1 /* 2 * Tiny Code Generator for QEMU 3 * 4 * Copyright (c) 2008 Fabrice Bellard 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to deal 8 * in the Software without restriction, including without limitation the rights 9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 10 * copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 22 * THE SOFTWARE. 23 */ 24 25 #include "qemu/osdep.h" 26 27 /* Define to jump the ELF file used to communicate with GDB. */ 28 #undef DEBUG_JIT 29 30 #include "qemu/error-report.h" 31 #include "qemu/cutils.h" 32 #include "qemu/host-utils.h" 33 #include "qemu/qemu-print.h" 34 #include "qemu/cacheflush.h" 35 #include "qemu/cacheinfo.h" 36 #include "qemu/timer.h" 37 #include "exec/translation-block.h" 38 #include "exec/tlb-common.h" 39 #include "tcg/startup.h" 40 #include "tcg/tcg-op-common.h" 41 42 #if UINTPTR_MAX == UINT32_MAX 43 # define ELF_CLASS ELFCLASS32 44 #else 45 # define ELF_CLASS ELFCLASS64 46 #endif 47 #if HOST_BIG_ENDIAN 48 # define ELF_DATA ELFDATA2MSB 49 #else 50 # define ELF_DATA ELFDATA2LSB 51 #endif 52 53 #include "elf.h" 54 #include "exec/log.h" 55 #include "tcg/tcg-ldst.h" 56 #include "tcg/tcg-temp-internal.h" 57 #include "tcg-internal.h" 58 #include "tcg/perf.h" 59 #include "tcg-has.h" 60 #ifdef CONFIG_USER_ONLY 61 #include "user/guest-base.h" 62 #endif 63 64 /* Forward declarations for functions declared in tcg-target.c.inc and 65 used here. */ 66 static void tcg_target_init(TCGContext *s); 67 static void tcg_target_qemu_prologue(TCGContext *s); 68 static bool patch_reloc(tcg_insn_unit *code_ptr, int type, 69 intptr_t value, intptr_t addend); 70 static void tcg_out_nop_fill(tcg_insn_unit *p, int count); 71 72 typedef struct TCGLabelQemuLdst TCGLabelQemuLdst; 73 static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l); 74 static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l); 75 76 /* The CIE and FDE header definitions will be common to all hosts. */ 77 typedef struct { 78 uint32_t len __attribute__((aligned((sizeof(void *))))); 79 uint32_t id; 80 uint8_t version; 81 char augmentation[1]; 82 uint8_t code_align; 83 uint8_t data_align; 84 uint8_t return_column; 85 } DebugFrameCIE; 86 87 typedef struct QEMU_PACKED { 88 uint32_t len __attribute__((aligned((sizeof(void *))))); 89 uint32_t cie_offset; 90 uintptr_t func_start; 91 uintptr_t func_len; 92 } DebugFrameFDEHeader; 93 94 typedef struct QEMU_PACKED { 95 DebugFrameCIE cie; 96 DebugFrameFDEHeader fde; 97 } DebugFrameHeader; 98 99 struct TCGLabelQemuLdst { 100 bool is_ld; /* qemu_ld: true, qemu_st: false */ 101 MemOpIdx oi; 102 TCGType type; /* result type of a load */ 103 TCGReg addr_reg; /* reg index for guest virtual addr */ 104 TCGReg datalo_reg; /* reg index for low word to be loaded or stored */ 105 TCGReg datahi_reg; /* reg index for high word to be loaded or stored */ 106 const tcg_insn_unit *raddr; /* addr of the next IR of qemu_ld/st IR */ 107 tcg_insn_unit *label_ptr[2]; /* label pointers to be updated */ 108 QSIMPLEQ_ENTRY(TCGLabelQemuLdst) next; 109 }; 110 111 static void tcg_register_jit_int(const void *buf, size_t size, 112 const void *debug_frame, 113 size_t debug_frame_size) 114 __attribute__((unused)); 115 116 /* Forward declarations for functions declared and used in tcg-target.c.inc. */ 117 static void tcg_out_tb_start(TCGContext *s); 118 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1, 119 intptr_t arg2); 120 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 121 static void tcg_out_movi(TCGContext *s, TCGType type, 122 TCGReg ret, tcg_target_long arg); 123 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 124 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 125 static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg); 126 static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg); 127 static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg); 128 static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg); 129 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 130 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 131 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg); 132 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long); 133 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2); 134 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg); 135 static void tcg_out_goto_tb(TCGContext *s, int which); 136 static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type, 137 const TCGArg args[TCG_MAX_OP_ARGS], 138 const int const_args[TCG_MAX_OP_ARGS]); 139 #if TCG_TARGET_MAYBE_vec 140 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 141 TCGReg dst, TCGReg src); 142 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 143 TCGReg dst, TCGReg base, intptr_t offset); 144 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 145 TCGReg dst, int64_t arg); 146 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 147 unsigned vecl, unsigned vece, 148 const TCGArg args[TCG_MAX_OP_ARGS], 149 const int const_args[TCG_MAX_OP_ARGS]); 150 #else 151 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 152 TCGReg dst, TCGReg src) 153 { 154 g_assert_not_reached(); 155 } 156 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 157 TCGReg dst, TCGReg base, intptr_t offset) 158 { 159 g_assert_not_reached(); 160 } 161 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 162 TCGReg dst, int64_t arg) 163 { 164 g_assert_not_reached(); 165 } 166 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 167 unsigned vecl, unsigned vece, 168 const TCGArg args[TCG_MAX_OP_ARGS], 169 const int const_args[TCG_MAX_OP_ARGS]) 170 { 171 g_assert_not_reached(); 172 } 173 int tcg_can_emit_vec_op(TCGOpcode o, TCGType t, unsigned ve) 174 { 175 return 0; 176 } 177 #endif 178 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1, 179 intptr_t arg2); 180 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, 181 TCGReg base, intptr_t ofs); 182 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target, 183 const TCGHelperInfo *info); 184 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot); 185 static bool tcg_target_const_match(int64_t val, int ct, 186 TCGType type, TCGCond cond, int vece); 187 188 #ifndef CONFIG_USER_ONLY 189 #define guest_base ({ qemu_build_not_reached(); (uintptr_t)0; }) 190 #endif 191 192 typedef struct TCGLdstHelperParam { 193 TCGReg (*ra_gen)(TCGContext *s, const TCGLabelQemuLdst *l, int arg_reg); 194 unsigned ntmp; 195 int tmp[3]; 196 } TCGLdstHelperParam; 197 198 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 199 const TCGLdstHelperParam *p) 200 __attribute__((unused)); 201 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *l, 202 bool load_sign, const TCGLdstHelperParam *p) 203 __attribute__((unused)); 204 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 205 const TCGLdstHelperParam *p) 206 __attribute__((unused)); 207 208 static void * const qemu_ld_helpers[MO_SSIZE + 1] __attribute__((unused)) = { 209 [MO_UB] = helper_ldub_mmu, 210 [MO_SB] = helper_ldsb_mmu, 211 [MO_UW] = helper_lduw_mmu, 212 [MO_SW] = helper_ldsw_mmu, 213 [MO_UL] = helper_ldul_mmu, 214 [MO_UQ] = helper_ldq_mmu, 215 #if TCG_TARGET_REG_BITS == 64 216 [MO_SL] = helper_ldsl_mmu, 217 [MO_128] = helper_ld16_mmu, 218 #endif 219 }; 220 221 static void * const qemu_st_helpers[MO_SIZE + 1] __attribute__((unused)) = { 222 [MO_8] = helper_stb_mmu, 223 [MO_16] = helper_stw_mmu, 224 [MO_32] = helper_stl_mmu, 225 [MO_64] = helper_stq_mmu, 226 #if TCG_TARGET_REG_BITS == 64 227 [MO_128] = helper_st16_mmu, 228 #endif 229 }; 230 231 typedef struct { 232 MemOp atom; /* lg2 bits of atomicity required */ 233 MemOp align; /* lg2 bits of alignment to use */ 234 } TCGAtomAlign; 235 236 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 237 MemOp host_atom, bool allow_two_ops) 238 __attribute__((unused)); 239 240 #ifdef CONFIG_USER_ONLY 241 bool tcg_use_softmmu; 242 #endif 243 244 TCGContext tcg_init_ctx; 245 __thread TCGContext *tcg_ctx; 246 247 TCGContext **tcg_ctxs; 248 unsigned int tcg_cur_ctxs; 249 unsigned int tcg_max_ctxs; 250 TCGv_env tcg_env; 251 const void *tcg_code_gen_epilogue; 252 uintptr_t tcg_splitwx_diff; 253 254 #ifndef CONFIG_TCG_INTERPRETER 255 tcg_prologue_fn *tcg_qemu_tb_exec; 256 #endif 257 258 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT]; 259 static TCGRegSet tcg_target_call_clobber_regs; 260 261 #if TCG_TARGET_INSN_UNIT_SIZE == 1 262 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v) 263 { 264 *s->code_ptr++ = v; 265 } 266 267 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p, 268 uint8_t v) 269 { 270 *p = v; 271 } 272 #endif 273 274 #if TCG_TARGET_INSN_UNIT_SIZE <= 2 275 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v) 276 { 277 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 278 *s->code_ptr++ = v; 279 } else { 280 tcg_insn_unit *p = s->code_ptr; 281 memcpy(p, &v, sizeof(v)); 282 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE); 283 } 284 } 285 286 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p, 287 uint16_t v) 288 { 289 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 290 *p = v; 291 } else { 292 memcpy(p, &v, sizeof(v)); 293 } 294 } 295 #endif 296 297 #if TCG_TARGET_INSN_UNIT_SIZE <= 4 298 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v) 299 { 300 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 301 *s->code_ptr++ = v; 302 } else { 303 tcg_insn_unit *p = s->code_ptr; 304 memcpy(p, &v, sizeof(v)); 305 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE); 306 } 307 } 308 309 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p, 310 uint32_t v) 311 { 312 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 313 *p = v; 314 } else { 315 memcpy(p, &v, sizeof(v)); 316 } 317 } 318 #endif 319 320 #if TCG_TARGET_INSN_UNIT_SIZE <= 8 321 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v) 322 { 323 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 324 *s->code_ptr++ = v; 325 } else { 326 tcg_insn_unit *p = s->code_ptr; 327 memcpy(p, &v, sizeof(v)); 328 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE); 329 } 330 } 331 332 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p, 333 uint64_t v) 334 { 335 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 336 *p = v; 337 } else { 338 memcpy(p, &v, sizeof(v)); 339 } 340 } 341 #endif 342 343 /* label relocation processing */ 344 345 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type, 346 TCGLabel *l, intptr_t addend) 347 { 348 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation)); 349 350 r->type = type; 351 r->ptr = code_ptr; 352 r->addend = addend; 353 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next); 354 } 355 356 static void tcg_out_label(TCGContext *s, TCGLabel *l) 357 { 358 tcg_debug_assert(!l->has_value); 359 l->has_value = 1; 360 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr); 361 } 362 363 TCGLabel *gen_new_label(void) 364 { 365 TCGContext *s = tcg_ctx; 366 TCGLabel *l = tcg_malloc(sizeof(TCGLabel)); 367 368 memset(l, 0, sizeof(TCGLabel)); 369 l->id = s->nb_labels++; 370 QSIMPLEQ_INIT(&l->branches); 371 QSIMPLEQ_INIT(&l->relocs); 372 373 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next); 374 375 return l; 376 } 377 378 static bool tcg_resolve_relocs(TCGContext *s) 379 { 380 TCGLabel *l; 381 382 QSIMPLEQ_FOREACH(l, &s->labels, next) { 383 TCGRelocation *r; 384 uintptr_t value = l->u.value; 385 386 QSIMPLEQ_FOREACH(r, &l->relocs, next) { 387 if (!patch_reloc(r->ptr, r->type, value, r->addend)) { 388 return false; 389 } 390 } 391 } 392 return true; 393 } 394 395 static void set_jmp_reset_offset(TCGContext *s, int which) 396 { 397 /* 398 * We will check for overflow at the end of the opcode loop in 399 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 400 */ 401 s->gen_tb->jmp_reset_offset[which] = tcg_current_code_size(s); 402 } 403 404 static void G_GNUC_UNUSED set_jmp_insn_offset(TCGContext *s, int which) 405 { 406 /* 407 * We will check for overflow at the end of the opcode loop in 408 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 409 */ 410 s->gen_tb->jmp_insn_offset[which] = tcg_current_code_size(s); 411 } 412 413 static uintptr_t G_GNUC_UNUSED get_jmp_target_addr(TCGContext *s, int which) 414 { 415 /* 416 * Return the read-execute version of the pointer, for the benefit 417 * of any pc-relative addressing mode. 418 */ 419 return (uintptr_t)tcg_splitwx_to_rx(&s->gen_tb->jmp_target_addr[which]); 420 } 421 422 static int __attribute__((unused)) 423 tlb_mask_table_ofs(TCGContext *s, int which) 424 { 425 return (offsetof(CPUNegativeOffsetState, tlb.f[which]) - 426 sizeof(CPUNegativeOffsetState)); 427 } 428 429 /* Signal overflow, starting over with fewer guest insns. */ 430 static G_NORETURN 431 void tcg_raise_tb_overflow(TCGContext *s) 432 { 433 siglongjmp(s->jmp_trans, -2); 434 } 435 436 /* 437 * Used by tcg_out_movext{1,2} to hold the arguments for tcg_out_movext. 438 * By the time we arrive at tcg_out_movext1, @dst is always a TCGReg. 439 * 440 * However, tcg_out_helper_load_slots reuses this field to hold an 441 * argument slot number (which may designate a argument register or an 442 * argument stack slot), converting to TCGReg once all arguments that 443 * are destined for the stack are processed. 444 */ 445 typedef struct TCGMovExtend { 446 unsigned dst; 447 TCGReg src; 448 TCGType dst_type; 449 TCGType src_type; 450 MemOp src_ext; 451 } TCGMovExtend; 452 453 /** 454 * tcg_out_movext -- move and extend 455 * @s: tcg context 456 * @dst_type: integral type for destination 457 * @dst: destination register 458 * @src_type: integral type for source 459 * @src_ext: extension to apply to source 460 * @src: source register 461 * 462 * Move or extend @src into @dst, depending on @src_ext and the types. 463 */ 464 static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst, 465 TCGType src_type, MemOp src_ext, TCGReg src) 466 { 467 switch (src_ext) { 468 case MO_UB: 469 tcg_out_ext8u(s, dst, src); 470 break; 471 case MO_SB: 472 tcg_out_ext8s(s, dst_type, dst, src); 473 break; 474 case MO_UW: 475 tcg_out_ext16u(s, dst, src); 476 break; 477 case MO_SW: 478 tcg_out_ext16s(s, dst_type, dst, src); 479 break; 480 case MO_UL: 481 case MO_SL: 482 if (dst_type == TCG_TYPE_I32) { 483 if (src_type == TCG_TYPE_I32) { 484 tcg_out_mov(s, TCG_TYPE_I32, dst, src); 485 } else { 486 tcg_out_extrl_i64_i32(s, dst, src); 487 } 488 } else if (src_type == TCG_TYPE_I32) { 489 if (src_ext & MO_SIGN) { 490 tcg_out_exts_i32_i64(s, dst, src); 491 } else { 492 tcg_out_extu_i32_i64(s, dst, src); 493 } 494 } else { 495 if (src_ext & MO_SIGN) { 496 tcg_out_ext32s(s, dst, src); 497 } else { 498 tcg_out_ext32u(s, dst, src); 499 } 500 } 501 break; 502 case MO_UQ: 503 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 504 if (dst_type == TCG_TYPE_I32) { 505 tcg_out_extrl_i64_i32(s, dst, src); 506 } else { 507 tcg_out_mov(s, TCG_TYPE_I64, dst, src); 508 } 509 break; 510 default: 511 g_assert_not_reached(); 512 } 513 } 514 515 /* Minor variations on a theme, using a structure. */ 516 static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i, 517 TCGReg src) 518 { 519 tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src); 520 } 521 522 static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i) 523 { 524 tcg_out_movext1_new_src(s, i, i->src); 525 } 526 527 /** 528 * tcg_out_movext2 -- move and extend two pair 529 * @s: tcg context 530 * @i1: first move description 531 * @i2: second move description 532 * @scratch: temporary register, or -1 for none 533 * 534 * As tcg_out_movext, for both @i1 and @i2, caring for overlap 535 * between the sources and destinations. 536 */ 537 538 static void tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1, 539 const TCGMovExtend *i2, int scratch) 540 { 541 TCGReg src1 = i1->src; 542 TCGReg src2 = i2->src; 543 544 if (i1->dst != src2) { 545 tcg_out_movext1(s, i1); 546 tcg_out_movext1(s, i2); 547 return; 548 } 549 if (i2->dst == src1) { 550 TCGType src1_type = i1->src_type; 551 TCGType src2_type = i2->src_type; 552 553 if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) { 554 /* The data is now in the correct registers, now extend. */ 555 src1 = i2->src; 556 src2 = i1->src; 557 } else { 558 tcg_debug_assert(scratch >= 0); 559 tcg_out_mov(s, src1_type, scratch, src1); 560 src1 = scratch; 561 } 562 } 563 tcg_out_movext1_new_src(s, i2, src2); 564 tcg_out_movext1_new_src(s, i1, src1); 565 } 566 567 /** 568 * tcg_out_movext3 -- move and extend three pair 569 * @s: tcg context 570 * @i1: first move description 571 * @i2: second move description 572 * @i3: third move description 573 * @scratch: temporary register, or -1 for none 574 * 575 * As tcg_out_movext, for all of @i1, @i2 and @i3, caring for overlap 576 * between the sources and destinations. 577 */ 578 579 static void tcg_out_movext3(TCGContext *s, const TCGMovExtend *i1, 580 const TCGMovExtend *i2, const TCGMovExtend *i3, 581 int scratch) 582 { 583 TCGReg src1 = i1->src; 584 TCGReg src2 = i2->src; 585 TCGReg src3 = i3->src; 586 587 if (i1->dst != src2 && i1->dst != src3) { 588 tcg_out_movext1(s, i1); 589 tcg_out_movext2(s, i2, i3, scratch); 590 return; 591 } 592 if (i2->dst != src1 && i2->dst != src3) { 593 tcg_out_movext1(s, i2); 594 tcg_out_movext2(s, i1, i3, scratch); 595 return; 596 } 597 if (i3->dst != src1 && i3->dst != src2) { 598 tcg_out_movext1(s, i3); 599 tcg_out_movext2(s, i1, i2, scratch); 600 return; 601 } 602 603 /* 604 * There is a cycle. Since there are only 3 nodes, the cycle is 605 * either "clockwise" or "anti-clockwise", and can be solved with 606 * a single scratch or two xchg. 607 */ 608 if (i1->dst == src2 && i2->dst == src3 && i3->dst == src1) { 609 /* "Clockwise" */ 610 if (tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2)) { 611 tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3); 612 /* The data is now in the correct registers, now extend. */ 613 tcg_out_movext1_new_src(s, i1, i1->dst); 614 tcg_out_movext1_new_src(s, i2, i2->dst); 615 tcg_out_movext1_new_src(s, i3, i3->dst); 616 } else { 617 tcg_debug_assert(scratch >= 0); 618 tcg_out_mov(s, i1->src_type, scratch, src1); 619 tcg_out_movext1(s, i3); 620 tcg_out_movext1(s, i2); 621 tcg_out_movext1_new_src(s, i1, scratch); 622 } 623 } else if (i1->dst == src3 && i2->dst == src1 && i3->dst == src2) { 624 /* "Anti-clockwise" */ 625 if (tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3)) { 626 tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2); 627 /* The data is now in the correct registers, now extend. */ 628 tcg_out_movext1_new_src(s, i1, i1->dst); 629 tcg_out_movext1_new_src(s, i2, i2->dst); 630 tcg_out_movext1_new_src(s, i3, i3->dst); 631 } else { 632 tcg_debug_assert(scratch >= 0); 633 tcg_out_mov(s, i1->src_type, scratch, src1); 634 tcg_out_movext1(s, i2); 635 tcg_out_movext1(s, i3); 636 tcg_out_movext1_new_src(s, i1, scratch); 637 } 638 } else { 639 g_assert_not_reached(); 640 } 641 } 642 643 /* 644 * Allocate a new TCGLabelQemuLdst entry. 645 */ 646 647 __attribute__((unused)) 648 static TCGLabelQemuLdst *new_ldst_label(TCGContext *s) 649 { 650 TCGLabelQemuLdst *l = tcg_malloc(sizeof(*l)); 651 652 memset(l, 0, sizeof(*l)); 653 QSIMPLEQ_INSERT_TAIL(&s->ldst_labels, l, next); 654 655 return l; 656 } 657 658 /* 659 * Allocate new constant pool entries. 660 */ 661 662 typedef struct TCGLabelPoolData { 663 struct TCGLabelPoolData *next; 664 tcg_insn_unit *label; 665 intptr_t addend; 666 int rtype; 667 unsigned nlong; 668 tcg_target_ulong data[]; 669 } TCGLabelPoolData; 670 671 static TCGLabelPoolData *new_pool_alloc(TCGContext *s, int nlong, int rtype, 672 tcg_insn_unit *label, intptr_t addend) 673 { 674 TCGLabelPoolData *n = tcg_malloc(sizeof(TCGLabelPoolData) 675 + sizeof(tcg_target_ulong) * nlong); 676 677 n->label = label; 678 n->addend = addend; 679 n->rtype = rtype; 680 n->nlong = nlong; 681 return n; 682 } 683 684 static void new_pool_insert(TCGContext *s, TCGLabelPoolData *n) 685 { 686 TCGLabelPoolData *i, **pp; 687 int nlong = n->nlong; 688 689 /* Insertion sort on the pool. */ 690 for (pp = &s->pool_labels; (i = *pp) != NULL; pp = &i->next) { 691 if (nlong > i->nlong) { 692 break; 693 } 694 if (nlong < i->nlong) { 695 continue; 696 } 697 if (memcmp(n->data, i->data, sizeof(tcg_target_ulong) * nlong) >= 0) { 698 break; 699 } 700 } 701 n->next = *pp; 702 *pp = n; 703 } 704 705 /* The "usual" for generic integer code. */ 706 __attribute__((unused)) 707 static void new_pool_label(TCGContext *s, tcg_target_ulong d, int rtype, 708 tcg_insn_unit *label, intptr_t addend) 709 { 710 TCGLabelPoolData *n = new_pool_alloc(s, 1, rtype, label, addend); 711 n->data[0] = d; 712 new_pool_insert(s, n); 713 } 714 715 /* For v64 or v128, depending on the host. */ 716 __attribute__((unused)) 717 static void new_pool_l2(TCGContext *s, int rtype, tcg_insn_unit *label, 718 intptr_t addend, tcg_target_ulong d0, 719 tcg_target_ulong d1) 720 { 721 TCGLabelPoolData *n = new_pool_alloc(s, 2, rtype, label, addend); 722 n->data[0] = d0; 723 n->data[1] = d1; 724 new_pool_insert(s, n); 725 } 726 727 /* For v128 or v256, depending on the host. */ 728 __attribute__((unused)) 729 static void new_pool_l4(TCGContext *s, int rtype, tcg_insn_unit *label, 730 intptr_t addend, tcg_target_ulong d0, 731 tcg_target_ulong d1, tcg_target_ulong d2, 732 tcg_target_ulong d3) 733 { 734 TCGLabelPoolData *n = new_pool_alloc(s, 4, rtype, label, addend); 735 n->data[0] = d0; 736 n->data[1] = d1; 737 n->data[2] = d2; 738 n->data[3] = d3; 739 new_pool_insert(s, n); 740 } 741 742 /* For v256, for 32-bit host. */ 743 __attribute__((unused)) 744 static void new_pool_l8(TCGContext *s, int rtype, tcg_insn_unit *label, 745 intptr_t addend, tcg_target_ulong d0, 746 tcg_target_ulong d1, tcg_target_ulong d2, 747 tcg_target_ulong d3, tcg_target_ulong d4, 748 tcg_target_ulong d5, tcg_target_ulong d6, 749 tcg_target_ulong d7) 750 { 751 TCGLabelPoolData *n = new_pool_alloc(s, 8, rtype, label, addend); 752 n->data[0] = d0; 753 n->data[1] = d1; 754 n->data[2] = d2; 755 n->data[3] = d3; 756 n->data[4] = d4; 757 n->data[5] = d5; 758 n->data[6] = d6; 759 n->data[7] = d7; 760 new_pool_insert(s, n); 761 } 762 763 /* 764 * Generate TB finalization at the end of block 765 */ 766 767 static int tcg_out_ldst_finalize(TCGContext *s) 768 { 769 TCGLabelQemuLdst *lb; 770 771 /* qemu_ld/st slow paths */ 772 QSIMPLEQ_FOREACH(lb, &s->ldst_labels, next) { 773 if (lb->is_ld 774 ? !tcg_out_qemu_ld_slow_path(s, lb) 775 : !tcg_out_qemu_st_slow_path(s, lb)) { 776 return -2; 777 } 778 779 /* 780 * Test for (pending) buffer overflow. The assumption is that any 781 * one operation beginning below the high water mark cannot overrun 782 * the buffer completely. Thus we can test for overflow after 783 * generating code without having to check during generation. 784 */ 785 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) { 786 return -1; 787 } 788 } 789 return 0; 790 } 791 792 static int tcg_out_pool_finalize(TCGContext *s) 793 { 794 TCGLabelPoolData *p = s->pool_labels; 795 TCGLabelPoolData *l = NULL; 796 void *a; 797 798 if (p == NULL) { 799 return 0; 800 } 801 802 /* 803 * ??? Round up to qemu_icache_linesize, but then do not round 804 * again when allocating the next TranslationBlock structure. 