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 /* define it to use liveness analysis (better code) */ 26 #define USE_TCG_OPTIMIZATIONS 27 28 #include "qemu/osdep.h" 29 30 /* Define to jump the ELF file used to communicate with GDB. */ 31 #undef DEBUG_JIT 32 33 #include "qemu/error-report.h" 34 #include "qemu/cutils.h" 35 #include "qemu/host-utils.h" 36 #include "qemu/qemu-print.h" 37 #include "qemu/timer.h" 38 #include "qemu/cacheflush.h" 39 #include "qemu/cacheinfo.h" 40 41 /* Note: the long term plan is to reduce the dependencies on the QEMU 42 CPU definitions. Currently they are used for qemu_ld/st 43 instructions */ 44 #define NO_CPU_IO_DEFS 45 46 #include "exec/exec-all.h" 47 #include "tcg/tcg-op.h" 48 49 #if UINTPTR_MAX == UINT32_MAX 50 # define ELF_CLASS ELFCLASS32 51 #else 52 # define ELF_CLASS ELFCLASS64 53 #endif 54 #if HOST_BIG_ENDIAN 55 # define ELF_DATA ELFDATA2MSB 56 #else 57 # define ELF_DATA ELFDATA2LSB 58 #endif 59 60 #include "elf.h" 61 #include "exec/log.h" 62 #include "tcg/tcg-ldst.h" 63 #include "tcg-internal.h" 64 65 #ifdef CONFIG_TCG_INTERPRETER 66 #include <ffi.h> 67 #endif 68 69 /* Forward declarations for functions declared in tcg-target.c.inc and 70 used here. */ 71 static void tcg_target_init(TCGContext *s); 72 static void tcg_target_qemu_prologue(TCGContext *s); 73 static bool patch_reloc(tcg_insn_unit *code_ptr, int type, 74 intptr_t value, intptr_t addend); 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 static void tcg_register_jit_int(const void *buf, size_t size, 100 const void *debug_frame, 101 size_t debug_frame_size) 102 __attribute__((unused)); 103 104 /* Forward declarations for functions declared and used in tcg-target.c.inc. */ 105 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1, 106 intptr_t arg2); 107 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 108 static void tcg_out_movi(TCGContext *s, TCGType type, 109 TCGReg ret, tcg_target_long arg); 110 static void tcg_out_op(TCGContext *s, TCGOpcode opc, 111 const TCGArg args[TCG_MAX_OP_ARGS], 112 const int const_args[TCG_MAX_OP_ARGS]); 113 #if TCG_TARGET_MAYBE_vec 114 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 115 TCGReg dst, TCGReg src); 116 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 117 TCGReg dst, TCGReg base, intptr_t offset); 118 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 119 TCGReg dst, int64_t arg); 120 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 121 unsigned vecl, unsigned vece, 122 const TCGArg args[TCG_MAX_OP_ARGS], 123 const int const_args[TCG_MAX_OP_ARGS]); 124 #else 125 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 126 TCGReg dst, TCGReg src) 127 { 128 g_assert_not_reached(); 129 } 130 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 131 TCGReg dst, TCGReg base, intptr_t offset) 132 { 133 g_assert_not_reached(); 134 } 135 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 136 TCGReg dst, int64_t arg) 137 { 138 g_assert_not_reached(); 139 } 140 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 141 unsigned vecl, unsigned vece, 142 const TCGArg args[TCG_MAX_OP_ARGS], 143 const int const_args[TCG_MAX_OP_ARGS]) 144 { 145 g_assert_not_reached(); 146 } 147 #endif 148 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1, 149 intptr_t arg2); 150 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, 151 TCGReg base, intptr_t ofs); 152 #ifdef CONFIG_TCG_INTERPRETER 153 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target, 154 ffi_cif *cif); 155 #else 156 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target); 157 #endif 158 static bool tcg_target_const_match(int64_t val, TCGType type, int ct); 159 #ifdef TCG_TARGET_NEED_LDST_LABELS 160 static int tcg_out_ldst_finalize(TCGContext *s); 161 #endif 162 163 TCGContext tcg_init_ctx; 164 __thread TCGContext *tcg_ctx; 165 166 TCGContext **tcg_ctxs; 167 unsigned int tcg_cur_ctxs; 168 unsigned int tcg_max_ctxs; 169 TCGv_env cpu_env = 0; 170 const void *tcg_code_gen_epilogue; 171 uintptr_t tcg_splitwx_diff; 172 173 #ifndef CONFIG_TCG_INTERPRETER 174 tcg_prologue_fn *tcg_qemu_tb_exec; 175 #endif 176 177 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT]; 178 static TCGRegSet tcg_target_call_clobber_regs; 179 180 #if TCG_TARGET_INSN_UNIT_SIZE == 1 181 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v) 182 { 183 *s->code_ptr++ = v; 184 } 185 186 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p, 187 uint8_t v) 188 { 189 *p = v; 190 } 191 #endif 192 193 #if TCG_TARGET_INSN_UNIT_SIZE <= 2 194 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v) 195 { 196 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 197 *s->code_ptr++ = v; 198 } else { 199 tcg_insn_unit *p = s->code_ptr; 200 memcpy(p, &v, sizeof(v)); 201 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE); 202 } 203 } 204 205 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p, 206 uint16_t v) 207 { 208 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 209 *p = v; 210 } else { 211 memcpy(p, &v, sizeof(v)); 212 } 213 } 214 #endif 215 216 #if TCG_TARGET_INSN_UNIT_SIZE <= 4 217 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v) 218 { 219 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 220 *s->code_ptr++ = v; 221 } else { 222 tcg_insn_unit *p = s->code_ptr; 223 memcpy(p, &v, sizeof(v)); 224 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE); 225 } 226 } 227 228 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p, 229 uint32_t v) 230 { 231 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 232 *p = v; 233 } else { 234 memcpy(p, &v, sizeof(v)); 235 } 236 } 237 #endif 238 239 #if TCG_TARGET_INSN_UNIT_SIZE <= 8 240 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v) 241 { 242 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 243 *s->code_ptr++ = v; 244 } else { 245 tcg_insn_unit *p = s->code_ptr; 246 memcpy(p, &v, sizeof(v)); 247 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE); 248 } 249 } 250 251 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p, 252 uint64_t v) 253 { 254 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 255 *p = v; 256 } else { 257 memcpy(p, &v, sizeof(v)); 258 } 259 } 260 #endif 261 262 /* label relocation processing */ 263 264 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type, 265 TCGLabel *l, intptr_t addend) 266 { 267 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation)); 268 269 r->type = type; 270 r->ptr = code_ptr; 271 r->addend = addend; 272 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next); 273 } 274 275 static void tcg_out_label(TCGContext *s, TCGLabel *l) 276 { 277 tcg_debug_assert(!l->has_value); 278 l->has_value = 1; 279 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr); 280 } 281 282 TCGLabel *gen_new_label(void) 283 { 284 TCGContext *s = tcg_ctx; 285 TCGLabel *l = tcg_malloc(sizeof(TCGLabel)); 286 287 memset(l, 0, sizeof(TCGLabel)); 288 l->id = s->nb_labels++; 289 QSIMPLEQ_INIT(&l->relocs); 290 291 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next); 292 293 return l; 294 } 295 296 static bool tcg_resolve_relocs(TCGContext *s) 297 { 298 TCGLabel *l; 299 300 QSIMPLEQ_FOREACH(l, &s->labels, next) { 301 TCGRelocation *r; 302 uintptr_t value = l->u.value; 303 304 QSIMPLEQ_FOREACH(r, &l->relocs, next) { 305 if (!patch_reloc(r->ptr, r->type, value, r->addend)) { 306 return false; 307 } 308 } 309 } 310 return true; 311 } 312 313 static void set_jmp_reset_offset(TCGContext *s, int which) 314 { 315 /* 316 * We will check for overflow at the end of the opcode loop in 317 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 318 */ 319 s->tb_jmp_reset_offset[which] = tcg_current_code_size(s); 320 } 321 322 /* Signal overflow, starting over with fewer guest insns. */ 323 static G_NORETURN 324 void tcg_raise_tb_overflow(TCGContext *s) 325 { 326 siglongjmp(s->jmp_trans, -2); 327 } 328 329 #define C_PFX1(P, A) P##A 330 #define C_PFX2(P, A, B) P##A##_##B 331 #define C_PFX3(P, A, B, C) P##A##_##B##_##C 332 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D 333 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E 334 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F 335 336 /* Define an enumeration for the various combinations. */ 337 338 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1), 339 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2), 340 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3), 341 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4), 342 343 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1), 344 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2), 345 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3), 346 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4), 347 348 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2), 349 350 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1), 351 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2), 352 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3), 353 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4), 354 355 typedef enum { 356 #include "tcg-target-con-set.h" 357 } TCGConstraintSetIndex; 358 359 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode); 360 361 #undef C_O0_I1 362 #undef C_O0_I2 363 #undef C_O0_I3 364 #undef C_O0_I4 365 #undef C_O1_I1 366 #undef C_O1_I2 367 #undef C_O1_I3 368 #undef C_O1_I4 369 #undef C_N1_I2 370 #undef C_O2_I1 371 #undef C_O2_I2 372 #undef C_O2_I3 373 #undef C_O2_I4 374 375 /* Put all of the constraint sets into an array, indexed by the enum. */ 376 377 #define C_O0_I1(I1) { .args_ct_str = { #I1 } }, 378 #define C_O0_I2(I1, I2) { .args_ct_str = { #I1, #I2 } }, 379 #define C_O0_I3(I1, I2, I3) { .args_ct_str = { #I1, #I2, #I3 } }, 380 #define C_O0_I4(I1, I2, I3, I4) { .args_ct_str = { #I1, #I2, #I3, #I4 } }, 381 382 #define C_O1_I1(O1, I1) { .args_ct_str = { #O1, #I1 } }, 383 #define C_O1_I2(O1, I1, I2) { .args_ct_str = { #O1, #I1, #I2 } }, 384 #define C_O1_I3(O1, I1, I2, I3) { .args_ct_str = { #O1, #I1, #I2, #I3 } }, 385 #define C_O1_I4(O1, I1, I2, I3, I4) { .args_ct_str = { #O1, #I1, #I2, #I3, #I4 } }, 386 387 #define C_N1_I2(O1, I1, I2) { .args_ct_str = { "&" #O1, #I1, #I2 } }, 388 389 #define C_O2_I1(O1, O2, I1) { .args_ct_str = { #O1, #O2, #I1 } }, 390 #define C_O2_I2(O1, O2, I1, I2) { .args_ct_str = { #O1, #O2, #I1, #I2 } }, 391 #define C_O2_I3(O1, O2, I1, I2, I3) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3 } }, 392 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3, #I4 } }, 393 394 static const TCGTargetOpDef constraint_sets[] = { 395 #include "tcg-target-con-set.h" 396 }; 397 398 399 #undef C_O0_I1 400 #undef C_O0_I2 401 #undef C_O0_I3 402 #undef C_O0_I4 403 #undef C_O1_I1 404 #undef C_O1_I2 405 #undef C_O1_I3 406 #undef C_O1_I4 407 #undef C_N1_I2 408 #undef C_O2_I1 409 #undef C_O2_I2 410 #undef C_O2_I3 411 #undef C_O2_I4 412 413 /* Expand the enumerator to be returned from tcg_target_op_def(). */ 414 415 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1) 416 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2) 417 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3) 418 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4) 419 420 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1) 421 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2) 422 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3) 423 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4) 424 425 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2) 426 427 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1) 428 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2) 429 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3) 430 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4) 431 432 #include "tcg-target.c.inc" 433 434 static void alloc_tcg_plugin_context(TCGContext *s) 435 { 436 #ifdef CONFIG_PLUGIN 437 s->plugin_tb = g_new0(struct qemu_plugin_tb, 1); 438 s->plugin_tb->insns = 439 g_ptr_array_new_with_free_func(qemu_plugin_insn_cleanup_fn); 440 #endif 441 } 442 443 /* 444 * All TCG threads except the parent (i.e. the one that called tcg_context_init 445 * and registered the target's TCG globals) must register with this function 446 * before initiating translation. 447 * 448 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation 449 * of tcg_region_init() for the reasoning behind this. 450 * 451 * In softmmu each caller registers its context in tcg_ctxs[]. Note that in 452 * softmmu tcg_ctxs[] does not track tcg_ctx_init, since the initial context 453 * is not used anymore for translation once this function is called. 454 * 455 * Not tracking tcg_init_ctx in tcg_ctxs[] in softmmu keeps code that iterates 456 * over the array (e.g. tcg_code_size() the same for both softmmu and user-mode. 457 */ 458 #ifdef CONFIG_USER_ONLY 459 void tcg_register_thread(void) 460 { 461 tcg_ctx = &tcg_init_ctx; 462 } 463 #else 464 void tcg_register_thread(void) 465 { 466 TCGContext *s = g_malloc(sizeof(*s)); 467 unsigned int i, n; 468 469 *s = tcg_init_ctx; 470 471 /* Relink mem_base. */ 472 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) { 473 if (tcg_init_ctx.temps[i].mem_base) { 474 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps; 475 tcg_debug_assert(b >= 0 && b < n); 476 s->temps[i].mem_base = &s->temps[b]; 477 } 478 } 479 480 /* Claim an entry in tcg_ctxs */ 481 n = qatomic_fetch_inc(&tcg_cur_ctxs); 482 g_assert(n < tcg_max_ctxs); 483 qatomic_set(&tcg_ctxs[n], s); 484 485 if (n > 0) { 486 alloc_tcg_plugin_context(s); 487 tcg_region_initial_alloc(s); 488 } 489 490 tcg_ctx = s; 491 } 492 #endif /* !CONFIG_USER_ONLY */ 493 494 /* pool based memory allocation */ 495 void *tcg_malloc_internal(TCGContext *s, int size) 496 { 497 TCGPool *p; 498 int pool_size; 499 500 if (size > TCG_POOL_CHUNK_SIZE) { 501 /* big malloc: insert a new pool (XXX: could optimize) */ 502 p = g_malloc(sizeof(TCGPool) + size); 503 p->size = size; 504 p->next = s->pool_first_large; 505 s->pool_first_large = p; 506 return p->data; 507 } else { 508 p = s->pool_current; 509 if (!p) { 510 p = s->pool_first; 511 if (!p) 512 goto new_pool; 513 } else { 514 if (!p->next) { 515 new_pool: 516 pool_size = TCG_POOL_CHUNK_SIZE; 517 p = g_malloc(sizeof(TCGPool) + pool_size); 518 p->size = pool_size; 519 p->next = NULL; 520 if (s->pool_current) 521 s->pool_current->next = p; 522 else 523 s->pool_first = p; 524 } else { 525 p = p->next; 526 } 527 } 528 } 529 s->pool_current = p; 530 s->pool_cur = p->data + size; 531 s->pool_end = p->data + p->size; 532 return p->data; 533 } 534 535 void tcg_pool_reset(TCGContext *s) 536 { 537 TCGPool *p, *t; 538 for (p = s->pool_first_large; p; p = t) { 539 t = p->next; 540 g_free(p); 541 } 542 s->pool_first_large = NULL; 543 s->pool_cur = s->pool_end = NULL; 544 s->pool_current = NULL; 545 } 546 547 #include "exec/helper-proto.h" 548 549 static const TCGHelperInfo all_helpers[] = { 550 #include "exec/helper-tcg.h" 551 }; 552 static GHashTable *helper_table; 553 554 #ifdef CONFIG_TCG_INTERPRETER 555 static GHashTable *ffi_table; 556 557 static ffi_type * const typecode_to_ffi[8] = { 558 [dh_typecode_void] = &ffi_type_void, 559 [dh_typecode_i32] = &ffi_type_uint32, 560 [dh_typecode_s32] = &ffi_type_sint32, 561 [dh_typecode_i64] = &ffi_type_uint64, 562 [dh_typecode_s64] = &ffi_type_sint64, 563 [dh_typecode_ptr] = &ffi_type_pointer, 564 }; 565 #endif 566 567 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)]; 568 static void process_op_defs(TCGContext *s); 569 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 570 TCGReg reg, const char *name); 571 572 static void tcg_context_init(unsigned max_cpus) 573 { 574 TCGContext *s = &tcg_init_ctx; 575 int op, total_args, n, i; 576 TCGOpDef *def; 577 TCGArgConstraint *args_ct; 578 TCGTemp *ts; 579 580 memset(s, 0, sizeof(*s)); 581 s->nb_globals = 0; 582 583 /* Count total number of arguments and allocate the corresponding 584 space */ 585 total_args = 0; 586 for(op = 0; op < NB_OPS; op++) { 587 def = &tcg_op_defs[op]; 588 n = def->nb_iargs + def->nb_oargs; 589 total_args += n; 590 } 591 592 args_ct = g_new0(TCGArgConstraint, total_args); 593 594 for(op = 0; op < NB_OPS; op++) { 595 def = &tcg_op_defs[op]; 596 def->args_ct = args_ct; 597 n = def->nb_iargs + def->nb_oargs; 598 args_ct += n; 599 } 600 601 /* Register helpers. */ 602 /* Use g_direct_hash/equal for direct pointer comparisons on func. */ 603 helper_table = g_hash_table_new(NULL, NULL); 604 605 for (i = 0; i < ARRAY_SIZE(all_helpers); ++i) { 606 g_hash_table_insert(helper_table, (gpointer)all_helpers[i].func, 607 (gpointer)&all_helpers[i]); 608 } 609 610 #ifdef CONFIG_TCG_INTERPRETER 611 /* g_direct_hash/equal for direct comparisons on uint32_t. */ 612 ffi_table = g_hash_table_new(NULL, NULL); 613 for (i = 0; i < ARRAY_SIZE(all_helpers); ++i) { 614 struct { 615 ffi_cif cif; 616 ffi_type *args[]; 617 } *ca; 618 uint32_t typemask = all_helpers[i].typemask; 619 gpointer hash = (gpointer)(uintptr_t)typemask; 620 ffi_status status; 621 int nargs; 622 623 if (g_hash_table_lookup(ffi_table, hash)) { 624 continue; 625 } 626 627 /* Ignoring the return type, find the last non-zero field. */ 628 nargs = 32 - clz32(typemask >> 3); 629 nargs = DIV_ROUND_UP(nargs, 3); 630 631 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *)); 632 ca->cif.rtype = typecode_to_ffi[typemask & 7]; 633 ca->cif.nargs = nargs; 634 635 if (nargs != 0) { 636 ca->cif.arg_types = ca->args; 637 for (int j = 0; j < nargs; ++j) { 638 int typecode = extract32(typemask, (j + 1) * 3, 3); 639 ca->args[j] = typecode_to_ffi[typecode]; 640 } 641 } 642 643 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs, 644 ca->cif.rtype, ca->cif.arg_types); 645 assert(status == FFI_OK); 646 647 g_hash_table_insert(ffi_table, hash, (gpointer)&ca->cif); 648 } 649 #endif 650 651 tcg_target_init(s); 652 process_op_defs(s); 653 654 /* Reverse the order of the saved registers, assuming they're all at 655 the start of tcg_target_reg_alloc_order. */ 656 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) { 657 int r = tcg_target_reg_alloc_order[n]; 658 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) { 659 break; 660 } 661 } 662 for (i = 0; i < n; ++i) { 663 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i]; 664 } 665 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) { 666 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i]; 667 } 668 669 alloc_tcg_plugin_context(s); 670 671 tcg_ctx = s; 672 /* 673 * In user-mode we simply share the init context among threads, since we 674 * use a single region. See the documentation tcg_region_init() for the 675 * reasoning behind this. 676 * In softmmu we will have at most max_cpus TCG threads. 