_regs,TCGRegSet preferred_regs,bool rev)4591 static TCGReg tcg_reg_alloc_pair(TCGContext *s, TCGRegSet required_regs,
4592 TCGRegSet allocated_regs,
4593 TCGRegSet preferred_regs, bool rev)
4594 {
4595 int i, j, k, fmin, n = ARRAY_SIZE(tcg_target_reg_alloc_order);
4596 TCGRegSet reg_ct[2];
4597 const int *order;
4598
4599 /* Ensure that if I is not in allocated_regs, I+1 is not either. */
4600 reg_ct[1] = required_regs & ~(allocated_regs | (allocated_regs >> 1));
4601 tcg_debug_assert(reg_ct[1] != 0);
4602 reg_ct[0] = reg_ct[1] & preferred_regs;
4603
4604 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order;
4605
4606 /*
4607 * Skip the preferred_regs option if it cannot be satisfied,
4608 * or if the preference made no difference.
4609 */
4610 k = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1];
4611
4612 /*
4613 * Minimize the number of flushes by looking for 2 free registers first,
4614 * then a single flush, then two flushes.
4615 */
4616 for (fmin = 2; fmin >= 0; fmin--) {
4617 for (j = k; j < 2; j++) {
4618 TCGRegSet set = reg_ct[j];
4619
4620 for (i = 0; i < n; i++) {
4621 TCGReg reg = order[i];
4622
4623 if (tcg_regset_test_reg(set, reg)) {
4624 int f = !s->reg_to_temp[reg] + !s->reg_to_temp[reg + 1];
4625 if (f >= fmin) {
4626 tcg_reg_free(s, reg, allocated_regs);
4627 tcg_reg_free(s, reg + 1, allocated_regs);
4628 return reg;
4629 }
4630 }
4631 }
4632 }
4633 }
4634 g_assert_not_reached();
4635 }
4636
4637 /* Make sure the temporary is in a register. If needed, allocate the register
4638 from DESIRED while avoiding ALLOCATED. */
temp_load(TCGContext * s,TCGTemp * ts,TCGRegSet desired_regs,TCGRegSet allocated_regs,TCGRegSet preferred_regs)4639 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs,
4640 TCGRegSet allocated_regs, TCGRegSet preferred_regs)
4641 {
4642 TCGReg reg;
4643
4644 switch (ts->val_type) {
4645 case TEMP_VAL_REG:
4646 return;
4647 case TEMP_VAL_CONST:
4648 reg = tcg_reg_alloc(s, desired_regs, allocated_regs,
4649 preferred_regs, ts->indirect_base);
4650 if (ts->type <= TCG_TYPE_I64) {
4651 tcg_out_movi(s, ts->type, reg, ts->val);
4652 } else {
4653 uint64_t val = ts->val;
4654 MemOp vece = MO_64;
4655
4656 /*
4657 * Find the minimal vector element that matches the constant.
4658 * The targets will, in general, have to do this search anyway,
4659 * do this generically.
4660 */
4661 if (val == dup_const(MO_8, val)) {
4662 vece = MO_8;
4663 } else if (val == dup_const(MO_16, val)) {
4664 vece = MO_16;
4665 } else if (val == dup_const(MO_32, val)) {
4666 vece = MO_32;
4667 }
4668
4669 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val);
4670 }
4671 ts->mem_coherent = 0;
4672 break;
4673 case TEMP_VAL_MEM:
4674 if (!ts->mem_allocated) {
4675 temp_allocate_frame(s, ts);
4676 }
4677 reg = tcg_reg_alloc(s, desired_regs, allocated_regs,
4678 preferred_regs, ts->indirect_base);
4679 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset);
4680 ts->mem_coherent = 1;
4681 break;
4682 case TEMP_VAL_DEAD:
4683 default:
4684 g_assert_not_reached();
4685 }
4686 set_temp_val_reg(s, ts, reg);
4687 }
4688
4689 /* Save a temporary to memory. 'allocated_regs' is used in case a
4690 temporary registers needs to be allocated to store a constant. */
temp_save(TCGContext * s,TCGTemp * ts,TCGRegSet allocated_regs)4691 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs)
4692 {
4693 /* The liveness analysis already ensures that globals are back
4694 in memory. Keep an tcg_debug_assert for safety. */
4695 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts));
4696 }
4697
4698 /* save globals to their canonical location and assume they can be
4699 modified be the following code. 'allocated_regs' is used in case a
4700 temporary registers needs to be allocated to store a constant. */
save_globals(TCGContext * s,TCGRegSet allocated_regs)4701 static void save_globals(TCGContext *s, TCGRegSet allocated_regs)
4702 {
4703 int i, n;
4704
4705 for (i = 0, n = s->nb_globals; i < n; i++) {
4706 temp_save(s, &s->temps[i], allocated_regs);
4707 }
4708 }
4709
4710 /* sync globals to their canonical location and assume they can be
4711 read by the following code. 'allocated_regs' is used in case a
4712 temporary registers needs to be allocated to store a constant. */
sync_globals(TCGContext * s,TCGRegSet allocated_regs)4713 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs)
4714 {
4715 int i, n;
4716
4717 for (i = 0, n = s->nb_globals; i < n; i++) {
4718 TCGTemp *ts = &s->temps[i];
4719 tcg_debug_assert(ts->val_type != TEMP_VAL_REG
4720 || ts->kind == TEMP_FIXED
4721 || ts->mem_coherent);
4722 }
4723 }
4724
4725 /* at the end of a basic block, we assume all temporaries are dead and
4726 all globals are stored at their canonical location. */
tcg_reg_alloc_bb_end(TCGContext * s,TCGRegSet allocated_regs)4727 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs)
4728 {
4729 int i;
4730
4731 for (i = s->nb_globals; i < s->nb_temps; i++) {
4732 TCGTemp *ts = &s->temps[i];
4733
4734 switch (ts->kind) {
4735 case TEMP_TB:
4736 temp_save(s, ts, allocated_regs);
4737 break;
4738 case TEMP_EBB:
4739 /* The liveness analysis already ensures that temps are dead.
4740 Keep an tcg_debug_assert for safety. */
4741 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD);
4742 break;
4743 case TEMP_CONST:
4744 /* Similarly, we should have freed any allocated register. */
4745 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST);
4746 break;
4747 default:
4748 g_assert_not_reached();
4749 }
4750 }
4751
4752 save_globals(s, allocated_regs);
4753 }
4754
4755 /*
4756 * At a conditional branch, we assume all temporaries are dead unless
4757 * explicitly live-across-conditional-branch; all globals and local
4758 * temps are synced to their location.
4759 */
tcg_reg_alloc_cbranch(TCGContext * s,TCGRegSet allocated_regs)4760 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs)
4761 {
4762 sync_globals(s, allocated_regs);
4763
4764 for (int i = s->nb_globals; i < s->nb_temps; i++) {
4765 TCGTemp *ts = &s->temps[i];
4766 /*
4767 * The liveness analysis already ensures that temps are dead.
4768 * Keep tcg_debug_asserts for safety.
4769 */
4770 switch (ts->kind) {
4771 case TEMP_TB:
4772 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent);
4773 break;
4774 case TEMP_EBB:
4775 case TEMP_CONST:
4776 break;
4777 default:
4778 g_assert_not_reached();
4779 }
4780 }
4781 }
4782
4783 /*
4784 * Specialized code generation for INDEX_op_mov_* with a constant.
4785 */
tcg_reg_alloc_do_movi(TCGContext * s,TCGTemp * ots,tcg_target_ulong val,TCGLifeData arg_life,TCGRegSet preferred_regs)4786 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots,
4787 tcg_target_ulong val, TCGLifeData arg_life,
4788 TCGRegSet preferred_regs)
4789 {
4790 /* ENV should not be modified. */
4791 tcg_debug_assert(!temp_readonly(ots));
4792
4793 /* The movi is not explicitly generated here. */
4794 set_temp_val_nonreg(s, ots, TEMP_VAL_CONST);
4795 ots->val = val;
4796 ots->mem_coherent = 0;
4797 if (NEED_SYNC_ARG(0)) {
4798 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0));
4799 } else if (IS_DEAD_ARG(0)) {
4800 temp_dead(s, ots);
4801 }
4802 }
4803
4804 /*
4805 * Specialized code generation for INDEX_op_mov_*.
4806 */
tcg_reg_alloc_mov(TCGContext * s,const TCGOp * op)4807 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op)
4808 {
4809 const TCGLifeData arg_life = op->life;
4810 TCGRegSet allocated_regs, preferred_regs;
4811 TCGTemp *ts, *ots;
4812 TCGType otype, itype;
4813 TCGReg oreg, ireg;
4814
4815 allocated_regs = s->reserved_regs;
4816 preferred_regs = output_pref(op, 0);
4817 ots = arg_temp(op->args[0]);
4818 ts = arg_temp(op->args[1]);
4819
4820 /* ENV should not be modified. */
4821 tcg_debug_assert(!temp_readonly(ots));
4822
4823 /* Note that otype != itype for no-op truncation. */
4824 otype = ots->type;
4825 itype = ts->type;
4826
4827 if (ts->val_type == TEMP_VAL_CONST) {
4828 /* propagate constant or generate sti */
4829 tcg_target_ulong val = ts->val;
4830 if (IS_DEAD_ARG(1)) {
4831 temp_dead(s, ts);
4832 }
4833 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs);
4834 return;
4835 }
4836
4837 /* If the source value is in memory we're going to be forced
4838 to have it in a register in order to perform the copy. Copy
4839 the SOURCE value into its own register first, that way we
4840 don't have to reload SOURCE the next time it is used. */
4841 if (ts->val_type == TEMP_VAL_MEM) {
4842 temp_load(s, ts, tcg_target_available_regs[itype],
4843 allocated_regs, preferred_regs);
4844 }
4845 tcg_debug_assert(ts->val_type == TEMP_VAL_REG);
4846 ireg = ts->reg;
4847
4848 if (IS_DEAD_ARG(0)) {
4849 /* mov to a non-saved dead register makes no sense (even with
4850 liveness analysis disabled). */
4851 tcg_debug_assert(NEED_SYNC_ARG(0));
4852 if (!ots->mem_allocated) {
4853 temp_allocate_frame(s, ots);
4854 }
4855 tcg_out_st(s, otype, ireg, ots->mem_base->reg, ots->mem_offset);
4856 if (IS_DEAD_ARG(1)) {
4857 temp_dead(s, ts);
4858 }
4859 temp_dead(s, ots);
4860 return;
4861 }
4862
4863 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) {
4864 /*
4865 * The mov can be suppressed. Kill input first, so that it
4866 * is unlinked from reg_to_temp, then set the output to the
4867 * reg that we saved from the input.