805 */ 806 a = (void *)ROUND_UP((uintptr_t)s->code_ptr, 807 sizeof(tcg_target_ulong) * p->nlong); 808 tcg_out_nop_fill(s->code_ptr, (tcg_insn_unit *)a - s->code_ptr); 809 s->data_gen_ptr = a; 810 811 for (; p != NULL; p = p->next) { 812 size_t size = sizeof(tcg_target_ulong) * p->nlong; 813 uintptr_t value; 814 815 if (!l || l->nlong != p->nlong || memcmp(l->data, p->data, size)) { 816 if (unlikely(a > s->code_gen_highwater)) { 817 return -1; 818 } 819 memcpy(a, p->data, size); 820 a += size; 821 l = p; 822 } 823 824 value = (uintptr_t)tcg_splitwx_to_rx(a) - size; 825 if (!patch_reloc(p->label, p->rtype, value, p->addend)) { 826 return -2; 827 } 828 } 829 830 s->code_ptr = a; 831 return 0; 832 } 833 834 #define C_PFX1(P, A) P##A 835 #define C_PFX2(P, A, B) P##A##_##B 836 #define C_PFX3(P, A, B, C) P##A##_##B##_##C 837 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D 838 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E 839 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F 840 841 /* Define an enumeration for the various combinations. */ 842 843 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1), 844 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2), 845 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3), 846 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4), 847 848 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1), 849 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2), 850 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3), 851 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4), 852 853 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2), 854 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1), 855 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1), 856 857 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1), 858 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2), 859 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3), 860 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4), 861 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4), 862 863 typedef enum { 864 C_Dynamic = -2, 865 C_NotImplemented = -1, 866 #include "tcg-target-con-set.h" 867 } TCGConstraintSetIndex; 868 869 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode, TCGType, unsigned); 870 871 #undef C_O0_I1 872 #undef C_O0_I2 873 #undef C_O0_I3 874 #undef C_O0_I4 875 #undef C_O1_I1 876 #undef C_O1_I2 877 #undef C_O1_I3 878 #undef C_O1_I4 879 #undef C_N1_I2 880 #undef C_N1O1_I1 881 #undef C_N2_I1 882 #undef C_O2_I1 883 #undef C_O2_I2 884 #undef C_O2_I3 885 #undef C_O2_I4 886 #undef C_N1_O1_I4 887 888 /* Put all of the constraint sets into an array, indexed by the enum. */ 889 890 typedef struct TCGConstraintSet { 891 uint8_t nb_oargs, nb_iargs; 892 const char *args_ct_str[TCG_MAX_OP_ARGS]; 893 } TCGConstraintSet; 894 895 #define C_O0_I1(I1) { 0, 1, { #I1 } }, 896 #define C_O0_I2(I1, I2) { 0, 2, { #I1, #I2 } }, 897 #define C_O0_I3(I1, I2, I3) { 0, 3, { #I1, #I2, #I3 } }, 898 #define C_O0_I4(I1, I2, I3, I4) { 0, 4, { #I1, #I2, #I3, #I4 } }, 899 900 #define C_O1_I1(O1, I1) { 1, 1, { #O1, #I1 } }, 901 #define C_O1_I2(O1, I1, I2) { 1, 2, { #O1, #I1, #I2 } }, 902 #define C_O1_I3(O1, I1, I2, I3) { 1, 3, { #O1, #I1, #I2, #I3 } }, 903 #define C_O1_I4(O1, I1, I2, I3, I4) { 1, 4, { #O1, #I1, #I2, #I3, #I4 } }, 904 905 #define C_N1_I2(O1, I1, I2) { 1, 2, { "&" #O1, #I1, #I2 } }, 906 #define C_N1O1_I1(O1, O2, I1) { 2, 1, { "&" #O1, #O2, #I1 } }, 907 #define C_N2_I1(O1, O2, I1) { 2, 1, { "&" #O1, "&" #O2, #I1 } }, 908 909 #define C_O2_I1(O1, O2, I1) { 2, 1, { #O1, #O2, #I1 } }, 910 #define C_O2_I2(O1, O2, I1, I2) { 2, 2, { #O1, #O2, #I1, #I2 } }, 911 #define C_O2_I3(O1, O2, I1, I2, I3) { 2, 3, { #O1, #O2, #I1, #I2, #I3 } }, 912 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { 2, 4, { #O1, #O2, #I1, #I2, #I3, #I4 } }, 913 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) { 2, 4, { "&" #O1, #O2, #I1, #I2, #I3, #I4 } }, 914 915 static const TCGConstraintSet constraint_sets[] = { 916 #include "tcg-target-con-set.h" 917 }; 918 919 #undef C_O0_I1 920 #undef C_O0_I2 921 #undef C_O0_I3 922 #undef C_O0_I4 923 #undef C_O1_I1 924 #undef C_O1_I2 925 #undef C_O1_I3 926 #undef C_O1_I4 927 #undef C_N1_I2 928 #undef C_N1O1_I1 929 #undef C_N2_I1 930 #undef C_O2_I1 931 #undef C_O2_I2 932 #undef C_O2_I3 933 #undef C_O2_I4 934 #undef C_N1_O1_I4 935 936 /* Expand the enumerator to be returned from tcg_target_op_def(). */ 937 938 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1) 939 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2) 940 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3) 941 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4) 942 943 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1) 944 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2) 945 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3) 946 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4) 947 948 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2) 949 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1) 950 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1) 951 952 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1) 953 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2) 954 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3) 955 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4) 956 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4) 957 958 /* 959 * TCGOutOp is the base class for a set of structures that describe how 960 * to generate code for a given TCGOpcode. 961 * 962 * @static_constraint: 963 * C_NotImplemented: The TCGOpcode is not supported by the backend. 964 * C_Dynamic: Use @dynamic_constraint to select a constraint set 965 * based on any of @type, @flags, or host isa. 966 * Otherwise: The register allocation constrains for the TCGOpcode. 967 * 968 * Subclasses of TCGOutOp will define a set of output routines that may 969 * be used. Such routines will often be selected by the set of registers 970 * and constants that come out of register allocation. The set of 971 * routines that are provided will guide the set of constraints that are 972 * legal. In particular, assume that tcg_optimize() has done its job in 973 * swapping commutative operands and folding operations for which all 974 * operands are constant. 975 */ 976 typedef struct TCGOutOp { 977 TCGConstraintSetIndex static_constraint; 978 TCGConstraintSetIndex (*dynamic_constraint)(TCGType type, unsigned flags); 979 } TCGOutOp; 980 981 typedef struct TCGOutOpBinary { 982 TCGOutOp base; 983 void (*out_rrr)(TCGContext *s, TCGType type, 984 TCGReg a0, TCGReg a1, TCGReg a2); 985 void (*out_rri)(TCGContext *s, TCGType type, 986 TCGReg a0, TCGReg a1, tcg_target_long a2); 987 } TCGOutOpBinary; 988 989 #include "tcg-target.c.inc" 990 991 #ifndef CONFIG_TCG_INTERPRETER 992 /* Validate CPUTLBDescFast placement. */ 993 QEMU_BUILD_BUG_ON((int)(offsetof(CPUNegativeOffsetState, tlb.f[0]) - 994 sizeof(CPUNegativeOffsetState)) 995 < MIN_TLB_MASK_TABLE_OFS); 996 #endif 997 998 /* 999 * Register V as the TCGOutOp for O. 1000 * This verifies that V is of type T, otherwise give a nice compiler error. 1001 * This prevents trivial mistakes within each arch/tcg-target.c.inc. 1002 */ 1003 #define OUTOP(O, T, V) [O] = _Generic(V, T: &V.base) 1004 1005 /* Register allocation descriptions for every TCGOpcode. */ 1006 static const TCGOutOp * const all_outop[NB_OPS] = { 1007 OUTOP(INDEX_op_add, TCGOutOpBinary, outop_add), 1008 OUTOP(INDEX_op_and, TCGOutOpBinary, outop_and), 1009 OUTOP(INDEX_op_andc, TCGOutOpBinary, outop_andc), 1010 }; 1011 1012 #undef OUTOP 1013 1014 /* 1015 * All TCG threads except the parent (i.e. the one that called tcg_context_init 1016 * and registered the target's TCG globals) must register with this function 1017 * before initiating translation. 1018 * 1019 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation 1020 * of tcg_region_init() for the reasoning behind this. 1021 * 1022 * In system-mode each caller registers its context in tcg_ctxs[]. Note that in 1023 * system-mode tcg_ctxs[] does not track tcg_ctx_init, since the initial context 1024 * is not used anymore for translation once this function is called. 1025 * 1026 * Not tracking tcg_init_ctx in tcg_ctxs[] in system-mode keeps code that 1027 * iterates over the array (e.g. tcg_code_size() the same for both system/user 1028 * modes. 1029 */ 1030 #ifdef CONFIG_USER_ONLY 1031 void tcg_register_thread(void) 1032 { 1033 tcg_ctx = &tcg_init_ctx; 1034 } 1035 #else 1036 void tcg_register_thread(void) 1037 { 1038 TCGContext *s = g_malloc(sizeof(*s)); 1039 unsigned int i, n; 1040 1041 *s = tcg_init_ctx; 1042 1043 /* Relink mem_base. */ 1044 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) { 1045 if (tcg_init_ctx.temps[i].mem_base) { 1046 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps; 1047 tcg_debug_assert(b >= 0 && b < n); 1048 s->temps[i].mem_base = &s->temps[b]; 1049 } 1050 } 1051 1052 /* Claim an entry in tcg_ctxs */ 1053 n = qatomic_fetch_inc(&tcg_cur_ctxs); 1054 g_assert(n < tcg_max_ctxs); 1055 qatomic_set(&tcg_ctxs[n], s); 1056 1057 if (n > 0) { 1058 tcg_region_initial_alloc(s); 1059 } 1060 1061 tcg_ctx = s; 1062 } 1063 #endif /* !CONFIG_USER_ONLY */ 1064 1065 /* pool based memory allocation */ 1066 void *tcg_malloc_internal(TCGContext *s, int size) 1067 { 1068 TCGPool *p; 1069 int pool_size; 1070 1071 if (size > TCG_POOL_CHUNK_SIZE) { 1072 /* big malloc: insert a new pool (XXX: could optimize) */ 1073 p = g_malloc(sizeof(TCGPool) + size); 1074 p->size = size; 1075 p->next = s->pool_first_large; 1076 s->pool_first_large = p; 1077 return p->data; 1078 } else { 1079 p = s->pool_current; 1080 if (!p) { 1081 p = s->pool_first; 1082 if (!p) 1083 goto new_pool; 1084 } else { 1085 if (!p->next) { 1086 new_pool: 1087 pool_size = TCG_POOL_CHUNK_SIZE; 1088 p = g_malloc(sizeof(TCGPool) + pool_size); 1089 p->size = pool_size; 1090 p->next = NULL; 1091 if (s->pool_current) { 1092 s->pool_current->next = p; 1093 } else { 1094 s->pool_first = p; 1095 } 1096 } else { 1097 p = p->next; 1098 } 1099 } 1100 } 1101 s->pool_current = p; 1102 s->pool_cur = p->data + size; 1103 s->pool_end = p->data + p->size; 1104 return p->data; 1105 } 1106 1107 void tcg_pool_reset(TCGContext *s) 1108 { 1109 TCGPool *p, *t; 1110 for (p = s->pool_first_large; p; p = t) { 1111 t = p->next; 1112 g_free(p); 1113 } 1114 s->pool_first_large = NULL; 1115 s->pool_cur = s->pool_end = NULL; 1116 s->pool_current = NULL; 1117 } 1118 1119 /* 1120 * Create TCGHelperInfo structures for "tcg/tcg-ldst.h" functions, 1121 * akin to what "exec/helper-tcg.h" does with DEF_HELPER_FLAGS_N. 1122 * We only use these for layout in tcg_out_ld_helper_ret and 1123 * tcg_out_st_helper_args, and share them between several of 1124 * the helpers, with the end result that it's easier to build manually. 1125 */ 1126 1127 #if TCG_TARGET_REG_BITS == 32 1128 # define dh_typecode_ttl dh_typecode_i32 1129 #else 1130 # define dh_typecode_ttl dh_typecode_i64 1131 #endif 1132 1133 static TCGHelperInfo info_helper_ld32_mmu = { 1134 .flags = TCG_CALL_NO_WG, 1135 .typemask = dh_typemask(ttl, 0) /* return tcg_target_ulong */ 1136 | dh_typemask(env, 1) 1137 | dh_typemask(i64, 2) /* uint64_t addr */ 1138 | dh_typemask(i32, 3) /* unsigned oi */ 1139 | dh_typemask(ptr, 4) /* uintptr_t ra */ 1140 }; 1141 1142 static TCGHelperInfo info_helper_ld64_mmu = { 1143 .flags = TCG_CALL_NO_WG, 1144 .typemask = dh_typemask(i64, 0) /* return uint64_t */ 1145 | dh_typemask(env, 1) 1146 | dh_typemask(i64, 2) /* uint64_t addr */ 1147 | dh_typemask(i32, 3) /* unsigned oi */ 1148 | dh_typemask(ptr, 4) /* uintptr_t ra */ 1149 }; 1150 1151 static TCGHelperInfo info_helper_ld128_mmu = { 1152 .flags = TCG_CALL_NO_WG, 1153 .typemask = dh_typemask(i128, 0) /* return Int128 */ 1154 | dh_typemask(env, 1) 1155 | dh_typemask(i64, 2) /* uint64_t addr */ 1156 | dh_typemask(i32, 3) /* unsigned oi */ 1157 | dh_typemask(ptr, 4) /* uintptr_t ra */ 1158 }; 1159 1160 static TCGHelperInfo info_helper_st32_mmu = { 1161 .flags = TCG_CALL_NO_WG, 1162 .typemask = dh_typemask(void, 0) 1163 | dh_typemask(env, 1) 1164 | dh_typemask(i64, 2) /* uint64_t addr */ 1165 | dh_typemask(i32, 3) /* uint32_t data */ 1166 | dh_typemask(i32, 4) /* unsigned oi */ 1167 | dh_typemask(ptr, 5) /* uintptr_t ra */ 1168 }; 1169 1170 static TCGHelperInfo info_helper_st64_mmu = { 1171 .flags = TCG_CALL_NO_WG, 1172 .typemask = dh_typemask(void, 0) 1173 | dh_typemask(env, 1) 1174 | dh_typemask(i64, 2) /* uint64_t addr */ 1175 | dh_typemask(i64, 3) /* uint64_t data */ 1176 | dh_typemask(i32, 4) /* unsigned oi */ 1177 | dh_typemask(ptr, 5) /* uintptr_t ra */ 1178 }; 1179 1180 static TCGHelperInfo info_helper_st128_mmu = { 1181 .flags = TCG_CALL_NO_WG, 1182 .typemask = dh_typemask(void, 0) 1183 | dh_typemask(env, 1) 1184 | dh_typemask(i64, 2) /* uint64_t addr */ 1185 | dh_typemask(i128, 3) /* Int128 data */ 1186 | dh_typemask(i32, 4) /* unsigned oi */ 1187 | dh_typemask(ptr, 5) /* uintptr_t ra */ 1188 }; 1189 1190 #ifdef CONFIG_TCG_INTERPRETER 1191 static ffi_type *typecode_to_ffi(int argmask) 1192 { 1193 /* 1194 * libffi does not support __int128_t, so we have forced Int128 1195 * to use the structure definition instead of the builtin type. 1196 */ 1197 static ffi_type *ffi_type_i128_elements[3] = { 1198 &ffi_type_uint64, 1199 &ffi_type_uint64, 1200 NULL 1201 }; 1202 static ffi_type ffi_type_i128 = { 1203 .size = 16, 1204 .alignment = __alignof__(Int128), 1205 .type = FFI_TYPE_STRUCT, 1206 .elements = ffi_type_i128_elements, 1207 }; 1208 1209 switch (argmask) { 1210 case dh_typecode_void: 1211 return &ffi_type_void; 1212 case dh_typecode_i32: 1213 return &ffi_type_uint32; 1214 case dh_typecode_s32: 1215 return &ffi_type_sint32; 1216 case dh_typecode_i64: 1217 return &ffi_type_uint64; 1218 case dh_typecode_s64: 1219 return &ffi_type_sint64; 1220 case dh_typecode_ptr: 1221 return &ffi_type_pointer; 1222 case dh_typecode_i128: 1223 return &ffi_type_i128; 1224 } 1225 g_assert_not_reached(); 1226 } 1227 1228 static ffi_cif *init_ffi_layout(TCGHelperInfo *info) 1229 { 1230 unsigned typemask = info->typemask; 1231 struct { 1232 ffi_cif cif; 1233 ffi_type *args[]; 1234 } *ca; 1235 ffi_status status; 1236 int nargs; 1237 1238 /* Ignoring the return type, find the last non-zero field. */ 1239 nargs = 32 - clz32(typemask >> 3); 1240 nargs = DIV_ROUND_UP(nargs, 3); 1241 assert(nargs <= MAX_CALL_IARGS); 1242 1243 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *)); 1244 ca->cif.rtype = typecode_to_ffi(typemask & 7); 1245 ca->cif.nargs = nargs; 1246 1247 if (nargs != 0) { 1248 ca->cif.arg_types = ca->args; 1249 for (int j = 0; j < nargs; ++j) { 1250 int typecode = extract32(typemask, (j + 1) * 3, 3); 1251 ca->args[j] = typecode_to_ffi(typecode); 1252 } 1253 } 1254 1255 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs, 1256 ca->cif.rtype, ca->cif.arg_types); 1257 assert(status == FFI_OK); 1258 1259 return &ca->cif; 1260 } 1261 1262 #define HELPER_INFO_INIT(I) (&(I)->cif) 1263 #define HELPER_INFO_INIT_VAL(I) init_ffi_layout(I) 1264 #else 1265 #define HELPER_INFO_INIT(I) (&(I)->init) 1266 #define HELPER_INFO_INIT_VAL(I) 1 1267 #endif /* CONFIG_TCG_INTERPRETER */ 1268 1269 static inline bool arg_slot_reg_p(unsigned arg_slot) 1270 { 1271 /* 1272 * Split the sizeof away from the comparison to avoid Werror from 1273 * "unsigned < 0 is always false", when iarg_regs is empty. 1274 */ 1275 unsigned nreg = ARRAY_SIZE(tcg_target_call_iarg_regs); 1276 return arg_slot < nreg; 1277 } 1278 1279 static inline int arg_slot_stk_ofs(unsigned arg_slot) 1280 { 1281 unsigned max = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1282 unsigned stk_slot = arg_slot - ARRAY_SIZE(tcg_target_call_iarg_regs); 1283 1284 tcg_debug_assert(stk_slot < max); 1285 return TCG_TARGET_CALL_STACK_OFFSET + stk_slot * sizeof(tcg_target_long); 1286 } 1287 1288 typedef struct TCGCumulativeArgs { 1289 int arg_idx; /* tcg_gen_callN args[] */ 1290 int info_in_idx; /* TCGHelperInfo in[] */ 1291 int arg_slot; /* regs+stack slot */ 1292 int ref_slot; /* stack slots for references */ 1293 } TCGCumulativeArgs; 1294 1295 static void layout_arg_even(TCGCumulativeArgs *cum) 1296 { 1297 cum->arg_slot += cum->arg_slot & 1; 1298 } 1299 1300 static void layout_arg_1(TCGCumulativeArgs *cum, TCGHelperInfo *info, 1301 TCGCallArgumentKind kind) 1302 { 1303 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1304 1305 *loc = (TCGCallArgumentLoc){ 1306 .kind = kind, 1307 .arg_idx = cum->arg_idx, 1308 .arg_slot = cum->arg_slot, 1309 }; 1310 cum->info_in_idx++; 1311 cum->arg_slot++; 1312 } 1313 1314 static void layout_arg_normal_n(TCGCumulativeArgs *cum, 1315 TCGHelperInfo *info, int n) 1316 { 1317 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1318 1319 for (int i = 0; i < n; ++i) { 1320 /* Layout all using the same arg_idx, adjusting the subindex. */ 1321 loc[i] = (TCGCallArgumentLoc){ 1322 .kind = TCG_CALL_ARG_NORMAL, 1323 .arg_idx = cum->arg_idx, 1324 .tmp_subindex = i, 1325 .arg_slot = cum->arg_slot + i, 1326 }; 1327 } 1328 cum->info_in_idx += n; 1329 cum->arg_slot += n; 1330 } 1331 1332 static void layout_arg_by_ref(TCGCumulativeArgs *cum, TCGHelperInfo *info) 1333 { 1334 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1335 int n = 128 / TCG_TARGET_REG_BITS; 1336 1337 /* The first subindex carries the pointer. */ 1338 layout_arg_1(cum, info, TCG_CALL_ARG_BY_REF); 1339 1340 /* 1341 * The callee is allowed to clobber memory associated with 1342 * structure pass by-reference. Therefore we must make copies. 1343 * Allocate space from "ref_slot", which will be adjusted to 1344 * follow the parameters on the stack. 1345 */ 1346 loc[0].ref_slot = cum->ref_slot; 1347 1348 /* 1349 * Subsequent words also go into the reference slot, but 1350 * do not accumulate into the regular arguments. 1351 */ 1352 for (int i = 1; i < n; ++i) { 1353 loc[i] = (TCGCallArgumentLoc){ 1354 .kind = TCG_CALL_ARG_BY_REF_N, 1355 .arg_idx = cum->arg_idx, 1356 .tmp_subindex = i, 1357 .ref_slot = cum->ref_slot + i, 1358 }; 1359 } 1360 cum->info_in_idx += n - 1; /* i=0 accounted for in layout_arg_1 */ 1361 cum->ref_slot += n; 1362 } 1363 1364 static void init_call_layout(TCGHelperInfo *info) 1365 { 1366 int max_reg_slots = ARRAY_SIZE(tcg_target_call_iarg_regs); 1367 int max_stk_slots = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1368 unsigned typemask = info->typemask; 1369 unsigned typecode; 1370 TCGCumulativeArgs cum = { }; 1371 1372 /* 1373 * Parse and place any function return value. 1374 */ 1375 typecode = typemask & 7; 1376 switch (typecode) { 1377 case dh_typecode_void: 1378 info->nr_out = 0; 1379 break; 1380 case dh_typecode_i32: 1381 case dh_typecode_s32: 1382 case dh_typecode_ptr: 1383 info->nr_out = 1; 1384 info->out_kind = TCG_CALL_RET_NORMAL; 1385 break; 1386 case dh_typecode_i64: 1387 case dh_typecode_s64: 1388 info->nr_out = 64 / TCG_TARGET_REG_BITS; 1389 info->out_kind = TCG_CALL_RET_NORMAL; 1390 /* Query the last register now to trigger any assert early. */ 1391 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1392 break; 1393 case dh_typecode_i128: 1394 info->nr_out = 128 / TCG_TARGET_REG_BITS; 1395 info->out_kind = TCG_TARGET_CALL_RET_I128; 1396 switch (TCG_TARGET_CALL_RET_I128) { 1397 case TCG_CALL_RET_NORMAL: 1398 /* Query the last register now to trigger any assert early. */ 1399 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1400 break; 1401 case TCG_CALL_RET_BY_VEC: 1402 /* Query the single register now to trigger any assert early. */ 1403 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0); 1404 break; 1405 case TCG_CALL_RET_BY_REF: 1406 /* 1407 * Allocate the first argument to the output. 1408 * We don't need to store this anywhere, just make it 1409 * unavailable for use in the input loop below. 1410 */ 1411 cum.arg_slot = 1; 1412 break; 1413 default: 1414 qemu_build_not_reached(); 1415 } 1416 break; 1417 default: 1418 g_assert_not_reached(); 1419 } 1420 1421 /* 1422 * Parse and place function arguments. 1423 */ 1424 for (typemask >>= 3; typemask; typemask >>= 3, cum.arg_idx++) { 1425 TCGCallArgumentKind kind; 1426 TCGType type; 1427 1428 typecode = typemask & 7; 1429 switch (typecode) { 1430 case dh_typecode_i32: 1431 case dh_typecode_s32: 1432 type = TCG_TYPE_I32; 1433 break; 1434 case dh_typecode_i64: 1435 case dh_typecode_s64: 1436 type = TCG_TYPE_I64; 1437 break; 1438 case dh_typecode_ptr: 1439 type = TCG_TYPE_PTR; 1440 break; 1441 case dh_typecode_i128: 1442 type = TCG_TYPE_I128; 1443 break; 1444 default: 1445 g_assert_not_reached(); 1446 } 1447 1448 switch (type) { 1449 case TCG_TYPE_I32: 1450 switch (TCG_TARGET_CALL_ARG_I32) { 1451 case TCG_CALL_ARG_EVEN: 1452 layout_arg_even(&cum); 1453 /* fall through */ 1454 case TCG_CALL_ARG_NORMAL: 1455 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1456 break; 1457 case TCG_CALL_ARG_EXTEND: 1458 kind = TCG_CALL_ARG_EXTEND_U + (typecode & 1); 1459 layout_arg_1(&cum, info, kind); 1460 break; 1461 default: 1462 qemu_build_not_reached(); 1463 } 1464 break; 1465 1466 case TCG_TYPE_I64: 1467 switch (TCG_TARGET_CALL_ARG_I64) { 1468 case TCG_CALL_ARG_EVEN: 1469 layout_arg_even(&cum); 1470 /* fall through */ 1471 case TCG_CALL_ARG_NORMAL: 1472 if (TCG_TARGET_REG_BITS == 32) { 1473 layout_arg_normal_n(&cum, info, 2); 1474 } else { 1475 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1476 } 1477 break; 1478 default: 1479 qemu_build_not_reached(); 1480 } 1481 break; 1482 1483 case TCG_TYPE_I128: 1484 switch (TCG_TARGET_CALL_ARG_I128) { 1485 case TCG_CALL_ARG_EVEN: 1486 layout_arg_even(&cum); 1487 /* fall through */ 1488 case TCG_CALL_ARG_NORMAL: 1489 layout_arg_normal_n(&cum, info, 128 / TCG_TARGET_REG_BITS); 1490 break; 1491 case TCG_CALL_ARG_BY_REF: 1492 layout_arg_by_ref(&cum, info); 1493 break; 1494 default: 1495 qemu_build_not_reached(); 1496 } 1497 break; 1498 1499 default: 1500 g_assert_not_reached(); 1501 } 1502 } 1503 info->nr_in = cum.info_in_idx; 1504 1505 /* Validate that we didn't overrun the input array. */ 1506 assert(cum.info_in_idx <= ARRAY_SIZE(info->in)); 1507 /* Validate the backend has enough argument space. */ 1508 assert(cum.arg_slot <= max_reg_slots + max_stk_slots); 1509 1510 /* 1511 * Relocate the "ref_slot" area to the end of the parameters. 1512 * Minimizing this stack offset helps code size for x86, 1513 * which has a signed 8-bit offset encoding. 1514 */ 1515 if (cum.ref_slot != 0) { 1516 int ref_base = 0; 1517 1518 if (cum.arg_slot > max_reg_slots) { 1519 int align = __alignof(Int128) / sizeof(tcg_target_long); 1520 1521 ref_base = cum.arg_slot - max_reg_slots; 1522 if (align > 1) { 1523 ref_base = ROUND_UP(ref_base, align); 1524 } 1525 } 1526 assert(ref_base + cum.ref_slot <= max_stk_slots); 1527 ref_base += max_reg_slots; 1528 1529 if (ref_base != 0) { 1530 for (int i = cum.info_in_idx - 1; i >= 0; --i) { 1531 TCGCallArgumentLoc *loc = &info->in[i]; 1532 switch (loc->kind) { 1533 case TCG_CALL_ARG_BY_REF: 1534 case TCG_CALL_ARG_BY_REF_N: 1535 loc->ref_slot += ref_base; 1536 break; 1537 default: 1538 break; 1539 } 1540 } 1541 } 1542 } 1543 } 1544 1545 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)]; 1546 static void process_constraint_sets(void); 1547 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1548 TCGReg reg, const char *name); 1549 1550 static void tcg_context_init(unsigned max_threads) 1551 { 1552 TCGContext *s = &tcg_init_ctx; 1553 int n, i; 1554 TCGTemp *ts; 1555 1556 memset(s, 0, sizeof(*s)); 1557 s->nb_globals = 0; 1558 1559 init_call_layout(&info_helper_ld32_mmu); 1560 init_call_layout(&info_helper_ld64_mmu); 1561 init_call_layout(&info_helper_ld128_mmu); 1562 init_call_layout(&info_helper_st32_mmu); 1563 init_call_layout(&info_helper_st64_mmu); 1564 init_call_layout(&info_helper_st128_mmu); 1565 1566 tcg_target_init(s); 1567 process_constraint_sets(); 1568 1569 /* Reverse the order of the saved registers, assuming they're all at 1570 the start of tcg_target_reg_alloc_order. */ 1571 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) { 1572 int r = tcg_target_reg_alloc_order[n]; 1573 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) { 1574 break; 1575 } 1576 } 1577 for (i = 0; i < n; ++i) { 1578 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i]; 1579 } 1580 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) { 1581 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i]; 1582 } 1583 1584 tcg_ctx = s; 1585 /* 1586 * In user-mode we simply share the init context among threads, since we 1587 * use a single region. See the documentation tcg_region_init() for the 1588 * reasoning behind this. 1589 * In system-mode we will have at most max_threads TCG threads. 1590 */ 1591 #ifdef CONFIG_USER_ONLY 1592 tcg_ctxs = &tcg_ctx; 1593 tcg_cur_ctxs = 1; 1594 tcg_max_ctxs = 1; 1595 #else 1596 tcg_max_ctxs = max_threads; 1597 tcg_ctxs = g_new0(TCGContext *, max_threads); 1598 #endif 1599 1600 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0)); 1601 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env"); 1602 tcg_env = temp_tcgv_ptr(ts); 1603 } 1604 1605 void tcg_init(size_t tb_size, int splitwx, unsigned max_threads) 1606 { 1607 tcg_context_init(max_threads); 1608 tcg_region_init(tb_size, splitwx, max_threads); 1609 } 1610 1611 /* 1612 * Allocate TBs right before their corresponding translated code, making 1613 * sure that TBs and code are on different cache lines. 