677 */ 678 #ifdef CONFIG_USER_ONLY 679 tcg_ctxs = &tcg_ctx; 680 tcg_cur_ctxs = 1; 681 tcg_max_ctxs = 1; 682 #else 683 tcg_max_ctxs = max_cpus; 684 tcg_ctxs = g_new0(TCGContext *, max_cpus); 685 #endif 686 687 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0)); 688 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env"); 689 cpu_env = temp_tcgv_ptr(ts); 690 } 691 692 void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus) 693 { 694 tcg_context_init(max_cpus); 695 tcg_region_init(tb_size, splitwx, max_cpus); 696 } 697 698 /* 699 * Allocate TBs right before their corresponding translated code, making 700 * sure that TBs and code are on different cache lines. 701 */ 702 TranslationBlock *tcg_tb_alloc(TCGContext *s) 703 { 704 uintptr_t align = qemu_icache_linesize; 705 TranslationBlock *tb; 706 void *next; 707 708 retry: 709 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align); 710 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align); 711 712 if (unlikely(next > s->code_gen_highwater)) { 713 if (tcg_region_alloc(s)) { 714 return NULL; 715 } 716 goto retry; 717 } 718 qatomic_set(&s->code_gen_ptr, next); 719 s->data_gen_ptr = NULL; 720 return tb; 721 } 722 723 void tcg_prologue_init(TCGContext *s) 724 { 725 size_t prologue_size; 726 727 s->code_ptr = s->code_gen_ptr; 728 s->code_buf = s->code_gen_ptr; 729 s->data_gen_ptr = NULL; 730 731 #ifndef CONFIG_TCG_INTERPRETER 732 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr); 733 #endif 734 735 #ifdef TCG_TARGET_NEED_POOL_LABELS 736 s->pool_labels = NULL; 737 #endif 738 739 qemu_thread_jit_write(); 740 /* Generate the prologue. */ 741 tcg_target_qemu_prologue(s); 742 743 #ifdef TCG_TARGET_NEED_POOL_LABELS 744 /* Allow the prologue to put e.g. guest_base into a pool entry. */ 745 { 746 int result = tcg_out_pool_finalize(s); 747 tcg_debug_assert(result == 0); 748 } 749 #endif 750 751 prologue_size = tcg_current_code_size(s); 752 753 #ifndef CONFIG_TCG_INTERPRETER 754 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 755 (uintptr_t)s->code_buf, prologue_size); 756 #endif 757 758 #ifdef DEBUG_DISAS 759 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { 760 FILE *logfile = qemu_log_trylock(); 761 if (logfile) { 762 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size); 763 if (s->data_gen_ptr) { 764 size_t code_size = s->data_gen_ptr - s->code_gen_ptr; 765 size_t data_size = prologue_size - code_size; 766 size_t i; 767 768 disas(logfile, s->code_gen_ptr, code_size); 769 770 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) { 771 if (sizeof(tcg_target_ulong) == 8) { 772 fprintf(logfile, 773 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n", 774 (uintptr_t)s->data_gen_ptr + i, 775 *(uint64_t *)(s->data_gen_ptr + i)); 776 } else { 777 fprintf(logfile, 778 "0x%08" PRIxPTR ": .long 0x%08x\n", 779 (uintptr_t)s->data_gen_ptr + i, 780 *(uint32_t *)(s->data_gen_ptr + i)); 781 } 782 } 783 } else { 784 disas(logfile, s->code_gen_ptr, prologue_size); 785 } 786 fprintf(logfile, "\n"); 787 qemu_log_unlock(logfile); 788 } 789 } 790 #endif 791 792 #ifndef CONFIG_TCG_INTERPRETER 793 /* 794 * Assert that goto_ptr is implemented completely, setting an epilogue. 795 * For tci, we use NULL as the signal to return from the interpreter, 796 * so skip this check. 797 */ 798 tcg_debug_assert(tcg_code_gen_epilogue != NULL); 799 #endif 800 801 tcg_region_prologue_set(s); 802 } 803 804 void tcg_func_start(TCGContext *s) 805 { 806 tcg_pool_reset(s); 807 s->nb_temps = s->nb_globals; 808 809 /* No temps have been previously allocated for size or locality. */ 810 memset(s->free_temps, 0, sizeof(s->free_temps)); 811 812 /* No constant temps have been previously allocated. */ 813 for (int i = 0; i < TCG_TYPE_COUNT; ++i) { 814 if (s->const_table[i]) { 815 g_hash_table_remove_all(s->const_table[i]); 816 } 817 } 818 819 s->nb_ops = 0; 820 s->nb_labels = 0; 821 s->current_frame_offset = s->frame_start; 822 823 #ifdef CONFIG_DEBUG_TCG 824 s->goto_tb_issue_mask = 0; 825 #endif 826 827 QTAILQ_INIT(&s->ops); 828 QTAILQ_INIT(&s->free_ops); 829 QSIMPLEQ_INIT(&s->labels); 830 } 831 832 static TCGTemp *tcg_temp_alloc(TCGContext *s) 833 { 834 int n = s->nb_temps++; 835 836 if (n >= TCG_MAX_TEMPS) { 837 tcg_raise_tb_overflow(s); 838 } 839 return memset(&s->temps[n], 0, sizeof(TCGTemp)); 840 } 841 842 static TCGTemp *tcg_global_alloc(TCGContext *s) 843 { 844 TCGTemp *ts; 845 846 tcg_debug_assert(s->nb_globals == s->nb_temps); 847 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS); 848 s->nb_globals++; 849 ts = tcg_temp_alloc(s); 850 ts->kind = TEMP_GLOBAL; 851 852 return ts; 853 } 854 855 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 856 TCGReg reg, const char *name) 857 { 858 TCGTemp *ts; 859 860 if (TCG_TARGET_REG_BITS == 32 && type != TCG_TYPE_I32) { 861 tcg_abort(); 862 } 863 864 ts = tcg_global_alloc(s); 865 ts->base_type = type; 866 ts->type = type; 867 ts->kind = TEMP_FIXED; 868 ts->reg = reg; 869 ts->name = name; 870 tcg_regset_set_reg(s->reserved_regs, reg); 871 872 return ts; 873 } 874 875 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size) 876 { 877 s->frame_start = start; 878 s->frame_end = start + size; 879 s->frame_temp 880 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame"); 881 } 882 883 TCGTemp *tcg_global_mem_new_internal(TCGType type, TCGv_ptr base, 884 intptr_t offset, const char *name) 885 { 886 TCGContext *s = tcg_ctx; 887 TCGTemp *base_ts = tcgv_ptr_temp(base); 888 TCGTemp *ts = tcg_global_alloc(s); 889 int indirect_reg = 0, bigendian = 0; 890 #if HOST_BIG_ENDIAN 891 bigendian = 1; 892 #endif 893 894 switch (base_ts->kind) { 895 case TEMP_FIXED: 896 break; 897 case TEMP_GLOBAL: 898 /* We do not support double-indirect registers. */ 899 tcg_debug_assert(!base_ts->indirect_reg); 900 base_ts->indirect_base = 1; 901 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64 902 ? 2 : 1); 903 indirect_reg = 1; 904 break; 905 default: 906 g_assert_not_reached(); 907 } 908 909 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 910 TCGTemp *ts2 = tcg_global_alloc(s); 911 char buf[64]; 912 913 ts->base_type = TCG_TYPE_I64; 914 ts->type = TCG_TYPE_I32; 915 ts->indirect_reg = indirect_reg; 916 ts->mem_allocated = 1; 917 ts->mem_base = base_ts; 918 ts->mem_offset = offset + bigendian * 4; 919 pstrcpy(buf, sizeof(buf), name); 920 pstrcat(buf, sizeof(buf), "_0"); 921 ts->name = strdup(buf); 922 923 tcg_debug_assert(ts2 == ts + 1); 924 ts2->base_type = TCG_TYPE_I64; 925 ts2->type = TCG_TYPE_I32; 926 ts2->indirect_reg = indirect_reg; 927 ts2->mem_allocated = 1; 928 ts2->mem_base = base_ts; 929 ts2->mem_offset = offset + (1 - bigendian) * 4; 930 pstrcpy(buf, sizeof(buf), name); 931 pstrcat(buf, sizeof(buf), "_1"); 932 ts2->name = strdup(buf); 933 } else { 934 ts->base_type = type; 935 ts->type = type; 936 ts->indirect_reg = indirect_reg; 937 ts->mem_allocated = 1; 938 ts->mem_base = base_ts; 939 ts->mem_offset = offset; 940 ts->name = name; 941 } 942 return ts; 943 } 944 945 TCGTemp *tcg_temp_new_internal(TCGType type, bool temp_local) 946 { 947 TCGContext *s = tcg_ctx; 948 TCGTempKind kind = temp_local ? TEMP_LOCAL : TEMP_NORMAL; 949 TCGTemp *ts; 950 int idx, k; 951 952 k = type + (temp_local ? TCG_TYPE_COUNT : 0); 953 idx = find_first_bit(s->free_temps[k].l, TCG_MAX_TEMPS); 954 if (idx < TCG_MAX_TEMPS) { 955 /* There is already an available temp with the right type. */ 956 clear_bit(idx, s->free_temps[k].l); 957 958 ts = &s->temps[idx]; 959 ts->temp_allocated = 1; 960 tcg_debug_assert(ts->base_type == type); 961 tcg_debug_assert(ts->kind == kind); 962 } else { 963 ts = tcg_temp_alloc(s); 964 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 965 TCGTemp *ts2 = tcg_temp_alloc(s); 966 967 ts->base_type = type; 968 ts->type = TCG_TYPE_I32; 969 ts->temp_allocated = 1; 970 ts->kind = kind; 971 972 tcg_debug_assert(ts2 == ts + 1); 973 ts2->base_type = TCG_TYPE_I64; 974 ts2->type = TCG_TYPE_I32; 975 ts2->temp_allocated = 1; 976 ts2->kind = kind; 977 } else { 978 ts->base_type = type; 979 ts->type = type; 980 ts->temp_allocated = 1; 981 ts->kind = kind; 982 } 983 } 984 985 #if defined(CONFIG_DEBUG_TCG) 986 s->temps_in_use++; 987 #endif 988 return ts; 989 } 990 991 TCGv_vec tcg_temp_new_vec(TCGType type) 992 { 993 TCGTemp *t; 994 995 #ifdef CONFIG_DEBUG_TCG 996 switch (type) { 997 case TCG_TYPE_V64: 998 assert(TCG_TARGET_HAS_v64); 999 break; 1000 case TCG_TYPE_V128: 1001 assert(TCG_TARGET_HAS_v128); 1002 break; 1003 case TCG_TYPE_V256: 1004 assert(TCG_TARGET_HAS_v256); 1005 break; 1006 default: 1007 g_assert_not_reached(); 1008 } 1009 #endif 1010 1011 t = tcg_temp_new_internal(type, 0); 1012 return temp_tcgv_vec(t); 1013 } 1014 1015 /* Create a new temp of the same type as an existing temp. */ 1016 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match) 1017 { 1018 TCGTemp *t = tcgv_vec_temp(match); 1019 1020 tcg_debug_assert(t->temp_allocated != 0); 1021 1022 t = tcg_temp_new_internal(t->base_type, 0); 1023 return temp_tcgv_vec(t); 1024 } 1025 1026 void tcg_temp_free_internal(TCGTemp *ts) 1027 { 1028 TCGContext *s = tcg_ctx; 1029 int k, idx; 1030 1031 switch (ts->kind) { 1032 case TEMP_CONST: 1033 /* 1034 * In order to simplify users of tcg_constant_*, 1035 * silently ignore free. 1036 */ 1037 return; 1038 case TEMP_NORMAL: 1039 case TEMP_LOCAL: 1040 break; 1041 default: 1042 g_assert_not_reached(); 1043 } 1044 1045 #if defined(CONFIG_DEBUG_TCG) 1046 s->temps_in_use--; 1047 if (s->temps_in_use < 0) { 1048 fprintf(stderr, "More temporaries freed than allocated!\n"); 1049 } 1050 #endif 1051 1052 tcg_debug_assert(ts->temp_allocated != 0); 1053 ts->temp_allocated = 0; 1054 1055 idx = temp_idx(ts); 1056 k = ts->base_type + (ts->kind == TEMP_NORMAL ? 0 : TCG_TYPE_COUNT); 1057 set_bit(idx, s->free_temps[k].l); 1058 } 1059 1060 TCGTemp *tcg_constant_internal(TCGType type, int64_t val) 1061 { 1062 TCGContext *s = tcg_ctx; 1063 GHashTable *h = s->const_table[type]; 1064 TCGTemp *ts; 1065 1066 if (h == NULL) { 1067 h = g_hash_table_new(g_int64_hash, g_int64_equal); 1068 s->const_table[type] = h; 1069 } 1070 1071 ts = g_hash_table_lookup(h, &val); 1072 if (ts == NULL) { 1073 ts = tcg_temp_alloc(s); 1074 1075 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1076 TCGTemp *ts2 = tcg_temp_alloc(s); 1077 1078 ts->base_type = TCG_TYPE_I64; 1079 ts->type = TCG_TYPE_I32; 1080 ts->kind = TEMP_CONST; 1081 ts->temp_allocated = 1; 1082 /* 1083 * Retain the full value of the 64-bit constant in the low 1084 * part, so that the hash table works. Actual uses will 1085 * truncate the value to the low part. 1086 */ 1087 ts->val = val; 1088 1089 tcg_debug_assert(ts2 == ts + 1); 1090 ts2->base_type = TCG_TYPE_I64; 1091 ts2->type = TCG_TYPE_I32; 1092 ts2->kind = TEMP_CONST; 1093 ts2->temp_allocated = 1; 1094 ts2->val = val >> 32; 1095 } else { 1096 ts->base_type = type; 1097 ts->type = type; 1098 ts->kind = TEMP_CONST; 1099 ts->temp_allocated = 1; 1100 ts->val = val; 1101 } 1102 g_hash_table_insert(h, &ts->val, ts); 1103 } 1104 1105 return ts; 1106 } 1107 1108 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val) 1109 { 1110 val = dup_const(vece, val); 1111 return temp_tcgv_vec(tcg_constant_internal(type, val)); 1112 } 1113 1114 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val) 1115 { 1116 TCGTemp *t = tcgv_vec_temp(match); 1117 1118 tcg_debug_assert(t->temp_allocated != 0); 1119 return tcg_constant_vec(t->base_type, vece, val); 1120 } 1121 1122 TCGv_i32 tcg_const_i32(int32_t val) 1123 { 1124 TCGv_i32 t0; 1125 t0 = tcg_temp_new_i32(); 1126 tcg_gen_movi_i32(t0, val); 1127 return t0; 1128 } 1129 1130 TCGv_i64 tcg_const_i64(int64_t val) 1131 { 1132 TCGv_i64 t0; 1133 t0 = tcg_temp_new_i64(); 1134 tcg_gen_movi_i64(t0, val); 1135 return t0; 1136 } 1137 1138 TCGv_i32 tcg_const_local_i32(int32_t val) 1139 { 1140 TCGv_i32 t0; 1141 t0 = tcg_temp_local_new_i32(); 1142 tcg_gen_movi_i32(t0, val); 1143 return t0; 1144 } 1145 1146 TCGv_i64 tcg_const_local_i64(int64_t val) 1147 { 1148 TCGv_i64 t0; 1149 t0 = tcg_temp_local_new_i64(); 1150 tcg_gen_movi_i64(t0, val); 1151 return t0; 1152 } 1153 1154 #if defined(CONFIG_DEBUG_TCG) 1155 void tcg_clear_temp_count(void) 1156 { 1157 TCGContext *s = tcg_ctx; 1158 s->temps_in_use = 0; 1159 } 1160 1161 int tcg_check_temp_count(void) 1162 { 1163 TCGContext *s = tcg_ctx; 1164 if (s->temps_in_use) { 1165 /* Clear the count so that we don't give another 1166 * warning immediately next time around. 1167 */ 1168 s->temps_in_use = 0; 1169 return 1; 1170 } 1171 return 0; 1172 } 1173 #endif 1174 1175 /* Return true if OP may appear in the opcode stream. 