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 38 /* Note: the long term plan is to reduce the dependencies on the QEMU 39 CPU definitions. Currently they are used for qemu_ld/st 40 instructions */ 41 #define NO_CPU_IO_DEFS 42 43 #include "exec/exec-all.h" 44 #include "tcg/tcg-op.h" 45 46 #if UINTPTR_MAX == UINT32_MAX 47 # define ELF_CLASS ELFCLASS32 48 #else 49 # define ELF_CLASS ELFCLASS64 50 #endif 51 #if HOST_BIG_ENDIAN 52 # define ELF_DATA ELFDATA2MSB 53 #else 54 # define ELF_DATA ELFDATA2LSB 55 #endif 56 57 #include "elf.h" 58 #include "exec/log.h" 59 #include "tcg/tcg-ldst.h" 60 #include "tcg/tcg-temp-internal.h" 61 #include "tcg-internal.h" 62 #include "accel/tcg/perf.h" 63 #ifdef CONFIG_USER_ONLY 64 #include "exec/user/guest-base.h" 65 #endif 66 67 /* Forward declarations for functions declared in tcg-target.c.inc and 68 used here. */ 69 static void tcg_target_init(TCGContext *s); 70 static void tcg_target_qemu_prologue(TCGContext *s); 71 static bool patch_reloc(tcg_insn_unit *code_ptr, int type, 72 intptr_t value, intptr_t addend); 73 74 /* The CIE and FDE header definitions will be common to all hosts. */ 75 typedef struct { 76 uint32_t len __attribute__((aligned((sizeof(void *))))); 77 uint32_t id; 78 uint8_t version; 79 char augmentation[1]; 80 uint8_t code_align; 81 uint8_t data_align; 82 uint8_t return_column; 83 } DebugFrameCIE; 84 85 typedef struct QEMU_PACKED { 86 uint32_t len __attribute__((aligned((sizeof(void *))))); 87 uint32_t cie_offset; 88 uintptr_t func_start; 89 uintptr_t func_len; 90 } DebugFrameFDEHeader; 91 92 typedef struct QEMU_PACKED { 93 DebugFrameCIE cie; 94 DebugFrameFDEHeader fde; 95 } DebugFrameHeader; 96 97 typedef struct TCGLabelQemuLdst { 98 bool is_ld; /* qemu_ld: true, qemu_st: false */ 99 MemOpIdx oi; 100 TCGType type; /* result type of a load */ 101 TCGReg addrlo_reg; /* reg index for low word of guest virtual addr */ 102 TCGReg addrhi_reg; /* reg index for high word of guest virtual addr */ 103 TCGReg datalo_reg; /* reg index for low word to be loaded or stored */ 104 TCGReg datahi_reg; /* reg index for high word to be loaded or stored */ 105 const tcg_insn_unit *raddr; /* addr of the next IR of qemu_ld/st IR */ 106 tcg_insn_unit *label_ptr[2]; /* label pointers to be updated */ 107 QSIMPLEQ_ENTRY(TCGLabelQemuLdst) next; 108 } TCGLabelQemuLdst; 109 110 static void tcg_register_jit_int(const void *buf, size_t size, 111 const void *debug_frame, 112 size_t debug_frame_size) 113 __attribute__((unused)); 114 115 /* Forward declarations for functions declared and used in tcg-target.c.inc. */ 116 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1, 117 intptr_t arg2); 118 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 119 static void tcg_out_movi(TCGContext *s, TCGType type, 120 TCGReg ret, tcg_target_long arg); 121 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 122 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 123 static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg); 124 static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg); 125 static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg); 126 static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg); 127 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 128 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 129 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg); 130 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long); 131 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2); 132 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg); 133 static void tcg_out_goto_tb(TCGContext *s, int which); 134 static void tcg_out_op(TCGContext *s, TCGOpcode opc, 135 const TCGArg args[TCG_MAX_OP_ARGS], 136 const int const_args[TCG_MAX_OP_ARGS]); 137 #if TCG_TARGET_MAYBE_vec 138 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 139 TCGReg dst, TCGReg src); 140 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 141 TCGReg dst, TCGReg base, intptr_t offset); 142 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 143 TCGReg dst, int64_t arg); 144 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 145 unsigned vecl, unsigned vece, 146 const TCGArg args[TCG_MAX_OP_ARGS], 147 const int const_args[TCG_MAX_OP_ARGS]); 148 #else 149 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 150 TCGReg dst, TCGReg src) 151 { 152 g_assert_not_reached(); 153 } 154 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 155 TCGReg dst, TCGReg base, intptr_t offset) 156 { 157 g_assert_not_reached(); 158 } 159 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 160 TCGReg dst, int64_t arg) 161 { 162 g_assert_not_reached(); 163 } 164 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 165 unsigned vecl, unsigned vece, 166 const TCGArg args[TCG_MAX_OP_ARGS], 167 const int const_args[TCG_MAX_OP_ARGS]) 168 { 169 g_assert_not_reached(); 170 } 171 #endif 172 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1, 173 intptr_t arg2); 174 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, 175 TCGReg base, intptr_t ofs); 176 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target, 177 const TCGHelperInfo *info); 178 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot); 179 static bool tcg_target_const_match(int64_t val, TCGType type, int ct); 180 #ifdef TCG_TARGET_NEED_LDST_LABELS 181 static int tcg_out_ldst_finalize(TCGContext *s); 182 #endif 183 184 typedef struct TCGLdstHelperParam { 185 TCGReg (*ra_gen)(TCGContext *s, const TCGLabelQemuLdst *l, int arg_reg); 186 unsigned ntmp; 187 int tmp[3]; 188 } TCGLdstHelperParam; 189 190 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 191 const TCGLdstHelperParam *p) 192 __attribute__((unused)); 193 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *l, 194 bool load_sign, const TCGLdstHelperParam *p) 195 __attribute__((unused)); 196 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 197 const TCGLdstHelperParam *p) 198 __attribute__((unused)); 199 200 static void * const qemu_ld_helpers[MO_SSIZE + 1] __attribute__((unused)) = { 201 [MO_UB] = helper_ldub_mmu, 202 [MO_SB] = helper_ldsb_mmu, 203 [MO_UW] = helper_lduw_mmu, 204 [MO_SW] = helper_ldsw_mmu, 205 [MO_UL] = helper_ldul_mmu, 206 [MO_UQ] = helper_ldq_mmu, 207 #if TCG_TARGET_REG_BITS == 64 208 [MO_SL] = helper_ldsl_mmu, 209 [MO_128] = helper_ld16_mmu, 210 #endif 211 }; 212 213 static void * const qemu_st_helpers[MO_SIZE + 1] __attribute__((unused)) = { 214 [MO_8] = helper_stb_mmu, 215 [MO_16] = helper_stw_mmu, 216 [MO_32] = helper_stl_mmu, 217 [MO_64] = helper_stq_mmu, 218 #if TCG_TARGET_REG_BITS == 64 219 [MO_128] = helper_st16_mmu, 220 #endif 221 }; 222 223 typedef struct { 224 MemOp atom; /* lg2 bits of atomicity required */ 225 MemOp align; /* lg2 bits of alignment to use */ 226 } TCGAtomAlign; 227 228 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 229 MemOp host_atom, bool allow_two_ops) 230 __attribute__((unused)); 231 232 TCGContext tcg_init_ctx; 233 __thread TCGContext *tcg_ctx; 234 235 TCGContext **tcg_ctxs; 236 unsigned int tcg_cur_ctxs; 237 unsigned int tcg_max_ctxs; 238 TCGv_env cpu_env = 0; 239 const void *tcg_code_gen_epilogue; 240 uintptr_t tcg_splitwx_diff; 241 242 #ifndef CONFIG_TCG_INTERPRETER 243 tcg_prologue_fn *tcg_qemu_tb_exec; 244 #endif 245 246 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT]; 247 static TCGRegSet tcg_target_call_clobber_regs; 248 249 #if TCG_TARGET_INSN_UNIT_SIZE == 1 250 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v) 251 { 252 *s->code_ptr++ = v; 253 } 254 255 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p, 256 uint8_t v) 257 { 258 *p = v; 259 } 260 #endif 261 262 #if TCG_TARGET_INSN_UNIT_SIZE <= 2 263 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v) 264 { 265 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 266 *s->code_ptr++ = v; 267 } else { 268 tcg_insn_unit *p = s->code_ptr; 269 memcpy(p, &v, sizeof(v)); 270 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE); 271 } 272 } 273 274 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p, 275 uint16_t v) 276 { 277 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 278 *p = v; 279 } else { 280 memcpy(p, &v, sizeof(v)); 281 } 282 } 283 #endif 284 285 #if TCG_TARGET_INSN_UNIT_SIZE <= 4 286 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v) 287 { 288 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 289 *s->code_ptr++ = v; 290 } else { 291 tcg_insn_unit *p = s->code_ptr; 292 memcpy(p, &v, sizeof(v)); 293 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE); 294 } 295 } 296 297 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p, 298 uint32_t v) 299 { 300 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 301 *p = v; 302 } else { 303 memcpy(p, &v, sizeof(v)); 304 } 305 } 306 #endif 307 308 #if TCG_TARGET_INSN_UNIT_SIZE <= 8 309 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v) 310 { 311 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 312 *s->code_ptr++ = v; 313 } else { 314 tcg_insn_unit *p = s->code_ptr; 315 memcpy(p, &v, sizeof(v)); 316 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE); 317 } 318 } 319 320 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p, 321 uint64_t v) 322 { 323 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 324 *p = v; 325 } else { 326 memcpy(p, &v, sizeof(v)); 327 } 328 } 329 #endif 330 331 /* label relocation processing */ 332 333 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type, 334 TCGLabel *l, intptr_t addend) 335 { 336 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation)); 337 338 r->type = type; 339 r->ptr = code_ptr; 340 r->addend = addend; 341 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next); 342 } 343 344 static void tcg_out_label(TCGContext *s, TCGLabel *l) 345 { 346 tcg_debug_assert(!l->has_value); 347 l->has_value = 1; 348 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr); 349 } 350 351 TCGLabel *gen_new_label(void) 352 { 353 TCGContext *s = tcg_ctx; 354 TCGLabel *l = tcg_malloc(sizeof(TCGLabel)); 355 356 memset(l, 0, sizeof(TCGLabel)); 357 l->id = s->nb_labels++; 358 QSIMPLEQ_INIT(&l->branches); 359 QSIMPLEQ_INIT(&l->relocs); 360 361 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next); 362 363 return l; 364 } 365 366 static bool tcg_resolve_relocs(TCGContext *s) 367 { 368 TCGLabel *l; 369 370 QSIMPLEQ_FOREACH(l, &s->labels, next) { 371 TCGRelocation *r; 372 uintptr_t value = l->u.value; 373 374 QSIMPLEQ_FOREACH(r, &l->relocs, next) { 375 if (!patch_reloc(r->ptr, r->type, value, r->addend)) { 376 return false; 377 } 378 } 379 } 380 return true; 381 } 382 383 static void set_jmp_reset_offset(TCGContext *s, int which) 384 { 385 /* 386 * We will check for overflow at the end of the opcode loop in 387 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 388 */ 389 s->gen_tb->jmp_reset_offset[which] = tcg_current_code_size(s); 390 } 391 392 static void G_GNUC_UNUSED set_jmp_insn_offset(TCGContext *s, int which) 393 { 394 /* 395 * We will check for overflow at the end of the opcode loop in 396 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 397 */ 398 s->gen_tb->jmp_insn_offset[which] = tcg_current_code_size(s); 399 } 400 401 static uintptr_t G_GNUC_UNUSED get_jmp_target_addr(TCGContext *s, int which) 402 { 403 /* 404 * Return the read-execute version of the pointer, for the benefit 405 * of any pc-relative addressing mode. 406 */ 407 return (uintptr_t)tcg_splitwx_to_rx(&s->gen_tb->jmp_target_addr[which]); 408 } 409 410 /* Signal overflow, starting over with fewer guest insns. */ 411 static G_NORETURN 412 void tcg_raise_tb_overflow(TCGContext *s) 413 { 414 siglongjmp(s->jmp_trans, -2); 415 } 416 417 /* 418 * Used by tcg_out_movext{1,2} to hold the arguments for tcg_out_movext. 419 * By the time we arrive at tcg_out_movext1, @dst is always a TCGReg. 420 * 421 * However, tcg_out_helper_load_slots reuses this field to hold an 422 * argument slot number (which may designate a argument register or an 423 * argument stack slot), converting to TCGReg once all arguments that 424 * are destined for the stack are processed. 425 */ 426 typedef struct TCGMovExtend { 427 unsigned dst; 428 TCGReg src; 429 TCGType dst_type; 430 TCGType src_type; 431 MemOp src_ext; 432 } TCGMovExtend; 433 434 /** 435 * tcg_out_movext -- move and extend 436 * @s: tcg context 437 * @dst_type: integral type for destination 438 * @dst: destination register 439 * @src_type: integral type for source 440 * @src_ext: extension to apply to source 441 * @src: source register 442 * 443 * Move or extend @src into @dst, depending on @src_ext and the types. 444 */ 445 static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst, 446 TCGType src_type, MemOp src_ext, TCGReg src) 447 { 448 switch (src_ext) { 449 case MO_UB: 450 tcg_out_ext8u(s, dst, src); 451 break; 452 case MO_SB: 453 tcg_out_ext8s(s, dst_type, dst, src); 454 break; 455 case MO_UW: 456 tcg_out_ext16u(s, dst, src); 457 break; 458 case MO_SW: 459 tcg_out_ext16s(s, dst_type, dst, src); 460 break; 461 case MO_UL: 462 case MO_SL: 463 if (dst_type == TCG_TYPE_I32) { 464 if (src_type == TCG_TYPE_I32) { 465 tcg_out_mov(s, TCG_TYPE_I32, dst, src); 466 } else { 467 tcg_out_extrl_i64_i32(s, dst, src); 468 } 469 } else if (src_type == TCG_TYPE_I32) { 470 if (src_ext & MO_SIGN) { 471 tcg_out_exts_i32_i64(s, dst, src); 472 } else { 473 tcg_out_extu_i32_i64(s, dst, src); 474 } 475 } else { 476 if (src_ext & MO_SIGN) { 477 tcg_out_ext32s(s, dst, src); 478 } else { 479 tcg_out_ext32u(s, dst, src); 480 } 481 } 482 break; 483 case MO_UQ: 484 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 485 if (dst_type == TCG_TYPE_I32) { 486 tcg_out_extrl_i64_i32(s, dst, src); 487 } else { 488 tcg_out_mov(s, TCG_TYPE_I64, dst, src); 489 } 490 break; 491 default: 492 g_assert_not_reached(); 493 } 494 } 495 496 /* Minor variations on a theme, using a structure. */ 497 static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i, 498 TCGReg src) 499 { 500 tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src); 501 } 502 503 static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i) 504 { 505 tcg_out_movext1_new_src(s, i, i->src); 506 } 507 508 /** 509 * tcg_out_movext2 -- move and extend two pair 510 * @s: tcg context 511 * @i1: first move description 512 * @i2: second move description 513 * @scratch: temporary register, or -1 for none 514 * 515 * As tcg_out_movext, for both @i1 and @i2, caring for overlap 516 * between the sources and destinations. 517 */ 518 519 static void tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1, 520 const TCGMovExtend *i2, int scratch) 521 { 522 TCGReg src1 = i1->src; 523 TCGReg src2 = i2->src; 524 525 if (i1->dst != src2) { 526 tcg_out_movext1(s, i1); 527 tcg_out_movext1(s, i2); 528 return; 529 } 530 if (i2->dst == src1) { 531 TCGType src1_type = i1->src_type; 532 TCGType src2_type = i2->src_type; 533 534 if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) { 535 /* The data is now in the correct registers, now extend. */ 536 src1 = i2->src; 537 src2 = i1->src; 538 } else { 539 tcg_debug_assert(scratch >= 0); 540 tcg_out_mov(s, src1_type, scratch, src1); 541 src1 = scratch; 542 } 543 } 544 tcg_out_movext1_new_src(s, i2, src2); 545 tcg_out_movext1_new_src(s, i1, src1); 546 } 547 548 /** 549 * tcg_out_movext3 -- move and extend three pair 550 * @s: tcg context 551 * @i1: first move description 552 * @i2: second move description 553 * @i3: third move description 554 * @scratch: temporary register, or -1 for none 555 * 556 * As tcg_out_movext, for all of @i1, @i2 and @i3, caring for overlap 557 * between the sources and destinations. 558 */ 559 560 static void tcg_out_movext3(TCGContext *s, const TCGMovExtend *i1, 561 const TCGMovExtend *i2, const TCGMovExtend *i3, 562 int scratch) 563 { 564 TCGReg src1 = i1->src; 565 TCGReg src2 = i2->src; 566 TCGReg src3 = i3->src; 567 568 if (i1->dst != src2 && i1->dst != src3) { 569 tcg_out_movext1(s, i1); 570 tcg_out_movext2(s, i2, i3, scratch); 571 return; 572 } 573 if (i2->dst != src1 && i2->dst != src3) { 574 tcg_out_movext1(s, i2); 575 tcg_out_movext2(s, i1, i3, scratch); 576 return; 577 } 578 if (i3->dst != src1 && i3->dst != src2) { 579 tcg_out_movext1(s, i3); 580 tcg_out_movext2(s, i1, i2, scratch); 581 return; 582 } 583 584 /* 585 * There is a cycle. Since there are only 3 nodes, the cycle is 586 * either "clockwise" or "anti-clockwise", and can be solved with 587 * a single scratch or two xchg. 588 */ 589 if (i1->dst == src2 && i2->dst == src3 && i3->dst == src1) { 590 /* "Clockwise" */ 591 if (tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2)) { 592 tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3); 593 /* The data is now in the correct registers, now extend. */ 594 tcg_out_movext1_new_src(s, i1, i1->dst); 595 tcg_out_movext1_new_src(s, i2, i2->dst); 596 tcg_out_movext1_new_src(s, i3, i3->dst); 597 } else { 598 tcg_debug_assert(scratch >= 0); 599 tcg_out_mov(s, i1->src_type, scratch, src1); 600 tcg_out_movext1(s, i3); 601 tcg_out_movext1(s, i2); 602 tcg_out_movext1_new_src(s, i1, scratch); 603 } 604 } else if (i1->dst == src3 && i2->dst == src1 && i3->dst == src2) { 605 /* "Anti-clockwise" */ 606 if (tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3)) { 607 tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2); 608 /* The data is now in the correct registers, now extend. */ 609 tcg_out_movext1_new_src(s, i1, i1->dst); 610 tcg_out_movext1_new_src(s, i2, i2->dst); 611 tcg_out_movext1_new_src(s, i3, i3->dst); 612 } else { 613 tcg_debug_assert(scratch >= 0); 614 tcg_out_mov(s, i1->src_type, scratch, src1); 615 tcg_out_movext1(s, i2); 616 tcg_out_movext1(s, i3); 617 tcg_out_movext1_new_src(s, i1, scratch); 618 } 619 } else { 620 g_assert_not_reached(); 621 } 622 } 623 624 #define C_PFX1(P, A) P##A 625 #define C_PFX2(P, A, B) P##A##_##B 626 #define C_PFX3(P, A, B, C) P##A##_##B##_##C 627 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D 628 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E 629 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F 630 631 /* Define an enumeration for the various combinations. */ 632 633 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1), 634 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2), 635 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3), 636 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4), 637 638 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1), 639 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2), 640 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3), 641 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4), 642 643 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2), 644 645 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1), 646 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2), 647 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3), 648 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4), 649 650 typedef enum { 651 #include "tcg-target-con-set.h" 652 } TCGConstraintSetIndex; 653 654 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode); 655 656 #undef C_O0_I1 657 #undef C_O0_I2 658 #undef C_O0_I3 659 #undef C_O0_I4 660 #undef C_O1_I1 661 #undef C_O1_I2 662 #undef C_O1_I3 663 #undef C_O1_I4 664 #undef C_N1_I2 665 #undef C_O2_I1 666 #undef C_O2_I2 667 #undef C_O2_I3 668 #undef C_O2_I4 669 670 /* Put all of the constraint sets into an array, indexed by the enum. */ 671 672 #define C_O0_I1(I1) { .args_ct_str = { #I1 } }, 673 #define C_O0_I2(I1, I2) { .args_ct_str = { #I1, #I2 } }, 674 #define C_O0_I3(I1, I2, I3) { .args_ct_str = { #I1, #I2, #I3 } }, 675 #define C_O0_I4(I1, I2, I3, I4) { .args_ct_str = { #I1, #I2, #I3, #I4 } }, 676 677 #define C_O1_I1(O1, I1) { .args_ct_str = { #O1, #I1 } }, 678 #define C_O1_I2(O1, I1, I2) { .args_ct_str = { #O1, #I1, #I2 } }, 679 #define C_O1_I3(O1, I1, I2, I3) { .args_ct_str = { #O1, #I1, #I2, #I3 } }, 680 #define C_O1_I4(O1, I1, I2, I3, I4) { .args_ct_str = { #O1, #I1, #I2, #I3, #I4 } }, 681 682 #define C_N1_I2(O1, I1, I2) { .args_ct_str = { "&" #O1, #I1, #I2 } }, 683 684 #define C_O2_I1(O1, O2, I1) { .args_ct_str = { #O1, #O2, #I1 } }, 685 #define C_O2_I2(O1, O2, I1, I2) { .args_ct_str = { #O1, #O2, #I1, #I2 } }, 686 #define C_O2_I3(O1, O2, I1, I2, I3) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3 } }, 687 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3, #I4 } }, 688 689 static const TCGTargetOpDef constraint_sets[] = { 690 #include "tcg-target-con-set.h" 691 }; 692 693 694 #undef C_O0_I1 695 #undef C_O0_I2 696 #undef C_O0_I3 697 #undef C_O0_I4 698 #undef C_O1_I1 699 #undef C_O1_I2 700 #undef C_O1_I3 701 #undef C_O1_I4 702 #undef C_N1_I2 703 #undef C_O2_I1 704 #undef C_O2_I2 705 #undef C_O2_I3 706 #undef C_O2_I4 707 708 /* Expand the enumerator to be returned from tcg_target_op_def(). */ 709 710 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1) 711 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2) 712 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3) 713 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4) 714 715 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1) 716 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2) 717 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3) 718 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4) 719 720 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2) 721 722 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1) 723 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2) 724 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3) 725 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4) 726 727 #include "tcg-target.c.inc" 728 729 static void alloc_tcg_plugin_context(TCGContext *s) 730 { 731 #ifdef CONFIG_PLUGIN 732 s->plugin_tb = g_new0(struct qemu_plugin_tb, 1); 733 s->plugin_tb->insns = 734 g_ptr_array_new_with_free_func(qemu_plugin_insn_cleanup_fn); 735 #endif 736 } 737 738 /* 739 * All TCG threads except the parent (i.e. the one that called tcg_context_init 740 * and registered the target's TCG globals) must register with this function 741 * before initiating translation. 