1614 */ 1615 TranslationBlock *tcg_tb_alloc(TCGContext *s) 1616 { 1617 uintptr_t align = qemu_icache_linesize; 1618 TranslationBlock *tb; 1619 void *next; 1620 1621 retry: 1622 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align); 1623 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align); 1624 1625 if (unlikely(next > s->code_gen_highwater)) { 1626 if (tcg_region_alloc(s)) { 1627 return NULL; 1628 } 1629 goto retry; 1630 } 1631 qatomic_set(&s->code_gen_ptr, next); 1632 return tb; 1633 } 1634 1635 void tcg_prologue_init(void) 1636 { 1637 TCGContext *s = tcg_ctx; 1638 size_t prologue_size; 1639 1640 s->code_ptr = s->code_gen_ptr; 1641 s->code_buf = s->code_gen_ptr; 1642 s->data_gen_ptr = NULL; 1643 1644 #ifndef CONFIG_TCG_INTERPRETER 1645 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr); 1646 #endif 1647 1648 s->pool_labels = NULL; 1649 1650 qemu_thread_jit_write(); 1651 /* Generate the prologue. */ 1652 tcg_target_qemu_prologue(s); 1653 1654 /* Allow the prologue to put e.g. guest_base into a pool entry. */ 1655 { 1656 int result = tcg_out_pool_finalize(s); 1657 tcg_debug_assert(result == 0); 1658 } 1659 1660 prologue_size = tcg_current_code_size(s); 1661 perf_report_prologue(s->code_gen_ptr, prologue_size); 1662 1663 #ifndef CONFIG_TCG_INTERPRETER 1664 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 1665 (uintptr_t)s->code_buf, prologue_size); 1666 #endif 1667 1668 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { 1669 FILE *logfile = qemu_log_trylock(); 1670 if (logfile) { 1671 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size); 1672 if (s->data_gen_ptr) { 1673 size_t code_size = s->data_gen_ptr - s->code_gen_ptr; 1674 size_t data_size = prologue_size - code_size; 1675 size_t i; 1676 1677 disas(logfile, s->code_gen_ptr, code_size); 1678 1679 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) { 1680 if (sizeof(tcg_target_ulong) == 8) { 1681 fprintf(logfile, 1682 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n", 1683 (uintptr_t)s->data_gen_ptr + i, 1684 *(uint64_t *)(s->data_gen_ptr + i)); 1685 } else { 1686 fprintf(logfile, 1687 "0x%08" PRIxPTR ": .long 0x%08x\n", 1688 (uintptr_t)s->data_gen_ptr + i, 1689 *(uint32_t *)(s->data_gen_ptr + i)); 1690 } 1691 } 1692 } else { 1693 disas(logfile, s->code_gen_ptr, prologue_size); 1694 } 1695 fprintf(logfile, "\n"); 1696 qemu_log_unlock(logfile); 1697 } 1698 } 1699 1700 #ifndef CONFIG_TCG_INTERPRETER 1701 /* 1702 * Assert that goto_ptr is implemented completely, setting an epilogue. 1703 * For tci, we use NULL as the signal to return from the interpreter, 1704 * so skip this check. 1705 */ 1706 tcg_debug_assert(tcg_code_gen_epilogue != NULL); 1707 #endif 1708 1709 tcg_region_prologue_set(s); 1710 } 1711 1712 void tcg_func_start(TCGContext *s) 1713 { 1714 tcg_pool_reset(s); 1715 s->nb_temps = s->nb_globals; 1716 1717 /* No temps have been previously allocated for size or locality. */ 1718 tcg_temp_ebb_reset_freed(s); 1719 1720 /* No constant temps have been previously allocated. */ 1721 for (int i = 0; i < TCG_TYPE_COUNT; ++i) { 1722 if (s->const_table[i]) { 1723 g_hash_table_remove_all(s->const_table[i]); 1724 } 1725 } 1726 1727 s->nb_ops = 0; 1728 s->nb_labels = 0; 1729 s->current_frame_offset = s->frame_start; 1730 1731 #ifdef CONFIG_DEBUG_TCG 1732 s->goto_tb_issue_mask = 0; 1733 #endif 1734 1735 QTAILQ_INIT(&s->ops); 1736 QTAILQ_INIT(&s->free_ops); 1737 s->emit_before_op = NULL; 1738 QSIMPLEQ_INIT(&s->labels); 1739 1740 tcg_debug_assert(s->addr_type <= TCG_TYPE_REG); 1741 tcg_debug_assert(s->insn_start_words > 0); 1742 } 1743 1744 static TCGTemp *tcg_temp_alloc(TCGContext *s) 1745 { 1746 int n = s->nb_temps++; 1747 1748 if (n >= TCG_MAX_TEMPS) { 1749 tcg_raise_tb_overflow(s); 1750 } 1751 return memset(&s->temps[n], 0, sizeof(TCGTemp)); 1752 } 1753 1754 static TCGTemp *tcg_global_alloc(TCGContext *s) 1755 { 1756 TCGTemp *ts; 1757 1758 tcg_debug_assert(s->nb_globals == s->nb_temps); 1759 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS); 1760 s->nb_globals++; 1761 ts = tcg_temp_alloc(s); 1762 ts->kind = TEMP_GLOBAL; 1763 1764 return ts; 1765 } 1766 1767 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1768 TCGReg reg, const char *name) 1769 { 1770 TCGTemp *ts; 1771 1772 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32); 1773 1774 ts = tcg_global_alloc(s); 1775 ts->base_type = type; 1776 ts->type = type; 1777 ts->kind = TEMP_FIXED; 1778 ts->reg = reg; 1779 ts->name = name; 1780 tcg_regset_set_reg(s->reserved_regs, reg); 1781 1782 return ts; 1783 } 1784 1785 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size) 1786 { 1787 s->frame_start = start; 1788 s->frame_end = start + size; 1789 s->frame_temp 1790 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame"); 1791 } 1792 1793 static TCGTemp *tcg_global_mem_new_internal(TCGv_ptr base, intptr_t offset, 1794 const char *name, TCGType type) 1795 { 1796 TCGContext *s = tcg_ctx; 1797 TCGTemp *base_ts = tcgv_ptr_temp(base); 1798 TCGTemp *ts = tcg_global_alloc(s); 1799 int indirect_reg = 0; 1800 1801 switch (base_ts->kind) { 1802 case TEMP_FIXED: 1803 break; 1804 case TEMP_GLOBAL: 1805 /* We do not support double-indirect registers. */ 1806 tcg_debug_assert(!base_ts->indirect_reg); 1807 base_ts->indirect_base = 1; 1808 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64 1809 ? 2 : 1); 1810 indirect_reg = 1; 1811 break; 1812 default: 1813 g_assert_not_reached(); 1814 } 1815 1816 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1817 TCGTemp *ts2 = tcg_global_alloc(s); 1818 char buf[64]; 1819 1820 ts->base_type = TCG_TYPE_I64; 1821 ts->type = TCG_TYPE_I32; 1822 ts->indirect_reg = indirect_reg; 1823 ts->mem_allocated = 1; 1824 ts->mem_base = base_ts; 1825 ts->mem_offset = offset; 1826 pstrcpy(buf, sizeof(buf), name); 1827 pstrcat(buf, sizeof(buf), "_0"); 1828 ts->name = strdup(buf); 1829 1830 tcg_debug_assert(ts2 == ts + 1); 1831 ts2->base_type = TCG_TYPE_I64; 1832 ts2->type = TCG_TYPE_I32; 1833 ts2->indirect_reg = indirect_reg; 1834 ts2->mem_allocated = 1; 1835 ts2->mem_base = base_ts; 1836 ts2->mem_offset = offset + 4; 1837 ts2->temp_subindex = 1; 1838 pstrcpy(buf, sizeof(buf), name); 1839 pstrcat(buf, sizeof(buf), "_1"); 1840 ts2->name = strdup(buf); 1841 } else { 1842 ts->base_type = type; 1843 ts->type = type; 1844 ts->indirect_reg = indirect_reg; 1845 ts->mem_allocated = 1; 1846 ts->mem_base = base_ts; 1847 ts->mem_offset = offset; 1848 ts->name = name; 1849 } 1850 return ts; 1851 } 1852 1853 TCGv_i32 tcg_global_mem_new_i32(TCGv_ptr reg, intptr_t off, const char *name) 1854 { 1855 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I32); 1856 return temp_tcgv_i32(ts); 1857 } 1858 1859 TCGv_i64 tcg_global_mem_new_i64(TCGv_ptr reg, intptr_t off, const char *name) 1860 { 1861 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I64); 1862 return temp_tcgv_i64(ts); 1863 } 1864 1865 TCGv_ptr tcg_global_mem_new_ptr(TCGv_ptr reg, intptr_t off, const char *name) 1866 { 1867 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_PTR); 1868 return temp_tcgv_ptr(ts); 1869 } 1870 1871 TCGTemp *tcg_temp_new_internal(TCGType type, TCGTempKind kind) 1872 { 1873 TCGContext *s = tcg_ctx; 1874 TCGTemp *ts; 1875 int n; 1876 1877 if (kind == TEMP_EBB) { 1878 int idx = find_first_bit(s->free_temps[type].l, TCG_MAX_TEMPS); 1879 1880 if (idx < TCG_MAX_TEMPS) { 1881 /* There is already an available temp with the right type. */ 1882 clear_bit(idx, s->free_temps[type].l); 1883 1884 ts = &s->temps[idx]; 1885 ts->temp_allocated = 1; 1886 tcg_debug_assert(ts->base_type == type); 1887 tcg_debug_assert(ts->kind == kind); 1888 return ts; 1889 } 1890 } else { 1891 tcg_debug_assert(kind == TEMP_TB); 1892 } 1893 1894 switch (type) { 1895 case TCG_TYPE_I32: 1896 case TCG_TYPE_V64: 1897 case TCG_TYPE_V128: 1898 case TCG_TYPE_V256: 1899 n = 1; 1900 break; 1901 case TCG_TYPE_I64: 1902 n = 64 / TCG_TARGET_REG_BITS; 1903 break; 1904 case TCG_TYPE_I128: 1905 n = 128 / TCG_TARGET_REG_BITS; 1906 break; 1907 default: 1908 g_assert_not_reached(); 1909 } 1910 1911 ts = tcg_temp_alloc(s); 1912 ts->base_type = type; 1913 ts->temp_allocated = 1; 1914 ts->kind = kind; 1915 1916 if (n == 1) { 1917 ts->type = type; 1918 } else { 1919 ts->type = TCG_TYPE_REG; 1920 1921 for (int i = 1; i < n; ++i) { 1922 TCGTemp *ts2 = tcg_temp_alloc(s); 1923 1924 tcg_debug_assert(ts2 == ts + i); 1925 ts2->base_type = type; 1926 ts2->type = TCG_TYPE_REG; 1927 ts2->temp_allocated = 1; 1928 ts2->temp_subindex = i; 1929 ts2->kind = kind; 1930 } 1931 } 1932 return ts; 1933 } 1934 1935 TCGv_i32 tcg_temp_new_i32(void) 1936 { 1937 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_TB)); 1938 } 1939 1940 TCGv_i32 tcg_temp_ebb_new_i32(void) 1941 { 1942 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_EBB)); 1943 } 1944 1945 TCGv_i64 tcg_temp_new_i64(void) 1946 { 1947 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_TB)); 1948 } 1949 1950 TCGv_i64 tcg_temp_ebb_new_i64(void) 1951 { 1952 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_EBB)); 1953 } 1954 1955 TCGv_ptr tcg_temp_new_ptr(void) 1956 { 1957 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_TB)); 1958 } 1959 1960 TCGv_ptr tcg_temp_ebb_new_ptr(void) 1961 { 1962 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_EBB)); 1963 } 1964 1965 TCGv_i128 tcg_temp_new_i128(void) 1966 { 1967 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_TB)); 1968 } 1969 1970 TCGv_i128 tcg_temp_ebb_new_i128(void) 1971 { 1972 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_EBB)); 1973 } 1974 1975 TCGv_vec tcg_temp_new_vec(TCGType type) 1976 { 1977 TCGTemp *t; 1978 1979 #ifdef CONFIG_DEBUG_TCG 1980 switch (type) { 1981 case TCG_TYPE_V64: 1982 assert(TCG_TARGET_HAS_v64); 1983 break; 1984 case TCG_TYPE_V128: 1985 assert(TCG_TARGET_HAS_v128); 1986 break; 1987 case TCG_TYPE_V256: 1988 assert(TCG_TARGET_HAS_v256); 1989 break; 1990 default: 1991 g_assert_not_reached(); 1992 } 1993 #endif 1994 1995 t = tcg_temp_new_internal(type, TEMP_EBB); 1996 return temp_tcgv_vec(t); 1997 } 1998 1999 /* Create a new temp of the same type as an existing temp. */ 2000 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match) 2001 { 2002 TCGTemp *t = tcgv_vec_temp(match); 2003 2004 tcg_debug_assert(t->temp_allocated != 0); 2005 2006 t = tcg_temp_new_internal(t->base_type, TEMP_EBB); 2007 return temp_tcgv_vec(t); 2008 } 2009 2010 void tcg_temp_free_internal(TCGTemp *ts) 2011 { 2012 TCGContext *s = tcg_ctx; 2013 2014 switch (ts->kind) { 2015 case TEMP_CONST: 2016 case TEMP_TB: 2017 /* Silently ignore free. */ 2018 break; 2019 case TEMP_EBB: 2020 tcg_debug_assert(ts->temp_allocated != 0); 2021 ts->temp_allocated = 0; 2022 set_bit(temp_idx(ts), s->free_temps[ts->base_type].l); 2023 break; 2024 default: 2025 /* It never made sense to free TEMP_FIXED or TEMP_GLOBAL. */ 2026 g_assert_not_reached(); 2027 } 2028 } 2029 2030 void tcg_temp_free_i32(TCGv_i32 arg) 2031 { 2032 tcg_temp_free_internal(tcgv_i32_temp(arg)); 2033 } 2034 2035 void tcg_temp_free_i64(TCGv_i64 arg) 2036 { 2037 tcg_temp_free_internal(tcgv_i64_temp(arg)); 2038 } 2039 2040 void tcg_temp_free_i128(TCGv_i128 arg) 2041 { 2042 tcg_temp_free_internal(tcgv_i128_temp(arg)); 2043 } 2044 2045 void tcg_temp_free_ptr(TCGv_ptr arg) 2046 { 2047 tcg_temp_free_internal(tcgv_ptr_temp(arg)); 2048 } 2049 2050 void tcg_temp_free_vec(TCGv_vec arg) 2051 { 2052 tcg_temp_free_internal(tcgv_vec_temp(arg)); 2053 } 2054 2055 TCGTemp *tcg_constant_internal(TCGType type, int64_t val) 2056 { 2057 TCGContext *s = tcg_ctx; 2058 GHashTable *h = s->const_table[type]; 2059 TCGTemp *ts; 2060 2061 if (h == NULL) { 2062 h = g_hash_table_new(g_int64_hash, g_int64_equal); 2063 s->const_table[type] = h; 2064 } 2065 2066 ts = g_hash_table_lookup(h, &val); 2067 if (ts == NULL) { 2068 int64_t *val_ptr; 2069 2070 ts = tcg_temp_alloc(s); 2071 2072 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 2073 TCGTemp *ts2 = tcg_temp_alloc(s); 2074 2075 tcg_debug_assert(ts2 == ts + 1); 2076 2077 ts->base_type = TCG_TYPE_I64; 2078 ts->type = TCG_TYPE_I32; 2079 ts->kind = TEMP_CONST; 2080 ts->temp_allocated = 1; 2081 2082 ts2->base_type = TCG_TYPE_I64; 2083 ts2->type = TCG_TYPE_I32; 2084 ts2->kind = TEMP_CONST; 2085 ts2->temp_allocated = 1; 2086 ts2->temp_subindex = 1; 2087 2088 /* 2089 * Retain the full value of the 64-bit constant in the low 2090 * part, so that the hash table works. Actual uses will 2091 * truncate the value to the low part. 2092 */ 2093 ts[HOST_BIG_ENDIAN].val = val; 2094 ts[!HOST_BIG_ENDIAN].val = val >> 32; 2095 val_ptr = &ts[HOST_BIG_ENDIAN].val; 2096 } else { 2097 ts->base_type = type; 2098 ts->type = type; 2099 ts->kind = TEMP_CONST; 2100 ts->temp_allocated = 1; 2101 ts->val = val; 2102 val_ptr = &ts->val; 2103 } 2104 g_hash_table_insert(h, val_ptr, ts); 2105 } 2106 2107 return ts; 2108 } 2109 2110 TCGv_i32 tcg_constant_i32(int32_t val) 2111 { 2112 return temp_tcgv_i32(tcg_constant_internal(TCG_TYPE_I32, val)); 2113 } 2114 2115 TCGv_i64 tcg_constant_i64(int64_t val) 2116 { 2117 return temp_tcgv_i64(tcg_constant_internal(TCG_TYPE_I64, val)); 2118 } 2119 2120 TCGv_ptr tcg_constant_ptr_int(intptr_t val) 2121 { 2122 return temp_tcgv_ptr(tcg_constant_internal(TCG_TYPE_PTR, val)); 2123 } 2124 2125 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val) 2126 { 2127 val = dup_const(vece, val); 2128 return temp_tcgv_vec(tcg_constant_internal(type, val)); 2129 } 2130 2131 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val) 2132 { 2133 TCGTemp *t = tcgv_vec_temp(match); 2134 2135 tcg_debug_assert(t->temp_allocated != 0); 2136 return tcg_constant_vec(t->base_type, vece, val); 2137 } 2138 2139 #ifdef CONFIG_DEBUG_TCG 2140 size_t temp_idx(TCGTemp *ts) 2141 { 2142 ptrdiff_t n = ts - tcg_ctx->temps; 2143 assert(n >= 0 && n < tcg_ctx->nb_temps); 2144 return n; 2145 } 2146 2147 TCGTemp *tcgv_i32_temp(TCGv_i32 v) 2148 { 2149 uintptr_t o = (uintptr_t)v - offsetof(TCGContext, temps); 2150 2151 assert(o < sizeof(TCGTemp) * tcg_ctx->nb_temps); 2152 assert(o % sizeof(TCGTemp) == 0); 2153 2154 return (void *)tcg_ctx + (uintptr_t)v; 2155 } 2156 #endif /* CONFIG_DEBUG_TCG */ 2157 2158 /* 2159 * Return true if OP may appear in the opcode stream with TYPE. 2160 * Test the runtime variable that controls each opcode. 2161 */ 2162 bool tcg_op_supported(TCGOpcode op, TCGType type, unsigned flags) 2163 { 2164 bool has_type; 2165 2166 switch (type) { 2167 case TCG_TYPE_I32: 2168 has_type = true; 2169 break; 2170 case TCG_TYPE_I64: 2171 has_type = TCG_TARGET_REG_BITS == 64; 2172 break; 2173 case TCG_TYPE_V64: 2174 has_type = TCG_TARGET_HAS_v64; 2175 break; 2176 case TCG_TYPE_V128: 2177 has_type = TCG_TARGET_HAS_v128; 2178 break; 2179 case TCG_TYPE_V256: 2180 has_type = TCG_TARGET_HAS_v256; 2181 break; 2182 default: 2183 has_type = false; 2184 break; 2185 } 2186 2187 switch (op) { 2188 case INDEX_op_discard: 2189 case INDEX_op_set_label: 2190 case INDEX_op_call: 2191 case INDEX_op_br: 2192 case INDEX_op_mb: 2193 case INDEX_op_insn_start: 2194 case INDEX_op_exit_tb: 2195 case INDEX_op_goto_tb: 2196 case INDEX_op_goto_ptr: 2197 case INDEX_op_qemu_ld_i32: 2198 case INDEX_op_qemu_st_i32: 2199 case INDEX_op_qemu_ld_i64: 2200 case INDEX_op_qemu_st_i64: 2201 return true; 2202 2203 case INDEX_op_qemu_st8_i32: 2204 return TCG_TARGET_HAS_qemu_st8_i32; 2205 2206 case INDEX_op_qemu_ld_i128: 2207 case INDEX_op_qemu_st_i128: 2208 return TCG_TARGET_HAS_qemu_ldst_i128; 2209 2210 case INDEX_op_add: 2211 case INDEX_op_and: 2212 case INDEX_op_mov: 2213 return has_type; 2214 2215 case INDEX_op_setcond_i32: 2216 case INDEX_op_brcond_i32: 2217 case INDEX_op_movcond_i32: 2218 case INDEX_op_ld8u_i32: 2219 case INDEX_op_ld8s_i32: 2220 case INDEX_op_ld16u_i32: 2221 case INDEX_op_ld16s_i32: 2222 case INDEX_op_ld_i32: 2223 case INDEX_op_st8_i32: 2224 case INDEX_op_st16_i32: 2225 case INDEX_op_st_i32: 2226 case INDEX_op_sub_i32: 2227 case INDEX_op_neg_i32: 2228 case INDEX_op_mul_i32: 2229 case INDEX_op_or_i32: 2230 case INDEX_op_xor_i32: 2231 case INDEX_op_shl_i32: 2232 case INDEX_op_shr_i32: 2233 case INDEX_op_sar_i32: 2234 case INDEX_op_extract_i32: 2235 case INDEX_op_sextract_i32: 2236 case INDEX_op_deposit_i32: 2237 return true; 2238 2239 case INDEX_op_negsetcond_i32: 2240 return TCG_TARGET_HAS_negsetcond_i32; 2241 case INDEX_op_div_i32: 2242 case INDEX_op_divu_i32: 2243 return TCG_TARGET_HAS_div_i32; 2244 case INDEX_op_rem_i32: 2245 case INDEX_op_remu_i32: 2246 return TCG_TARGET_HAS_rem_i32; 2247 case INDEX_op_div2_i32: 2248 case INDEX_op_divu2_i32: 2249 return TCG_TARGET_HAS_div2_i32; 2250 case INDEX_op_rotl_i32: 2251 case INDEX_op_rotr_i32: 2252 return TCG_TARGET_HAS_rot_i32; 2253 case INDEX_op_extract2_i32: 2254 return TCG_TARGET_HAS_extract2_i32; 2255 case INDEX_op_add2_i32: 2256 return TCG_TARGET_HAS_add2_i32; 2257 case INDEX_op_sub2_i32: 2258 return TCG_TARGET_HAS_sub2_i32; 2259 case INDEX_op_mulu2_i32: 2260 return TCG_TARGET_HAS_mulu2_i32; 2261 case INDEX_op_muls2_i32: 2262 return TCG_TARGET_HAS_muls2_i32; 2263 case INDEX_op_muluh_i32: 2264 return TCG_TARGET_HAS_muluh_i32; 2265 case INDEX_op_mulsh_i32: 2266 return TCG_TARGET_HAS_mulsh_i32; 2267 case INDEX_op_bswap16_i32: 2268 return TCG_TARGET_HAS_bswap16_i32; 2269 case INDEX_op_bswap32_i32: 2270 return TCG_TARGET_HAS_bswap32_i32; 2271 case INDEX_op_not_i32: 2272 return TCG_TARGET_HAS_not_i32; 2273 case INDEX_op_orc_i32: 2274 return TCG_TARGET_HAS_orc_i32; 2275 case INDEX_op_eqv_i32: 2276 return TCG_TARGET_HAS_eqv_i32; 2277 case INDEX_op_nand_i32: 2278 return TCG_TARGET_HAS_nand_i32; 2279 case INDEX_op_nor_i32: 2280 return TCG_TARGET_HAS_nor_i32; 2281 case INDEX_op_clz_i32: 2282 return TCG_TARGET_HAS_clz_i32; 2283 case INDEX_op_ctz_i32: 2284 return TCG_TARGET_HAS_ctz_i32; 2285 case INDEX_op_ctpop_i32: 2286 return TCG_TARGET_HAS_ctpop_i32; 2287 2288 case INDEX_op_brcond2_i32: 2289 case INDEX_op_setcond2_i32: 2290 return TCG_TARGET_REG_BITS == 32; 2291 2292 case INDEX_op_setcond_i64: 2293 case INDEX_op_brcond_i64: 2294 case INDEX_op_movcond_i64: 2295 case INDEX_op_ld8u_i64: 2296 case INDEX_op_ld8s_i64: 2297 case INDEX_op_ld16u_i64: 2298 case INDEX_op_ld16s_i64: 2299 case INDEX_op_ld32u_i64: 2300 case INDEX_op_ld32s_i64: 2301 case INDEX_op_ld_i64: 2302 case INDEX_op_st8_i64: 2303 case INDEX_op_st16_i64: 2304 case INDEX_op_st32_i64: 2305 case INDEX_op_st_i64: 2306 case INDEX_op_sub_i64: 2307 case INDEX_op_neg_i64: 2308 case INDEX_op_mul_i64: 2309 case INDEX_op_or_i64: 2310 case INDEX_op_xor_i64: 2311 case INDEX_op_shl_i64: 2312 case INDEX_op_shr_i64: 2313 case INDEX_op_sar_i64: 2314 case INDEX_op_ext_i32_i64: 2315 case INDEX_op_extu_i32_i64: 2316 case INDEX_op_extract_i64: 2317 case INDEX_op_sextract_i64: 2318 case INDEX_op_deposit_i64: 2319 return TCG_TARGET_REG_BITS == 64; 2320 2321 case INDEX_op_negsetcond_i64: 2322 return TCG_TARGET_HAS_negsetcond_i64; 2323 case INDEX_op_div_i64: 2324 case INDEX_op_divu_i64: 2325 return TCG_TARGET_HAS_div_i64; 2326 case INDEX_op_rem_i64: 2327 case INDEX_op_remu_i64: 2328 return TCG_TARGET_HAS_rem_i64; 2329 case INDEX_op_div2_i64: 2330 case INDEX_op_divu2_i64: 2331 return TCG_TARGET_HAS_div2_i64; 2332 case INDEX_op_rotl_i64: 2333 case INDEX_op_rotr_i64: 2334 return TCG_TARGET_HAS_rot_i64; 2335 case INDEX_op_extract2_i64: 2336 return TCG_TARGET_HAS_extract2_i64; 2337 case INDEX_op_extrl_i64_i32: 2338 case INDEX_op_extrh_i64_i32: 2339 return TCG_TARGET_HAS_extr_i64_i32; 2340 case INDEX_op_bswap16_i64: 2341 return TCG_TARGET_HAS_bswap16_i64; 2342 case INDEX_op_bswap32_i64: 2343 return TCG_TARGET_HAS_bswap32_i64; 2344 case INDEX_op_bswap64_i64: 2345 return TCG_TARGET_HAS_bswap64_i64; 2346 case INDEX_op_not_i64: 2347 return TCG_TARGET_HAS_not_i64; 2348 case INDEX_op_orc_i64: 2349 return TCG_TARGET_HAS_orc_i64; 2350 case INDEX_op_eqv_i64: 2351 return TCG_TARGET_HAS_eqv_i64; 2352 case INDEX_op_nand_i64: 2353 return TCG_TARGET_HAS_nand_i64; 2354 case INDEX_op_nor_i64: 2355 return TCG_TARGET_HAS_nor_i64; 2356 case INDEX_op_clz_i64: 2357 return TCG_TARGET_HAS_clz_i64; 2358 case INDEX_op_ctz_i64: 2359 return TCG_TARGET_HAS_ctz_i64; 2360 case INDEX_op_ctpop_i64: 2361 return TCG_TARGET_HAS_ctpop_i64; 2362 case INDEX_op_add2_i64: 2363 return TCG_TARGET_HAS_add2_i64; 2364 case INDEX_op_sub2_i64: 2365 return TCG_TARGET_HAS_sub2_i64; 2366 case INDEX_op_mulu2_i64: 2367 return TCG_TARGET_HAS_mulu2_i64; 2368 case INDEX_op_muls2_i64: 2369 return TCG_TARGET_HAS_muls2_i64; 2370 case INDEX_op_muluh_i64: 2371 return TCG_TARGET_HAS_muluh_i64; 2372 case INDEX_op_mulsh_i64: 2373 return TCG_TARGET_HAS_mulsh_i64; 2374 2375 case INDEX_op_mov_vec: 2376 case INDEX_op_dup_vec: 2377 case INDEX_op_dupm_vec: 2378 case INDEX_op_ld_vec: 2379 case INDEX_op_st_vec: 2380 case INDEX_op_add_vec: 2381 case INDEX_op_sub_vec: 2382 case INDEX_op_and_vec: 2383 case INDEX_op_or_vec: 2384 case INDEX_op_xor_vec: 2385 case INDEX_op_cmp_vec: 2386 return has_type; 2387 case INDEX_op_dup2_vec: 2388 return has_type && TCG_TARGET_REG_BITS == 32; 2389 case INDEX_op_not_vec: 2390 return has_type && TCG_TARGET_HAS_not_vec; 2391 case INDEX_op_neg_vec: 2392 return has_type && TCG_TARGET_HAS_neg_vec; 2393 case INDEX_op_abs_vec: 2394 return has_type && TCG_TARGET_HAS_abs_vec; 2395 case INDEX_op_andc_vec: 2396 return has_type && TCG_TARGET_HAS_andc_vec; 2397 case INDEX_op_orc_vec: 2398 return has_type && TCG_TARGET_HAS_orc_vec; 2399 case INDEX_op_nand_vec: 2400 return has_type && TCG_TARGET_HAS_nand_vec; 2401 case INDEX_op_nor_vec: 2402 return has_type && TCG_TARGET_HAS_nor_vec; 2403 case INDEX_op_eqv_vec: 2404 return has_type && TCG_TARGET_HAS_eqv_vec; 2405 case INDEX_op_mul_vec: 2406 return has_type && TCG_TARGET_HAS_mul_vec; 2407 case INDEX_op_shli_vec: 2408 case INDEX_op_shri_vec: 2409 case INDEX_op_sari_vec: 2410 return has_type && TCG_TARGET_HAS_shi_vec; 2411 case INDEX_op_shls_vec: 2412 case INDEX_op_shrs_vec: 2413 case INDEX_op_sars_vec: 2414 return has_type && TCG_TARGET_HAS_shs_vec; 2415 case INDEX_op_shlv_vec: 2416 case INDEX_op_shrv_vec: 2417 case INDEX_op_sarv_vec: 2418 return has_type && TCG_TARGET_HAS_shv_vec; 2419 case INDEX_op_rotli_vec: 2420 return has_type && TCG_TARGET_HAS_roti_vec; 2421 case INDEX_op_rotls_vec: 2422 return has_type && TCG_TARGET_HAS_rots_vec; 2423 case INDEX_op_rotlv_vec: 2424 case INDEX_op_rotrv_vec: 2425 return has_type && TCG_TARGET_HAS_rotv_vec; 2426 case INDEX_op_ssadd_vec: 2427 case INDEX_op_usadd_vec: 2428 case INDEX_op_sssub_vec: 2429 case INDEX_op_ussub_vec: 2430 return has_type && TCG_TARGET_HAS_sat_vec; 2431 case INDEX_op_smin_vec: 2432 case INDEX_op_umin_vec: 2433 case INDEX_op_smax_vec: 2434 case INDEX_op_umax_vec: 2435 return has_type && TCG_TARGET_HAS_minmax_vec; 2436 case INDEX_op_bitsel_vec: 2437 return has_type && TCG_TARGET_HAS_bitsel_vec; 2438 case INDEX_op_cmpsel_vec: 2439 return has_type && TCG_TARGET_HAS_cmpsel_vec; 2440 2441 default: 2442 if (op < INDEX_op_last_generic) { 2443 const TCGOutOp *outop; 2444 TCGConstraintSetIndex con_set; 2445 2446 if (!has_type) { 2447 return false; 2448 } 2449 2450 outop = all_outop[op]; 2451 tcg_debug_assert(outop != NULL); 2452 2453 con_set = outop->static_constraint; 2454 if (con_set == C_Dynamic) { 2455 con_set = outop->dynamic_constraint(type, flags); 2456 } 2457 if (con_set >= 0) { 2458 return true; 2459 } 2460 tcg_debug_assert(con_set == C_NotImplemented); 2461 return false; 2462 } 2463 tcg_debug_assert(op < NB_OPS); 2464 return true; 2465 2466 case INDEX_op_last_generic: 2467 g_assert_not_reached(); 2468 } 2469 } 2470 2471 bool tcg_op_deposit_valid(TCGType type, unsigned ofs, unsigned len) 2472 { 2473 unsigned width; 2474 2475 tcg_debug_assert(type == TCG_TYPE_I32 || type == TCG_TYPE_I64); 2476 width = (type == TCG_TYPE_I32 ? 32 : 64); 2477 2478 tcg_debug_assert(ofs < width); 2479 tcg_debug_assert(len > 0); 2480 tcg_debug_assert(len <= width - ofs); 2481 2482 return TCG_TARGET_deposit_valid(type, ofs, len); 2483 } 2484 2485 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs); 2486 2487 static void tcg_gen_callN(void *func, TCGHelperInfo *info, 2488 TCGTemp *ret, TCGTemp **args) 2489 { 2490 TCGv_i64 extend_free[MAX_CALL_IARGS]; 2491 int n_extend = 0; 2492 TCGOp *op; 2493 int i, n, pi = 0, total_args; 2494 2495 if (unlikely(g_once_init_enter(HELPER_INFO_INIT(info)))) { 2496 init_call_layout(info); 2497 g_once_init_leave(HELPER_INFO_INIT(info), HELPER_INFO_INIT_VAL(info)); 2498 } 2499 2500 total_args = info->nr_out + info->nr_in + 2; 2501 op = tcg_op_alloc(INDEX_op_call, total_args); 2502 2503 #ifdef CONFIG_PLUGIN 2504 /* Flag helpers that may affect guest state */ 2505 if (tcg_ctx->plugin_insn && !