1176 Test the runtime variable that controls each opcode. */ 1177 bool tcg_op_supported(TCGOpcode op) 1178 { 1179 const bool have_vec 1180 = TCG_TARGET_HAS_v64 | TCG_TARGET_HAS_v128 | TCG_TARGET_HAS_v256; 1181 1182 switch (op) { 1183 case INDEX_op_discard: 1184 case INDEX_op_set_label: 1185 case INDEX_op_call: 1186 case INDEX_op_br: 1187 case INDEX_op_mb: 1188 case INDEX_op_insn_start: 1189 case INDEX_op_exit_tb: 1190 case INDEX_op_goto_tb: 1191 case INDEX_op_goto_ptr: 1192 case INDEX_op_qemu_ld_i32: 1193 case INDEX_op_qemu_st_i32: 1194 case INDEX_op_qemu_ld_i64: 1195 case INDEX_op_qemu_st_i64: 1196 return true; 1197 1198 case INDEX_op_qemu_st8_i32: 1199 return TCG_TARGET_HAS_qemu_st8_i32; 1200 1201 case INDEX_op_mov_i32: 1202 case INDEX_op_setcond_i32: 1203 case INDEX_op_brcond_i32: 1204 case INDEX_op_ld8u_i32: 1205 case INDEX_op_ld8s_i32: 1206 case INDEX_op_ld16u_i32: 1207 case INDEX_op_ld16s_i32: 1208 case INDEX_op_ld_i32: 1209 case INDEX_op_st8_i32: 1210 case INDEX_op_st16_i32: 1211 case INDEX_op_st_i32: 1212 case INDEX_op_add_i32: 1213 case INDEX_op_sub_i32: 1214 case INDEX_op_mul_i32: 1215 case INDEX_op_and_i32: 1216 case INDEX_op_or_i32: 1217 case INDEX_op_xor_i32: 1218 case INDEX_op_shl_i32: 1219 case INDEX_op_shr_i32: 1220 case INDEX_op_sar_i32: 1221 return true; 1222 1223 case INDEX_op_movcond_i32: 1224 return TCG_TARGET_HAS_movcond_i32; 1225 case INDEX_op_div_i32: 1226 case INDEX_op_divu_i32: 1227 return TCG_TARGET_HAS_div_i32; 1228 case INDEX_op_rem_i32: 1229 case INDEX_op_remu_i32: 1230 return TCG_TARGET_HAS_rem_i32; 1231 case INDEX_op_div2_i32: 1232 case INDEX_op_divu2_i32: 1233 return TCG_TARGET_HAS_div2_i32; 1234 case INDEX_op_rotl_i32: 1235 case INDEX_op_rotr_i32: 1236 return TCG_TARGET_HAS_rot_i32; 1237 case INDEX_op_deposit_i32: 1238 return TCG_TARGET_HAS_deposit_i32; 1239 case INDEX_op_extract_i32: 1240 return TCG_TARGET_HAS_extract_i32; 1241 case INDEX_op_sextract_i32: 1242 return TCG_TARGET_HAS_sextract_i32; 1243 case INDEX_op_extract2_i32: 1244 return TCG_TARGET_HAS_extract2_i32; 1245 case INDEX_op_add2_i32: 1246 return TCG_TARGET_HAS_add2_i32; 1247 case INDEX_op_sub2_i32: 1248 return TCG_TARGET_HAS_sub2_i32; 1249 case INDEX_op_mulu2_i32: 1250 return TCG_TARGET_HAS_mulu2_i32; 1251 case INDEX_op_muls2_i32: 1252 return TCG_TARGET_HAS_muls2_i32; 1253 case INDEX_op_muluh_i32: 1254 return TCG_TARGET_HAS_muluh_i32; 1255 case INDEX_op_mulsh_i32: 1256 return TCG_TARGET_HAS_mulsh_i32; 1257 case INDEX_op_ext8s_i32: 1258 return TCG_TARGET_HAS_ext8s_i32; 1259 case INDEX_op_ext16s_i32: 1260 return TCG_TARGET_HAS_ext16s_i32; 1261 case INDEX_op_ext8u_i32: 1262 return TCG_TARGET_HAS_ext8u_i32; 1263 case INDEX_op_ext16u_i32: 1264 return TCG_TARGET_HAS_ext16u_i32; 1265 case INDEX_op_bswap16_i32: 1266 return TCG_TARGET_HAS_bswap16_i32; 1267 case INDEX_op_bswap32_i32: 1268 return TCG_TARGET_HAS_bswap32_i32; 1269 case INDEX_op_not_i32: 1270 return TCG_TARGET_HAS_not_i32; 1271 case INDEX_op_neg_i32: 1272 return TCG_TARGET_HAS_neg_i32; 1273 case INDEX_op_andc_i32: 1274 return TCG_TARGET_HAS_andc_i32; 1275 case INDEX_op_orc_i32: 1276 return TCG_TARGET_HAS_orc_i32; 1277 case INDEX_op_eqv_i32: 1278 return TCG_TARGET_HAS_eqv_i32; 1279 case INDEX_op_nand_i32: 1280 return TCG_TARGET_HAS_nand_i32; 1281 case INDEX_op_nor_i32: 1282 return TCG_TARGET_HAS_nor_i32; 1283 case INDEX_op_clz_i32: 1284 return TCG_TARGET_HAS_clz_i32; 1285 case INDEX_op_ctz_i32: 1286 return TCG_TARGET_HAS_ctz_i32; 1287 case INDEX_op_ctpop_i32: 1288 return TCG_TARGET_HAS_ctpop_i32; 1289 1290 case INDEX_op_brcond2_i32: 1291 case INDEX_op_setcond2_i32: 1292 return TCG_TARGET_REG_BITS == 32; 1293 1294 case INDEX_op_mov_i64: 1295 case INDEX_op_setcond_i64: 1296 case INDEX_op_brcond_i64: 1297 case INDEX_op_ld8u_i64: 1298 case INDEX_op_ld8s_i64: 1299 case INDEX_op_ld16u_i64: 1300 case INDEX_op_ld16s_i64: 1301 case INDEX_op_ld32u_i64: 1302 case INDEX_op_ld32s_i64: 1303 case INDEX_op_ld_i64: 1304 case INDEX_op_st8_i64: 1305 case INDEX_op_st16_i64: 1306 case INDEX_op_st32_i64: 1307 case INDEX_op_st_i64: 1308 case INDEX_op_add_i64: 1309 case INDEX_op_sub_i64: 1310 case INDEX_op_mul_i64: 1311 case INDEX_op_and_i64: 1312 case INDEX_op_or_i64: 1313 case INDEX_op_xor_i64: 1314 case INDEX_op_shl_i64: 1315 case INDEX_op_shr_i64: 1316 case INDEX_op_sar_i64: 1317 case INDEX_op_ext_i32_i64: 1318 case INDEX_op_extu_i32_i64: 1319 return TCG_TARGET_REG_BITS == 64; 1320 1321 case INDEX_op_movcond_i64: 1322 return TCG_TARGET_HAS_movcond_i64; 1323 case INDEX_op_div_i64: 1324 case INDEX_op_divu_i64: 1325 return TCG_TARGET_HAS_div_i64; 1326 case INDEX_op_rem_i64: 1327 case INDEX_op_remu_i64: 1328 return TCG_TARGET_HAS_rem_i64; 1329 case INDEX_op_div2_i64: 1330 case INDEX_op_divu2_i64: 1331 return TCG_TARGET_HAS_div2_i64; 1332 case INDEX_op_rotl_i64: 1333 case INDEX_op_rotr_i64: 1334 return TCG_TARGET_HAS_rot_i64; 1335 case INDEX_op_deposit_i64: 1336 return TCG_TARGET_HAS_deposit_i64; 1337 case INDEX_op_extract_i64: 1338 return TCG_TARGET_HAS_extract_i64; 1339 case INDEX_op_sextract_i64: 1340 return TCG_TARGET_HAS_sextract_i64; 1341 case INDEX_op_extract2_i64: 1342 return TCG_TARGET_HAS_extract2_i64; 1343 case INDEX_op_extrl_i64_i32: 1344 return TCG_TARGET_HAS_extrl_i64_i32; 1345 case INDEX_op_extrh_i64_i32: 1346 return TCG_TARGET_HAS_extrh_i64_i32; 1347 case INDEX_op_ext8s_i64: 1348 return TCG_TARGET_HAS_ext8s_i64; 1349 case INDEX_op_ext16s_i64: 1350 return TCG_TARGET_HAS_ext16s_i64; 1351 case INDEX_op_ext32s_i64: 1352 return TCG_TARGET_HAS_ext32s_i64; 1353 case INDEX_op_ext8u_i64: 1354 return TCG_TARGET_HAS_ext8u_i64; 1355 case INDEX_op_ext16u_i64: 1356 return TCG_TARGET_HAS_ext16u_i64; 1357 case INDEX_op_ext32u_i64: 1358 return TCG_TARGET_HAS_ext32u_i64; 1359 case INDEX_op_bswap16_i64: 1360 return TCG_TARGET_HAS_bswap16_i64; 1361 case INDEX_op_bswap32_i64: 1362 return TCG_TARGET_HAS_bswap32_i64; 1363 case INDEX_op_bswap64_i64: 1364 return TCG_TARGET_HAS_bswap64_i64; 1365 case INDEX_op_not_i64: 1366 return TCG_TARGET_HAS_not_i64; 1367 case INDEX_op_neg_i64: 1368 return TCG_TARGET_HAS_neg_i64; 1369 case INDEX_op_andc_i64: 1370 return TCG_TARGET_HAS_andc_i64; 1371 case INDEX_op_orc_i64: 1372 return TCG_TARGET_HAS_orc_i64; 1373 case INDEX_op_eqv_i64: 1374 return TCG_TARGET_HAS_eqv_i64; 1375 case INDEX_op_nand_i64: 1376 return TCG_TARGET_HAS_nand_i64; 1377 case INDEX_op_nor_i64: 1378 return TCG_TARGET_HAS_nor_i64; 1379 case INDEX_op_clz_i64: 1380 return TCG_TARGET_HAS_clz_i64; 1381 case INDEX_op_ctz_i64: 1382 return TCG_TARGET_HAS_ctz_i64; 1383 case INDEX_op_ctpop_i64: 1384 return TCG_TARGET_HAS_ctpop_i64; 1385 case INDEX_op_add2_i64: 1386 return TCG_TARGET_HAS_add2_i64; 1387 case INDEX_op_sub2_i64: 1388 return TCG_TARGET_HAS_sub2_i64; 1389 case INDEX_op_mulu2_i64: 1390 return TCG_TARGET_HAS_mulu2_i64; 1391 case INDEX_op_muls2_i64: 1392 return TCG_TARGET_HAS_muls2_i64; 1393 case INDEX_op_muluh_i64: 1394 return TCG_TARGET_HAS_muluh_i64; 1395 case INDEX_op_mulsh_i64: 1396 return TCG_TARGET_HAS_mulsh_i64; 1397 1398 case INDEX_op_mov_vec: 1399 case INDEX_op_dup_vec: 1400 case INDEX_op_dupm_vec: 1401 case INDEX_op_ld_vec: 1402 case INDEX_op_st_vec: 1403 case INDEX_op_add_vec: 1404 case INDEX_op_sub_vec: 1405 case INDEX_op_and_vec: 1406 case INDEX_op_or_vec: 1407 case INDEX_op_xor_vec: 1408 case INDEX_op_cmp_vec: 1409 return have_vec; 1410 case INDEX_op_dup2_vec: 1411 return have_vec && TCG_TARGET_REG_BITS == 32; 1412 case INDEX_op_not_vec: 1413 return have_vec && TCG_TARGET_HAS_not_vec; 1414 case INDEX_op_neg_vec: 1415 return have_vec && TCG_TARGET_HAS_neg_vec; 1416 case INDEX_op_abs_vec: 1417 return have_vec && TCG_TARGET_HAS_abs_vec; 1418 case INDEX_op_andc_vec: 1419 return have_vec && TCG_TARGET_HAS_andc_vec; 1420 case INDEX_op_orc_vec: 1421 return have_vec && TCG_TARGET_HAS_orc_vec; 1422 case INDEX_op_nand_vec: 1423 return have_vec && TCG_TARGET_HAS_nand_vec; 1424 case INDEX_op_nor_vec: 1425 return have_vec && TCG_TARGET_HAS_nor_vec; 1426 case INDEX_op_eqv_vec: 1427 return have_vec && TCG_TARGET_HAS_eqv_vec; 1428 case INDEX_op_mul_vec: 1429 return have_vec && TCG_TARGET_HAS_mul_vec; 1430 case INDEX_op_shli_vec: 1431 case INDEX_op_shri_vec: 1432 case INDEX_op_sari_vec: 1433 return have_vec && TCG_TARGET_HAS_shi_vec; 1434 case INDEX_op_shls_vec: 1435 case INDEX_op_shrs_vec: 1436 case INDEX_op_sars_vec: 1437 return have_vec && TCG_TARGET_HAS_shs_vec; 1438 case INDEX_op_shlv_vec: 1439 case INDEX_op_shrv_vec: 1440 case INDEX_op_sarv_vec: 1441 return have_vec && TCG_TARGET_HAS_shv_vec; 1442 case INDEX_op_rotli_vec: 1443 return have_vec && TCG_TARGET_HAS_roti_vec; 1444 case INDEX_op_rotls_vec: 1445 return have_vec && TCG_TARGET_HAS_rots_vec; 1446 case INDEX_op_rotlv_vec: 1447 case INDEX_op_rotrv_vec: 1448 return have_vec && TCG_TARGET_HAS_rotv_vec; 1449 case INDEX_op_ssadd_vec: 1450 case INDEX_op_usadd_vec: 1451 case INDEX_op_sssub_vec: 1452 case INDEX_op_ussub_vec: 1453 return have_vec && TCG_TARGET_HAS_sat_vec; 1454 case INDEX_op_smin_vec: 1455 case INDEX_op_umin_vec: 1456 case INDEX_op_smax_vec: 1457 case INDEX_op_umax_vec: 1458 return have_vec && TCG_TARGET_HAS_minmax_vec; 1459 case INDEX_op_bitsel_vec: 1460 return have_vec && TCG_TARGET_HAS_bitsel_vec; 1461 case INDEX_op_cmpsel_vec: 1462 return have_vec && TCG_TARGET_HAS_cmpsel_vec; 1463 1464 default: 1465 tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS); 1466 return true; 1467 } 1468 } 1469 1470 /* Note: we convert the 64 bit args to 32 bit and do some alignment 1471 and endian swap. Maybe it would be better to do the alignment 1472 and endian swap in tcg_reg_alloc_call(). */ 1473 void tcg_gen_callN(void *func, TCGTemp *ret, int nargs, TCGTemp **args) 1474 { 1475 int i, real_args, nb_rets, pi; 1476 unsigned typemask; 1477 const TCGHelperInfo *info; 1478 TCGOp *op; 1479 1480 info = g_hash_table_lookup(helper_table, (gpointer)func); 1481 typemask = info->typemask; 1482 1483 #ifdef CONFIG_PLUGIN 1484 /* detect non-plugin helpers */ 1485 if (tcg_ctx->plugin_insn && unlikely(strncmp(info->name, "plugin_", 7))) { 1486 tcg_ctx->plugin_insn->calls_helpers = true; 1487 } 1488 #endif 1489 1490 #if defined(TCG_TARGET_EXTEND_ARGS) && TCG_TARGET_REG_BITS == 64 1491 for (i = 0; i < nargs; ++i) { 1492 int argtype = extract32(typemask, (i + 1) * 3, 3); 1493 bool is_32bit = (argtype & ~1) == dh_typecode_i32; 1494 bool is_signed = argtype & 1; 1495 1496 if (is_32bit) { 1497 TCGv_i64 temp = tcg_temp_new_i64(); 1498 TCGv_i32 orig = temp_tcgv_i32(args[i]); 1499 if (is_signed) { 1500 tcg_gen_ext_i32_i64(temp, orig); 1501 } else { 1502 tcg_gen_extu_i32_i64(temp, orig); 1503 } 1504 args[i] = tcgv_i64_temp(temp); 1505 } 1506 } 1507 #endif /* TCG_TARGET_EXTEND_ARGS */ 1508 1509 op = tcg_emit_op(INDEX_op_call); 1510 1511 pi = 0; 1512 if (ret != NULL) { 1513 if (TCG_TARGET_REG_BITS < 64 && (typemask & 6) == dh_typecode_i64) { 1514 #if HOST_BIG_ENDIAN 1515 op->args[pi++] = temp_arg(ret + 1); 1516 op->args[pi++] = temp_arg(ret); 1517 #else 1518 op->args[pi++] = temp_arg(ret); 1519 op->args[pi++] = temp_arg(ret + 1); 1520 #endif 1521 nb_rets = 2; 1522 } else { 1523 op->args[pi++] = temp_arg(ret); 1524 nb_rets = 1; 1525 } 1526 } else { 1527 nb_rets = 0; 1528 } 1529 TCGOP_CALLO(op) = nb_rets; 1530 1531 real_args = 0; 1532 for (i = 0; i < nargs; i++) { 1533 int argtype = extract32(typemask, (i + 1) * 3, 3); 1534 bool is_64bit = (argtype & ~1) == dh_typecode_i64; 1535 bool want_align = false; 1536 1537 #if defined(CONFIG_TCG_INTERPRETER) 1538 /* 1539 * Align all arguments, so that they land in predictable places 1540 * for passing off to ffi_call. 1541 */ 1542 want_align = true; 1543 #elif defined(TCG_TARGET_CALL_ALIGN_ARGS) 1544 /* Some targets want aligned 64 bit args */ 1545 want_align = is_64bit; 1546 #endif 1547 1548 if (TCG_TARGET_REG_BITS < 64 && want_align && (real_args & 1)) { 1549 op->args[pi++] = TCG_CALL_DUMMY_ARG; 1550 real_args++; 1551 } 1552 1553 if (TCG_TARGET_REG_BITS < 64 && is_64bit) { 1554 /* 1555 * If stack grows up, then we will be placing successive 1556 * arguments at lower addresses, which means we need to 1557 * reverse the order compared to how we would normally 1558 * treat either big or little-endian. For those arguments 1559 * that will wind up in registers, this still works for 1560 * HPPA (the only current STACK_GROWSUP target) since the 1561 * argument registers are *also* allocated in decreasing 1562 * order. If another such target is added, this logic may 1563 * have to get more complicated to differentiate between 1564 * stack arguments and register arguments. 1565 */ 1566 #if HOST_BIG_ENDIAN != defined(TCG_TARGET_STACK_GROWSUP) 1567 op->args[pi++] = temp_arg(args[i] + 1); 1568 op->args[pi++] = temp_arg(args[i]); 1569 #else 1570 op->args[pi++] = temp_arg(args[i]); 1571 op->args[pi++] = temp_arg(args[i] + 1); 1572 #endif 1573 real_args += 2; 1574 continue; 1575 } 1576 1577 op->args[pi++] = temp_arg(args[i]); 1578 real_args++; 1579 } 1580 op->args[pi++] = (uintptr_t)func; 1581 op->args[pi++] = (uintptr_t)info; 1582 TCGOP_CALLI(op) = real_args; 1583 1584 /* Make sure the fields didn't overflow. */ 1585 tcg_debug_assert(TCGOP_CALLI(op) == real_args); 1586 tcg_debug_assert(pi <= ARRAY_SIZE(op->args)); 1587 1588 #if defined(TCG_TARGET_EXTEND_ARGS) && TCG_TARGET_REG_BITS == 64 1589 for (i = 0; i < nargs; ++i) { 1590 int argtype = extract32(typemask, (i + 1) * 3, 3); 1591 bool is_32bit = (argtype & ~1) == dh_typecode_i32; 1592 1593 if (is_32bit) { 1594 tcg_temp_free_internal(args[i]); 1595 } 1596 } 1597 #endif /* TCG_TARGET_EXTEND_ARGS */ 1598 } 1599 1600 static void tcg_reg_alloc_start(TCGContext *s) 1601 { 1602 int i, n; 1603 1604 for (i = 0, n = s->nb_temps; i < n; i++) { 1605 TCGTemp *ts = &s->temps[i]; 1606 TCGTempVal val = TEMP_VAL_MEM; 1607 1608 switch (ts->kind) { 1609 case TEMP_CONST: 1610 val = TEMP_VAL_CONST; 1611 break; 1612 case TEMP_FIXED: 1613 val = TEMP_VAL_REG; 1614 break; 1615 case TEMP_GLOBAL: 1616 break; 1617 case TEMP_NORMAL: 1618 case TEMP_EBB: 1619 val = TEMP_VAL_DEAD; 1620 /* fall through */ 1621 case TEMP_LOCAL: 1622 ts->mem_allocated = 0; 1623 break; 1624 default: 1625 g_assert_not_reached(); 1626 } 1627 ts->val_type = val; 1628 } 1629 1630 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp)); 1631 } 1632 1633 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size, 1634 TCGTemp *ts) 1635 { 1636 int idx = temp_idx(ts); 1637 1638 switch (ts->kind) { 1639 case TEMP_FIXED: 1640 case TEMP_GLOBAL: 1641 pstrcpy(buf, buf_size, ts->name); 1642 break; 1643 case TEMP_LOCAL: 1644 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals); 1645 break; 1646 case TEMP_EBB: 1647 snprintf(buf, buf_size, "ebb%d", idx - s->nb_globals); 1648 break; 1649 case TEMP_NORMAL: 1650 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals); 1651 break; 1652 case TEMP_CONST: 1653 switch (ts->type) { 1654 case TCG_TYPE_I32: 1655 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val); 1656 break; 1657 #if TCG_TARGET_REG_BITS > 32 1658 case TCG_TYPE_I64: 1659 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val); 1660 break; 1661 #endif 1662 case TCG_TYPE_V64: 1663 case TCG_TYPE_V128: 1664 case TCG_TYPE_V256: 1665 snprintf(buf, buf_size, "v%d$0x%" PRIx64, 1666 64 << (ts->type - TCG_TYPE_V64), ts->val); 1667 break; 1668 default: 1669 g_assert_not_reached(); 1670 } 1671 break; 1672 } 1673 return buf; 1674 } 1675 1676 static char *tcg_get_arg_str(TCGContext *s, char *buf, 1677 int buf_size, TCGArg arg) 1678 { 1679 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg)); 1680 } 1681 1682 static const char * const cond_name[] = 1683 { 1684 [TCG_COND_NEVER] = "never", 1685 [TCG_COND_ALWAYS] = "always", 1686 [TCG_COND_EQ] = "eq", 1687 [TCG_COND_NE] = "ne", 1688 [TCG_COND_LT] = "lt", 1689 [TCG_COND_GE] = "ge", 1690 [TCG_COND_LE] = "le", 1691 [TCG_COND_GT] = "gt", 1692 [TCG_COND_LTU] = "ltu", 1693 [TCG_COND_GEU] = "geu", 1694 [TCG_COND_LEU] = "leu", 1695 [TCG_COND_GTU] = "gtu" 1696 }; 1697 1698 static const char * const ldst_name[] = 1699 { 1700 [MO_UB] = "ub", 1701 [MO_SB] = "sb", 1702 [MO_LEUW] = "leuw", 1703 [MO_LESW] = "lesw", 1704 [MO_LEUL] = "leul", 1705 [MO_LESL] = "lesl", 1706 [MO_LEUQ] = "leq", 1707 [MO_BEUW] = "beuw", 1708 [MO_BESW] = "besw", 1709 [MO_BEUL] = "beul", 1710 [MO_BESL] = "besl", 1711 [MO_BEUQ] = "beq", 1712 }; 1713 1714 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = { 1715 #ifdef TARGET_ALIGNED_ONLY 1716 [MO_UNALN >> MO_ASHIFT] = "un+", 1717 [MO_ALIGN >> MO_ASHIFT] = "", 1718 #else 1719 [MO_UNALN >> MO_ASHIFT] = "", 1720 [MO_ALIGN >> MO_ASHIFT] = "al+", 1721 #endif 1722 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+", 1723 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+", 1724 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+", 1725 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+", 1726 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+", 1727 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+", 1728 }; 1729 1730 static const char bswap_flag_name[][6] = { 1731 [TCG_BSWAP_IZ] = "iz", 1732 [TCG_BSWAP_OZ] = "oz", 1733 [TCG_BSWAP_OS] = "os", 1734 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz", 1735 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os", 1736 }; 1737 1738 static inline bool tcg_regset_single(TCGRegSet d) 1739 { 1740 return (d & (d - 1)) == 0; 1741 } 1742 1743 static inline TCGReg tcg_regset_first(TCGRegSet d) 1744 { 1745 if (TCG_TARGET_NB_REGS <= 32) { 1746 return ctz32(d); 1747 } else { 1748 return ctz64(d); 1749 } 1750 } 1751 1752 /* Return only the number of characters output -- no error return. */ 1753 #define ne_fprintf(...) \ 1754 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; }) 1755 1756 static void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs) 1757 { 1758 char buf[128]; 1759 TCGOp *op; 1760 1761 QTAILQ_FOREACH(op, &s->ops, link) { 1762 int i, k, nb_oargs, nb_iargs, nb_cargs; 1763 const TCGOpDef *def; 1764 TCGOpcode c; 1765 int col = 0; 1766 1767 c = op->opc; 1768 def = &tcg_op_defs[c]; 1769 1770 if (c == INDEX_op_insn_start) { 1771 nb_oargs = 0; 1772 col += ne_fprintf(f, "\n ----"); 1773 1774 for (i = 0; i < TARGET_INSN_START_WORDS; ++i) { 1775 target_ulong a; 1776 #if TARGET_LONG_BITS > TCG_TARGET_REG_BITS 1777 a = deposit64(op->args[i * 2], 32, 32, op->args[i * 2 + 1]); 1778 #else 1779 a = op->args[i]; 1780 #endif 1781 col += ne_fprintf(f, " " TARGET_FMT_lx, a); 1782 } 1783 } else if (c == INDEX_op_call) { 1784 const TCGHelperInfo *info = tcg_call_info(op); 1785 void *func = tcg_call_func(op); 1786 1787 /* variable number of arguments */ 1788 nb_oargs = TCGOP_CALLO(op); 1789 nb_iargs = TCGOP_CALLI(op); 1790 nb_cargs = def->nb_cargs; 1791 1792 col += ne_fprintf(f, " %s ", def->name); 1793 1794 /* 1795 * Print the function name from TCGHelperInfo, if available. 1796 * Note that plugins have a template function for the info, 1797 * but the actual function pointer comes from the plugin. 