742 * 743 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation 744 * of tcg_region_init() for the reasoning behind this. 745 * 746 * In softmmu each caller registers its context in tcg_ctxs[]. Note that in 747 * softmmu tcg_ctxs[] does not track tcg_ctx_init, since the initial context 748 * is not used anymore for translation once this function is called. 749 * 750 * Not tracking tcg_init_ctx in tcg_ctxs[] in softmmu keeps code that iterates 751 * over the array (e.g. tcg_code_size() the same for both softmmu and user-mode. 752 */ 753 #ifdef CONFIG_USER_ONLY 754 void tcg_register_thread(void) 755 { 756 tcg_ctx = &tcg_init_ctx; 757 } 758 #else 759 void tcg_register_thread(void) 760 { 761 TCGContext *s = g_malloc(sizeof(*s)); 762 unsigned int i, n; 763 764 *s = tcg_init_ctx; 765 766 /* Relink mem_base. */ 767 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) { 768 if (tcg_init_ctx.temps[i].mem_base) { 769 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps; 770 tcg_debug_assert(b >= 0 && b < n); 771 s->temps[i].mem_base = &s->temps[b]; 772 } 773 } 774 775 /* Claim an entry in tcg_ctxs */ 776 n = qatomic_fetch_inc(&tcg_cur_ctxs); 777 g_assert(n < tcg_max_ctxs); 778 qatomic_set(&tcg_ctxs[n], s); 779 780 if (n > 0) { 781 alloc_tcg_plugin_context(s); 782 tcg_region_initial_alloc(s); 783 } 784 785 tcg_ctx = s; 786 } 787 #endif /* !CONFIG_USER_ONLY */ 788 789 /* pool based memory allocation */ 790 void *tcg_malloc_internal(TCGContext *s, int size) 791 { 792 TCGPool *p; 793 int pool_size; 794 795 if (size > TCG_POOL_CHUNK_SIZE) { 796 /* big malloc: insert a new pool (XXX: could optimize) */ 797 p = g_malloc(sizeof(TCGPool) + size); 798 p->size = size; 799 p->next = s->pool_first_large; 800 s->pool_first_large = p; 801 return p->data; 802 } else { 803 p = s->pool_current; 804 if (!p) { 805 p = s->pool_first; 806 if (!p) 807 goto new_pool; 808 } else { 809 if (!p->next) { 810 new_pool: 811 pool_size = TCG_POOL_CHUNK_SIZE; 812 p = g_malloc(sizeof(TCGPool) + pool_size); 813 p->size = pool_size; 814 p->next = NULL; 815 if (s->pool_current) { 816 s->pool_current->next = p; 817 } else { 818 s->pool_first = p; 819 } 820 } else { 821 p = p->next; 822 } 823 } 824 } 825 s->pool_current = p; 826 s->pool_cur = p->data + size; 827 s->pool_end = p->data + p->size; 828 return p->data; 829 } 830 831 void tcg_pool_reset(TCGContext *s) 832 { 833 TCGPool *p, *t; 834 for (p = s->pool_first_large; p; p = t) { 835 t = p->next; 836 g_free(p); 837 } 838 s->pool_first_large = NULL; 839 s->pool_cur = s->pool_end = NULL; 840 s->pool_current = NULL; 841 } 842 843 #include "exec/helper-proto.h" 844 845 static TCGHelperInfo all_helpers[] = { 846 #include "exec/helper-tcg.h" 847 }; 848 static GHashTable *helper_table; 849 850 /* 851 * Create TCGHelperInfo structures for "tcg/tcg-ldst.h" functions, 852 * akin to what "exec/helper-tcg.h" does with DEF_HELPER_FLAGS_N. 853 * We only use these for layout in tcg_out_ld_helper_ret and 854 * tcg_out_st_helper_args, and share them between several of 855 * the helpers, with the end result that it's easier to build manually. 856 */ 857 858 #if TCG_TARGET_REG_BITS == 32 859 # define dh_typecode_ttl dh_typecode_i32 860 #else 861 # define dh_typecode_ttl dh_typecode_i64 862 #endif 863 864 static TCGHelperInfo info_helper_ld32_mmu = { 865 .flags = TCG_CALL_NO_WG, 866 .typemask = dh_typemask(ttl, 0) /* return tcg_target_ulong */ 867 | dh_typemask(env, 1) 868 | dh_typemask(i64, 2) /* uint64_t addr */ 869 | dh_typemask(i32, 3) /* unsigned oi */ 870 | dh_typemask(ptr, 4) /* uintptr_t ra */ 871 }; 872 873 static TCGHelperInfo info_helper_ld64_mmu = { 874 .flags = TCG_CALL_NO_WG, 875 .typemask = dh_typemask(i64, 0) /* return uint64_t */ 876 | dh_typemask(env, 1) 877 | dh_typemask(i64, 2) /* uint64_t addr */ 878 | dh_typemask(i32, 3) /* unsigned oi */ 879 | dh_typemask(ptr, 4) /* uintptr_t ra */ 880 }; 881 882 static TCGHelperInfo info_helper_ld128_mmu = { 883 .flags = TCG_CALL_NO_WG, 884 .typemask = dh_typemask(i128, 0) /* return Int128 */ 885 | dh_typemask(env, 1) 886 | dh_typemask(i64, 2) /* uint64_t addr */ 887 | dh_typemask(i32, 3) /* unsigned oi */ 888 | dh_typemask(ptr, 4) /* uintptr_t ra */ 889 }; 890 891 static TCGHelperInfo info_helper_st32_mmu = { 892 .flags = TCG_CALL_NO_WG, 893 .typemask = dh_typemask(void, 0) 894 | dh_typemask(env, 1) 895 | dh_typemask(i64, 2) /* uint64_t addr */ 896 | dh_typemask(i32, 3) /* uint32_t data */ 897 | dh_typemask(i32, 4) /* unsigned oi */ 898 | dh_typemask(ptr, 5) /* uintptr_t ra */ 899 }; 900 901 static TCGHelperInfo info_helper_st64_mmu = { 902 .flags = TCG_CALL_NO_WG, 903 .typemask = dh_typemask(void, 0) 904 | dh_typemask(env, 1) 905 | dh_typemask(i64, 2) /* uint64_t addr */ 906 | dh_typemask(i64, 3) /* uint64_t data */ 907 | dh_typemask(i32, 4) /* unsigned oi */ 908 | dh_typemask(ptr, 5) /* uintptr_t ra */ 909 }; 910 911 static TCGHelperInfo info_helper_st128_mmu = { 912 .flags = TCG_CALL_NO_WG, 913 .typemask = dh_typemask(void, 0) 914 | dh_typemask(env, 1) 915 | dh_typemask(i64, 2) /* uint64_t addr */ 916 | dh_typemask(i128, 3) /* Int128 data */ 917 | dh_typemask(i32, 4) /* unsigned oi */ 918 | dh_typemask(ptr, 5) /* uintptr_t ra */ 919 }; 920 921 #ifdef CONFIG_TCG_INTERPRETER 922 static ffi_type *typecode_to_ffi(int argmask) 923 { 924 /* 925 * libffi does not support __int128_t, so we have forced Int128 926 * to use the structure definition instead of the builtin type. 927 */ 928 static ffi_type *ffi_type_i128_elements[3] = { 929 &ffi_type_uint64, 930 &ffi_type_uint64, 931 NULL 932 }; 933 static ffi_type ffi_type_i128 = { 934 .size = 16, 935 .alignment = __alignof__(Int128), 936 .type = FFI_TYPE_STRUCT, 937 .elements = ffi_type_i128_elements, 938 }; 939 940 switch (argmask) { 941 case dh_typecode_void: 942 return &ffi_type_void; 943 case dh_typecode_i32: 944 return &ffi_type_uint32; 945 case dh_typecode_s32: 946 return &ffi_type_sint32; 947 case dh_typecode_i64: 948 return &ffi_type_uint64; 949 case dh_typecode_s64: 950 return &ffi_type_sint64; 951 case dh_typecode_ptr: 952 return &ffi_type_pointer; 953 case dh_typecode_i128: 954 return &ffi_type_i128; 955 } 956 g_assert_not_reached(); 957 } 958 959 static void init_ffi_layouts(void) 960 { 961 /* g_direct_hash/equal for direct comparisons on uint32_t. */ 962 GHashTable *ffi_table = g_hash_table_new(NULL, NULL); 963 964 for (int i = 0; i < ARRAY_SIZE(all_helpers); ++i) { 965 TCGHelperInfo *info = &all_helpers[i]; 966 unsigned typemask = info->typemask; 967 gpointer hash = (gpointer)(uintptr_t)typemask; 968 struct { 969 ffi_cif cif; 970 ffi_type *args[]; 971 } *ca; 972 ffi_status status; 973 int nargs; 974 ffi_cif *cif; 975 976 cif = g_hash_table_lookup(ffi_table, hash); 977 if (cif) { 978 info->cif = cif; 979 continue; 980 } 981 982 /* Ignoring the return type, find the last non-zero field. */ 983 nargs = 32 - clz32(typemask >> 3); 984 nargs = DIV_ROUND_UP(nargs, 3); 985 assert(nargs <= MAX_CALL_IARGS); 986 987 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *)); 988 ca->cif.rtype = typecode_to_ffi(typemask & 7); 989 ca->cif.nargs = nargs; 990 991 if (nargs != 0) { 992 ca->cif.arg_types = ca->args; 993 for (int j = 0; j < nargs; ++j) { 994 int typecode = extract32(typemask, (j + 1) * 3, 3); 995 ca->args[j] = typecode_to_ffi(typecode); 996 } 997 } 998 999 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs, 1000 ca->cif.rtype, ca->cif.arg_types); 1001 assert(status == FFI_OK); 1002 1003 cif = &ca->cif; 1004 info->cif = cif; 1005 g_hash_table_insert(ffi_table, hash, (gpointer)cif); 1006 } 1007 1008 g_hash_table_destroy(ffi_table); 1009 } 1010 #endif /* CONFIG_TCG_INTERPRETER */ 1011 1012 static inline bool arg_slot_reg_p(unsigned arg_slot) 1013 { 1014 /* 1015 * Split the sizeof away from the comparison to avoid Werror from 1016 * "unsigned < 0 is always false", when iarg_regs is empty. 1017 */ 1018 unsigned nreg = ARRAY_SIZE(tcg_target_call_iarg_regs); 1019 return arg_slot < nreg; 1020 } 1021 1022 static inline int arg_slot_stk_ofs(unsigned arg_slot) 1023 { 1024 unsigned max = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1025 unsigned stk_slot = arg_slot - ARRAY_SIZE(tcg_target_call_iarg_regs); 1026 1027 tcg_debug_assert(stk_slot < max); 1028 return TCG_TARGET_CALL_STACK_OFFSET + stk_slot * sizeof(tcg_target_long); 1029 } 1030 1031 typedef struct TCGCumulativeArgs { 1032 int arg_idx; /* tcg_gen_callN args[] */ 1033 int info_in_idx; /* TCGHelperInfo in[] */ 1034 int arg_slot; /* regs+stack slot */ 1035 int ref_slot; /* stack slots for references */ 1036 } TCGCumulativeArgs; 1037 1038 static void layout_arg_even(TCGCumulativeArgs *cum) 1039 { 1040 cum->arg_slot += cum->arg_slot & 1; 1041 } 1042 1043 static void layout_arg_1(TCGCumulativeArgs *cum, TCGHelperInfo *info, 1044 TCGCallArgumentKind kind) 1045 { 1046 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1047 1048 *loc = (TCGCallArgumentLoc){ 1049 .kind = kind, 1050 .arg_idx = cum->arg_idx, 1051 .arg_slot = cum->arg_slot, 1052 }; 1053 cum->info_in_idx++; 1054 cum->arg_slot++; 1055 } 1056 1057 static void layout_arg_normal_n(TCGCumulativeArgs *cum, 1058 TCGHelperInfo *info, int n) 1059 { 1060 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1061 1062 for (int i = 0; i < n; ++i) { 1063 /* Layout all using the same arg_idx, adjusting the subindex. */ 1064 loc[i] = (TCGCallArgumentLoc){ 1065 .kind = TCG_CALL_ARG_NORMAL, 1066 .arg_idx = cum->arg_idx, 1067 .tmp_subindex = i, 1068 .arg_slot = cum->arg_slot + i, 1069 }; 1070 } 1071 cum->info_in_idx += n; 1072 cum->arg_slot += n; 1073 } 1074 1075 static void layout_arg_by_ref(TCGCumulativeArgs *cum, TCGHelperInfo *info) 1076 { 1077 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1078 int n = 128 / TCG_TARGET_REG_BITS; 1079 1080 /* The first subindex carries the pointer. */ 1081 layout_arg_1(cum, info, TCG_CALL_ARG_BY_REF); 1082 1083 /* 1084 * The callee is allowed to clobber memory associated with 1085 * structure pass by-reference. Therefore we must make copies. 1086 * Allocate space from "ref_slot", which will be adjusted to 1087 * follow the parameters on the stack. 1088 */ 1089 loc[0].ref_slot = cum->ref_slot; 1090 1091 /* 1092 * Subsequent words also go into the reference slot, but 1093 * do not accumulate into the regular arguments. 1094 */ 1095 for (int i = 1; i < n; ++i) { 1096 loc[i] = (TCGCallArgumentLoc){ 1097 .kind = TCG_CALL_ARG_BY_REF_N, 1098 .arg_idx = cum->arg_idx, 1099 .tmp_subindex = i, 1100 .ref_slot = cum->ref_slot + i, 1101 }; 1102 } 1103 cum->info_in_idx += n; 1104 cum->ref_slot += n; 1105 } 1106 1107 static void init_call_layout(TCGHelperInfo *info) 1108 { 1109 int max_reg_slots = ARRAY_SIZE(tcg_target_call_iarg_regs); 1110 int max_stk_slots = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1111 unsigned typemask = info->typemask; 1112 unsigned typecode; 1113 TCGCumulativeArgs cum = { }; 1114 1115 /* 1116 * Parse and place any function return value. 1117 */ 1118 typecode = typemask & 7; 1119 switch (typecode) { 1120 case dh_typecode_void: 1121 info->nr_out = 0; 1122 break; 1123 case dh_typecode_i32: 1124 case dh_typecode_s32: 1125 case dh_typecode_ptr: 1126 info->nr_out = 1; 1127 info->out_kind = TCG_CALL_RET_NORMAL; 1128 break; 1129 case dh_typecode_i64: 1130 case dh_typecode_s64: 1131 info->nr_out = 64 / TCG_TARGET_REG_BITS; 1132 info->out_kind = TCG_CALL_RET_NORMAL; 1133 /* Query the last register now to trigger any assert early. */ 1134 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1135 break; 1136 case dh_typecode_i128: 1137 info->nr_out = 128 / TCG_TARGET_REG_BITS; 1138 info->out_kind = TCG_TARGET_CALL_RET_I128; 1139 switch (TCG_TARGET_CALL_RET_I128) { 1140 case TCG_CALL_RET_NORMAL: 1141 /* Query the last register now to trigger any assert early. */ 1142 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1143 break; 1144 case TCG_CALL_RET_BY_VEC: 1145 /* Query the single register now to trigger any assert early. */ 1146 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0); 1147 break; 1148 case TCG_CALL_RET_BY_REF: 1149 /* 1150 * Allocate the first argument to the output. 1151 * We don't need to store this anywhere, just make it 1152 * unavailable for use in the input loop below. 1153 */ 1154 cum.arg_slot = 1; 1155 break; 1156 default: 1157 qemu_build_not_reached(); 1158 } 1159 break; 1160 default: 1161 g_assert_not_reached(); 1162 } 1163 1164 /* 1165 * Parse and place function arguments. 1166 */ 1167 for (typemask >>= 3; typemask; typemask >>= 3, cum.arg_idx++) { 1168 TCGCallArgumentKind kind; 1169 TCGType type; 1170 1171 typecode = typemask & 7; 1172 switch (typecode) { 1173 case dh_typecode_i32: 1174 case dh_typecode_s32: 1175 type = TCG_TYPE_I32; 1176 break; 1177 case dh_typecode_i64: 1178 case dh_typecode_s64: 1179 type = TCG_TYPE_I64; 1180 break; 1181 case dh_typecode_ptr: 1182 type = TCG_TYPE_PTR; 1183 break; 1184 case dh_typecode_i128: 1185 type = TCG_TYPE_I128; 1186 break; 1187 default: 1188 g_assert_not_reached(); 1189 } 1190 1191 switch (type) { 1192 case TCG_TYPE_I32: 1193 switch (TCG_TARGET_CALL_ARG_I32) { 1194 case TCG_CALL_ARG_EVEN: 1195 layout_arg_even(&cum); 1196 /* fall through */ 1197 case TCG_CALL_ARG_NORMAL: 1198 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1199 break; 1200 case TCG_CALL_ARG_EXTEND: 1201 kind = TCG_CALL_ARG_EXTEND_U + (typecode & 1); 1202 layout_arg_1(&cum, info, kind); 1203 break; 1204 default: 1205 qemu_build_not_reached(); 1206 } 1207 break; 1208 1209 case TCG_TYPE_I64: 1210 switch (TCG_TARGET_CALL_ARG_I64) { 1211 case TCG_CALL_ARG_EVEN: 1212 layout_arg_even(&cum); 1213 /* fall through */ 1214 case TCG_CALL_ARG_NORMAL: 1215 if (TCG_TARGET_REG_BITS == 32) { 1216 layout_arg_normal_n(&cum, info, 2); 1217 } else { 1218 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1219 } 1220 break; 1221 default: 1222 qemu_build_not_reached(); 1223 } 1224 break; 1225 1226 case TCG_TYPE_I128: 1227 switch (TCG_TARGET_CALL_ARG_I128) { 1228 case TCG_CALL_ARG_EVEN: 1229 layout_arg_even(&cum); 1230 /* fall through */ 1231 case TCG_CALL_ARG_NORMAL: 1232 layout_arg_normal_n(&cum, info, 128 / TCG_TARGET_REG_BITS); 1233 break; 1234 case TCG_CALL_ARG_BY_REF: 1235 layout_arg_by_ref(&cum, info); 1236 break; 1237 default: 1238 qemu_build_not_reached(); 1239 } 1240 break; 1241 1242 default: 1243 g_assert_not_reached(); 1244 } 1245 } 1246 info->nr_in = cum.info_in_idx; 1247 1248 /* Validate that we didn't overrun the input array. */ 1249 assert(cum.info_in_idx <= ARRAY_SIZE(info->in)); 1250 /* Validate the backend has enough argument space. */ 1251 assert(cum.arg_slot <= max_reg_slots + max_stk_slots); 1252 1253 /* 1254 * Relocate the "ref_slot" area to the end of the parameters. 1255 * Minimizing this stack offset helps code size for x86, 1256 * which has a signed 8-bit offset encoding. 1257 */ 1258 if (cum.ref_slot != 0) { 1259 int ref_base = 0; 1260 1261 if (cum.arg_slot > max_reg_slots) { 1262 int align = __alignof(Int128) / sizeof(tcg_target_long); 1263 1264 ref_base = cum.arg_slot - max_reg_slots; 1265 if (align > 1) { 1266 ref_base = ROUND_UP(ref_base, align); 1267 } 1268 } 1269 assert(ref_base + cum.ref_slot <= max_stk_slots); 1270 ref_base += max_reg_slots; 1271 1272 if (ref_base != 0) { 1273 for (int i = cum.info_in_idx - 1; i >= 0; --i) { 1274 TCGCallArgumentLoc *loc = &info->in[i]; 1275 switch (loc->kind) { 1276 case TCG_CALL_ARG_BY_REF: 1277 case TCG_CALL_ARG_BY_REF_N: 1278 loc->ref_slot += ref_base; 1279 break; 1280 default: 1281 break; 1282 } 1283 } 1284 } 1285 } 1286 } 1287 1288 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)]; 1289 static void process_op_defs(TCGContext *s); 1290 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1291 TCGReg reg, const char *name); 1292 1293 static void tcg_context_init(unsigned max_cpus) 1294 { 1295 TCGContext *s = &tcg_init_ctx; 1296 int op, total_args, n, i; 1297 TCGOpDef *def; 1298 TCGArgConstraint *args_ct; 1299 TCGTemp *ts; 1300 1301 memset(s, 0, sizeof(*s)); 1302 s->nb_globals = 0; 1303 1304 /* Count total number of arguments and allocate the corresponding 1305 space */ 1306 total_args = 0; 1307 for(op = 0; op < NB_OPS; op++) { 1308 def = &tcg_op_defs[op]; 1309 n = def->nb_iargs + def->nb_oargs; 1310 total_args += n; 1311 } 1312 1313 args_ct = g_new0(TCGArgConstraint, total_args); 1314 1315 for(op = 0; op < NB_OPS; op++) { 1316 def = &tcg_op_defs[op]; 1317 def->args_ct = args_ct; 1318 n = def->nb_iargs + def->nb_oargs; 1319 args_ct += n; 1320 } 1321 1322 /* Register helpers. */ 1323 /* Use g_direct_hash/equal for direct pointer comparisons on func. */ 1324 helper_table = g_hash_table_new(NULL, NULL); 1325 1326 for (i = 0; i < ARRAY_SIZE(all_helpers); ++i) { 1327 init_call_layout(&all_helpers[i]); 1328 g_hash_table_insert(helper_table, (gpointer)all_helpers[i].func, 1329 (gpointer)&all_helpers[i]); 1330 } 1331 1332 init_call_layout(&info_helper_ld32_mmu); 1333 init_call_layout(&info_helper_ld64_mmu); 1334 init_call_layout(&info_helper_ld128_mmu); 1335 init_call_layout(&info_helper_st32_mmu); 1336 init_call_layout(&info_helper_st64_mmu); 1337 init_call_layout(&info_helper_st128_mmu); 1338 1339 #ifdef CONFIG_TCG_INTERPRETER 1340 init_ffi_layouts(); 1341 #endif 1342 1343 tcg_target_init(s); 1344 process_op_defs(s); 1345 1346 /* Reverse the order of the saved registers, assuming they're all at 1347 the start of tcg_target_reg_alloc_order. */ 1348 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) { 1349 int r = tcg_target_reg_alloc_order[n]; 1350 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) { 1351 break; 1352 } 1353 } 1354 for (i = 0; i < n; ++i) { 1355 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i]; 1356 } 1357 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) { 1358 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i]; 1359 } 1360 1361 alloc_tcg_plugin_context(s); 1362 1363 tcg_ctx = s; 1364 /* 1365 * In user-mode we simply share the init context among threads, since we 1366 * use a single region. See the documentation tcg_region_init() for the 1367 * reasoning behind this. 1368 * In softmmu we will have at most max_cpus TCG threads. 1369 */ 1370 #ifdef CONFIG_USER_ONLY 1371 tcg_ctxs = &tcg_ctx; 1372 tcg_cur_ctxs = 1; 1373 tcg_max_ctxs = 1; 1374 #else 1375 tcg_max_ctxs = max_cpus; 1376 tcg_ctxs = g_new0(TCGContext *, max_cpus); 1377 #endif 1378 1379 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0)); 1380 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env"); 1381 cpu_env = temp_tcgv_ptr(ts); 1382 } 1383 1384 void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus) 1385 { 1386 tcg_context_init(max_cpus); 1387 tcg_region_init(tb_size, splitwx, max_cpus); 1388 } 1389 1390 /* 1391 * Allocate TBs right before their corresponding translated code, making 1392 * sure that TBs and code are on different cache lines. 1393 */ 1394 TranslationBlock *tcg_tb_alloc(TCGContext *s) 1395 { 1396 uintptr_t align = qemu_icache_linesize; 1397 TranslationBlock *tb; 1398 void *next; 1399 1400 retry: 1401 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align); 1402 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align); 1403 1404 if (unlikely(next > s->code_gen_highwater)) { 1405 if (tcg_region_alloc(s)) { 1406 return NULL; 1407 } 1408 goto retry; 1409 } 1410 qatomic_set(&s->code_gen_ptr, next); 1411 s->data_gen_ptr = NULL; 1412 return tb; 1413 } 1414 1415 void tcg_prologue_init(TCGContext *s) 1416 { 1417 size_t prologue_size; 1418 1419 s->code_ptr = s->code_gen_ptr; 1420 s->code_buf = s->code_gen_ptr; 1421 s->data_gen_ptr = NULL; 1422 1423 #ifndef CONFIG_TCG_INTERPRETER 1424 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr); 1425 #endif 1426 1427 #ifdef TCG_TARGET_NEED_POOL_LABELS 1428 s->pool_labels = NULL; 1429 #endif 1430 1431 qemu_thread_jit_write(); 1432 /* Generate the prologue. */ 1433 tcg_target_qemu_prologue(s); 1434 1435 #ifdef TCG_TARGET_NEED_POOL_LABELS 1436 /* Allow the prologue to put e.g. guest_base into a pool entry. */ 1437 { 1438 int result = tcg_out_pool_finalize(s); 1439 tcg_debug_assert(result == 0); 1440 } 1441 #endif 1442 1443 prologue_size = tcg_current_code_size(s); 1444 perf_report_prologue(s->code_gen_ptr, prologue_size); 1445 1446 #ifndef CONFIG_TCG_INTERPRETER 1447 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 1448 (uintptr_t)s->code_buf, prologue_size); 1449 #endif 1450 1451 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { 1452 FILE *logfile = qemu_log_trylock(); 1453 if (logfile) { 1454 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size); 1455 if (s->data_gen_ptr) { 1456 size_t code_size = s->data_gen_ptr - s->code_gen_ptr; 1457 size_t data_size = prologue_size - code_size; 1458 size_t i; 1459 1460 disas(logfile, s->code_gen_ptr, code_size); 1461 1462 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) { 1463 if (sizeof(tcg_target_ulong) == 8) { 1464 fprintf(logfile, 1465 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n", 1466 (uintptr_t)s->data_gen_ptr + i, 1467 *(uint64_t *)(s->data_gen_ptr + i)); 1468 } else { 1469 fprintf(logfile, 1470 "0x%08" PRIxPTR ": .long 0x%08x\n", 1471 (uintptr_t)s->data_gen_ptr + i, 1472 *(uint32_t *)(s->data_gen_ptr + i)); 1473 } 1474 } 1475 } else { 1476 disas(logfile, s->code_gen_ptr, prologue_size); 1477 } 1478 fprintf(logfile, "\n"); 1479 qemu_log_unlock(logfile); 1480 } 1481 } 1482 1483 #ifndef CONFIG_TCG_INTERPRETER 1484 /* 1485 * Assert that goto_ptr is implemented completely, setting an epilogue. 