(info->flags & TCG_CALL_NO_SIDE_EFFECTS)) { 2506 tcg_ctx->plugin_insn->calls_helpers = true; 2507 } 2508 #endif 2509 2510 TCGOP_CALLO(op) = n = info->nr_out; 2511 switch (n) { 2512 case 0: 2513 tcg_debug_assert(ret == NULL); 2514 break; 2515 case 1: 2516 tcg_debug_assert(ret != NULL); 2517 op->args[pi++] = temp_arg(ret); 2518 break; 2519 case 2: 2520 case 4: 2521 tcg_debug_assert(ret != NULL); 2522 tcg_debug_assert(ret->base_type == ret->type + ctz32(n)); 2523 tcg_debug_assert(ret->temp_subindex == 0); 2524 for (i = 0; i < n; ++i) { 2525 op->args[pi++] = temp_arg(ret + i); 2526 } 2527 break; 2528 default: 2529 g_assert_not_reached(); 2530 } 2531 2532 TCGOP_CALLI(op) = n = info->nr_in; 2533 for (i = 0; i < n; i++) { 2534 const TCGCallArgumentLoc *loc = &info->in[i]; 2535 TCGTemp *ts = args[loc->arg_idx] + loc->tmp_subindex; 2536 2537 switch (loc->kind) { 2538 case TCG_CALL_ARG_NORMAL: 2539 case TCG_CALL_ARG_BY_REF: 2540 case TCG_CALL_ARG_BY_REF_N: 2541 op->args[pi++] = temp_arg(ts); 2542 break; 2543 2544 case TCG_CALL_ARG_EXTEND_U: 2545 case TCG_CALL_ARG_EXTEND_S: 2546 { 2547 TCGv_i64 temp = tcg_temp_ebb_new_i64(); 2548 TCGv_i32 orig = temp_tcgv_i32(ts); 2549 2550 if (loc->kind == TCG_CALL_ARG_EXTEND_S) { 2551 tcg_gen_ext_i32_i64(temp, orig); 2552 } else { 2553 tcg_gen_extu_i32_i64(temp, orig); 2554 } 2555 op->args[pi++] = tcgv_i64_arg(temp); 2556 extend_free[n_extend++] = temp; 2557 } 2558 break; 2559 2560 default: 2561 g_assert_not_reached(); 2562 } 2563 } 2564 op->args[pi++] = (uintptr_t)func; 2565 op->args[pi++] = (uintptr_t)info; 2566 tcg_debug_assert(pi == total_args); 2567 2568 if (tcg_ctx->emit_before_op) { 2569 QTAILQ_INSERT_BEFORE(tcg_ctx->emit_before_op, op, link); 2570 } else { 2571 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 2572 } 2573 2574 tcg_debug_assert(n_extend < ARRAY_SIZE(extend_free)); 2575 for (i = 0; i < n_extend; ++i) { 2576 tcg_temp_free_i64(extend_free[i]); 2577 } 2578 } 2579 2580 void tcg_gen_call0(void *func, TCGHelperInfo *info, TCGTemp *ret) 2581 { 2582 tcg_gen_callN(func, info, ret, NULL); 2583 } 2584 2585 void tcg_gen_call1(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1) 2586 { 2587 tcg_gen_callN(func, info, ret, &t1); 2588 } 2589 2590 void tcg_gen_call2(void *func, TCGHelperInfo *info, TCGTemp *ret, 2591 TCGTemp *t1, TCGTemp *t2) 2592 { 2593 TCGTemp *args[2] = { t1, t2 }; 2594 tcg_gen_callN(func, info, ret, args); 2595 } 2596 2597 void tcg_gen_call3(void *func, TCGHelperInfo *info, TCGTemp *ret, 2598 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3) 2599 { 2600 TCGTemp *args[3] = { t1, t2, t3 }; 2601 tcg_gen_callN(func, info, ret, args); 2602 } 2603 2604 void tcg_gen_call4(void *func, TCGHelperInfo *info, TCGTemp *ret, 2605 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3, TCGTemp *t4) 2606 { 2607 TCGTemp *args[4] = { t1, t2, t3, t4 }; 2608 tcg_gen_callN(func, info, ret, args); 2609 } 2610 2611 void tcg_gen_call5(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2612 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, TCGTemp *t5) 2613 { 2614 TCGTemp *args[5] = { t1, t2, t3, t4, t5 }; 2615 tcg_gen_callN(func, info, ret, args); 2616 } 2617 2618 void tcg_gen_call6(void *func, TCGHelperInfo *info, TCGTemp *ret, 2619 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3, 2620 TCGTemp *t4, TCGTemp *t5, TCGTemp *t6) 2621 { 2622 TCGTemp *args[6] = { t1, t2, t3, t4, t5, t6 }; 2623 tcg_gen_callN(func, info, ret, args); 2624 } 2625 2626 void tcg_gen_call7(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2627 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, 2628 TCGTemp *t5, TCGTemp *t6, TCGTemp *t7) 2629 { 2630 TCGTemp *args[7] = { t1, t2, t3, t4, t5, t6, t7 }; 2631 tcg_gen_callN(func, info, ret, args); 2632 } 2633 2634 static void tcg_reg_alloc_start(TCGContext *s) 2635 { 2636 int i, n; 2637 2638 for (i = 0, n = s->nb_temps; i < n; i++) { 2639 TCGTemp *ts = &s->temps[i]; 2640 TCGTempVal val = TEMP_VAL_MEM; 2641 2642 switch (ts->kind) { 2643 case TEMP_CONST: 2644 val = TEMP_VAL_CONST; 2645 break; 2646 case TEMP_FIXED: 2647 val = TEMP_VAL_REG; 2648 break; 2649 case TEMP_GLOBAL: 2650 break; 2651 case TEMP_EBB: 2652 val = TEMP_VAL_DEAD; 2653 /* fall through */ 2654 case TEMP_TB: 2655 ts->mem_allocated = 0; 2656 break; 2657 default: 2658 g_assert_not_reached(); 2659 } 2660 ts->val_type = val; 2661 } 2662 2663 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp)); 2664 } 2665 2666 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size, 2667 TCGTemp *ts) 2668 { 2669 int idx = temp_idx(ts); 2670 2671 switch (ts->kind) { 2672 case TEMP_FIXED: 2673 case TEMP_GLOBAL: 2674 pstrcpy(buf, buf_size, ts->name); 2675 break; 2676 case TEMP_TB: 2677 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals); 2678 break; 2679 case TEMP_EBB: 2680 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals); 2681 break; 2682 case TEMP_CONST: 2683 switch (ts->type) { 2684 case TCG_TYPE_I32: 2685 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val); 2686 break; 2687 #if TCG_TARGET_REG_BITS > 32 2688 case TCG_TYPE_I64: 2689 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val); 2690 break; 2691 #endif 2692 case TCG_TYPE_V64: 2693 case TCG_TYPE_V128: 2694 case TCG_TYPE_V256: 2695 snprintf(buf, buf_size, "v%d$0x%" PRIx64, 2696 64 << (ts->type - TCG_TYPE_V64), ts->val); 2697 break; 2698 default: 2699 g_assert_not_reached(); 2700 } 2701 break; 2702 } 2703 return buf; 2704 } 2705 2706 static char *tcg_get_arg_str(TCGContext *s, char *buf, 2707 int buf_size, TCGArg arg) 2708 { 2709 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg)); 2710 } 2711 2712 static const char * const cond_name[] = 2713 { 2714 [TCG_COND_NEVER] = "never", 2715 [TCG_COND_ALWAYS] = "always", 2716 [TCG_COND_EQ] = "eq", 2717 [TCG_COND_NE] = "ne", 2718 [TCG_COND_LT] = "lt", 2719 [TCG_COND_GE] = "ge", 2720 [TCG_COND_LE] = "le", 2721 [TCG_COND_GT] = "gt", 2722 [TCG_COND_LTU] = "ltu", 2723 [TCG_COND_GEU] = "geu", 2724 [TCG_COND_LEU] = "leu", 2725 [TCG_COND_GTU] = "gtu", 2726 [TCG_COND_TSTEQ] = "tsteq", 2727 [TCG_COND_TSTNE] = "tstne", 2728 }; 2729 2730 static const char * const ldst_name[(MO_BSWAP | MO_SSIZE) + 1] = 2731 { 2732 [MO_UB] = "ub", 2733 [MO_SB] = "sb", 2734 [MO_LEUW] = "leuw", 2735 [MO_LESW] = "lesw", 2736 [MO_LEUL] = "leul", 2737 [MO_LESL] = "lesl", 2738 [MO_LEUQ] = "leq", 2739 [MO_BEUW] = "beuw", 2740 [MO_BESW] = "besw", 2741 [MO_BEUL] = "beul", 2742 [MO_BESL] = "besl", 2743 [MO_BEUQ] = "beq", 2744 [MO_128 + MO_BE] = "beo", 2745 [MO_128 + MO_LE] = "leo", 2746 }; 2747 2748 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = { 2749 [MO_UNALN >> MO_ASHIFT] = "un+", 2750 [MO_ALIGN >> MO_ASHIFT] = "al+", 2751 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+", 2752 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+", 2753 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+", 2754 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+", 2755 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+", 2756 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+", 2757 }; 2758 2759 static const char * const atom_name[(MO_ATOM_MASK >> MO_ATOM_SHIFT) + 1] = { 2760 [MO_ATOM_IFALIGN >> MO_ATOM_SHIFT] = "", 2761 [MO_ATOM_IFALIGN_PAIR >> MO_ATOM_SHIFT] = "pair+", 2762 [MO_ATOM_WITHIN16 >> MO_ATOM_SHIFT] = "w16+", 2763 [MO_ATOM_WITHIN16_PAIR >> MO_ATOM_SHIFT] = "w16p+", 2764 [MO_ATOM_SUBALIGN >> MO_ATOM_SHIFT] = "sub+", 2765 [MO_ATOM_NONE >> MO_ATOM_SHIFT] = "noat+", 2766 }; 2767 2768 static const char bswap_flag_name[][6] = { 2769 [TCG_BSWAP_IZ] = "iz", 2770 [TCG_BSWAP_OZ] = "oz", 2771 [TCG_BSWAP_OS] = "os", 2772 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz", 2773 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os", 2774 }; 2775 2776 #ifdef CONFIG_PLUGIN 2777 static const char * const plugin_from_name[] = { 2778 "from-tb", 2779 "from-insn", 2780 "after-insn", 2781 "after-tb", 2782 }; 2783 #endif 2784 2785 static inline bool tcg_regset_single(TCGRegSet d) 2786 { 2787 return (d & (d - 1)) == 0; 2788 } 2789 2790 static inline TCGReg tcg_regset_first(TCGRegSet d) 2791 { 2792 if (TCG_TARGET_NB_REGS <= 32) { 2793 return ctz32(d); 2794 } else { 2795 return ctz64(d); 2796 } 2797 } 2798 2799 /* Return only the number of characters output -- no error return. */ 2800 #define ne_fprintf(...) \ 2801 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; }) 2802 2803 void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs) 2804 { 2805 char buf[128]; 2806 TCGOp *op; 2807 2808 QTAILQ_FOREACH(op, &s->ops, link) { 2809 int i, k, nb_oargs, nb_iargs, nb_cargs; 2810 const TCGOpDef *def; 2811 TCGOpcode c; 2812 int col = 0; 2813 2814 c = op->opc; 2815 def = &tcg_op_defs[c]; 2816 2817 if (c == INDEX_op_insn_start) { 2818 nb_oargs = 0; 2819 col += ne_fprintf(f, "\n ----"); 2820 2821 for (i = 0, k = s->insn_start_words; i < k; ++i) { 2822 col += ne_fprintf(f, " %016" PRIx64, 2823 tcg_get_insn_start_param(op, i)); 2824 } 2825 } else if (c == INDEX_op_call) { 2826 const TCGHelperInfo *info = tcg_call_info(op); 2827 void *func = tcg_call_func(op); 2828 2829 /* variable number of arguments */ 2830 nb_oargs = TCGOP_CALLO(op); 2831 nb_iargs = TCGOP_CALLI(op); 2832 nb_cargs = def->nb_cargs; 2833 2834 col += ne_fprintf(f, " %s ", def->name); 2835 2836 /* 2837 * Print the function name from TCGHelperInfo, if available. 2838 * Note that plugins have a template function for the info, 2839 * but the actual function pointer comes from the plugin. 2840 */ 2841 if (func == info->func) { 2842 col += ne_fprintf(f, "%s", info->name); 2843 } else { 2844 col += ne_fprintf(f, "plugin(%p)", func); 2845 } 2846 2847 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs); 2848 for (i = 0; i < nb_oargs; i++) { 2849 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf), 2850 op->args[i])); 2851 } 2852 for (i = 0; i < nb_iargs; i++) { 2853 TCGArg arg = op->args[nb_oargs + i]; 2854 const char *t = tcg_get_arg_str(s, buf, sizeof(buf), arg); 2855 col += ne_fprintf(f, ",%s", t); 2856 } 2857 } else { 2858 if (def->flags & TCG_OPF_INT) { 2859 col += ne_fprintf(f, " %s_i%d ", 2860 def->name, 2861 8 * tcg_type_size(TCGOP_TYPE(op))); 2862 } else if (def->flags & TCG_OPF_VECTOR) { 2863 col += ne_fprintf(f, "%s v%d,e%d,", 2864 def->name, 2865 8 * tcg_type_size(TCGOP_TYPE(op)), 2866 8 << TCGOP_VECE(op)); 2867 } else { 2868 col += ne_fprintf(f, " %s ", def->name); 2869 } 2870 2871 nb_oargs = def->nb_oargs; 2872 nb_iargs = def->nb_iargs; 2873 nb_cargs = def->nb_cargs; 2874 2875 k = 0; 2876 for (i = 0; i < nb_oargs; i++) { 2877 const char *sep = k ? "," : ""; 2878 col += ne_fprintf(f, "%s%s", sep, 2879 tcg_get_arg_str(s, buf, sizeof(buf), 2880 op->args[k++])); 2881 } 2882 for (i = 0; i < nb_iargs; i++) { 2883 const char *sep = k ? "," : ""; 2884 col += ne_fprintf(f, "%s%s", sep, 2885 tcg_get_arg_str(s, buf, sizeof(buf), 2886 op->args[k++])); 2887 } 2888 switch (c) { 2889 case INDEX_op_brcond_i32: 2890 case INDEX_op_setcond_i32: 2891 case INDEX_op_negsetcond_i32: 2892 case INDEX_op_movcond_i32: 2893 case INDEX_op_brcond2_i32: 2894 case INDEX_op_setcond2_i32: 2895 case INDEX_op_brcond_i64: 2896 case INDEX_op_setcond_i64: 2897 case INDEX_op_negsetcond_i64: 2898 case INDEX_op_movcond_i64: 2899 case INDEX_op_cmp_vec: 2900 case INDEX_op_cmpsel_vec: 2901 if (op->args[k] < ARRAY_SIZE(cond_name) 2902 && cond_name[op->args[k]]) { 2903 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]); 2904 } else { 2905 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]); 2906 } 2907 i = 1; 2908 break; 2909 case INDEX_op_qemu_ld_i32: 2910 case INDEX_op_qemu_st_i32: 2911 case INDEX_op_qemu_st8_i32: 2912 case INDEX_op_qemu_ld_i64: 2913 case INDEX_op_qemu_st_i64: 2914 case INDEX_op_qemu_ld_i128: 2915 case INDEX_op_qemu_st_i128: 2916 { 2917 const char *s_al, *s_op, *s_at; 2918 MemOpIdx oi = op->args[k++]; 2919 MemOp mop = get_memop(oi); 2920 unsigned ix = get_mmuidx(oi); 2921 2922 s_al = alignment_name[(mop & MO_AMASK) >> MO_ASHIFT]; 2923 s_op = ldst_name[mop & (MO_BSWAP | MO_SSIZE)]; 2924 s_at = atom_name[(mop & MO_ATOM_MASK) >> MO_ATOM_SHIFT]; 2925 mop &= ~(MO_AMASK | MO_BSWAP | MO_SSIZE | MO_ATOM_MASK); 2926 2927 /* If all fields are accounted for, print symbolically. */ 2928 if (!mop && s_al && s_op && s_at) { 2929 col += ne_fprintf(f, ",%s%s%s,%u", 2930 s_at, s_al, s_op, ix); 2931 } else { 2932 mop = get_memop(oi); 2933 col += ne_fprintf(f, ",$0x%x,%u", mop, ix); 2934 } 2935 i = 1; 2936 } 2937 break; 2938 case INDEX_op_bswap16_i32: 2939 case INDEX_op_bswap16_i64: 2940 case INDEX_op_bswap32_i32: 2941 case INDEX_op_bswap32_i64: 2942 case INDEX_op_bswap64_i64: 2943 { 2944 TCGArg flags = op->args[k]; 2945 const char *name = NULL; 2946 2947 if (flags < ARRAY_SIZE(bswap_flag_name)) { 2948 name = bswap_flag_name[flags]; 2949 } 2950 if (name) { 2951 col += ne_fprintf(f, ",%s", name); 2952 } else { 2953 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags); 2954 } 2955 i = k = 1; 2956 } 2957 break; 2958 #ifdef CONFIG_PLUGIN 2959 case INDEX_op_plugin_cb: 2960 { 2961 TCGArg from = op->args[k++]; 2962 const char *name = NULL; 2963 2964 if (from < ARRAY_SIZE(plugin_from_name)) { 2965 name = plugin_from_name[from]; 2966 } 2967 if (name) { 2968 col += ne_fprintf(f, "%s", name); 2969 } else { 2970 col += ne_fprintf(f, "$0x%" TCG_PRIlx, from); 2971 } 2972 i = 1; 2973 } 2974 break; 2975 #endif 2976 default: 2977 i = 0; 2978 break; 2979 } 2980 switch (c) { 2981 case INDEX_op_set_label: 2982 case INDEX_op_br: 2983 case INDEX_op_brcond_i32: 2984 case INDEX_op_brcond_i64: 2985 case INDEX_op_brcond2_i32: 2986 col += ne_fprintf(f, "%s$L%d", k ? "," : "", 2987 arg_label(op->args[k])->id); 2988 i++, k++; 2989 break; 2990 case INDEX_op_mb: 2991 { 2992 TCGBar membar = op->args[k]; 2993 const char *b_op, *m_op; 2994 2995 switch (membar & TCG_BAR_SC) { 2996 case 0: 2997 b_op = "none"; 2998 break; 2999 case TCG_BAR_LDAQ: 3000 b_op = "acq"; 3001 break; 3002 case TCG_BAR_STRL: 3003 b_op = "rel"; 3004 break; 3005 case TCG_BAR_SC: 3006 b_op = "seq"; 3007 break; 3008 default: 3009 g_assert_not_reached(); 3010 } 3011 3012 switch (membar & TCG_MO_ALL) { 3013 case 0: 3014 m_op = "none"; 3015 break; 3016 case TCG_MO_LD_LD: 3017 m_op = "rr"; 3018 break; 3019 case TCG_MO_LD_ST: 3020 m_op = "rw"; 3021 break; 3022 case TCG_MO_ST_LD: 3023 m_op = "wr"; 3024 break; 3025 case TCG_MO_ST_ST: 3026 m_op = "ww"; 3027 break; 3028 case TCG_MO_LD_LD | TCG_MO_LD_ST: 3029 m_op = "rr+rw"; 3030 break; 3031 case TCG_MO_LD_LD | TCG_MO_ST_LD: 3032 m_op = "rr+wr"; 3033 break; 3034 case TCG_MO_LD_LD | TCG_MO_ST_ST: 3035 m_op = "rr+ww"; 3036 break; 3037 case TCG_MO_LD_ST | TCG_MO_ST_LD: 3038 m_op = "rw+wr"; 3039 break; 3040 case TCG_MO_LD_ST | TCG_MO_ST_ST: 3041 m_op = "rw+ww"; 3042 break; 3043 case TCG_MO_ST_LD | TCG_MO_ST_ST: 3044 m_op = "wr+ww"; 3045 break; 3046 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_LD: 3047 m_op = "rr+rw+wr"; 3048 break; 3049 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_ST: 3050 m_op = "rr+rw+ww"; 3051 break; 3052 case TCG_MO_LD_LD | TCG_MO_ST_LD | TCG_MO_ST_ST: 3053 m_op = "rr+wr+ww"; 3054 break; 3055 case TCG_MO_LD_ST | TCG_MO_ST_LD | TCG_MO_ST_ST: 3056 m_op = "rw+wr+ww"; 3057 break; 3058 case TCG_MO_ALL: 3059 m_op = "all"; 3060 break; 3061 default: 3062 g_assert_not_reached(); 3063 } 3064 3065 col += ne_fprintf(f, "%s%s:%s", (k ? "," : ""), b_op, m_op); 3066 i++, k++; 3067 } 3068 break; 3069 default: 3070 break; 3071 } 3072 for (; i < nb_cargs; i++, k++) { 3073 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "", 3074 op->args[k]); 3075 } 3076 } 3077 3078 if (have_prefs || op->life) { 3079 for (; col < 40; ++col) { 3080 putc(' ', f); 3081 } 3082 } 3083 3084 if (op->life) { 3085 unsigned life = op->life; 3086 3087 if (life & (SYNC_ARG * 3)) { 3088 ne_fprintf(f, " sync:"); 3089 for (i = 0; i < 2; ++i) { 3090 if (life & (SYNC_ARG << i)) { 3091 ne_fprintf(f, " %d", i); 3092 } 3093 } 3094 } 3095 life /= DEAD_ARG; 3096 if (life) { 3097 ne_fprintf(f, " dead:"); 3098 for (i = 0; life; ++i, life >>= 1) { 3099 if (life & 1) { 3100 ne_fprintf(f, " %d", i); 3101 } 3102 } 3103 } 3104 } 3105 3106 if (have_prefs) { 3107 for (i = 0; i < nb_oargs; ++i) { 3108 TCGRegSet set = output_pref(op, i); 3109 3110 if (i == 0) { 3111 ne_fprintf(f, " pref="); 3112 } else { 3113 ne_fprintf(f, ","); 3114 } 3115 if (set == 0) { 3116 ne_fprintf(f, "none"); 3117 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) { 3118 ne_fprintf(f, "all"); 3119 #ifdef CONFIG_DEBUG_TCG 3120 } else if (tcg_regset_single(set)) { 3121 TCGReg reg = tcg_regset_first(set); 3122 ne_fprintf(f, "%s", tcg_target_reg_names[reg]); 3123 #endif 3124 } else if (TCG_TARGET_NB_REGS <= 32) { 3125 ne_fprintf(f, "0x%x", (uint32_t)set); 3126 } else { 3127 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set); 3128 } 3129 } 3130 } 3131 3132 putc('\n', f); 3133 } 3134 } 3135 3136 /* we give more priority to constraints with less registers */ 3137 static int get_constraint_priority(const TCGArgConstraint *arg_ct, int k) 3138 { 3139 int n; 3140 3141 arg_ct += k; 3142 n = ctpop64(arg_ct->regs); 3143 3144 /* 3145 * Sort constraints of a single register first, which includes output 3146 * aliases (which must exactly match the input already allocated). 3147 */ 3148 if (n == 1 || arg_ct->oalias) { 3149 return INT_MAX; 3150 } 3151 3152 /* 3153 * Sort register pairs next, first then second immediately after. 3154 * Arbitrarily sort multiple pairs by the index of the first reg; 3155 * there shouldn't be many pairs. 3156 */ 3157 switch (arg_ct->pair) { 3158 case 1: 3159 case 3: 3160 return (k + 1) * 2; 3161 case 2: 3162 return (arg_ct->pair_index + 1) * 2 - 1; 3163 } 3164 3165 /* Finally, sort by decreasing register count. */ 3166 assert(n > 1); 3167 return -n; 3168 } 3169 3170 /* sort from highest priority to lowest */ 3171 static void sort_constraints(TCGArgConstraint *a, int start, int n) 3172 { 3173 int i, j; 3174 3175 for (i = 0; i < n; i++) { 3176 a[start + i].sort_index = start + i; 3177 } 3178 if (n <= 1) { 3179 return; 3180 } 3181 for (i = 0; i < n - 1; i++) { 3182 for (j = i + 1; j < n; j++) { 3183 int p1 = get_constraint_priority(a, a[start + i].sort_index); 3184 int p2 = get_constraint_priority(a, a[start + j].sort_index); 3185 if (p1 < p2) { 3186 int tmp = a[start + i].sort_index; 3187 a[start + i].sort_index = a[start + j].sort_index; 3188 a[start + j].sort_index = tmp; 3189 } 3190 } 3191 } 3192 } 3193 3194 static const TCGArgConstraint empty_cts[TCG_MAX_OP_ARGS]; 3195 static TCGArgConstraint all_cts[ARRAY_SIZE(constraint_sets)][TCG_MAX_OP_ARGS]; 3196 3197 static void process_constraint_sets(void) 3198 { 3199 for (size_t c = 0; c < ARRAY_SIZE(constraint_sets); ++c) { 3200 const TCGConstraintSet *tdefs = &constraint_sets[c]; 3201 TCGArgConstraint *args_ct = all_cts[c]; 3202 int nb_oargs = tdefs->nb_oargs; 3203 int nb_iargs = tdefs->nb_iargs; 3204 int nb_args = nb_oargs + nb_iargs; 3205 bool saw_alias_pair = false; 3206 3207 for (int i = 0; i < nb_args; i++) { 3208 const char *ct_str = tdefs->args_ct_str[i]; 3209 bool input_p = i >= nb_oargs; 3210 int o; 3211 3212 switch (*ct_str) { 3213 case '0' ... '9': 3214 o = *ct_str - '0'; 3215 tcg_debug_assert(input_p); 3216 tcg_debug_assert(o < nb_oargs); 3217 tcg_debug_assert(args_ct[o].regs != 0); 3218 tcg_debug_assert(!args_ct[o].oalias); 3219 args_ct[i] = args_ct[o]; 3220 /* The output sets oalias. */ 3221 args_ct[o].oalias = 1; 3222 args_ct[o].alias_index = i; 3223 /* The input sets ialias. */ 3224 args_ct[i].ialias = 1; 3225 args_ct[i].alias_index = o; 3226 if (args_ct[i].pair) { 3227 saw_alias_pair = true; 3228 } 3229 tcg_debug_assert(ct_str[1] == '\0'); 3230 continue; 3231 3232 case '&': 3233 tcg_debug_assert(!input_p); 3234 args_ct[i].newreg = true; 3235 ct_str++; 3236 break; 3237 3238 case 'p': /* plus */ 3239 /* Allocate to the register after the previous. */ 3240 tcg_debug_assert(i > (input_p ? nb_oargs : 0)); 3241 o = i - 1; 3242 tcg_debug_assert(!args_ct[o].pair); 3243 tcg_debug_assert(!args_ct[o].ct); 3244 args_ct[i] = (TCGArgConstraint){ 3245 .pair = 2, 3246 .pair_index = o, 3247 .regs = args_ct[o].regs << 1, 3248 .newreg = args_ct[o].newreg, 3249 }; 3250 args_ct[o].pair = 1; 3251 args_ct[o].pair_index = i; 3252 tcg_debug_assert(ct_str[1] == '\0'); 3253 continue; 3254 3255 case 'm': /* minus */ 3256 /* Allocate to the register before the previous. */ 3257 tcg_debug_assert(i > (input_p ? nb_oargs : 0)); 3258 o = i - 1; 3259 tcg_debug_assert(!args_ct[o].pair); 3260 tcg_debug_assert(!args_ct[o].ct); 3261 args_ct[i] = (TCGArgConstraint){ 3262 .pair = 1, 3263 .pair_index = o, 3264 .regs = args_ct[o].regs >> 1, 3265 .newreg = args_ct[o].newreg, 3266 }; 3267 args_ct[o].pair = 2; 3268 args_ct[o].pair_index = i; 3269 tcg_debug_assert(ct_str[1] == '\0'); 3270 continue; 3271 } 3272 3273 do { 3274 switch (*ct_str) { 3275 case 'i': 3276 args_ct[i].ct |= TCG_CT_CONST; 3277 break; 3278 #ifdef TCG_REG_ZERO 3279 case 'z': 3280 args_ct[i].ct |= TCG_CT_REG_ZERO; 3281 break; 3282 #endif 3283 3284 /* Include all of the target-specific constraints. */ 3285 3286 #undef CONST 3287 #define CONST(CASE, MASK) \ 3288 case CASE: args_ct[i].ct |= MASK; break; 3289 #define REGS(CASE, MASK) \ 3290 case CASE: args_ct[i].regs |= MASK; break; 3291 3292 #include "tcg-target-con-str.h" 3293 3294 #undef REGS 3295 #undef CONST 3296 default: 3297 case '0' ... '9': 3298 case '&': 3299 case 'p': 3300 case 'm': 3301 /* Typo in TCGConstraintSet constraint. */ 3302 g_assert_not_reached(); 3303 } 3304 } while (*++ct_str != '\0'); 3305 } 3306 3307 /* 3308 * Fix up output pairs that are aliased with inputs. 3309 * When we created the alias, we copied pair from the output. 3310 * There are three cases: 3311 * (1a) Pairs of inputs alias pairs of outputs. 3312 * (1b) One input aliases the first of a pair of outputs. 3313 * (2) One input aliases the second of a pair of outputs. 3314 * 3315 * Case 1a is handled by making sure that the pair_index'es are 3316 * properly updated so that they appear the same as a pair of inputs. 3317 * 3318 * Case 1b is handled by setting the pair_index of the input to 3319 * itself, simply so it doesn't point to an unrelated argument. 3320 * Since we don't encounter the "second" during the input allocation 3321 * phase, nothing happens with the second half of the input pair. 3322 * 3323 * Case 2 is handled by setting the second input to pair=3, the 3324 * first output to pair=3, and the pair_index'es to match. 3325 */ 3326 if (saw_alias_pair) { 3327 for (int i = nb_oargs; i < nb_args; i++) { 3328 int o, o2, i2; 3329 3330 /* 3331 * Since [0-9pm] must be alone in the constraint string, 3332 * the only way they can both be set is if the pair comes 3333 * from the output alias. 3334 */ 3335 if (!args_ct[i].ialias) { 3336 continue; 3337 } 3338 switch (args_ct[i].pair) { 3339 case 0: 3340 break; 3341 case 1: 3342 o = args_ct[i].alias_index; 3343 o2 = args_ct[o].pair_index; 3344 tcg_debug_assert(args_ct[o].pair == 1); 3345 tcg_debug_assert(args_ct[o2].pair == 2); 3346 if (args_ct[o2].oalias) { 3347 /* Case 1a */ 3348 i2 = args_ct[o2].alias_index; 3349 tcg_debug_assert(args_ct[i2].pair == 2); 3350 args_ct[i2].pair_index = i; 3351 args_ct[i].pair_index = i2; 3352 } else { 3353 /* Case 1b */ 3354 args_ct[i].pair_index = i; 3355 } 3356 break; 3357 case 2: 3358 o = args_ct[i].alias_index; 3359 o2 = args_ct[o].pair_index; 3360 tcg_debug_assert(args_ct[o].pair == 2); 3361 tcg_debug_assert(args_ct[o2].pair == 1); 3362 if (args_ct[o2].oalias) { 3363 /* Case 1a */ 3364 i2 = args_ct[o2].alias_index; 3365 tcg_debug_assert(args_ct[i2].pair == 1); 3366 args_ct[i2].pair_index = i; 3367 args_ct[i].pair_index = i2; 3368 } else { 3369 /* Case 2 */ 3370 args_ct[i].pair = 3; 3371 args_ct[o2].pair = 3; 3372 args_ct[i].pair_index = o2; 3373 args_ct[o2].pair_index = i; 3374 } 3375 break; 3376 default: 3377 g_assert_not_reached(); 3378 } 3379 } 3380 } 3381 3382 /* sort the constraints (XXX: this is just an heuristic) */ 3383 sort_constraints(args_ct, 0, nb_oargs); 3384 sort_constraints(args_ct, nb_oargs, nb_iargs); 3385 } 3386 } 3387 3388 static const TCGArgConstraint *opcode_args_ct(const TCGOp *op) 3389 { 3390 TCGOpcode opc = op->opc; 3391 TCGType type = TCGOP_TYPE(op); 3392 unsigned flags = TCGOP_FLAGS(op); 3393 const TCGOpDef *def = &tcg_op_defs[opc]; 3394 const TCGOutOp *outop = all_outop[opc]; 3395 TCGConstraintSetIndex con_set; 3396 3397 if (def->flags & TCG_OPF_NOT_PRESENT) { 3398 return empty_cts; 3399 } 3400 3401 if (outop) { 3402 con_set = outop->static_constraint; 3403 if (con_set == C_Dynamic) { 3404 con_set = outop->dynamic_constraint(type, flags); 3405 } 3406 } else { 3407 con_set = tcg_target_op_def(opc, type, flags); 3408 } 3409 tcg_debug_assert(con_set >= 0); 3410 tcg_debug_assert(con_set < ARRAY_SIZE(constraint_sets)); 3411 3412 /* The constraint arguments must match TCGOpcode arguments. */ 3413 tcg_debug_assert(constraint_sets[con_set].nb_oargs == def->nb_oargs); 3414 tcg_debug_assert(constraint_sets[con_set].nb_iargs == def->nb_iargs); 3415 3416 return all_cts[con_set]; 3417 } 3418 3419 static void remove_label_use(TCGOp *op, int idx) 3420 { 3421 TCGLabel *label = arg_label(op->args[idx]); 3422 TCGLabelUse *use; 3423 3424 QSIMPLEQ_FOREACH(use, &label->branches, next) { 3425 if (use->op == op) { 3426 QSIMPLEQ_REMOVE(&label->branches, use, TCGLabelUse, next); 3427 return; 3428 } 3429 } 3430 g_assert_not_reached(); 3431 } 3432 3433 void tcg_op_remove(TCGContext *s, TCGOp *op) 3434 { 3435 switch (op->opc) { 3436 case INDEX_op_br: 3437 remove_label_use(op, 0); 3438 break; 3439 case INDEX_op_brcond_i32: 3440 case INDEX_op_brcond_i64: 3441 remove_label_use(op, 3); 3442 break; 3443 case INDEX_op_brcond2_i32: 3444 remove_label_use(op, 5); 3445 break; 3446 default: 3447 break; 3448 } 3449 3450 QTAILQ_REMOVE(&s->ops, op, link); 3451 QTAILQ_INSERT_TAIL(&s->free_ops, op, link); 3452 s->nb_ops--; 3453 } 3454 3455 void tcg_remove_ops_after(TCGOp *op) 3456 { 3457 TCGContext *s = tcg_ctx; 3458 3459 while (true) { 3460 TCGOp *last = tcg_last_op(); 3461 if (last == op) { 3462 return; 3463 } 3464 tcg_op_remove(s, last); 3465 } 3466 } 3467 3468 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs) 3469 { 3470 TCGContext *s = tcg_ctx; 3471 TCGOp *op = NULL; 3472 3473 if (unlikely(!QTAILQ_EMPTY(&s->free_ops))) { 3474 QTAILQ_FOREACH(op, &s->free_ops, link) { 3475 if (nargs <= op->nargs) { 3476 QTAILQ_REMOVE(&s->free_ops, op, link); 3477 nargs = op->nargs; 3478 goto found; 3479 } 3480 } 3481 } 3482 3483 /* Most opcodes have 3 or 4 operands: reduce fragmentation. */ 3484 nargs = MAX(4, nargs); 3485 op = tcg_malloc(sizeof(TCGOp) + sizeof(TCGArg) * nargs); 3486 3487 found: 3488 memset(op, 0, offsetof(TCGOp, link)); 3489 op->opc = opc; 3490 op->nargs = nargs; 3491 3492 /* Check for bitfield overflow. */ 3493 tcg_debug_assert(op->nargs == nargs); 3494 3495 s->nb_ops++; 3496 return op; 3497 } 3498 3499 TCGOp *tcg_emit_op(TCGOpcode opc, unsigned nargs) 3500 { 3501 TCGOp *op = tcg_op_alloc(opc, nargs); 3502 3503 if (tcg_ctx->emit_before_op) { 3504 QTAILQ_INSERT_BEFORE(tcg_ctx->emit_before_op, op, link); 3505 } else { 3506 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 3507 } 3508 return op; 3509 } 3510 3511 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op, 3512 TCGOpcode opc, TCGType type, unsigned nargs) 3513 { 3514 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3515 3516 TCGOP_TYPE(new_op) = type; 3517 QTAILQ_INSERT_BEFORE(old_op, new_op, link); 3518 return new_op; 3519 } 3520 3521 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op, 3522 TCGOpcode opc, TCGType type, unsigned nargs) 3523 { 3524 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3525 3526 TCGOP_TYPE(new_op) = type; 3527 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link); 3528 return new_op; 3529 } 3530 3531 static void move_label_uses(TCGLabel *to, TCGLabel *from) 3532 { 3533 TCGLabelUse *u; 3534 3535 QSIMPLEQ_FOREACH(u, &from->branches, next) { 3536 TCGOp *op = u->op; 3537 switch (op->opc) { 3538 case INDEX_op_br: 3539 op->args[0] = label_arg(to); 3540 break; 3541 case INDEX_op_brcond_i32: 3542 case INDEX_op_brcond_i64: 3543 op->args[3] = label_arg(to); 3544 break; 3545 case INDEX_op_brcond2_i32: 3546 op->args[5] = label_arg(to); 3547 break; 3548 default: 3549 g_assert_not_reached(); 3550 } 3551 } 3552 3553 QSIMPLEQ_CONCAT(&to->branches, &from->branches); 3554 } 3555 3556 /* Reachable analysis : remove unreachable code. */ 3557 static void __attribute__((noinline)) 3558 reachable_code_pass(TCGContext *s) 3559 { 3560 TCGOp *op, *op_next, *op_prev; 3561 bool dead = false; 3562 3563 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3564 bool remove = dead; 3565 TCGLabel *label; 3566 3567 switch (op->opc) { 3568 case INDEX_op_set_label: 3569 label = arg_label(op->args[0]); 3570 3571 /* 3572 * Note that the first op in the TB is always a load, 3573 * so there is always something before a label. 