4868 */
4869 temp_dead(s, ts);
4870 oreg = ireg;
4871 } else {
4872 if (ots->val_type == TEMP_VAL_REG) {
4873 oreg = ots->reg;
4874 } else {
4875 /* Make sure to not spill the input register during allocation. */
4876 oreg = tcg_reg_alloc(s, tcg_target_available_regs[otype],
4877 allocated_regs | ((TCGRegSet)1 << ireg),
4878 preferred_regs, ots->indirect_base);
4879 }
4880 if (!tcg_out_mov(s, otype, oreg, ireg)) {
4881 /*
4882 * Cross register class move not supported.
4883 * Store the source register into the destination slot
4884 * and leave the destination temp as TEMP_VAL_MEM.
4885 */
4886 assert(!temp_readonly(ots));
4887 if (!ts->mem_allocated) {
4888 temp_allocate_frame(s, ots);
4889 }
4890 tcg_out_st(s, ts->type, ireg, ots->mem_base->reg, ots->mem_offset);
4891 set_temp_val_nonreg(s, ts, TEMP_VAL_MEM);
4892 ots->mem_coherent = 1;
4893 return;
4894 }
4895 }
4896 set_temp_val_reg(s, ots, oreg);
4897 ots->mem_coherent = 0;
4898
4899 if (NEED_SYNC_ARG(0)) {
4900 temp_sync(s, ots, allocated_regs, 0, 0);
4901 }
4902 }
4903
4904 /*
4905 * Specialized code generation for INDEX_op_dup_vec.
4906 */
tcg_reg_alloc_dup(TCGContext * s,const TCGOp * op)4907 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op)
4908 {
4909 const TCGLifeData arg_life = op->life;
4910 TCGRegSet dup_out_regs, dup_in_regs;
4911 const TCGArgConstraint *dup_args_ct;
4912 TCGTemp *its, *ots;
4913 TCGType itype, vtype;
4914 unsigned vece;
4915 int lowpart_ofs;
4916 bool ok;
4917
4918 ots = arg_temp(op->args[0]);
4919 its = arg_temp(op->args[1]);
4920
4921 /* ENV should not be modified. */
4922 tcg_debug_assert(!temp_readonly(ots));
4923
4924 itype = its->type;
4925 vece = TCGOP_VECE(op);
4926 vtype = TCGOP_TYPE(op);
4927
4928 if (its->val_type == TEMP_VAL_CONST) {
4929 /* Propagate constant via movi -> dupi. */
4930 tcg_target_ulong val = its->val;
4931 if (IS_DEAD_ARG(1)) {
4932 temp_dead(s, its);
4933 }
4934 tcg_reg_alloc_do_movi(s, ots, val, arg_life, output_pref(op, 0));
4935 return;
4936 }
4937
4938 dup_args_ct = opcode_args_ct(op);
4939 dup_out_regs = dup_args_ct[0].regs;
4940 dup_in_regs = dup_args_ct[1].regs;
4941
4942 /* Allocate the output register now. */
4943 if (ots->val_type != TEMP_VAL_REG) {
4944 TCGRegSet allocated_regs = s->reserved_regs;
4945 TCGReg oreg;
4946
4947 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) {
4948 /* Make sure to not spill the input register. */
4949 tcg_regset_set_reg(allocated_regs, its->reg);
4950 }
4951 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs,
4952 output_pref(op, 0), ots->indirect_base);
4953 set_temp_val_reg(s, ots, oreg);
4954 }
4955
4956 switch (its->val_type) {
4957 case TEMP_VAL_REG:
4958 /*
4959 * The dup constriaints must be broad, covering all possible VECE.
4960 * However, tcg_op_dup_vec() gets to see the VECE and we allow it
4961 * to fail, indicating that extra moves are required for that case.
4962 */
4963 if (tcg_regset_test_reg(dup_in_regs, its->reg)) {
4964 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) {
4965 goto done;
4966 }
4967 /* Try again from memory or a vector input register. */
4968 }
4969 if (!its->mem_coherent) {
4970 /*
4971 * The input register is not synced, and so an extra store
4972 * would be required to use memory. Attempt an integer-vector
4973 * register move first. We do not have a TCGRegSet for this.
4974 */
4975 if (tcg_out_mov(s, itype, ots->reg, its->reg)) {
4976 break;
4977 }
4978 /* Sync the temp back to its slot and load from there. */
4979 temp_sync(s, its, s->reserved_regs, 0, 0);
4980 }
4981 /* fall through */
4982
4983 case TEMP_VAL_MEM:
4984 lowpart_ofs = 0;
4985 if (HOST_BIG_ENDIAN) {
4986 lowpart_ofs = tcg_type_size(itype) - (1 << vece);
4987 }
4988 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg,
4989 its->mem_offset + lowpart_ofs)) {
4990 goto done;
4991 }
4992 /* Load the input into the destination vector register. */
4993 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset);
4994 break;
4995
4996 default:
4997 g_assert_not_reached();
4998 }
4999
5000 /* We now have a vector input register, so dup must succeed. */
5001 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg);
5002 tcg_debug_assert(ok);
5003
5004 done:
5005 ots->mem_coherent = 0;
5006 if (IS_DEAD_ARG(1)) {
5007 temp_dead(s, its);
5008 }
5009 if (NEED_SYNC_ARG(0)) {
5010 temp_sync(s, ots, s->reserved_regs, 0, 0);
5011 }
5012 if (IS_DEAD_ARG(0)) {
5013 temp_dead(s, ots);
5014 }
5015 }
5016
tcg_reg_alloc_op(TCGContext * s,const TCGOp * op)5017 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op)
5018 {
5019 const TCGLifeData arg_life = op->life;
5020 const TCGOpDef * const def = &tcg_op_defs[op->opc];
5021 TCGRegSet i_allocated_regs;
5022 TCGRegSet o_allocated_regs;
5023 int i, k, nb_iargs, nb_oargs;
5024 TCGReg reg;
5025 TCGArg arg;
5026 const TCGArgConstraint *args_ct;
5027 const TCGArgConstraint *arg_ct;
5028 TCGTemp *ts;
5029 TCGArg new_args[TCG_MAX_OP_ARGS];
5030 int const_args[TCG_MAX_OP_ARGS];
5031 TCGCond op_cond;
5032
5033 nb_oargs = def->nb_oargs;
5034 nb_iargs = def->nb_iargs;
5035
5036 /* copy constants */
5037 memcpy(new_args + nb_oargs + nb_iargs,
5038 op->args + nb_oargs + nb_iargs,
5039 sizeof(TCGArg) * def->nb_cargs);
5040
5041 i_allocated_regs = s->reserved_regs;
5042 o_allocated_regs = s->reserved_regs;
5043
5044 switch (op->opc) {
5045 case INDEX_op_brcond_i32:
5046 case INDEX_op_brcond_i64:
5047 op_cond = op->args[2];
5048 break;
5049 case INDEX_op_setcond_i32:
5050 case INDEX_op_setcond_i64:
5051 case INDEX_op_negsetcond_i32:
5052 case INDEX_op_negsetcond_i64:
5053 case INDEX_op_cmp_vec:
5054 op_cond = op->args[3];
5055 break;
5056 case INDEX_op_brcond2_i32:
5057 op_cond = op->args[4];
5058 break;
5059 case INDEX_op_movcond_i32:
5060 case INDEX_op_movcond_i64:
5061 case INDEX_op_setcond2_i32:
5062 case INDEX_op_cmpsel_vec:
5063 op_cond = op->args[5];
5064 break;
5065 default:
5066 /* No condition within opcode. */
5067 op_cond = TCG_COND_ALWAYS;
5068 break;
5069 }
5070
5071 args_ct = opcode_args_ct(op);
5072
5073 /* satisfy input constraints */
5074 for (k = 0; k < nb_iargs; k++) {
5075 TCGRegSet i_preferred_regs, i_required_regs;
5076 bool allocate_new_reg, copyto_new_reg;
5077 TCGTemp *ts2;
5078 int i1, i2;
5079
5080 i = args_ct[nb_oargs + k].sort_index;
5081 arg = op->args[i];
5082 arg_ct = &args_ct[i];
5083 ts = arg_temp(arg);
5084
5085 if (ts->val_type == TEMP_VAL_CONST) {
5086 #ifdef TCG_REG_ZERO
5087 if (ts->val == 0 && (arg_ct->ct & TCG_CT_REG_ZERO)) {
5088 /* Hardware zero register: indicate register via non-const. */
5089 const_args[i] = 0;
5090 new_args[i] = TCG_REG_ZERO;
5091 continue;
5092 }
5093 #endif
5094
5095 if (tcg_target_const_match(ts->val, arg_ct->ct, ts->type,
5096 op_cond, TCGOP_VECE(op))) {
5097 /* constant is OK for instruction */
5098 const_args[i] = 1;
5099 new_args[i] = ts->val;
5100 continue;
5101 }
5102 }
5103
5104 reg = ts->reg;
5105 i_preferred_regs = 0;
5106 i_required_regs = arg_ct->regs;
5107 allocate_new_reg = false;
5108 copyto_new_reg = false;
5109
5110 switch (arg_ct->pair) {
5111 case 0: /* not paired */
5112 if (arg_ct->ialias) {
5113 i_preferred_regs = output_pref(op, arg_ct->alias_index);
5114
5115 /*
5116 * If the input is readonly, then it cannot also be an
5117 * output and aliased to itself. If the input is not
5118 * dead after the instruction, we must allocate a new
5119 * register and move it.
5120 */
5121 if (temp_readonly(ts) || !IS_DEAD_ARG(i)
5122 || args_ct[arg_ct->alias_index].newreg) {
5123 allocate_new_reg = true;
5124 } else if (ts->val_type == TEMP_VAL_REG) {
5125 /*
5126 * Check if the current register has already been
5127 * allocated for another input.
5128 */
5129 allocate_new_reg =
5130 tcg_regset_test_reg(i_allocated_regs, reg);
5131 }
5132 }
5133 if (!allocate_new_reg) {
5134 temp_load(s, ts, i_required_regs, i_allocated_regs,
5135 i_preferred_regs);
5136 reg = ts->reg;
5137 allocate_new_reg = !tcg_regset_test_reg(i_required_regs, reg);
5138 }
5139 if (allocate_new_reg) {
5140 /*
5141 * Allocate a new register matching the constraint
5142 * and move the temporary register into it.
5143 */
5144 temp_load(s, ts, tcg_target_available_regs[ts->type],
5145 i_allocated_regs, 0);
5146 reg = tcg_reg_alloc(s, i_required_regs, i_allocated_regs,
5147 i_preferred_regs, ts->indirect_base);
5148 copyto_new_reg = true;
5149 }
5150 break;
5151
5152 case 1:
5153 /* First of an input pair; if i1 == i2, the second is an output. */
5154 i1 = i;
5155 i2 = arg_ct->pair_index;
5156 ts2 = i1 != i2 ? arg_temp(op->args[i2]) : NULL;
5157
5158 /*
5159 * It is easier to default to allocating a new pair
5160 * and to identify a few cases where it's not required.