1486 * For tci, we use NULL as the signal to return from the interpreter, 1487 * so skip this check. 1488 */ 1489 tcg_debug_assert(tcg_code_gen_epilogue != NULL); 1490 #endif 1491 1492 tcg_region_prologue_set(s); 1493 } 1494 1495 void tcg_func_start(TCGContext *s) 1496 { 1497 tcg_pool_reset(s); 1498 s->nb_temps = s->nb_globals; 1499 1500 /* No temps have been previously allocated for size or locality. */ 1501 memset(s->free_temps, 0, sizeof(s->free_temps)); 1502 1503 /* No constant temps have been previously allocated. */ 1504 for (int i = 0; i < TCG_TYPE_COUNT; ++i) { 1505 if (s->const_table[i]) { 1506 g_hash_table_remove_all(s->const_table[i]); 1507 } 1508 } 1509 1510 s->nb_ops = 0; 1511 s->nb_labels = 0; 1512 s->current_frame_offset = s->frame_start; 1513 1514 #ifdef CONFIG_DEBUG_TCG 1515 s->goto_tb_issue_mask = 0; 1516 #endif 1517 1518 QTAILQ_INIT(&s->ops); 1519 QTAILQ_INIT(&s->free_ops); 1520 QSIMPLEQ_INIT(&s->labels); 1521 1522 tcg_debug_assert(s->addr_type == TCG_TYPE_I32 || 1523 s->addr_type == TCG_TYPE_I64); 1524 } 1525 1526 static TCGTemp *tcg_temp_alloc(TCGContext *s) 1527 { 1528 int n = s->nb_temps++; 1529 1530 if (n >= TCG_MAX_TEMPS) { 1531 tcg_raise_tb_overflow(s); 1532 } 1533 return memset(&s->temps[n], 0, sizeof(TCGTemp)); 1534 } 1535 1536 static TCGTemp *tcg_global_alloc(TCGContext *s) 1537 { 1538 TCGTemp *ts; 1539 1540 tcg_debug_assert(s->nb_globals == s->nb_temps); 1541 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS); 1542 s->nb_globals++; 1543 ts = tcg_temp_alloc(s); 1544 ts->kind = TEMP_GLOBAL; 1545 1546 return ts; 1547 } 1548 1549 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1550 TCGReg reg, const char *name) 1551 { 1552 TCGTemp *ts; 1553 1554 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32); 1555 1556 ts = tcg_global_alloc(s); 1557 ts->base_type = type; 1558 ts->type = type; 1559 ts->kind = TEMP_FIXED; 1560 ts->reg = reg; 1561 ts->name = name; 1562 tcg_regset_set_reg(s->reserved_regs, reg); 1563 1564 return ts; 1565 } 1566 1567 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size) 1568 { 1569 s->frame_start = start; 1570 s->frame_end = start + size; 1571 s->frame_temp 1572 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame"); 1573 } 1574 1575 TCGTemp *tcg_global_mem_new_internal(TCGType type, TCGv_ptr base, 1576 intptr_t offset, const char *name) 1577 { 1578 TCGContext *s = tcg_ctx; 1579 TCGTemp *base_ts = tcgv_ptr_temp(base); 1580 TCGTemp *ts = tcg_global_alloc(s); 1581 int indirect_reg = 0; 1582 1583 switch (base_ts->kind) { 1584 case TEMP_FIXED: 1585 break; 1586 case TEMP_GLOBAL: 1587 /* We do not support double-indirect registers. */ 1588 tcg_debug_assert(!base_ts->indirect_reg); 1589 base_ts->indirect_base = 1; 1590 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64 1591 ? 2 : 1); 1592 indirect_reg = 1; 1593 break; 1594 default: 1595 g_assert_not_reached(); 1596 } 1597 1598 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1599 TCGTemp *ts2 = tcg_global_alloc(s); 1600 char buf[64]; 1601 1602 ts->base_type = TCG_TYPE_I64; 1603 ts->type = TCG_TYPE_I32; 1604 ts->indirect_reg = indirect_reg; 1605 ts->mem_allocated = 1; 1606 ts->mem_base = base_ts; 1607 ts->mem_offset = offset; 1608 pstrcpy(buf, sizeof(buf), name); 1609 pstrcat(buf, sizeof(buf), "_0"); 1610 ts->name = strdup(buf); 1611 1612 tcg_debug_assert(ts2 == ts + 1); 1613 ts2->base_type = TCG_TYPE_I64; 1614 ts2->type = TCG_TYPE_I32; 1615 ts2->indirect_reg = indirect_reg; 1616 ts2->mem_allocated = 1; 1617 ts2->mem_base = base_ts; 1618 ts2->mem_offset = offset + 4; 1619 ts2->temp_subindex = 1; 1620 pstrcpy(buf, sizeof(buf), name); 1621 pstrcat(buf, sizeof(buf), "_1"); 1622 ts2->name = strdup(buf); 1623 } else { 1624 ts->base_type = type; 1625 ts->type = type; 1626 ts->indirect_reg = indirect_reg; 1627 ts->mem_allocated = 1; 1628 ts->mem_base = base_ts; 1629 ts->mem_offset = offset; 1630 ts->name = name; 1631 } 1632 return ts; 1633 } 1634 1635 TCGTemp *tcg_temp_new_internal(TCGType type, TCGTempKind kind) 1636 { 1637 TCGContext *s = tcg_ctx; 1638 TCGTemp *ts; 1639 int n; 1640 1641 if (kind == TEMP_EBB) { 1642 int idx = find_first_bit(s->free_temps[type].l, TCG_MAX_TEMPS); 1643 1644 if (idx < TCG_MAX_TEMPS) { 1645 /* There is already an available temp with the right type. */ 1646 clear_bit(idx, s->free_temps[type].l); 1647 1648 ts = &s->temps[idx]; 1649 ts->temp_allocated = 1; 1650 tcg_debug_assert(ts->base_type == type); 1651 tcg_debug_assert(ts->kind == kind); 1652 return ts; 1653 } 1654 } else { 1655 tcg_debug_assert(kind == TEMP_TB); 1656 } 1657 1658 switch (type) { 1659 case TCG_TYPE_I32: 1660 case TCG_TYPE_V64: 1661 case TCG_TYPE_V128: 1662 case TCG_TYPE_V256: 1663 n = 1; 1664 break; 1665 case TCG_TYPE_I64: 1666 n = 64 / TCG_TARGET_REG_BITS; 1667 break; 1668 case TCG_TYPE_I128: 1669 n = 128 / TCG_TARGET_REG_BITS; 1670 break; 1671 default: 1672 g_assert_not_reached(); 1673 } 1674 1675 ts = tcg_temp_alloc(s); 1676 ts->base_type = type; 1677 ts->temp_allocated = 1; 1678 ts->kind = kind; 1679 1680 if (n == 1) { 1681 ts->type = type; 1682 } else { 1683 ts->type = TCG_TYPE_REG; 1684 1685 for (int i = 1; i < n; ++i) { 1686 TCGTemp *ts2 = tcg_temp_alloc(s); 1687 1688 tcg_debug_assert(ts2 == ts + i); 1689 ts2->base_type = type; 1690 ts2->type = TCG_TYPE_REG; 1691 ts2->temp_allocated = 1; 1692 ts2->temp_subindex = i; 1693 ts2->kind = kind; 1694 } 1695 } 1696 return ts; 1697 } 1698 1699 TCGv_vec tcg_temp_new_vec(TCGType type) 1700 { 1701 TCGTemp *t; 1702 1703 #ifdef CONFIG_DEBUG_TCG 1704 switch (type) { 1705 case TCG_TYPE_V64: 1706 assert(TCG_TARGET_HAS_v64); 1707 break; 1708 case TCG_TYPE_V128: 1709 assert(TCG_TARGET_HAS_v128); 1710 break; 1711 case TCG_TYPE_V256: 1712 assert(TCG_TARGET_HAS_v256); 1713 break; 1714 default: 1715 g_assert_not_reached(); 1716 } 1717 #endif 1718 1719 t = tcg_temp_new_internal(type, TEMP_EBB); 1720 return temp_tcgv_vec(t); 1721 } 1722 1723 /* Create a new temp of the same type as an existing temp. */ 1724 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match) 1725 { 1726 TCGTemp *t = tcgv_vec_temp(match); 1727 1728 tcg_debug_assert(t->temp_allocated != 0); 1729 1730 t = tcg_temp_new_internal(t->base_type, TEMP_EBB); 1731 return temp_tcgv_vec(t); 1732 } 1733 1734 void tcg_temp_free_internal(TCGTemp *ts) 1735 { 1736 TCGContext *s = tcg_ctx; 1737 1738 switch (ts->kind) { 1739 case TEMP_CONST: 1740 case TEMP_TB: 1741 /* Silently ignore free. */ 1742 break; 1743 case TEMP_EBB: 1744 tcg_debug_assert(ts->temp_allocated != 0); 1745 ts->temp_allocated = 0; 1746 set_bit(temp_idx(ts), s->free_temps[ts->base_type].l); 1747 break; 1748 default: 1749 /* It never made sense to free TEMP_FIXED or TEMP_GLOBAL. */ 1750 g_assert_not_reached(); 1751 } 1752 } 1753 1754 TCGTemp *tcg_constant_internal(TCGType type, int64_t val) 1755 { 1756 TCGContext *s = tcg_ctx; 1757 GHashTable *h = s->const_table[type]; 1758 TCGTemp *ts; 1759 1760 if (h == NULL) { 1761 h = g_hash_table_new(g_int64_hash, g_int64_equal); 1762 s->const_table[type] = h; 1763 } 1764 1765 ts = g_hash_table_lookup(h, &val); 1766 if (ts == NULL) { 1767 int64_t *val_ptr; 1768 1769 ts = tcg_temp_alloc(s); 1770 1771 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1772 TCGTemp *ts2 = tcg_temp_alloc(s); 1773 1774 tcg_debug_assert(ts2 == ts + 1); 1775 1776 ts->base_type = TCG_TYPE_I64; 1777 ts->type = TCG_TYPE_I32; 1778 ts->kind = TEMP_CONST; 1779 ts->temp_allocated = 1; 1780 1781 ts2->base_type = TCG_TYPE_I64; 1782 ts2->type = TCG_TYPE_I32; 1783 ts2->kind = TEMP_CONST; 1784 ts2->temp_allocated = 1; 1785 ts2->temp_subindex = 1; 1786 1787 /* 1788 * Retain the full value of the 64-bit constant in the low 1789 * part, so that the hash table works. Actual uses will 1790 * truncate the value to the low part. 1791 */ 1792 ts[HOST_BIG_ENDIAN].val = val; 1793 ts[!HOST_BIG_ENDIAN].val = val >> 32; 1794 val_ptr = &ts[HOST_BIG_ENDIAN].val; 1795 } else { 1796 ts->base_type = type; 1797 ts->type = type; 1798 ts->kind = TEMP_CONST; 1799 ts->temp_allocated = 1; 1800 ts->val = val; 1801 val_ptr = &ts->val; 1802 } 1803 g_hash_table_insert(h, val_ptr, ts); 1804 } 1805 1806 return ts; 1807 } 1808 1809 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val) 1810 { 1811 val = dup_const(vece, val); 1812 return temp_tcgv_vec(tcg_constant_internal(type, val)); 1813 } 1814 1815 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val) 1816 { 1817 TCGTemp *t = tcgv_vec_temp(match); 1818 1819 tcg_debug_assert(t->temp_allocated != 0); 1820 return tcg_constant_vec(t->base_type, vece, val); 1821 } 1822 1823 /* Return true if OP may appear in the opcode stream. 1824 Test the runtime variable that controls each opcode. */ 1825 bool tcg_op_supported(TCGOpcode op) 1826 { 1827 const bool have_vec 1828 = TCG_TARGET_HAS_v64 | TCG_TARGET_HAS_v128 | TCG_TARGET_HAS_v256; 1829 1830 switch (op) { 1831 case INDEX_op_discard: 1832 case INDEX_op_set_label: 1833 case INDEX_op_call: 1834 case INDEX_op_br: 1835 case INDEX_op_mb: 1836 case INDEX_op_insn_start: 1837 case INDEX_op_exit_tb: 1838 case INDEX_op_goto_tb: 1839 case INDEX_op_goto_ptr: 1840 case INDEX_op_qemu_ld_a32_i32: 1841 case INDEX_op_qemu_ld_a64_i32: 1842 case INDEX_op_qemu_st_a32_i32: 1843 case INDEX_op_qemu_st_a64_i32: 1844 case INDEX_op_qemu_ld_a32_i64: 1845 case INDEX_op_qemu_ld_a64_i64: 1846 case INDEX_op_qemu_st_a32_i64: 1847 case INDEX_op_qemu_st_a64_i64: 1848 return true; 1849 1850 case INDEX_op_qemu_st8_a32_i32: 1851 case INDEX_op_qemu_st8_a64_i32: 1852 return TCG_TARGET_HAS_qemu_st8_i32; 1853 1854 case INDEX_op_qemu_ld_a32_i128: 1855 case INDEX_op_qemu_ld_a64_i128: 1856 case INDEX_op_qemu_st_a32_i128: 1857 case INDEX_op_qemu_st_a64_i128: 1858 return TCG_TARGET_HAS_qemu_ldst_i128; 1859 1860 case INDEX_op_mov_i32: 1861 case INDEX_op_setcond_i32: 1862 case INDEX_op_brcond_i32: 1863 case INDEX_op_ld8u_i32: 1864 case INDEX_op_ld8s_i32: 1865 case INDEX_op_ld16u_i32: 1866 case INDEX_op_ld16s_i32: 1867 case INDEX_op_ld_i32: 1868 case INDEX_op_st8_i32: 1869 case INDEX_op_st16_i32: 1870 case INDEX_op_st_i32: 1871 case INDEX_op_add_i32: 1872 case INDEX_op_sub_i32: 1873 case INDEX_op_mul_i32: 1874 case INDEX_op_and_i32: 1875 case INDEX_op_or_i32: 1876 case INDEX_op_xor_i32: 1877 case INDEX_op_shl_i32: 1878 case INDEX_op_shr_i32: 1879 case INDEX_op_sar_i32: 1880 return true; 1881 1882 case INDEX_op_movcond_i32: 1883 return TCG_TARGET_HAS_movcond_i32; 1884 case INDEX_op_div_i32: 1885 case INDEX_op_divu_i32: 1886 return TCG_TARGET_HAS_div_i32; 1887 case INDEX_op_rem_i32: 1888 case INDEX_op_remu_i32: 1889 return TCG_TARGET_HAS_rem_i32; 1890 case INDEX_op_div2_i32: 1891 case INDEX_op_divu2_i32: 1892 return TCG_TARGET_HAS_div2_i32; 1893 case INDEX_op_rotl_i32: 1894 case INDEX_op_rotr_i32: 1895 return TCG_TARGET_HAS_rot_i32; 1896 case INDEX_op_deposit_i32: 1897 return TCG_TARGET_HAS_deposit_i32; 1898 case INDEX_op_extract_i32: 1899 return TCG_TARGET_HAS_extract_i32; 1900 case INDEX_op_sextract_i32: 1901 return TCG_TARGET_HAS_sextract_i32; 1902 case INDEX_op_extract2_i32: 1903 return TCG_TARGET_HAS_extract2_i32; 1904 case INDEX_op_add2_i32: 1905 return TCG_TARGET_HAS_add2_i32; 1906 case INDEX_op_sub2_i32: 1907 return TCG_TARGET_HAS_sub2_i32; 1908 case INDEX_op_mulu2_i32: 1909 return TCG_TARGET_HAS_mulu2_i32; 1910 case INDEX_op_muls2_i32: 1911 return TCG_TARGET_HAS_muls2_i32; 1912 case INDEX_op_muluh_i32: 1913 return TCG_TARGET_HAS_muluh_i32; 1914 case INDEX_op_mulsh_i32: 1915 return TCG_TARGET_HAS_mulsh_i32; 1916 case INDEX_op_ext8s_i32: 1917 return TCG_TARGET_HAS_ext8s_i32; 1918 case INDEX_op_ext16s_i32: 1919 return TCG_TARGET_HAS_ext16s_i32; 1920 case INDEX_op_ext8u_i32: 1921 return TCG_TARGET_HAS_ext8u_i32; 1922 case INDEX_op_ext16u_i32: 1923 return TCG_TARGET_HAS_ext16u_i32; 1924 case INDEX_op_bswap16_i32: 1925 return TCG_TARGET_HAS_bswap16_i32; 1926 case INDEX_op_bswap32_i32: 1927 return TCG_TARGET_HAS_bswap32_i32; 1928 case INDEX_op_not_i32: 1929 return TCG_TARGET_HAS_not_i32; 1930 case INDEX_op_neg_i32: 1931 return TCG_TARGET_HAS_neg_i32; 1932 case INDEX_op_andc_i32: 1933 return TCG_TARGET_HAS_andc_i32; 1934 case INDEX_op_orc_i32: 1935 return TCG_TARGET_HAS_orc_i32; 1936 case INDEX_op_eqv_i32: 1937 return TCG_TARGET_HAS_eqv_i32; 1938 case INDEX_op_nand_i32: 1939 return TCG_TARGET_HAS_nand_i32; 1940 case INDEX_op_nor_i32: 1941 return TCG_TARGET_HAS_nor_i32; 1942 case INDEX_op_clz_i32: 1943 return TCG_TARGET_HAS_clz_i32; 1944 case INDEX_op_ctz_i32: 1945 return TCG_TARGET_HAS_ctz_i32; 1946 case INDEX_op_ctpop_i32: 1947 return TCG_TARGET_HAS_ctpop_i32; 1948 1949 case INDEX_op_brcond2_i32: 1950 case INDEX_op_setcond2_i32: 1951 return TCG_TARGET_REG_BITS == 32; 1952 1953 case INDEX_op_mov_i64: 1954 case INDEX_op_setcond_i64: 1955 case INDEX_op_brcond_i64: 1956 case INDEX_op_ld8u_i64: 1957 case INDEX_op_ld8s_i64: 1958 case INDEX_op_ld16u_i64: 1959 case INDEX_op_ld16s_i64: 1960 case INDEX_op_ld32u_i64: 1961 case INDEX_op_ld32s_i64: 1962 case INDEX_op_ld_i64: 1963 case INDEX_op_st8_i64: 1964 case INDEX_op_st16_i64: 1965 case INDEX_op_st32_i64: 1966 case INDEX_op_st_i64: 1967 case INDEX_op_add_i64: 1968 case INDEX_op_sub_i64: 1969 case INDEX_op_mul_i64: 1970 case INDEX_op_and_i64: 1971 case INDEX_op_or_i64: 1972 case INDEX_op_xor_i64: 1973 case INDEX_op_shl_i64: 1974 case INDEX_op_shr_i64: 1975 case INDEX_op_sar_i64: 1976 case INDEX_op_ext_i32_i64: 1977 case INDEX_op_extu_i32_i64: 1978 return TCG_TARGET_REG_BITS == 64; 1979 1980 case INDEX_op_movcond_i64: 1981 return TCG_TARGET_HAS_movcond_i64; 1982 case INDEX_op_div_i64: 1983 case INDEX_op_divu_i64: 1984 return TCG_TARGET_HAS_div_i64; 1985 case INDEX_op_rem_i64: 1986 case INDEX_op_remu_i64: 1987 return TCG_TARGET_HAS_rem_i64; 1988 case INDEX_op_div2_i64: 1989 case INDEX_op_divu2_i64: 1990 return TCG_TARGET_HAS_div2_i64; 1991 case INDEX_op_rotl_i64: 1992 case INDEX_op_rotr_i64: 1993 return TCG_TARGET_HAS_rot_i64; 1994 case INDEX_op_deposit_i64: 1995 return TCG_TARGET_HAS_deposit_i64; 1996 case INDEX_op_extract_i64: 1997 return TCG_TARGET_HAS_extract_i64; 1998 case INDEX_op_sextract_i64: 1999 return TCG_TARGET_HAS_sextract_i64; 2000 case INDEX_op_extract2_i64: 2001 return TCG_TARGET_HAS_extract2_i64; 2002 case INDEX_op_extrl_i64_i32: 2003 return TCG_TARGET_HAS_extrl_i64_i32; 2004 case INDEX_op_extrh_i64_i32: 2005 return TCG_TARGET_HAS_extrh_i64_i32; 2006 case INDEX_op_ext8s_i64: 2007 return TCG_TARGET_HAS_ext8s_i64; 2008 case INDEX_op_ext16s_i64: 2009 return TCG_TARGET_HAS_ext16s_i64; 2010 case INDEX_op_ext32s_i64: 2011 return TCG_TARGET_HAS_ext32s_i64; 2012 case INDEX_op_ext8u_i64: 2013 return TCG_TARGET_HAS_ext8u_i64; 2014 case INDEX_op_ext16u_i64: 2015 return TCG_TARGET_HAS_ext16u_i64; 2016 case INDEX_op_ext32u_i64: 2017 return TCG_TARGET_HAS_ext32u_i64; 2018 case INDEX_op_bswap16_i64: 2019 return TCG_TARGET_HAS_bswap16_i64; 2020 case INDEX_op_bswap32_i64: 2021 return TCG_TARGET_HAS_bswap32_i64; 2022 case INDEX_op_bswap64_i64: 2023 return TCG_TARGET_HAS_bswap64_i64; 2024 case INDEX_op_not_i64: 2025 return TCG_TARGET_HAS_not_i64; 2026 case INDEX_op_neg_i64: 2027 return TCG_TARGET_HAS_neg_i64; 2028 case INDEX_op_andc_i64: 2029 return TCG_TARGET_HAS_andc_i64; 2030 case INDEX_op_orc_i64: 2031 return TCG_TARGET_HAS_orc_i64; 2032 case INDEX_op_eqv_i64: 2033 return TCG_TARGET_HAS_eqv_i64; 2034 case INDEX_op_nand_i64: 2035 return TCG_TARGET_HAS_nand_i64; 2036 case INDEX_op_nor_i64: 2037 return TCG_TARGET_HAS_nor_i64; 2038 case INDEX_op_clz_i64: 2039 return TCG_TARGET_HAS_clz_i64; 2040 case INDEX_op_ctz_i64: 2041 return TCG_TARGET_HAS_ctz_i64; 2042 case INDEX_op_ctpop_i64: 2043 return TCG_TARGET_HAS_ctpop_i64; 2044 case INDEX_op_add2_i64: 2045 return TCG_TARGET_HAS_add2_i64; 2046 case INDEX_op_sub2_i64: 2047 return TCG_TARGET_HAS_sub2_i64; 2048 case INDEX_op_mulu2_i64: 2049 return TCG_TARGET_HAS_mulu2_i64; 2050 case INDEX_op_muls2_i64: 2051 return TCG_TARGET_HAS_muls2_i64; 2052 case INDEX_op_muluh_i64: 2053 return TCG_TARGET_HAS_muluh_i64; 2054 case INDEX_op_mulsh_i64: 2055 return TCG_TARGET_HAS_mulsh_i64; 2056 2057 case INDEX_op_mov_vec: 2058 case INDEX_op_dup_vec: 2059 case INDEX_op_dupm_vec: 2060 case INDEX_op_ld_vec: 2061 case INDEX_op_st_vec: 2062 case INDEX_op_add_vec: 2063 case INDEX_op_sub_vec: 2064 case INDEX_op_and_vec: 2065 case INDEX_op_or_vec: 2066 case INDEX_op_xor_vec: 2067 case INDEX_op_cmp_vec: 2068 return have_vec; 2069 case INDEX_op_dup2_vec: 2070 return have_vec && TCG_TARGET_REG_BITS == 32; 2071 case INDEX_op_not_vec: 2072 return have_vec && TCG_TARGET_HAS_not_vec; 2073 case INDEX_op_neg_vec: 2074 return have_vec && TCG_TARGET_HAS_neg_vec; 2075 case INDEX_op_abs_vec: 2076 return have_vec && TCG_TARGET_HAS_abs_vec; 2077 case INDEX_op_andc_vec: 2078 return have_vec && TCG_TARGET_HAS_andc_vec; 2079 case INDEX_op_orc_vec: 2080 return have_vec && TCG_TARGET_HAS_orc_vec; 2081 case INDEX_op_nand_vec: 2082 return have_vec && TCG_TARGET_HAS_nand_vec; 2083 case INDEX_op_nor_vec: 2084 return have_vec && TCG_TARGET_HAS_nor_vec; 2085 case INDEX_op_eqv_vec: 2086 return have_vec && TCG_TARGET_HAS_eqv_vec; 2087 case INDEX_op_mul_vec: 2088 return have_vec && TCG_TARGET_HAS_mul_vec; 2089 case INDEX_op_shli_vec: 2090 case INDEX_op_shri_vec: 2091 case INDEX_op_sari_vec: 2092 return have_vec && TCG_TARGET_HAS_shi_vec; 2093 case INDEX_op_shls_vec: 2094 case INDEX_op_shrs_vec: 2095 case INDEX_op_sars_vec: 2096 return have_vec && TCG_TARGET_HAS_shs_vec; 2097 case INDEX_op_shlv_vec: 2098 case INDEX_op_shrv_vec: 2099 case INDEX_op_sarv_vec: 2100 return have_vec && TCG_TARGET_HAS_shv_vec; 2101 case INDEX_op_rotli_vec: 2102 return have_vec && TCG_TARGET_HAS_roti_vec; 2103 case INDEX_op_rotls_vec: 2104 return have_vec && TCG_TARGET_HAS_rots_vec; 2105 case INDEX_op_rotlv_vec: 2106 case INDEX_op_rotrv_vec: 2107 return have_vec && TCG_TARGET_HAS_rotv_vec; 2108 case INDEX_op_ssadd_vec: 2109 case INDEX_op_usadd_vec: 2110 case INDEX_op_sssub_vec: 2111 case INDEX_op_ussub_vec: 2112 return have_vec && TCG_TARGET_HAS_sat_vec; 2113 case INDEX_op_smin_vec: 2114 case INDEX_op_umin_vec: 2115 case INDEX_op_smax_vec: 2116 case INDEX_op_umax_vec: 2117 return have_vec && TCG_TARGET_HAS_minmax_vec; 2118 case INDEX_op_bitsel_vec: 2119 return have_vec && TCG_TARGET_HAS_bitsel_vec; 2120 case INDEX_op_cmpsel_vec: 2121 return have_vec && TCG_TARGET_HAS_cmpsel_vec; 2122 2123 default: 2124 tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS); 2125 return true; 2126 } 2127 } 2128 2129 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs); 2130 2131 void tcg_gen_callN(void *func, TCGTemp *ret, int nargs, TCGTemp **args) 2132 { 2133 const TCGHelperInfo *info; 2134 TCGv_i64 extend_free[MAX_CALL_IARGS]; 2135 int n_extend = 0; 2136 TCGOp *op; 2137 int i, n, pi = 0, total_args; 2138 2139 info = g_hash_table_lookup(helper_table, (gpointer)func); 2140 total_args = info->nr_out + info->nr_in + 2; 2141 op = tcg_op_alloc(INDEX_op_call, total_args); 2142 2143 #ifdef CONFIG_PLUGIN 2144 /* Flag helpers that may affect guest state */ 2145 if (tcg_ctx->plugin_insn && 2146 !(info->flags & TCG_CALL_PLUGIN) && 2147 !