3574 */ 3575 op_prev = QTAILQ_PREV(op, link); 3576 3577 /* 3578 * If we find two sequential labels, move all branches to 3579 * reference the second label and remove the first label. 3580 * Do this before branch to next optimization, so that the 3581 * middle label is out of the way. 3582 */ 3583 if (op_prev->opc == INDEX_op_set_label) { 3584 move_label_uses(label, arg_label(op_prev->args[0])); 3585 tcg_op_remove(s, op_prev); 3586 op_prev = QTAILQ_PREV(op, link); 3587 } 3588 3589 /* 3590 * Optimization can fold conditional branches to unconditional. 3591 * If we find a label which is preceded by an unconditional 3592 * branch to next, remove the branch. We couldn't do this when 3593 * processing the branch because any dead code between the branch 3594 * and label had not yet been removed. 3595 */ 3596 if (op_prev->opc == INDEX_op_br && 3597 label == arg_label(op_prev->args[0])) { 3598 tcg_op_remove(s, op_prev); 3599 /* Fall through means insns become live again. */ 3600 dead = false; 3601 } 3602 3603 if (QSIMPLEQ_EMPTY(&label->branches)) { 3604 /* 3605 * While there is an occasional backward branch, virtually 3606 * all branches generated by the translators are forward. 3607 * Which means that generally we will have already removed 3608 * all references to the label that will be, and there is 3609 * little to be gained by iterating. 3610 */ 3611 remove = true; 3612 } else { 3613 /* Once we see a label, insns become live again. */ 3614 dead = false; 3615 remove = false; 3616 } 3617 break; 3618 3619 case INDEX_op_br: 3620 case INDEX_op_exit_tb: 3621 case INDEX_op_goto_ptr: 3622 /* Unconditional branches; everything following is dead. */ 3623 dead = true; 3624 break; 3625 3626 case INDEX_op_call: 3627 /* Notice noreturn helper calls, raising exceptions. */ 3628 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) { 3629 dead = true; 3630 } 3631 break; 3632 3633 case INDEX_op_insn_start: 3634 /* Never remove -- we need to keep these for unwind. */ 3635 remove = false; 3636 break; 3637 3638 default: 3639 break; 3640 } 3641 3642 if (remove) { 3643 tcg_op_remove(s, op); 3644 } 3645 } 3646 } 3647 3648 #define TS_DEAD 1 3649 #define TS_MEM 2 3650 3651 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n))) 3652 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n))) 3653 3654 /* For liveness_pass_1, the register preferences for a given temp. */ 3655 static inline TCGRegSet *la_temp_pref(TCGTemp *ts) 3656 { 3657 return ts->state_ptr; 3658 } 3659 3660 /* For liveness_pass_1, reset the preferences for a given temp to the 3661 * maximal regset for its type. 3662 */ 3663 static inline void la_reset_pref(TCGTemp *ts) 3664 { 3665 *la_temp_pref(ts) 3666 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]); 3667 } 3668 3669 /* liveness analysis: end of function: all temps are dead, and globals 3670 should be in memory. */ 3671 static void la_func_end(TCGContext *s, int ng, int nt) 3672 { 3673 int i; 3674 3675 for (i = 0; i < ng; ++i) { 3676 s->temps[i].state = TS_DEAD | TS_MEM; 3677 la_reset_pref(&s->temps[i]); 3678 } 3679 for (i = ng; i < nt; ++i) { 3680 s->temps[i].state = TS_DEAD; 3681 la_reset_pref(&s->temps[i]); 3682 } 3683 } 3684 3685 /* liveness analysis: end of basic block: all temps are dead, globals 3686 and local temps should be in memory. */ 3687 static void la_bb_end(TCGContext *s, int ng, int nt) 3688 { 3689 int i; 3690 3691 for (i = 0; i < nt; ++i) { 3692 TCGTemp *ts = &s->temps[i]; 3693 int state; 3694 3695 switch (ts->kind) { 3696 case TEMP_FIXED: 3697 case TEMP_GLOBAL: 3698 case TEMP_TB: 3699 state = TS_DEAD | TS_MEM; 3700 break; 3701 case TEMP_EBB: 3702 case TEMP_CONST: 3703 state = TS_DEAD; 3704 break; 3705 default: 3706 g_assert_not_reached(); 3707 } 3708 ts->state = state; 3709 la_reset_pref(ts); 3710 } 3711 } 3712 3713 /* liveness analysis: sync globals back to memory. */ 3714 static void la_global_sync(TCGContext *s, int ng) 3715 { 3716 int i; 3717 3718 for (i = 0; i < ng; ++i) { 3719 int state = s->temps[i].state; 3720 s->temps[i].state = state | TS_MEM; 3721 if (state == TS_DEAD) { 3722 /* If the global was previously dead, reset prefs. */ 3723 la_reset_pref(&s->temps[i]); 3724 } 3725 } 3726 } 3727 3728 /* 3729 * liveness analysis: conditional branch: all temps are dead unless 3730 * explicitly live-across-conditional-branch, globals and local temps 3731 * should be synced. 3732 */ 3733 static void la_bb_sync(TCGContext *s, int ng, int nt) 3734 { 3735 la_global_sync(s, ng); 3736 3737 for (int i = ng; i < nt; ++i) { 3738 TCGTemp *ts = &s->temps[i]; 3739 int state; 3740 3741 switch (ts->kind) { 3742 case TEMP_TB: 3743 state = ts->state; 3744 ts->state = state | TS_MEM; 3745 if (state != TS_DEAD) { 3746 continue; 3747 } 3748 break; 3749 case TEMP_EBB: 3750 case TEMP_CONST: 3751 continue; 3752 default: 3753 g_assert_not_reached(); 3754 } 3755 la_reset_pref(&s->temps[i]); 3756 } 3757 } 3758 3759 /* liveness analysis: sync globals back to memory and kill. */ 3760 static void la_global_kill(TCGContext *s, int ng) 3761 { 3762 int i; 3763 3764 for (i = 0; i < ng; i++) { 3765 s->temps[i].state = TS_DEAD | TS_MEM; 3766 la_reset_pref(&s->temps[i]); 3767 } 3768 } 3769 3770 /* liveness analysis: note live globals crossing calls. */ 3771 static void la_cross_call(TCGContext *s, int nt) 3772 { 3773 TCGRegSet mask = ~tcg_target_call_clobber_regs; 3774 int i; 3775 3776 for (i = 0; i < nt; i++) { 3777 TCGTemp *ts = &s->temps[i]; 3778 if (!(ts->state & TS_DEAD)) { 3779 TCGRegSet *pset = la_temp_pref(ts); 3780 TCGRegSet set = *pset; 3781 3782 set &= mask; 3783 /* If the combination is not possible, restart. */ 3784 if (set == 0) { 3785 set = tcg_target_available_regs[ts->type] & mask; 3786 } 3787 *pset = set; 3788 } 3789 } 3790 } 3791 3792 /* 3793 * Liveness analysis: Verify the lifetime of TEMP_TB, and reduce 3794 * to TEMP_EBB, if possible. 3795 */ 3796 static void __attribute__((noinline)) 3797 liveness_pass_0(TCGContext *s) 3798 { 3799 void * const multiple_ebb = (void *)(uintptr_t)-1; 3800 int nb_temps = s->nb_temps; 3801 TCGOp *op, *ebb; 3802 3803 for (int i = s->nb_globals; i < nb_temps; ++i) { 3804 s->temps[i].state_ptr = NULL; 3805 } 3806 3807 /* 3808 * Represent each EBB by the op at which it begins. In the case of 3809 * the first EBB, this is the first op, otherwise it is a label. 3810 * Collect the uses of each TEMP_TB: NULL for unused, EBB for use 3811 * within a single EBB, else MULTIPLE_EBB. 3812 */ 3813 ebb = QTAILQ_FIRST(&s->ops); 3814 QTAILQ_FOREACH(op, &s->ops, link) { 3815 const TCGOpDef *def; 3816 int nb_oargs, nb_iargs; 3817 3818 switch (op->opc) { 3819 case INDEX_op_set_label: 3820 ebb = op; 3821 continue; 3822 case INDEX_op_discard: 3823 continue; 3824 case INDEX_op_call: 3825 nb_oargs = TCGOP_CALLO(op); 3826 nb_iargs = TCGOP_CALLI(op); 3827 break; 3828 default: 3829 def = &tcg_op_defs[op->opc]; 3830 nb_oargs = def->nb_oargs; 3831 nb_iargs = def->nb_iargs; 3832 break; 3833 } 3834 3835 for (int i = 0; i < nb_oargs + nb_iargs; ++i) { 3836 TCGTemp *ts = arg_temp(op->args[i]); 3837 3838 if (ts->kind != TEMP_TB) { 3839 continue; 3840 } 3841 if (ts->state_ptr == NULL) { 3842 ts->state_ptr = ebb; 3843 } else if (ts->state_ptr != ebb) { 3844 ts->state_ptr = multiple_ebb; 3845 } 3846 } 3847 } 3848 3849 /* 3850 * For TEMP_TB that turned out not to be used beyond one EBB, 3851 * reduce the liveness to TEMP_EBB. 3852 */ 3853 for (int i = s->nb_globals; i < nb_temps; ++i) { 3854 TCGTemp *ts = &s->temps[i]; 3855 if (ts->kind == TEMP_TB && ts->state_ptr != multiple_ebb) { 3856 ts->kind = TEMP_EBB; 3857 } 3858 } 3859 } 3860 3861 /* Liveness analysis : update the opc_arg_life array to tell if a 3862 given input arguments is dead. Instructions updating dead 3863 temporaries are removed. */ 3864 static void __attribute__((noinline)) 3865 liveness_pass_1(TCGContext *s) 3866 { 3867 int nb_globals = s->nb_globals; 3868 int nb_temps = s->nb_temps; 3869 TCGOp *op, *op_prev; 3870 TCGRegSet *prefs; 3871 int i; 3872 3873 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps); 3874 for (i = 0; i < nb_temps; ++i) { 3875 s->temps[i].state_ptr = prefs + i; 3876 } 3877 3878 /* ??? Should be redundant with the exit_tb that ends the TB. */ 3879 la_func_end(s, nb_globals, nb_temps); 3880 3881 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) { 3882 int nb_iargs, nb_oargs; 3883 TCGOpcode opc_new, opc_new2; 3884 bool have_opc_new2; 3885 TCGLifeData arg_life = 0; 3886 TCGTemp *ts; 3887 TCGOpcode opc = op->opc; 3888 const TCGOpDef *def = &tcg_op_defs[opc]; 3889 const TCGArgConstraint *args_ct; 3890 3891 switch (opc) { 3892 case INDEX_op_call: 3893 { 3894 const TCGHelperInfo *info = tcg_call_info(op); 3895 int call_flags = tcg_call_flags(op); 3896 3897 nb_oargs = TCGOP_CALLO(op); 3898 nb_iargs = TCGOP_CALLI(op); 3899 3900 /* pure functions can be removed if their result is unused */ 3901 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) { 3902 for (i = 0; i < nb_oargs; i++) { 3903 ts = arg_temp(op->args[i]); 3904 if (ts->state != TS_DEAD) { 3905 goto do_not_remove_call; 3906 } 3907 } 3908 goto do_remove; 3909 } 3910 do_not_remove_call: 3911 3912 /* Output args are dead. */ 3913 for (i = 0; i < nb_oargs; i++) { 3914 ts = arg_temp(op->args[i]); 3915 if (ts->state & TS_DEAD) { 3916 arg_life |= DEAD_ARG << i; 3917 } 3918 if (ts->state & TS_MEM) { 3919 arg_life |= SYNC_ARG << i; 3920 } 3921 ts->state = TS_DEAD; 3922 la_reset_pref(ts); 3923 } 3924 3925 /* Not used -- it will be tcg_target_call_oarg_reg(). */ 3926 memset(op->output_pref, 0, sizeof(op->output_pref)); 3927 3928 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS | 3929 TCG_CALL_NO_READ_GLOBALS))) { 3930 la_global_kill(s, nb_globals); 3931 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) { 3932 la_global_sync(s, nb_globals); 3933 } 3934 3935 /* Record arguments that die in this helper. */ 3936 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3937 ts = arg_temp(op->args[i]); 3938 if (ts->state & TS_DEAD) { 3939 arg_life |= DEAD_ARG << i; 3940 } 3941 } 3942 3943 /* For all live registers, remove call-clobbered prefs. */ 3944 la_cross_call(s, nb_temps); 3945 3946 /* 3947 * Input arguments are live for preceding opcodes. 3948 * 3949 * For those arguments that die, and will be allocated in 3950 * registers, clear the register set for that arg, to be 3951 * filled in below. For args that will be on the stack, 3952 * reset to any available reg. Process arguments in reverse 3953 * order so that if a temp is used more than once, the stack 3954 * reset to max happens before the register reset to 0. 3955 */ 3956 for (i = nb_iargs - 1; i >= 0; i--) { 3957 const TCGCallArgumentLoc *loc = &info->in[i]; 3958 ts = arg_temp(op->args[nb_oargs + i]); 3959 3960 if (ts->state & TS_DEAD) { 3961 switch (loc->kind) { 3962 case TCG_CALL_ARG_NORMAL: 3963 case TCG_CALL_ARG_EXTEND_U: 3964 case TCG_CALL_ARG_EXTEND_S: 3965 if (arg_slot_reg_p(loc->arg_slot)) { 3966 *la_temp_pref(ts) = 0; 3967 break; 3968 } 3969 /* fall through */ 3970 default: 3971 *la_temp_pref(ts) = 3972 tcg_target_available_regs[ts->type]; 3973 break; 3974 } 3975 ts->state &= ~TS_DEAD; 3976 } 3977 } 3978 3979 /* 3980 * For each input argument, add its input register to prefs. 3981 * If a temp is used once, this produces a single set bit; 3982 * if a temp is used multiple times, this produces a set. 3983 */ 3984 for (i = 0; i < nb_iargs; i++) { 3985 const TCGCallArgumentLoc *loc = &info->in[i]; 3986 ts = arg_temp(op->args[nb_oargs + i]); 3987 3988 switch (loc->kind) { 3989 case TCG_CALL_ARG_NORMAL: 3990 case TCG_CALL_ARG_EXTEND_U: 3991 case TCG_CALL_ARG_EXTEND_S: 3992 if (arg_slot_reg_p(loc->arg_slot)) { 3993 tcg_regset_set_reg(*la_temp_pref(ts), 3994 tcg_target_call_iarg_regs[loc->arg_slot]); 3995 } 3996 break; 3997 default: 3998 break; 3999 } 4000 } 4001 } 4002 break; 4003 case INDEX_op_insn_start: 4004 break; 4005 case INDEX_op_discard: 4006 /* mark the temporary as dead */ 4007 ts = arg_temp(op->args[0]); 4008 ts->state = TS_DEAD; 4009 la_reset_pref(ts); 4010 break; 4011 4012 case INDEX_op_add2_i32: 4013 case INDEX_op_add2_i64: 4014 opc_new = INDEX_op_add; 4015 goto do_addsub2; 4016 case INDEX_op_sub2_i32: 4017 opc_new = INDEX_op_sub_i32; 4018 goto do_addsub2; 4019 case INDEX_op_sub2_i64: 4020 opc_new = INDEX_op_sub_i64; 4021 do_addsub2: 4022 nb_iargs = 4; 4023 nb_oargs = 2; 4024 /* Test if the high part of the operation is dead, but not 4025 the low part. The result can be optimized to a simple 4026 add or sub. This happens often for x86_64 guest when the 4027 cpu mode is set to 32 bit. */ 4028 if (arg_temp(op->args[1])->state == TS_DEAD) { 4029 if (arg_temp(op->args[0])->state == TS_DEAD) { 4030 goto do_remove; 4031 } 4032 /* Replace the opcode and adjust the args in place, 4033 leaving 3 unused args at the end. */ 4034 op->opc = opc = opc_new; 4035 op->args[1] = op->args[2]; 4036 op->args[2] = op->args[4]; 4037 /* Fall through and mark the single-word operation live. */ 4038 nb_iargs = 2; 4039 nb_oargs = 1; 4040 } 4041 goto do_not_remove; 4042 4043 case INDEX_op_mulu2_i32: 4044 opc_new = INDEX_op_mul_i32; 4045 opc_new2 = INDEX_op_muluh_i32; 4046 have_opc_new2 = TCG_TARGET_HAS_muluh_i32; 4047 goto do_mul2; 4048 case INDEX_op_muls2_i32: 4049 opc_new = INDEX_op_mul_i32; 4050 opc_new2 = INDEX_op_mulsh_i32; 4051 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32; 4052 goto do_mul2; 4053 case INDEX_op_mulu2_i64: 4054 opc_new = INDEX_op_mul_i64; 4055 opc_new2 = INDEX_op_muluh_i64; 4056 have_opc_new2 = TCG_TARGET_HAS_muluh_i64; 4057 goto do_mul2; 4058 case INDEX_op_muls2_i64: 4059 opc_new = INDEX_op_mul_i64; 4060 opc_new2 = INDEX_op_mulsh_i64; 4061 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64; 4062 goto do_mul2; 4063 do_mul2: 4064 nb_iargs = 2; 4065 nb_oargs = 2; 4066 if (arg_temp(op->args[1])->state == TS_DEAD) { 4067 if (arg_temp(op->args[0])->state == TS_DEAD) { 4068 /* Both parts of the operation are dead. */ 4069 goto do_remove; 4070 } 4071 /* The high part of the operation is dead; generate the low. */ 4072 op->opc = opc = opc_new; 4073 op->args[1] = op->args[2]; 4074 op->args[2] = op->args[3]; 4075 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) { 4076 /* The low part of the operation is dead; generate the high. */ 4077 op->opc = opc = opc_new2; 4078 op->args[0] = op->args[1]; 4079 op->args[1] = op->args[2]; 4080 op->args[2] = op->args[3]; 4081 } else { 4082 goto do_not_remove; 4083 } 4084 /* Mark the single-word operation live. */ 4085 nb_oargs = 1; 4086 goto do_not_remove; 4087 4088 default: 4089 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */ 4090 nb_iargs = def->nb_iargs; 4091 nb_oargs = def->nb_oargs; 4092 4093 /* Test if the operation can be removed because all 4094 its outputs are dead. We assume that nb_oargs == 0 4095 implies side effects */ 4096 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { 4097 for (i = 0; i < nb_oargs; i++) { 4098 if (arg_temp(op->args[i])->state != TS_DEAD) { 4099 goto do_not_remove; 4100 } 4101 } 4102 goto do_remove; 4103 } 4104 goto do_not_remove; 4105 4106 do_remove: 4107 tcg_op_remove(s, op); 4108 break; 4109 4110 do_not_remove: 4111 for (i = 0; i < nb_oargs; i++) { 4112 ts = arg_temp(op->args[i]); 4113 4114 /* Remember the preference of the uses that followed. */ 4115 if (i < ARRAY_SIZE(op->output_pref)) { 4116 op->output_pref[i] = *la_temp_pref(ts); 4117 } 4118 4119 /* Output args are dead. */ 4120 if (ts->state & TS_DEAD) { 4121 arg_life |= DEAD_ARG << i; 4122 } 4123 if (ts->state & TS_MEM) { 4124 arg_life |= SYNC_ARG << i; 4125 } 4126 ts->state = TS_DEAD; 4127 la_reset_pref(ts); 4128 } 4129 4130 /* If end of basic block, update. */ 4131 if (def->flags & TCG_OPF_BB_EXIT) { 4132 la_func_end(s, nb_globals, nb_temps); 4133 } else if (def->flags & TCG_OPF_COND_BRANCH) { 4134 la_bb_sync(s, nb_globals, nb_temps); 4135 } else if (def->flags & TCG_OPF_BB_END) { 4136 la_bb_end(s, nb_globals, nb_temps); 4137 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 4138 la_global_sync(s, nb_globals); 4139 if (def->flags & TCG_OPF_CALL_CLOBBER) { 4140 la_cross_call(s, nb_temps); 4141 } 4142 } 4143 4144 /* Record arguments that die in this opcode. */ 4145 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 4146 ts = arg_temp(op->args[i]); 4147 if (ts->state & TS_DEAD) { 4148 arg_life |= DEAD_ARG << i; 4149 } 4150 } 4151 4152 /* Input arguments are live for preceding opcodes. */ 4153 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 4154 ts = arg_temp(op->args[i]); 4155 if (ts->state & TS_DEAD) { 4156 /* For operands that were dead, initially allow 4157 all regs for the type. */ 4158 *la_temp_pref(ts) = tcg_target_available_regs[ts->type]; 4159 ts->state &= ~TS_DEAD; 4160 } 4161 } 4162 4163 /* Incorporate constraints for this operand. */ 4164 switch (opc) { 4165 case INDEX_op_mov: 4166 /* Note that these are TCG_OPF_NOT_PRESENT and do not 4167 have proper constraints. That said, special case 4168 moves to propagate preferences backward. */ 4169 if (IS_DEAD_ARG(1)) { 4170 *la_temp_pref(arg_temp(op->args[0])) 4171 = *la_temp_pref(arg_temp(op->args[1])); 4172 } 4173 break; 4174 4175 default: 4176 args_ct = opcode_args_ct(op); 4177 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 4178 const TCGArgConstraint *ct = &args_ct[i]; 4179 TCGRegSet set, *pset; 4180 4181 ts = arg_temp(op->args[i]); 4182 pset = la_temp_pref(ts); 4183 set = *pset; 4184 4185 set &= ct->regs; 4186 if (ct->ialias) { 4187 set &= output_pref(op, ct->alias_index); 4188 } 4189 /* If the combination is not possible, restart. */ 4190 if (set == 0) { 4191 set = ct->regs; 4192 } 4193 *pset = set; 4194 } 4195 break; 4196 } 4197 break; 4198 } 4199 op->life = arg_life; 4200 } 4201 } 4202 4203 /* Liveness analysis: Convert indirect regs to direct temporaries. */ 4204 static bool __attribute__((noinline)) 4205 liveness_pass_2(TCGContext *s) 4206 { 4207 int nb_globals = s->nb_globals; 4208 int nb_temps, i; 4209 bool changes = false; 4210 TCGOp *op, *op_next; 4211 4212 /* Create a temporary for each indirect global. */ 4213 for (i = 0; i < nb_globals; ++i) { 4214 TCGTemp *its = &s->temps[i]; 4215 if (its->indirect_reg) { 4216 TCGTemp *dts = tcg_temp_alloc(s); 4217 dts->type = its->type; 4218 dts->base_type = its->base_type; 4219 dts->temp_subindex = its->temp_subindex; 4220 dts->kind = TEMP_EBB; 4221 its->state_ptr = dts; 4222 } else { 4223 its->state_ptr = NULL; 4224 } 4225 /* All globals begin dead. */ 4226 its->state = TS_DEAD; 4227 } 4228 for (nb_temps = s->nb_temps; i < nb_temps; ++i) { 4229 TCGTemp *its = &s->temps[i]; 4230 its->state_ptr = NULL; 4231 its->state = TS_DEAD; 4232 } 4233 4234 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 4235 TCGOpcode opc = op->opc; 4236 const TCGOpDef *def = &tcg_op_defs[opc]; 4237 TCGLifeData arg_life = op->life; 4238 int nb_iargs, nb_oargs, call_flags; 4239 TCGTemp *arg_ts, *dir_ts; 4240 4241 if (opc == INDEX_op_call) { 4242 nb_oargs = TCGOP_CALLO(op); 4243 nb_iargs = TCGOP_CALLI(op); 4244 call_flags = tcg_call_flags(op); 4245 } else { 4246 nb_iargs = def->nb_iargs; 4247 nb_oargs = def->nb_oargs; 4248 4249 /* Set flags similar to how calls require. */ 4250 if (def->flags & TCG_OPF_COND_BRANCH) { 4251 /* Like reading globals: sync_globals */ 4252 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 4253 } else if (def->flags & TCG_OPF_BB_END) { 4254 /* Like writing globals: save_globals */ 4255 call_flags = 0; 4256 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 4257 /* Like reading globals: sync_globals */ 4258 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 4259 } else { 4260 /* No effect on globals. */ 4261 call_flags = (TCG_CALL_NO_READ_GLOBALS | 4262 TCG_CALL_NO_WRITE_GLOBALS); 4263 } 4264 } 4265 4266 /* Make sure that input arguments are available. */ 4267 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 4268 arg_ts = arg_temp(op->args[i]); 4269 dir_ts = arg_ts->state_ptr; 4270 if (dir_ts && arg_ts->state == TS_DEAD) { 4271 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32 4272 ? INDEX_op_ld_i32 4273 : INDEX_op_ld_i64); 4274 TCGOp *lop = tcg_op_insert_before(s, op, lopc, 4275 arg_ts->type, 3); 4276 4277 lop->args[0] = temp_arg(dir_ts); 4278 lop->args[1] = temp_arg(arg_ts->mem_base); 4279 lop->args[2] = arg_ts->mem_offset; 4280 4281 /* Loaded, but synced with memory. */ 4282 arg_ts->state = TS_MEM; 4283 } 4284 } 4285 4286 /* Perform input replacement, and mark inputs that became dead. 4287 No action is required except keeping temp_state up to date 4288 so that we reload when needed. */ 4289 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 4290 arg_ts = arg_temp(op->args[i]); 4291 dir_ts = arg_ts->state_ptr; 4292 if (dir_ts) { 4293 op->args[i] = temp_arg(dir_ts); 4294 changes = true; 4295 if (IS_DEAD_ARG(i)) { 4296 arg_ts->state = TS_DEAD; 4297 } 4298 } 4299 } 4300 4301 /* Liveness analysis should ensure that the following are 4302 all correct, for call sites and basic block end points. */ 4303 if (call_flags & TCG_CALL_NO_READ_GLOBALS) { 4304 /* Nothing to do */ 4305 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) { 4306 for (i = 0; i < nb_globals; ++i) { 4307 /* Liveness should see that globals are synced back, 4308 that is, either TS_DEAD or TS_MEM. */ 4309 arg_ts = &s->temps[i]; 4310 tcg_debug_assert(arg_ts->state_ptr == 0 4311 || arg_ts->state != 0); 4312 } 4313 } else { 4314 for (i = 0; i < nb_globals; ++i) { 4315 /* Liveness should see that globals are saved back, 4316 that is, TS_DEAD, waiting to be reloaded. */ 4317 arg_ts = &s->temps[i]; 4318 tcg_debug_assert(arg_ts->state_ptr == 0 4319 || arg_ts->state == TS_DEAD); 4320 } 4321 } 4322 4323 /* Outputs become available. */ 4324 if (opc == INDEX_op_mov) { 4325 arg_ts = arg_temp(op->args[0]); 4326 dir_ts = arg_ts->state_ptr; 4327 if (dir_ts) { 4328 op->args[0] = temp_arg(dir_ts); 4329 changes = true; 4330 4331 /* The output is now live and modified. */ 4332 arg_ts->state = 0; 4333 4334 if (NEED_SYNC_ARG(0)) { 4335 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 4336 ? INDEX_op_st_i32 4337 : INDEX_op_st_i64); 4338 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 4339 arg_ts->type, 3); 4340 TCGTemp *out_ts = dir_ts; 4341 4342 if (IS_DEAD_ARG(0)) { 4343 out_ts = arg_temp(op->args[1]); 4344 arg_ts->state = TS_DEAD; 4345 tcg_op_remove(s, op); 4346 } else { 4347 arg_ts->state = TS_MEM; 4348 } 4349 4350 sop->args[0] = temp_arg(out_ts); 4351 sop->args[1] = temp_arg(arg_ts->mem_base); 4352 sop->args[2] = arg_ts->mem_offset; 4353 } else { 4354 tcg_debug_assert(!IS_DEAD_ARG(0)); 4355 } 4356 } 4357 } else { 4358 for (i = 0; i < nb_oargs; i++) { 4359 arg_ts = arg_temp(op->args[i]); 4360 dir_ts = arg_ts->state_ptr; 4361 if (!dir_ts) { 4362 continue; 4363 } 4364 op->args[i] = temp_arg(dir_ts); 4365 changes = true; 4366 4367 /* The output is now live and modified. */ 4368 arg_ts->state = 0; 4369 4370 /* Sync outputs upon their last write. */ 4371 if (NEED_SYNC_ARG(i)) { 4372 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 4373 ? INDEX_op_st_i32 4374 : INDEX_op_st_i64); 4375 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 4376 arg_ts->type, 3); 4377 4378 sop->args[0] = temp_arg(dir_ts); 4379 sop->args[1] = temp_arg(arg_ts->mem_base); 4380 sop->args[2] = arg_ts->mem_offset; 4381 4382 arg_ts->state = TS_MEM; 4383 } 4384 /* Drop outputs that are dead. */ 4385 if (IS_DEAD_ARG(i)) { 4386 arg_ts->state = TS_DEAD; 4387 } 4388 } 4389 } 4390 } 4391 4392 return changes; 4393 } 4394 4395 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts) 4396 { 4397 intptr_t off; 4398 int size, align; 4399 4400 /* When allocating an object, look at the full type. */ 4401 size = tcg_type_size(ts->base_type); 4402 switch (ts->base_type) { 4403 case TCG_TYPE_I32: 4404 align = 4; 4405 break; 4406 case TCG_TYPE_I64: 4407 case TCG_TYPE_V64: 4408 align = 8; 4409 break; 4410 case TCG_TYPE_I128: 4411 case TCG_TYPE_V128: 4412 case TCG_TYPE_V256: 4413 /* 4414 * Note that we do not require aligned storage for V256, 4415 * and that we provide alignment for I128 to match V128, 4416 * even if that's above what the host ABI requires. 