5161 */
5162 if (arg_ct->ialias) {
5163 i_preferred_regs = output_pref(op, arg_ct->alias_index);
5164 if (IS_DEAD_ARG(i1) &&
5165 IS_DEAD_ARG(i2) &&
5166 !temp_readonly(ts) &&
5167 ts->val_type == TEMP_VAL_REG &&
5168 ts->reg < TCG_TARGET_NB_REGS - 1 &&
5169 tcg_regset_test_reg(i_required_regs, reg) &&
5170 !tcg_regset_test_reg(i_allocated_regs, reg) &&
5171 !tcg_regset_test_reg(i_allocated_regs, reg + 1) &&
5172 (ts2
5173 ? ts2->val_type == TEMP_VAL_REG &&
5174 ts2->reg == reg + 1 &&
5175 !temp_readonly(ts2)
5176 : s->reg_to_temp[reg + 1] == NULL)) {
5177 break;
5178 }
5179 } else {
5180 /* Without aliasing, the pair must also be an input. */
5181 tcg_debug_assert(ts2);
5182 if (ts->val_type == TEMP_VAL_REG &&
5183 ts2->val_type == TEMP_VAL_REG &&
5184 ts2->reg == reg + 1 &&
5185 tcg_regset_test_reg(i_required_regs, reg)) {
5186 break;
5187 }
5188 }
5189 reg = tcg_reg_alloc_pair(s, i_required_regs, i_allocated_regs,
5190 0, ts->indirect_base);
5191 goto do_pair;
5192
5193 case 2: /* pair second */
5194 reg = new_args[arg_ct->pair_index] + 1;
5195 goto do_pair;
5196
5197 case 3: /* ialias with second output, no first input */
5198 tcg_debug_assert(arg_ct->ialias);
5199 i_preferred_regs = output_pref(op, arg_ct->alias_index);
5200
5201 if (IS_DEAD_ARG(i) &&
5202 !temp_readonly(ts) &&
5203 ts->val_type == TEMP_VAL_REG &&
5204 reg > 0 &&
5205 s->reg_to_temp[reg - 1] == NULL &&
5206 tcg_regset_test_reg(i_required_regs, reg) &&
5207 !tcg_regset_test_reg(i_allocated_regs, reg) &&
5208 !tcg_regset_test_reg(i_allocated_regs, reg - 1)) {
5209 tcg_regset_set_reg(i_allocated_regs, reg - 1);
5210 break;
5211 }
5212 reg = tcg_reg_alloc_pair(s, i_required_regs >> 1,
5213 i_allocated_regs, 0,
5214 ts->indirect_base);
5215 tcg_regset_set_reg(i_allocated_regs, reg);
5216 reg += 1;
5217 goto do_pair;
5218
5219 do_pair:
5220 /*
5221 * If an aliased input is not dead after the instruction,
5222 * we must allocate a new register and move it.
5223 */
5224 if (arg_ct->ialias && (!IS_DEAD_ARG(i) || temp_readonly(ts))) {
5225 TCGRegSet t_allocated_regs = i_allocated_regs;
5226
5227 /*
5228 * Because of the alias, and the continued life, make sure
5229 * that the temp is somewhere *other* than the reg pair,
5230 * and we get a copy in reg.
5231 */
5232 tcg_regset_set_reg(t_allocated_regs, reg);
5233 tcg_regset_set_reg(t_allocated_regs, reg + 1);
5234 if (ts->val_type == TEMP_VAL_REG && ts->reg == reg) {
5235 /* If ts was already in reg, copy it somewhere else. */
5236 TCGReg nr;
5237 bool ok;
5238
5239 tcg_debug_assert(ts->kind != TEMP_FIXED);
5240 nr = tcg_reg_alloc(s, tcg_target_available_regs[ts->type],
5241 t_allocated_regs, 0, ts->indirect_base);
5242 ok = tcg_out_mov(s, ts->type, nr, reg);
5243 tcg_debug_assert(ok);
5244
5245 set_temp_val_reg(s, ts, nr);
5246 } else {
5247 temp_load(s, ts, tcg_target_available_regs[ts->type],
5248 t_allocated_regs, 0);
5249 copyto_new_reg = true;
5250 }
5251 } else {
5252 /* Preferably allocate to reg, otherwise copy. */
5253 i_required_regs = (TCGRegSet)1 << reg;
5254 temp_load(s, ts, i_required_regs, i_allocated_regs,
5255 i_preferred_regs);
5256 copyto_new_reg = ts->reg != reg;
5257 }
5258 break;
5259
5260 default:
5261 g_assert_not_reached();
5262 }
5263
5264 if (copyto_new_reg) {
5265 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) {
5266 /*
5267 * Cross register class move not supported. Sync the
5268 * temp back to its slot and load from there.
5269 */
5270 temp_sync(s, ts, i_allocated_regs, 0, 0);
5271 tcg_out_ld(s, ts->type, reg,
5272 ts->mem_base->reg, ts->mem_offset);
5273 }
5274 }
5275 new_args[i] = reg;
5276 const_args[i] = 0;
5277 tcg_regset_set_reg(i_allocated_regs, reg);
5278 }
5279
5280 /* mark dead temporaries and free the associated registers */
5281 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
5282 if (IS_DEAD_ARG(i)) {
5283 temp_dead(s, arg_temp(op->args[i]));
5284 }
5285 }
5286
5287 if (def->flags & TCG_OPF_COND_BRANCH) {
5288 tcg_reg_alloc_cbranch(s, i_allocated_regs);
5289 } else if (def->flags & TCG_OPF_BB_END) {
5290 tcg_reg_alloc_bb_end(s, i_allocated_regs);
5291 } else {
5292 if (def->flags & TCG_OPF_CALL_CLOBBER) {
5293 /* XXX: permit generic clobber register list ? */
5294 for (i = 0; i < TCG_TARGET_NB_REGS; i++) {
5295 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) {
5296 tcg_reg_free(s, i, i_allocated_regs);
5297 }
5298 }
5299 }
5300 if (def->flags & TCG_OPF_SIDE_EFFECTS) {
5301 /* sync globals if the op has side effects and might trigger
5302 an exception. */
5303 sync_globals(s, i_allocated_regs);
5304 }
5305
5306 /* satisfy the output constraints */
5307 for (k = 0; k < nb_oargs; k++) {
5308 i = args_ct[k].sort_index;
5309 arg = op->args[i];
5310 arg_ct = &args_ct[i];
5311 ts = arg_temp(arg);
5312
5313 /* ENV should not be modified. */
5314 tcg_debug_assert(!temp_readonly(ts));
5315
5316 switch (arg_ct->pair) {
5317 case 0: /* not paired */
5318 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) {
5319 reg = new_args[arg_ct->alias_index];
5320 } else if (arg_ct->newreg) {
5321 reg = tcg_reg_alloc(s, arg_ct->regs,
5322 i_allocated_regs | o_allocated_regs,
5323 output_pref(op, k), ts->indirect_base);
5324 } else {
5325 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs,
5326 output_pref(op, k), ts->indirect_base);
5327 }
5328 break;
5329
5330 case 1: /* first of pair */
5331 if (arg_ct->oalias) {
5332 reg = new_args[arg_ct->alias_index];
5333 } else if (arg_ct->newreg) {
5334 reg = tcg_reg_alloc_pair(s, arg_ct->regs,
5335 i_allocated_regs | o_allocated_regs,
5336 output_pref(op, k),
5337 ts->indirect_base);
5338 } else {
5339 reg = tcg_reg_alloc_pair(s, arg_ct->regs, o_allocated_regs,
5340 output_pref(op, k),
5341 ts->indirect_base);
5342 }
5343 break;
5344
5345 case 2: /* second of pair */
5346 if (arg_ct->oalias) {
5347 reg = new_args[arg_ct->alias_index];
5348 } else {
5349 reg = new_args[arg_ct->pair_index] + 1;
5350 }
5351 break;
5352
5353 case 3: /* first of pair, aliasing with a second input */
5354 tcg_debug_assert(!arg_ct->newreg);
5355 reg = new_args[arg_ct->pair_index] - 1;
5356 break;
5357
5358 default:
5359 g_assert_not_reached();
5360 }
5361 tcg_regset_set_reg(o_allocated_regs, reg);
5362 set_temp_val_reg(s, ts, reg);
5363 ts->mem_coherent = 0;
5364 new_args[i] = reg;
5365 }
5366 }
5367
5368 /* emit instruction */
5369 switch (op->opc) {
5370 case INDEX_op_ext8s_i32:
5371 tcg_out_ext8s(s, TCG_TYPE_I32, new_args[0], new_args[1]);
5372 break;
5373 case INDEX_op_ext8s_i64:
5374 tcg_out_ext8s(s, TCG_TYPE_I64, new_args[0], new_args[1]);
5375 break;
5376 case INDEX_op_ext8u_i32:
5377 case INDEX_op_ext8u_i64:
5378 tcg_out_ext8u(s, new_args[0], new_args[1]);
5379 break;
5380 case INDEX_op_ext16s_i32:
5381 tcg_out_ext16s(s, TCG_TYPE_I32, new_args[0], new_args[1]);
5382 break;
5383 case INDEX_op_ext16s_i64:
5384 tcg_out_ext16s(s, TCG_TYPE_I64, new_args[0], new_args[1]);
5385 break;
5386 case INDEX_op_ext16u_i32:
5387 case INDEX_op_ext16u_i64:
5388 tcg_out_ext16u(s, new_args[0], new_args[1]);
5389 break;
5390 case INDEX_op_ext32s_i64:
5391 tcg_out_ext32s(s, new_args[0], new_args[1]);
5392 break;
5393 case INDEX_op_ext32u_i64:
5394 tcg_out_ext32u(s, new_args[0], new_args[1]);
5395 break;
5396 case INDEX_op_ext_i32_i64:
5397 tcg_out_exts_i32_i64(s, new_args[0], new_args[1]);
5398 break;
5399 case INDEX_op_extu_i32_i64:
5400 tcg_out_extu_i32_i64(s, new_args[0], new_args[1]);
5401 break;
5402 case INDEX_op_extrl_i64_i32:
5403 tcg_out_extrl_i64_i32(s, new_args[0], new_args[1]);
5404 break;
5405 default:
5406 if (def->flags & TCG_OPF_VECTOR) {
5407 tcg_out_vec_op(s, op->opc, TCGOP_TYPE(op) - TCG_TYPE_V64,
5408 TCGOP_VECE(op), new_args, const_args);
5409 } else {
5410 tcg_out_op(s, op->opc, TCGOP_TYPE(op), new_args, const_args);
5411 }
5412 break;
5413 }
5414
5415 /* move the outputs in the correct register if needed */