1798 */ 1799 if (func == info->func) { 1800 col += ne_fprintf(f, "%s", info->name); 1801 } else { 1802 col += ne_fprintf(f, "plugin(%p)", func); 1803 } 1804 1805 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs); 1806 for (i = 0; i < nb_oargs; i++) { 1807 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf), 1808 op->args[i])); 1809 } 1810 for (i = 0; i < nb_iargs; i++) { 1811 TCGArg arg = op->args[nb_oargs + i]; 1812 const char *t = "<dummy>"; 1813 if (arg != TCG_CALL_DUMMY_ARG) { 1814 t = tcg_get_arg_str(s, buf, sizeof(buf), arg); 1815 } 1816 col += ne_fprintf(f, ",%s", t); 1817 } 1818 } else { 1819 col += ne_fprintf(f, " %s ", def->name); 1820 1821 nb_oargs = def->nb_oargs; 1822 nb_iargs = def->nb_iargs; 1823 nb_cargs = def->nb_cargs; 1824 1825 if (def->flags & TCG_OPF_VECTOR) { 1826 col += ne_fprintf(f, "v%d,e%d,", 64 << TCGOP_VECL(op), 1827 8 << TCGOP_VECE(op)); 1828 } 1829 1830 k = 0; 1831 for (i = 0; i < nb_oargs; i++) { 1832 const char *sep = k ? "," : ""; 1833 col += ne_fprintf(f, "%s%s", sep, 1834 tcg_get_arg_str(s, buf, sizeof(buf), 1835 op->args[k++])); 1836 } 1837 for (i = 0; i < nb_iargs; i++) { 1838 const char *sep = k ? "," : ""; 1839 col += ne_fprintf(f, "%s%s", sep, 1840 tcg_get_arg_str(s, buf, sizeof(buf), 1841 op->args[k++])); 1842 } 1843 switch (c) { 1844 case INDEX_op_brcond_i32: 1845 case INDEX_op_setcond_i32: 1846 case INDEX_op_movcond_i32: 1847 case INDEX_op_brcond2_i32: 1848 case INDEX_op_setcond2_i32: 1849 case INDEX_op_brcond_i64: 1850 case INDEX_op_setcond_i64: 1851 case INDEX_op_movcond_i64: 1852 case INDEX_op_cmp_vec: 1853 case INDEX_op_cmpsel_vec: 1854 if (op->args[k] < ARRAY_SIZE(cond_name) 1855 && cond_name[op->args[k]]) { 1856 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]); 1857 } else { 1858 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]); 1859 } 1860 i = 1; 1861 break; 1862 case INDEX_op_qemu_ld_i32: 1863 case INDEX_op_qemu_st_i32: 1864 case INDEX_op_qemu_st8_i32: 1865 case INDEX_op_qemu_ld_i64: 1866 case INDEX_op_qemu_st_i64: 1867 { 1868 MemOpIdx oi = op->args[k++]; 1869 MemOp op = get_memop(oi); 1870 unsigned ix = get_mmuidx(oi); 1871 1872 if (op & ~(MO_AMASK | MO_BSWAP | MO_SSIZE)) { 1873 col += ne_fprintf(f, ",$0x%x,%u", op, ix); 1874 } else { 1875 const char *s_al, *s_op; 1876 s_al = alignment_name[(op & MO_AMASK) >> MO_ASHIFT]; 1877 s_op = ldst_name[op & (MO_BSWAP | MO_SSIZE)]; 1878 col += ne_fprintf(f, ",%s%s,%u", s_al, s_op, ix); 1879 } 1880 i = 1; 1881 } 1882 break; 1883 case INDEX_op_bswap16_i32: 1884 case INDEX_op_bswap16_i64: 1885 case INDEX_op_bswap32_i32: 1886 case INDEX_op_bswap32_i64: 1887 case INDEX_op_bswap64_i64: 1888 { 1889 TCGArg flags = op->args[k]; 1890 const char *name = NULL; 1891 1892 if (flags < ARRAY_SIZE(bswap_flag_name)) { 1893 name = bswap_flag_name[flags]; 1894 } 1895 if (name) { 1896 col += ne_fprintf(f, ",%s", name); 1897 } else { 1898 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags); 1899 } 1900 i = k = 1; 1901 } 1902 break; 1903 default: 1904 i = 0; 1905 break; 1906 } 1907 switch (c) { 1908 case INDEX_op_set_label: 1909 case INDEX_op_br: 1910 case INDEX_op_brcond_i32: 1911 case INDEX_op_brcond_i64: 1912 case INDEX_op_brcond2_i32: 1913 col += ne_fprintf(f, "%s$L%d", k ? "," : "", 1914 arg_label(op->args[k])->id); 1915 i++, k++; 1916 break; 1917 default: 1918 break; 1919 } 1920 for (; i < nb_cargs; i++, k++) { 1921 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "", 1922 op->args[k]); 1923 } 1924 } 1925 1926 if (have_prefs || op->life) { 1927 for (; col < 40; ++col) { 1928 putc(' ', f); 1929 } 1930 } 1931 1932 if (op->life) { 1933 unsigned life = op->life; 1934 1935 if (life & (SYNC_ARG * 3)) { 1936 ne_fprintf(f, " sync:"); 1937 for (i = 0; i < 2; ++i) { 1938 if (life & (SYNC_ARG << i)) { 1939 ne_fprintf(f, " %d", i); 1940 } 1941 } 1942 } 1943 life /= DEAD_ARG; 1944 if (life) { 1945 ne_fprintf(f, " dead:"); 1946 for (i = 0; life; ++i, life >>= 1) { 1947 if (life & 1) { 1948 ne_fprintf(f, " %d", i); 1949 } 1950 } 1951 } 1952 } 1953 1954 if (have_prefs) { 1955 for (i = 0; i < nb_oargs; ++i) { 1956 TCGRegSet set = op->output_pref[i]; 1957 1958 if (i == 0) { 1959 ne_fprintf(f, " pref="); 1960 } else { 1961 ne_fprintf(f, ","); 1962 } 1963 if (set == 0) { 1964 ne_fprintf(f, "none"); 1965 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) { 1966 ne_fprintf(f, "all"); 1967 #ifdef CONFIG_DEBUG_TCG 1968 } else if (tcg_regset_single(set)) { 1969 TCGReg reg = tcg_regset_first(set); 1970 ne_fprintf(f, "%s", tcg_target_reg_names[reg]); 1971 #endif 1972 } else if (TCG_TARGET_NB_REGS <= 32) { 1973 ne_fprintf(f, "0x%x", (uint32_t)set); 1974 } else { 1975 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set); 1976 } 1977 } 1978 } 1979 1980 putc('\n', f); 1981 } 1982 } 1983 1984 /* we give more priority to constraints with less registers */ 1985 static int get_constraint_priority(const TCGOpDef *def, int k) 1986 { 1987 const TCGArgConstraint *arg_ct = &def->args_ct[k]; 1988 int n; 1989 1990 if (arg_ct->oalias) { 1991 /* an alias is equivalent to a single register */ 1992 n = 1; 1993 } else { 1994 n = ctpop64(arg_ct->regs); 1995 } 1996 return TCG_TARGET_NB_REGS - n + 1; 1997 } 1998 1999 /* sort from highest priority to lowest */ 2000 static void sort_constraints(TCGOpDef *def, int start, int n) 2001 { 2002 int i, j; 2003 TCGArgConstraint *a = def->args_ct; 2004 2005 for (i = 0; i < n; i++) { 2006 a[start + i].sort_index = start + i; 2007 } 2008 if (n <= 1) { 2009 return; 2010 } 2011 for (i = 0; i < n - 1; i++) { 2012 for (j = i + 1; j < n; j++) { 2013 int p1 = get_constraint_priority(def, a[start + i].sort_index); 2014 int p2 = get_constraint_priority(def, a[start + j].sort_index); 2015 if (p1 < p2) { 2016 int tmp = a[start + i].sort_index; 2017 a[start + i].sort_index = a[start + j].sort_index; 2018 a[start + j].sort_index = tmp; 2019 } 2020 } 2021 } 2022 } 2023 2024 static void process_op_defs(TCGContext *s) 2025 { 2026 TCGOpcode op; 2027 2028 for (op = 0; op < NB_OPS; op++) { 2029 TCGOpDef *def = &tcg_op_defs[op]; 2030 const TCGTargetOpDef *tdefs; 2031 int i, nb_args; 2032 2033 if (def->flags & TCG_OPF_NOT_PRESENT) { 2034 continue; 2035 } 2036 2037 nb_args = def->nb_iargs + def->nb_oargs; 2038 if (nb_args == 0) { 2039 continue; 2040 } 2041 2042 /* 2043 * Macro magic should make it impossible, but double-check that 2044 * the array index is in range. Since the signness of an enum 2045 * is implementation defined, force the result to unsigned. 2046 */ 2047 unsigned con_set = tcg_target_op_def(op); 2048 tcg_debug_assert(con_set < ARRAY_SIZE(constraint_sets)); 2049 tdefs = &constraint_sets[con_set]; 2050 2051 for (i = 0; i < nb_args; i++) { 2052 const char *ct_str = tdefs->args_ct_str[i]; 2053 /* Incomplete TCGTargetOpDef entry. */ 2054 tcg_debug_assert(ct_str != NULL); 2055 2056 while (*ct_str != '\0') { 2057 switch(*ct_str) { 2058 case '0' ... '9': 2059 { 2060 int oarg = *ct_str - '0'; 2061 tcg_debug_assert(ct_str == tdefs->args_ct_str[i]); 2062 tcg_debug_assert(oarg < def->nb_oargs); 2063 tcg_debug_assert(def->args_ct[oarg].regs != 0); 2064 def->args_ct[i] = def->args_ct[oarg]; 2065 /* The output sets oalias. */ 2066 def->args_ct[oarg].oalias = true; 2067 def->args_ct[oarg].alias_index = i; 2068 /* The input sets ialias. */ 2069 def->args_ct[i].ialias = true; 2070 def->args_ct[i].alias_index = oarg; 2071 } 2072 ct_str++; 2073 break; 2074 case '&': 2075 def->args_ct[i].newreg = true; 2076 ct_str++; 2077 break; 2078 case 'i': 2079 def->args_ct[i].ct |= TCG_CT_CONST; 2080 ct_str++; 2081 break; 2082 2083 /* Include all of the target-specific constraints. */ 2084 2085 #undef CONST 2086 #define CONST(CASE, MASK) \ 2087 case CASE: def->args_ct[i].ct |= MASK; ct_str++; break; 2088 #define REGS(CASE, MASK) \ 2089 case CASE: def->args_ct[i].regs |= MASK; ct_str++; break; 2090 2091 #include "tcg-target-con-str.h" 2092 2093 #undef REGS 2094 #undef CONST 2095 default: 2096 /* Typo in TCGTargetOpDef constraint. */ 2097 g_assert_not_reached(); 2098 } 2099 } 2100 } 2101 2102 /* TCGTargetOpDef entry with too much information? */ 2103 tcg_debug_assert(i == TCG_MAX_OP_ARGS || tdefs->args_ct_str[i] == NULL); 2104 2105 /* sort the constraints (XXX: this is just an heuristic) */ 2106 sort_constraints(def, 0, def->nb_oargs); 2107 sort_constraints(def, def->nb_oargs, def->nb_iargs); 2108 } 2109 } 2110 2111 void tcg_op_remove(TCGContext *s, TCGOp *op) 2112 { 2113 TCGLabel *label; 2114 2115 switch (op->opc) { 2116 case INDEX_op_br: 2117 label = arg_label(op->args[0]); 2118 label->refs--; 2119 break; 2120 case INDEX_op_brcond_i32: 2121 case INDEX_op_brcond_i64: 2122 label = arg_label(op->args[3]); 2123 label->refs--; 2124 break; 2125 case INDEX_op_brcond2_i32: 2126 label = arg_label(op->args[5]); 2127 label->refs--; 2128 break; 2129 default: 2130 break; 2131 } 2132 2133 QTAILQ_REMOVE(&s->ops, op, link); 2134 QTAILQ_INSERT_TAIL(&s->free_ops, op, link); 2135 s->nb_ops--; 2136 2137 #ifdef CONFIG_PROFILER 2138 qatomic_set(&s->prof.del_op_count, s->prof.del_op_count + 1); 2139 #endif 2140 } 2141 2142 void tcg_remove_ops_after(TCGOp *op) 2143 { 2144 TCGContext *s = tcg_ctx; 2145 2146 while (true) { 2147 TCGOp *last = tcg_last_op(); 2148 if (last == op) { 2149 return; 2150 } 2151 tcg_op_remove(s, last); 2152 } 2153 } 2154 2155 static TCGOp *tcg_op_alloc(TCGOpcode opc) 2156 { 2157 TCGContext *s = tcg_ctx; 2158 TCGOp *op; 2159 2160 if (likely(QTAILQ_EMPTY(&s->free_ops))) { 2161 op = tcg_malloc(sizeof(TCGOp)); 2162 } else { 2163 op = QTAILQ_FIRST(&s->free_ops); 2164 QTAILQ_REMOVE(&s->free_ops, op, link); 2165 } 2166 memset(op, 0, offsetof(TCGOp, link)); 2167 op->opc = opc; 2168 s->nb_ops++; 2169 2170 return op; 2171 } 2172 2173 TCGOp *tcg_emit_op(TCGOpcode opc) 2174 { 2175 TCGOp *op = tcg_op_alloc(opc); 2176 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 2177 return op; 2178 } 2179 2180 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op, TCGOpcode opc) 2181 { 2182 TCGOp *new_op = tcg_op_alloc(opc); 2183 QTAILQ_INSERT_BEFORE(old_op, new_op, link); 2184 return new_op; 2185 } 2186 2187 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op, TCGOpcode opc) 2188 { 2189 TCGOp *new_op = tcg_op_alloc(opc); 2190 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link); 2191 return new_op; 2192 } 2193 2194 /* Reachable analysis : remove unreachable code. */ 2195 static void reachable_code_pass(TCGContext *s) 2196 { 2197 TCGOp *op, *op_next; 2198 bool dead = false; 2199 2200 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 2201 bool remove = dead; 2202 TCGLabel *label; 2203 2204 switch (op->opc) { 2205 case INDEX_op_set_label: 2206 label = arg_label(op->args[0]); 2207 if (label->refs == 0) { 2208 /* 2209 * While there is an occasional backward branch, virtually 2210 * all branches generated by the translators are forward. 2211 * Which means that generally we will have already removed 2212 * all references to the label that will be, and there is 2213 * little to be gained by iterating. 2214 */ 2215 remove = true; 2216 } else { 2217 /* Once we see a label, insns become live again. */ 2218 dead = false; 2219 remove = false; 2220 2221 /* 2222 * Optimization can fold conditional branches to unconditional. 2223 * If we find a label with one reference which is preceded by 2224 * an unconditional branch to it, remove both. This needed to 2225 * wait until the dead code in between them was removed. 2226 */ 2227 if (label->refs == 1) { 2228 TCGOp *op_prev = QTAILQ_PREV(op, link); 2229 if (op_prev->opc == INDEX_op_br && 2230 label == arg_label(op_prev->args[0])) { 2231 tcg_op_remove(s, op_prev); 2232 remove = true; 2233 } 2234 } 2235 } 2236 break; 2237 2238 case INDEX_op_br: 2239 case INDEX_op_exit_tb: 2240 case INDEX_op_goto_ptr: 2241 /* Unconditional branches; everything following is dead. */ 2242 dead = true; 2243 break; 2244 2245 case INDEX_op_call: 2246 /* Notice noreturn helper calls, raising exceptions. */ 2247 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) { 2248 dead = true; 2249 } 2250 break; 2251 2252 case INDEX_op_insn_start: 2253 /* Never remove -- we need to keep these for unwind. */ 2254 remove = false; 2255 break; 2256 2257 default: 2258 break; 2259 } 2260 2261 if (remove) { 2262 tcg_op_remove(s, op); 2263 } 2264 } 2265 } 2266 2267 #define TS_DEAD 1 2268 #define TS_MEM 2 2269 2270 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n))) 2271 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n))) 2272 2273 /* For liveness_pass_1, the register preferences for a given temp. */ 2274 static inline TCGRegSet *la_temp_pref(TCGTemp *ts) 2275 { 2276 return ts->state_ptr; 2277 } 2278 2279 /* For liveness_pass_1, reset the preferences for a given temp to the 2280 * maximal regset for its type. 2281 */ 2282 static inline void la_reset_pref(TCGTemp *ts) 2283 { 2284 *la_temp_pref(ts) 2285 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]); 2286 } 2287 2288 /* liveness analysis: end of function: all temps are dead, and globals 2289 should be in memory. */ 2290 static void la_func_end(TCGContext *s, int ng, int nt) 2291 { 2292 int i; 2293 2294 for (i = 0; i < ng; ++i) { 2295 s->temps[i].state = TS_DEAD | TS_MEM; 2296 la_reset_pref(&s->temps[i]); 2297 } 2298 for (i = ng; i < nt; ++i) { 2299 s->temps[i].state = TS_DEAD; 2300 la_reset_pref(&s->temps[i]); 2301 } 2302 } 2303 2304 /* liveness analysis: end of basic block: all temps are dead, globals 2305 and local temps should be in memory. */ 2306 static void la_bb_end(TCGContext *s, int ng, int nt) 2307 { 2308 int i; 2309 2310 for (i = 0; i < nt; ++i) { 2311 TCGTemp *ts = &s->temps[i]; 2312 int state; 2313 2314 switch (ts->kind) { 2315 case TEMP_FIXED: 2316 case TEMP_GLOBAL: 2317 case TEMP_LOCAL: 2318 state = TS_DEAD | TS_MEM; 2319 break; 2320 case TEMP_NORMAL: 2321 case TEMP_EBB: 2322 case TEMP_CONST: 2323 state = TS_DEAD; 2324 break; 2325 default: 2326 g_assert_not_reached(); 2327 } 2328 ts->state = state; 2329 la_reset_pref(ts); 2330 } 2331 } 2332 2333 /* liveness analysis: sync globals back to memory. */ 2334 static void la_global_sync(TCGContext *s, int ng) 2335 { 2336 int i; 2337 2338 for (i = 0; i < ng; ++i) { 2339 int state = s->temps[i].state; 2340 s->temps[i].state = state | TS_MEM; 2341 if (state == TS_DEAD) { 2342 /* If the global was previously dead, reset prefs. */ 2343 la_reset_pref(&s->temps[i]); 2344 } 2345 } 2346 } 2347 2348 /* 2349 * liveness analysis: conditional branch: all temps are dead unless 2350 * explicitly live-across-conditional-branch, globals and local temps 2351 * should be synced. 2352 */ 2353 static void la_bb_sync(TCGContext *s, int ng, int nt) 2354 { 2355 la_global_sync(s, ng); 2356 2357 for (int i = ng; i < nt; ++i) { 2358 TCGTemp *ts = &s->temps[i]; 2359 int state; 2360 2361 switch (ts->kind) { 2362 case TEMP_LOCAL: 2363 state = ts->state; 2364 ts->state = state | TS_MEM; 2365 if (state != TS_DEAD) { 2366 continue; 2367 } 2368 break; 2369 case TEMP_NORMAL: 2370 s->temps[i].state = TS_DEAD; 2371 break; 2372 case TEMP_EBB: 2373 case TEMP_CONST: 2374 continue; 2375 default: 2376 g_assert_not_reached(); 2377 } 2378 la_reset_pref(&s->temps[i]); 2379 } 2380 } 2381 2382 /* liveness analysis: sync globals back to memory and kill. */ 2383 static void la_global_kill(TCGContext *s, int ng) 2384 { 2385 int i; 2386 2387 for (i = 0; i < ng; i++) { 2388 s->temps[i].state = TS_DEAD | TS_MEM; 2389 la_reset_pref(&s->temps[i]); 2390 } 2391 } 2392 2393 /* liveness analysis: note live globals crossing calls. */ 2394 static void la_cross_call(TCGContext *s, int nt) 2395 { 2396 TCGRegSet mask = ~tcg_target_call_clobber_regs; 2397 int i; 2398 2399 for (i = 0; i < nt; i++) { 2400 TCGTemp *ts = &s->temps[i]; 2401 if (!