(info->flags & TCG_CALL_NO_SIDE_EFFECTS)) { 2148 tcg_ctx->plugin_insn->calls_helpers = true; 2149 } 2150 #endif 2151 2152 TCGOP_CALLO(op) = n = info->nr_out; 2153 switch (n) { 2154 case 0: 2155 tcg_debug_assert(ret == NULL); 2156 break; 2157 case 1: 2158 tcg_debug_assert(ret != NULL); 2159 op->args[pi++] = temp_arg(ret); 2160 break; 2161 case 2: 2162 case 4: 2163 tcg_debug_assert(ret != NULL); 2164 tcg_debug_assert(ret->base_type == ret->type + ctz32(n)); 2165 tcg_debug_assert(ret->temp_subindex == 0); 2166 for (i = 0; i < n; ++i) { 2167 op->args[pi++] = temp_arg(ret + i); 2168 } 2169 break; 2170 default: 2171 g_assert_not_reached(); 2172 } 2173 2174 TCGOP_CALLI(op) = n = info->nr_in; 2175 for (i = 0; i < n; i++) { 2176 const TCGCallArgumentLoc *loc = &info->in[i]; 2177 TCGTemp *ts = args[loc->arg_idx] + loc->tmp_subindex; 2178 2179 switch (loc->kind) { 2180 case TCG_CALL_ARG_NORMAL: 2181 case TCG_CALL_ARG_BY_REF: 2182 case TCG_CALL_ARG_BY_REF_N: 2183 op->args[pi++] = temp_arg(ts); 2184 break; 2185 2186 case TCG_CALL_ARG_EXTEND_U: 2187 case TCG_CALL_ARG_EXTEND_S: 2188 { 2189 TCGv_i64 temp = tcg_temp_ebb_new_i64(); 2190 TCGv_i32 orig = temp_tcgv_i32(ts); 2191 2192 if (loc->kind == TCG_CALL_ARG_EXTEND_S) { 2193 tcg_gen_ext_i32_i64(temp, orig); 2194 } else { 2195 tcg_gen_extu_i32_i64(temp, orig); 2196 } 2197 op->args[pi++] = tcgv_i64_arg(temp); 2198 extend_free[n_extend++] = temp; 2199 } 2200 break; 2201 2202 default: 2203 g_assert_not_reached(); 2204 } 2205 } 2206 op->args[pi++] = (uintptr_t)func; 2207 op->args[pi++] = (uintptr_t)info; 2208 tcg_debug_assert(pi == total_args); 2209 2210 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 2211 2212 tcg_debug_assert(n_extend < ARRAY_SIZE(extend_free)); 2213 for (i = 0; i < n_extend; ++i) { 2214 tcg_temp_free_i64(extend_free[i]); 2215 } 2216 } 2217 2218 static void tcg_reg_alloc_start(TCGContext *s) 2219 { 2220 int i, n; 2221 2222 for (i = 0, n = s->nb_temps; i < n; i++) { 2223 TCGTemp *ts = &s->temps[i]; 2224 TCGTempVal val = TEMP_VAL_MEM; 2225 2226 switch (ts->kind) { 2227 case TEMP_CONST: 2228 val = TEMP_VAL_CONST; 2229 break; 2230 case TEMP_FIXED: 2231 val = TEMP_VAL_REG; 2232 break; 2233 case TEMP_GLOBAL: 2234 break; 2235 case TEMP_EBB: 2236 val = TEMP_VAL_DEAD; 2237 /* fall through */ 2238 case TEMP_TB: 2239 ts->mem_allocated = 0; 2240 break; 2241 default: 2242 g_assert_not_reached(); 2243 } 2244 ts->val_type = val; 2245 } 2246 2247 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp)); 2248 } 2249 2250 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size, 2251 TCGTemp *ts) 2252 { 2253 int idx = temp_idx(ts); 2254 2255 switch (ts->kind) { 2256 case TEMP_FIXED: 2257 case TEMP_GLOBAL: 2258 pstrcpy(buf, buf_size, ts->name); 2259 break; 2260 case TEMP_TB: 2261 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals); 2262 break; 2263 case TEMP_EBB: 2264 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals); 2265 break; 2266 case TEMP_CONST: 2267 switch (ts->type) { 2268 case TCG_TYPE_I32: 2269 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val); 2270 break; 2271 #if TCG_TARGET_REG_BITS > 32 2272 case TCG_TYPE_I64: 2273 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val); 2274 break; 2275 #endif 2276 case TCG_TYPE_V64: 2277 case TCG_TYPE_V128: 2278 case TCG_TYPE_V256: 2279 snprintf(buf, buf_size, "v%d$0x%" PRIx64, 2280 64 << (ts->type - TCG_TYPE_V64), ts->val); 2281 break; 2282 default: 2283 g_assert_not_reached(); 2284 } 2285 break; 2286 } 2287 return buf; 2288 } 2289 2290 static char *tcg_get_arg_str(TCGContext *s, char *buf, 2291 int buf_size, TCGArg arg) 2292 { 2293 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg)); 2294 } 2295 2296 static const char * const cond_name[] = 2297 { 2298 [TCG_COND_NEVER] = "never", 2299 [TCG_COND_ALWAYS] = "always", 2300 [TCG_COND_EQ] = "eq", 2301 [TCG_COND_NE] = "ne", 2302 [TCG_COND_LT] = "lt", 2303 [TCG_COND_GE] = "ge", 2304 [TCG_COND_LE] = "le", 2305 [TCG_COND_GT] = "gt", 2306 [TCG_COND_LTU] = "ltu", 2307 [TCG_COND_GEU] = "geu", 2308 [TCG_COND_LEU] = "leu", 2309 [TCG_COND_GTU] = "gtu" 2310 }; 2311 2312 static const char * const ldst_name[(MO_BSWAP | MO_SSIZE) + 1] = 2313 { 2314 [MO_UB] = "ub", 2315 [MO_SB] = "sb", 2316 [MO_LEUW] = "leuw", 2317 [MO_LESW] = "lesw", 2318 [MO_LEUL] = "leul", 2319 [MO_LESL] = "lesl", 2320 [MO_LEUQ] = "leq", 2321 [MO_BEUW] = "beuw", 2322 [MO_BESW] = "besw", 2323 [MO_BEUL] = "beul", 2324 [MO_BESL] = "besl", 2325 [MO_BEUQ] = "beq", 2326 [MO_128 + MO_BE] = "beo", 2327 [MO_128 + MO_LE] = "leo", 2328 }; 2329 2330 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = { 2331 [MO_UNALN >> MO_ASHIFT] = "un+", 2332 [MO_ALIGN >> MO_ASHIFT] = "al+", 2333 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+", 2334 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+", 2335 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+", 2336 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+", 2337 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+", 2338 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+", 2339 }; 2340 2341 static const char * const atom_name[(MO_ATOM_MASK >> MO_ATOM_SHIFT) + 1] = { 2342 [MO_ATOM_IFALIGN >> MO_ATOM_SHIFT] = "", 2343 [MO_ATOM_IFALIGN_PAIR >> MO_ATOM_SHIFT] = "pair+", 2344 [MO_ATOM_WITHIN16 >> MO_ATOM_SHIFT] = "w16+", 2345 [MO_ATOM_WITHIN16_PAIR >> MO_ATOM_SHIFT] = "w16p+", 2346 [MO_ATOM_SUBALIGN >> MO_ATOM_SHIFT] = "sub+", 2347 [MO_ATOM_NONE >> MO_ATOM_SHIFT] = "noat+", 2348 }; 2349 2350 static const char bswap_flag_name[][6] = { 2351 [TCG_BSWAP_IZ] = "iz", 2352 [TCG_BSWAP_OZ] = "oz", 2353 [TCG_BSWAP_OS] = "os", 2354 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz", 2355 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os", 2356 }; 2357 2358 static inline bool tcg_regset_single(TCGRegSet d) 2359 { 2360 return (d & (d - 1)) == 0; 2361 } 2362 2363 static inline TCGReg tcg_regset_first(TCGRegSet d) 2364 { 2365 if (TCG_TARGET_NB_REGS <= 32) { 2366 return ctz32(d); 2367 } else { 2368 return ctz64(d); 2369 } 2370 } 2371 2372 /* Return only the number of characters output -- no error return. */ 2373 #define ne_fprintf(...) \ 2374 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; }) 2375 2376 static void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs) 2377 { 2378 char buf[128]; 2379 TCGOp *op; 2380 2381 QTAILQ_FOREACH(op, &s->ops, link) { 2382 int i, k, nb_oargs, nb_iargs, nb_cargs; 2383 const TCGOpDef *def; 2384 TCGOpcode c; 2385 int col = 0; 2386 2387 c = op->opc; 2388 def = &tcg_op_defs[c]; 2389 2390 if (c == INDEX_op_insn_start) { 2391 nb_oargs = 0; 2392 col += ne_fprintf(f, "\n ----"); 2393 2394 for (i = 0; i < TARGET_INSN_START_WORDS; ++i) { 2395 col += ne_fprintf(f, " %016" PRIx64, 2396 tcg_get_insn_start_param(op, i)); 2397 } 2398 } else if (c == INDEX_op_call) { 2399 const TCGHelperInfo *info = tcg_call_info(op); 2400 void *func = tcg_call_func(op); 2401 2402 /* variable number of arguments */ 2403 nb_oargs = TCGOP_CALLO(op); 2404 nb_iargs = TCGOP_CALLI(op); 2405 nb_cargs = def->nb_cargs; 2406 2407 col += ne_fprintf(f, " %s ", def->name); 2408 2409 /* 2410 * Print the function name from TCGHelperInfo, if available. 2411 * Note that plugins have a template function for the info, 2412 * but the actual function pointer comes from the plugin. 2413 */ 2414 if (func == info->func) { 2415 col += ne_fprintf(f, "%s", info->name); 2416 } else { 2417 col += ne_fprintf(f, "plugin(%p)", func); 2418 } 2419 2420 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs); 2421 for (i = 0; i < nb_oargs; i++) { 2422 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf), 2423 op->args[i])); 2424 } 2425 for (i = 0; i < nb_iargs; i++) { 2426 TCGArg arg = op->args[nb_oargs + i]; 2427 const char *t = tcg_get_arg_str(s, buf, sizeof(buf), arg); 2428 col += ne_fprintf(f, ",%s", t); 2429 } 2430 } else { 2431 col += ne_fprintf(f, " %s ", def->name); 2432 2433 nb_oargs = def->nb_oargs; 2434 nb_iargs = def->nb_iargs; 2435 nb_cargs = def->nb_cargs; 2436 2437 if (def->flags & TCG_OPF_VECTOR) { 2438 col += ne_fprintf(f, "v%d,e%d,", 64 << TCGOP_VECL(op), 2439 8 << TCGOP_VECE(op)); 2440 } 2441 2442 k = 0; 2443 for (i = 0; i < nb_oargs; i++) { 2444 const char *sep = k ? "," : ""; 2445 col += ne_fprintf(f, "%s%s", sep, 2446 tcg_get_arg_str(s, buf, sizeof(buf), 2447 op->args[k++])); 2448 } 2449 for (i = 0; i < nb_iargs; i++) { 2450 const char *sep = k ? "," : ""; 2451 col += ne_fprintf(f, "%s%s", sep, 2452 tcg_get_arg_str(s, buf, sizeof(buf), 2453 op->args[k++])); 2454 } 2455 switch (c) { 2456 case INDEX_op_brcond_i32: 2457 case INDEX_op_setcond_i32: 2458 case INDEX_op_movcond_i32: 2459 case INDEX_op_brcond2_i32: 2460 case INDEX_op_setcond2_i32: 2461 case INDEX_op_brcond_i64: 2462 case INDEX_op_setcond_i64: 2463 case INDEX_op_movcond_i64: 2464 case INDEX_op_cmp_vec: 2465 case INDEX_op_cmpsel_vec: 2466 if (op->args[k] < ARRAY_SIZE(cond_name) 2467 && cond_name[op->args[k]]) { 2468 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]); 2469 } else { 2470 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]); 2471 } 2472 i = 1; 2473 break; 2474 case INDEX_op_qemu_ld_a32_i32: 2475 case INDEX_op_qemu_ld_a64_i32: 2476 case INDEX_op_qemu_st_a32_i32: 2477 case INDEX_op_qemu_st_a64_i32: 2478 case INDEX_op_qemu_st8_a32_i32: 2479 case INDEX_op_qemu_st8_a64_i32: 2480 case INDEX_op_qemu_ld_a32_i64: 2481 case INDEX_op_qemu_ld_a64_i64: 2482 case INDEX_op_qemu_st_a32_i64: 2483 case INDEX_op_qemu_st_a64_i64: 2484 case INDEX_op_qemu_ld_a32_i128: 2485 case INDEX_op_qemu_ld_a64_i128: 2486 case INDEX_op_qemu_st_a32_i128: 2487 case INDEX_op_qemu_st_a64_i128: 2488 { 2489 const char *s_al, *s_op, *s_at; 2490 MemOpIdx oi = op->args[k++]; 2491 MemOp op = get_memop(oi); 2492 unsigned ix = get_mmuidx(oi); 2493 2494 s_al = alignment_name[(op & MO_AMASK) >> MO_ASHIFT]; 2495 s_op = ldst_name[op & (MO_BSWAP | MO_SSIZE)]; 2496 s_at = atom_name[(op & MO_ATOM_MASK) >> MO_ATOM_SHIFT]; 2497 op &= ~(MO_AMASK | MO_BSWAP | MO_SSIZE | MO_ATOM_MASK); 2498 2499 /* If all fields are accounted for, print symbolically. */ 2500 if (!op && s_al && s_op && s_at) { 2501 col += ne_fprintf(f, ",%s%s%s,%u", 2502 s_at, s_al, s_op, ix); 2503 } else { 2504 op = get_memop(oi); 2505 col += ne_fprintf(f, ",$0x%x,%u", op, ix); 2506 } 2507 i = 1; 2508 } 2509 break; 2510 case INDEX_op_bswap16_i32: 2511 case INDEX_op_bswap16_i64: 2512 case INDEX_op_bswap32_i32: 2513 case INDEX_op_bswap32_i64: 2514 case INDEX_op_bswap64_i64: 2515 { 2516 TCGArg flags = op->args[k]; 2517 const char *name = NULL; 2518 2519 if (flags < ARRAY_SIZE(bswap_flag_name)) { 2520 name = bswap_flag_name[flags]; 2521 } 2522 if (name) { 2523 col += ne_fprintf(f, ",%s", name); 2524 } else { 2525 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags); 2526 } 2527 i = k = 1; 2528 } 2529 break; 2530 default: 2531 i = 0; 2532 break; 2533 } 2534 switch (c) { 2535 case INDEX_op_set_label: 2536 case INDEX_op_br: 2537 case INDEX_op_brcond_i32: 2538 case INDEX_op_brcond_i64: 2539 case INDEX_op_brcond2_i32: 2540 col += ne_fprintf(f, "%s$L%d", k ? "," : "", 2541 arg_label(op->args[k])->id); 2542 i++, k++; 2543 break; 2544 case INDEX_op_mb: 2545 { 2546 TCGBar membar = op->args[k]; 2547 const char *b_op, *m_op; 2548 2549 switch (membar & TCG_BAR_SC) { 2550 case 0: 2551 b_op = "none"; 2552 break; 2553 case TCG_BAR_LDAQ: 2554 b_op = "acq"; 2555 break; 2556 case TCG_BAR_STRL: 2557 b_op = "rel"; 2558 break; 2559 case TCG_BAR_SC: 2560 b_op = "seq"; 2561 break; 2562 default: 2563 g_assert_not_reached(); 2564 } 2565 2566 switch (membar & TCG_MO_ALL) { 2567 case 0: 2568 m_op = "none"; 2569 break; 2570 case TCG_MO_LD_LD: 2571 m_op = "rr"; 2572 break; 2573 case TCG_MO_LD_ST: 2574 m_op = "rw"; 2575 break; 2576 case TCG_MO_ST_LD: 2577 m_op = "wr"; 2578 break; 2579 case TCG_MO_ST_ST: 2580 m_op = "ww"; 2581 break; 2582 case TCG_MO_LD_LD | TCG_MO_LD_ST: 2583 m_op = "rr+rw"; 2584 break; 2585 case TCG_MO_LD_LD | TCG_MO_ST_LD: 2586 m_op = "rr+wr"; 2587 break; 2588 case TCG_MO_LD_LD | TCG_MO_ST_ST: 2589 m_op = "rr+ww"; 2590 break; 2591 case TCG_MO_LD_ST | TCG_MO_ST_LD: 2592 m_op = "rw+wr"; 2593 break; 2594 case TCG_MO_LD_ST | TCG_MO_ST_ST: 2595 m_op = "rw+ww"; 2596 break; 2597 case TCG_MO_ST_LD | TCG_MO_ST_ST: 2598 m_op = "wr+ww"; 2599 break; 2600 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_LD: 2601 m_op = "rr+rw+wr"; 2602 break; 2603 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_ST: 2604 m_op = "rr+rw+ww"; 2605 break; 2606 case TCG_MO_LD_LD | TCG_MO_ST_LD | TCG_MO_ST_ST: 2607 m_op = "rr+wr+ww"; 2608 break; 2609 case TCG_MO_LD_ST | TCG_MO_ST_LD | TCG_MO_ST_ST: 2610 m_op = "rw+wr+ww"; 2611 break; 2612 case TCG_MO_ALL: 2613 m_op = "all"; 2614 break; 2615 default: 2616 g_assert_not_reached(); 2617 } 2618 2619 col += ne_fprintf(f, "%s%s:%s", (k ? "," : ""), b_op, m_op); 2620 i++, k++; 2621 } 2622 break; 2623 default: 2624 break; 2625 } 2626 for (; i < nb_cargs; i++, k++) { 2627 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "", 2628 op->args[k]); 2629 } 2630 } 2631 2632 if (have_prefs || op->life) { 2633 for (; col < 40; ++col) { 2634 putc(' ', f); 2635 } 2636 } 2637 2638 if (op->life) { 2639 unsigned life = op->life; 2640 2641 if (life & (SYNC_ARG * 3)) { 2642 ne_fprintf(f, " sync:"); 2643 for (i = 0; i < 2; ++i) { 2644 if (life & (SYNC_ARG << i)) { 2645 ne_fprintf(f, " %d", i); 2646 } 2647 } 2648 } 2649 life /= DEAD_ARG; 2650 if (life) { 2651 ne_fprintf(f, " dead:"); 2652 for (i = 0; life; ++i, life >>= 1) { 2653 if (life & 1) { 2654 ne_fprintf(f, " %d", i); 2655 } 2656 } 2657 } 2658 } 2659 2660 if (have_prefs) { 2661 for (i = 0; i < nb_oargs; ++i) { 2662 TCGRegSet set = output_pref(op, i); 2663 2664 if (i == 0) { 2665 ne_fprintf(f, " pref="); 2666 } else { 2667 ne_fprintf(f, ","); 2668 } 2669 if (set == 0) { 2670 ne_fprintf(f, "none"); 2671 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) { 2672 ne_fprintf(f, "all"); 2673 #ifdef CONFIG_DEBUG_TCG 2674 } else if (tcg_regset_single(set)) { 2675 TCGReg reg = tcg_regset_first(set); 2676 ne_fprintf(f, "%s", tcg_target_reg_names[reg]); 2677 #endif 2678 } else if (TCG_TARGET_NB_REGS <= 32) { 2679 ne_fprintf(f, "0x%x", (uint32_t)set); 2680 } else { 2681 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set); 2682 } 2683 } 2684 } 2685 2686 putc('\n', f); 2687 } 2688 } 2689 2690 /* we give more priority to constraints with less registers */ 2691 static int get_constraint_priority(const TCGOpDef *def, int k) 2692 { 2693 const TCGArgConstraint *arg_ct = &def->args_ct[k]; 2694 int n = ctpop64(arg_ct->regs); 2695 2696 /* 2697 * Sort constraints of a single register first, which includes output 2698 * aliases (which must exactly match the input already allocated). 2699 */ 2700 if (n == 1 || arg_ct->oalias) { 2701 return INT_MAX; 2702 } 2703 2704 /* 2705 * Sort register pairs next, first then second immediately after. 2706 * Arbitrarily sort multiple pairs by the index of the first reg; 2707 * there shouldn't be many pairs. 2708 */ 2709 switch (arg_ct->pair) { 2710 case 1: 2711 case 3: 2712 return (k + 1) * 2; 2713 case 2: 2714 return (arg_ct->pair_index + 1) * 2 - 1; 2715 } 2716 2717 /* Finally, sort by decreasing register count. */ 2718 assert(n > 1); 2719 return -n; 2720 } 2721 2722 /* sort from highest priority to lowest */ 2723 static void sort_constraints(TCGOpDef *def, int start, int n) 2724 { 2725 int i, j; 2726 TCGArgConstraint *a = def->args_ct; 2727 2728 for (i = 0; i < n; i++) { 2729 a[start + i].sort_index = start + i; 2730 } 2731 if (n <= 1) { 2732 return; 2733 } 2734 for (i = 0; i < n - 1; i++) { 2735 for (j = i + 1; j < n; j++) { 2736 int p1 = get_constraint_priority(def, a[start + i].sort_index); 2737 int p2 = get_constraint_priority(def, a[start + j].sort_index); 2738 if (p1 < p2) { 2739 int tmp = a[start + i].sort_index; 2740 a[start + i].sort_index = a[start + j].sort_index; 2741 a[start + j].sort_index = tmp; 2742 } 2743 } 2744 } 2745 } 2746 2747 static void process_op_defs(TCGContext *s) 2748 { 2749 TCGOpcode op; 2750 2751 for (op = 0; op < NB_OPS; op++) { 2752 TCGOpDef *def = &tcg_op_defs[op]; 2753 const TCGTargetOpDef *tdefs; 2754 bool saw_alias_pair = false; 2755 int i, o, i2, o2, nb_args; 2756 2757 if (def->flags & TCG_OPF_NOT_PRESENT) { 2758 continue; 2759 } 2760 2761 nb_args = def->nb_iargs + def->nb_oargs; 2762 if (nb_args == 0) { 2763 continue; 2764 } 2765 2766 /* 2767 * Macro magic should make it impossible, but double-check that 2768 * the array index is in range. Since the signness of an enum 2769 * is implementation defined, force the result to unsigned. 2770 */ 2771 unsigned con_set = tcg_target_op_def(op); 2772 tcg_debug_assert(con_set < ARRAY_SIZE(constraint_sets)); 2773 tdefs = &constraint_sets[con_set]; 2774 2775 for (i = 0; i < nb_args; i++) { 2776 const char *ct_str = tdefs->args_ct_str[i]; 2777 bool input_p = i >= def->nb_oargs; 2778 2779 /* Incomplete TCGTargetOpDef entry. */ 2780 tcg_debug_assert(ct_str != NULL); 2781 2782 switch (*ct_str) { 2783 case '0' ... '9': 2784 o = *ct_str - '0'; 2785 tcg_debug_assert(input_p); 2786 tcg_debug_assert(o < def->nb_oargs); 2787 tcg_debug_assert(def->args_ct[o].regs != 0); 2788 tcg_debug_assert(!def->args_ct[o].oalias); 2789 def->args_ct[i] = def->args_ct[o]; 2790 /* The output sets oalias. */ 2791 def->args_ct[o].oalias = 1; 2792 def->args_ct[o].alias_index = i; 2793 /* The input sets ialias. */ 2794 def->args_ct[i].ialias = 1; 2795 def->args_ct[i].alias_index = o; 2796 if (def->args_ct[i].pair) { 2797 saw_alias_pair = true; 2798 } 2799 tcg_debug_assert(ct_str[1] == '\0'); 2800 continue; 2801 2802 case '&': 2803 tcg_debug_assert(!input_p); 2804 def->args_ct[i].newreg = true; 2805 ct_str++; 2806 break; 2807 2808 case 'p': /* plus */ 2809 /* Allocate to the register after the previous. */ 2810 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0)); 2811 o = i - 1; 2812 tcg_debug_assert(!def->args_ct[o].pair); 2813 tcg_debug_assert(!def->args_ct[o].ct); 2814 def->args_ct[i] = (TCGArgConstraint){ 2815 .pair = 2, 2816 .pair_index = o, 2817 .regs = def->args_ct[o].regs << 1, 2818 }; 2819 def->args_ct[o].pair = 1; 2820 def->args_ct[o].pair_index = i; 2821 tcg_debug_assert(ct_str[1] == '\0'); 2822 continue; 2823 2824 case 'm': /* minus */ 2825 /* Allocate to the register before the previous. */ 2826 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0)); 2827 o = i - 1; 2828 tcg_debug_assert(!def->args_ct[o].pair); 2829 tcg_debug_assert(!def->args_ct[o].ct); 2830 def->args_ct[i] = (TCGArgConstraint){ 2831 .pair = 1, 2832 .pair_index = o, 2833 .regs = def->args_ct[o].regs >> 1, 2834 }; 2835 def->args_ct[o].pair = 2; 2836 def->args_ct[o].pair_index = i; 2837 tcg_debug_assert(ct_str[1] == '\0'); 2838 continue; 2839 } 2840 2841 do { 2842 switch (*ct_str) { 2843 case 'i': 2844 def->args_ct[i].ct |= TCG_CT_CONST; 2845 break; 2846 2847 /* Include all of the target-specific constraints. */ 2848 2849 #undef CONST 2850 #define CONST(CASE, MASK) \ 2851 case CASE: def->args_ct[i].ct |= MASK; break; 2852 #define REGS(CASE, MASK) \ 2853 case CASE: def->args_ct[i].regs |= MASK; break; 2854 2855 #include "tcg-target-con-str.h" 2856 2857 #undef REGS 2858 #undef CONST 2859 default: 2860 case '0' ... '9': 2861 case '&': 2862 case 'p': 2863 case 'm': 2864 /* Typo in TCGTargetOpDef constraint. */ 2865 g_assert_not_reached(); 2866 } 2867 } while (*++ct_str != '\0'); 2868 } 2869 2870 /* TCGTargetOpDef entry with too much information? */ 2871 tcg_debug_assert(i == TCG_MAX_OP_ARGS || tdefs->args_ct_str[i] == NULL); 2872 2873 /* 2874 * Fix up output pairs that are aliased with inputs. 2875 * When we created the alias, we copied pair from the output. 2876 * There are three cases: 2877 * (1a) Pairs of inputs alias pairs of outputs. 2878 * (1b) One input aliases the first of a pair of outputs. 2879 * (2) One input aliases the second of a pair of outputs. 2880 * 2881 * Case 1a is handled by making sure that the pair_index'es are 2882 * properly updated so that they appear the same as a pair of inputs. 2883 * 2884 * Case 1b is handled by setting the pair_index of the input to 2885 * itself, simply so it doesn't point to an unrelated argument. 2886 * Since we don't encounter the "second" during the input allocation 2887 * phase, nothing happens with the second half of the input pair. 2888 * 2889 * Case 2 is handled by setting the second input to pair=3, the 2890 * first output to pair=3, and the pair_index'es to match. 2891 */ 2892 if (saw_alias_pair) { 2893 for (i = def->nb_oargs; i < nb_args; i++) { 2894 /* 2895 * Since [0-9pm] must be alone in the constraint string, 2896 * the only way they can both be set is if the pair comes 2897 * from the output alias. 