5416 for(i = 0; i < nb_oargs; i++) {
5417 ts = arg_temp(op->args[i]);
5418
5419 /* ENV should not be modified. */
5420 tcg_debug_assert(!temp_readonly(ts));
5421
5422 if (NEED_SYNC_ARG(i)) {
5423 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i));
5424 } else if (IS_DEAD_ARG(i)) {
5425 temp_dead(s, ts);
5426 }
5427 }
5428 }
5429
tcg_reg_alloc_dup2(TCGContext * s,const TCGOp * op)5430 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op)
5431 {
5432 const TCGLifeData arg_life = op->life;
5433 TCGTemp *ots, *itsl, *itsh;
5434 TCGType vtype = TCGOP_TYPE(op);
5435
5436 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */
5437 tcg_debug_assert(TCG_TARGET_REG_BITS == 32);
5438 tcg_debug_assert(TCGOP_VECE(op) == MO_64);
5439
5440 ots = arg_temp(op->args[0]);
5441 itsl = arg_temp(op->args[1]);
5442 itsh = arg_temp(op->args[2]);
5443
5444 /* ENV should not be modified. */
5445 tcg_debug_assert(!temp_readonly(ots));
5446
5447 /* Allocate the output register now. */
5448 if (ots->val_type != TEMP_VAL_REG) {
5449 TCGRegSet allocated_regs = s->reserved_regs;
5450 TCGRegSet dup_out_regs = opcode_args_ct(op)[0].regs;
5451 TCGReg oreg;
5452
5453 /* Make sure to not spill the input registers. */
5454 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) {
5455 tcg_regset_set_reg(allocated_regs, itsl->reg);
5456 }
5457 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) {
5458 tcg_regset_set_reg(allocated_regs, itsh->reg);
5459 }
5460
5461 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs,
5462 output_pref(op, 0), ots->indirect_base);
5463 set_temp_val_reg(s, ots, oreg);
5464 }
5465
5466 /* Promote dup2 of immediates to dupi_vec. */
5467 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) {
5468 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val);
5469 MemOp vece = MO_64;
5470
5471 if (val == dup_const(MO_8, val)) {
5472 vece = MO_8;
5473 } else if (val == dup_const(MO_16, val)) {
5474 vece = MO_16;
5475 } else if (val == dup_const(MO_32, val)) {
5476 vece = MO_32;
5477 }
5478
5479 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val);
5480 goto done;
5481 }
5482
5483 /* If the two inputs form one 64-bit value, try dupm_vec. */
5484 if (itsl->temp_subindex == HOST_BIG_ENDIAN &&
5485 itsh->temp_subindex == !HOST_BIG_ENDIAN &&
5486 itsl == itsh + (HOST_BIG_ENDIAN ? 1 : -1)) {
5487 TCGTemp *its = itsl - HOST_BIG_ENDIAN;
5488
5489 temp_sync(s, its + 0, s->reserved_regs, 0, 0);
5490 temp_sync(s, its + 1, s->reserved_regs, 0, 0);
5491
5492 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg,
5493 its->mem_base->reg, its->mem_offset)) {
5494 goto done;
5495 }
5496 }
5497
5498 /* Fall back to generic expansion. */
5499 return false;
5500
5501 done:
5502 ots->mem_coherent = 0;
5503 if (IS_DEAD_ARG(1)) {
5504 temp_dead(s, itsl);
5505 }
5506 if (IS_DEAD_ARG(2)) {
5507 temp_dead(s, itsh);
5508 }
5509 if (NEED_SYNC_ARG(0)) {
5510 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0));
5511 } else if (IS_DEAD_ARG(0)) {
5512 temp_dead(s, ots);
5513 }
5514 return true;
5515 }
5516
load_arg_reg(TCGContext * s,TCGReg reg,TCGTemp * ts,TCGRegSet allocated_regs)5517 static void load_arg_reg(TCGContext *s, TCGReg reg, TCGTemp *ts,
5518 TCGRegSet allocated_regs)
5519 {
5520 if (ts->val_type == TEMP_VAL_REG) {
5521 if (ts->reg != reg) {
5522 tcg_reg_free(s, reg, allocated_regs);
5523 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) {
5524 /*
5525 * Cross register class move not supported. Sync the
5526 * temp back to its slot and load from there.
5527 */
5528 temp_sync(s, ts, allocated_regs, 0, 0);
5529 tcg_out_ld(s, ts->type, reg,
5530 ts->mem_base->reg, ts->mem_offset);
5531 }
5532 }
5533 } else {
5534 TCGRegSet arg_set = 0;
5535
5536 tcg_reg_free(s, reg, allocated_regs);
5537 tcg_regset_set_reg(arg_set, reg);
5538 temp_load(s, ts, arg_set, allocated_regs, 0);
5539 }
5540 }
5541
load_arg_stk(TCGContext * s,unsigned arg_slot,TCGTemp * ts,TCGRegSet allocated_regs)5542 static void load_arg_stk(TCGContext *s, unsigned arg_slot, TCGTemp *ts,
5543 TCGRegSet allocated_regs)
5544 {
5545 /*
5546 * When the destination is on the stack, load up the temp and store.
5547 * If there are many call-saved registers, the temp might live to
5548 * see another use; otherwise it'll be discarded.
5549 */
5550 temp_load(s, ts, tcg_target_available_regs[ts->type], allocated_regs, 0);
5551 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK,
5552 arg_slot_stk_ofs(arg_slot));
5553 }
5554
load_arg_normal(TCGContext * s,const TCGCallArgumentLoc * l,TCGTemp * ts,TCGRegSet * allocated_regs)5555 static void load_arg_normal(TCGContext *s, const TCGCallArgumentLoc *l,
5556 TCGTemp *ts, TCGRegSet *allocated_regs)
5557 {
5558 if (arg_slot_reg_p(l->arg_slot)) {
5559 TCGReg reg = tcg_target_call_iarg_regs[l->arg_slot];
5560 load_arg_reg(s, reg, ts, *allocated_regs);
5561 tcg_regset_set_reg(*allocated_regs, reg);
5562 } else {
5563 load_arg_stk(s, l->arg_slot, ts, *allocated_regs);
5564 }
5565 }
5566
load_arg_ref(TCGContext * s,unsigned arg_slot,TCGReg ref_base,intptr_t ref_off,TCGRegSet * allocated_regs)5567 static void load_arg_ref(TCGContext *s, unsigned arg_slot, TCGReg ref_base,
5568 intptr_t ref_off, TCGRegSet *allocated_regs)
5569 {
5570 TCGReg reg;
5571
5572 if (arg_slot_reg_p(arg_slot)) {
5573 reg = tcg_target_call_iarg_regs[arg_slot];
5574 tcg_reg_free(s, reg, *allocated_regs);
5575 tcg_out_addi_ptr(s, reg, ref_base, ref_off);
5576 tcg_regset_set_reg(*allocated_regs, reg);
5577 } else {
5578 reg = tcg_reg_alloc(s, tcg_target_available_regs[TCG_TYPE_PTR],
5579 *allocated_regs, 0, false);
5580 tcg_out_addi_ptr(s, reg, ref_base, ref_off);
5581 tcg_out_st(s, TCG_TYPE_PTR, reg, TCG_REG_CALL_STACK,
5582 arg_slot_stk_ofs(arg_slot));
5583 }
5584 }
5585
tcg_reg_alloc_call(TCGContext * s,TCGOp * op)5586 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op)
5587 {
5588 const int nb_oargs = TCGOP_CALLO(op);
5589 const int nb_iargs = TCGOP_CALLI(op);
5590 const TCGLifeData arg_life = op->life;
5591 const TCGHelperInfo *info = tcg_call_info(op);
5592 TCGRegSet allocated_regs = s->reserved_regs;
5593 int i;
5594
5595 /*
5596 * Move inputs into place in reverse order,
5597 * so that we place stacked arguments first.
5598 */
5599 for (i = nb_iargs - 1; i >= 0; --i) {
5600 const TCGCallArgumentLoc *loc = &info->in[i];
5601 TCGTemp *ts = arg_temp(op->args[nb_oargs + i]);
5602
5603 switch (loc->kind) {
5604 case TCG_CALL_ARG_NORMAL:
5605 case TCG_CALL_ARG_EXTEND_U:
5606 case TCG_CALL_ARG_EXTEND_S:
5607 load_arg_normal(s, loc, ts, &allocated_regs);
5608 break;
5609 case TCG_CALL_ARG_BY_REF:
5610 load_arg_stk(s, loc->ref_slot, ts, allocated_regs);
5611 load_arg_ref(s, loc->arg_slot, TCG_REG_CALL_STACK,
5612 arg_slot_stk_ofs(loc->ref_slot),
5613 &allocated_regs);
5614 break;
5615 case TCG_CALL_ARG_BY_REF_N:
5616 load_arg_stk(s, loc->ref_slot, ts, allocated_regs);
5617 break;
5618 default:
5619 g_assert_not_reached();
5620 }
5621 }
5622
5623 /* Mark dead temporaries and free the associated registers. */
5624 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) {
5625 if (IS_DEAD_ARG(i)) {
5626 temp_dead(s, arg_temp(op->args[i]));
5627 }
5628 }
5629
5630 /* Clobber call registers. */
5631 for (i = 0; i < TCG_TARGET_NB_REGS; i++) {
5632 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) {
5633 tcg_reg_free(s, i, allocated_regs);
5634 }
5635 }
5636
5637 /*
5638 * Save globals if they might be written by the helper,
5639 * sync them if they might be read.
5640 */
5641 if (info->flags & TCG_CALL_NO_READ_GLOBALS) {
5642 /* Nothing to do */
5643 } else if (info->flags & TCG_CALL_NO_WRITE_GLOBALS) {
5644 sync_globals(s, allocated_regs);
5645 } else {
5646 save_globals(s, allocated_regs);
5647 }
5648
5649 /*
5650 * If the ABI passes a pointer to the returned struct as the first
5651 * argument, load that now. Pass a pointer to the output home slot.