(ts->state & TS_DEAD)) { 2402 TCGRegSet *pset = la_temp_pref(ts); 2403 TCGRegSet set = *pset; 2404 2405 set &= mask; 2406 /* If the combination is not possible, restart. */ 2407 if (set == 0) { 2408 set = tcg_target_available_regs[ts->type] & mask; 2409 } 2410 *pset = set; 2411 } 2412 } 2413 } 2414 2415 /* Liveness analysis : update the opc_arg_life array to tell if a 2416 given input arguments is dead. Instructions updating dead 2417 temporaries are removed. */ 2418 static void liveness_pass_1(TCGContext *s) 2419 { 2420 int nb_globals = s->nb_globals; 2421 int nb_temps = s->nb_temps; 2422 TCGOp *op, *op_prev; 2423 TCGRegSet *prefs; 2424 int i; 2425 2426 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps); 2427 for (i = 0; i < nb_temps; ++i) { 2428 s->temps[i].state_ptr = prefs + i; 2429 } 2430 2431 /* ??? Should be redundant with the exit_tb that ends the TB. */ 2432 la_func_end(s, nb_globals, nb_temps); 2433 2434 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) { 2435 int nb_iargs, nb_oargs; 2436 TCGOpcode opc_new, opc_new2; 2437 bool have_opc_new2; 2438 TCGLifeData arg_life = 0; 2439 TCGTemp *ts; 2440 TCGOpcode opc = op->opc; 2441 const TCGOpDef *def = &tcg_op_defs[opc]; 2442 2443 switch (opc) { 2444 case INDEX_op_call: 2445 { 2446 int call_flags; 2447 int nb_call_regs; 2448 2449 nb_oargs = TCGOP_CALLO(op); 2450 nb_iargs = TCGOP_CALLI(op); 2451 call_flags = tcg_call_flags(op); 2452 2453 /* pure functions can be removed if their result is unused */ 2454 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) { 2455 for (i = 0; i < nb_oargs; i++) { 2456 ts = arg_temp(op->args[i]); 2457 if (ts->state != TS_DEAD) { 2458 goto do_not_remove_call; 2459 } 2460 } 2461 goto do_remove; 2462 } 2463 do_not_remove_call: 2464 2465 /* Output args are dead. */ 2466 for (i = 0; i < nb_oargs; i++) { 2467 ts = arg_temp(op->args[i]); 2468 if (ts->state & TS_DEAD) { 2469 arg_life |= DEAD_ARG << i; 2470 } 2471 if (ts->state & TS_MEM) { 2472 arg_life |= SYNC_ARG << i; 2473 } 2474 ts->state = TS_DEAD; 2475 la_reset_pref(ts); 2476 2477 /* Not used -- it will be tcg_target_call_oarg_regs[i]. */ 2478 op->output_pref[i] = 0; 2479 } 2480 2481 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS | 2482 TCG_CALL_NO_READ_GLOBALS))) { 2483 la_global_kill(s, nb_globals); 2484 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) { 2485 la_global_sync(s, nb_globals); 2486 } 2487 2488 /* Record arguments that die in this helper. */ 2489 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 2490 ts = arg_temp(op->args[i]); 2491 if (ts && ts->state & TS_DEAD) { 2492 arg_life |= DEAD_ARG << i; 2493 } 2494 } 2495 2496 /* For all live registers, remove call-clobbered prefs. */ 2497 la_cross_call(s, nb_temps); 2498 2499 nb_call_regs = ARRAY_SIZE(tcg_target_call_iarg_regs); 2500 2501 /* Input arguments are live for preceding opcodes. */ 2502 for (i = 0; i < nb_iargs; i++) { 2503 ts = arg_temp(op->args[i + nb_oargs]); 2504 if (ts && ts->state & TS_DEAD) { 2505 /* For those arguments that die, and will be allocated 2506 * in registers, clear the register set for that arg, 2507 * to be filled in below. For args that will be on 2508 * the stack, reset to any available reg. 2509 */ 2510 *la_temp_pref(ts) 2511 = (i < nb_call_regs ? 0 : 2512 tcg_target_available_regs[ts->type]); 2513 ts->state &= ~TS_DEAD; 2514 } 2515 } 2516 2517 /* For each input argument, add its input register to prefs. 2518 If a temp is used once, this produces a single set bit. */ 2519 for (i = 0; i < MIN(nb_call_regs, nb_iargs); i++) { 2520 ts = arg_temp(op->args[i + nb_oargs]); 2521 if (ts) { 2522 tcg_regset_set_reg(*la_temp_pref(ts), 2523 tcg_target_call_iarg_regs[i]); 2524 } 2525 } 2526 } 2527 break; 2528 case INDEX_op_insn_start: 2529 break; 2530 case INDEX_op_discard: 2531 /* mark the temporary as dead */ 2532 ts = arg_temp(op->args[0]); 2533 ts->state = TS_DEAD; 2534 la_reset_pref(ts); 2535 break; 2536 2537 case INDEX_op_add2_i32: 2538 opc_new = INDEX_op_add_i32; 2539 goto do_addsub2; 2540 case INDEX_op_sub2_i32: 2541 opc_new = INDEX_op_sub_i32; 2542 goto do_addsub2; 2543 case INDEX_op_add2_i64: 2544 opc_new = INDEX_op_add_i64; 2545 goto do_addsub2; 2546 case INDEX_op_sub2_i64: 2547 opc_new = INDEX_op_sub_i64; 2548 do_addsub2: 2549 nb_iargs = 4; 2550 nb_oargs = 2; 2551 /* Test if the high part of the operation is dead, but not 2552 the low part. The result can be optimized to a simple 2553 add or sub. This happens often for x86_64 guest when the 2554 cpu mode is set to 32 bit. */ 2555 if (arg_temp(op->args[1])->state == TS_DEAD) { 2556 if (arg_temp(op->args[0])->state == TS_DEAD) { 2557 goto do_remove; 2558 } 2559 /* Replace the opcode and adjust the args in place, 2560 leaving 3 unused args at the end. */ 2561 op->opc = opc = opc_new; 2562 op->args[1] = op->args[2]; 2563 op->args[2] = op->args[4]; 2564 /* Fall through and mark the single-word operation live. */ 2565 nb_iargs = 2; 2566 nb_oargs = 1; 2567 } 2568 goto do_not_remove; 2569 2570 case INDEX_op_mulu2_i32: 2571 opc_new = INDEX_op_mul_i32; 2572 opc_new2 = INDEX_op_muluh_i32; 2573 have_opc_new2 = TCG_TARGET_HAS_muluh_i32; 2574 goto do_mul2; 2575 case INDEX_op_muls2_i32: 2576 opc_new = INDEX_op_mul_i32; 2577 opc_new2 = INDEX_op_mulsh_i32; 2578 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32; 2579 goto do_mul2; 2580 case INDEX_op_mulu2_i64: 2581 opc_new = INDEX_op_mul_i64; 2582 opc_new2 = INDEX_op_muluh_i64; 2583 have_opc_new2 = TCG_TARGET_HAS_muluh_i64; 2584 goto do_mul2; 2585 case INDEX_op_muls2_i64: 2586 opc_new = INDEX_op_mul_i64; 2587 opc_new2 = INDEX_op_mulsh_i64; 2588 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64; 2589 goto do_mul2; 2590 do_mul2: 2591 nb_iargs = 2; 2592 nb_oargs = 2; 2593 if (arg_temp(op->args[1])->state == TS_DEAD) { 2594 if (arg_temp(op->args[0])->state == TS_DEAD) { 2595 /* Both parts of the operation are dead. */ 2596 goto do_remove; 2597 } 2598 /* The high part of the operation is dead; generate the low. */ 2599 op->opc = opc = opc_new; 2600 op->args[1] = op->args[2]; 2601 op->args[2] = op->args[3]; 2602 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) { 2603 /* The low part of the operation is dead; generate the high. */ 2604 op->opc = opc = opc_new2; 2605 op->args[0] = op->args[1]; 2606 op->args[1] = op->args[2]; 2607 op->args[2] = op->args[3]; 2608 } else { 2609 goto do_not_remove; 2610 } 2611 /* Mark the single-word operation live. */ 2612 nb_oargs = 1; 2613 goto do_not_remove; 2614 2615 default: 2616 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */ 2617 nb_iargs = def->nb_iargs; 2618 nb_oargs = def->nb_oargs; 2619 2620 /* Test if the operation can be removed because all 2621 its outputs are dead. We assume that nb_oargs == 0 2622 implies side effects */ 2623 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { 2624 for (i = 0; i < nb_oargs; i++) { 2625 if (arg_temp(op->args[i])->state != TS_DEAD) { 2626 goto do_not_remove; 2627 } 2628 } 2629 goto do_remove; 2630 } 2631 goto do_not_remove; 2632 2633 do_remove: 2634 tcg_op_remove(s, op); 2635 break; 2636 2637 do_not_remove: 2638 for (i = 0; i < nb_oargs; i++) { 2639 ts = arg_temp(op->args[i]); 2640 2641 /* Remember the preference of the uses that followed. */ 2642 op->output_pref[i] = *la_temp_pref(ts); 2643 2644 /* Output args are dead. */ 2645 if (ts->state & TS_DEAD) { 2646 arg_life |= DEAD_ARG << i; 2647 } 2648 if (ts->state & TS_MEM) { 2649 arg_life |= SYNC_ARG << i; 2650 } 2651 ts->state = TS_DEAD; 2652 la_reset_pref(ts); 2653 } 2654 2655 /* If end of basic block, update. */ 2656 if (def->flags & TCG_OPF_BB_EXIT) { 2657 la_func_end(s, nb_globals, nb_temps); 2658 } else if (def->flags & TCG_OPF_COND_BRANCH) { 2659 la_bb_sync(s, nb_globals, nb_temps); 2660 } else if (def->flags & TCG_OPF_BB_END) { 2661 la_bb_end(s, nb_globals, nb_temps); 2662 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 2663 la_global_sync(s, nb_globals); 2664 if (def->flags & TCG_OPF_CALL_CLOBBER) { 2665 la_cross_call(s, nb_temps); 2666 } 2667 } 2668 2669 /* Record arguments that die in this opcode. */ 2670 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 2671 ts = arg_temp(op->args[i]); 2672 if (ts->state & TS_DEAD) { 2673 arg_life |= DEAD_ARG << i; 2674 } 2675 } 2676 2677 /* Input arguments are live for preceding opcodes. */ 2678 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 2679 ts = arg_temp(op->args[i]); 2680 if (ts->state & TS_DEAD) { 2681 /* For operands that were dead, initially allow 2682 all regs for the type. */ 2683 *la_temp_pref(ts) = tcg_target_available_regs[ts->type]; 2684 ts->state &= ~TS_DEAD; 2685 } 2686 } 2687 2688 /* Incorporate constraints for this operand. */ 2689 switch (opc) { 2690 case INDEX_op_mov_i32: 2691 case INDEX_op_mov_i64: 2692 /* Note that these are TCG_OPF_NOT_PRESENT and do not 2693 have proper constraints. That said, special case 2694 moves to propagate preferences backward. */ 2695 if (IS_DEAD_ARG(1)) { 2696 *la_temp_pref(arg_temp(op->args[0])) 2697 = *la_temp_pref(arg_temp(op->args[1])); 2698 } 2699 break; 2700 2701 default: 2702 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 2703 const TCGArgConstraint *ct = &def->args_ct[i]; 2704 TCGRegSet set, *pset; 2705 2706 ts = arg_temp(op->args[i]); 2707 pset = la_temp_pref(ts); 2708 set = *pset; 2709 2710 set &= ct->regs; 2711 if (ct->ialias) { 2712 set &= op->output_pref[ct->alias_index]; 2713 } 2714 /* If the combination is not possible, restart. */ 2715 if (set == 0) { 2716 set = ct->regs; 2717 } 2718 *pset = set; 2719 } 2720 break; 2721 } 2722 break; 2723 } 2724 op->life = arg_life; 2725 } 2726 } 2727 2728 /* Liveness analysis: Convert indirect regs to direct temporaries. */ 2729 static bool liveness_pass_2(TCGContext *s) 2730 { 2731 int nb_globals = s->nb_globals; 2732 int nb_temps, i; 2733 bool changes = false; 2734 TCGOp *op, *op_next; 2735 2736 /* Create a temporary for each indirect global. */ 2737 for (i = 0; i < nb_globals; ++i) { 2738 TCGTemp *its = &s->temps[i]; 2739 if (its->indirect_reg) { 2740 TCGTemp *dts = tcg_temp_alloc(s); 2741 dts->type = its->type; 2742 dts->base_type = its->base_type; 2743 dts->kind = TEMP_EBB; 2744 its->state_ptr = dts; 2745 } else { 2746 its->state_ptr = NULL; 2747 } 2748 /* All globals begin dead. */ 2749 its->state = TS_DEAD; 2750 } 2751 for (nb_temps = s->nb_temps; i < nb_temps; ++i) { 2752 TCGTemp *its = &s->temps[i]; 2753 its->state_ptr = NULL; 2754 its->state = TS_DEAD; 2755 } 2756 2757 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 2758 TCGOpcode opc = op->opc; 2759 const TCGOpDef *def = &tcg_op_defs[opc]; 2760 TCGLifeData arg_life = op->life; 2761 int nb_iargs, nb_oargs, call_flags; 2762 TCGTemp *arg_ts, *dir_ts; 2763 2764 if (opc == INDEX_op_call) { 2765 nb_oargs = TCGOP_CALLO(op); 2766 nb_iargs = TCGOP_CALLI(op); 2767 call_flags = tcg_call_flags(op); 2768 } else { 2769 nb_iargs = def->nb_iargs; 2770 nb_oargs = def->nb_oargs; 2771 2772 /* Set flags similar to how calls require. */ 2773 if (def->flags & TCG_OPF_COND_BRANCH) { 2774 /* Like reading globals: sync_globals */ 2775 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 2776 } else if (def->flags & TCG_OPF_BB_END) { 2777 /* Like writing globals: save_globals */ 2778 call_flags = 0; 2779 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 2780 /* Like reading globals: sync_globals */ 2781 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 2782 } else { 2783 /* No effect on globals. */ 2784 call_flags = (TCG_CALL_NO_READ_GLOBALS | 2785 TCG_CALL_NO_WRITE_GLOBALS); 2786 } 2787 } 2788 2789 /* Make sure that input arguments are available. */ 2790 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 2791 arg_ts = arg_temp(op->args[i]); 2792 if (arg_ts) { 2793 dir_ts = arg_ts->state_ptr; 2794 if (dir_ts && arg_ts->state == TS_DEAD) { 2795 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32 2796 ? INDEX_op_ld_i32 2797 : INDEX_op_ld_i64); 2798 TCGOp *lop = tcg_op_insert_before(s, op, lopc); 2799 2800 lop->args[0] = temp_arg(dir_ts); 2801 lop->args[1] = temp_arg(arg_ts->mem_base); 2802 lop->args[2] = arg_ts->mem_offset; 2803 2804 /* Loaded, but synced with memory. */ 2805 arg_ts->state = TS_MEM; 2806 } 2807 } 2808 } 2809 2810 /* Perform input replacement, and mark inputs that became dead. 2811 No action is required except keeping temp_state up to date 2812 so that we reload when needed. */ 2813 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 2814 arg_ts = arg_temp(op->args[i]); 2815 if (arg_ts) { 2816 dir_ts = arg_ts->state_ptr; 2817 if (dir_ts) { 2818 op->args[i] = temp_arg(dir_ts); 2819 changes = true; 2820 if (IS_DEAD_ARG(i)) { 2821 arg_ts->state = TS_DEAD; 2822 } 2823 } 2824 } 2825 } 2826 2827 /* Liveness analysis should ensure that the following are 2828 all correct, for call sites and basic block end points. */ 2829 if (call_flags & TCG_CALL_NO_READ_GLOBALS) { 2830 /* Nothing to do */ 2831 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) { 2832 for (i = 0; i < nb_globals; ++i) { 2833 /* Liveness should see that globals are synced back, 2834 that is, either TS_DEAD or TS_MEM. */ 2835 arg_ts = &s->temps[i]; 2836 tcg_debug_assert(arg_ts->state_ptr == 0 2837 || arg_ts->state != 0); 2838 } 2839 } else { 2840 for (i = 0; i < nb_globals; ++i) { 2841 /* Liveness should see that globals are saved back, 2842 that is, TS_DEAD, waiting to be reloaded. */ 2843 arg_ts = &s->temps[i]; 2844 tcg_debug_assert(arg_ts->state_ptr == 0 2845 || arg_ts->state == TS_DEAD); 2846 } 2847 } 2848 2849 /* Outputs become available. */ 2850 if (opc == INDEX_op_mov_i32 || opc == INDEX_op_mov_i64) { 2851 arg_ts = arg_temp(op->args[0]); 2852 dir_ts = arg_ts->state_ptr; 2853 if (dir_ts) { 2854 op->args[0] = temp_arg(dir_ts); 2855 changes = true; 2856 2857 /* The output is now live and modified. */ 2858 arg_ts->state = 0; 2859 2860 if (NEED_SYNC_ARG(0)) { 2861 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 2862 ? INDEX_op_st_i32 2863 : INDEX_op_st_i64); 2864 TCGOp *sop = tcg_op_insert_after(s, op, sopc); 2865 TCGTemp *out_ts = dir_ts; 2866 2867 if (IS_DEAD_ARG(0)) { 2868 out_ts = arg_temp(op->args[1]); 2869 arg_ts->state = TS_DEAD; 2870 tcg_op_remove(s, op); 2871 } else { 2872 arg_ts->state = TS_MEM; 2873 } 2874 2875 sop->args[0] = temp_arg(out_ts); 2876 sop->args[1] = temp_arg(arg_ts->mem_base); 2877 sop->args[2] = arg_ts->mem_offset; 2878 } else { 2879 tcg_debug_assert(!IS_DEAD_ARG(0)); 2880 } 2881 } 2882 } else { 2883 for (i = 0; i < nb_oargs; i++) { 2884 arg_ts = arg_temp(op->args[i]); 2885 dir_ts = arg_ts->state_ptr; 2886 if (!dir_ts) { 2887 continue; 2888 } 2889 op->args[i] = temp_arg(dir_ts); 2890 changes = true; 2891 2892 /* The output is now live and modified. */ 2893 arg_ts->state = 0; 2894 2895 /* Sync outputs upon their last write. */ 2896 if (NEED_SYNC_ARG(i)) { 2897 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 2898 ? INDEX_op_st_i32 2899 : INDEX_op_st_i64); 2900 TCGOp *sop = tcg_op_insert_after(s, op, sopc); 2901 2902 sop->args[0] = temp_arg(dir_ts); 2903 sop->args[1] = temp_arg(arg_ts->mem_base); 2904 sop->args[2] = arg_ts->mem_offset; 2905 2906 arg_ts->state = TS_MEM; 2907 } 2908 /* Drop outputs that are dead. */ 2909 if (IS_DEAD_ARG(i)) { 2910 arg_ts->state = TS_DEAD; 2911 } 2912 } 2913 } 2914 } 2915 2916 return changes; 2917 } 2918 2919 #ifdef CONFIG_DEBUG_TCG 2920 static void dump_regs(TCGContext *s) 2921 { 2922 TCGTemp *ts; 2923 int i; 2924 char buf[64]; 2925 2926 for(i = 0; i < s->nb_temps; i++) { 2927 ts = &s->temps[i]; 2928 printf(" %10s: ", tcg_get_arg_str_ptr(s, buf, sizeof(buf), ts)); 2929 switch(ts->val_type) { 2930 case TEMP_VAL_REG: 2931 printf("%s", tcg_target_reg_names[ts->reg]); 2932 break; 2933 case TEMP_VAL_MEM: 2934 printf("%d(%s)", (int)ts->mem_offset, 2935 tcg_target_reg_names[ts->mem_base->reg]); 2936 break; 2937 case TEMP_VAL_CONST: 2938 printf("$0x%" PRIx64, ts->val); 2939 break; 2940 case TEMP_VAL_DEAD: 2941 printf("D"); 2942 break; 2943 default: 2944 printf("???"); 2945 break; 2946 } 2947 printf("\n"); 2948 } 2949 2950 for(i = 0; i < TCG_TARGET_NB_REGS; i++) { 2951 if (s->reg_to_temp[i] != NULL) { 2952 printf("%s: %s\n", 2953 tcg_target_reg_names[i], 2954 tcg_get_arg_str_ptr(s, buf, sizeof(buf), s->reg_to_temp[i])); 2955 } 2956 } 2957 } 2958 2959 static void check_regs(TCGContext *s) 2960 { 2961 int reg; 2962 int k; 2963 TCGTemp *ts; 2964 char buf[64]; 2965 2966 for (reg = 0; reg < TCG_TARGET_NB_REGS; reg++) { 2967 ts = s->reg_to_temp[reg]; 2968 if (ts != NULL) { 2969 if (ts->val_type != TEMP_VAL_REG || ts->reg != reg) { 2970 printf("Inconsistency for register %s:\n", 2971 tcg_target_reg_names[reg]); 2972 goto fail; 2973 } 2974 } 2975 } 2976 for (k = 0; k < s->nb_temps; k++) { 2977 ts = &s->temps[k]; 2978 if (ts->val_type == TEMP_VAL_REG 2979 && ts->kind != TEMP_FIXED 2980 && s->reg_to_temp[ts->reg] != ts) { 2981 printf("Inconsistency for temp %s:\n", 2982 tcg_get_arg_str_ptr(s, buf, sizeof(buf), ts)); 2983 fail: 2984 printf("reg state:\n"); 2985 dump_regs(s); 2986 tcg_abort(); 2987 } 2988 } 2989 } 2990 #endif 2991 2992 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts) 2993 { 2994 intptr_t off, size, align; 2995 2996 switch (ts->type) { 2997 case TCG_TYPE_I32: 2998 size = align = 4; 2999 break; 3000 case TCG_TYPE_I64: 3001 case TCG_TYPE_V64: 3002 size = align = 8; 3003 break; 3004 case TCG_TYPE_V128: 3005 size = align = 16; 3006 break; 3007 case TCG_TYPE_V256: 3008 /* Note that we do not require aligned storage for V256. */ 3009 size = 32, align = 16; 3010 break; 3011 default: 3012 g_assert_not_reached(); 3013 } 3014 3015 /* 3016 * Assume the stack is sufficiently aligned. 3017 * This affects e.g. ARM NEON, where we have 8 byte stack alignment 3018 * and do not require 16 byte vector alignment. This seems slightly 3019 * easier than fully parameterizing the above switch statement. 