2898 */ 2899 if (!def->args_ct[i].ialias) { 2900 continue; 2901 } 2902 switch (def->args_ct[i].pair) { 2903 case 0: 2904 break; 2905 case 1: 2906 o = def->args_ct[i].alias_index; 2907 o2 = def->args_ct[o].pair_index; 2908 tcg_debug_assert(def->args_ct[o].pair == 1); 2909 tcg_debug_assert(def->args_ct[o2].pair == 2); 2910 if (def->args_ct[o2].oalias) { 2911 /* Case 1a */ 2912 i2 = def->args_ct[o2].alias_index; 2913 tcg_debug_assert(def->args_ct[i2].pair == 2); 2914 def->args_ct[i2].pair_index = i; 2915 def->args_ct[i].pair_index = i2; 2916 } else { 2917 /* Case 1b */ 2918 def->args_ct[i].pair_index = i; 2919 } 2920 break; 2921 case 2: 2922 o = def->args_ct[i].alias_index; 2923 o2 = def->args_ct[o].pair_index; 2924 tcg_debug_assert(def->args_ct[o].pair == 2); 2925 tcg_debug_assert(def->args_ct[o2].pair == 1); 2926 if (def->args_ct[o2].oalias) { 2927 /* Case 1a */ 2928 i2 = def->args_ct[o2].alias_index; 2929 tcg_debug_assert(def->args_ct[i2].pair == 1); 2930 def->args_ct[i2].pair_index = i; 2931 def->args_ct[i].pair_index = i2; 2932 } else { 2933 /* Case 2 */ 2934 def->args_ct[i].pair = 3; 2935 def->args_ct[o2].pair = 3; 2936 def->args_ct[i].pair_index = o2; 2937 def->args_ct[o2].pair_index = i; 2938 } 2939 break; 2940 default: 2941 g_assert_not_reached(); 2942 } 2943 } 2944 } 2945 2946 /* sort the constraints (XXX: this is just an heuristic) */ 2947 sort_constraints(def, 0, def->nb_oargs); 2948 sort_constraints(def, def->nb_oargs, def->nb_iargs); 2949 } 2950 } 2951 2952 static void remove_label_use(TCGOp *op, int idx) 2953 { 2954 TCGLabel *label = arg_label(op->args[idx]); 2955 TCGLabelUse *use; 2956 2957 QSIMPLEQ_FOREACH(use, &label->branches, next) { 2958 if (use->op == op) { 2959 QSIMPLEQ_REMOVE(&label->branches, use, TCGLabelUse, next); 2960 return; 2961 } 2962 } 2963 g_assert_not_reached(); 2964 } 2965 2966 void tcg_op_remove(TCGContext *s, TCGOp *op) 2967 { 2968 switch (op->opc) { 2969 case INDEX_op_br: 2970 remove_label_use(op, 0); 2971 break; 2972 case INDEX_op_brcond_i32: 2973 case INDEX_op_brcond_i64: 2974 remove_label_use(op, 3); 2975 break; 2976 case INDEX_op_brcond2_i32: 2977 remove_label_use(op, 5); 2978 break; 2979 default: 2980 break; 2981 } 2982 2983 QTAILQ_REMOVE(&s->ops, op, link); 2984 QTAILQ_INSERT_TAIL(&s->free_ops, op, link); 2985 s->nb_ops--; 2986 2987 #ifdef CONFIG_PROFILER 2988 qatomic_set(&s->prof.del_op_count, s->prof.del_op_count + 1); 2989 #endif 2990 } 2991 2992 void tcg_remove_ops_after(TCGOp *op) 2993 { 2994 TCGContext *s = tcg_ctx; 2995 2996 while (true) { 2997 TCGOp *last = tcg_last_op(); 2998 if (last == op) { 2999 return; 3000 } 3001 tcg_op_remove(s, last); 3002 } 3003 } 3004 3005 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs) 3006 { 3007 TCGContext *s = tcg_ctx; 3008 TCGOp *op = NULL; 3009 3010 if (unlikely(!QTAILQ_EMPTY(&s->free_ops))) { 3011 QTAILQ_FOREACH(op, &s->free_ops, link) { 3012 if (nargs <= op->nargs) { 3013 QTAILQ_REMOVE(&s->free_ops, op, link); 3014 nargs = op->nargs; 3015 goto found; 3016 } 3017 } 3018 } 3019 3020 /* Most opcodes have 3 or 4 operands: reduce fragmentation. */ 3021 nargs = MAX(4, nargs); 3022 op = tcg_malloc(sizeof(TCGOp) + sizeof(TCGArg) * nargs); 3023 3024 found: 3025 memset(op, 0, offsetof(TCGOp, link)); 3026 op->opc = opc; 3027 op->nargs = nargs; 3028 3029 /* Check for bitfield overflow. */ 3030 tcg_debug_assert(op->nargs == nargs); 3031 3032 s->nb_ops++; 3033 return op; 3034 } 3035 3036 TCGOp *tcg_emit_op(TCGOpcode opc, unsigned nargs) 3037 { 3038 TCGOp *op = tcg_op_alloc(opc, nargs); 3039 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 3040 return op; 3041 } 3042 3043 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op, 3044 TCGOpcode opc, unsigned nargs) 3045 { 3046 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3047 QTAILQ_INSERT_BEFORE(old_op, new_op, link); 3048 return new_op; 3049 } 3050 3051 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op, 3052 TCGOpcode opc, unsigned nargs) 3053 { 3054 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3055 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link); 3056 return new_op; 3057 } 3058 3059 static void move_label_uses(TCGLabel *to, TCGLabel *from) 3060 { 3061 TCGLabelUse *u; 3062 3063 QSIMPLEQ_FOREACH(u, &from->branches, next) { 3064 TCGOp *op = u->op; 3065 switch (op->opc) { 3066 case INDEX_op_br: 3067 op->args[0] = label_arg(to); 3068 break; 3069 case INDEX_op_brcond_i32: 3070 case INDEX_op_brcond_i64: 3071 op->args[3] = label_arg(to); 3072 break; 3073 case INDEX_op_brcond2_i32: 3074 op->args[5] = label_arg(to); 3075 break; 3076 default: 3077 g_assert_not_reached(); 3078 } 3079 } 3080 3081 QSIMPLEQ_CONCAT(&to->branches, &from->branches); 3082 } 3083 3084 /* Reachable analysis : remove unreachable code. */ 3085 static void __attribute__((noinline)) 3086 reachable_code_pass(TCGContext *s) 3087 { 3088 TCGOp *op, *op_next, *op_prev; 3089 bool dead = false; 3090 3091 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3092 bool remove = dead; 3093 TCGLabel *label; 3094 3095 switch (op->opc) { 3096 case INDEX_op_set_label: 3097 label = arg_label(op->args[0]); 3098 3099 /* 3100 * Note that the first op in the TB is always a load, 3101 * so there is always something before a label. 3102 */ 3103 op_prev = QTAILQ_PREV(op, link); 3104 3105 /* 3106 * If we find two sequential labels, move all branches to 3107 * reference the second label and remove the first label. 3108 * Do this before branch to next optimization, so that the 3109 * middle label is out of the way. 3110 */ 3111 if (op_prev->opc == INDEX_op_set_label) { 3112 move_label_uses(label, arg_label(op_prev->args[0])); 3113 tcg_op_remove(s, op_prev); 3114 op_prev = QTAILQ_PREV(op, link); 3115 } 3116 3117 /* 3118 * Optimization can fold conditional branches to unconditional. 3119 * If we find a label which is preceded by an unconditional 3120 * branch to next, remove the branch. We couldn't do this when 3121 * processing the branch because any dead code between the branch 3122 * and label had not yet been removed. 3123 */ 3124 if (op_prev->opc == INDEX_op_br && 3125 label == arg_label(op_prev->args[0])) { 3126 tcg_op_remove(s, op_prev); 3127 /* Fall through means insns become live again. */ 3128 dead = false; 3129 } 3130 3131 if (QSIMPLEQ_EMPTY(&label->branches)) { 3132 /* 3133 * While there is an occasional backward branch, virtually 3134 * all branches generated by the translators are forward. 3135 * Which means that generally we will have already removed 3136 * all references to the label that will be, and there is 3137 * little to be gained by iterating. 3138 */ 3139 remove = true; 3140 } else { 3141 /* Once we see a label, insns become live again. */ 3142 dead = false; 3143 remove = false; 3144 } 3145 break; 3146 3147 case INDEX_op_br: 3148 case INDEX_op_exit_tb: 3149 case INDEX_op_goto_ptr: 3150 /* Unconditional branches; everything following is dead. */ 3151 dead = true; 3152 break; 3153 3154 case INDEX_op_call: 3155 /* Notice noreturn helper calls, raising exceptions. */ 3156 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) { 3157 dead = true; 3158 } 3159 break; 3160 3161 case INDEX_op_insn_start: 3162 /* Never remove -- we need to keep these for unwind. */ 3163 remove = false; 3164 break; 3165 3166 default: 3167 break; 3168 } 3169 3170 if (remove) { 3171 tcg_op_remove(s, op); 3172 } 3173 } 3174 } 3175 3176 #define TS_DEAD 1 3177 #define TS_MEM 2 3178 3179 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n))) 3180 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n))) 3181 3182 /* For liveness_pass_1, the register preferences for a given temp. */ 3183 static inline TCGRegSet *la_temp_pref(TCGTemp *ts) 3184 { 3185 return ts->state_ptr; 3186 } 3187 3188 /* For liveness_pass_1, reset the preferences for a given temp to the 3189 * maximal regset for its type. 3190 */ 3191 static inline void la_reset_pref(TCGTemp *ts) 3192 { 3193 *la_temp_pref(ts) 3194 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]); 3195 } 3196 3197 /* liveness analysis: end of function: all temps are dead, and globals 3198 should be in memory. */ 3199 static void la_func_end(TCGContext *s, int ng, int nt) 3200 { 3201 int i; 3202 3203 for (i = 0; i < ng; ++i) { 3204 s->temps[i].state = TS_DEAD | TS_MEM; 3205 la_reset_pref(&s->temps[i]); 3206 } 3207 for (i = ng; i < nt; ++i) { 3208 s->temps[i].state = TS_DEAD; 3209 la_reset_pref(&s->temps[i]); 3210 } 3211 } 3212 3213 /* liveness analysis: end of basic block: all temps are dead, globals 3214 and local temps should be in memory. */ 3215 static void la_bb_end(TCGContext *s, int ng, int nt) 3216 { 3217 int i; 3218 3219 for (i = 0; i < nt; ++i) { 3220 TCGTemp *ts = &s->temps[i]; 3221 int state; 3222 3223 switch (ts->kind) { 3224 case TEMP_FIXED: 3225 case TEMP_GLOBAL: 3226 case TEMP_TB: 3227 state = TS_DEAD | TS_MEM; 3228 break; 3229 case TEMP_EBB: 3230 case TEMP_CONST: 3231 state = TS_DEAD; 3232 break; 3233 default: 3234 g_assert_not_reached(); 3235 } 3236 ts->state = state; 3237 la_reset_pref(ts); 3238 } 3239 } 3240 3241 /* liveness analysis: sync globals back to memory. */ 3242 static void la_global_sync(TCGContext *s, int ng) 3243 { 3244 int i; 3245 3246 for (i = 0; i < ng; ++i) { 3247 int state = s->temps[i].state; 3248 s->temps[i].state = state | TS_MEM; 3249 if (state == TS_DEAD) { 3250 /* If the global was previously dead, reset prefs. */ 3251 la_reset_pref(&s->temps[i]); 3252 } 3253 } 3254 } 3255 3256 /* 3257 * liveness analysis: conditional branch: all temps are dead unless 3258 * explicitly live-across-conditional-branch, globals and local temps 3259 * should be synced. 3260 */ 3261 static void la_bb_sync(TCGContext *s, int ng, int nt) 3262 { 3263 la_global_sync(s, ng); 3264 3265 for (int i = ng; i < nt; ++i) { 3266 TCGTemp *ts = &s->temps[i]; 3267 int state; 3268 3269 switch (ts->kind) { 3270 case TEMP_TB: 3271 state = ts->state; 3272 ts->state = state | TS_MEM; 3273 if (state != TS_DEAD) { 3274 continue; 3275 } 3276 break; 3277 case TEMP_EBB: 3278 case TEMP_CONST: 3279 continue; 3280 default: 3281 g_assert_not_reached(); 3282 } 3283 la_reset_pref(&s->temps[i]); 3284 } 3285 } 3286 3287 /* liveness analysis: sync globals back to memory and kill. */ 3288 static void la_global_kill(TCGContext *s, int ng) 3289 { 3290 int i; 3291 3292 for (i = 0; i < ng; i++) { 3293 s->temps[i].state = TS_DEAD | TS_MEM; 3294 la_reset_pref(&s->temps[i]); 3295 } 3296 } 3297 3298 /* liveness analysis: note live globals crossing calls. */ 3299 static void la_cross_call(TCGContext *s, int nt) 3300 { 3301 TCGRegSet mask = ~tcg_target_call_clobber_regs; 3302 int i; 3303 3304 for (i = 0; i < nt; i++) { 3305 TCGTemp *ts = &s->temps[i]; 3306 if (!(ts->state & TS_DEAD)) { 3307 TCGRegSet *pset = la_temp_pref(ts); 3308 TCGRegSet set = *pset; 3309 3310 set &= mask; 3311 /* If the combination is not possible, restart. */ 3312 if (set == 0) { 3313 set = tcg_target_available_regs[ts->type] & mask; 3314 } 3315 *pset = set; 3316 } 3317 } 3318 } 3319 3320 /* 3321 * Liveness analysis: Verify the lifetime of TEMP_TB, and reduce 3322 * to TEMP_EBB, if possible. 3323 */ 3324 static void __attribute__((noinline)) 3325 liveness_pass_0(TCGContext *s) 3326 { 3327 void * const multiple_ebb = (void *)(uintptr_t)-1; 3328 int nb_temps = s->nb_temps; 3329 TCGOp *op, *ebb; 3330 3331 for (int i = s->nb_globals; i < nb_temps; ++i) { 3332 s->temps[i].state_ptr = NULL; 3333 } 3334 3335 /* 3336 * Represent each EBB by the op at which it begins. In the case of 3337 * the first EBB, this is the first op, otherwise it is a label. 3338 * Collect the uses of each TEMP_TB: NULL for unused, EBB for use 3339 * within a single EBB, else MULTIPLE_EBB. 3340 */ 3341 ebb = QTAILQ_FIRST(&s->ops); 3342 QTAILQ_FOREACH(op, &s->ops, link) { 3343 const TCGOpDef *def; 3344 int nb_oargs, nb_iargs; 3345 3346 switch (op->opc) { 3347 case INDEX_op_set_label: 3348 ebb = op; 3349 continue; 3350 case INDEX_op_discard: 3351 continue; 3352 case INDEX_op_call: 3353 nb_oargs = TCGOP_CALLO(op); 3354 nb_iargs = TCGOP_CALLI(op); 3355 break; 3356 default: 3357 def = &tcg_op_defs[op->opc]; 3358 nb_oargs = def->nb_oargs; 3359 nb_iargs = def->nb_iargs; 3360 break; 3361 } 3362 3363 for (int i = 0; i < nb_oargs + nb_iargs; ++i) { 3364 TCGTemp *ts = arg_temp(op->args[i]); 3365 3366 if (ts->kind != TEMP_TB) { 3367 continue; 3368 } 3369 if (ts->state_ptr == NULL) { 3370 ts->state_ptr = ebb; 3371 } else if (ts->state_ptr != ebb) { 3372 ts->state_ptr = multiple_ebb; 3373 } 3374 } 3375 } 3376 3377 /* 3378 * For TEMP_TB that turned out not to be used beyond one EBB, 3379 * reduce the liveness to TEMP_EBB. 3380 */ 3381 for (int i = s->nb_globals; i < nb_temps; ++i) { 3382 TCGTemp *ts = &s->temps[i]; 3383 if (ts->kind == TEMP_TB && ts->state_ptr != multiple_ebb) { 3384 ts->kind = TEMP_EBB; 3385 } 3386 } 3387 } 3388 3389 /* Liveness analysis : update the opc_arg_life array to tell if a 3390 given input arguments is dead. Instructions updating dead 3391 temporaries are removed. */ 3392 static void __attribute__((noinline)) 3393 liveness_pass_1(TCGContext *s) 3394 { 3395 int nb_globals = s->nb_globals; 3396 int nb_temps = s->nb_temps; 3397 TCGOp *op, *op_prev; 3398 TCGRegSet *prefs; 3399 int i; 3400 3401 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps); 3402 for (i = 0; i < nb_temps; ++i) { 3403 s->temps[i].state_ptr = prefs + i; 3404 } 3405 3406 /* ??? Should be redundant with the exit_tb that ends the TB. */ 3407 la_func_end(s, nb_globals, nb_temps); 3408 3409 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) { 3410 int nb_iargs, nb_oargs; 3411 TCGOpcode opc_new, opc_new2; 3412 bool have_opc_new2; 3413 TCGLifeData arg_life = 0; 3414 TCGTemp *ts; 3415 TCGOpcode opc = op->opc; 3416 const TCGOpDef *def = &tcg_op_defs[opc]; 3417 3418 switch (opc) { 3419 case INDEX_op_call: 3420 { 3421 const TCGHelperInfo *info = tcg_call_info(op); 3422 int call_flags = tcg_call_flags(op); 3423 3424 nb_oargs = TCGOP_CALLO(op); 3425 nb_iargs = TCGOP_CALLI(op); 3426 3427 /* pure functions can be removed if their result is unused */ 3428 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) { 3429 for (i = 0; i < nb_oargs; i++) { 3430 ts = arg_temp(op->args[i]); 3431 if (ts->state != TS_DEAD) { 3432 goto do_not_remove_call; 3433 } 3434 } 3435 goto do_remove; 3436 } 3437 do_not_remove_call: 3438 3439 /* Output args are dead. */ 3440 for (i = 0; i < nb_oargs; i++) { 3441 ts = arg_temp(op->args[i]); 3442 if (ts->state & TS_DEAD) { 3443 arg_life |= DEAD_ARG << i; 3444 } 3445 if (ts->state & TS_MEM) { 3446 arg_life |= SYNC_ARG << i; 3447 } 3448 ts->state = TS_DEAD; 3449 la_reset_pref(ts); 3450 } 3451 3452 /* Not used -- it will be tcg_target_call_oarg_reg(). */ 3453 memset(op->output_pref, 0, sizeof(op->output_pref)); 3454 3455 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS | 3456 TCG_CALL_NO_READ_GLOBALS))) { 3457 la_global_kill(s, nb_globals); 3458 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) { 3459 la_global_sync(s, nb_globals); 3460 } 3461 3462 /* Record arguments that die in this helper. */ 3463 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3464 ts = arg_temp(op->args[i]); 3465 if (ts->state & TS_DEAD) { 3466 arg_life |= DEAD_ARG << i; 3467 } 3468 } 3469 3470 /* For all live registers, remove call-clobbered prefs. */ 3471 la_cross_call(s, nb_temps); 3472 3473 /* 3474 * Input arguments are live for preceding opcodes. 3475 * 3476 * For those arguments that die, and will be allocated in 3477 * registers, clear the register set for that arg, to be 3478 * filled in below. For args that will be on the stack, 3479 * reset to any available reg. Process arguments in reverse 3480 * order so that if a temp is used more than once, the stack 3481 * reset to max happens before the register reset to 0. 3482 */ 3483 for (i = nb_iargs - 1; i >= 0; i--) { 3484 const TCGCallArgumentLoc *loc = &info->in[i]; 3485 ts = arg_temp(op->args[nb_oargs + i]); 3486 3487 if (ts->state & TS_DEAD) { 3488 switch (loc->kind) { 3489 case TCG_CALL_ARG_NORMAL: 3490 case TCG_CALL_ARG_EXTEND_U: 3491 case TCG_CALL_ARG_EXTEND_S: 3492 if (arg_slot_reg_p(loc->arg_slot)) { 3493 *la_temp_pref(ts) = 0; 3494 break; 3495 } 3496 /* fall through */ 3497 default: 3498 *la_temp_pref(ts) = 3499 tcg_target_available_regs[ts->type]; 3500 break; 3501 } 3502 ts->state &= ~TS_DEAD; 3503 } 3504 } 3505 3506 /* 3507 * For each input argument, add its input register to prefs. 3508 * If a temp is used once, this produces a single set bit; 3509 * if a temp is used multiple times, this produces a set. 3510 */ 3511 for (i = 0; i < nb_iargs; i++) { 3512 const TCGCallArgumentLoc *loc = &info->in[i]; 3513 ts = arg_temp(op->args[nb_oargs + i]); 3514 3515 switch (loc->kind) { 3516 case TCG_CALL_ARG_NORMAL: 3517 case TCG_CALL_ARG_EXTEND_U: 3518 case TCG_CALL_ARG_EXTEND_S: 3519 if (arg_slot_reg_p(loc->arg_slot)) { 3520 tcg_regset_set_reg(*la_temp_pref(ts), 3521 tcg_target_call_iarg_regs[loc->arg_slot]); 3522 } 3523 break; 3524 default: 3525 break; 3526 } 3527 } 3528 } 3529 break; 3530 case INDEX_op_insn_start: 3531 break; 3532 case INDEX_op_discard: 3533 /* mark the temporary as dead */ 3534 ts = arg_temp(op->args[0]); 3535 ts->state = TS_DEAD; 3536 la_reset_pref(ts); 3537 break; 3538 3539 case INDEX_op_add2_i32: 3540 opc_new = INDEX_op_add_i32; 3541 goto do_addsub2; 3542 case INDEX_op_sub2_i32: 3543 opc_new = INDEX_op_sub_i32; 3544 goto do_addsub2; 3545 case INDEX_op_add2_i64: 3546 opc_new = INDEX_op_add_i64; 3547 goto do_addsub2; 3548 case INDEX_op_sub2_i64: 3549 opc_new = INDEX_op_sub_i64; 3550 do_addsub2: 3551 nb_iargs = 4; 3552 nb_oargs = 2; 3553 /* Test if the high part of the operation is dead, but not 3554 the low part. The result can be optimized to a simple 3555 add or sub. This happens often for x86_64 guest when the 3556 cpu mode is set to 32 bit. */ 3557 if (arg_temp(op->args[1])->state == TS_DEAD) { 3558 if (arg_temp(op->args[0])->state == TS_DEAD) { 3559 goto do_remove; 3560 } 3561 /* Replace the opcode and adjust the args in place, 3562 leaving 3 unused args at the end. */ 3563 op->opc = opc = opc_new; 3564 op->args[1] = op->args[2]; 3565 op->args[2] = op->args[4]; 3566 /* Fall through and mark the single-word operation live. */ 3567 nb_iargs = 2; 3568 nb_oargs = 1; 3569 } 3570 goto do_not_remove; 3571 3572 case INDEX_op_mulu2_i32: 3573 opc_new = INDEX_op_mul_i32; 3574 opc_new2 = INDEX_op_muluh_i32; 3575 have_opc_new2 = TCG_TARGET_HAS_muluh_i32; 3576 goto do_mul2; 3577 case INDEX_op_muls2_i32: 3578 opc_new = INDEX_op_mul_i32; 3579 opc_new2 = INDEX_op_mulsh_i32; 3580 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32; 3581 goto do_mul2; 3582 case INDEX_op_mulu2_i64: 3583 opc_new = INDEX_op_mul_i64; 3584 opc_new2 = INDEX_op_muluh_i64; 3585 have_opc_new2 = TCG_TARGET_HAS_muluh_i64; 3586 goto do_mul2; 3587 case INDEX_op_muls2_i64: 3588 opc_new = INDEX_op_mul_i64; 3589 opc_new2 = INDEX_op_mulsh_i64; 3590 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64; 3591 goto do_mul2; 3592 do_mul2: 3593 nb_iargs = 2; 3594 nb_oargs = 2; 3595 if (arg_temp(op->args[1])->state == TS_DEAD) { 3596 if (arg_temp(op->args[0])->state == TS_DEAD) { 3597 /* Both parts of the operation are dead. */ 3598 goto do_remove; 3599 } 3600 /* The high part of the operation is dead; generate the low. */ 3601 op->opc = opc = opc_new; 3602 op->args[1] = op->args[2]; 3603 op->args[2] = op->args[3]; 3604 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) { 3605 /* The low part of the operation is dead; generate the high. */ 3606 op->opc = opc = opc_new2; 3607 op->args[0] = op->args[1]; 3608 op->args[1] = op->args[2]; 3609 op->args[2] = op->args[3]; 3610 } else { 3611 goto do_not_remove; 3612 } 3613 /* Mark the single-word operation live. */ 3614 nb_oargs = 1; 3615 goto do_not_remove; 3616 3617 default: 3618 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */ 3619 nb_iargs = def->nb_iargs; 3620 nb_oargs = def->nb_oargs; 3621 3622 /* Test if the operation can be removed because all 3623 its outputs are dead. We assume that nb_oargs == 0 3624 implies side effects */ 3625 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { 3626 for (i = 0; i < nb_oargs; i++) { 3627 if (arg_temp(op->args[i])->state != TS_DEAD) { 3628 goto do_not_remove; 3629 } 3630 } 3631 goto do_remove; 3632 } 3633 goto do_not_remove; 3634 3635 do_remove: 3636 tcg_op_remove(s, op); 3637 break; 3638 3639 do_not_remove: 3640 for (i = 0; i < nb_oargs; i++) { 3641 ts = arg_temp(op->args[i]); 3642 3643 /* Remember the preference of the uses that followed. */ 3644 if (i < ARRAY_SIZE(op->output_pref)) { 3645 op->output_pref[i] = *la_temp_pref(ts); 3646 } 3647 3648 /* Output args are dead. */ 3649 if (ts->state & TS_DEAD) { 3650 arg_life |= DEAD_ARG << i; 3651 } 3652 if (ts->state & TS_MEM) { 3653 arg_life |= SYNC_ARG << i; 3654 } 3655 ts->state = TS_DEAD; 3656 la_reset_pref(ts); 3657 } 3658 3659 /* If end of basic block, update. */ 3660 if (def->flags & TCG_OPF_BB_EXIT) { 3661 la_func_end(s, nb_globals, nb_temps); 3662 } else if (def->flags & TCG_OPF_COND_BRANCH) { 3663 la_bb_sync(s, nb_globals, nb_temps); 3664 } else if (def->flags & TCG_OPF_BB_END) { 3665 la_bb_end(s, nb_globals, nb_temps); 3666 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3667 la_global_sync(s, nb_globals); 3668 if (def->flags & TCG_OPF_CALL_CLOBBER) { 3669 la_cross_call(s, nb_temps); 3670 } 3671 } 3672 3673 /* Record arguments that die in this opcode. */ 3674 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3675 ts = arg_temp(op->args[i]); 3676 if (ts->state & TS_DEAD) { 3677 arg_life |= DEAD_ARG << i; 3678 } 3679 } 3680 3681 /* Input arguments are live for preceding opcodes. */ 3682 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3683 ts = arg_temp(op->args[i]); 3684 if (ts->state & TS_DEAD) { 3685 /* For operands that were dead, initially allow 3686 all regs for the type. */ 3687 *la_temp_pref(ts) = tcg_target_available_regs[ts->type]; 3688 ts->state &= ~TS_DEAD; 3689 } 3690 } 3691 3692 /* Incorporate constraints for this operand. */ 3693 switch (opc) { 3694 case INDEX_op_mov_i32: 3695 case INDEX_op_mov_i64: 3696 /* Note that these are TCG_OPF_NOT_PRESENT and do not 3697 have proper constraints. That said, special case 3698 moves to propagate preferences backward. */ 3699 if (IS_DEAD_ARG(1)) { 3700 *la_temp_pref(arg_temp(op->args[0])) 3701 = *la_temp_pref(arg_temp(op->args[1])); 3702 } 3703 break; 3704 3705 default: 3706 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3707 const TCGArgConstraint *ct = &def->args_ct[i]; 3708 TCGRegSet set, *pset; 3709 3710 ts = arg_temp(op->args[i]); 3711 pset = la_temp_pref(ts); 3712 set = *pset; 3713 3714 set &= ct->regs; 3715 if (ct->ialias) { 3716 set &= output_pref(op, ct->alias_index); 3717 } 3718 /* If the combination is not possible, restart. */ 3719 if (set == 0) { 3720 set = ct->regs; 3721 } 3722 *pset = set; 3723 } 3724 break; 3725 } 3726 break; 3727 } 3728 op->life = arg_life; 3729 } 3730 } 3731 3732 /* Liveness analysis: Convert indirect regs to direct temporaries. */ 3733 static bool __attribute__((noinline)) 3734 liveness_pass_2(TCGContext *s) 3735 { 3736 int nb_globals = s->nb_globals; 3737 int nb_temps, i; 3738 bool changes = false; 3739 TCGOp *op, *op_next; 3740 3741 /* Create a temporary for each indirect global. */ 3742 for (i = 0; i < nb_globals; ++i) { 3743 TCGTemp *its = &s->temps[i]; 3744 if (its->indirect_reg) { 3745 TCGTemp *dts = tcg_temp_alloc(s); 3746 dts->type = its->type; 3747 dts->base_type = its->base_type; 3748 dts->temp_subindex = its->temp_subindex; 3749 dts->kind = TEMP_EBB; 3750 its->state_ptr = dts; 3751 } else { 3752 its->state_ptr = NULL; 3753 } 3754 /* All globals begin dead. */ 3755 its->state = TS_DEAD; 3756 } 3757 for (nb_temps = s->nb_temps; i < nb_temps; ++i) { 3758 TCGTemp *its = &s->temps[i]; 3759 its->state_ptr = NULL; 3760 its->state = TS_DEAD; 3761 } 3762 3763 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3764 TCGOpcode opc = op->opc; 3765 