5652 */
5653 if (info->out_kind == TCG_CALL_RET_BY_REF) {
5654 TCGTemp *ts = arg_temp(op->args[0]);
5655
5656 if (!ts->mem_allocated) {
5657 temp_allocate_frame(s, ts);
5658 }
5659 load_arg_ref(s, 0, ts->mem_base->reg, ts->mem_offset, &allocated_regs);
5660 }
5661
5662 tcg_out_call(s, tcg_call_func(op), info);
5663
5664 /* Assign output registers and emit moves if needed. */
5665 switch (info->out_kind) {
5666 case TCG_CALL_RET_NORMAL:
5667 for (i = 0; i < nb_oargs; i++) {
5668 TCGTemp *ts = arg_temp(op->args[i]);
5669 TCGReg reg = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, i);
5670
5671 /* ENV should not be modified. */
5672 tcg_debug_assert(!temp_readonly(ts));
5673
5674 set_temp_val_reg(s, ts, reg);
5675 ts->mem_coherent = 0;
5676 }
5677 break;
5678
5679 case TCG_CALL_RET_BY_VEC:
5680 {
5681 TCGTemp *ts = arg_temp(op->args[0]);
5682
5683 tcg_debug_assert(ts->base_type == TCG_TYPE_I128);
5684 tcg_debug_assert(ts->temp_subindex == 0);
5685 if (!ts->mem_allocated) {
5686 temp_allocate_frame(s, ts);
5687 }
5688 tcg_out_st(s, TCG_TYPE_V128,
5689 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0),
5690 ts->mem_base->reg, ts->mem_offset);
5691 }
5692 /* fall through to mark all parts in memory */
5693
5694 case TCG_CALL_RET_BY_REF:
5695 /* The callee has performed a write through the reference. */
5696 for (i = 0; i < nb_oargs; i++) {
5697 TCGTemp *ts = arg_temp(op->args[i]);
5698 ts->val_type = TEMP_VAL_MEM;
5699 }
5700 break;
5701
5702 default:
5703 g_assert_not_reached();
5704 }
5705
5706 /* Flush or discard output registers as needed. */
5707 for (i = 0; i < nb_oargs; i++) {
5708 TCGTemp *ts = arg_temp(op->args[i]);
5709 if (NEED_SYNC_ARG(i)) {
5710 temp_sync(s, ts, s->reserved_regs, 0, IS_DEAD_ARG(i));
5711 } else if (IS_DEAD_ARG(i)) {
5712 temp_dead(s, ts);
5713 }
5714 }
5715 }
5716
5717 /**
5718 * atom_and_align_for_opc:
5719 * @s: tcg context
5720 * @opc: memory operation code
5721 * @host_atom: MO_ATOM_{IFALIGN,WITHIN16,SUBALIGN} for host operations
5722 * @allow_two_ops: true if we are prepared to issue two operations
5723 *
5724 * Return the alignment and atomicity to use for the inline fast path
5725 * for the given memory operation. The alignment may be larger than
5726 * that specified in @opc, and the correct alignment will be diagnosed
5727 * by the slow path helper.
5728 *
5729 * If @allow_two_ops, the host is prepared to test for 2x alignment,
5730 * and issue two loads or stores for subalignment.
5731 */
atom_and_align_for_opc(TCGContext * s,MemOp opc,MemOp host_atom,bool allow_two_ops)5732 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc,
5733 MemOp host_atom, bool allow_two_ops)
5734 {
5735 MemOp align = memop_alignment_bits(opc);
5736 MemOp size = opc & MO_SIZE;
5737 MemOp half = size ? size - 1 : 0;
5738 MemOp atom = opc & MO_ATOM_MASK;
5739 MemOp atmax;
5740
5741 switch (atom) {
5742 case MO_ATOM_NONE:
5743 /* The operation requires no specific atomicity. */
5744 atmax = MO_8;
5745 break;
5746
5747 case MO_ATOM_IFALIGN:
5748 atmax = size;
5749 break;
5750
5751 case MO_ATOM_IFALIGN_PAIR:
5752 atmax = half;
5753 break;
5754
5755 case MO_ATOM_WITHIN16:
5756 atmax = size;
5757 if (size == MO_128) {
5758 /* Misalignment implies !within16, and therefore no atomicity. */
5759 } else if (host_atom != MO_ATOM_WITHIN16) {
5760 /* The host does not implement within16, so require alignment. */
5761 align = MAX(align, size);
5762 }
5763 break;
5764
5765 case MO_ATOM_WITHIN16_PAIR:
5766 atmax = size;
5767 /*
5768 * Misalignment implies !within16, and therefore half atomicity.
5769 * Any host prepared for two operations can implement this with
5770 * half alignment.
5771 */
5772 if (host_atom != MO_ATOM_WITHIN16 && allow_two_ops) {
5773 align = MAX(align, half);
5774 }
5775 break;
5776
5777 case MO_ATOM_SUBALIGN:
5778 atmax = size;
5779 if (host_atom != MO_ATOM_SUBALIGN) {
5780 /* If unaligned but not odd, there are subobjects up to half. */
5781 if (allow_two_ops) {
5782 align = MAX(align, half);
5783 } else {
5784 align = MAX(align, size);
5785 }
5786 }
5787 break;
5788
5789 default:
5790 g_assert_not_reached();
5791 }
5792
5793 return (TCGAtomAlign){ .atom = atmax, .align = align };
5794 }
5795
5796 /*
5797 * Similarly for qemu_ld/st slow path helpers.
5798 * We must re-implement tcg_gen_callN and tcg_reg_alloc_call simultaneously,
5799 * using only the provided backend tcg_out_* functions.
5800 */
5801
tcg_out_helper_stk_ofs(TCGType type,unsigned slot)5802 static int tcg_out_helper_stk_ofs(TCGType type, unsigned slot)
5803 {
5804 int ofs = arg_slot_stk_ofs(slot);
5805
5806 /*
5807 * Each stack slot is TCG_TARGET_LONG_BITS. If the host does not
5808 * require extension to uint64_t, adjust the address for uint32_t.
5809 */
5810 if (HOST_BIG_ENDIAN &&
5811 TCG_TARGET_REG_BITS == 64 &&
5812 type == TCG_TYPE_I32) {
5813 ofs += 4;
5814 }
5815 return ofs;
5816 }
5817
tcg_out_helper_load_slots(TCGContext * s,unsigned nmov,TCGMovExtend * mov,const TCGLdstHelperParam * parm)5818 static void tcg_out_helper_load_slots(TCGContext *s,
5819 unsigned nmov, TCGMovExtend *mov,
5820 const TCGLdstHelperParam *parm)
5821 {
5822 unsigned i;
5823 TCGReg dst3;
5824
5825 /*
5826 * Start from the end, storing to the stack first.
5827 * This frees those registers, so we need not consider overlap.
5828 */
5829 for (i = nmov; i-- > 0; ) {
5830 unsigned slot = mov[i].dst;
5831
5832 if (arg_slot_reg_p(slot)) {
5833 goto found_reg;
5834 }
5835
5836 TCGReg src = mov[i].src;
5837 TCGType dst_type = mov[i].dst_type;
5838 MemOp dst_mo = dst_type == TCG_TYPE_I32 ? MO_32 : MO_64;
5839
5840 /* The argument is going onto the stack; extend into scratch. */
5841 if ((mov[i].src_ext & MO_SIZE) != dst_mo) {
5842 tcg_debug_assert(parm->ntmp != 0);
5843 mov[i].dst = src = parm->tmp[0];
5844 tcg_out_movext1(s, &mov[i]);
5845 }
5846
5847 tcg_out_st(s, dst_type, src, TCG_REG_CALL_STACK,
5848 tcg_out_helper_stk_ofs(dst_type, slot));
5849 }
5850 return;
5851
5852 found_reg:
5853 /*
5854 * The remaining arguments are in registers.
5855 * Convert slot numbers to argument registers.
5856 */
5857 nmov = i + 1;
5858 for (i = 0; i < nmov; ++i) {
5859 mov[i].dst = tcg_target_call_iarg_regs[mov[i].dst];
5860 }
5861
5862 switch (nmov) {
5863 case 4:
5864 /* The backend must have provided enough temps for the worst case. */
5865 tcg_debug_assert(parm->ntmp >= 2);
5866
5867 dst3 = mov[3].dst;
5868 for (unsigned j = 0; j < 3; ++j) {
5869 if (dst3 == mov[j].src) {
5870 /*
5871 * Conflict. Copy the source to a temporary, perform the
5872 * remaining moves, then the extension from our scratch
5873 * on the way out.
5874 */
5875 TCGReg scratch = parm->tmp[1];
5876
5877 tcg_out_mov(s, mov[3].src_type, scratch, mov[3].src);
5878 tcg_out_movext3(s, mov, mov + 1, mov + 2, parm->tmp[0]);
5879 tcg_out_movext1_new_src(s, &mov[3], scratch);
5880 break;
5881 }
5882 }
5883
5884 /* No conflicts: perform this move and continue. */
5885 tcg_out_movext1(s, &mov[3]);
5886 /* fall through */
5887
5888 case 3:
5889 tcg_out_movext3(s, mov, mov + 1, mov + 2,
5890 parm->ntmp ? parm->tmp[0] : -1);
5891 break;
5892 case 2:
5893 tcg_out_movext2(s, mov, mov + 1,
5894 parm->ntmp ? parm->tmp[0] : -1);
5895 break;
5896 case 1:
5897 tcg_out_movext1(s, mov);
5898 break;
5899 default:
5900 g_assert_not_reached();
5901 }
5902 }
5903
tcg_out_helper_load_imm(TCGContext * s,unsigned slot,TCGType type,tcg_target_long imm,const TCGLdstHelperParam * parm)5904 static void tcg_out_helper_load_imm(TCGContext *s, unsigned slot,
5905 TCGType type, tcg_target_long imm,
5906 const TCGLdstHelperParam *parm)
5907 {
5908 if (arg_slot_reg_p(slot)) {
5909 tcg_out_movi(s, type, tcg_target_call_iarg_regs[slot], imm);
5910 } else {
5911 int ofs = tcg_out_helper_stk_ofs(type, slot);
5912 if (!tcg_out_sti(s, type, imm, TCG_REG_CALL_STACK, ofs)) {
5913 tcg_debug_assert(parm->ntmp != 0);
5914 tcg_out_movi(s, type, parm->tmp[0], imm);
5915 tcg_out_st(s, type, parm->tmp[0], TCG_REG_CALL_STACK, ofs);
5916 }
5917 }
5918 }
5919
tcg_out_helper_load_common_args(TCGContext * s,const TCGLabelQemuLdst * ldst,const TCGLdstHelperParam * parm,const TCGHelperInfo * info,unsigned next_arg)5920 static void tcg_out_helper_load_common_args(TCGContext *s,
5921 const TCGLabelQemuLdst *ldst,
5922 const TCGLdstHelperParam *parm,
5923 const TCGHelperInfo *info,
5924 unsigned next_arg)
5925 {
5926 TCGMovExtend ptr_mov = {
5927 .dst_type = TCG_TYPE_PTR,
5928 .src_type = TCG_TYPE_PTR,
5929 .src_ext = sizeof(void *) == 4 ? MO_32 : MO_64
5930 };
5931 const TCGCallArgumentLoc *loc = &info->in[0];
5932 TCGType type;
5933 unsigned slot;
5934 tcg_target_ulong imm;
5935
5936 /*
5937 * Handle env, which is always first.
5938 */
5939 ptr_mov.dst = loc->arg_slot;
5940 ptr_mov.src = TCG_AREG0;
5941 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm);
5942
5943 /*
5944 * Handle oi.