3020 */ 3021 align = MIN(TCG_TARGET_STACK_ALIGN, align); 3022 off = ROUND_UP(s->current_frame_offset, align); 3023 3024 /* If we've exhausted the stack frame, restart with a smaller TB. */ 3025 if (off + size > s->frame_end) { 3026 tcg_raise_tb_overflow(s); 3027 } 3028 s->current_frame_offset = off + size; 3029 3030 ts->mem_offset = off; 3031 #if defined(__sparc__) 3032 ts->mem_offset += TCG_TARGET_STACK_BIAS; 3033 #endif 3034 ts->mem_base = s->frame_temp; 3035 ts->mem_allocated = 1; 3036 } 3037 3038 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet); 3039 3040 /* Mark a temporary as free or dead. If 'free_or_dead' is negative, 3041 mark it free; otherwise mark it dead. */ 3042 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead) 3043 { 3044 TCGTempVal new_type; 3045 3046 switch (ts->kind) { 3047 case TEMP_FIXED: 3048 return; 3049 case TEMP_GLOBAL: 3050 case TEMP_LOCAL: 3051 new_type = TEMP_VAL_MEM; 3052 break; 3053 case TEMP_NORMAL: 3054 case TEMP_EBB: 3055 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD; 3056 break; 3057 case TEMP_CONST: 3058 new_type = TEMP_VAL_CONST; 3059 break; 3060 default: 3061 g_assert_not_reached(); 3062 } 3063 if (ts->val_type == TEMP_VAL_REG) { 3064 s->reg_to_temp[ts->reg] = NULL; 3065 } 3066 ts->val_type = new_type; 3067 } 3068 3069 /* Mark a temporary as dead. */ 3070 static inline void temp_dead(TCGContext *s, TCGTemp *ts) 3071 { 3072 temp_free_or_dead(s, ts, 1); 3073 } 3074 3075 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary 3076 registers needs to be allocated to store a constant. If 'free_or_dead' 3077 is non-zero, subsequently release the temporary; if it is positive, the 3078 temp is dead; if it is negative, the temp is free. */ 3079 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs, 3080 TCGRegSet preferred_regs, int free_or_dead) 3081 { 3082 if (!temp_readonly(ts) && !ts->mem_coherent) { 3083 if (!ts->mem_allocated) { 3084 temp_allocate_frame(s, ts); 3085 } 3086 switch (ts->val_type) { 3087 case TEMP_VAL_CONST: 3088 /* If we're going to free the temp immediately, then we won't 3089 require it later in a register, so attempt to store the 3090 constant to memory directly. */ 3091 if (free_or_dead 3092 && tcg_out_sti(s, ts->type, ts->val, 3093 ts->mem_base->reg, ts->mem_offset)) { 3094 break; 3095 } 3096 temp_load(s, ts, tcg_target_available_regs[ts->type], 3097 allocated_regs, preferred_regs); 3098 /* fallthrough */ 3099 3100 case TEMP_VAL_REG: 3101 tcg_out_st(s, ts->type, ts->reg, 3102 ts->mem_base->reg, ts->mem_offset); 3103 break; 3104 3105 case TEMP_VAL_MEM: 3106 break; 3107 3108 case TEMP_VAL_DEAD: 3109 default: 3110 tcg_abort(); 3111 } 3112 ts->mem_coherent = 1; 3113 } 3114 if (free_or_dead) { 3115 temp_free_or_dead(s, ts, free_or_dead); 3116 } 3117 } 3118 3119 /* free register 'reg' by spilling the corresponding temporary if necessary */ 3120 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs) 3121 { 3122 TCGTemp *ts = s->reg_to_temp[reg]; 3123 if (ts != NULL) { 3124 temp_sync(s, ts, allocated_regs, 0, -1); 3125 } 3126 } 3127 3128 /** 3129 * tcg_reg_alloc: 3130 * @required_regs: Set of registers in which we must allocate. 3131 * @allocated_regs: Set of registers which must be avoided. 3132 * @preferred_regs: Set of registers we should prefer. 3133 * @rev: True if we search the registers in "indirect" order. 3134 * 3135 * The allocated register must be in @required_regs & ~@allocated_regs, 3136 * but if we can put it in @preferred_regs we may save a move later. 3137 */ 3138 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs, 3139 TCGRegSet allocated_regs, 3140 TCGRegSet preferred_regs, bool rev) 3141 { 3142 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 3143 TCGRegSet reg_ct[2]; 3144 const int *order; 3145 3146 reg_ct[1] = required_regs & ~allocated_regs; 3147 tcg_debug_assert(reg_ct[1] != 0); 3148 reg_ct[0] = reg_ct[1] & preferred_regs; 3149 3150 /* Skip the preferred_regs option if it cannot be satisfied, 3151 or if the preference made no difference. */ 3152 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 3153 3154 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 3155 3156 /* Try free registers, preferences first. */ 3157 for (j = f; j < 2; j++) { 3158 TCGRegSet set = reg_ct[j]; 3159 3160 if (tcg_regset_single(set)) { 3161 /* One register in the set. */ 3162 TCGReg reg = tcg_regset_first(set); 3163 if (s->reg_to_temp[reg] == NULL) { 3164 return reg; 3165 } 3166 } else { 3167 for (i = 0; i < n; i++) { 3168 TCGReg reg = order[i]; 3169 if (s->reg_to_temp[reg] == NULL && 3170 tcg_regset_test_reg(set, reg)) { 3171 return reg; 3172 } 3173 } 3174 } 3175 } 3176 3177 /* We must spill something. */ 3178 for (j = f; j < 2; j++) { 3179 TCGRegSet set = reg_ct[j]; 3180 3181 if (tcg_regset_single(set)) { 3182 /* One register in the set. */ 3183 TCGReg reg = tcg_regset_first(set); 3184 tcg_reg_free(s, reg, allocated_regs); 3185 return reg; 3186 } else { 3187 for (i = 0; i < n; i++) { 3188 TCGReg reg = order[i]; 3189 if (tcg_regset_test_reg(set, reg)) { 3190 tcg_reg_free(s, reg, allocated_regs); 3191 return reg; 3192 } 3193 } 3194 } 3195 } 3196 3197 tcg_abort(); 3198 } 3199 3200 /* Make sure the temporary is in a register. If needed, allocate the register 3201 from DESIRED while avoiding ALLOCATED. */ 3202 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs, 3203 TCGRegSet allocated_regs, TCGRegSet preferred_regs) 3204 { 3205 TCGReg reg; 3206 3207 switch (ts->val_type) { 3208 case TEMP_VAL_REG: 3209 return; 3210 case TEMP_VAL_CONST: 3211 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 3212 preferred_regs, ts->indirect_base); 3213 if (ts->type <= TCG_TYPE_I64) { 3214 tcg_out_movi(s, ts->type, reg, ts->val); 3215 } else { 3216 uint64_t val = ts->val; 3217 MemOp vece = MO_64; 3218 3219 /* 3220 * Find the minimal vector element that matches the constant. 3221 * The targets will, in general, have to do this search anyway, 3222 * do this generically. 3223 */ 3224 if (val == dup_const(MO_8, val)) { 3225 vece = MO_8; 3226 } else if (val == dup_const(MO_16, val)) { 3227 vece = MO_16; 3228 } else if (val == dup_const(MO_32, val)) { 3229 vece = MO_32; 3230 } 3231 3232 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val); 3233 } 3234 ts->mem_coherent = 0; 3235 break; 3236 case TEMP_VAL_MEM: 3237 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 3238 preferred_regs, ts->indirect_base); 3239 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset); 3240 ts->mem_coherent = 1; 3241 break; 3242 case TEMP_VAL_DEAD: 3243 default: 3244 tcg_abort(); 3245 } 3246 ts->reg = reg; 3247 ts->val_type = TEMP_VAL_REG; 3248 s->reg_to_temp[reg] = ts; 3249 } 3250 3251 /* Save a temporary to memory. 'allocated_regs' is used in case a 3252 temporary registers needs to be allocated to store a constant. */ 3253 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs) 3254 { 3255 /* The liveness analysis already ensures that globals are back 3256 in memory. Keep an tcg_debug_assert for safety. */ 3257 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts)); 3258 } 3259 3260 /* save globals to their canonical location and assume they can be 3261 modified be the following code. 'allocated_regs' is used in case a 3262 temporary registers needs to be allocated to store a constant. */ 3263 static void save_globals(TCGContext *s, TCGRegSet allocated_regs) 3264 { 3265 int i, n; 3266 3267 for (i = 0, n = s->nb_globals; i < n; i++) { 3268 temp_save(s, &s->temps[i], allocated_regs); 3269 } 3270 } 3271 3272 /* sync globals to their canonical location and assume they can be 3273 read by the following code. 'allocated_regs' is used in case a 3274 temporary registers needs to be allocated to store a constant. */ 3275 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs) 3276 { 3277 int i, n; 3278 3279 for (i = 0, n = s->nb_globals; i < n; i++) { 3280 TCGTemp *ts = &s->temps[i]; 3281 tcg_debug_assert(ts->val_type != TEMP_VAL_REG 3282 || ts->kind == TEMP_FIXED 3283 || ts->mem_coherent); 3284 } 3285 } 3286 3287 /* at the end of a basic block, we assume all temporaries are dead and 3288 all globals are stored at their canonical location. */ 3289 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) 3290 { 3291 int i; 3292 3293 for (i = s->nb_globals; i < s->nb_temps; i++) { 3294 TCGTemp *ts = &s->temps[i]; 3295 3296 switch (ts->kind) { 3297 case TEMP_LOCAL: 3298 temp_save(s, ts, allocated_regs); 3299 break; 3300 case TEMP_NORMAL: 3301 case TEMP_EBB: 3302 /* The liveness analysis already ensures that temps are dead. 3303 Keep an tcg_debug_assert for safety. */ 3304 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD); 3305 break; 3306 case TEMP_CONST: 3307 /* Similarly, we should have freed any allocated register. */ 3308 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST); 3309 break; 3310 default: 3311 g_assert_not_reached(); 3312 } 3313 } 3314 3315 save_globals(s, allocated_regs); 3316 } 3317 3318 /* 3319 * At a conditional branch, we assume all temporaries are dead unless 3320 * explicitly live-across-conditional-branch; all globals and local 3321 * temps are synced to their location. 3322 */ 3323 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs) 3324 { 3325 sync_globals(s, allocated_regs); 3326 3327 for (int i = s->nb_globals; i < s->nb_temps; i++) { 3328 TCGTemp *ts = &s->temps[i]; 3329 /* 3330 * The liveness analysis already ensures that temps are dead. 3331 * Keep tcg_debug_asserts for safety. 3332 */ 3333 switch (ts->kind) { 3334 case TEMP_LOCAL: 3335 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent); 3336 break; 3337 case TEMP_NORMAL: 3338 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD); 3339 break; 3340 case TEMP_EBB: 3341 case TEMP_CONST: 3342 break; 3343 default: 3344 g_assert_not_reached(); 3345 } 3346 } 3347 } 3348 3349 /* 3350 * Specialized code generation for INDEX_op_mov_* with a constant. 3351 */ 3352 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots, 3353 tcg_target_ulong val, TCGLifeData arg_life, 3354 TCGRegSet preferred_regs) 3355 { 3356 /* ENV should not be modified. */ 3357 tcg_debug_assert(!temp_readonly(ots)); 3358 3359 /* The movi is not explicitly generated here. */ 3360 if (ots->val_type == TEMP_VAL_REG) { 3361 s->reg_to_temp[ots->reg] = NULL; 3362 } 3363 ots->val_type = TEMP_VAL_CONST; 3364 ots->val = val; 3365 ots->mem_coherent = 0; 3366 if (NEED_SYNC_ARG(0)) { 3367 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0)); 3368 } else if (IS_DEAD_ARG(0)) { 3369 temp_dead(s, ots); 3370 } 3371 } 3372 3373 /* 3374 * Specialized code generation for INDEX_op_mov_*. 3375 */ 3376 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op) 3377 { 3378 const TCGLifeData arg_life = op->life; 3379 TCGRegSet allocated_regs, preferred_regs; 3380 TCGTemp *ts, *ots; 3381 TCGType otype, itype; 3382 3383 allocated_regs = s->reserved_regs; 3384 preferred_regs = op->output_pref[0]; 3385 ots = arg_temp(op->args[0]); 3386 ts = arg_temp(op->args[1]); 3387 3388 /* ENV should not be modified. */ 3389 tcg_debug_assert(!temp_readonly(ots)); 3390 3391 /* Note that otype != itype for no-op truncation. */ 3392 otype = ots->type; 3393 itype = ts->type; 3394 3395 if (ts->val_type == TEMP_VAL_CONST) { 3396 /* propagate constant or generate sti */ 3397 tcg_target_ulong val = ts->val; 3398 if (IS_DEAD_ARG(1)) { 3399 temp_dead(s, ts); 3400 } 3401 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs); 3402 return; 3403 } 3404 3405 /* If the source value is in memory we're going to be forced 3406 to have it in a register in order to perform the copy. Copy 3407 the SOURCE value into its own register first, that way we 3408 don't have to reload SOURCE the next time it is used. */ 3409 if (ts->val_type == TEMP_VAL_MEM) { 3410 temp_load(s, ts, tcg_target_available_regs[itype], 3411 allocated_regs, preferred_regs); 3412 } 3413 3414 tcg_debug_assert(ts->val_type == TEMP_VAL_REG); 3415 if (IS_DEAD_ARG(0)) { 3416 /* mov to a non-saved dead register makes no sense (even with 3417 liveness analysis disabled). */ 3418 tcg_debug_assert(NEED_SYNC_ARG(0)); 3419 if (!ots->mem_allocated) { 3420 temp_allocate_frame(s, ots); 3421 } 3422 tcg_out_st(s, otype, ts->reg, ots->mem_base->reg, ots->mem_offset); 3423 if (IS_DEAD_ARG(1)) { 3424 temp_dead(s, ts); 3425 } 3426 temp_dead(s, ots); 3427 } else { 3428 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) { 3429 /* the mov can be suppressed */ 3430 if (ots->val_type == TEMP_VAL_REG) { 3431 s->reg_to_temp[ots->reg] = NULL; 3432 } 3433 ots->reg = ts->reg; 3434 temp_dead(s, ts); 3435 } else { 3436 if (ots->val_type != TEMP_VAL_REG) { 3437 /* When allocating a new register, make sure to not spill the 3438 input one. */ 3439 tcg_regset_set_reg(allocated_regs, ts->reg); 3440 ots->reg = tcg_reg_alloc(s, tcg_target_available_regs[otype], 3441 allocated_regs, preferred_regs, 3442 ots->indirect_base); 3443 } 3444 if (!tcg_out_mov(s, otype, ots->reg, ts->reg)) { 3445 /* 3446 * Cross register class move not supported. 3447 * Store the source register into the destination slot 3448 * and leave the destination temp as TEMP_VAL_MEM. 3449 */ 3450 assert(!temp_readonly(ots)); 3451 if (!ts->mem_allocated) { 3452 temp_allocate_frame(s, ots); 3453 } 3454 tcg_out_st(s, ts->type, ts->reg, 3455 ots->mem_base->reg, ots->mem_offset); 3456 ots->mem_coherent = 1; 3457 temp_free_or_dead(s, ots, -1); 3458 return; 3459 } 3460 } 3461 ots->val_type = TEMP_VAL_REG; 3462 ots->mem_coherent = 0; 3463 s->reg_to_temp[ots->reg] = ots; 3464 if (NEED_SYNC_ARG(0)) { 3465 temp_sync(s, ots, allocated_regs, 0, 0); 3466 } 3467 } 3468 } 3469 3470 /* 3471 * Specialized code generation for INDEX_op_dup_vec. 3472 */ 3473 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op) 3474 { 3475 const TCGLifeData arg_life = op->life; 3476 TCGRegSet dup_out_regs, dup_in_regs; 3477 TCGTemp *its, *ots; 3478 TCGType itype, vtype; 3479 intptr_t endian_fixup; 3480 unsigned vece; 3481 bool ok; 3482 3483 ots = arg_temp(op->args[0]); 3484 its = arg_temp(op->args[1]); 3485 3486 /* ENV should not be modified. */ 3487 tcg_debug_assert(!temp_readonly(ots)); 3488 3489 itype = its->type; 3490 vece = TCGOP_VECE(op); 3491 vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 3492 3493 if (its->val_type == TEMP_VAL_CONST) { 3494 /* Propagate constant via movi -> dupi. */ 3495 tcg_target_ulong val = its->val; 3496 if (IS_DEAD_ARG(1)) { 3497 temp_dead(s, its); 3498 } 3499 tcg_reg_alloc_do_movi(s, ots, val, arg_life, op->output_pref[0]); 3500 return; 3501 } 3502 3503 dup_out_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 3504 dup_in_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[1].regs; 3505 3506 /* Allocate the output register now. */ 3507 if (ots->val_type != TEMP_VAL_REG) { 3508 TCGRegSet allocated_regs = s->reserved_regs; 3509 3510 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) { 3511 /* Make sure to not spill the input register. */ 3512 tcg_regset_set_reg(allocated_regs, its->reg); 3513 } 3514 ots->reg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 3515 op->output_pref[0], ots->indirect_base); 3516 ots->val_type = TEMP_VAL_REG; 3517 ots->mem_coherent = 0; 3518 s->reg_to_temp[ots->reg] = ots; 3519 } 3520 3521 switch (its->val_type) { 3522 case TEMP_VAL_REG: 3523 /* 3524 * The dup constriaints must be broad, covering all possible VECE. 3525 * However, tcg_op_dup_vec() gets to see the VECE and we allow it 3526 * to fail, indicating that extra moves are required for that case. 3527 */ 3528 if (tcg_regset_test_reg(dup_in_regs, its->reg)) { 3529 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) { 3530 goto done; 3531 } 3532 /* Try again from memory or a vector input register. */ 3533 } 3534 if (!its->mem_coherent) { 3535 /* 3536 * The input register is not synced, and so an extra store 3537 * would be required to use memory. Attempt an integer-vector 3538 * register move first. We do not have a TCGRegSet for this. 3539 */ 3540 if (tcg_out_mov(s, itype, ots->reg, its->reg)) { 3541 break; 3542 } 3543 /* Sync the temp back to its slot and load from there. */ 3544 temp_sync(s, its, s->reserved_regs, 0, 0); 3545 } 3546 /* fall through */ 3547 3548 case TEMP_VAL_MEM: 3549 #if HOST_BIG_ENDIAN 3550 endian_fixup = itype == TCG_TYPE_I32 ? 