4417 */ 4418 align = 16; 4419 break; 4420 default: 4421 g_assert_not_reached(); 4422 } 4423 4424 /* 4425 * Assume the stack is sufficiently aligned. 4426 * This affects e.g. ARM NEON, where we have 8 byte stack alignment 4427 * and do not require 16 byte vector alignment. This seems slightly 4428 * easier than fully parameterizing the above switch statement. 4429 */ 4430 align = MIN(TCG_TARGET_STACK_ALIGN, align); 4431 off = ROUND_UP(s->current_frame_offset, align); 4432 4433 /* If we've exhausted the stack frame, restart with a smaller TB. */ 4434 if (off + size > s->frame_end) { 4435 tcg_raise_tb_overflow(s); 4436 } 4437 s->current_frame_offset = off + size; 4438 #if defined(__sparc__) 4439 off += TCG_TARGET_STACK_BIAS; 4440 #endif 4441 4442 /* If the object was subdivided, assign memory to all the parts. */ 4443 if (ts->base_type != ts->type) { 4444 int part_size = tcg_type_size(ts->type); 4445 int part_count = size / part_size; 4446 4447 /* 4448 * Each part is allocated sequentially in tcg_temp_new_internal. 4449 * Jump back to the first part by subtracting the current index. 4450 */ 4451 ts -= ts->temp_subindex; 4452 for (int i = 0; i < part_count; ++i) { 4453 ts[i].mem_offset = off + i * part_size; 4454 ts[i].mem_base = s->frame_temp; 4455 ts[i].mem_allocated = 1; 4456 } 4457 } else { 4458 ts->mem_offset = off; 4459 ts->mem_base = s->frame_temp; 4460 ts->mem_allocated = 1; 4461 } 4462 } 4463 4464 /* Assign @reg to @ts, and update reg_to_temp[]. */ 4465 static void set_temp_val_reg(TCGContext *s, TCGTemp *ts, TCGReg reg) 4466 { 4467 if (ts->val_type == TEMP_VAL_REG) { 4468 TCGReg old = ts->reg; 4469 tcg_debug_assert(s->reg_to_temp[old] == ts); 4470 if (old == reg) { 4471 return; 4472 } 4473 s->reg_to_temp[old] = NULL; 4474 } 4475 tcg_debug_assert(s->reg_to_temp[reg] == NULL); 4476 s->reg_to_temp[reg] = ts; 4477 ts->val_type = TEMP_VAL_REG; 4478 ts->reg = reg; 4479 } 4480 4481 /* Assign a non-register value type to @ts, and update reg_to_temp[]. */ 4482 static void set_temp_val_nonreg(TCGContext *s, TCGTemp *ts, TCGTempVal type) 4483 { 4484 tcg_debug_assert(type != TEMP_VAL_REG); 4485 if (ts->val_type == TEMP_VAL_REG) { 4486 TCGReg reg = ts->reg; 4487 tcg_debug_assert(s->reg_to_temp[reg] == ts); 4488 s->reg_to_temp[reg] = NULL; 4489 } 4490 ts->val_type = type; 4491 } 4492 4493 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet); 4494 4495 /* Mark a temporary as free or dead. If 'free_or_dead' is negative, 4496 mark it free; otherwise mark it dead. */ 4497 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead) 4498 { 4499 TCGTempVal new_type; 4500 4501 switch (ts->kind) { 4502 case TEMP_FIXED: 4503 return; 4504 case TEMP_GLOBAL: 4505 case TEMP_TB: 4506 new_type = TEMP_VAL_MEM; 4507 break; 4508 case TEMP_EBB: 4509 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD; 4510 break; 4511 case TEMP_CONST: 4512 new_type = TEMP_VAL_CONST; 4513 break; 4514 default: 4515 g_assert_not_reached(); 4516 } 4517 set_temp_val_nonreg(s, ts, new_type); 4518 } 4519 4520 /* Mark a temporary as dead. */ 4521 static inline void temp_dead(TCGContext *s, TCGTemp *ts) 4522 { 4523 temp_free_or_dead(s, ts, 1); 4524 } 4525 4526 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary 4527 registers needs to be allocated to store a constant. If 'free_or_dead' 4528 is non-zero, subsequently release the temporary; if it is positive, the 4529 temp is dead; if it is negative, the temp is free. */ 4530 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs, 4531 TCGRegSet preferred_regs, int free_or_dead) 4532 { 4533 if (!temp_readonly(ts) && !ts->mem_coherent) { 4534 if (!ts->mem_allocated) { 4535 temp_allocate_frame(s, ts); 4536 } 4537 switch (ts->val_type) { 4538 case TEMP_VAL_CONST: 4539 /* If we're going to free the temp immediately, then we won't 4540 require it later in a register, so attempt to store the 4541 constant to memory directly. */ 4542 if (free_or_dead 4543 && tcg_out_sti(s, ts->type, ts->val, 4544 ts->mem_base->reg, ts->mem_offset)) { 4545 break; 4546 } 4547 temp_load(s, ts, tcg_target_available_regs[ts->type], 4548 allocated_regs, preferred_regs); 4549 /* fallthrough */ 4550 4551 case TEMP_VAL_REG: 4552 tcg_out_st(s, ts->type, ts->reg, 4553 ts->mem_base->reg, ts->mem_offset); 4554 break; 4555 4556 case TEMP_VAL_MEM: 4557 break; 4558 4559 case TEMP_VAL_DEAD: 4560 default: 4561 g_assert_not_reached(); 4562 } 4563 ts->mem_coherent = 1; 4564 } 4565 if (free_or_dead) { 4566 temp_free_or_dead(s, ts, free_or_dead); 4567 } 4568 } 4569 4570 /* free register 'reg' by spilling the corresponding temporary if necessary */ 4571 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs) 4572 { 4573 TCGTemp *ts = s->reg_to_temp[reg]; 4574 if (ts != NULL) { 4575 temp_sync(s, ts, allocated_regs, 0, -1); 4576 } 4577 } 4578 4579 /** 4580 * tcg_reg_alloc: 4581 * @required_regs: Set of registers in which we must allocate. 4582 * @allocated_regs: Set of registers which must be avoided. 4583 * @preferred_regs: Set of registers we should prefer. 4584 * @rev: True if we search the registers in "indirect" order. 4585 * 4586 * The allocated register must be in @required_regs & ~@allocated_regs, 4587 * but if we can put it in @preferred_regs we may save a move later. 4588 */ 4589 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs, 4590 TCGRegSet allocated_regs, 4591 TCGRegSet preferred_regs, bool rev) 4592 { 4593 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4594 TCGRegSet reg_ct[2]; 4595 const int *order; 4596 4597 reg_ct[1] = required_regs & ~allocated_regs; 4598 tcg_debug_assert(reg_ct[1] != 0); 4599 reg_ct[0] = reg_ct[1] & preferred_regs; 4600 4601 /* Skip the preferred_regs option if it cannot be satisfied, 4602 or if the preference made no difference. */ 4603 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4604 4605 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4606 4607 /* Try free registers, preferences first. */ 4608 for (j = f; j < 2; j++) { 4609 TCGRegSet set = reg_ct[j]; 4610 4611 if (tcg_regset_single(set)) { 4612 /* One register in the set. */ 4613 TCGReg reg = tcg_regset_first(set); 4614 if (s->reg_to_temp[reg] == NULL) { 4615 return reg; 4616 } 4617 } else { 4618 for (i = 0; i < n; i++) { 4619 TCGReg reg = order[i]; 4620 if (s->reg_to_temp[reg] == NULL && 4621 tcg_regset_test_reg(set, reg)) { 4622 return reg; 4623 } 4624 } 4625 } 4626 } 4627 4628 /* We must spill something. */ 4629 for (j = f; j < 2; j++) { 4630 TCGRegSet set = reg_ct[j]; 4631 4632 if (tcg_regset_single(set)) { 4633 /* One register in the set. */ 4634 TCGReg reg = tcg_regset_first(set); 4635 tcg_reg_free(s, reg, allocated_regs); 4636 return reg; 4637 } else { 4638 for (i = 0; i < n; i++) { 4639 TCGReg reg = order[i]; 4640 if (tcg_regset_test_reg(set, reg)) { 4641 tcg_reg_free(s, reg, allocated_regs); 4642 return reg; 4643 } 4644 } 4645 } 4646 } 4647 4648 g_assert_not_reached(); 4649 } 4650 4651 static TCGReg tcg_reg_alloc_pair(TCGContext *s, TCGRegSet required_regs, 4652 TCGRegSet allocated_regs, 4653 TCGRegSet preferred_regs, bool rev) 4654 { 4655 int i, j, k, fmin, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4656 TCGRegSet reg_ct[2]; 4657 const int *order; 4658 4659 /* Ensure that if I is not in allocated_regs, I+1 is not either. */ 4660 reg_ct[1] = required_regs & ~(allocated_regs | (allocated_regs >> 1)); 4661 tcg_debug_assert(reg_ct[1] != 0); 4662 reg_ct[0] = reg_ct[1] & preferred_regs; 4663 4664 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4665 4666 /* 4667 * Skip the preferred_regs option if it cannot be satisfied, 4668 * or if the preference made no difference. 4669 */ 4670 k = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4671 4672 /* 4673 * Minimize the number of flushes by looking for 2 free registers first, 4674 * then a single flush, then two flushes. 4675 */ 4676 for (fmin = 2; fmin >= 0; fmin--) { 4677 for (j = k; j < 2; j++) { 4678 TCGRegSet set = reg_ct[j]; 4679 4680 for (i = 0; i < n; i++) { 4681 TCGReg reg = order[i]; 4682 4683 if (tcg_regset_test_reg(set, reg)) { 4684 int f = !s->reg_to_temp[reg] + !s->reg_to_temp[reg + 1]; 4685 if (f >= fmin) { 4686 tcg_reg_free(s, reg, allocated_regs); 4687 tcg_reg_free(s, reg + 1, allocated_regs); 4688 return reg; 4689 } 4690 } 4691 } 4692 } 4693 } 4694 g_assert_not_reached(); 4695 } 4696 4697 /* Make sure the temporary is in a register. If needed, allocate the register 4698 from DESIRED while avoiding ALLOCATED. */ 4699 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs, 4700 TCGRegSet allocated_regs, TCGRegSet preferred_regs) 4701 { 4702 TCGReg reg; 4703 4704 switch (ts->val_type) { 4705 case TEMP_VAL_REG: 4706 return; 4707 case TEMP_VAL_CONST: 4708 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4709 preferred_regs, ts->indirect_base); 4710 if (ts->type <= TCG_TYPE_I64) { 4711 tcg_out_movi(s, ts->type, reg, ts->val); 4712 } else { 4713 uint64_t val = ts->val; 4714 MemOp vece = MO_64; 4715 4716 /* 4717 * Find the minimal vector element that matches the constant. 4718 * The targets will, in general, have to do this search anyway, 4719 * do this generically. 4720 */ 4721 if (val == dup_const(MO_8, val)) { 4722 vece = MO_8; 4723 } else if (val == dup_const(MO_16, val)) { 4724 vece = MO_16; 4725 } else if (val == dup_const(MO_32, val)) { 4726 vece = MO_32; 4727 } 4728 4729 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val); 4730 } 4731 ts->mem_coherent = 0; 4732 break; 4733 case TEMP_VAL_MEM: 4734 if (!ts->mem_allocated) { 4735 temp_allocate_frame(s, ts); 4736 } 4737 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4738 preferred_regs, ts->indirect_base); 4739 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset); 4740 ts->mem_coherent = 1; 4741 break; 4742 case TEMP_VAL_DEAD: 4743 default: 4744 g_assert_not_reached(); 4745 } 4746 set_temp_val_reg(s, ts, reg); 4747 } 4748 4749 /* Save a temporary to memory. 'allocated_regs' is used in case a 4750 temporary registers needs to be allocated to store a constant. */ 4751 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs) 4752 { 4753 /* The liveness analysis already ensures that globals are back 4754 in memory. Keep an tcg_debug_assert for safety. */ 4755 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts)); 4756 } 4757 4758 /* save globals to their canonical location and assume they can be 4759 modified be the following code. 'allocated_regs' is used in case a 4760 temporary registers needs to be allocated to store a constant. */ 4761 static void save_globals(TCGContext *s, TCGRegSet allocated_regs) 4762 { 4763 int i, n; 4764 4765 for (i = 0, n = s->nb_globals; i < n; i++) { 4766 temp_save(s, &s->temps[i], allocated_regs); 4767 } 4768 } 4769 4770 /* sync globals to their canonical location and assume they can be 4771 read by the following code. 'allocated_regs' is used in case a 4772 temporary registers needs to be allocated to store a constant. */ 4773 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs) 4774 { 4775 int i, n; 4776 4777 for (i = 0, n = s->nb_globals; i < n; i++) { 4778 TCGTemp *ts = &s->temps[i]; 4779 tcg_debug_assert(ts->val_type != TEMP_VAL_REG 4780 || ts->kind == TEMP_FIXED 4781 || ts->mem_coherent); 4782 } 4783 } 4784 4785 /* at the end of a basic block, we assume all temporaries are dead and 4786 all globals are stored at their canonical location. */ 4787 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) 4788 { 4789 int i; 4790 4791 for (i = s->nb_globals; i < s->nb_temps; i++) { 4792 TCGTemp *ts = &s->temps[i]; 4793 4794 switch (ts->kind) { 4795 case TEMP_TB: 4796 temp_save(s, ts, allocated_regs); 4797 break; 4798 case TEMP_EBB: 4799 /* The liveness analysis already ensures that temps are dead. 4800 Keep an tcg_debug_assert for safety. */ 4801 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD); 4802 break; 4803 case TEMP_CONST: 4804 /* Similarly, we should have freed any allocated register. */ 4805 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST); 4806 break; 4807 default: 4808 g_assert_not_reached(); 4809 } 4810 } 4811 4812 save_globals(s, allocated_regs); 4813 } 4814 4815 /* 4816 * At a conditional branch, we assume all temporaries are dead unless 4817 * explicitly live-across-conditional-branch; all globals and local 4818 * temps are synced to their location. 4819 */ 4820 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs) 4821 { 4822 sync_globals(s, allocated_regs); 4823 4824 for (int i = s->nb_globals; i < s->nb_temps; i++) { 4825 TCGTemp *ts = &s->temps[i]; 4826 /* 4827 * The liveness analysis already ensures that temps are dead. 4828 * Keep tcg_debug_asserts for safety. 4829 */ 4830 switch (ts->kind) { 4831 case TEMP_TB: 4832 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent); 4833 break; 4834 case TEMP_EBB: 4835 case TEMP_CONST: 4836 break; 4837 default: 4838 g_assert_not_reached(); 4839 } 4840 } 4841 } 4842 4843 /* 4844 * Specialized code generation for INDEX_op_mov_* with a constant. 4845 */ 4846 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots, 4847 tcg_target_ulong val, TCGLifeData arg_life, 4848 TCGRegSet preferred_regs) 4849 { 4850 /* ENV should not be modified. */ 4851 tcg_debug_assert(!temp_readonly(ots)); 4852 4853 /* The movi is not explicitly generated here. */ 4854 set_temp_val_nonreg(s, ots, TEMP_VAL_CONST); 4855 ots->val = val; 4856 ots->mem_coherent = 0; 4857 if (NEED_SYNC_ARG(0)) { 4858 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0)); 4859 } else if (IS_DEAD_ARG(0)) { 4860 temp_dead(s, ots); 4861 } 4862 } 4863 4864 /* 4865 * Specialized code generation for INDEX_op_mov_*. 4866 */ 4867 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op) 4868 { 4869 const TCGLifeData arg_life = op->life; 4870 TCGRegSet allocated_regs, preferred_regs; 4871 TCGTemp *ts, *ots; 4872 TCGType otype, itype; 4873 TCGReg oreg, ireg; 4874 4875 allocated_regs = s->reserved_regs; 4876 preferred_regs = output_pref(op, 0); 4877 ots = arg_temp(op->args[0]); 4878 ts = arg_temp(op->args[1]); 4879 4880 /* ENV should not be modified. */ 4881 tcg_debug_assert(!temp_readonly(ots)); 4882 4883 /* Note that otype != itype for no-op truncation. */ 4884 otype = ots->type; 4885 itype = ts->type; 4886 4887 if (ts->val_type == TEMP_VAL_CONST) { 4888 /* propagate constant or generate sti */ 4889 tcg_target_ulong val = ts->val; 4890 if (IS_DEAD_ARG(1)) { 4891 temp_dead(s, ts); 4892 } 4893 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs); 4894 return; 4895 } 4896 4897 /* If the source value is in memory we're going to be forced 4898 to have it in a register in order to perform the copy. Copy 4899 the SOURCE value into its own register first, that way we 4900 don't have to reload SOURCE the next time it is used. */ 4901 if (ts->val_type == TEMP_VAL_MEM) { 4902 temp_load(s, ts, tcg_target_available_regs[itype], 4903 allocated_regs, preferred_regs); 4904 } 4905 tcg_debug_assert(ts->val_type == TEMP_VAL_REG); 4906 ireg = ts->reg; 4907 4908 if (IS_DEAD_ARG(0)) { 4909 /* mov to a non-saved dead register makes no sense (even with 4910 liveness analysis disabled). */ 4911 tcg_debug_assert(NEED_SYNC_ARG(0)); 4912 if (!ots->mem_allocated) { 4913 temp_allocate_frame(s, ots); 4914 } 4915 tcg_out_st(s, otype, ireg, ots->mem_base->reg, ots->mem_offset); 4916 if (IS_DEAD_ARG(1)) { 4917 temp_dead(s, ts); 4918 } 4919 temp_dead(s, ots); 4920 return; 4921 } 4922 4923 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) { 4924 /* 4925 * The mov can be suppressed. Kill input first, so that it 4926 * is unlinked from reg_to_temp, then set the output to the 4927 * reg that we saved from the input. 4928 */ 4929 temp_dead(s, ts); 4930 oreg = ireg; 4931 } else { 4932 if (ots->val_type == TEMP_VAL_REG) { 4933 oreg = ots->reg; 4934 } else { 4935 /* Make sure to not spill the input register during allocation. */ 4936 oreg = tcg_reg_alloc(s, tcg_target_available_regs[otype], 4937 allocated_regs | ((TCGRegSet)1 << ireg), 4938 preferred_regs, ots->indirect_base); 4939 } 4940 if (!tcg_out_mov(s, otype, oreg, ireg)) { 4941 /* 4942 * Cross register class move not supported. 4943 * Store the source register into the destination slot 4944 * and leave the destination temp as TEMP_VAL_MEM. 4945 */ 4946 assert(!temp_readonly(ots)); 4947 if (!ts->mem_allocated) { 4948 temp_allocate_frame(s, ots); 4949 } 4950 tcg_out_st(s, ts->type, ireg, ots->mem_base->reg, ots->mem_offset); 4951 set_temp_val_nonreg(s, ts, TEMP_VAL_MEM); 4952 ots->mem_coherent = 1; 4953 return; 4954 } 4955 } 4956 set_temp_val_reg(s, ots, oreg); 4957 ots->mem_coherent = 0; 4958 4959 if (NEED_SYNC_ARG(0)) { 4960 temp_sync(s, ots, allocated_regs, 0, 0); 4961 } 4962 } 4963 4964 /* 4965 * Specialized code generation for INDEX_op_dup_vec. 4966 */ 4967 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op) 4968 { 4969 const TCGLifeData arg_life = op->life; 4970 TCGRegSet dup_out_regs, dup_in_regs; 4971 const TCGArgConstraint *dup_args_ct; 4972 TCGTemp *its, *ots; 4973 TCGType itype, vtype; 4974 unsigned vece; 4975 int lowpart_ofs; 4976 bool ok; 4977 4978 ots = arg_temp(op->args[0]); 4979 its = arg_temp(op->args[1]); 4980 4981 /* ENV should not be modified. */ 4982 tcg_debug_assert(!temp_readonly(ots)); 4983 4984 itype = its->type; 4985 vece = TCGOP_VECE(op); 4986 vtype = TCGOP_TYPE(op); 4987 4988 if (its->val_type == TEMP_VAL_CONST) { 4989 /* Propagate constant via movi -> dupi. */ 4990 tcg_target_ulong val = its->val; 4991 if (IS_DEAD_ARG(1)) { 4992 temp_dead(s, its); 4993 } 4994 tcg_reg_alloc_do_movi(s, ots, val, arg_life, output_pref(op, 0)); 4995 return; 4996 } 4997 4998 dup_args_ct = opcode_args_ct(op); 4999 dup_out_regs = dup_args_ct[0].regs; 5000 dup_in_regs = dup_args_ct[1].regs; 5001 5002 /* Allocate the output register now. */ 5003 if (ots->val_type != TEMP_VAL_REG) { 5004 TCGRegSet allocated_regs = s->reserved_regs; 5005 TCGReg oreg; 5006 5007 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) { 5008 /* Make sure to not spill the input register. */ 5009 tcg_regset_set_reg(allocated_regs, its->reg); 5010 } 5011 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 5012 output_pref(op, 0), ots->indirect_base); 5013 set_temp_val_reg(s, ots, oreg); 5014 } 5015 5016 switch (its->val_type) { 5017 case TEMP_VAL_REG: 5018 /* 5019 * The dup constriaints must be broad, covering all possible VECE. 5020 * However, tcg_op_dup_vec() gets to see the VECE and we allow it 5021 * to fail, indicating that extra moves are required for that case. 5022 */ 5023 if (tcg_regset_test_reg(dup_in_regs, its->reg)) { 5024 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) { 5025 goto done; 5026 } 5027 /* Try again from memory or a vector input register. */ 5028 } 5029 if (!its->mem_coherent) { 5030 /* 5031 * The input register is not synced, and so an extra store 5032 * would be required to use memory. Attempt an integer-vector 5033 * register move first. We do not have a TCGRegSet for this. 5034 */ 5035 if (tcg_out_mov(s, itype, ots->reg, its->reg)) { 5036 break; 5037 } 5038 /* Sync the temp back to its slot and load from there. */ 5039 temp_sync(s, its, s->reserved_regs, 0, 0); 5040 } 5041 /* fall through */ 5042 5043 case TEMP_VAL_MEM: 5044 lowpart_ofs = 0; 5045 if (HOST_BIG_ENDIAN) { 5046 lowpart_ofs = tcg_type_size(itype) - (1 << vece); 5047 } 5048 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg, 5049 its->mem_offset + lowpart_ofs)) { 5050 goto done; 5051 } 5052 /* Load the input into the destination vector register. */ 5053 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset); 5054 break; 5055 5056 default: 5057 g_assert_not_reached(); 5058 } 5059 5060 /* We now have a vector input register, so dup must succeed. */ 5061 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg); 5062 tcg_debug_assert(ok); 5063 5064 done: 5065 ots->mem_coherent = 0; 5066 if (IS_DEAD_ARG(1)) { 5067 temp_dead(s, its); 5068 } 5069 if (NEED_SYNC_ARG(0)) { 5070 temp_sync(s, ots, s->reserved_regs, 0, 0); 5071 } 5072 if (IS_DEAD_ARG(0)) { 5073 temp_dead(s, ots); 5074 } 5075 } 5076 5077 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op) 5078 { 5079 const TCGLifeData arg_life = op->life; 5080 const TCGOpDef * const def = &tcg_op_defs[op->opc]; 5081 TCGRegSet i_allocated_regs; 5082 TCGRegSet o_allocated_regs; 5083 int i, k, nb_iargs, nb_oargs; 5084 TCGReg reg; 5085 TCGArg arg; 5086 const TCGArgConstraint *args_ct; 5087 const TCGArgConstraint *arg_ct; 5088 TCGTemp *ts; 5089 TCGArg new_args[TCG_MAX_OP_ARGS]; 5090 int const_args[TCG_MAX_OP_ARGS]; 5091 TCGCond op_cond; 5092 5093 nb_oargs = def->nb_oargs; 5094 nb_iargs = def->nb_iargs; 5095 5096 /* copy constants */ 5097 memcpy(new_args + nb_oargs + nb_iargs, 5098 op->args + nb_oargs + nb_iargs, 5099 sizeof(TCGArg) * def->nb_cargs); 5100 5101 i_allocated_regs = s->reserved_regs; 5102 o_allocated_regs = s->reserved_regs; 5103 5104 switch (op->opc) { 5105 case INDEX_op_brcond_i32: 5106 case INDEX_op_brcond_i64: 5107 op_cond = op->args[2]; 5108 break; 5109 case INDEX_op_setcond_i32: 5110 case INDEX_op_setcond_i64: 5111 case INDEX_op_negsetcond_i32: 5112 case INDEX_op_negsetcond_i64: 5113 case INDEX_op_cmp_vec: 5114 op_cond = op->args[3]; 5115 break; 5116 case INDEX_op_brcond2_i32: 5117 op_cond = op->args[4]; 5118 break; 5119 case INDEX_op_movcond_i32: 5120 case INDEX_op_movcond_i64: 5121 case INDEX_op_setcond2_i32: 5122 case INDEX_op_cmpsel_vec: 5123 op_cond = op->args[5]; 5124 break; 5125 default: 5126 /* No condition within opcode. */ 5127 op_cond = TCG_COND_ALWAYS; 5128 break; 5129 } 5130 5131 args_ct = opcode_args_ct(op); 5132 5133 /* satisfy input constraints */ 5134 for (k = 0; k < nb_iargs; k++) { 5135 TCGRegSet i_preferred_regs, i_required_regs; 5136 bool allocate_new_reg, copyto_new_reg; 5137 TCGTemp *ts2; 5138 int i1, i2; 5139 5140 i = args_ct[nb_oargs + k].sort_index; 5141 arg = op->args[i]; 5142 arg_ct = &args_ct[i]; 5143 ts = arg_temp(arg); 5144 5145 if (ts->val_type == TEMP_VAL_CONST) { 5146 #ifdef TCG_REG_ZERO 5147 if (ts->val == 0 && (arg_ct->ct & TCG_CT_REG_ZERO)) { 5148 /* Hardware zero register: indicate register via non-const. */ 5149 const_args[i] = 0; 5150 new_args[i] = TCG_REG_ZERO; 5151 continue; 5152 } 5153 #endif 5154 5155 if (tcg_target_const_match(ts->val, arg_ct->ct, ts->type, 5156 op_cond, TCGOP_VECE(op))) { 5157 /* constant is OK for instruction */ 5158 const_args[i] = 1; 5159 new_args[i] = ts->val; 5160 continue; 5161 } 5162 } 5163 5164 reg = ts->reg; 5165 i_preferred_regs = 0; 5166 i_required_regs = arg_ct->regs; 5167 allocate_new_reg = false; 5168 copyto_new_reg = false; 5169 5170 switch (arg_ct->pair) { 5171 case 0: /* not paired */ 5172 if (arg_ct->ialias) { 5173 i_preferred_regs = output_pref(op, arg_ct->alias_index); 5174 5175 /* 5176 * If the input is readonly, then it cannot also be an 5177 * output and aliased to itself. If the input is not 5178 * dead after the instruction, we must allocate a new 5179 * register and move it. 5180 */ 5181 if (temp_readonly(ts) || !IS_DEAD_ARG(i) 5182 || args_ct[arg_ct->alias_index].newreg) { 5183 allocate_new_reg = true; 5184 } else if (ts->val_type == TEMP_VAL_REG) { 5185 /* 5186 * Check if the current register has already been 5187 * allocated for another input. 