5945 */
5946 imm = ldst->oi;
5947 loc = &info->in[next_arg];
5948 type = TCG_TYPE_I32;
5949 switch (loc->kind) {
5950 case TCG_CALL_ARG_NORMAL:
5951 break;
5952 case TCG_CALL_ARG_EXTEND_U:
5953 case TCG_CALL_ARG_EXTEND_S:
5954 /* No extension required for MemOpIdx. */
5955 tcg_debug_assert(imm <= INT32_MAX);
5956 type = TCG_TYPE_REG;
5957 break;
5958 default:
5959 g_assert_not_reached();
5960 }
5961 tcg_out_helper_load_imm(s, loc->arg_slot, type, imm, parm);
5962 next_arg++;
5963
5964 /*
5965 * Handle ra.
5966 */
5967 loc = &info->in[next_arg];
5968 slot = loc->arg_slot;
5969 if (parm->ra_gen) {
5970 int arg_reg = -1;
5971 TCGReg ra_reg;
5972
5973 if (arg_slot_reg_p(slot)) {
5974 arg_reg = tcg_target_call_iarg_regs[slot];
5975 }
5976 ra_reg = parm->ra_gen(s, ldst, arg_reg);
5977
5978 ptr_mov.dst = slot;
5979 ptr_mov.src = ra_reg;
5980 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm);
5981 } else {
5982 imm = (uintptr_t)ldst->raddr;
5983 tcg_out_helper_load_imm(s, slot, TCG_TYPE_PTR, imm, parm);
5984 }
5985 }
5986
tcg_out_helper_add_mov(TCGMovExtend * mov,const TCGCallArgumentLoc * loc,TCGType dst_type,TCGType src_type,TCGReg lo,TCGReg hi)5987 static unsigned tcg_out_helper_add_mov(TCGMovExtend *mov,
5988 const TCGCallArgumentLoc *loc,
5989 TCGType dst_type, TCGType src_type,
5990 TCGReg lo, TCGReg hi)
5991 {
5992 MemOp reg_mo;
5993
5994 if (dst_type <= TCG_TYPE_REG) {
5995 MemOp src_ext;
5996
5997 switch (loc->kind) {
5998 case TCG_CALL_ARG_NORMAL:
5999 src_ext = src_type == TCG_TYPE_I32 ? MO_32 : MO_64;
6000 break;
6001 case TCG_CALL_ARG_EXTEND_U:
6002 dst_type = TCG_TYPE_REG;
6003 src_ext = MO_UL;
6004 break;
6005 case TCG_CALL_ARG_EXTEND_S:
6006 dst_type = TCG_TYPE_REG;
6007 src_ext = MO_SL;
6008 break;
6009 default:
6010 g_assert_not_reached();
6011 }
6012
6013 mov[0].dst = loc->arg_slot;
6014 mov[0].dst_type = dst_type;
6015 mov[0].src = lo;
6016 mov[0].src_type = src_type;
6017 mov[0].src_ext = src_ext;
6018 return 1;
6019 }
6020
6021 if (TCG_TARGET_REG_BITS == 32) {
6022 assert(dst_type == TCG_TYPE_I64);
6023 reg_mo = MO_32;
6024 } else {
6025 assert(dst_type == TCG_TYPE_I128);
6026 reg_mo = MO_64;
6027 }
6028
6029 mov[0].dst = loc[HOST_BIG_ENDIAN].arg_slot;
6030 mov[0].src = lo;
6031 mov[0].dst_type = TCG_TYPE_REG;
6032 mov[0].src_type = TCG_TYPE_REG;
6033 mov[0].src_ext = reg_mo;
6034
6035 mov[1].dst = loc[!HOST_BIG_ENDIAN].arg_slot;
6036 mov[1].src = hi;
6037 mov[1].dst_type = TCG_TYPE_REG;
6038 mov[1].src_type = TCG_TYPE_REG;
6039 mov[1].src_ext = reg_mo;
6040
6041 return 2;
6042 }
6043
tcg_out_ld_helper_args(TCGContext * s,const TCGLabelQemuLdst * ldst,const TCGLdstHelperParam * parm)6044 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst,
6045 const TCGLdstHelperParam *parm)
6046 {
6047 const TCGHelperInfo *info;
6048 const TCGCallArgumentLoc *loc;
6049 TCGMovExtend mov[2];
6050 unsigned next_arg, nmov;
6051 MemOp mop = get_memop(ldst->oi);
6052
6053 switch (mop & MO_SIZE) {
6054 case MO_8:
6055 case MO_16:
6056 case MO_32:
6057 info = &info_helper_ld32_mmu;
6058 break;
6059 case MO_64:
6060 info = &info_helper_ld64_mmu;
6061 break;
6062 case MO_128:
6063 info = &info_helper_ld128_mmu;
6064 break;
6065 default:
6066 g_assert_not_reached();
6067 }
6068
6069 /* Defer env argument. */
6070 next_arg = 1;
6071
6072 loc = &info->in[next_arg];
6073 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) {
6074 /*
6075 * 32-bit host with 32-bit guest: zero-extend the guest address
6076 * to 64-bits for the helper by storing the low part, then
6077 * load a zero for the high part.
6078 */
6079 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN,
6080 TCG_TYPE_I32, TCG_TYPE_I32,
6081 ldst->addr_reg, -1);
6082 tcg_out_helper_load_slots(s, 1, mov, parm);
6083
6084 tcg_out_helper_load_imm(s, loc[!HOST_BIG_ENDIAN].arg_slot,
6085 TCG_TYPE_I32, 0, parm);
6086 next_arg += 2;
6087 } else {
6088 nmov = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type,
6089 ldst->addr_reg, -1);
6090 tcg_out_helper_load_slots(s, nmov, mov, parm);
6091 next_arg += nmov;
6092 }
6093
6094 switch (info->out_kind) {
6095 case TCG_CALL_RET_NORMAL:
6096 case TCG_CALL_RET_BY_VEC:
6097 break;
6098 case TCG_CALL_RET_BY_REF:
6099 /*
6100 * The return reference is in the first argument slot.
6101 * We need memory in which to return: re-use the top of stack.
6102 */
6103 {
6104 int ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET;
6105
6106 if (arg_slot_reg_p(0)) {
6107 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[0],
6108 TCG_REG_CALL_STACK, ofs_slot0);
6109 } else {
6110 tcg_debug_assert(parm->ntmp != 0);
6111 tcg_out_addi_ptr(s, parm->tmp[0],
6112 TCG_REG_CALL_STACK, ofs_slot0);
6113 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0],
6114 TCG_REG_CALL_STACK, ofs_slot0);
6115 }
6116 }
6117 break;
6118 default:
6119 g_assert_not_reached();
6120 }
6121
6122 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg);
6123 }
6124
tcg_out_ld_helper_ret(TCGContext * s,const TCGLabelQemuLdst * ldst,bool load_sign,const TCGLdstHelperParam * parm)6125 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *ldst,
6126 bool load_sign,
6127 const TCGLdstHelperParam *parm)
6128 {
6129 MemOp mop = get_memop(ldst->oi);
6130 TCGMovExtend mov[2];
6131 int ofs_slot0;
6132
6133 switch (ldst->type) {
6134 case TCG_TYPE_I64:
6135 if (TCG_TARGET_REG_BITS == 32) {
6136 break;
6137 }
6138 /* fall through */
6139
6140 case TCG_TYPE_I32:
6141 mov[0].dst = ldst->datalo_reg;
6142 mov[0].src = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, 0);
6143 mov[0].dst_type = ldst->type;
6144 mov[0].src_type = TCG_TYPE_REG;
6145
6146 /*
6147 * If load_sign, then we allowed the helper to perform the
6148 * appropriate sign extension to tcg_target_ulong, and all
6149 * we need now is a plain move.
6150 *
6151 * If they do not, then we expect the relevant extension
6152 * instruction to be no more expensive than a move, and
6153 * we thus save the icache etc by only using one of two
6154 * helper functions.
6155 */
6156 if (load_sign || !(mop & MO_SIGN)) {
6157 if (TCG_TARGET_REG_BITS == 32 || ldst->type == TCG_TYPE_I32) {
6158 mov[0].src_ext = MO_32;
6159 } else {
6160 mov[0].src_ext = MO_64;
6161 }
6162 } else {
6163 mov[0].src_ext = mop & MO_SSIZE;
6164 }
6165 tcg_out_movext1(s, mov);
6166 return;
6167
6168 case TCG_TYPE_I128:
6169 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
6170 ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET;
6171 switch (TCG_TARGET_CALL_RET_I128) {
6172 case TCG_CALL_RET_NORMAL:
6173 break;
6174 case TCG_CALL_RET_BY_VEC:
6175 tcg_out_st(s, TCG_TYPE_V128,
6176 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0),
6177 TCG_REG_CALL_STACK, ofs_slot0);
6178 /* fall through */
6179 case TCG_CALL_RET_BY_REF:
6180 tcg_out_ld(s, TCG_TYPE_I64, ldst->datalo_reg,
6181 TCG_REG_CALL_STACK, ofs_slot0 + 8 * HOST_BIG_ENDIAN);
6182 tcg_out_ld(s, TCG_TYPE_I64, ldst->datahi_reg,
6183 TCG_REG_CALL_STACK, ofs_slot0 + 8 * !HOST_BIG_ENDIAN);
6184 return;
6185 default:
6186 g_assert_not_reached();
6187 }
6188 break;
6189
6190 default:
6191 g_assert_not_reached();
6192 }
6193
6194 mov[0].dst = ldst->datalo_reg;
6195 mov[0].src =
6196 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, HOST_BIG_ENDIAN);
6197 mov[0].dst_type = TCG_TYPE_REG;
6198 mov[0].src_type = TCG_TYPE_REG;
6199 mov[0].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64;
6200
6201 mov[1].dst = ldst->datahi_reg;
6202 mov[1].src =
6203 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, !HOST_BIG_ENDIAN);
6204 mov[1].dst_type = TCG_TYPE_REG;
6205 mov[1].src_type = TCG_TYPE_REG;
6206 mov[1].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64;
6207
6208 tcg_out_movext2(s, mov, mov + 1, parm->ntmp ? parm->tmp[0] : -1);
6209 }
6210
tcg_out_st_helper_args(TCGContext * s,const TCGLabelQemuLdst * ldst,const TCGLdstHelperParam * parm)6211 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst,
6212 const TCGLdstHelperParam *parm)
6213 {
6214 const TCGHelperInfo *info;
6215 const TCGCallArgumentLoc *loc;
6216 TCGMovExtend mov[4];
6217 TCGType data_type;
6218 unsigned next_arg, nmov, n;
6219 MemOp mop = get_memop(ldst->oi);
6220
6221 switch (mop & MO_SIZE) {
6222 case MO_8:
6223 case MO_16:
6224 case MO_32:
6225 info = &info_helper_st32_mmu;
6226 data_type = TCG_TYPE_I32;
6227 break;
6228 case MO_64:
6229 info = &info_helper_st64_mmu;
6230 data_type = TCG_TYPE_I64;
6231 break;
6232 case MO_128:
6233 info = &info_helper_st128_mmu;
6234 data_type = TCG_TYPE_I128;
6235 break;
6236 default:
6237 g_assert_not_reached();
6238 }
6239
6240 /* Defer env argument. */
6241 next_arg = 1;
6242 nmov = 0;
6243
6244 /* Handle addr argument. */
6245 loc = &info->in[next_arg];
6246 tcg_debug_assert(s->addr_type <= TCG_TYPE_REG);
6247 if (TCG_TARGET_REG_BITS == 32) {
6248 /*
6249 * 32-bit host (and thus 32-bit guest): zero-extend the guest address
6250 * to 64-bits for the helper by storing the low part. Later,
6251 * after we have processed the register inputs, we will load a
6252 * zero for the high part.