const TCGOpDef *def = &tcg_op_defs[opc]; 3766 TCGLifeData arg_life = op->life; 3767 int nb_iargs, nb_oargs, call_flags; 3768 TCGTemp *arg_ts, *dir_ts; 3769 3770 if (opc == INDEX_op_call) { 3771 nb_oargs = TCGOP_CALLO(op); 3772 nb_iargs = TCGOP_CALLI(op); 3773 call_flags = tcg_call_flags(op); 3774 } else { 3775 nb_iargs = def->nb_iargs; 3776 nb_oargs = def->nb_oargs; 3777 3778 /* Set flags similar to how calls require. */ 3779 if (def->flags & TCG_OPF_COND_BRANCH) { 3780 /* Like reading globals: sync_globals */ 3781 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 3782 } else if (def->flags & TCG_OPF_BB_END) { 3783 /* Like writing globals: save_globals */ 3784 call_flags = 0; 3785 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3786 /* Like reading globals: sync_globals */ 3787 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 3788 } else { 3789 /* No effect on globals. */ 3790 call_flags = (TCG_CALL_NO_READ_GLOBALS | 3791 TCG_CALL_NO_WRITE_GLOBALS); 3792 } 3793 } 3794 3795 /* Make sure that input arguments are available. */ 3796 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3797 arg_ts = arg_temp(op->args[i]); 3798 dir_ts = arg_ts->state_ptr; 3799 if (dir_ts && arg_ts->state == TS_DEAD) { 3800 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32 3801 ? INDEX_op_ld_i32 3802 : INDEX_op_ld_i64); 3803 TCGOp *lop = tcg_op_insert_before(s, op, lopc, 3); 3804 3805 lop->args[0] = temp_arg(dir_ts); 3806 lop->args[1] = temp_arg(arg_ts->mem_base); 3807 lop->args[2] = arg_ts->mem_offset; 3808 3809 /* Loaded, but synced with memory. */ 3810 arg_ts->state = TS_MEM; 3811 } 3812 } 3813 3814 /* Perform input replacement, and mark inputs that became dead. 3815 No action is required except keeping temp_state up to date 3816 so that we reload when needed. */ 3817 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3818 arg_ts = arg_temp(op->args[i]); 3819 dir_ts = arg_ts->state_ptr; 3820 if (dir_ts) { 3821 op->args[i] = temp_arg(dir_ts); 3822 changes = true; 3823 if (IS_DEAD_ARG(i)) { 3824 arg_ts->state = TS_DEAD; 3825 } 3826 } 3827 } 3828 3829 /* Liveness analysis should ensure that the following are 3830 all correct, for call sites and basic block end points. */ 3831 if (call_flags & TCG_CALL_NO_READ_GLOBALS) { 3832 /* Nothing to do */ 3833 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) { 3834 for (i = 0; i < nb_globals; ++i) { 3835 /* Liveness should see that globals are synced back, 3836 that is, either TS_DEAD or TS_MEM. */ 3837 arg_ts = &s->temps[i]; 3838 tcg_debug_assert(arg_ts->state_ptr == 0 3839 || arg_ts->state != 0); 3840 } 3841 } else { 3842 for (i = 0; i < nb_globals; ++i) { 3843 /* Liveness should see that globals are saved back, 3844 that is, TS_DEAD, waiting to be reloaded. */ 3845 arg_ts = &s->temps[i]; 3846 tcg_debug_assert(arg_ts->state_ptr == 0 3847 || arg_ts->state == TS_DEAD); 3848 } 3849 } 3850 3851 /* Outputs become available. */ 3852 if (opc == INDEX_op_mov_i32 || opc == INDEX_op_mov_i64) { 3853 arg_ts = arg_temp(op->args[0]); 3854 dir_ts = arg_ts->state_ptr; 3855 if (dir_ts) { 3856 op->args[0] = temp_arg(dir_ts); 3857 changes = true; 3858 3859 /* The output is now live and modified. */ 3860 arg_ts->state = 0; 3861 3862 if (NEED_SYNC_ARG(0)) { 3863 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 3864 ? INDEX_op_st_i32 3865 : INDEX_op_st_i64); 3866 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3); 3867 TCGTemp *out_ts = dir_ts; 3868 3869 if (IS_DEAD_ARG(0)) { 3870 out_ts = arg_temp(op->args[1]); 3871 arg_ts->state = TS_DEAD; 3872 tcg_op_remove(s, op); 3873 } else { 3874 arg_ts->state = TS_MEM; 3875 } 3876 3877 sop->args[0] = temp_arg(out_ts); 3878 sop->args[1] = temp_arg(arg_ts->mem_base); 3879 sop->args[2] = arg_ts->mem_offset; 3880 } else { 3881 tcg_debug_assert(!IS_DEAD_ARG(0)); 3882 } 3883 } 3884 } else { 3885 for (i = 0; i < nb_oargs; i++) { 3886 arg_ts = arg_temp(op->args[i]); 3887 dir_ts = arg_ts->state_ptr; 3888 if (!dir_ts) { 3889 continue; 3890 } 3891 op->args[i] = temp_arg(dir_ts); 3892 changes = true; 3893 3894 /* The output is now live and modified. */ 3895 arg_ts->state = 0; 3896 3897 /* Sync outputs upon their last write. */ 3898 if (NEED_SYNC_ARG(i)) { 3899 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 3900 ? INDEX_op_st_i32 3901 : INDEX_op_st_i64); 3902 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3); 3903 3904 sop->args[0] = temp_arg(dir_ts); 3905 sop->args[1] = temp_arg(arg_ts->mem_base); 3906 sop->args[2] = arg_ts->mem_offset; 3907 3908 arg_ts->state = TS_MEM; 3909 } 3910 /* Drop outputs that are dead. */ 3911 if (IS_DEAD_ARG(i)) { 3912 arg_ts->state = TS_DEAD; 3913 } 3914 } 3915 } 3916 } 3917 3918 return changes; 3919 } 3920 3921 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts) 3922 { 3923 intptr_t off; 3924 int size, align; 3925 3926 /* When allocating an object, look at the full type. */ 3927 size = tcg_type_size(ts->base_type); 3928 switch (ts->base_type) { 3929 case TCG_TYPE_I32: 3930 align = 4; 3931 break; 3932 case TCG_TYPE_I64: 3933 case TCG_TYPE_V64: 3934 align = 8; 3935 break; 3936 case TCG_TYPE_I128: 3937 case TCG_TYPE_V128: 3938 case TCG_TYPE_V256: 3939 /* 3940 * Note that we do not require aligned storage for V256, 3941 * and that we provide alignment for I128 to match V128, 3942 * even if that's above what the host ABI requires. 3943 */ 3944 align = 16; 3945 break; 3946 default: 3947 g_assert_not_reached(); 3948 } 3949 3950 /* 3951 * Assume the stack is sufficiently aligned. 3952 * This affects e.g. ARM NEON, where we have 8 byte stack alignment 3953 * and do not require 16 byte vector alignment. This seems slightly 3954 * easier than fully parameterizing the above switch statement. 3955 */ 3956 align = MIN(TCG_TARGET_STACK_ALIGN, align); 3957 off = ROUND_UP(s->current_frame_offset, align); 3958 3959 /* If we've exhausted the stack frame, restart with a smaller TB. */ 3960 if (off + size > s->frame_end) { 3961 tcg_raise_tb_overflow(s); 3962 } 3963 s->current_frame_offset = off + size; 3964 #if defined(__sparc__) 3965 off += TCG_TARGET_STACK_BIAS; 3966 #endif 3967 3968 /* If the object was subdivided, assign memory to all the parts. */ 3969 if (ts->base_type != ts->type) { 3970 int part_size = tcg_type_size(ts->type); 3971 int part_count = size / part_size; 3972 3973 /* 3974 * Each part is allocated sequentially in tcg_temp_new_internal. 3975 * Jump back to the first part by subtracting the current index. 3976 */ 3977 ts -= ts->temp_subindex; 3978 for (int i = 0; i < part_count; ++i) { 3979 ts[i].mem_offset = off + i * part_size; 3980 ts[i].mem_base = s->frame_temp; 3981 ts[i].mem_allocated = 1; 3982 } 3983 } else { 3984 ts->mem_offset = off; 3985 ts->mem_base = s->frame_temp; 3986 ts->mem_allocated = 1; 3987 } 3988 } 3989 3990 /* Assign @reg to @ts, and update reg_to_temp[]. */ 3991 static void set_temp_val_reg(TCGContext *s, TCGTemp *ts, TCGReg reg) 3992 { 3993 if (ts->val_type == TEMP_VAL_REG) { 3994 TCGReg old = ts->reg; 3995 tcg_debug_assert(s->reg_to_temp[old] == ts); 3996 if (old == reg) { 3997 return; 3998 } 3999 s->reg_to_temp[old] = NULL; 4000 } 4001 tcg_debug_assert(s->reg_to_temp[reg] == NULL); 4002 s->reg_to_temp[reg] = ts; 4003 ts->val_type = TEMP_VAL_REG; 4004 ts->reg = reg; 4005 } 4006 4007 /* Assign a non-register value type to @ts, and update reg_to_temp[]. */ 4008 static void set_temp_val_nonreg(TCGContext *s, TCGTemp *ts, TCGTempVal type) 4009 { 4010 tcg_debug_assert(type != TEMP_VAL_REG); 4011 if (ts->val_type == TEMP_VAL_REG) { 4012 TCGReg reg = ts->reg; 4013 tcg_debug_assert(s->reg_to_temp[reg] == ts); 4014 s->reg_to_temp[reg] = NULL; 4015 } 4016 ts->val_type = type; 4017 } 4018 4019 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet); 4020 4021 /* Mark a temporary as free or dead. If 'free_or_dead' is negative, 4022 mark it free; otherwise mark it dead. */ 4023 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead) 4024 { 4025 TCGTempVal new_type; 4026 4027 switch (ts->kind) { 4028 case TEMP_FIXED: 4029 return; 4030 case TEMP_GLOBAL: 4031 case TEMP_TB: 4032 new_type = TEMP_VAL_MEM; 4033 break; 4034 case TEMP_EBB: 4035 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD; 4036 break; 4037 case TEMP_CONST: 4038 new_type = TEMP_VAL_CONST; 4039 break; 4040 default: 4041 g_assert_not_reached(); 4042 } 4043 set_temp_val_nonreg(s, ts, new_type); 4044 } 4045 4046 /* Mark a temporary as dead. */ 4047 static inline void temp_dead(TCGContext *s, TCGTemp *ts) 4048 { 4049 temp_free_or_dead(s, ts, 1); 4050 } 4051 4052 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary 4053 registers needs to be allocated to store a constant. If 'free_or_dead' 4054 is non-zero, subsequently release the temporary; if it is positive, the 4055 temp is dead; if it is negative, the temp is free. */ 4056 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs, 4057 TCGRegSet preferred_regs, int free_or_dead) 4058 { 4059 if (!temp_readonly(ts) && !ts->mem_coherent) { 4060 if (!ts->mem_allocated) { 4061 temp_allocate_frame(s, ts); 4062 } 4063 switch (ts->val_type) { 4064 case TEMP_VAL_CONST: 4065 /* If we're going to free the temp immediately, then we won't 4066 require it later in a register, so attempt to store the 4067 constant to memory directly. */ 4068 if (free_or_dead 4069 && tcg_out_sti(s, ts->type, ts->val, 4070 ts->mem_base->reg, ts->mem_offset)) { 4071 break; 4072 } 4073 temp_load(s, ts, tcg_target_available_regs[ts->type], 4074 allocated_regs, preferred_regs); 4075 /* fallthrough */ 4076 4077 case TEMP_VAL_REG: 4078 tcg_out_st(s, ts->type, ts->reg, 4079 ts->mem_base->reg, ts->mem_offset); 4080 break; 4081 4082 case TEMP_VAL_MEM: 4083 break; 4084 4085 case TEMP_VAL_DEAD: 4086 default: 4087 g_assert_not_reached(); 4088 } 4089 ts->mem_coherent = 1; 4090 } 4091 if (free_or_dead) { 4092 temp_free_or_dead(s, ts, free_or_dead); 4093 } 4094 } 4095 4096 /* free register 'reg' by spilling the corresponding temporary if necessary */ 4097 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs) 4098 { 4099 TCGTemp *ts = s->reg_to_temp[reg]; 4100 if (ts != NULL) { 4101 temp_sync(s, ts, allocated_regs, 0, -1); 4102 } 4103 } 4104 4105 /** 4106 * tcg_reg_alloc: 4107 * @required_regs: Set of registers in which we must allocate. 4108 * @allocated_regs: Set of registers which must be avoided. 4109 * @preferred_regs: Set of registers we should prefer. 4110 * @rev: True if we search the registers in "indirect" order. 4111 * 4112 * The allocated register must be in @required_regs & ~@allocated_regs, 4113 * but if we can put it in @preferred_regs we may save a move later. 4114 */ 4115 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs, 4116 TCGRegSet allocated_regs, 4117 TCGRegSet preferred_regs, bool rev) 4118 { 4119 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4120 TCGRegSet reg_ct[2]; 4121 const int *order; 4122 4123 reg_ct[1] = required_regs & ~allocated_regs; 4124 tcg_debug_assert(reg_ct[1] != 0); 4125 reg_ct[0] = reg_ct[1] & preferred_regs; 4126 4127 /* Skip the preferred_regs option if it cannot be satisfied, 4128 or if the preference made no difference. */ 4129 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4130 4131 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4132 4133 /* Try free registers, preferences first. */ 4134 for (j = f; j < 2; j++) { 4135 TCGRegSet set = reg_ct[j]; 4136 4137 if (tcg_regset_single(set)) { 4138 /* One register in the set. */ 4139 TCGReg reg = tcg_regset_first(set); 4140 if (s->reg_to_temp[reg] == NULL) { 4141 return reg; 4142 } 4143 } else { 4144 for (i = 0; i < n; i++) { 4145 TCGReg reg = order[i]; 4146 if (s->reg_to_temp[reg] == NULL && 4147 tcg_regset_test_reg(set, reg)) { 4148 return reg; 4149 } 4150 } 4151 } 4152 } 4153 4154 /* We must spill something. */ 4155 for (j = f; j < 2; j++) { 4156 TCGRegSet set = reg_ct[j]; 4157 4158 if (tcg_regset_single(set)) { 4159 /* One register in the set. */ 4160 TCGReg reg = tcg_regset_first(set); 4161 tcg_reg_free(s, reg, allocated_regs); 4162 return reg; 4163 } else { 4164 for (i = 0; i < n; i++) { 4165 TCGReg reg = order[i]; 4166 if (tcg_regset_test_reg(set, reg)) { 4167 tcg_reg_free(s, reg, allocated_regs); 4168 return reg; 4169 } 4170 } 4171 } 4172 } 4173 4174 g_assert_not_reached(); 4175 } 4176 4177 static TCGReg tcg_reg_alloc_pair(TCGContext *s, TCGRegSet required_regs, 4178 TCGRegSet allocated_regs, 4179 TCGRegSet preferred_regs, bool rev) 4180 { 4181 int i, j, k, fmin, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4182 TCGRegSet reg_ct[2]; 4183 const int *order; 4184 4185 /* Ensure that if I is not in allocated_regs, I+1 is not either. */ 4186 reg_ct[1] = required_regs & ~(allocated_regs | (allocated_regs >> 1)); 4187 tcg_debug_assert(reg_ct[1] != 0); 4188 reg_ct[0] = reg_ct[1] & preferred_regs; 4189 4190 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4191 4192 /* 4193 * Skip the preferred_regs option if it cannot be satisfied, 4194 * or if the preference made no difference. 4195 */ 4196 k = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4197 4198 /* 4199 * Minimize the number of flushes by looking for 2 free registers first, 4200 * then a single flush, then two flushes. 4201 */ 4202 for (fmin = 2; fmin >= 0; fmin--) { 4203 for (j = k; j < 2; j++) { 4204 TCGRegSet set = reg_ct[j]; 4205 4206 for (i = 0; i < n; i++) { 4207 TCGReg reg = order[i]; 4208 4209 if (tcg_regset_test_reg(set, reg)) { 4210 int f = !s->reg_to_temp[reg] + !s->reg_to_temp[reg + 1]; 4211 if (f >= fmin) { 4212 tcg_reg_free(s, reg, allocated_regs); 4213 tcg_reg_free(s, reg + 1, allocated_regs); 4214 return reg; 4215 } 4216 } 4217 } 4218 } 4219 } 4220 g_assert_not_reached(); 4221 } 4222 4223 /* Make sure the temporary is in a register. If needed, allocate the register 4224 from DESIRED while avoiding ALLOCATED. */ 4225 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs, 4226 TCGRegSet allocated_regs, TCGRegSet preferred_regs) 4227 { 4228 TCGReg reg; 4229 4230 switch (ts->val_type) { 4231 case TEMP_VAL_REG: 4232 return; 4233 case TEMP_VAL_CONST: 4234 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4235 preferred_regs, ts->indirect_base); 4236 if (ts->type <= TCG_TYPE_I64) { 4237 tcg_out_movi(s, ts->type, reg, ts->val); 4238 } else { 4239 uint64_t val = ts->val; 4240 MemOp vece = MO_64; 4241 4242 /* 4243 * Find the minimal vector element that matches the constant. 4244 * The targets will, in general, have to do this search anyway, 4245 * do this generically. 4246 */ 4247 if (val == dup_const(MO_8, val)) { 4248 vece = MO_8; 4249 } else if (val == dup_const(MO_16, val)) { 4250 vece = MO_16; 4251 } else if (val == dup_const(MO_32, val)) { 4252 vece = MO_32; 4253 } 4254 4255 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val); 4256 } 4257 ts->mem_coherent = 0; 4258 break; 4259 case TEMP_VAL_MEM: 4260 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4261 preferred_regs, ts->indirect_base); 4262 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset); 4263 ts->mem_coherent = 1; 4264 break; 4265 case TEMP_VAL_DEAD: 4266 default: 4267 g_assert_not_reached(); 4268 } 4269 set_temp_val_reg(s, ts, reg); 4270 } 4271 4272 /* Save a temporary to memory. 'allocated_regs' is used in case a 4273 temporary registers needs to be allocated to store a constant. */ 4274 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs) 4275 { 4276 /* The liveness analysis already ensures that globals are back 4277 in memory. Keep an tcg_debug_assert for safety. */ 4278 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts)); 4279 } 4280 4281 /* save globals to their canonical location and assume they can be 4282 modified be the following code. 'allocated_regs' is used in case a 4283 temporary registers needs to be allocated to store a constant. */ 4284 static void save_globals(TCGContext *s, TCGRegSet allocated_regs) 4285 { 4286 int i, n; 4287 4288 for (i = 0, n = s->nb_globals; i < n; i++) { 4289 temp_save(s, &s->temps[i], allocated_regs); 4290 } 4291 } 4292 4293 /* sync globals to their canonical location and assume they can be 4294 read by the following code. 'allocated_regs' is used in case a 4295 temporary registers needs to be allocated to store a constant. */ 4296 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs) 4297 { 4298 int i, n; 4299 4300 for (i = 0, n = s->nb_globals; i < n; i++) { 4301 TCGTemp *ts = &s->temps[i]; 4302 tcg_debug_assert(ts->val_type != TEMP_VAL_REG 4303 || ts->kind == TEMP_FIXED 4304 || ts->mem_coherent); 4305 } 4306 } 4307 4308 /* at the end of a basic block, we assume all temporaries are dead and 4309 all globals are stored at their canonical location. */ 4310 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) 4311 { 4312 int i; 4313 4314 for (i = s->nb_globals; i < s->nb_temps; i++) { 4315 TCGTemp *ts = &s->temps[i]; 4316 4317 switch (ts->kind) { 4318 case TEMP_TB: 4319 temp_save(s, ts, allocated_regs); 4320 break; 4321 case TEMP_EBB: 4322 /* The liveness analysis already ensures that temps are dead. 4323 Keep an tcg_debug_assert for safety. */ 4324 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD); 4325 break; 4326 case TEMP_CONST: 4327 /* Similarly, we should have freed any allocated register. */ 4328 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST); 4329 break; 4330 default: 4331 g_assert_not_reached(); 4332 } 4333 } 4334 4335 save_globals(s, allocated_regs); 4336 } 4337 4338 /* 4339 * At a conditional branch, we assume all temporaries are dead unless 4340 * explicitly live-across-conditional-branch; all globals and local 4341 * temps are synced to their location. 4342 */ 4343 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs) 4344 { 4345 sync_globals(s, allocated_regs); 4346 4347 for (int i = s->nb_globals; i < s->nb_temps; i++) { 4348 TCGTemp *ts = &s->temps[i]; 4349 /* 4350 * The liveness analysis already ensures that temps are dead. 4351 * Keep tcg_debug_asserts for safety. 4352 */ 4353 switch (ts->kind) { 4354 case TEMP_TB: 4355 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent); 4356 break; 4357 case TEMP_EBB: 4358 case TEMP_CONST: 4359 break; 4360 default: 4361 g_assert_not_reached(); 4362 } 4363 } 4364 } 4365 4366 /* 4367 * Specialized code generation for INDEX_op_mov_* with a constant. 4368 */ 4369 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots, 4370 tcg_target_ulong val, TCGLifeData arg_life, 4371 TCGRegSet preferred_regs) 4372 { 4373 /* ENV should not be modified. */ 4374 tcg_debug_assert(!temp_readonly(ots)); 4375 4376 /* The movi is not explicitly generated here. */ 4377 set_temp_val_nonreg(s, ots, TEMP_VAL_CONST); 4378 ots->val = val; 4379 ots->mem_coherent = 0; 4380 if (NEED_SYNC_ARG(0)) { 4381 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0)); 4382 } else if (IS_DEAD_ARG(0)) { 4383 temp_dead(s, ots); 4384 } 4385 } 4386 4387 /* 4388 * Specialized code generation for INDEX_op_mov_*. 4389 */ 4390 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op) 4391 { 4392 const TCGLifeData arg_life = op->life; 4393 TCGRegSet allocated_regs, preferred_regs; 4394 TCGTemp *ts, *ots; 4395 TCGType otype, itype; 4396 TCGReg oreg, ireg; 4397 4398 allocated_regs = s->reserved_regs; 4399 preferred_regs = output_pref(op, 0); 4400 ots = arg_temp(op->args[0]); 4401 ts = arg_temp(op->args[1]); 4402 4403 /* ENV should not be modified. */ 4404 tcg_debug_assert(!temp_readonly(ots)); 4405 4406 /* Note that otype != itype for no-op truncation. */ 4407 otype = ots->type; 4408 itype = ts->type; 4409 4410 if (ts->val_type == TEMP_VAL_CONST) { 4411 /* propagate constant or generate sti */ 4412 tcg_target_ulong val = ts->val; 4413 if (IS_DEAD_ARG(1)) { 4414 temp_dead(s, ts); 4415 } 4416 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs); 4417 return; 4418 } 4419 4420 /* If the source value is in memory we're going to be forced 4421 to have it in a register in order to perform the copy. Copy 4422 the SOURCE value into its own register first, that way we 4423 don't have to reload SOURCE the next time it is used. */ 4424 if (ts->val_type == TEMP_VAL_MEM) { 4425 temp_load(s, ts, tcg_target_available_regs[itype], 4426 allocated_regs, preferred_regs); 4427 } 4428 tcg_debug_assert(ts->val_type == TEMP_VAL_REG); 4429 ireg = ts->reg; 4430 4431 if (IS_DEAD_ARG(0)) { 4432 /* mov to a non-saved dead register makes no sense (even with 4433 liveness analysis disabled). */ 4434 tcg_debug_assert(NEED_SYNC_ARG(0)); 4435 if (!ots->mem_allocated) { 4436 temp_allocate_frame(s, ots); 4437 } 4438 tcg_out_st(s, otype, ireg, ots->mem_base->reg, ots->mem_offset); 4439 if (IS_DEAD_ARG(1)) { 4440 temp_dead(s, ts); 4441 } 4442 temp_dead(s, ots); 4443 return; 4444 } 4445 4446 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) { 4447 /* 4448 * The mov can be suppressed. Kill input first, so that it 4449 * is unlinked from reg_to_temp, then set the output to the 4450 * reg that we saved from the input. 4451 */ 4452 temp_dead(s, ts); 4453 oreg = ireg; 4454 } else { 4455 if (ots->val_type == TEMP_VAL_REG) { 4456 oreg = ots->reg; 4457 } else { 4458 /* Make sure to not spill the input register during allocation. */ 4459 oreg = tcg_reg_alloc(s, tcg_target_available_regs[otype], 4460 allocated_regs | ((TCGRegSet)1 << ireg), 4461 preferred_regs, ots->indirect_base); 4462 } 4463 if (!tcg_out_mov(s, otype, oreg, ireg)) { 4464 /* 4465 * Cross register class move not supported. 