4 : 8; 3551 endian_fixup -= 1 << vece; 3552 #else 3553 endian_fixup = 0; 3554 #endif 3555 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg, 3556 its->mem_offset + endian_fixup)) { 3557 goto done; 3558 } 3559 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset); 3560 break; 3561 3562 default: 3563 g_assert_not_reached(); 3564 } 3565 3566 /* We now have a vector input register, so dup must succeed. */ 3567 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg); 3568 tcg_debug_assert(ok); 3569 3570 done: 3571 if (IS_DEAD_ARG(1)) { 3572 temp_dead(s, its); 3573 } 3574 if (NEED_SYNC_ARG(0)) { 3575 temp_sync(s, ots, s->reserved_regs, 0, 0); 3576 } 3577 if (IS_DEAD_ARG(0)) { 3578 temp_dead(s, ots); 3579 } 3580 } 3581 3582 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op) 3583 { 3584 const TCGLifeData arg_life = op->life; 3585 const TCGOpDef * const def = &tcg_op_defs[op->opc]; 3586 TCGRegSet i_allocated_regs; 3587 TCGRegSet o_allocated_regs; 3588 int i, k, nb_iargs, nb_oargs; 3589 TCGReg reg; 3590 TCGArg arg; 3591 const TCGArgConstraint *arg_ct; 3592 TCGTemp *ts; 3593 TCGArg new_args[TCG_MAX_OP_ARGS]; 3594 int const_args[TCG_MAX_OP_ARGS]; 3595 3596 nb_oargs = def->nb_oargs; 3597 nb_iargs = def->nb_iargs; 3598 3599 /* copy constants */ 3600 memcpy(new_args + nb_oargs + nb_iargs, 3601 op->args + nb_oargs + nb_iargs, 3602 sizeof(TCGArg) * def->nb_cargs); 3603 3604 i_allocated_regs = s->reserved_regs; 3605 o_allocated_regs = s->reserved_regs; 3606 3607 /* satisfy input constraints */ 3608 for (k = 0; k < nb_iargs; k++) { 3609 TCGRegSet i_preferred_regs, o_preferred_regs; 3610 3611 i = def->args_ct[nb_oargs + k].sort_index; 3612 arg = op->args[i]; 3613 arg_ct = &def->args_ct[i]; 3614 ts = arg_temp(arg); 3615 3616 if (ts->val_type == TEMP_VAL_CONST 3617 && tcg_target_const_match(ts->val, ts->type, arg_ct->ct)) { 3618 /* constant is OK for instruction */ 3619 const_args[i] = 1; 3620 new_args[i] = ts->val; 3621 continue; 3622 } 3623 3624 i_preferred_regs = o_preferred_regs = 0; 3625 if (arg_ct->ialias) { 3626 o_preferred_regs = op->output_pref[arg_ct->alias_index]; 3627 3628 /* 3629 * If the input is readonly, then it cannot also be an 3630 * output and aliased to itself. If the input is not 3631 * dead after the instruction, we must allocate a new 3632 * register and move it. 3633 */ 3634 if (temp_readonly(ts) || !IS_DEAD_ARG(i)) { 3635 goto allocate_in_reg; 3636 } 3637 3638 /* 3639 * Check if the current register has already been allocated 3640 * for another input aliased to an output. 3641 */ 3642 if (ts->val_type == TEMP_VAL_REG) { 3643 reg = ts->reg; 3644 for (int k2 = 0; k2 < k; k2++) { 3645 int i2 = def->args_ct[nb_oargs + k2].sort_index; 3646 if (def->args_ct[i2].ialias && reg == new_args[i2]) { 3647 goto allocate_in_reg; 3648 } 3649 } 3650 } 3651 i_preferred_regs = o_preferred_regs; 3652 } 3653 3654 temp_load(s, ts, arg_ct->regs, i_allocated_regs, i_preferred_regs); 3655 reg = ts->reg; 3656 3657 if (!tcg_regset_test_reg(arg_ct->regs, reg)) { 3658 allocate_in_reg: 3659 /* 3660 * Allocate a new register matching the constraint 3661 * and move the temporary register into it. 3662 */ 3663 temp_load(s, ts, tcg_target_available_regs[ts->type], 3664 i_allocated_regs, 0); 3665 reg = tcg_reg_alloc(s, arg_ct->regs, i_allocated_regs, 3666 o_preferred_regs, ts->indirect_base); 3667 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 3668 /* 3669 * Cross register class move not supported. Sync the 3670 * temp back to its slot and load from there. 3671 */ 3672 temp_sync(s, ts, i_allocated_regs, 0, 0); 3673 tcg_out_ld(s, ts->type, reg, 3674 ts->mem_base->reg, ts->mem_offset); 3675 } 3676 } 3677 new_args[i] = reg; 3678 const_args[i] = 0; 3679 tcg_regset_set_reg(i_allocated_regs, reg); 3680 } 3681 3682 /* mark dead temporaries and free the associated registers */ 3683 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3684 if (IS_DEAD_ARG(i)) { 3685 temp_dead(s, arg_temp(op->args[i])); 3686 } 3687 } 3688 3689 if (def->flags & TCG_OPF_COND_BRANCH) { 3690 tcg_reg_alloc_cbranch(s, i_allocated_regs); 3691 } else if (def->flags & TCG_OPF_BB_END) { 3692 tcg_reg_alloc_bb_end(s, i_allocated_regs); 3693 } else { 3694 if (def->flags & TCG_OPF_CALL_CLOBBER) { 3695 /* XXX: permit generic clobber register list ? */ 3696 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 3697 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 3698 tcg_reg_free(s, i, i_allocated_regs); 3699 } 3700 } 3701 } 3702 if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3703 /* sync globals if the op has side effects and might trigger 3704 an exception. */ 3705 sync_globals(s, i_allocated_regs); 3706 } 3707 3708 /* satisfy the output constraints */ 3709 for(k = 0; k < nb_oargs; k++) { 3710 i = def->args_ct[k].sort_index; 3711 arg = op->args[i]; 3712 arg_ct = &def->args_ct[i]; 3713 ts = arg_temp(arg); 3714 3715 /* ENV should not be modified. */ 3716 tcg_debug_assert(!temp_readonly(ts)); 3717 3718 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) { 3719 reg = new_args[arg_ct->alias_index]; 3720 } else if (arg_ct->newreg) { 3721 reg = tcg_reg_alloc(s, arg_ct->regs, 3722 i_allocated_regs | o_allocated_regs, 3723 op->output_pref[k], ts->indirect_base); 3724 } else { 3725 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs, 3726 op->output_pref[k], ts->indirect_base); 3727 } 3728 tcg_regset_set_reg(o_allocated_regs, reg); 3729 if (ts->val_type == TEMP_VAL_REG) { 3730 s->reg_to_temp[ts->reg] = NULL; 3731 } 3732 ts->val_type = TEMP_VAL_REG; 3733 ts->reg = reg; 3734 /* 3735 * Temp value is modified, so the value kept in memory is 3736 * potentially not the same. 3737 */ 3738 ts->mem_coherent = 0; 3739 s->reg_to_temp[reg] = ts; 3740 new_args[i] = reg; 3741 } 3742 } 3743 3744 /* emit instruction */ 3745 if (def->flags & TCG_OPF_VECTOR) { 3746 tcg_out_vec_op(s, op->opc, TCGOP_VECL(op), TCGOP_VECE(op), 3747 new_args, const_args); 3748 } else { 3749 tcg_out_op(s, op->opc, new_args, const_args); 3750 } 3751 3752 /* move the outputs in the correct register if needed */ 3753 for(i = 0; i < nb_oargs; i++) { 3754 ts = arg_temp(op->args[i]); 3755 3756 /* ENV should not be modified. */ 3757 tcg_debug_assert(!temp_readonly(ts)); 3758 3759 if (NEED_SYNC_ARG(i)) { 3760 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i)); 3761 } else if (IS_DEAD_ARG(i)) { 3762 temp_dead(s, ts); 3763 } 3764 } 3765 } 3766 3767 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op) 3768 { 3769 const TCGLifeData arg_life = op->life; 3770 TCGTemp *ots, *itsl, *itsh; 3771 TCGType vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 3772 3773 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */ 3774 tcg_debug_assert(TCG_TARGET_REG_BITS == 32); 3775 tcg_debug_assert(TCGOP_VECE(op) == MO_64); 3776 3777 ots = arg_temp(op->args[0]); 3778 itsl = arg_temp(op->args[1]); 3779 itsh = arg_temp(op->args[2]); 3780 3781 /* ENV should not be modified. */ 3782 tcg_debug_assert(!temp_readonly(ots)); 3783 3784 /* Allocate the output register now. */ 3785 if (ots->val_type != TEMP_VAL_REG) { 3786 TCGRegSet allocated_regs = s->reserved_regs; 3787 TCGRegSet dup_out_regs = 3788 tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 3789 3790 /* Make sure to not spill the input registers. */ 3791 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) { 3792 tcg_regset_set_reg(allocated_regs, itsl->reg); 3793 } 3794 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) { 3795 tcg_regset_set_reg(allocated_regs, itsh->reg); 3796 } 3797 3798 ots->reg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 3799 op->output_pref[0], ots->indirect_base); 3800 ots->val_type = TEMP_VAL_REG; 3801 ots->mem_coherent = 0; 3802 s->reg_to_temp[ots->reg] = ots; 3803 } 3804 3805 /* Promote dup2 of immediates to dupi_vec. */ 3806 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) { 3807 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val); 3808 MemOp vece = MO_64; 3809 3810 if (val == dup_const(MO_8, val)) { 3811 vece = MO_8; 3812 } else if (val == dup_const(MO_16, val)) { 3813 vece = MO_16; 3814 } else if (val == dup_const(MO_32, val)) { 3815 vece = MO_32; 3816 } 3817 3818 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val); 3819 goto done; 3820 } 3821 3822 /* If the two inputs form one 64-bit value, try dupm_vec. */ 3823 if (itsl + 1 == itsh && itsl->base_type == TCG_TYPE_I64) { 3824 if (!itsl->mem_coherent) { 3825 temp_sync(s, itsl, s->reserved_regs, 0, 0); 3826 } 3827 if (!itsh->mem_coherent) { 3828 temp_sync(s, itsh, s->reserved_regs, 0, 0); 3829 } 3830 #if HOST_BIG_ENDIAN 3831 TCGTemp *its = itsh; 3832 #else 3833 TCGTemp *its = itsl; 3834 #endif 3835 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg, 3836 its->mem_base->reg, its->mem_offset)) { 3837 goto done; 3838 } 3839 } 3840 3841 /* Fall back to generic expansion. */ 3842 return false; 3843 3844 done: 3845 if (IS_DEAD_ARG(1)) { 3846 temp_dead(s, itsl); 3847 } 3848 if (IS_DEAD_ARG(2)) { 3849 temp_dead(s, itsh); 3850 } 3851 if (NEED_SYNC_ARG(0)) { 3852 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0)); 3853 } else if (IS_DEAD_ARG(0)) { 3854 temp_dead(s, ots); 3855 } 3856 return true; 3857 } 3858 3859 #ifdef TCG_TARGET_STACK_GROWSUP 3860 #define STACK_DIR(x) (-(x)) 3861 #else 3862 #define STACK_DIR(x) (x) 3863 #endif 3864 3865 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op) 3866 { 3867 const int nb_oargs = TCGOP_CALLO(op); 3868 const int nb_iargs = TCGOP_CALLI(op); 3869 const TCGLifeData arg_life = op->life; 3870 const TCGHelperInfo *info; 3871 int flags, nb_regs, i; 3872 TCGReg reg; 3873 TCGArg arg; 3874 TCGTemp *ts; 3875 intptr_t stack_offset; 3876 size_t call_stack_size; 3877 tcg_insn_unit *func_addr; 3878 int allocate_args; 3879 TCGRegSet allocated_regs; 3880 3881 func_addr = tcg_call_func(op); 3882 info = tcg_call_info(op); 3883 flags = info->flags; 3884 3885 nb_regs = ARRAY_SIZE(tcg_target_call_iarg_regs); 3886 if (nb_regs > nb_iargs) { 3887 nb_regs = nb_iargs; 3888 } 3889 3890 /* assign stack slots first */ 3891 call_stack_size = (nb_iargs - nb_regs) * sizeof(tcg_target_long); 3892 call_stack_size = (call_stack_size + TCG_TARGET_STACK_ALIGN - 1) & 3893 ~(TCG_TARGET_STACK_ALIGN - 1); 3894 allocate_args = (call_stack_size > TCG_STATIC_CALL_ARGS_SIZE); 3895 if (allocate_args) { 3896 /* XXX: if more than TCG_STATIC_CALL_ARGS_SIZE is needed, 3897 preallocate call stack */ 3898 tcg_abort(); 3899 } 3900 3901 stack_offset = TCG_TARGET_CALL_STACK_OFFSET; 3902 for (i = nb_regs; i < nb_iargs; i++) { 3903 arg = op->args[nb_oargs + i]; 3904 #ifdef TCG_TARGET_STACK_GROWSUP 3905 stack_offset -= sizeof(tcg_target_long); 3906 #endif 3907 if (arg != TCG_CALL_DUMMY_ARG) { 3908 ts = arg_temp(arg); 3909 temp_load(s, ts, tcg_target_available_regs[ts->type], 3910 s->reserved_regs, 0); 3911 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, stack_offset); 3912 } 3913 #ifndef TCG_TARGET_STACK_GROWSUP 3914 stack_offset += sizeof(tcg_target_long); 3915 #endif 3916 } 3917 3918 /* assign input registers */ 3919 allocated_regs = s->reserved_regs; 3920 for (i = 0; i < nb_regs; i++) { 3921 arg = op->args[nb_oargs + i]; 3922 if (arg != TCG_CALL_DUMMY_ARG) { 3923 ts = arg_temp(arg); 3924 reg = tcg_target_call_iarg_regs[i]; 3925 3926 if (ts->val_type == TEMP_VAL_REG) { 3927 if (ts->reg != reg) { 3928 tcg_reg_free(s, reg, allocated_regs); 3929 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 3930 /* 3931 * Cross register class move not supported. Sync the 3932 * temp back to its slot and load from there. 3933 */ 3934 temp_sync(s, ts, allocated_regs, 0, 0); 3935 tcg_out_ld(s, ts->type, reg, 3936 ts->mem_base->reg, ts->mem_offset); 3937 } 3938 } 3939 } else { 3940 TCGRegSet arg_set = 0; 3941 3942 tcg_reg_free(s, reg, allocated_regs); 3943 tcg_regset_set_reg(arg_set, reg); 3944 temp_load(s, ts, arg_set, allocated_regs, 0); 3945 } 3946 3947 tcg_regset_set_reg(allocated_regs, reg); 3948 } 3949 } 3950 3951 /* mark dead temporaries and free the associated registers */ 3952 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3953 if (IS_DEAD_ARG(i)) { 3954 temp_dead(s, arg_temp(op->args[i])); 3955 } 3956 } 3957 3958 /* clobber call registers */ 3959 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 3960 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 3961 tcg_reg_free(s, i, allocated_regs); 3962 } 3963 } 3964 3965 /* Save globals if they might be written by the helper, sync them if 3966 they might be read. */ 3967 if (flags & TCG_CALL_NO_READ_GLOBALS) { 3968 /* Nothing to do */ 3969 } else if (flags & TCG_CALL_NO_WRITE_GLOBALS) { 3970 sync_globals(s, allocated_regs); 3971 } else { 3972 save_globals(s, allocated_regs); 3973 } 3974 3975 #ifdef CONFIG_TCG_INTERPRETER 3976 { 3977 gpointer hash = (gpointer)(uintptr_t)info->typemask; 3978 ffi_cif *cif = g_hash_table_lookup(ffi_table, hash); 3979 assert(cif != NULL); 3980 tcg_out_call(s, func_addr, cif); 3981 } 3982 #else 3983 tcg_out_call(s, func_addr); 3984 #endif 3985 3986 /* assign output registers and emit moves if needed */ 3987 for(i = 0; i < nb_oargs; i++) { 3988 arg = op->args[i]; 3989 ts = arg_temp(arg); 3990 3991 /* ENV should not be modified. */ 3992 tcg_debug_assert(!temp_readonly(ts)); 3993 3994 reg = tcg_target_call_oarg_regs[i]; 3995 tcg_debug_assert(s->reg_to_temp[reg] == NULL); 3996 if (ts->val_type == TEMP_VAL_REG) { 3997 s->reg_to_temp[ts->reg] = NULL; 3998 } 3999 ts->val_type = TEMP_VAL_REG; 4000 ts->reg = reg; 4001 ts->mem_coherent = 0; 4002 s->reg_to_temp[reg] = ts; 4003 if (NEED_SYNC_ARG(i)) { 4004 temp_sync(s, ts, allocated_regs, 0, IS_DEAD_ARG(i)); 4005 } else if (IS_DEAD_ARG(i)) { 4006 temp_dead(s, ts); 4007 } 4008 } 4009 } 4010 4011 #ifdef CONFIG_PROFILER 4012 4013 /* avoid copy/paste errors */ 4014 #define PROF_ADD(to, from, field) \ 4015 do { \ 4016 (to)->field += qatomic_read(&((from)->field)); \ 4017 } while (0) 4018 4019 #define PROF_MAX(to, from, field) \ 4020 do { \ 4021 typeof((from)->field) val__ = qatomic_read(&((from)->field)); \ 4022 if (val__ > (to)->field) { \ 4023 (to)->field = val__; \ 4024 } \ 4025 } while (0) 4026 4027 /* Pass in a zero'ed @prof */ 4028 static inline 4029 void tcg_profile_snapshot(TCGProfile *prof, bool counters, bool table) 4030 { 4031 unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); 4032 unsigned int i; 4033 4034 for (i = 0; i < n_ctxs; i++) { 4035 TCGContext *s = qatomic_read(&tcg_ctxs[i]); 4036 const TCGProfile *orig = &s->prof; 4037 4038 if (counters) { 4039 PROF_ADD(prof, orig, cpu_exec_time); 4040 PROF_ADD(prof, orig, tb_count1); 4041 PROF_ADD(prof, orig, tb_count); 4042 PROF_ADD(prof, orig, op_count); 4043 PROF_MAX(prof, orig, op_count_max); 4044 PROF_ADD(prof, orig, temp_count); 4045 PROF_MAX(prof, orig, temp_count_max); 4046 PROF_ADD(prof, orig, del_op_count); 4047 PROF_ADD(prof, orig, code_in_len); 4048 PROF_ADD(prof, orig, code_out_len); 4049 PROF_ADD(prof, orig, search_out_len); 4050 PROF_ADD(prof, orig, interm_time); 4051 PROF_ADD(prof, orig, code_time); 4052 PROF_ADD(prof, orig, la_time); 4053 PROF_ADD(prof, orig, opt_time); 4054 PROF_ADD(prof, orig, restore_count); 4055 PROF_ADD(prof, orig, restore_time); 4056 } 4057 if (table) { 4058 int i; 4059 4060 for (i = 0; i < NB_OPS; i++) { 4061 PROF_ADD(prof, orig, table_op_count[i]); 4062 } 4063 } 4064 } 4065 } 4066 4067 #undef PROF_ADD 4068 #undef PROF_MAX 4069 4070 static void tcg_profile_snapshot_counters(TCGProfile *prof) 4071 { 4072 tcg_profile_snapshot(prof, true, false); 4073 } 4074 4075 static void tcg_profile_snapshot_table(TCGProfile *prof) 4076 { 4077 tcg_profile_snapshot(prof, false, true); 4078 } 4079 4080 void tcg_dump_op_count(GString *buf) 4081 { 4082 TCGProfile prof = {}; 4083 int i; 4084 4085 tcg_profile_snapshot_table(&prof); 4086 for (i = 0; i < NB_OPS; i++) { 4087 g_string_append_printf(buf, "%s %" PRId64 "\n", tcg_op_defs[i].name, 4088 prof.table_op_count[i]); 4089 } 4090 } 4091 4092 int64_t tcg_cpu_exec_time(void) 4093 { 4094 unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); 4095 unsigned int i; 4096 int64_t ret = 0; 4097 4098 for (i = 0; i < n_ctxs; i++) { 4099 const TCGContext *s = qatomic_read(&tcg_ctxs[i]); 4100 const TCGProfile *prof = &s->prof; 4101 4102 ret += qatomic_read(&prof->cpu_exec_time); 4103 } 4104 return ret; 4105 } 4106 #else 4107 void tcg_dump_op_count(GString *buf) 4108 { 4109 g_string_append_printf(buf, "[TCG profiler not compiled]\n"); 4110 } 4111 4112 int64_t tcg_cpu_exec_time(void) 4113 { 4114 error_report("%s: TCG profiler not compiled", __func__); 4115 exit(EXIT_FAILURE); 4116 } 4117 #endif 4118 4119 4120 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, target_ulong pc_start) 4121 { 4122 #ifdef CONFIG_PROFILER 4123 TCGProfile *prof = &s->prof; 4124 #endif 4125 int i, num_insns; 4126 TCGOp *op; 4127 4128 #ifdef CONFIG_PROFILER 4129 { 4130 int n = 0; 4131 4132 QTAILQ_FOREACH(op, &s->ops, link) { 4133 n++; 4134 } 4135 qatomic_set(&prof->op_count, prof->op_count + n); 4136 if (n > prof->op_count_max) { 4137 qatomic_set(&prof->op_count_max, n); 4138 } 4139 4140 n = s->nb_temps; 4141 qatomic_set(&prof->temp_count, prof->temp_count + n); 4142 if (n > prof->temp_count_max) { 4143 qatomic_set(&prof->temp_count_max, n); 4144 } 4145 } 4146 #endif 4147 4148 #ifdef DEBUG_DISAS 4149 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP) 4150 && qemu_log_in_addr_range(pc_start))) { 4151 FILE *logfile = qemu_log_trylock(); 4152 if (logfile) { 4153 fprintf(logfile, "OP:\n"); 4154 tcg_dump_ops(s, logfile, false); 4155 fprintf(logfile, "\n"); 4156 qemu_log_unlock(logfile); 4157 } 4158 } 4159 #endif 4160 4161 #ifdef CONFIG_DEBUG_TCG 4162 /* Ensure all labels referenced have been emitted. */ 4163 { 4164 TCGLabel *l; 4165 bool error = false; 4166 4167 QSIMPLEQ_FOREACH(l, &s->labels, next) { 4168 if (unlikely(!l->present) && l->refs) { 4169 qemu_log_mask(CPU_LOG_TB_OP, 4170 "$L%d referenced but not present.