5188 */ 5189 allocate_new_reg = 5190 tcg_regset_test_reg(i_allocated_regs, reg); 5191 } 5192 } 5193 if (!allocate_new_reg) { 5194 temp_load(s, ts, i_required_regs, i_allocated_regs, 5195 i_preferred_regs); 5196 reg = ts->reg; 5197 allocate_new_reg = !tcg_regset_test_reg(i_required_regs, reg); 5198 } 5199 if (allocate_new_reg) { 5200 /* 5201 * Allocate a new register matching the constraint 5202 * and move the temporary register into it. 5203 */ 5204 temp_load(s, ts, tcg_target_available_regs[ts->type], 5205 i_allocated_regs, 0); 5206 reg = tcg_reg_alloc(s, i_required_regs, i_allocated_regs, 5207 i_preferred_regs, ts->indirect_base); 5208 copyto_new_reg = true; 5209 } 5210 break; 5211 5212 case 1: 5213 /* First of an input pair; if i1 == i2, the second is an output. */ 5214 i1 = i; 5215 i2 = arg_ct->pair_index; 5216 ts2 = i1 != i2 ? arg_temp(op->args[i2]) : NULL; 5217 5218 /* 5219 * It is easier to default to allocating a new pair 5220 * and to identify a few cases where it's not required. 5221 */ 5222 if (arg_ct->ialias) { 5223 i_preferred_regs = output_pref(op, arg_ct->alias_index); 5224 if (IS_DEAD_ARG(i1) && 5225 IS_DEAD_ARG(i2) && 5226 !temp_readonly(ts) && 5227 ts->val_type == TEMP_VAL_REG && 5228 ts->reg < TCG_TARGET_NB_REGS - 1 && 5229 tcg_regset_test_reg(i_required_regs, reg) && 5230 !tcg_regset_test_reg(i_allocated_regs, reg) && 5231 !tcg_regset_test_reg(i_allocated_regs, reg + 1) && 5232 (ts2 5233 ? ts2->val_type == TEMP_VAL_REG && 5234 ts2->reg == reg + 1 && 5235 !temp_readonly(ts2) 5236 : s->reg_to_temp[reg + 1] == NULL)) { 5237 break; 5238 } 5239 } else { 5240 /* Without aliasing, the pair must also be an input. */ 5241 tcg_debug_assert(ts2); 5242 if (ts->val_type == TEMP_VAL_REG && 5243 ts2->val_type == TEMP_VAL_REG && 5244 ts2->reg == reg + 1 && 5245 tcg_regset_test_reg(i_required_regs, reg)) { 5246 break; 5247 } 5248 } 5249 reg = tcg_reg_alloc_pair(s, i_required_regs, i_allocated_regs, 5250 0, ts->indirect_base); 5251 goto do_pair; 5252 5253 case 2: /* pair second */ 5254 reg = new_args[arg_ct->pair_index] + 1; 5255 goto do_pair; 5256 5257 case 3: /* ialias with second output, no first input */ 5258 tcg_debug_assert(arg_ct->ialias); 5259 i_preferred_regs = output_pref(op, arg_ct->alias_index); 5260 5261 if (IS_DEAD_ARG(i) && 5262 !temp_readonly(ts) && 5263 ts->val_type == TEMP_VAL_REG && 5264 reg > 0 && 5265 s->reg_to_temp[reg - 1] == NULL && 5266 tcg_regset_test_reg(i_required_regs, reg) && 5267 !tcg_regset_test_reg(i_allocated_regs, reg) && 5268 !tcg_regset_test_reg(i_allocated_regs, reg - 1)) { 5269 tcg_regset_set_reg(i_allocated_regs, reg - 1); 5270 break; 5271 } 5272 reg = tcg_reg_alloc_pair(s, i_required_regs >> 1, 5273 i_allocated_regs, 0, 5274 ts->indirect_base); 5275 tcg_regset_set_reg(i_allocated_regs, reg); 5276 reg += 1; 5277 goto do_pair; 5278 5279 do_pair: 5280 /* 5281 * If an aliased input is not dead after the instruction, 5282 * we must allocate a new register and move it. 5283 */ 5284 if (arg_ct->ialias && (!IS_DEAD_ARG(i) || temp_readonly(ts))) { 5285 TCGRegSet t_allocated_regs = i_allocated_regs; 5286 5287 /* 5288 * Because of the alias, and the continued life, make sure 5289 * that the temp is somewhere *other* than the reg pair, 5290 * and we get a copy in reg. 5291 */ 5292 tcg_regset_set_reg(t_allocated_regs, reg); 5293 tcg_regset_set_reg(t_allocated_regs, reg + 1); 5294 if (ts->val_type == TEMP_VAL_REG && ts->reg == reg) { 5295 /* If ts was already in reg, copy it somewhere else. */ 5296 TCGReg nr; 5297 bool ok; 5298 5299 tcg_debug_assert(ts->kind != TEMP_FIXED); 5300 nr = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], 5301 t_allocated_regs, 0, ts->indirect_base); 5302 ok = tcg_out_mov(s, ts->type, nr, reg); 5303 tcg_debug_assert(ok); 5304 5305 set_temp_val_reg(s, ts, nr); 5306 } else { 5307 temp_load(s, ts, tcg_target_available_regs[ts->type], 5308 t_allocated_regs, 0); 5309 copyto_new_reg = true; 5310 } 5311 } else { 5312 /* Preferably allocate to reg, otherwise copy. */ 5313 i_required_regs = (TCGRegSet)1 << reg; 5314 temp_load(s, ts, i_required_regs, i_allocated_regs, 5315 i_preferred_regs); 5316 copyto_new_reg = ts->reg != reg; 5317 } 5318 break; 5319 5320 default: 5321 g_assert_not_reached(); 5322 } 5323 5324 if (copyto_new_reg) { 5325 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 5326 /* 5327 * Cross register class move not supported. Sync the 5328 * temp back to its slot and load from there. 5329 */ 5330 temp_sync(s, ts, i_allocated_regs, 0, 0); 5331 tcg_out_ld(s, ts->type, reg, 5332 ts->mem_base->reg, ts->mem_offset); 5333 } 5334 } 5335 new_args[i] = reg; 5336 const_args[i] = 0; 5337 tcg_regset_set_reg(i_allocated_regs, reg); 5338 } 5339 5340 /* mark dead temporaries and free the associated registers */ 5341 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 5342 if (IS_DEAD_ARG(i)) { 5343 temp_dead(s, arg_temp(op->args[i])); 5344 } 5345 } 5346 5347 if (def->flags & TCG_OPF_COND_BRANCH) { 5348 tcg_reg_alloc_cbranch(s, i_allocated_regs); 5349 } else if (def->flags & TCG_OPF_BB_END) { 5350 tcg_reg_alloc_bb_end(s, i_allocated_regs); 5351 } else { 5352 if (def->flags & TCG_OPF_CALL_CLOBBER) { 5353 /* XXX: permit generic clobber register list ? */ 5354 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 5355 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 5356 tcg_reg_free(s, i, i_allocated_regs); 5357 } 5358 } 5359 } 5360 if (def->flags & TCG_OPF_SIDE_EFFECTS) { 5361 /* sync globals if the op has side effects and might trigger 5362 an exception. */ 5363 sync_globals(s, i_allocated_regs); 5364 } 5365 5366 /* satisfy the output constraints */ 5367 for (k = 0; k < nb_oargs; k++) { 5368 i = args_ct[k].sort_index; 5369 arg = op->args[i]; 5370 arg_ct = &args_ct[i]; 5371 ts = arg_temp(arg); 5372 5373 /* ENV should not be modified. */ 5374 tcg_debug_assert(!temp_readonly(ts)); 5375 5376 switch (arg_ct->pair) { 5377 case 0: /* not paired */ 5378 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) { 5379 reg = new_args[arg_ct->alias_index]; 5380 } else if (arg_ct->newreg) { 5381 reg = tcg_reg_alloc(s, arg_ct->regs, 5382 i_allocated_regs | o_allocated_regs, 5383 output_pref(op, k), ts->indirect_base); 5384 } else { 5385 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs, 5386 output_pref(op, k), ts->indirect_base); 5387 } 5388 break; 5389 5390 case 1: /* first of pair */ 5391 if (arg_ct->oalias) { 5392 reg = new_args[arg_ct->alias_index]; 5393 } else if (arg_ct->newreg) { 5394 reg = tcg_reg_alloc_pair(s, arg_ct->regs, 5395 i_allocated_regs | o_allocated_regs, 5396 output_pref(op, k), 5397 ts->indirect_base); 5398 } else { 5399 reg = tcg_reg_alloc_pair(s, arg_ct->regs, o_allocated_regs, 5400 output_pref(op, k), 5401 ts->indirect_base); 5402 } 5403 break; 5404 5405 case 2: /* second of pair */ 5406 if (arg_ct->oalias) { 5407 reg = new_args[arg_ct->alias_index]; 5408 } else { 5409 reg = new_args[arg_ct->pair_index] + 1; 5410 } 5411 break; 5412 5413 case 3: /* first of pair, aliasing with a second input */ 5414 tcg_debug_assert(!arg_ct->newreg); 5415 reg = new_args[arg_ct->pair_index] - 1; 5416 break; 5417 5418 default: 5419 g_assert_not_reached(); 5420 } 5421 tcg_regset_set_reg(o_allocated_regs, reg); 5422 set_temp_val_reg(s, ts, reg); 5423 ts->mem_coherent = 0; 5424 new_args[i] = reg; 5425 } 5426 } 5427 5428 /* emit instruction */ 5429 TCGType type = TCGOP_TYPE(op); 5430 switch (op->opc) { 5431 case INDEX_op_ext_i32_i64: 5432 tcg_out_exts_i32_i64(s, new_args[0], new_args[1]); 5433 break; 5434 case INDEX_op_extu_i32_i64: 5435 tcg_out_extu_i32_i64(s, new_args[0], new_args[1]); 5436 break; 5437 case INDEX_op_extrl_i64_i32: 5438 tcg_out_extrl_i64_i32(s, new_args[0], new_args[1]); 5439 break; 5440 5441 case INDEX_op_add: 5442 case INDEX_op_and: 5443 case INDEX_op_andc: 5444 { 5445 const TCGOutOpBinary *out = 5446 container_of(all_outop[op->opc], TCGOutOpBinary, base); 5447 5448 /* Constants should never appear in the first source operand. */ 5449 tcg_debug_assert(!const_args[1]); 5450 if (const_args[2]) { 5451 out->out_rri(s, type, new_args[0], new_args[1], new_args[2]); 5452 } else { 5453 out->out_rrr(s, type, new_args[0], new_args[1], new_args[2]); 5454 } 5455 } 5456 break; 5457 5458 default: 5459 if (def->flags & TCG_OPF_VECTOR) { 5460 tcg_out_vec_op(s, op->opc, type - TCG_TYPE_V64, 5461 TCGOP_VECE(op), new_args, const_args); 5462 } else { 5463 tcg_out_op(s, op->opc, type, new_args, const_args); 5464 } 5465 break; 5466 } 5467 5468 /* move the outputs in the correct register if needed */ 5469 for(i = 0; i < nb_oargs; i++) { 5470 ts = arg_temp(op->args[i]); 5471 5472 /* ENV should not be modified. */ 5473 tcg_debug_assert(!temp_readonly(ts)); 5474 5475 if (NEED_SYNC_ARG(i)) { 5476 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i)); 5477 } else if (IS_DEAD_ARG(i)) { 5478 temp_dead(s, ts); 5479 } 5480 } 5481 } 5482 5483 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op) 5484 { 5485 const TCGLifeData arg_life = op->life; 5486 TCGTemp *ots, *itsl, *itsh; 5487 TCGType vtype = TCGOP_TYPE(op); 5488 5489 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */ 5490 tcg_debug_assert(TCG_TARGET_REG_BITS == 32); 5491 tcg_debug_assert(TCGOP_VECE(op) == MO_64); 5492 5493 ots = arg_temp(op->args[0]); 5494 itsl = arg_temp(op->args[1]); 5495 itsh = arg_temp(op->args[2]); 5496 5497 /* ENV should not be modified. */ 5498 tcg_debug_assert(!temp_readonly(ots)); 5499 5500 /* Allocate the output register now. */ 5501 if (ots->val_type != TEMP_VAL_REG) { 5502 TCGRegSet allocated_regs = s->reserved_regs; 5503 TCGRegSet dup_out_regs = opcode_args_ct(op)[0].regs; 5504 TCGReg oreg; 5505 5506 /* Make sure to not spill the input registers. */ 5507 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) { 5508 tcg_regset_set_reg(allocated_regs, itsl->reg); 5509 } 5510 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) { 5511 tcg_regset_set_reg(allocated_regs, itsh->reg); 5512 } 5513 5514 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 5515 output_pref(op, 0), ots->indirect_base); 5516 set_temp_val_reg(s, ots, oreg); 5517 } 5518 5519 /* Promote dup2 of immediates to dupi_vec. */ 5520 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) { 5521 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val); 5522 MemOp vece = MO_64; 5523 5524 if (val == dup_const(MO_8, val)) { 5525 vece = MO_8; 5526 } else if (val == dup_const(MO_16, val)) { 5527 vece = MO_16; 5528 } else if (val == dup_const(MO_32, val)) { 5529 vece = MO_32; 5530 } 5531 5532 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val); 5533 goto done; 5534 } 5535 5536 /* If the two inputs form one 64-bit value, try dupm_vec. */ 5537 if (itsl->temp_subindex == HOST_BIG_ENDIAN && 5538 itsh->temp_subindex == !HOST_BIG_ENDIAN && 5539 itsl == itsh + (HOST_BIG_ENDIAN ? 1 : -1)) { 5540 TCGTemp *its = itsl - HOST_BIG_ENDIAN; 5541 5542 temp_sync(s, its + 0, s->reserved_regs, 0, 0); 5543 temp_sync(s, its + 1, s->reserved_regs, 0, 0); 5544 5545 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg, 5546 its->mem_base->reg, its->mem_offset)) { 5547 goto done; 5548 } 5549 } 5550 5551 /* Fall back to generic expansion. */ 5552 return false; 5553 5554 done: 5555 ots->mem_coherent = 0; 5556 if (IS_DEAD_ARG(1)) { 5557 temp_dead(s, itsl); 5558 } 5559 if (IS_DEAD_ARG(2)) { 5560 temp_dead(s, itsh); 5561 } 5562 if (NEED_SYNC_ARG(0)) { 5563 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0)); 5564 } else if (IS_DEAD_ARG(0)) { 5565 temp_dead(s, ots); 5566 } 5567 return true; 5568 } 5569 5570 static void load_arg_reg(TCGContext *s, TCGReg reg, TCGTemp *ts, 5571 TCGRegSet allocated_regs) 5572 { 5573 if (ts->val_type == TEMP_VAL_REG) { 5574 if (ts->reg != reg) { 5575 tcg_reg_free(s, reg, allocated_regs); 5576 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 5577 /* 5578 * Cross register class move not supported. Sync the 5579 * temp back to its slot and load from there. 5580 */ 5581 temp_sync(s, ts, allocated_regs, 0, 0); 5582 tcg_out_ld(s, ts->type, reg, 5583 ts->mem_base->reg, ts->mem_offset); 5584 } 5585 } 5586 } else { 5587 TCGRegSet arg_set = 0; 5588 5589 tcg_reg_free(s, reg, allocated_regs); 5590 tcg_regset_set_reg(arg_set, reg); 5591 temp_load(s, ts, arg_set, allocated_regs, 0); 5592 } 5593 } 5594 5595 static void load_arg_stk(TCGContext *s, unsigned arg_slot, TCGTemp *ts, 5596 TCGRegSet allocated_regs) 5597 { 5598 /* 5599 * When the destination is on the stack, load up the temp and store. 5600 * If there are many call-saved registers, the temp might live to 5601 * see another use; otherwise it'll be discarded. 5602 */ 5603 temp_load(s, ts, tcg_target_available_regs[ts->type], allocated_regs, 0); 5604 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, 5605 arg_slot_stk_ofs(arg_slot)); 5606 } 5607 5608 static void load_arg_normal(TCGContext *s, const TCGCallArgumentLoc *l, 5609 TCGTemp *ts, TCGRegSet *allocated_regs) 5610 { 5611 if (arg_slot_reg_p(l->arg_slot)) { 5612 TCGReg reg = tcg_target_call_iarg_regs[l->arg_slot]; 5613 load_arg_reg(s, reg, ts, *allocated_regs); 5614 tcg_regset_set_reg(*allocated_regs, reg); 5615 } else { 5616 load_arg_stk(s, l->arg_slot, ts, *allocated_regs); 5617 } 5618 } 5619 5620 static void load_arg_ref(TCGContext *s, unsigned arg_slot, TCGReg ref_base, 5621 intptr_t ref_off, TCGRegSet *allocated_regs) 5622 { 5623 TCGReg reg; 5624 5625 if (arg_slot_reg_p(arg_slot)) { 5626 reg = tcg_target_call_iarg_regs[arg_slot]; 5627 tcg_reg_free(s, reg, *allocated_regs); 5628 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5629 tcg_regset_set_reg(*allocated_regs, reg); 5630 } else { 5631 reg = tcg_reg_alloc(s, tcg_target_available_regs[TCG_TYPE_PTR], 5632 *allocated_regs, 0, false); 5633 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5634 tcg_out_st(s, TCG_TYPE_PTR, reg, TCG_REG_CALL_STACK, 5635 arg_slot_stk_ofs(arg_slot)); 5636 } 5637 } 5638 5639 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op) 5640 { 5641 const int nb_oargs = TCGOP_CALLO(op); 5642 const int nb_iargs = TCGOP_CALLI(op); 5643 const TCGLifeData arg_life = op->life; 5644 const TCGHelperInfo *info = tcg_call_info(op); 5645 TCGRegSet allocated_regs = s->reserved_regs; 5646 int i; 5647 5648 /* 5649 * Move inputs into place in reverse order, 5650 * so that we place stacked arguments first. 5651 */ 5652 for (i = nb_iargs - 1; i >= 0; --i) { 5653 const TCGCallArgumentLoc *loc = &info->in[i]; 5654 TCGTemp *ts = arg_temp(op->args[nb_oargs + i]); 5655 5656 switch (loc->kind) { 5657 case TCG_CALL_ARG_NORMAL: 5658 case TCG_CALL_ARG_EXTEND_U: 5659 case TCG_CALL_ARG_EXTEND_S: 5660 load_arg_normal(s, loc, ts, &allocated_regs); 5661 break; 5662 case TCG_CALL_ARG_BY_REF: 5663 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5664 load_arg_ref(s, loc->arg_slot, TCG_REG_CALL_STACK, 5665 arg_slot_stk_ofs(loc->ref_slot), 5666 &allocated_regs); 5667 break; 5668 case TCG_CALL_ARG_BY_REF_N: 5669 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5670 break; 5671 default: 5672 g_assert_not_reached(); 5673 } 5674 } 5675 5676 /* Mark dead temporaries and free the associated registers. */ 5677 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 5678 if (IS_DEAD_ARG(i)) { 5679 temp_dead(s, arg_temp(op->args[i])); 5680 } 5681 } 5682 5683 /* Clobber call registers. */ 5684 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 5685 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 5686 tcg_reg_free(s, i, allocated_regs); 5687 } 5688 } 5689 5690 /* 5691 * Save globals if they might be written by the helper, 5692 * sync them if they might be read. 5693 */ 5694 if (info->flags & TCG_CALL_NO_READ_GLOBALS) { 5695 /* Nothing to do */ 5696 } else if (info->flags & TCG_CALL_NO_WRITE_GLOBALS) { 5697 sync_globals(s, allocated_regs); 5698 } else { 5699 save_globals(s, allocated_regs); 5700 } 5701 5702 /* 5703 * If the ABI passes a pointer to the returned struct as the first 5704 * argument, load that now. Pass a pointer to the output home slot. 5705 */ 5706 if (info->out_kind == TCG_CALL_RET_BY_REF) { 5707 TCGTemp *ts = arg_temp(op->args[0]); 5708 5709 if (!ts->mem_allocated) { 5710 temp_allocate_frame(s, ts); 5711 } 5712 load_arg_ref(s, 0, ts->mem_base->reg, ts->mem_offset, &allocated_regs); 5713 } 5714 5715 tcg_out_call(s, tcg_call_func(op), info); 5716 5717 /* Assign output registers and emit moves if needed. */ 5718 switch (info->out_kind) { 5719 case TCG_CALL_RET_NORMAL: 5720 for (i = 0; i < nb_oargs; i++) { 5721 TCGTemp *ts = arg_temp(op->args[i]); 5722 TCGReg reg = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, i); 5723 5724 /* ENV should not be modified. */ 5725 tcg_debug_assert(!temp_readonly(ts)); 5726 5727 set_temp_val_reg(s, ts, reg); 5728 ts->mem_coherent = 0; 5729 } 5730 break; 5731 5732 case TCG_CALL_RET_BY_VEC: 5733 { 5734 TCGTemp *ts = arg_temp(op->args[0]); 5735 5736 tcg_debug_assert(ts->base_type == TCG_TYPE_I128); 5737 tcg_debug_assert(ts->temp_subindex == 0); 5738 if (!ts->mem_allocated) { 5739 temp_allocate_frame(s, ts); 5740 } 5741 tcg_out_st(s, TCG_TYPE_V128, 5742 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5743 ts->mem_base->reg, ts->mem_offset); 5744 } 5745 /* fall through to mark all parts in memory */ 5746 5747 case TCG_CALL_RET_BY_REF: 5748 /* The callee has performed a write through the reference. */ 5749 for (i = 0; i < nb_oargs; i++) { 5750 TCGTemp *ts = arg_temp(op->args[i]); 5751 ts->val_type = TEMP_VAL_MEM; 5752 } 5753 break; 5754 5755 default: 5756 g_assert_not_reached(); 5757 } 5758 5759 /* Flush or discard output registers as needed. */ 5760 for (i = 0; i < nb_oargs; i++) { 5761 TCGTemp *ts = arg_temp(op->args[i]); 5762 if (NEED_SYNC_ARG(i)) { 5763 temp_sync(s, ts, s->reserved_regs, 0, IS_DEAD_ARG(i)); 5764 } else if (IS_DEAD_ARG(i)) { 5765 temp_dead(s, ts); 5766 } 5767 } 5768 } 5769 5770 /** 5771 * atom_and_align_for_opc: 5772 * @s: tcg context 5773 * @opc: memory operation code 5774 * @host_atom: MO_ATOM_{IFALIGN,WITHIN16,SUBALIGN} for host operations 5775 * @allow_two_ops: true if we are prepared to issue two operations 5776 * 5777 * Return the alignment and atomicity to use for the inline fast path 5778 * for the given memory operation. The alignment may be larger than 5779 * that specified in @opc, and the correct alignment will be diagnosed 5780 * by the slow path helper. 5781 * 5782 * If @allow_two_ops, the host is prepared to test for 2x alignment, 5783 * and issue two loads or stores for subalignment. 5784 */ 5785 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 5786 MemOp host_atom, bool allow_two_ops) 5787 { 5788 MemOp align = memop_alignment_bits(opc); 5789 MemOp size = opc & MO_SIZE; 5790 MemOp half = size ? size - 1 : 0; 5791 MemOp atom = opc & MO_ATOM_MASK; 5792 MemOp atmax; 5793 5794 switch (atom) { 5795 case MO_ATOM_NONE: 5796 /* The operation requires no specific atomicity. */ 5797 atmax = MO_8; 5798 break; 5799 5800 case MO_ATOM_IFALIGN: 5801 atmax = size; 5802 break; 5803 5804 case MO_ATOM_IFALIGN_PAIR: 5805 atmax = half; 5806 break; 5807 5808 case MO_ATOM_WITHIN16: 5809 atmax = size; 5810 if (size == MO_128) { 5811 /* Misalignment implies !within16, and therefore no atomicity. */ 5812 } else if (host_atom != MO_ATOM_WITHIN16) { 5813 /* The host does not implement within16, so require alignment. */ 5814 align = MAX(align, size); 5815 } 5816 break; 5817 5818 case MO_ATOM_WITHIN16_PAIR: 5819 atmax = size; 5820 /* 5821 * Misalignment implies !within16, and therefore half atomicity. 5822 * Any host prepared for two operations can implement this with 5823 * half alignment. 5824 */ 5825 if (host_atom != MO_ATOM_WITHIN16 && allow_two_ops) { 5826 align = MAX(align, half); 5827 } 5828 break; 5829 5830 case MO_ATOM_SUBALIGN: 5831 atmax = size; 5832 if (host_atom != MO_ATOM_SUBALIGN) { 5833 /* If unaligned but not odd, there are subobjects up to half. */ 5834 if (allow_two_ops) { 5835 align = MAX(align, half); 5836 } else { 5837 align = MAX(align, size); 5838 } 5839 } 5840 break; 5841 5842 default: 5843 g_assert_not_reached(); 5844 } 5845 5846 return (TCGAtomAlign){ .atom = atmax, .align = align }; 5847 } 5848 5849 /* 5850 * Similarly for qemu_ld/st slow path helpers. 5851 * We must re-implement tcg_gen_callN and tcg_reg_alloc_call simultaneously, 5852 * using only the provided backend tcg_out_* functions. 5853 */ 5854 5855 static int tcg_out_helper_stk_ofs(TCGType type, unsigned slot) 5856 { 5857 int ofs = arg_slot_stk_ofs(slot); 5858 5859 /* 5860 * Each stack slot is TCG_TARGET_LONG_BITS. If the host does not 5861 * require extension to uint64_t, adjust the address for uint32_t. 5862 */ 5863 if (HOST_BIG_ENDIAN && 5864 TCG_TARGET_REG_BITS == 64 && 5865 type == TCG_TYPE_I32) { 5866 ofs += 4; 5867 } 5868 return ofs; 5869 } 5870 5871 static void tcg_out_helper_load_slots(TCGContext *s, 5872 unsigned nmov, TCGMovExtend *mov, 5873 const TCGLdstHelperParam *parm) 5874 { 5875 unsigned i; 5876 TCGReg dst3; 5877 5878 /* 5879 * Start from the end, storing to the stack first. 5880 * This frees those registers, so we need not consider overlap. 5881 */ 5882 for (i = nmov; i-- > 0; ) { 5883 unsigned slot = mov[i].dst; 5884 5885 if (arg_slot_reg_p(slot)) { 5886 goto found_reg; 5887 } 5888 5889 TCGReg src = mov[i].src; 5890 TCGType dst_type = mov[i].dst_type; 5891 MemOp dst_mo = dst_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5892 5893 /* The argument is going onto the stack; extend into scratch. */ 5894 if ((mov[i].src_ext & MO_SIZE) != dst_mo) { 5895 tcg_debug_assert(parm->ntmp != 0); 5896 mov[i].dst = src = parm->tmp[0]; 5897 tcg_out_movext1(s, &mov[i]); 5898 } 5899 5900 tcg_out_st(s, dst_type, src, TCG_REG_CALL_STACK, 5901 tcg_out_helper_stk_ofs(dst_type, slot)); 5902 } 5903 return; 5904 5905 found_reg: 5906 /* 5907 * The remaining arguments are in registers. 5908 * Convert slot numbers to argument registers. 5909 */ 5910 nmov = i + 1; 5911 for (i = 0; i < nmov; ++i) { 5912 mov[i].dst = tcg_target_call_iarg_regs[mov[i].dst]; 5913 } 5914 5915 switch (nmov) { 5916 case 4: 5917 /* The backend must have provided enough temps for the worst case. */ 5918 tcg_debug_assert(parm->ntmp >= 2); 5919 5920 dst3 = mov[3].dst; 5921 for (unsigned j = 0; j < 3; ++j) { 5922 if (dst3 == mov[j].src) { 5923 /* 5924 * Conflict. Copy the source to a temporary, perform the 5925 * remaining moves, then the extension from our scratch 5926 * on the way out. 5927 */ 5928 TCGReg scratch = parm->tmp[1]; 5929 5930 tcg_out_mov(s, mov[3].src_type, scratch, mov[3].src); 5931 tcg_out_movext3(s, mov, mov + 1, mov + 2, parm->tmp[0]); 5932 tcg_out_movext1_new_src(s, &mov[3], scratch); 5933 break; 5934 } 5935 } 5936 5937 /* No conflicts: perform this move and continue. */ 5938 tcg_out_movext1(s, &mov[3]); 5939 /* fall through */ 5940 5941 case 3: 5942 tcg_out_movext3(s, mov, mov + 1, mov + 2, 5943 parm->ntmp ? parm->tmp[0] : -1); 5944 break; 5945 case 2: 5946 tcg_out_movext2(s, mov, mov + 1, 5947 parm->ntmp ? parm->tmp[0] : -1); 5948 break; 5949 case 1: 5950 tcg_out_movext1(s, mov); 5951 break; 5952 default: 5953 g_assert_not_reached(); 5954 } 5955 } 5956 5957 static void tcg_out_helper_load_imm(TCGContext *s, unsigned slot, 5958 TCGType type, tcg_target_long imm, 5959 const TCGLdstHelperParam *parm) 5960 { 5961 if (arg_slot_reg_p(slot)) { 5962 tcg_out_movi(s, type, tcg_target_call_iarg_regs[slot], imm); 5963 } else { 5964 int ofs = tcg_out_helper_stk_ofs(type, slot); 5965 if (!tcg_out_sti(s, type, imm, TCG_REG_CALL_STACK, ofs)) { 5966 tcg_debug_assert(parm->ntmp != 0); 5967 tcg_out_movi(s, type, parm->tmp[0], imm); 5968 tcg_out_st(s, type, parm->tmp[0], TCG_REG_CALL_STACK, ofs); 5969 } 5970 } 5971 } 5972 5973 static void tcg_out_helper_load_common_args(TCGContext *s, 5974 const TCGLabelQemuLdst *ldst, 5975 const TCGLdstHelperParam *parm, 5976 const TCGHelperInfo *info, 5977 unsigned next_arg) 5978 { 5979 TCGMovExtend ptr_mov = { 5980 .dst_type = TCG_TYPE_PTR, 5981 .src_type = TCG_TYPE_PTR, 5982 .src_ext = sizeof(void *) == 4 ? MO_32 : MO_64 5983 }; 5984 const TCGCallArgumentLoc *loc = &info->in[0]; 5985 TCGType type; 5986 unsigned slot; 5987 tcg_target_ulong imm; 5988 5989 /* 5990 * Handle env, which is always first. 