4466 * Store the source register into the destination slot 4467 * and leave the destination temp as TEMP_VAL_MEM. 4468 */ 4469 assert(!temp_readonly(ots)); 4470 if (!ts->mem_allocated) { 4471 temp_allocate_frame(s, ots); 4472 } 4473 tcg_out_st(s, ts->type, ireg, ots->mem_base->reg, ots->mem_offset); 4474 set_temp_val_nonreg(s, ts, TEMP_VAL_MEM); 4475 ots->mem_coherent = 1; 4476 return; 4477 } 4478 } 4479 set_temp_val_reg(s, ots, oreg); 4480 ots->mem_coherent = 0; 4481 4482 if (NEED_SYNC_ARG(0)) { 4483 temp_sync(s, ots, allocated_regs, 0, 0); 4484 } 4485 } 4486 4487 /* 4488 * Specialized code generation for INDEX_op_dup_vec. 4489 */ 4490 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op) 4491 { 4492 const TCGLifeData arg_life = op->life; 4493 TCGRegSet dup_out_regs, dup_in_regs; 4494 TCGTemp *its, *ots; 4495 TCGType itype, vtype; 4496 unsigned vece; 4497 int lowpart_ofs; 4498 bool ok; 4499 4500 ots = arg_temp(op->args[0]); 4501 its = arg_temp(op->args[1]); 4502 4503 /* ENV should not be modified. */ 4504 tcg_debug_assert(!temp_readonly(ots)); 4505 4506 itype = its->type; 4507 vece = TCGOP_VECE(op); 4508 vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 4509 4510 if (its->val_type == TEMP_VAL_CONST) { 4511 /* Propagate constant via movi -> dupi. */ 4512 tcg_target_ulong val = its->val; 4513 if (IS_DEAD_ARG(1)) { 4514 temp_dead(s, its); 4515 } 4516 tcg_reg_alloc_do_movi(s, ots, val, arg_life, output_pref(op, 0)); 4517 return; 4518 } 4519 4520 dup_out_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 4521 dup_in_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[1].regs; 4522 4523 /* Allocate the output register now. */ 4524 if (ots->val_type != TEMP_VAL_REG) { 4525 TCGRegSet allocated_regs = s->reserved_regs; 4526 TCGReg oreg; 4527 4528 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) { 4529 /* Make sure to not spill the input register. */ 4530 tcg_regset_set_reg(allocated_regs, its->reg); 4531 } 4532 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 4533 output_pref(op, 0), ots->indirect_base); 4534 set_temp_val_reg(s, ots, oreg); 4535 } 4536 4537 switch (its->val_type) { 4538 case TEMP_VAL_REG: 4539 /* 4540 * The dup constriaints must be broad, covering all possible VECE. 4541 * However, tcg_op_dup_vec() gets to see the VECE and we allow it 4542 * to fail, indicating that extra moves are required for that case. 4543 */ 4544 if (tcg_regset_test_reg(dup_in_regs, its->reg)) { 4545 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) { 4546 goto done; 4547 } 4548 /* Try again from memory or a vector input register. */ 4549 } 4550 if (!its->mem_coherent) { 4551 /* 4552 * The input register is not synced, and so an extra store 4553 * would be required to use memory. Attempt an integer-vector 4554 * register move first. We do not have a TCGRegSet for this. 4555 */ 4556 if (tcg_out_mov(s, itype, ots->reg, its->reg)) { 4557 break; 4558 } 4559 /* Sync the temp back to its slot and load from there. */ 4560 temp_sync(s, its, s->reserved_regs, 0, 0); 4561 } 4562 /* fall through */ 4563 4564 case TEMP_VAL_MEM: 4565 lowpart_ofs = 0; 4566 if (HOST_BIG_ENDIAN) { 4567 lowpart_ofs = tcg_type_size(itype) - (1 << vece); 4568 } 4569 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg, 4570 its->mem_offset + lowpart_ofs)) { 4571 goto done; 4572 } 4573 /* Load the input into the destination vector register. */ 4574 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset); 4575 break; 4576 4577 default: 4578 g_assert_not_reached(); 4579 } 4580 4581 /* We now have a vector input register, so dup must succeed. */ 4582 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg); 4583 tcg_debug_assert(ok); 4584 4585 done: 4586 ots->mem_coherent = 0; 4587 if (IS_DEAD_ARG(1)) { 4588 temp_dead(s, its); 4589 } 4590 if (NEED_SYNC_ARG(0)) { 4591 temp_sync(s, ots, s->reserved_regs, 0, 0); 4592 } 4593 if (IS_DEAD_ARG(0)) { 4594 temp_dead(s, ots); 4595 } 4596 } 4597 4598 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op) 4599 { 4600 const TCGLifeData arg_life = op->life; 4601 const TCGOpDef * const def = &tcg_op_defs[op->opc]; 4602 TCGRegSet i_allocated_regs; 4603 TCGRegSet o_allocated_regs; 4604 int i, k, nb_iargs, nb_oargs; 4605 TCGReg reg; 4606 TCGArg arg; 4607 const TCGArgConstraint *arg_ct; 4608 TCGTemp *ts; 4609 TCGArg new_args[TCG_MAX_OP_ARGS]; 4610 int const_args[TCG_MAX_OP_ARGS]; 4611 4612 nb_oargs = def->nb_oargs; 4613 nb_iargs = def->nb_iargs; 4614 4615 /* copy constants */ 4616 memcpy(new_args + nb_oargs + nb_iargs, 4617 op->args + nb_oargs + nb_iargs, 4618 sizeof(TCGArg) * def->nb_cargs); 4619 4620 i_allocated_regs = s->reserved_regs; 4621 o_allocated_regs = s->reserved_regs; 4622 4623 /* satisfy input constraints */ 4624 for (k = 0; k < nb_iargs; k++) { 4625 TCGRegSet i_preferred_regs, i_required_regs; 4626 bool allocate_new_reg, copyto_new_reg; 4627 TCGTemp *ts2; 4628 int i1, i2; 4629 4630 i = def->args_ct[nb_oargs + k].sort_index; 4631 arg = op->args[i]; 4632 arg_ct = &def->args_ct[i]; 4633 ts = arg_temp(arg); 4634 4635 if (ts->val_type == TEMP_VAL_CONST 4636 && tcg_target_const_match(ts->val, ts->type, arg_ct->ct)) { 4637 /* constant is OK for instruction */ 4638 const_args[i] = 1; 4639 new_args[i] = ts->val; 4640 continue; 4641 } 4642 4643 reg = ts->reg; 4644 i_preferred_regs = 0; 4645 i_required_regs = arg_ct->regs; 4646 allocate_new_reg = false; 4647 copyto_new_reg = false; 4648 4649 switch (arg_ct->pair) { 4650 case 0: /* not paired */ 4651 if (arg_ct->ialias) { 4652 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4653 4654 /* 4655 * If the input is readonly, then it cannot also be an 4656 * output and aliased to itself. If the input is not 4657 * dead after the instruction, we must allocate a new 4658 * register and move it. 4659 */ 4660 if (temp_readonly(ts) || !IS_DEAD_ARG(i)) { 4661 allocate_new_reg = true; 4662 } else if (ts->val_type == TEMP_VAL_REG) { 4663 /* 4664 * Check if the current register has already been 4665 * allocated for another input. 4666 */ 4667 allocate_new_reg = 4668 tcg_regset_test_reg(i_allocated_regs, reg); 4669 } 4670 } 4671 if (!allocate_new_reg) { 4672 temp_load(s, ts, i_required_regs, i_allocated_regs, 4673 i_preferred_regs); 4674 reg = ts->reg; 4675 allocate_new_reg = !tcg_regset_test_reg(i_required_regs, reg); 4676 } 4677 if (allocate_new_reg) { 4678 /* 4679 * Allocate a new register matching the constraint 4680 * and move the temporary register into it. 4681 */ 4682 temp_load(s, ts, tcg_target_available_regs[ts->type], 4683 i_allocated_regs, 0); 4684 reg = tcg_reg_alloc(s, i_required_regs, i_allocated_regs, 4685 i_preferred_regs, ts->indirect_base); 4686 copyto_new_reg = true; 4687 } 4688 break; 4689 4690 case 1: 4691 /* First of an input pair; if i1 == i2, the second is an output. */ 4692 i1 = i; 4693 i2 = arg_ct->pair_index; 4694 ts2 = i1 != i2 ? arg_temp(op->args[i2]) : NULL; 4695 4696 /* 4697 * It is easier to default to allocating a new pair 4698 * and to identify a few cases where it's not required. 4699 */ 4700 if (arg_ct->ialias) { 4701 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4702 if (IS_DEAD_ARG(i1) && 4703 IS_DEAD_ARG(i2) && 4704 !temp_readonly(ts) && 4705 ts->val_type == TEMP_VAL_REG && 4706 ts->reg < TCG_TARGET_NB_REGS - 1 && 4707 tcg_regset_test_reg(i_required_regs, reg) && 4708 !tcg_regset_test_reg(i_allocated_regs, reg) && 4709 !tcg_regset_test_reg(i_allocated_regs, reg + 1) && 4710 (ts2 4711 ? ts2->val_type == TEMP_VAL_REG && 4712 ts2->reg == reg + 1 && 4713 !temp_readonly(ts2) 4714 : s->reg_to_temp[reg + 1] == NULL)) { 4715 break; 4716 } 4717 } else { 4718 /* Without aliasing, the pair must also be an input. */ 4719 tcg_debug_assert(ts2); 4720 if (ts->val_type == TEMP_VAL_REG && 4721 ts2->val_type == TEMP_VAL_REG && 4722 ts2->reg == reg + 1 && 4723 tcg_regset_test_reg(i_required_regs, reg)) { 4724 break; 4725 } 4726 } 4727 reg = tcg_reg_alloc_pair(s, i_required_regs, i_allocated_regs, 4728 0, ts->indirect_base); 4729 goto do_pair; 4730 4731 case 2: /* pair second */ 4732 reg = new_args[arg_ct->pair_index] + 1; 4733 goto do_pair; 4734 4735 case 3: /* ialias with second output, no first input */ 4736 tcg_debug_assert(arg_ct->ialias); 4737 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4738 4739 if (IS_DEAD_ARG(i) && 4740 !temp_readonly(ts) && 4741 ts->val_type == TEMP_VAL_REG && 4742 reg > 0 && 4743 s->reg_to_temp[reg - 1] == NULL && 4744 tcg_regset_test_reg(i_required_regs, reg) && 4745 !tcg_regset_test_reg(i_allocated_regs, reg) && 4746 !tcg_regset_test_reg(i_allocated_regs, reg - 1)) { 4747 tcg_regset_set_reg(i_allocated_regs, reg - 1); 4748 break; 4749 } 4750 reg = tcg_reg_alloc_pair(s, i_required_regs >> 1, 4751 i_allocated_regs, 0, 4752 ts->indirect_base); 4753 tcg_regset_set_reg(i_allocated_regs, reg); 4754 reg += 1; 4755 goto do_pair; 4756 4757 do_pair: 4758 /* 4759 * If an aliased input is not dead after the instruction, 4760 * we must allocate a new register and move it. 4761 */ 4762 if (arg_ct->ialias && (!IS_DEAD_ARG(i) || temp_readonly(ts))) { 4763 TCGRegSet t_allocated_regs = i_allocated_regs; 4764 4765 /* 4766 * Because of the alias, and the continued life, make sure 4767 * that the temp is somewhere *other* than the reg pair, 4768 * and we get a copy in reg. 4769 */ 4770 tcg_regset_set_reg(t_allocated_regs, reg); 4771 tcg_regset_set_reg(t_allocated_regs, reg + 1); 4772 if (ts->val_type == TEMP_VAL_REG && ts->reg == reg) { 4773 /* If ts was already in reg, copy it somewhere else. */ 4774 TCGReg nr; 4775 bool ok; 4776 4777 tcg_debug_assert(ts->kind != TEMP_FIXED); 4778 nr = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], 4779 t_allocated_regs, 0, ts->indirect_base); 4780 ok = tcg_out_mov(s, ts->type, nr, reg); 4781 tcg_debug_assert(ok); 4782 4783 set_temp_val_reg(s, ts, nr); 4784 } else { 4785 temp_load(s, ts, tcg_target_available_regs[ts->type], 4786 t_allocated_regs, 0); 4787 copyto_new_reg = true; 4788 } 4789 } else { 4790 /* Preferably allocate to reg, otherwise copy. */ 4791 i_required_regs = (TCGRegSet)1 << reg; 4792 temp_load(s, ts, i_required_regs, i_allocated_regs, 4793 i_preferred_regs); 4794 copyto_new_reg = ts->reg != reg; 4795 } 4796 break; 4797 4798 default: 4799 g_assert_not_reached(); 4800 } 4801 4802 if (copyto_new_reg) { 4803 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 4804 /* 4805 * Cross register class move not supported. Sync the 4806 * temp back to its slot and load from there. 4807 */ 4808 temp_sync(s, ts, i_allocated_regs, 0, 0); 4809 tcg_out_ld(s, ts->type, reg, 4810 ts->mem_base->reg, ts->mem_offset); 4811 } 4812 } 4813 new_args[i] = reg; 4814 const_args[i] = 0; 4815 tcg_regset_set_reg(i_allocated_regs, reg); 4816 } 4817 4818 /* mark dead temporaries and free the associated registers */ 4819 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 4820 if (IS_DEAD_ARG(i)) { 4821 temp_dead(s, arg_temp(op->args[i])); 4822 } 4823 } 4824 4825 if (def->flags & TCG_OPF_COND_BRANCH) { 4826 tcg_reg_alloc_cbranch(s, i_allocated_regs); 4827 } else if (def->flags & TCG_OPF_BB_END) { 4828 tcg_reg_alloc_bb_end(s, i_allocated_regs); 4829 } else { 4830 if (def->flags & TCG_OPF_CALL_CLOBBER) { 4831 /* XXX: permit generic clobber register list ? */ 4832 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 4833 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 4834 tcg_reg_free(s, i, i_allocated_regs); 4835 } 4836 } 4837 } 4838 if (def->flags & TCG_OPF_SIDE_EFFECTS) { 4839 /* sync globals if the op has side effects and might trigger 4840 an exception. */ 4841 sync_globals(s, i_allocated_regs); 4842 } 4843 4844 /* satisfy the output constraints */ 4845 for(k = 0; k < nb_oargs; k++) { 4846 i = def->args_ct[k].sort_index; 4847 arg = op->args[i]; 4848 arg_ct = &def->args_ct[i]; 4849 ts = arg_temp(arg); 4850 4851 /* ENV should not be modified. */ 4852 tcg_debug_assert(!temp_readonly(ts)); 4853 4854 switch (arg_ct->pair) { 4855 case 0: /* not paired */ 4856 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) { 4857 reg = new_args[arg_ct->alias_index]; 4858 } else if (arg_ct->newreg) { 4859 reg = tcg_reg_alloc(s, arg_ct->regs, 4860 i_allocated_regs | o_allocated_regs, 4861 output_pref(op, k), ts->indirect_base); 4862 } else { 4863 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs, 4864 output_pref(op, k), ts->indirect_base); 4865 } 4866 break; 4867 4868 case 1: /* first of pair */ 4869 tcg_debug_assert(!arg_ct->newreg); 4870 if (arg_ct->oalias) { 4871 reg = new_args[arg_ct->alias_index]; 4872 break; 4873 } 4874 reg = tcg_reg_alloc_pair(s, arg_ct->regs, o_allocated_regs, 4875 output_pref(op, k), ts->indirect_base); 4876 break; 4877 4878 case 2: /* second of pair */ 4879 tcg_debug_assert(!arg_ct->newreg); 4880 if (arg_ct->oalias) { 4881 reg = new_args[arg_ct->alias_index]; 4882 } else { 4883 reg = new_args[arg_ct->pair_index] + 1; 4884 } 4885 break; 4886 4887 case 3: /* first of pair, aliasing with a second input */ 4888 tcg_debug_assert(!arg_ct->newreg); 4889 reg = new_args[arg_ct->pair_index] - 1; 4890 break; 4891 4892 default: 4893 g_assert_not_reached(); 4894 } 4895 tcg_regset_set_reg(o_allocated_regs, reg); 4896 set_temp_val_reg(s, ts, reg); 4897 ts->mem_coherent = 0; 4898 new_args[i] = reg; 4899 } 4900 } 4901 4902 /* emit instruction */ 4903 switch (op->opc) { 4904 case INDEX_op_ext8s_i32: 4905 tcg_out_ext8s(s, TCG_TYPE_I32, new_args[0], new_args[1]); 4906 break; 4907 case INDEX_op_ext8s_i64: 4908 tcg_out_ext8s(s, TCG_TYPE_I64, new_args[0], new_args[1]); 4909 break; 4910 case INDEX_op_ext8u_i32: 4911 case INDEX_op_ext8u_i64: 4912 tcg_out_ext8u(s, new_args[0], new_args[1]); 4913 break; 4914 case INDEX_op_ext16s_i32: 4915 tcg_out_ext16s(s, TCG_TYPE_I32, new_args[0], new_args[1]); 4916 break; 4917 case INDEX_op_ext16s_i64: 4918 tcg_out_ext16s(s, TCG_TYPE_I64, new_args[0], new_args[1]); 4919 break; 4920 case INDEX_op_ext16u_i32: 4921 case INDEX_op_ext16u_i64: 4922 tcg_out_ext16u(s, new_args[0], new_args[1]); 4923 break; 4924 case INDEX_op_ext32s_i64: 4925 tcg_out_ext32s(s, new_args[0], new_args[1]); 4926 break; 4927 case INDEX_op_ext32u_i64: 4928 tcg_out_ext32u(s, new_args[0], new_args[1]); 4929 break; 4930 case INDEX_op_ext_i32_i64: 4931 tcg_out_exts_i32_i64(s, new_args[0], new_args[1]); 4932 break; 4933 case INDEX_op_extu_i32_i64: 4934 tcg_out_extu_i32_i64(s, new_args[0], new_args[1]); 4935 break; 4936 case INDEX_op_extrl_i64_i32: 4937 tcg_out_extrl_i64_i32(s, new_args[0], new_args[1]); 4938 break; 4939 default: 4940 if (def->flags & TCG_OPF_VECTOR) { 4941 tcg_out_vec_op(s, op->opc, TCGOP_VECL(op), TCGOP_VECE(op), 4942 new_args, const_args); 4943 } else { 4944 tcg_out_op(s, op->opc, new_args, const_args); 4945 } 4946 break; 4947 } 4948 4949 /* move the outputs in the correct register if needed */ 4950 for(i = 0; i < nb_oargs; i++) { 4951 ts = arg_temp(op->args[i]); 4952 4953 /* ENV should not be modified. */ 4954 tcg_debug_assert(!temp_readonly(ts)); 4955 4956 if (NEED_SYNC_ARG(i)) { 4957 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i)); 4958 } else if (IS_DEAD_ARG(i)) { 4959 temp_dead(s, ts); 4960 } 4961 } 4962 } 4963 4964 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op) 4965 { 4966 const TCGLifeData arg_life = op->life; 4967 TCGTemp *ots, *itsl, *itsh; 4968 TCGType vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 4969 4970 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */ 4971 tcg_debug_assert(TCG_TARGET_REG_BITS == 32); 4972 tcg_debug_assert(TCGOP_VECE(op) == MO_64); 4973 4974 ots = arg_temp(op->args[0]); 4975 itsl = arg_temp(op->args[1]); 4976 itsh = arg_temp(op->args[2]); 4977 4978 /* ENV should not be modified. */ 4979 tcg_debug_assert(!temp_readonly(ots)); 4980 4981 /* Allocate the output register now. */ 4982 if (ots->val_type != TEMP_VAL_REG) { 4983 TCGRegSet allocated_regs = s->reserved_regs; 4984 TCGRegSet dup_out_regs = 4985 tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 4986 TCGReg oreg; 4987 4988 /* Make sure to not spill the input registers. */ 4989 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) { 4990 tcg_regset_set_reg(allocated_regs, itsl->reg); 4991 } 4992 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) { 4993 tcg_regset_set_reg(allocated_regs, itsh->reg); 4994 } 4995 4996 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 4997 output_pref(op, 0), ots->indirect_base); 4998 set_temp_val_reg(s, ots, oreg); 4999 } 5000 5001 /* Promote dup2 of immediates to dupi_vec. */ 5002 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) { 5003 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val); 5004 MemOp vece = MO_64; 5005 5006 if (val == dup_const(MO_8, val)) { 5007 vece = MO_8; 5008 } else if (val == dup_const(MO_16, val)) { 5009 vece = MO_16; 5010 } else if (val == dup_const(MO_32, val)) { 5011 vece = MO_32; 5012 } 5013 5014 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val); 5015 goto done; 5016 } 5017 5018 /* If the two inputs form one 64-bit value, try dupm_vec. */ 5019 if (itsl->temp_subindex == HOST_BIG_ENDIAN && 5020 itsh->temp_subindex == !HOST_BIG_ENDIAN && 5021 itsl == itsh + (HOST_BIG_ENDIAN ? 1 : -1)) { 5022 TCGTemp *its = itsl - HOST_BIG_ENDIAN; 5023 5024 temp_sync(s, its + 0, s->reserved_regs, 0, 0); 5025 temp_sync(s, its + 1, s->reserved_regs, 0, 0); 5026 5027 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg, 5028 its->mem_base->reg, its->mem_offset)) { 5029 goto done; 5030 } 5031 } 5032 5033 /* Fall back to generic expansion. */ 5034 return false; 5035 5036 done: 5037 ots->mem_coherent = 0; 5038 if (IS_DEAD_ARG(1)) { 5039 temp_dead(s, itsl); 5040 } 5041 if (IS_DEAD_ARG(2)) { 5042 temp_dead(s, itsh); 5043 } 5044 if (NEED_SYNC_ARG(0)) { 5045 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0)); 5046 } else if (IS_DEAD_ARG(0)) { 5047 temp_dead(s, ots); 5048 } 5049 return true; 5050 } 5051 5052 static void load_arg_reg(TCGContext *s, TCGReg reg, TCGTemp *ts, 5053 TCGRegSet allocated_regs) 5054 { 5055 if (ts->val_type == TEMP_VAL_REG) { 5056 if (ts->reg != reg) { 5057 tcg_reg_free(s, reg, allocated_regs); 5058 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 5059 /* 5060 * Cross register class move not supported. Sync the 5061 * temp back to its slot and load from there. 5062 */ 5063 temp_sync(s, ts, allocated_regs, 0, 0); 5064 tcg_out_ld(s, ts->type, reg, 5065 ts->mem_base->reg, ts->mem_offset); 5066 } 5067 } 5068 } else { 5069 TCGRegSet arg_set = 0; 5070 5071 tcg_reg_free(s, reg, allocated_regs); 5072 tcg_regset_set_reg(arg_set, reg); 5073 temp_load(s, ts, arg_set, allocated_regs, 0); 5074 } 5075 } 5076 5077 static void load_arg_stk(TCGContext *s, unsigned arg_slot, TCGTemp *ts, 5078 TCGRegSet allocated_regs) 5079 { 5080 /* 5081 * When the destination is on the stack, load up the temp and store. 5082 * If there are many call-saved registers, the temp might live to 5083 * see another use; otherwise it'll be discarded. 5084 */ 5085 temp_load(s, ts, tcg_target_available_regs[ts->type], allocated_regs, 0); 5086 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, 5087 arg_slot_stk_ofs(arg_slot)); 5088 } 5089 5090 static void load_arg_normal(TCGContext *s, const TCGCallArgumentLoc *l, 5091 TCGTemp *ts, TCGRegSet *allocated_regs) 5092 { 5093 if (arg_slot_reg_p(l->arg_slot)) { 5094 TCGReg reg = tcg_target_call_iarg_regs[l->arg_slot]; 5095 load_arg_reg(s, reg, ts, *allocated_regs); 5096 tcg_regset_set_reg(*allocated_regs, reg); 5097 } else { 5098 load_arg_stk(s, l->arg_slot, ts, *allocated_regs); 5099 } 5100 } 5101 5102 static void load_arg_ref(TCGContext *s, unsigned arg_slot, TCGReg ref_base, 5103 intptr_t ref_off, TCGRegSet *allocated_regs) 5104 { 5105 TCGReg reg; 5106 5107 if (arg_slot_reg_p(arg_slot)) { 5108 reg = tcg_target_call_iarg_regs[arg_slot]; 5109 tcg_reg_free(s, reg, *allocated_regs); 5110 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5111 tcg_regset_set_reg(*allocated_regs, reg); 5112 } else { 5113 reg = tcg_reg_alloc(s, tcg_target_available_regs[TCG_TYPE_PTR], 5114 *allocated_regs, 0, false); 5115 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5116 tcg_out_st(s, TCG_TYPE_PTR, reg, TCG_REG_CALL_STACK, 5117 arg_slot_stk_ofs(arg_slot)); 5118 } 5119 } 5120 5121 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op) 5122 { 5123 const int nb_oargs = TCGOP_CALLO(op); 5124 const int nb_iargs = TCGOP_CALLI(op); 5125 const TCGLifeData arg_life = op->life; 5126 const TCGHelperInfo *info = tcg_call_info(op); 5127 TCGRegSet allocated_regs = s->reserved_regs; 5128 int i; 5129 5130 /* 5131 * Move inputs into place in reverse order, 5132 * so that we place stacked arguments first. 