6253 */
6254 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN,
6255 TCG_TYPE_I32, TCG_TYPE_I32,
6256 ldst->addr_reg, -1);
6257 next_arg += 2;
6258 nmov += 1;
6259 } else {
6260 n = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type,
6261 ldst->addr_reg, -1);
6262 next_arg += n;
6263 nmov += n;
6264 }
6265
6266 /* Handle data argument. */
6267 loc = &info->in[next_arg];
6268 switch (loc->kind) {
6269 case TCG_CALL_ARG_NORMAL:
6270 case TCG_CALL_ARG_EXTEND_U:
6271 case TCG_CALL_ARG_EXTEND_S:
6272 n = tcg_out_helper_add_mov(mov + nmov, loc, data_type, ldst->type,
6273 ldst->datalo_reg, ldst->datahi_reg);
6274 next_arg += n;
6275 nmov += n;
6276 tcg_out_helper_load_slots(s, nmov, mov, parm);
6277 break;
6278
6279 case TCG_CALL_ARG_BY_REF:
6280 tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
6281 tcg_debug_assert(data_type == TCG_TYPE_I128);
6282 tcg_out_st(s, TCG_TYPE_I64,
6283 HOST_BIG_ENDIAN ? ldst->datahi_reg : ldst->datalo_reg,
6284 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[0].ref_slot));
6285 tcg_out_st(s, TCG_TYPE_I64,
6286 HOST_BIG_ENDIAN ? ldst->datalo_reg : ldst->datahi_reg,
6287 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[1].ref_slot));
6288
6289 tcg_out_helper_load_slots(s, nmov, mov, parm);
6290
6291 if (arg_slot_reg_p(loc->arg_slot)) {
6292 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[loc->arg_slot],
6293 TCG_REG_CALL_STACK,
6294 arg_slot_stk_ofs(loc->ref_slot));
6295 } else {
6296 tcg_debug_assert(parm->ntmp != 0);
6297 tcg_out_addi_ptr(s, parm->tmp[0], TCG_REG_CALL_STACK,
6298 arg_slot_stk_ofs(loc->ref_slot));
6299 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0],
6300 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc->arg_slot));
6301 }
6302 next_arg += 2;
6303 break;
6304
6305 default:
6306 g_assert_not_reached();
6307 }
6308
6309 if (TCG_TARGET_REG_BITS == 32) {
6310 /* Zero extend the address by loading a zero for the high part. */
6311 loc = &info->in[1 + !HOST_BIG_ENDIAN];
6312 tcg_out_helper_load_imm(s, loc->arg_slot, TCG_TYPE_I32, 0, parm);
6313 }
6314
6315 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg);
6316 }
6317
tcg_gen_code(TCGContext * s,TranslationBlock * tb,uint64_t pc_start)6318 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start)
6319 {
6320 int i, start_words, num_insns;
6321 TCGOp *op;
6322
6323 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP)
6324 && qemu_log_in_addr_range(pc_start))) {
6325 FILE *logfile = qemu_log_trylock();
6326 if (logfile) {
6327 fprintf(logfile, "OP:\n");
6328 tcg_dump_ops(s, logfile, false);
6329 fprintf(logfile, "\n");
6330 qemu_log_unlock(logfile);
6331 }
6332 }
6333
6334 #ifdef CONFIG_DEBUG_TCG
6335 /* Ensure all labels referenced have been emitted. */
6336 {
6337 TCGLabel *l;
6338 bool error = false;
6339
6340 QSIMPLEQ_FOREACH(l, &s->labels, next) {
6341 if (unlikely(!l->present) && !QSIMPLEQ_EMPTY(&l->branches)) {
6342 qemu_log_mask(CPU_LOG_TB_OP,
6343 "$L%d referenced but not present.\n", l->id);
6344 error = true;
6345 }
6346 }
6347 assert(!error);
6348 }
6349 #endif
6350
6351 /* Do not reuse any EBB that may be allocated within the TB. */
6352 tcg_temp_ebb_reset_freed(s);
6353
6354 tcg_optimize(s);
6355
6356 reachable_code_pass(s);
6357 liveness_pass_0(s);
6358 liveness_pass_1(s);
6359
6360 if (s->nb_indirects > 0) {
6361 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND)
6362 && qemu_log_in_addr_range(pc_start))) {
6363 FILE *logfile = qemu_log_trylock();
6364 if (logfile) {
6365 fprintf(logfile, "OP before indirect lowering:\n");
6366 tcg_dump_ops(s, logfile, false);
6367 fprintf(logfile, "\n");
6368 qemu_log_unlock(logfile);
6369 }
6370 }
6371
6372 /* Replace indirect temps with direct temps. */
6373 if (liveness_pass_2(s)) {
6374 /* If changes were made, re-run liveness. */
6375 liveness_pass_1(s);
6376 }
6377 }
6378
6379 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT)
6380 && qemu_log_in_addr_range(pc_start))) {
6381 FILE *logfile = qemu_log_trylock();
6382 if (logfile) {
6383 fprintf(logfile, "OP after optimization and liveness analysis:\n");
6384 tcg_dump_ops(s, logfile, true);
6385 fprintf(logfile, "\n");
6386 qemu_log_unlock(logfile);
6387 }
6388 }
6389
6390 /* Initialize goto_tb jump offsets. */
6391 tb->jmp_reset_offset[0] = TB_JMP_OFFSET_INVALID;
6392 tb->jmp_reset_offset[1] = TB_JMP_OFFSET_INVALID;
6393 tb->jmp_insn_offset[0] = TB_JMP_OFFSET_INVALID;
6394 tb->jmp_insn_offset[1] = TB_JMP_OFFSET_INVALID;
6395
6396 tcg_reg_alloc_start(s);
6397
6398 /*
6399 * Reset the buffer pointers when restarting after overflow.
6400 * TODO: Move this into translate-all.c with the rest of the
6401 * buffer management. Having only this done here is confusing.
6402 */
6403 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr);
6404 s->code_ptr = s->code_buf;
6405 s->data_gen_ptr = NULL;
6406
6407 QSIMPLEQ_INIT(&s->ldst_labels);
6408 s->pool_labels = NULL;
6409
6410 start_words = s->insn_start_words;
6411 s->gen_insn_data =
6412 tcg_malloc(sizeof(uint64_t) * s->gen_tb->icount * start_words);
6413
6414 tcg_out_tb_start(s);
6415
6416 num_insns = -1;
6417 QTAILQ_FOREACH(op, &s->ops, link) {
6418 TCGOpcode opc = op->opc;
6419
6420 switch (opc) {
6421 case INDEX_op_mov_i32:
6422 case INDEX_op_mov_i64:
6423 case INDEX_op_mov_vec:
6424 tcg_reg_alloc_mov(s, op);
6425 break;
6426 case INDEX_op_dup_vec:
6427 tcg_reg_alloc_dup(s, op);
6428 break;
6429 case INDEX_op_insn_start:
6430 if (num_insns >= 0) {
6431 size_t off = tcg_current_code_size(s);
6432 s->gen_insn_end_off[num_insns] = off;
6433 /* Assert that we do not overflow our stored offset. */
6434 assert(s->gen_insn_end_off[num_insns] == off);
6435 }
6436 num_insns++;
6437 for (i = 0; i < start_words; ++i) {
6438 s->gen_insn_data[num_insns * start_words + i] =
6439 tcg_get_insn_start_param(op, i);
6440 }
6441 break;
6442 case INDEX_op_discard:
6443 temp_dead(s, arg_temp(op->args[0]));
6444 break;
6445 case INDEX_op_set_label:
6446 tcg_reg_alloc_bb_end(s, s->reserved_regs);
6447 tcg_out_label(s, arg_label(op->args[0]));
6448 break;
6449 case INDEX_op_call:
6450 tcg_reg_alloc_call(s, op);
6451 break;
6452 case INDEX_op_exit_tb:
6453 tcg_out_exit_tb(s, op->args[0]);
6454 break;
6455 case INDEX_op_goto_tb:
6456 tcg_out_goto_tb(s, op->args[0]);
6457 break;
6458 case INDEX_op_dup2_vec:
6459 if (tcg_reg_alloc_dup2(s, op)) {
6460 break;
6461 }
6462 /* fall through */
6463 default:
6464 /* Sanity check that we've not introduced any unhandled opcodes. */
6465 tcg_debug_assert(tcg_op_supported(opc, TCGOP_TYPE(op),
6466 TCGOP_FLAGS(op)));
6467 /* Note: in order to speed up the code, it would be much
6468 faster to have specialized register allocator functions for
6469 some common argument patterns */
6470 tcg_reg_alloc_op(s, op);
6471 break;
6472 }
6473 /* Test for (pending) buffer overflow. The assumption is that any
6474 one operation beginning below the high water mark cannot overrun
6475 the buffer completely. Thus we can test for overflow after
6476 generating code without having to check during generation. */
6477 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) {
6478 return -1;
6479 }
6480 /* Test for TB overflow, as seen by gen_insn_end_off. */
6481 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) {
6482 return -2;
6483 }
6484 }
6485 tcg_debug_assert(num_insns + 1 == s->gen_tb->icount);
6486 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s);
6487
6488 /* Generate TB finalization at the end of block */
6489 i = tcg_out_ldst_finalize(s);
6490 if (i < 0) {
6491 return i;
6492 }
6493 i = tcg_out_pool_finalize(s);
6494 if (i < 0) {
6495 return i;
6496 }
6497 if (!tcg_resolve_relocs(s)) {
6498 return -2;
6499 }
6500
6501 #ifndef CONFIG_TCG_INTERPRETER
6502 /* flush instruction cache */
6503 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf),
6504 (uintptr_t)s->code_buf,
6505 tcg_ptr_byte_diff(s->code_ptr, s->code_buf));
6506 #endif
6507
6508 return tcg_current_code_size(s);
6509 }
6510
6511 #ifdef ELF_HOST_MACHINE
6512 /* In order to use this feature, the backend needs to do three things:
6513
6514 (1) Define ELF_HOST_MACHINE to indicate both what value to
6515 put into the ELF image and to indicate support for the feature.