5991 */ 5992 ptr_mov.dst = loc->arg_slot; 5993 ptr_mov.src = TCG_AREG0; 5994 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5995 5996 /* 5997 * Handle oi. 5998 */ 5999 imm = ldst->oi; 6000 loc = &info->in[next_arg]; 6001 type = TCG_TYPE_I32; 6002 switch (loc->kind) { 6003 case TCG_CALL_ARG_NORMAL: 6004 break; 6005 case TCG_CALL_ARG_EXTEND_U: 6006 case TCG_CALL_ARG_EXTEND_S: 6007 /* No extension required for MemOpIdx. */ 6008 tcg_debug_assert(imm <= INT32_MAX); 6009 type = TCG_TYPE_REG; 6010 break; 6011 default: 6012 g_assert_not_reached(); 6013 } 6014 tcg_out_helper_load_imm(s, loc->arg_slot, type, imm, parm); 6015 next_arg++; 6016 6017 /* 6018 * Handle ra. 6019 */ 6020 loc = &info->in[next_arg]; 6021 slot = loc->arg_slot; 6022 if (parm->ra_gen) { 6023 int arg_reg = -1; 6024 TCGReg ra_reg; 6025 6026 if (arg_slot_reg_p(slot)) { 6027 arg_reg = tcg_target_call_iarg_regs[slot]; 6028 } 6029 ra_reg = parm->ra_gen(s, ldst, arg_reg); 6030 6031 ptr_mov.dst = slot; 6032 ptr_mov.src = ra_reg; 6033 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 6034 } else { 6035 imm = (uintptr_t)ldst->raddr; 6036 tcg_out_helper_load_imm(s, slot, TCG_TYPE_PTR, imm, parm); 6037 } 6038 } 6039 6040 static unsigned tcg_out_helper_add_mov(TCGMovExtend *mov, 6041 const TCGCallArgumentLoc *loc, 6042 TCGType dst_type, TCGType src_type, 6043 TCGReg lo, TCGReg hi) 6044 { 6045 MemOp reg_mo; 6046 6047 if (dst_type <= TCG_TYPE_REG) { 6048 MemOp src_ext; 6049 6050 switch (loc->kind) { 6051 case TCG_CALL_ARG_NORMAL: 6052 src_ext = src_type == TCG_TYPE_I32 ? MO_32 : MO_64; 6053 break; 6054 case TCG_CALL_ARG_EXTEND_U: 6055 dst_type = TCG_TYPE_REG; 6056 src_ext = MO_UL; 6057 break; 6058 case TCG_CALL_ARG_EXTEND_S: 6059 dst_type = TCG_TYPE_REG; 6060 src_ext = MO_SL; 6061 break; 6062 default: 6063 g_assert_not_reached(); 6064 } 6065 6066 mov[0].dst = loc->arg_slot; 6067 mov[0].dst_type = dst_type; 6068 mov[0].src = lo; 6069 mov[0].src_type = src_type; 6070 mov[0].src_ext = src_ext; 6071 return 1; 6072 } 6073 6074 if (TCG_TARGET_REG_BITS == 32) { 6075 assert(dst_type == TCG_TYPE_I64); 6076 reg_mo = MO_32; 6077 } else { 6078 assert(dst_type == TCG_TYPE_I128); 6079 reg_mo = MO_64; 6080 } 6081 6082 mov[0].dst = loc[HOST_BIG_ENDIAN].arg_slot; 6083 mov[0].src = lo; 6084 mov[0].dst_type = TCG_TYPE_REG; 6085 mov[0].src_type = TCG_TYPE_REG; 6086 mov[0].src_ext = reg_mo; 6087 6088 mov[1].dst = loc[!HOST_BIG_ENDIAN].arg_slot; 6089 mov[1].src = hi; 6090 mov[1].dst_type = TCG_TYPE_REG; 6091 mov[1].src_type = TCG_TYPE_REG; 6092 mov[1].src_ext = reg_mo; 6093 6094 return 2; 6095 } 6096 6097 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 6098 const TCGLdstHelperParam *parm) 6099 { 6100 const TCGHelperInfo *info; 6101 const TCGCallArgumentLoc *loc; 6102 TCGMovExtend mov[2]; 6103 unsigned next_arg, nmov; 6104 MemOp mop = get_memop(ldst->oi); 6105 6106 switch (mop & MO_SIZE) { 6107 case MO_8: 6108 case MO_16: 6109 case MO_32: 6110 info = &info_helper_ld32_mmu; 6111 break; 6112 case MO_64: 6113 info = &info_helper_ld64_mmu; 6114 break; 6115 case MO_128: 6116 info = &info_helper_ld128_mmu; 6117 break; 6118 default: 6119 g_assert_not_reached(); 6120 } 6121 6122 /* Defer env argument. */ 6123 next_arg = 1; 6124 6125 loc = &info->in[next_arg]; 6126 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 6127 /* 6128 * 32-bit host with 32-bit guest: zero-extend the guest address 6129 * to 64-bits for the helper by storing the low part, then 6130 * load a zero for the high part. 6131 */ 6132 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 6133 TCG_TYPE_I32, TCG_TYPE_I32, 6134 ldst->addr_reg, -1); 6135 tcg_out_helper_load_slots(s, 1, mov, parm); 6136 6137 tcg_out_helper_load_imm(s, loc[!HOST_BIG_ENDIAN].arg_slot, 6138 TCG_TYPE_I32, 0, parm); 6139 next_arg += 2; 6140 } else { 6141 nmov = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 6142 ldst->addr_reg, -1); 6143 tcg_out_helper_load_slots(s, nmov, mov, parm); 6144 next_arg += nmov; 6145 } 6146 6147 switch (info->out_kind) { 6148 case TCG_CALL_RET_NORMAL: 6149 case TCG_CALL_RET_BY_VEC: 6150 break; 6151 case TCG_CALL_RET_BY_REF: 6152 /* 6153 * The return reference is in the first argument slot. 6154 * We need memory in which to return: re-use the top of stack. 6155 */ 6156 { 6157 int ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 6158 6159 if (arg_slot_reg_p(0)) { 6160 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[0], 6161 TCG_REG_CALL_STACK, ofs_slot0); 6162 } else { 6163 tcg_debug_assert(parm->ntmp != 0); 6164 tcg_out_addi_ptr(s, parm->tmp[0], 6165 TCG_REG_CALL_STACK, ofs_slot0); 6166 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 6167 TCG_REG_CALL_STACK, ofs_slot0); 6168 } 6169 } 6170 break; 6171 default: 6172 g_assert_not_reached(); 6173 } 6174 6175 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 6176 } 6177 6178 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *ldst, 6179 bool load_sign, 6180 const TCGLdstHelperParam *parm) 6181 { 6182 MemOp mop = get_memop(ldst->oi); 6183 TCGMovExtend mov[2]; 6184 int ofs_slot0; 6185 6186 switch (ldst->type) { 6187 case TCG_TYPE_I64: 6188 if (TCG_TARGET_REG_BITS == 32) { 6189 break; 6190 } 6191 /* fall through */ 6192 6193 case TCG_TYPE_I32: 6194 mov[0].dst = ldst->datalo_reg; 6195 mov[0].src = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, 0); 6196 mov[0].dst_type = ldst->type; 6197 mov[0].src_type = TCG_TYPE_REG; 6198 6199 /* 6200 * If load_sign, then we allowed the helper to perform the 6201 * appropriate sign extension to tcg_target_ulong, and all 6202 * we need now is a plain move. 6203 * 6204 * If they do not, then we expect the relevant extension 6205 * instruction to be no more expensive than a move, and 6206 * we thus save the icache etc by only using one of two 6207 * helper functions. 6208 */ 6209 if (load_sign || !(mop & MO_SIGN)) { 6210 if (TCG_TARGET_REG_BITS == 32 || ldst->type == TCG_TYPE_I32) { 6211 mov[0].src_ext = MO_32; 6212 } else { 6213 mov[0].src_ext = MO_64; 6214 } 6215 } else { 6216 mov[0].src_ext = mop & MO_SSIZE; 6217 } 6218 tcg_out_movext1(s, mov); 6219 return; 6220 6221 case TCG_TYPE_I128: 6222 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 6223 ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 6224 switch (TCG_TARGET_CALL_RET_I128) { 6225 case TCG_CALL_RET_NORMAL: 6226 break; 6227 case TCG_CALL_RET_BY_VEC: 6228 tcg_out_st(s, TCG_TYPE_V128, 6229 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 6230 TCG_REG_CALL_STACK, ofs_slot0); 6231 /* fall through */ 6232 case TCG_CALL_RET_BY_REF: 6233 tcg_out_ld(s, TCG_TYPE_I64, ldst->datalo_reg, 6234 TCG_REG_CALL_STACK, ofs_slot0 + 8 * HOST_BIG_ENDIAN); 6235 tcg_out_ld(s, TCG_TYPE_I64, ldst->datahi_reg, 6236 TCG_REG_CALL_STACK, ofs_slot0 + 8 * !HOST_BIG_ENDIAN); 6237 return; 6238 default: 6239 g_assert_not_reached(); 6240 } 6241 break; 6242 6243 default: 6244 g_assert_not_reached(); 6245 } 6246 6247 mov[0].dst = ldst->datalo_reg; 6248 mov[0].src = 6249 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, HOST_BIG_ENDIAN); 6250 mov[0].dst_type = TCG_TYPE_REG; 6251 mov[0].src_type = TCG_TYPE_REG; 6252 mov[0].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 6253 6254 mov[1].dst = ldst->datahi_reg; 6255 mov[1].src = 6256 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, !HOST_BIG_ENDIAN); 6257 mov[1].dst_type = TCG_TYPE_REG; 6258 mov[1].src_type = TCG_TYPE_REG; 6259 mov[1].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 6260 6261 tcg_out_movext2(s, mov, mov + 1, parm->ntmp ? parm->tmp[0] : -1); 6262 } 6263 6264 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 6265 const TCGLdstHelperParam *parm) 6266 { 6267 const TCGHelperInfo *info; 6268 const TCGCallArgumentLoc *loc; 6269 TCGMovExtend mov[4]; 6270 TCGType data_type; 6271 unsigned next_arg, nmov, n; 6272 MemOp mop = get_memop(ldst->oi); 6273 6274 switch (mop & MO_SIZE) { 6275 case MO_8: 6276 case MO_16: 6277 case MO_32: 6278 info = &info_helper_st32_mmu; 6279 data_type = TCG_TYPE_I32; 6280 break; 6281 case MO_64: 6282 info = &info_helper_st64_mmu; 6283 data_type = TCG_TYPE_I64; 6284 break; 6285 case MO_128: 6286 info = &info_helper_st128_mmu; 6287 data_type = TCG_TYPE_I128; 6288 break; 6289 default: 6290 g_assert_not_reached(); 6291 } 6292 6293 /* Defer env argument. */ 6294 next_arg = 1; 6295 nmov = 0; 6296 6297 /* Handle addr argument. */ 6298 loc = &info->in[next_arg]; 6299 tcg_debug_assert(s->addr_type <= TCG_TYPE_REG); 6300 if (TCG_TARGET_REG_BITS == 32) { 6301 /* 6302 * 32-bit host (and thus 32-bit guest): zero-extend the guest address 6303 * to 64-bits for the helper by storing the low part. Later, 6304 * after we have processed the register inputs, we will load a 6305 * zero for the high part. 6306 */ 6307 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 6308 TCG_TYPE_I32, TCG_TYPE_I32, 6309 ldst->addr_reg, -1); 6310 next_arg += 2; 6311 nmov += 1; 6312 } else { 6313 n = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 6314 ldst->addr_reg, -1); 6315 next_arg += n; 6316 nmov += n; 6317 } 6318 6319 /* Handle data argument. */ 6320 loc = &info->in[next_arg]; 6321 switch (loc->kind) { 6322 case TCG_CALL_ARG_NORMAL: 6323 case TCG_CALL_ARG_EXTEND_U: 6324 case TCG_CALL_ARG_EXTEND_S: 6325 n = tcg_out_helper_add_mov(mov + nmov, loc, data_type, ldst->type, 6326 ldst->datalo_reg, ldst->datahi_reg); 6327 next_arg += n; 6328 nmov += n; 6329 tcg_out_helper_load_slots(s, nmov, mov, parm); 6330 break; 6331 6332 case TCG_CALL_ARG_BY_REF: 6333 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 6334 tcg_debug_assert(data_type == TCG_TYPE_I128); 6335 tcg_out_st(s, TCG_TYPE_I64, 6336 HOST_BIG_ENDIAN ? ldst->datahi_reg : ldst->datalo_reg, 6337 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[0].ref_slot)); 6338 tcg_out_st(s, TCG_TYPE_I64, 6339 HOST_BIG_ENDIAN ? ldst->datalo_reg : ldst->datahi_reg, 6340 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[1].ref_slot)); 6341 6342 tcg_out_helper_load_slots(s, nmov, mov, parm); 6343 6344 if (arg_slot_reg_p(loc->arg_slot)) { 6345 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[loc->arg_slot], 6346 TCG_REG_CALL_STACK, 6347 arg_slot_stk_ofs(loc->ref_slot)); 6348 } else { 6349 tcg_debug_assert(parm->ntmp != 0); 6350 tcg_out_addi_ptr(s, parm->tmp[0], TCG_REG_CALL_STACK, 6351 arg_slot_stk_ofs(loc->ref_slot)); 6352 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 6353 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc->arg_slot)); 6354 } 6355 next_arg += 2; 6356 break; 6357 6358 default: 6359 g_assert_not_reached(); 6360 } 6361 6362 if (TCG_TARGET_REG_BITS == 32) { 6363 /* Zero extend the address by loading a zero for the high part. */ 6364 loc = &info->in[1 + !HOST_BIG_ENDIAN]; 6365 tcg_out_helper_load_imm(s, loc->arg_slot, TCG_TYPE_I32, 0, parm); 6366 } 6367 6368 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 6369 } 6370 6371 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start) 6372 { 6373 int i, start_words, num_insns; 6374 TCGOp *op; 6375 6376 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP) 6377 && qemu_log_in_addr_range(pc_start))) { 6378 FILE *logfile = qemu_log_trylock(); 6379 if (logfile) { 6380 fprintf(logfile, "OP:\n"); 6381 tcg_dump_ops(s, logfile, false); 6382 fprintf(logfile, "\n"); 6383 qemu_log_unlock(logfile); 6384 } 6385 } 6386 6387 #ifdef CONFIG_DEBUG_TCG 6388 /* Ensure all labels referenced have been emitted. */ 6389 { 6390 TCGLabel *l; 6391 bool error = false; 6392 6393 QSIMPLEQ_FOREACH(l, &s->labels, next) { 6394 if (unlikely(!l->present) && !QSIMPLEQ_EMPTY(&l->branches)) { 6395 qemu_log_mask(CPU_LOG_TB_OP, 6396 "$L%d referenced but not present.\n", l->id); 6397 error = true; 6398 } 6399 } 6400 assert(!error); 6401 } 6402 #endif 6403 6404 /* Do not reuse any EBB that may be allocated within the TB. */ 6405 tcg_temp_ebb_reset_freed(s); 6406 6407 tcg_optimize(s); 6408 6409 reachable_code_pass(s); 6410 liveness_pass_0(s); 6411 liveness_pass_1(s); 6412 6413 if (s->nb_indirects > 0) { 6414 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND) 6415 && qemu_log_in_addr_range(pc_start))) { 6416 FILE *logfile = qemu_log_trylock(); 6417 if (logfile) { 6418 fprintf(logfile, "OP before indirect lowering:\n"); 6419 tcg_dump_ops(s, logfile, false); 6420 fprintf(logfile, "\n"); 6421 qemu_log_unlock(logfile); 6422 } 6423 } 6424 6425 /* Replace indirect temps with direct temps. */ 6426 if (liveness_pass_2(s)) { 6427 /* If changes were made, re-run liveness. */ 6428 liveness_pass_1(s); 6429 } 6430 } 6431 6432 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT) 6433 && qemu_log_in_addr_range(pc_start))) { 6434 FILE *logfile = qemu_log_trylock(); 6435 if (logfile) { 6436 fprintf(logfile, "OP after optimization and liveness analysis:\n"); 6437 tcg_dump_ops(s, logfile, true); 6438 fprintf(logfile, "\n"); 6439 qemu_log_unlock(logfile); 6440 } 6441 } 6442 6443 /* Initialize goto_tb jump offsets. */ 6444 tb->jmp_reset_offset[0] = TB_JMP_OFFSET_INVALID; 6445 tb->jmp_reset_offset[1] = TB_JMP_OFFSET_INVALID; 6446 tb->jmp_insn_offset[0] = TB_JMP_OFFSET_INVALID; 6447 tb->jmp_insn_offset[1] = TB_JMP_OFFSET_INVALID; 6448 6449 tcg_reg_alloc_start(s); 6450 6451 /* 6452 * Reset the buffer pointers when restarting after overflow. 6453 * TODO: Move this into translate-all.c with the rest of the 6454 * buffer management. Having only this done here is confusing. 6455 */ 6456 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr); 6457 s->code_ptr = s->code_buf; 6458 s->data_gen_ptr = NULL; 6459 6460 QSIMPLEQ_INIT(&s->ldst_labels); 6461 s->pool_labels = NULL; 6462 6463 start_words = s->insn_start_words; 6464 s->gen_insn_data = 6465 tcg_malloc(sizeof(uint64_t) * s->gen_tb->icount * start_words); 6466 6467 tcg_out_tb_start(s); 6468 6469 num_insns = -1; 6470 QTAILQ_FOREACH(op, &s->ops, link) { 6471 TCGOpcode opc = op->opc; 6472 6473 switch (opc) { 6474 case INDEX_op_mov: 6475 case INDEX_op_mov_vec: 6476 tcg_reg_alloc_mov(s, op); 6477 break; 6478 case INDEX_op_dup_vec: 6479 tcg_reg_alloc_dup(s, op); 6480 break; 6481 case INDEX_op_insn_start: 6482 if (num_insns >= 0) { 6483 size_t off = tcg_current_code_size(s); 6484 s->gen_insn_end_off[num_insns] = off; 6485 /* Assert that we do not overflow our stored offset. */ 6486 assert(s->gen_insn_end_off[num_insns] == off); 6487 } 6488 num_insns++; 6489 for (i = 0; i < start_words; ++i) { 6490 s->gen_insn_data[num_insns * start_words + i] = 6491 tcg_get_insn_start_param(op, i); 6492 } 6493 break; 6494 case INDEX_op_discard: 6495 temp_dead(s, arg_temp(op->args[0])); 6496 break; 6497 case INDEX_op_set_label: 6498 tcg_reg_alloc_bb_end(s, s->reserved_regs); 6499 tcg_out_label(s, arg_label(op->args[0])); 6500 break; 6501 case INDEX_op_call: 6502 tcg_reg_alloc_call(s, op); 6503 break; 6504 case INDEX_op_exit_tb: 6505 tcg_out_exit_tb(s, op->args[0]); 6506 break; 6507 case INDEX_op_goto_tb: 6508 tcg_out_goto_tb(s, op->args[0]); 6509 break; 6510 case INDEX_op_dup2_vec: 6511 if (tcg_reg_alloc_dup2(s, op)) { 6512 break; 6513 } 6514 /* fall through */ 6515 default: 6516 /* Sanity check that we've not introduced any unhandled opcodes. */ 6517 tcg_debug_assert(tcg_op_supported(opc, TCGOP_TYPE(op), 6518 TCGOP_FLAGS(op))); 6519 /* Note: in order to speed up the code, it would be much 6520 faster to have specialized register allocator functions for 6521 some common argument patterns */ 6522 tcg_reg_alloc_op(s, op); 6523 break; 6524 } 6525 /* Test for (pending) buffer overflow. The assumption is that any 6526 one operation beginning below the high water mark cannot overrun 6527 the buffer completely. Thus we can test for overflow after 6528 generating code without having to check during generation. */ 6529 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) { 6530 return -1; 6531 } 6532 /* Test for TB overflow, as seen by gen_insn_end_off. */ 6533 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) { 6534 return -2; 6535 } 6536 } 6537 tcg_debug_assert(num_insns + 1 == s->gen_tb->icount); 6538 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s); 6539 6540 /* Generate TB finalization at the end of block */ 6541 i = tcg_out_ldst_finalize(s); 6542 if (i < 0) { 6543 return i; 6544 } 6545 i = tcg_out_pool_finalize(s); 6546 if (i < 0) { 6547 return i; 6548 } 6549 if (!tcg_resolve_relocs(s)) { 6550 return -2; 6551 } 6552 6553 #ifndef CONFIG_TCG_INTERPRETER 6554 /* flush instruction cache */ 6555 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 6556 (uintptr_t)s->code_buf, 6557 tcg_ptr_byte_diff(s->code_ptr, s->code_buf)); 6558 #endif 6559 6560 return tcg_current_code_size(s); 6561 } 6562 6563 #ifdef ELF_HOST_MACHINE 6564 /* In order to use this feature, the backend needs to do three things: 6565 6566 (1) Define ELF_HOST_MACHINE to indicate both what value to 6567 put into the ELF image and to indicate support for the feature. 6568 6569 (2) Define tcg_register_jit. This should create a buffer containing 6570 the contents of a .debug_frame section that describes the post- 6571 prologue unwind info for the tcg machine. 6572 6573 (3) Call tcg_register_jit_int, with the constructed .debug_frame. 6574 */ 6575 6576 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */ 6577 typedef enum { 6578 JIT_NOACTION = 0, 6579 JIT_REGISTER_FN, 6580 JIT_UNREGISTER_FN 6581 } jit_actions_t; 6582 6583 struct jit_code_entry { 6584 struct jit_code_entry *next_entry; 6585 struct jit_code_entry *prev_entry; 6586 const void *symfile_addr; 6587 uint64_t symfile_size; 6588 }; 6589 6590 struct jit_descriptor { 6591 uint32_t version; 6592 uint32_t action_flag; 6593 struct jit_code_entry *relevant_entry; 6594 struct jit_code_entry *first_entry; 6595 }; 6596 6597 void __jit_debug_register_code(void) __attribute__((noinline)); 6598 void __jit_debug_register_code(void) 6599 { 6600 asm(""); 6601 } 6602 6603 /* Must statically initialize the version, because GDB may check 6604 the version before we can set it. */ 6605 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 }; 6606 6607 /* End GDB interface. */ 6608 6609 static int find_string(const char *strtab, const char *str) 6610 { 6611 const char *p = strtab + 1; 6612 6613 while (1) { 6614 if (strcmp(p, str) == 0) { 6615 return p - strtab; 6616 } 6617 p += strlen(p) + 1; 6618 } 6619 } 6620 6621 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size, 6622 const void *debug_frame, 6623 size_t debug_frame_size) 6624 { 6625 struct __attribute__((packed)) DebugInfo { 6626 uint32_t len; 6627 uint16_t version; 6628 uint32_t abbrev; 6629 uint8_t ptr_size; 6630 uint8_t cu_die; 6631 uint16_t cu_lang; 6632 uintptr_t cu_low_pc; 6633 uintptr_t cu_high_pc; 6634 uint8_t fn_die; 6635 char fn_name[16]; 6636 uintptr_t fn_low_pc; 6637 uintptr_t fn_high_pc; 6638 uint8_t cu_eoc; 6639 }; 6640 6641 struct ElfImage { 6642 ElfW(Ehdr) ehdr; 6643 ElfW(Phdr) phdr; 6644 ElfW(Shdr) shdr[7]; 6645 ElfW(Sym) sym[2]; 6646 struct DebugInfo di; 6647 uint8_t da[24]; 6648 char str[80]; 6649 }; 6650 6651 struct ElfImage *img; 6652 6653 static const struct ElfImage img_template = { 6654 .ehdr = { 6655 .e_ident[EI_MAG0] = ELFMAG0, 6656 .e_ident[EI_MAG1] = ELFMAG1, 6657 .e_ident[EI_MAG2] = ELFMAG2, 6658 .e_ident[EI_MAG3] = ELFMAG3, 6659 .e_ident[EI_CLASS] = ELF_CLASS, 6660 .e_ident[EI_DATA] = ELF_DATA, 6661 .e_ident[EI_VERSION] = EV_CURRENT, 6662 .e_type = ET_EXEC, 6663 .e_machine = ELF_HOST_MACHINE, 6664 .e_version = EV_CURRENT, 6665 .e_phoff = offsetof(struct ElfImage, phdr), 6666 .e_shoff = offsetof(struct ElfImage, shdr), 6667 .e_ehsize = sizeof(ElfW(Shdr)), 6668 .e_phentsize = sizeof(ElfW(Phdr)), 6669 .e_phnum = 1, 6670 .e_shentsize = sizeof(ElfW(Shdr)), 6671 .e_shnum = ARRAY_SIZE(img->shdr), 6672 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1, 6673 #ifdef ELF_HOST_FLAGS 6674 .e_flags = ELF_HOST_FLAGS, 6675 #endif 6676 #ifdef ELF_OSABI 6677 .e_ident[EI_OSABI] = ELF_OSABI, 6678 #endif 6679 }, 6680 .phdr = { 6681 .p_type = PT_LOAD, 6682 .p_flags = PF_X, 6683 }, 6684 .shdr = { 6685 [0] = { .sh_type = SHT_NULL }, 6686 /* Trick: The contents of code_gen_buffer are not present in 6687 this fake ELF file; that got allocated elsewhere. Therefore 6688 we mark .text as SHT_NOBITS (similar to .bss) so that readers 6689 will not look for contents. We can record any address. */ 6690 [1] = { /* .text */ 6691 .sh_type = SHT_NOBITS, 6692 .sh_flags = SHF_EXECINSTR | SHF_ALLOC, 6693 }, 6694 [2] = { /* .debug_info */ 6695 .sh_type = SHT_PROGBITS, 6696 .sh_offset = offsetof(struct ElfImage, di), 6697 .sh_size = sizeof(struct DebugInfo), 6698 }, 6699 [3] = { /* .debug_abbrev */ 6700 .sh_type = SHT_PROGBITS, 6701 .sh_offset = offsetof(struct ElfImage, da), 6702 .sh_size = sizeof(img->da), 6703 }, 6704 [4] = { /* .debug_frame */ 6705 .sh_type = SHT_PROGBITS, 6706 .sh_offset = sizeof(struct ElfImage), 6707 }, 6708 [5] = { /* .symtab */ 6709 .sh_type = SHT_SYMTAB, 6710 .sh_offset = offsetof(struct ElfImage, sym), 6711 .sh_size = sizeof(img->sym), 6712 .sh_info = 1, 6713 .sh_link = ARRAY_SIZE(img->shdr) - 1, 6714 .sh_entsize = sizeof(ElfW(Sym)), 6715 }, 6716 [6] = { /* .strtab */ 6717 .sh_type = SHT_STRTAB, 6718 .sh_offset = offsetof(struct ElfImage, str), 6719 .sh_size = sizeof(img->str), 6720 } 6721 }, 6722 .sym = { 6723 [1] = { /* code_gen_buffer */ 6724 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC), 6725 .st_shndx = 1, 6726 } 6727 }, 6728 .di = { 6729 .len = sizeof(struct DebugInfo) - 4, 6730 .version = 2, 6731 .ptr_size = sizeof(void *), 6732 .cu_die = 1, 6733 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */ 6734 .fn_die = 2, 6735 .fn_name = "code_gen_buffer" 6736 }, 6737 .da = { 6738 1, /* abbrev number (the cu) */ 6739 0x11, 1, /* DW_TAG_compile_unit, has children */ 6740 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */ 6741 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6742 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6743 0, 0, /* end of abbrev */ 6744 2, /* abbrev number (the fn) */ 6745 0x2e, 0, /* DW_TAG_subprogram, no children */ 6746 0x3, 0x8, /* DW_AT_name, DW_FORM_string */ 6747 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6748 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6749 0, 0, /* end of abbrev */ 6750 0 /* no more abbrev */ 6751 }, 6752 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0" 6753 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer", 6754 }; 6755 6756 /* We only need a single jit entry; statically allocate it. */ 6757 static struct jit_code_entry one_entry; 6758 6759 uintptr_t buf = (uintptr_t)buf_ptr; 6760 size_t img_size = sizeof(struct ElfImage) + debug_frame_size; 6761 DebugFrameHeader *dfh; 6762 6763 img = g_malloc(img_size); 6764 *img = img_template; 6765 6766 img->phdr.p_vaddr = buf; 6767 img->phdr.p_paddr = buf; 6768 img->phdr.p_memsz = buf_size; 6769 6770 img->shdr[1].sh_name = find_string(img->str, ".text"); 6771 img->shdr[1].sh_addr = buf; 6772 img->shdr[1].sh_size = buf_size; 6773 6774 img->shdr[2].sh_name = find_string(img->str, ".debug_info"); 6775 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev"); 6776 6777 img->shdr[4].sh_name = find_string(img->str, ".debug_frame"); 6778 img->shdr[4].sh_size = debug_frame_size; 6779 6780 img->shdr[5].sh_name = find_string(img->str, ".symtab"); 6781 img->shdr[6].sh_name = find_string(img->str, ".strtab"); 6782 6783 img->sym[1].st_name = find_string(img->str, "code_gen_buffer"); 6784 img->sym[1].st_value = buf; 6785 img->sym[1].st_size = buf_size; 6786 6787 img->di.cu_low_pc = buf; 6788 img->di.cu_high_pc = buf + buf_size; 6789 img->di.fn_low_pc = buf; 6790 img->di.fn_high_pc = buf + buf_size; 6791 6792 dfh = (DebugFrameHeader *)(img + 1); 6793 memcpy(dfh, debug_frame, debug_frame_size); 6794 dfh->fde.func_start = buf; 6795 dfh->fde.func_len = buf_size; 6796 6797 #ifdef DEBUG_JIT 6798 /* Enable this block to be able to debug the ELF image file creation. 6799 One can use readelf, objdump, or other inspection utilities. */ 6800 { 6801 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir()); 6802 FILE *f = fopen(jit, "w+b"); 6803 if (f) { 6804 if (fwrite(img, img_size, 1, f) != img_size) { 6805 /* Avoid stupid unused return value warning for fwrite. */ 6806 } 6807 fclose(f); 6808 } 6809 } 6810 #endif 6811 6812 one_entry.symfile_addr = img; 6813 one_entry.symfile_size = img_size; 6814 6815 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN; 6816 __jit_debug_descriptor.relevant_entry = &one_entry; 6817 __jit_debug_descriptor.first_entry = &one_entry; 6818 __jit_debug_register_code(); 6819 } 6820 #else 6821 /* No support for the feature. Provide the entry point expected by exec.c, 6822 and implement the internal function we declared earlier. */ 6823 6824 static void tcg_register_jit_int(const void *buf, size_t size, 6825 const void *debug_frame, 6826 size_t debug_frame_size) 6827 { 6828 } 6829 6830 void tcg_register_jit(const void *buf, size_t buf_size) 6831 { 6832 } 6833 #endif /* ELF_HOST_MACHINE */ 6834 6835 #if !TCG_TARGET_MAYBE_vec 6836 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...) 6837 { 6838 g_assert_not_reached(); 6839 } 6840 #endif 6841