\n", l->id); 4171 error = true; 4172 } 4173 } 4174 assert(!error); 4175 } 4176 #endif 4177 4178 #ifdef CONFIG_PROFILER 4179 qatomic_set(&prof->opt_time, prof->opt_time - profile_getclock()); 4180 #endif 4181 4182 #ifdef USE_TCG_OPTIMIZATIONS 4183 tcg_optimize(s); 4184 #endif 4185 4186 #ifdef CONFIG_PROFILER 4187 qatomic_set(&prof->opt_time, prof->opt_time + profile_getclock()); 4188 qatomic_set(&prof->la_time, prof->la_time - profile_getclock()); 4189 #endif 4190 4191 reachable_code_pass(s); 4192 liveness_pass_1(s); 4193 4194 if (s->nb_indirects > 0) { 4195 #ifdef DEBUG_DISAS 4196 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND) 4197 && qemu_log_in_addr_range(pc_start))) { 4198 FILE *logfile = qemu_log_trylock(); 4199 if (logfile) { 4200 fprintf(logfile, "OP before indirect lowering:\n"); 4201 tcg_dump_ops(s, logfile, false); 4202 fprintf(logfile, "\n"); 4203 qemu_log_unlock(logfile); 4204 } 4205 } 4206 #endif 4207 /* Replace indirect temps with direct temps. */ 4208 if (liveness_pass_2(s)) { 4209 /* If changes were made, re-run liveness. */ 4210 liveness_pass_1(s); 4211 } 4212 } 4213 4214 #ifdef CONFIG_PROFILER 4215 qatomic_set(&prof->la_time, prof->la_time + profile_getclock()); 4216 #endif 4217 4218 #ifdef DEBUG_DISAS 4219 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT) 4220 && qemu_log_in_addr_range(pc_start))) { 4221 FILE *logfile = qemu_log_trylock(); 4222 if (logfile) { 4223 fprintf(logfile, "OP after optimization and liveness analysis:\n"); 4224 tcg_dump_ops(s, logfile, true); 4225 fprintf(logfile, "\n"); 4226 qemu_log_unlock(logfile); 4227 } 4228 } 4229 #endif 4230 4231 tcg_reg_alloc_start(s); 4232 4233 /* 4234 * Reset the buffer pointers when restarting after overflow. 4235 * TODO: Move this into translate-all.c with the rest of the 4236 * buffer management. Having only this done here is confusing. 4237 */ 4238 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr); 4239 s->code_ptr = s->code_buf; 4240 4241 #ifdef TCG_TARGET_NEED_LDST_LABELS 4242 QSIMPLEQ_INIT(&s->ldst_labels); 4243 #endif 4244 #ifdef TCG_TARGET_NEED_POOL_LABELS 4245 s->pool_labels = NULL; 4246 #endif 4247 4248 num_insns = -1; 4249 QTAILQ_FOREACH(op, &s->ops, link) { 4250 TCGOpcode opc = op->opc; 4251 4252 #ifdef CONFIG_PROFILER 4253 qatomic_set(&prof->table_op_count[opc], prof->table_op_count[opc] + 1); 4254 #endif 4255 4256 switch (opc) { 4257 case INDEX_op_mov_i32: 4258 case INDEX_op_mov_i64: 4259 case INDEX_op_mov_vec: 4260 tcg_reg_alloc_mov(s, op); 4261 break; 4262 case INDEX_op_dup_vec: 4263 tcg_reg_alloc_dup(s, op); 4264 break; 4265 case INDEX_op_insn_start: 4266 if (num_insns >= 0) { 4267 size_t off = tcg_current_code_size(s); 4268 s->gen_insn_end_off[num_insns] = off; 4269 /* Assert that we do not overflow our stored offset. */ 4270 assert(s->gen_insn_end_off[num_insns] == off); 4271 } 4272 num_insns++; 4273 for (i = 0; i < TARGET_INSN_START_WORDS; ++i) { 4274 target_ulong a; 4275 #if TARGET_LONG_BITS > TCG_TARGET_REG_BITS 4276 a = deposit64(op->args[i * 2], 32, 32, op->args[i * 2 + 1]); 4277 #else 4278 a = op->args[i]; 4279 #endif 4280 s->gen_insn_data[num_insns][i] = a; 4281 } 4282 break; 4283 case INDEX_op_discard: 4284 temp_dead(s, arg_temp(op->args[0])); 4285 break; 4286 case INDEX_op_set_label: 4287 tcg_reg_alloc_bb_end(s, s->reserved_regs); 4288 tcg_out_label(s, arg_label(op->args[0])); 4289 break; 4290 case INDEX_op_call: 4291 tcg_reg_alloc_call(s, op); 4292 break; 4293 case INDEX_op_dup2_vec: 4294 if (tcg_reg_alloc_dup2(s, op)) { 4295 break; 4296 } 4297 /* fall through */ 4298 default: 4299 /* Sanity check that we've not introduced any unhandled opcodes. */ 4300 tcg_debug_assert(tcg_op_supported(opc)); 4301 /* Note: in order to speed up the code, it would be much 4302 faster to have specialized register allocator functions for 4303 some common argument patterns */ 4304 tcg_reg_alloc_op(s, op); 4305 break; 4306 } 4307 #ifdef CONFIG_DEBUG_TCG 4308 check_regs(s); 4309 #endif 4310 /* Test for (pending) buffer overflow. The assumption is that any 4311 one operation beginning below the high water mark cannot overrun 4312 the buffer completely. Thus we can test for overflow after 4313 generating code without having to check during generation. */ 4314 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) { 4315 return -1; 4316 } 4317 /* Test for TB overflow, as seen by gen_insn_end_off. */ 4318 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) { 4319 return -2; 4320 } 4321 } 4322 tcg_debug_assert(num_insns >= 0); 4323 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s); 4324 4325 /* Generate TB finalization at the end of block */ 4326 #ifdef TCG_TARGET_NEED_LDST_LABELS 4327 i = tcg_out_ldst_finalize(s); 4328 if (i < 0) { 4329 return i; 4330 } 4331 #endif 4332 #ifdef TCG_TARGET_NEED_POOL_LABELS 4333 i = tcg_out_pool_finalize(s); 4334 if (i < 0) { 4335 return i; 4336 } 4337 #endif 4338 if (!tcg_resolve_relocs(s)) { 4339 return -2; 4340 } 4341 4342 #ifndef CONFIG_TCG_INTERPRETER 4343 /* flush instruction cache */ 4344 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 4345 (uintptr_t)s->code_buf, 4346 tcg_ptr_byte_diff(s->code_ptr, s->code_buf)); 4347 #endif 4348 4349 return tcg_current_code_size(s); 4350 } 4351 4352 #ifdef CONFIG_PROFILER 4353 void tcg_dump_info(GString *buf) 4354 { 4355 TCGProfile prof = {}; 4356 const TCGProfile *s; 4357 int64_t tb_count; 4358 int64_t tb_div_count; 4359 int64_t tot; 4360 4361 tcg_profile_snapshot_counters(&prof); 4362 s = &prof; 4363 tb_count = s->tb_count; 4364 tb_div_count = tb_count ? tb_count : 1; 4365 tot = s->interm_time + s->code_time; 4366 4367 g_string_append_printf(buf, "JIT cycles %" PRId64 4368 " (%0.3f s at 2.4 GHz)\n", 4369 tot, tot / 2.4e9); 4370 g_string_append_printf(buf, "translated TBs %" PRId64 4371 " (aborted=%" PRId64 " %0.1f%%)\n", 4372 tb_count, s->tb_count1 - tb_count, 4373 (double)(s->tb_count1 - s->tb_count) 4374 / (s->tb_count1 ? s->tb_count1 : 1) * 100.0); 4375 g_string_append_printf(buf, "avg ops/TB %0.1f max=%d\n", 4376 (double)s->op_count / tb_div_count, s->op_count_max); 4377 g_string_append_printf(buf, "deleted ops/TB %0.2f\n", 4378 (double)s->del_op_count / tb_div_count); 4379 g_string_append_printf(buf, "avg temps/TB %0.2f max=%d\n", 4380 (double)s->temp_count / tb_div_count, 4381 s->temp_count_max); 4382 g_string_append_printf(buf, "avg host code/TB %0.1f\n", 4383 (double)s->code_out_len / tb_div_count); 4384 g_string_append_printf(buf, "avg search data/TB %0.1f\n", 4385 (double)s->search_out_len / tb_div_count); 4386 4387 g_string_append_printf(buf, "cycles/op %0.1f\n", 4388 s->op_count ? (double)tot / s->op_count : 0); 4389 g_string_append_printf(buf, "cycles/in byte %0.1f\n", 4390 s->code_in_len ? (double)tot / s->code_in_len : 0); 4391 g_string_append_printf(buf, "cycles/out byte %0.1f\n", 4392 s->code_out_len ? (double)tot / s->code_out_len : 0); 4393 g_string_append_printf(buf, "cycles/search byte %0.1f\n", 4394 s->search_out_len ? 4395 (double)tot / s->search_out_len : 0); 4396 if (tot == 0) { 4397 tot = 1; 4398 } 4399 g_string_append_printf(buf, " gen_interm time %0.1f%%\n", 4400 (double)s->interm_time / tot * 100.0); 4401 g_string_append_printf(buf, " gen_code time %0.1f%%\n", 4402 (double)s->code_time / tot * 100.0); 4403 g_string_append_printf(buf, "optim./code time %0.1f%%\n", 4404 (double)s->opt_time / (s->code_time ? 4405 s->code_time : 1) 4406 * 100.0); 4407 g_string_append_printf(buf, "liveness/code time %0.1f%%\n", 4408 (double)s->la_time / (s->code_time ? 4409 s->code_time : 1) * 100.0); 4410 g_string_append_printf(buf, "cpu_restore count %" PRId64 "\n", 4411 s->restore_count); 4412 g_string_append_printf(buf, " avg cycles %0.1f\n", 4413 s->restore_count ? 4414 (double)s->restore_time / s->restore_count : 0); 4415 } 4416 #else 4417 void tcg_dump_info(GString *buf) 4418 { 4419 g_string_append_printf(buf, "[TCG profiler not compiled]\n"); 4420 } 4421 #endif 4422 4423 #ifdef ELF_HOST_MACHINE 4424 /* In order to use this feature, the backend needs to do three things: 4425 4426 (1) Define ELF_HOST_MACHINE to indicate both what value to 4427 put into the ELF image and to indicate support for the feature. 4428 4429 (2) Define tcg_register_jit. This should create a buffer containing 4430 the contents of a .debug_frame section that describes the post- 4431 prologue unwind info for the tcg machine. 4432 4433 (3) Call tcg_register_jit_int, with the constructed .debug_frame. 4434 */ 4435 4436 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */ 4437 typedef enum { 4438 JIT_NOACTION = 0, 4439 JIT_REGISTER_FN, 4440 JIT_UNREGISTER_FN 4441 } jit_actions_t; 4442 4443 struct jit_code_entry { 4444 struct jit_code_entry *next_entry; 4445 struct jit_code_entry *prev_entry; 4446 const void *symfile_addr; 4447 uint64_t symfile_size; 4448 }; 4449 4450 struct jit_descriptor { 4451 uint32_t version; 4452 uint32_t action_flag; 4453 struct jit_code_entry *relevant_entry; 4454 struct jit_code_entry *first_entry; 4455 }; 4456 4457 void __jit_debug_register_code(void) __attribute__((noinline)); 4458 void __jit_debug_register_code(void) 4459 { 4460 asm(""); 4461 } 4462 4463 /* Must statically initialize the version, because GDB may check 4464 the version before we can set it. */ 4465 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 }; 4466 4467 /* End GDB interface. */ 4468 4469 static int find_string(const char *strtab, const char *str) 4470 { 4471 const char *p = strtab + 1; 4472 4473 while (1) { 4474 if (strcmp(p, str) == 0) { 4475 return p - strtab; 4476 } 4477 p += strlen(p) + 1; 4478 } 4479 } 4480 4481 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size, 4482 const void *debug_frame, 4483 size_t debug_frame_size) 4484 { 4485 struct __attribute__((packed)) DebugInfo { 4486 uint32_t len; 4487 uint16_t version; 4488 uint32_t abbrev; 4489 uint8_t ptr_size; 4490 uint8_t cu_die; 4491 uint16_t cu_lang; 4492 uintptr_t cu_low_pc; 4493 uintptr_t cu_high_pc; 4494 uint8_t fn_die; 4495 char fn_name[16]; 4496 uintptr_t fn_low_pc; 4497 uintptr_t fn_high_pc; 4498 uint8_t cu_eoc; 4499 }; 4500 4501 struct ElfImage { 4502 ElfW(Ehdr) ehdr; 4503 ElfW(Phdr) phdr; 4504 ElfW(Shdr) shdr[7]; 4505 ElfW(Sym) sym[2]; 4506 struct DebugInfo di; 4507 uint8_t da[24]; 4508 char str[80]; 4509 }; 4510 4511 struct ElfImage *img; 4512 4513 static const struct ElfImage img_template = { 4514 .ehdr = { 4515 .e_ident[EI_MAG0] = ELFMAG0, 4516 .e_ident[EI_MAG1] = ELFMAG1, 4517 .e_ident[EI_MAG2] = ELFMAG2, 4518 .e_ident[EI_MAG3] = ELFMAG3, 4519 .e_ident[EI_CLASS] = ELF_CLASS, 4520 .e_ident[EI_DATA] = ELF_DATA, 4521 .e_ident[EI_VERSION] = EV_CURRENT, 4522 .e_type = ET_EXEC, 4523 .e_machine = ELF_HOST_MACHINE, 4524 .e_version = EV_CURRENT, 4525 .e_phoff = offsetof(struct ElfImage, phdr), 4526 .e_shoff = offsetof(struct ElfImage, shdr), 4527 .e_ehsize = sizeof(ElfW(Shdr)), 4528 .e_phentsize = sizeof(ElfW(Phdr)), 4529 .e_phnum = 1, 4530 .e_shentsize = sizeof(ElfW(Shdr)), 4531 .e_shnum = ARRAY_SIZE(img->shdr), 4532 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1, 4533 #ifdef ELF_HOST_FLAGS 4534 .e_flags = ELF_HOST_FLAGS, 4535 #endif 4536 #ifdef ELF_OSABI 4537 .e_ident[EI_OSABI] = ELF_OSABI, 4538 #endif 4539 }, 4540 .phdr = { 4541 .p_type = PT_LOAD, 4542 .p_flags = PF_X, 4543 }, 4544 .shdr = { 4545 [0] = { .sh_type = SHT_NULL }, 4546 /* Trick: The contents of code_gen_buffer are not present in 4547 this fake ELF file; that got allocated elsewhere. Therefore 4548 we mark .text as SHT_NOBITS (similar to .bss) so that readers 4549 will not look for contents. We can record any address. */ 4550 [1] = { /* .text */ 4551 .sh_type = SHT_NOBITS, 4552 .sh_flags = SHF_EXECINSTR | SHF_ALLOC, 4553 }, 4554 [2] = { /* .debug_info */ 4555 .sh_type = SHT_PROGBITS, 4556 .sh_offset = offsetof(struct ElfImage, di), 4557 .sh_size = sizeof(struct DebugInfo), 4558 }, 4559 [3] = { /* .debug_abbrev */ 4560 .sh_type = SHT_PROGBITS, 4561 .sh_offset = offsetof(struct ElfImage, da), 4562 .sh_size = sizeof(img->da), 4563 }, 4564 [4] = { /* .debug_frame */ 4565 .sh_type = SHT_PROGBITS, 4566 .sh_offset = sizeof(struct ElfImage), 4567 }, 4568 [5] = { /* .symtab */ 4569 .sh_type = SHT_SYMTAB, 4570 .sh_offset = offsetof(struct ElfImage, sym), 4571 .sh_size = sizeof(img->sym), 4572 .sh_info = 1, 4573 .sh_link = ARRAY_SIZE(img->shdr) - 1, 4574 .sh_entsize = sizeof(ElfW(Sym)), 4575 }, 4576 [6] = { /* .strtab */ 4577 .sh_type = SHT_STRTAB, 4578 .sh_offset = offsetof(struct ElfImage, str), 4579 .sh_size = sizeof(img->str), 4580 } 4581 }, 4582 .sym = { 4583 [1] = { /* code_gen_buffer */ 4584 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC), 4585 .st_shndx = 1, 4586 } 4587 }, 4588 .di = { 4589 .len = sizeof(struct DebugInfo) - 4, 4590 .version = 2, 4591 .ptr_size = sizeof(void *), 4592 .cu_die = 1, 4593 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */ 4594 .fn_die = 2, 4595 .fn_name = "code_gen_buffer" 4596 }, 4597 .da = { 4598 1, /* abbrev number (the cu) */ 4599 0x11, 1, /* DW_TAG_compile_unit, has children */ 4600 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */ 4601 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 4602 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 4603 0, 0, /* end of abbrev */ 4604 2, /* abbrev number (the fn) */ 4605 0x2e, 0, /* DW_TAG_subprogram, no children */ 4606 0x3, 0x8, /* DW_AT_name, DW_FORM_string */ 4607 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 4608 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 4609 0, 0, /* end of abbrev */ 4610 0 /* no more abbrev */ 4611 }, 4612 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0" 4613 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer", 4614 }; 4615 4616 /* We only need a single jit entry; statically allocate it. */ 4617 static struct jit_code_entry one_entry; 4618 4619 uintptr_t buf = (uintptr_t)buf_ptr; 4620 size_t img_size = sizeof(struct ElfImage) + debug_frame_size; 4621 DebugFrameHeader *dfh; 4622 4623 img = g_malloc(img_size); 4624 *img = img_template; 4625 4626 img->phdr.p_vaddr = buf; 4627 img->phdr.p_paddr = buf; 4628 img->phdr.p_memsz = buf_size; 4629 4630 img->shdr[1].sh_name = find_string(img->str, ".text"); 4631 img->shdr[1].sh_addr = buf; 4632 img->shdr[1].sh_size = buf_size; 4633 4634 img->shdr[2].sh_name = find_string(img->str, ".debug_info"); 4635 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev"); 4636 4637 img->shdr[4].sh_name = find_string(img->str, ".debug_frame"); 4638 img->shdr[4].sh_size = debug_frame_size; 4639 4640 img->shdr[5].sh_name = find_string(img->str, ".symtab"); 4641 img->shdr[6].sh_name = find_string(img->str, ".strtab"); 4642 4643 img->sym[1].st_name = find_string(img->str, "code_gen_buffer"); 4644 img->sym[1].st_value = buf; 4645 img->sym[1].st_size = buf_size; 4646 4647 img->di.cu_low_pc = buf; 4648 img->di.cu_high_pc = buf + buf_size; 4649 img->di.fn_low_pc = buf; 4650 img->di.fn_high_pc = buf + buf_size; 4651 4652 dfh = (DebugFrameHeader *)(img + 1); 4653 memcpy(dfh, debug_frame, debug_frame_size); 4654 dfh->fde.func_start = buf; 4655 dfh->fde.func_len = buf_size; 4656 4657 #ifdef DEBUG_JIT 4658 /* Enable this block to be able to debug the ELF image file creation. 4659 One can use readelf, objdump, or other inspection utilities. */ 4660 { 4661 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir()); 4662 FILE *f = fopen(jit, "w+b"); 4663 if (f) { 4664 if (fwrite(img, img_size, 1, f) != img_size) { 4665 /* Avoid stupid unused return value warning for fwrite. */ 4666 } 4667 fclose(f); 4668 } 4669 } 4670 #endif 4671 4672 one_entry.symfile_addr = img; 4673 one_entry.symfile_size = img_size; 4674 4675 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN; 4676 __jit_debug_descriptor.relevant_entry = &one_entry; 4677 __jit_debug_descriptor.first_entry = &one_entry; 4678 __jit_debug_register_code(); 4679 } 4680 #else 4681 /* No support for the feature. Provide the entry point expected by exec.c, 4682 and implement the internal function we declared earlier. */ 4683 4684 static void tcg_register_jit_int(const void *buf, size_t size, 4685 const void *debug_frame, 4686 size_t debug_frame_size) 4687 { 4688 } 4689 4690 void tcg_register_jit(const void *buf, size_t buf_size) 4691 { 4692 } 4693 #endif /* ELF_HOST_MACHINE */ 4694 4695 #if !TCG_TARGET_MAYBE_vec 4696 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...) 4697 { 4698 g_assert_not_reached(); 4699 } 4700 #endif 4701