6516
6517 (2) Define tcg_register_jit. This should create a buffer containing
6518 the contents of a .debug_frame section that describes the post-
6519 prologue unwind info for the tcg machine.
6520
6521 (3) Call tcg_register_jit_int, with the constructed .debug_frame.
6522 */
6523
6524 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */
6525 typedef enum {
6526 JIT_NOACTION = 0,
6527 JIT_REGISTER_FN,
6528 JIT_UNREGISTER_FN
6529 } jit_actions_t;
6530
6531 struct jit_code_entry {
6532 struct jit_code_entry *next_entry;
6533 struct jit_code_entry *prev_entry;
6534 const void *symfile_addr;
6535 uint64_t symfile_size;
6536 };
6537
6538 struct jit_descriptor {
6539 uint32_t version;
6540 uint32_t action_flag;
6541 struct jit_code_entry *relevant_entry;
6542 struct jit_code_entry *first_entry;
6543 };
6544
6545 void __jit_debug_register_code(void) __attribute__((noinline));
__jit_debug_register_code(void)6546 void __jit_debug_register_code(void)
6547 {
6548 asm("");
6549 }
6550
6551 /* Must statically initialize the version, because GDB may check
6552 the version before we can set it. */
6553 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 };
6554
6555 /* End GDB interface. */
6556
find_string(const char * strtab,const char * str)6557 static int find_string(const char *strtab, const char *str)
6558 {
6559 const char *p = strtab + 1;
6560
6561 while (1) {
6562 if (strcmp(p, str) == 0) {
6563 return p - strtab;
6564 }
6565 p += strlen(p) + 1;
6566 }
6567 }
6568
tcg_register_jit_int(const void * buf_ptr,size_t buf_size,const void * debug_frame,size_t debug_frame_size)6569 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size,
6570 const void *debug_frame,
6571 size_t debug_frame_size)
6572 {
6573 struct __attribute__((packed)) DebugInfo {
6574 uint32_t len;
6575 uint16_t version;
6576 uint32_t abbrev;
6577 uint8_t ptr_size;
6578 uint8_t cu_die;
6579 uint16_t cu_lang;
6580 uintptr_t cu_low_pc;
6581 uintptr_t cu_high_pc;
6582 uint8_t fn_die;
6583 char fn_name[16];
6584 uintptr_t fn_low_pc;
6585 uintptr_t fn_high_pc;
6586 uint8_t cu_eoc;
6587 };
6588
6589 struct ElfImage {
6590 ElfW(Ehdr) ehdr;
6591 ElfW(Phdr) phdr;
6592 ElfW(Shdr) shdr[7];
6593 ElfW(Sym) sym[2];
6594 struct DebugInfo di;
6595 uint8_t da[24];
6596 char str[80];
6597 };
6598
6599 struct ElfImage *img;
6600
6601 static const struct ElfImage img_template = {
6602 .ehdr = {
6603 .e_ident[EI_MAG0] = ELFMAG0,
6604 .e_ident[EI_MAG1] = ELFMAG1,
6605 .e_ident[EI_MAG2] = ELFMAG2,
6606 .e_ident[EI_MAG3] = ELFMAG3,
6607 .e_ident[EI_CLASS] = ELF_CLASS,
6608 .e_ident[EI_DATA] = ELF_DATA,
6609 .e_ident[EI_VERSION] = EV_CURRENT,
6610 .e_type = ET_EXEC,
6611 .e_machine = ELF_HOST_MACHINE,
6612 .e_version = EV_CURRENT,
6613 .e_phoff = offsetof(struct ElfImage, phdr),
6614 .e_shoff = offsetof(struct ElfImage, shdr),
6615 .e_ehsize = sizeof(ElfW(Shdr)),
6616 .e_phentsize = sizeof(ElfW(Phdr)),
6617 .e_phnum = 1,
6618 .e_shentsize = sizeof(ElfW(Shdr)),
6619 .e_shnum = ARRAY_SIZE(img->shdr),
6620 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1,
6621 #ifdef ELF_HOST_FLAGS
6622 .e_flags = ELF_HOST_FLAGS,
6623 #endif
6624 #ifdef ELF_OSABI
6625 .e_ident[EI_OSABI] = ELF_OSABI,
6626 #endif
6627 },
6628 .phdr = {
6629 .p_type = PT_LOAD,
6630 .p_flags = PF_X,
6631 },
6632 .shdr = {
6633 [0] = { .sh_type = SHT_NULL },
6634 /* Trick: The contents of code_gen_buffer are not present in
6635 this fake ELF file; that got allocated elsewhere. Therefore
6636 we mark .text as SHT_NOBITS (similar to .bss) so that readers
6637 will not look for contents. We can record any address. */
6638 [1] = { /* .text */
6639 .sh_type = SHT_NOBITS,
6640 .sh_flags = SHF_EXECINSTR | SHF_ALLOC,
6641 },
6642 [2] = { /* .debug_info */
6643 .sh_type = SHT_PROGBITS,
6644 .sh_offset = offsetof(struct ElfImage, di),
6645 .sh_size = sizeof(struct DebugInfo),
6646 },
6647 [3] = { /* .debug_abbrev */
6648 .sh_type = SHT_PROGBITS,
6649 .sh_offset = offsetof(struct ElfImage, da),
6650 .sh_size = sizeof(img->da),
6651 },
6652 [4] = { /* .debug_frame */
6653 .sh_type = SHT_PROGBITS,
6654 .sh_offset = sizeof(struct ElfImage),
6655 },
6656 [5] = { /* .symtab */
6657 .sh_type = SHT_SYMTAB,
6658 .sh_offset = offsetof(struct ElfImage, sym),
6659 .sh_size = sizeof(img->sym),
6660 .sh_info = 1,
6661 .sh_link = ARRAY_SIZE(img->shdr) - 1,
6662 .sh_entsize = sizeof(ElfW(Sym)),
6663 },
6664 [6] = { /* .strtab */
6665 .sh_type = SHT_STRTAB,
6666 .sh_offset = offsetof(struct ElfImage, str),
6667 .sh_size = sizeof(img->str),
6668 }
6669 },
6670 .sym = {
6671 [1] = { /* code_gen_buffer */
6672 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC),
6673 .st_shndx = 1,
6674 }
6675 },
6676 .di = {
6677 .len = sizeof(struct DebugInfo) - 4,
6678 .version = 2,
6679 .ptr_size = sizeof(void *),
6680 .cu_die = 1,
6681 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */
6682 .fn_die = 2,
6683 .fn_name = "code_gen_buffer"
6684 },
6685 .da = {
6686 1, /* abbrev number (the cu) */
6687 0x11, 1, /* DW_TAG_compile_unit, has children */
6688 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */
6689 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */
6690 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */
6691 0, 0, /* end of abbrev */
6692 2, /* abbrev number (the fn) */
6693 0x2e, 0, /* DW_TAG_subprogram, no children */
6694 0x3, 0x8, /* DW_AT_name, DW_FORM_string */
6695 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */
6696 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */
6697 0, 0, /* end of abbrev */
6698 0 /* no more abbrev */
6699 },
6700 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0"
6701 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer",
6702 };
6703
6704 /* We only need a single jit entry; statically allocate it. */
6705 static struct jit_code_entry one_entry;
6706
6707 uintptr_t buf = (uintptr_t)buf_ptr;
6708 size_t img_size = sizeof(struct ElfImage) + debug_frame_size;
6709 DebugFrameHeader *dfh;
6710
6711 img = g_malloc(img_size);
6712 *img = img_template;
6713
6714 img->phdr.p_vaddr = buf;
6715 img->phdr.p_paddr = buf;
6716 img->phdr.p_memsz = buf_size;
6717
6718 img->shdr[1].sh_name = find_string(img->str, ".text");
6719 img->shdr[1].sh_addr = buf;
6720 img->shdr[1].sh_size = buf_size;
6721
6722 img->shdr[2].sh_name = find_string(img->str, ".debug_info");
6723 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev");
6724
6725 img->shdr[4].sh_name = find_string(img->str, ".debug_frame");
6726 img->shdr[4].sh_size = debug_frame_size;
6727
6728 img->shdr[5].sh_name = find_string(img->str, ".symtab");
6729 img->shdr[6].sh_name = find_string(img->str, ".strtab");
6730
6731 img->sym[1].st_name = find_string(img->str, "code_gen_buffer");
6732 img->sym[1].st_value = buf;
6733 img->sym[1].st_size = buf_size;
6734
6735 img->di.cu_low_pc = buf;
6736 img->di.cu_high_pc = buf + buf_size;
6737 img->di.fn_low_pc = buf;
6738 img->di.fn_high_pc = buf + buf_size;
6739
6740 dfh = (DebugFrameHeader *)(img + 1);
6741 memcpy(dfh, debug_frame, debug_frame_size);
6742 dfh->fde.func_start = buf;
6743 dfh->fde.func_len = buf_size;
6744
6745 #ifdef DEBUG_JIT
6746 /* Enable this block to be able to debug the ELF image file creation.
6747 One can use readelf, objdump, or other inspection utilities. */
6748 {
6749 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir());
6750 FILE *f = fopen(jit, "w+b");
6751 if (f) {
6752 if (fwrite(img, img_size, 1, f) != img_size) {
6753 /* Avoid stupid unused return value warning for fwrite. */
6754 }
6755 fclose(f);
6756 }
6757 }
6758 #endif
6759
6760 one_entry.symfile_addr = img;
6761 one_entry.symfile_size = img_size;
6762
6763 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN;
6764 __jit_debug_descriptor.relevant_entry = &one_entry;
6765 __jit_debug_descriptor.first_entry = &one_entry;
6766 __jit_debug_register_code();
6767 }
6768 #else
6769 /* No support for the feature. Provide the entry point expected by exec.c,
6770 and implement the internal function we declared earlier. */
6771
tcg_register_jit_int(const void * buf,size_t size,const void * debug_frame,size_t debug_frame_size)6772 static void tcg_register_jit_int(const void *buf, size_t size,
6773 const void *debug_frame,
6774 size_t debug_frame_size)
6775 {
6776 }
6777
tcg_register_jit(const void * buf,size_t buf_size)6778 void tcg_register_jit(const void *buf, size_t buf_size)
6779 {
6780 }
6781 #endif /* ELF_HOST_MACHINE */
6782
6783 #if !TCG_TARGET_MAYBE_vec
tcg_expand_vec_op(TCGOpcode o,TCGType t,unsigned e,TCGArg a0,...)6784 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...)
6785 {
6786 g_assert_not_reached();
6787 }
6788 #endif
6789