5133 */ 5134 for (i = nb_iargs - 1; i >= 0; --i) { 5135 const TCGCallArgumentLoc *loc = &info->in[i]; 5136 TCGTemp *ts = arg_temp(op->args[nb_oargs + i]); 5137 5138 switch (loc->kind) { 5139 case TCG_CALL_ARG_NORMAL: 5140 case TCG_CALL_ARG_EXTEND_U: 5141 case TCG_CALL_ARG_EXTEND_S: 5142 load_arg_normal(s, loc, ts, &allocated_regs); 5143 break; 5144 case TCG_CALL_ARG_BY_REF: 5145 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5146 load_arg_ref(s, loc->arg_slot, TCG_REG_CALL_STACK, 5147 arg_slot_stk_ofs(loc->ref_slot), 5148 &allocated_regs); 5149 break; 5150 case TCG_CALL_ARG_BY_REF_N: 5151 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5152 break; 5153 default: 5154 g_assert_not_reached(); 5155 } 5156 } 5157 5158 /* Mark dead temporaries and free the associated registers. */ 5159 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 5160 if (IS_DEAD_ARG(i)) { 5161 temp_dead(s, arg_temp(op->args[i])); 5162 } 5163 } 5164 5165 /* Clobber call registers. */ 5166 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 5167 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 5168 tcg_reg_free(s, i, allocated_regs); 5169 } 5170 } 5171 5172 /* 5173 * Save globals if they might be written by the helper, 5174 * sync them if they might be read. 5175 */ 5176 if (info->flags & TCG_CALL_NO_READ_GLOBALS) { 5177 /* Nothing to do */ 5178 } else if (info->flags & TCG_CALL_NO_WRITE_GLOBALS) { 5179 sync_globals(s, allocated_regs); 5180 } else { 5181 save_globals(s, allocated_regs); 5182 } 5183 5184 /* 5185 * If the ABI passes a pointer to the returned struct as the first 5186 * argument, load that now. Pass a pointer to the output home slot. 5187 */ 5188 if (info->out_kind == TCG_CALL_RET_BY_REF) { 5189 TCGTemp *ts = arg_temp(op->args[0]); 5190 5191 if (!ts->mem_allocated) { 5192 temp_allocate_frame(s, ts); 5193 } 5194 load_arg_ref(s, 0, ts->mem_base->reg, ts->mem_offset, &allocated_regs); 5195 } 5196 5197 tcg_out_call(s, tcg_call_func(op), info); 5198 5199 /* Assign output registers and emit moves if needed. */ 5200 switch (info->out_kind) { 5201 case TCG_CALL_RET_NORMAL: 5202 for (i = 0; i < nb_oargs; i++) { 5203 TCGTemp *ts = arg_temp(op->args[i]); 5204 TCGReg reg = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, i); 5205 5206 /* ENV should not be modified. */ 5207 tcg_debug_assert(!temp_readonly(ts)); 5208 5209 set_temp_val_reg(s, ts, reg); 5210 ts->mem_coherent = 0; 5211 } 5212 break; 5213 5214 case TCG_CALL_RET_BY_VEC: 5215 { 5216 TCGTemp *ts = arg_temp(op->args[0]); 5217 5218 tcg_debug_assert(ts->base_type == TCG_TYPE_I128); 5219 tcg_debug_assert(ts->temp_subindex == 0); 5220 if (!ts->mem_allocated) { 5221 temp_allocate_frame(s, ts); 5222 } 5223 tcg_out_st(s, TCG_TYPE_V128, 5224 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5225 ts->mem_base->reg, ts->mem_offset); 5226 } 5227 /* fall through to mark all parts in memory */ 5228 5229 case TCG_CALL_RET_BY_REF: 5230 /* The callee has performed a write through the reference. */ 5231 for (i = 0; i < nb_oargs; i++) { 5232 TCGTemp *ts = arg_temp(op->args[i]); 5233 ts->val_type = TEMP_VAL_MEM; 5234 } 5235 break; 5236 5237 default: 5238 g_assert_not_reached(); 5239 } 5240 5241 /* Flush or discard output registers as needed. */ 5242 for (i = 0; i < nb_oargs; i++) { 5243 TCGTemp *ts = arg_temp(op->args[i]); 5244 if (NEED_SYNC_ARG(i)) { 5245 temp_sync(s, ts, s->reserved_regs, 0, IS_DEAD_ARG(i)); 5246 } else if (IS_DEAD_ARG(i)) { 5247 temp_dead(s, ts); 5248 } 5249 } 5250 } 5251 5252 /** 5253 * atom_and_align_for_opc: 5254 * @s: tcg context 5255 * @opc: memory operation code 5256 * @host_atom: MO_ATOM_{IFALIGN,WITHIN16,SUBALIGN} for host operations 5257 * @allow_two_ops: true if we are prepared to issue two operations 5258 * 5259 * Return the alignment and atomicity to use for the inline fast path 5260 * for the given memory operation. The alignment may be larger than 5261 * that specified in @opc, and the correct alignment will be diagnosed 5262 * by the slow path helper. 5263 * 5264 * If @allow_two_ops, the host is prepared to test for 2x alignment, 5265 * and issue two loads or stores for subalignment. 5266 */ 5267 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 5268 MemOp host_atom, bool allow_two_ops) 5269 { 5270 MemOp align = get_alignment_bits(opc); 5271 MemOp size = opc & MO_SIZE; 5272 MemOp half = size ? size - 1 : 0; 5273 MemOp atmax; 5274 MemOp atom; 5275 5276 /* When serialized, no further atomicity required. */ 5277 if (s->gen_tb->cflags & CF_PARALLEL) { 5278 atom = opc & MO_ATOM_MASK; 5279 } else { 5280 atom = MO_ATOM_NONE; 5281 } 5282 5283 switch (atom) { 5284 case MO_ATOM_NONE: 5285 /* The operation requires no specific atomicity. */ 5286 atmax = MO_8; 5287 break; 5288 5289 case MO_ATOM_IFALIGN: 5290 atmax = size; 5291 break; 5292 5293 case MO_ATOM_IFALIGN_PAIR: 5294 atmax = half; 5295 break; 5296 5297 case MO_ATOM_WITHIN16: 5298 atmax = size; 5299 if (size == MO_128) { 5300 /* Misalignment implies !within16, and therefore no atomicity. */ 5301 } else if (host_atom != MO_ATOM_WITHIN16) { 5302 /* The host does not implement within16, so require alignment. */ 5303 align = MAX(align, size); 5304 } 5305 break; 5306 5307 case MO_ATOM_WITHIN16_PAIR: 5308 atmax = size; 5309 /* 5310 * Misalignment implies !within16, and therefore half atomicity. 5311 * Any host prepared for two operations can implement this with 5312 * half alignment. 5313 */ 5314 if (host_atom != MO_ATOM_WITHIN16 && allow_two_ops) { 5315 align = MAX(align, half); 5316 } 5317 break; 5318 5319 case MO_ATOM_SUBALIGN: 5320 atmax = size; 5321 if (host_atom != MO_ATOM_SUBALIGN) { 5322 /* If unaligned but not odd, there are subobjects up to half. */ 5323 if (allow_two_ops) { 5324 align = MAX(align, half); 5325 } else { 5326 align = MAX(align, size); 5327 } 5328 } 5329 break; 5330 5331 default: 5332 g_assert_not_reached(); 5333 } 5334 5335 return (TCGAtomAlign){ .atom = atmax, .align = align }; 5336 } 5337 5338 /* 5339 * Similarly for qemu_ld/st slow path helpers. 5340 * We must re-implement tcg_gen_callN and tcg_reg_alloc_call simultaneously, 5341 * using only the provided backend tcg_out_* functions. 5342 */ 5343 5344 static int tcg_out_helper_stk_ofs(TCGType type, unsigned slot) 5345 { 5346 int ofs = arg_slot_stk_ofs(slot); 5347 5348 /* 5349 * Each stack slot is TCG_TARGET_LONG_BITS. If the host does not 5350 * require extension to uint64_t, adjust the address for uint32_t. 5351 */ 5352 if (HOST_BIG_ENDIAN && 5353 TCG_TARGET_REG_BITS == 64 && 5354 type == TCG_TYPE_I32) { 5355 ofs += 4; 5356 } 5357 return ofs; 5358 } 5359 5360 static void tcg_out_helper_load_slots(TCGContext *s, 5361 unsigned nmov, TCGMovExtend *mov, 5362 const TCGLdstHelperParam *parm) 5363 { 5364 unsigned i; 5365 TCGReg dst3; 5366 5367 /* 5368 * Start from the end, storing to the stack first. 5369 * This frees those registers, so we need not consider overlap. 5370 */ 5371 for (i = nmov; i-- > 0; ) { 5372 unsigned slot = mov[i].dst; 5373 5374 if (arg_slot_reg_p(slot)) { 5375 goto found_reg; 5376 } 5377 5378 TCGReg src = mov[i].src; 5379 TCGType dst_type = mov[i].dst_type; 5380 MemOp dst_mo = dst_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5381 5382 /* The argument is going onto the stack; extend into scratch. */ 5383 if ((mov[i].src_ext & MO_SIZE) != dst_mo) { 5384 tcg_debug_assert(parm->ntmp != 0); 5385 mov[i].dst = src = parm->tmp[0]; 5386 tcg_out_movext1(s, &mov[i]); 5387 } 5388 5389 tcg_out_st(s, dst_type, src, TCG_REG_CALL_STACK, 5390 tcg_out_helper_stk_ofs(dst_type, slot)); 5391 } 5392 return; 5393 5394 found_reg: 5395 /* 5396 * The remaining arguments are in registers. 5397 * Convert slot numbers to argument registers. 5398 */ 5399 nmov = i + 1; 5400 for (i = 0; i < nmov; ++i) { 5401 mov[i].dst = tcg_target_call_iarg_regs[mov[i].dst]; 5402 } 5403 5404 switch (nmov) { 5405 case 4: 5406 /* The backend must have provided enough temps for the worst case. */ 5407 tcg_debug_assert(parm->ntmp >= 2); 5408 5409 dst3 = mov[3].dst; 5410 for (unsigned j = 0; j < 3; ++j) { 5411 if (dst3 == mov[j].src) { 5412 /* 5413 * Conflict. Copy the source to a temporary, perform the 5414 * remaining moves, then the extension from our scratch 5415 * on the way out. 5416 */ 5417 TCGReg scratch = parm->tmp[1]; 5418 5419 tcg_out_mov(s, mov[3].src_type, scratch, mov[3].src); 5420 tcg_out_movext3(s, mov, mov + 1, mov + 2, parm->tmp[0]); 5421 tcg_out_movext1_new_src(s, &mov[3], scratch); 5422 break; 5423 } 5424 } 5425 5426 /* No conflicts: perform this move and continue. */ 5427 tcg_out_movext1(s, &mov[3]); 5428 /* fall through */ 5429 5430 case 3: 5431 tcg_out_movext3(s, mov, mov + 1, mov + 2, 5432 parm->ntmp ? parm->tmp[0] : -1); 5433 break; 5434 case 2: 5435 tcg_out_movext2(s, mov, mov + 1, 5436 parm->ntmp ? parm->tmp[0] : -1); 5437 break; 5438 case 1: 5439 tcg_out_movext1(s, mov); 5440 break; 5441 default: 5442 g_assert_not_reached(); 5443 } 5444 } 5445 5446 static void tcg_out_helper_load_imm(TCGContext *s, unsigned slot, 5447 TCGType type, tcg_target_long imm, 5448 const TCGLdstHelperParam *parm) 5449 { 5450 if (arg_slot_reg_p(slot)) { 5451 tcg_out_movi(s, type, tcg_target_call_iarg_regs[slot], imm); 5452 } else { 5453 int ofs = tcg_out_helper_stk_ofs(type, slot); 5454 if (!tcg_out_sti(s, type, imm, TCG_REG_CALL_STACK, ofs)) { 5455 tcg_debug_assert(parm->ntmp != 0); 5456 tcg_out_movi(s, type, parm->tmp[0], imm); 5457 tcg_out_st(s, type, parm->tmp[0], TCG_REG_CALL_STACK, ofs); 5458 } 5459 } 5460 } 5461 5462 static void tcg_out_helper_load_common_args(TCGContext *s, 5463 const TCGLabelQemuLdst *ldst, 5464 const TCGLdstHelperParam *parm, 5465 const TCGHelperInfo *info, 5466 unsigned next_arg) 5467 { 5468 TCGMovExtend ptr_mov = { 5469 .dst_type = TCG_TYPE_PTR, 5470 .src_type = TCG_TYPE_PTR, 5471 .src_ext = sizeof(void *) == 4 ? MO_32 : MO_64 5472 }; 5473 const TCGCallArgumentLoc *loc = &info->in[0]; 5474 TCGType type; 5475 unsigned slot; 5476 tcg_target_ulong imm; 5477 5478 /* 5479 * Handle env, which is always first. 5480 */ 5481 ptr_mov.dst = loc->arg_slot; 5482 ptr_mov.src = TCG_AREG0; 5483 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5484 5485 /* 5486 * Handle oi. 5487 */ 5488 imm = ldst->oi; 5489 loc = &info->in[next_arg]; 5490 type = TCG_TYPE_I32; 5491 switch (loc->kind) { 5492 case TCG_CALL_ARG_NORMAL: 5493 break; 5494 case TCG_CALL_ARG_EXTEND_U: 5495 case TCG_CALL_ARG_EXTEND_S: 5496 /* No extension required for MemOpIdx. */ 5497 tcg_debug_assert(imm <= INT32_MAX); 5498 type = TCG_TYPE_REG; 5499 break; 5500 default: 5501 g_assert_not_reached(); 5502 } 5503 tcg_out_helper_load_imm(s, loc->arg_slot, type, imm, parm); 5504 next_arg++; 5505 5506 /* 5507 * Handle ra. 5508 */ 5509 loc = &info->in[next_arg]; 5510 slot = loc->arg_slot; 5511 if (parm->ra_gen) { 5512 int arg_reg = -1; 5513 TCGReg ra_reg; 5514 5515 if (arg_slot_reg_p(slot)) { 5516 arg_reg = tcg_target_call_iarg_regs[slot]; 5517 } 5518 ra_reg = parm->ra_gen(s, ldst, arg_reg); 5519 5520 ptr_mov.dst = slot; 5521 ptr_mov.src = ra_reg; 5522 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5523 } else { 5524 imm = (uintptr_t)ldst->raddr; 5525 tcg_out_helper_load_imm(s, slot, TCG_TYPE_PTR, imm, parm); 5526 } 5527 } 5528 5529 static unsigned tcg_out_helper_add_mov(TCGMovExtend *mov, 5530 const TCGCallArgumentLoc *loc, 5531 TCGType dst_type, TCGType src_type, 5532 TCGReg lo, TCGReg hi) 5533 { 5534 MemOp reg_mo; 5535 5536 if (dst_type <= TCG_TYPE_REG) { 5537 MemOp src_ext; 5538 5539 switch (loc->kind) { 5540 case TCG_CALL_ARG_NORMAL: 5541 src_ext = src_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5542 break; 5543 case TCG_CALL_ARG_EXTEND_U: 5544 dst_type = TCG_TYPE_REG; 5545 src_ext = MO_UL; 5546 break; 5547 case TCG_CALL_ARG_EXTEND_S: 5548 dst_type = TCG_TYPE_REG; 5549 src_ext = MO_SL; 5550 break; 5551 default: 5552 g_assert_not_reached(); 5553 } 5554 5555 mov[0].dst = loc->arg_slot; 5556 mov[0].dst_type = dst_type; 5557 mov[0].src = lo; 5558 mov[0].src_type = src_type; 5559 mov[0].src_ext = src_ext; 5560 return 1; 5561 } 5562 5563 if (TCG_TARGET_REG_BITS == 32) { 5564 assert(dst_type == TCG_TYPE_I64); 5565 reg_mo = MO_32; 5566 } else { 5567 assert(dst_type == TCG_TYPE_I128); 5568 reg_mo = MO_64; 5569 } 5570 5571 mov[0].dst = loc[HOST_BIG_ENDIAN].arg_slot; 5572 mov[0].src = lo; 5573 mov[0].dst_type = TCG_TYPE_REG; 5574 mov[0].src_type = TCG_TYPE_REG; 5575 mov[0].src_ext = reg_mo; 5576 5577 mov[1].dst = loc[!HOST_BIG_ENDIAN].arg_slot; 5578 mov[1].src = hi; 5579 mov[1].dst_type = TCG_TYPE_REG; 5580 mov[1].src_type = TCG_TYPE_REG; 5581 mov[1].src_ext = reg_mo; 5582 5583 return 2; 5584 } 5585 5586 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 5587 const TCGLdstHelperParam *parm) 5588 { 5589 const TCGHelperInfo *info; 5590 const TCGCallArgumentLoc *loc; 5591 TCGMovExtend mov[2]; 5592 unsigned next_arg, nmov; 5593 MemOp mop = get_memop(ldst->oi); 5594 5595 switch (mop & MO_SIZE) { 5596 case MO_8: 5597 case MO_16: 5598 case MO_32: 5599 info = &info_helper_ld32_mmu; 5600 break; 5601 case MO_64: 5602 info = &info_helper_ld64_mmu; 5603 break; 5604 case MO_128: 5605 info = &info_helper_ld128_mmu; 5606 break; 5607 default: 5608 g_assert_not_reached(); 5609 } 5610 5611 /* Defer env argument. */ 5612 next_arg = 1; 5613 5614 loc = &info->in[next_arg]; 5615 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5616 /* 5617 * 32-bit host with 32-bit guest: zero-extend the guest address 5618 * to 64-bits for the helper by storing the low part, then 5619 * load a zero for the high part. 5620 */ 5621 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 5622 TCG_TYPE_I32, TCG_TYPE_I32, 5623 ldst->addrlo_reg, -1); 5624 tcg_out_helper_load_slots(s, 1, mov, parm); 5625 5626 tcg_out_helper_load_imm(s, loc[!HOST_BIG_ENDIAN].arg_slot, 5627 TCG_TYPE_I32, 0, parm); 5628 next_arg += 2; 5629 } else { 5630 nmov = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 5631 ldst->addrlo_reg, ldst->addrhi_reg); 5632 tcg_out_helper_load_slots(s, nmov, mov, parm); 5633 next_arg += nmov; 5634 } 5635 5636 switch (info->out_kind) { 5637 case TCG_CALL_RET_NORMAL: 5638 case TCG_CALL_RET_BY_VEC: 5639 break; 5640 case TCG_CALL_RET_BY_REF: 5641 /* 5642 * The return reference is in the first argument slot. 5643 * We need memory in which to return: re-use the top of stack. 5644 */ 5645 { 5646 int ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 5647 5648 if (arg_slot_reg_p(0)) { 5649 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[0], 5650 TCG_REG_CALL_STACK, ofs_slot0); 5651 } else { 5652 tcg_debug_assert(parm->ntmp != 0); 5653 tcg_out_addi_ptr(s, parm->tmp[0], 5654 TCG_REG_CALL_STACK, ofs_slot0); 5655 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 5656 TCG_REG_CALL_STACK, ofs_slot0); 5657 } 5658 } 5659 break; 5660 default: 5661 g_assert_not_reached(); 5662 } 5663 5664 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 5665 } 5666 5667 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *ldst, 5668 bool load_sign, 5669 const TCGLdstHelperParam *parm) 5670 { 5671 MemOp mop = get_memop(ldst->oi); 5672 TCGMovExtend mov[2]; 5673 int ofs_slot0; 5674 5675 switch (ldst->type) { 5676 case TCG_TYPE_I64: 5677 if (TCG_TARGET_REG_BITS == 32) { 5678 break; 5679 } 5680 /* fall through */ 5681 5682 case TCG_TYPE_I32: 5683 mov[0].dst = ldst->datalo_reg; 5684 mov[0].src = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, 0); 5685 mov[0].dst_type = ldst->type; 5686 mov[0].src_type = TCG_TYPE_REG; 5687 5688 /* 5689 * If load_sign, then we allowed the helper to perform the 5690 * appropriate sign extension to tcg_target_ulong, and all 5691 * we need now is a plain move. 5692 * 5693 * If they do not, then we expect the relevant extension 5694 * instruction to be no more expensive than a move, and 5695 * we thus save the icache etc by only using one of two 5696 * helper functions. 5697 */ 5698 if (load_sign || !(mop & MO_SIGN)) { 5699 if (TCG_TARGET_REG_BITS == 32 || ldst->type == TCG_TYPE_I32) { 5700 mov[0].src_ext = MO_32; 5701 } else { 5702 mov[0].src_ext = MO_64; 5703 } 5704 } else { 5705 mov[0].src_ext = mop & MO_SSIZE; 5706 } 5707 tcg_out_movext1(s, mov); 5708 return; 5709 5710 case TCG_TYPE_I128: 5711 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 5712 ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 5713 switch (TCG_TARGET_CALL_RET_I128) { 5714 case TCG_CALL_RET_NORMAL: 5715 break; 5716 case TCG_CALL_RET_BY_VEC: 5717 tcg_out_st(s, TCG_TYPE_V128, 5718 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5719 TCG_REG_CALL_STACK, ofs_slot0); 5720 /* fall through */ 5721 case TCG_CALL_RET_BY_REF: 5722 tcg_out_ld(s, TCG_TYPE_I64, ldst->datalo_reg, 5723 TCG_REG_CALL_STACK, ofs_slot0 + 8 * HOST_BIG_ENDIAN); 5724 tcg_out_ld(s, TCG_TYPE_I64, ldst->datahi_reg, 5725 TCG_REG_CALL_STACK, ofs_slot0 + 8 * !HOST_BIG_ENDIAN); 5726 return; 5727 default: 5728 g_assert_not_reached(); 5729 } 5730 break; 5731 5732 default: 5733 g_assert_not_reached(); 5734 } 5735 5736 mov[0].dst = ldst->datalo_reg; 5737 mov[0].src = 5738 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, HOST_BIG_ENDIAN); 5739 mov[0].dst_type = TCG_TYPE_I32; 5740 mov[0].src_type = TCG_TYPE_I32; 5741 mov[0].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 5742 5743 mov[1].dst = ldst->datahi_reg; 5744 mov[1].src = 5745 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, !HOST_BIG_ENDIAN); 5746 mov[1].dst_type = TCG_TYPE_REG; 5747 mov[1].src_type = TCG_TYPE_REG; 5748 mov[1].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 5749 5750 tcg_out_movext2(s, mov, mov + 1, parm->ntmp ? parm->tmp[0] : -1); 5751 } 5752 5753 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 5754 const TCGLdstHelperParam *parm) 5755 { 5756 const TCGHelperInfo *info; 5757 const TCGCallArgumentLoc *loc; 5758 TCGMovExtend mov[4]; 5759 TCGType data_type; 5760 unsigned next_arg, nmov, n; 5761 MemOp mop = get_memop(ldst->oi); 5762 5763 switch (mop & MO_SIZE) { 5764 case MO_8: 5765 case MO_16: 5766 case MO_32: 5767 info = &info_helper_st32_mmu; 5768 data_type = TCG_TYPE_I32; 5769 break; 5770 case MO_64: 5771 info = &info_helper_st64_mmu; 5772 data_type = TCG_TYPE_I64; 5773 break; 5774 case MO_128: 5775 info = &info_helper_st128_mmu; 5776 data_type = TCG_TYPE_I128; 5777 break; 5778 default: 5779 g_assert_not_reached(); 5780 } 5781 5782 /* Defer env argument. */ 5783 next_arg = 1; 5784 nmov = 0; 5785 5786 /* Handle addr argument. */ 5787 loc = &info->in[next_arg]; 5788 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5789 /* 5790 * 32-bit host with 32-bit guest: zero-extend the guest address 5791 * to 64-bits for the helper by storing the low part. Later, 5792 * after we have processed the register inputs, we will load a 5793 * zero for the high part. 5794 */ 5795 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 5796 TCG_TYPE_I32, TCG_TYPE_I32, 5797 ldst->addrlo_reg, -1); 5798 next_arg += 2; 5799 nmov += 1; 5800 } else { 5801 n = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 5802 ldst->addrlo_reg, ldst->addrhi_reg); 5803 next_arg += n; 5804 nmov += n; 5805 } 5806 5807 /* Handle data argument. */ 5808 loc = &info->in[next_arg]; 5809 switch (loc->kind) { 5810 case TCG_CALL_ARG_NORMAL: 5811 case TCG_CALL_ARG_EXTEND_U: 5812 case TCG_CALL_ARG_EXTEND_S: 5813 n = tcg_out_helper_add_mov(mov + nmov, loc, data_type, ldst->type, 5814 ldst->datalo_reg, ldst->datahi_reg); 5815 next_arg += n; 5816 nmov += n; 5817 tcg_out_helper_load_slots(s, nmov, mov, parm); 5818 break; 5819 5820 case TCG_CALL_ARG_BY_REF: 5821 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 5822 tcg_debug_assert(data_type == TCG_TYPE_I128); 5823 tcg_out_st(s, TCG_TYPE_I64, 5824 HOST_BIG_ENDIAN ? ldst->datahi_reg : ldst->datalo_reg, 5825 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[0].ref_slot)); 5826 tcg_out_st(s, TCG_TYPE_I64, 5827 HOST_BIG_ENDIAN ? ldst->datalo_reg : ldst->datahi_reg, 5828 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[1].ref_slot)); 5829 5830 tcg_out_helper_load_slots(s, nmov, mov, parm); 5831 5832 if (arg_slot_reg_p(loc->arg_slot)) { 5833 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[loc->arg_slot], 5834 TCG_REG_CALL_STACK, 5835 arg_slot_stk_ofs(loc->ref_slot)); 5836 } else { 5837 tcg_debug_assert(parm->ntmp != 0); 5838 tcg_out_addi_ptr(s, parm->tmp[0], TCG_REG_CALL_STACK, 5839 arg_slot_stk_ofs(loc->ref_slot)); 5840 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 5841 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc->arg_slot)); 5842 } 5843 next_arg += 2; 5844 break; 5845 5846 default: 5847 g_assert_not_reached(); 5848 } 5849 5850 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5851 /* Zero extend the address by loading a zero for the high part. */ 5852 loc = &info->in[1 + !HOST_BIG_ENDIAN]; 5853 tcg_out_helper_load_imm(s, loc->arg_slot, TCG_TYPE_I32, 0, parm); 5854 } 5855 5856 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 5857 } 5858 5859 #ifdef CONFIG_PROFILER 5860 5861 /* avoid copy/paste errors */ 5862 #define PROF_ADD(to, from, field) \ 5863 do { \ 5864 (to)->field += qatomic_read(&((from)->field)); \ 5865 } while (0) 5866 5867 #define PROF_MAX(to, from, field) \ 5868 do { \ 5869 typeof((from)->field) val__ = qatomic_read(&((from)->field)); \ 5870 if (val__ > (to)->field) { \ 5871 (to)->field = val__; \ 5872 } \ 5873 } while (0) 5874 5875 /* Pass in a zero'ed @prof */ 5876 static inline 5877 void tcg_profile_snapshot(TCGProfile *prof, bool counters, bool table) 5878 { 5879 unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); 5880 unsigned int i; 5881 5882 for (i = 0; i < n_ctxs; i++) { 5883 TCGContext *s = qatomic_read(&tcg_ctxs[i]); 5884 const TCGProfile *orig = &s->prof; 5885 5886 if (counters) { 5887 PROF_ADD(prof, orig, cpu_exec_time); 5888 PROF_ADD(prof, orig, tb_count1); 5889 PROF_ADD(prof, orig, tb_count); 5890 PROF_ADD(prof, orig, op_count); 5891 PROF_MAX(prof, orig, op_count_max); 5892 PROF_ADD(prof, orig, temp_count); 5893 PROF_MAX(prof, orig, temp_count_max); 5894 PROF_ADD(prof, orig, del_op_count); 5895 PROF_ADD(prof, orig, code_in_len); 5896 PROF_ADD(prof, orig, code_out_len); 5897 PROF_ADD(prof, orig, search_out_len); 5898 PROF_ADD(prof, orig, interm_time); 5899 PROF_ADD(prof, orig, code_time); 5900 PROF_ADD(prof, orig, la_time); 5901 PROF_ADD(prof, orig, opt_time); 5902 PROF_ADD(prof, orig, restore_count); 5903 PROF_ADD(prof, orig, restore_time); 5904 } 5905 if (table) { 5906 int i; 5907 5908 for (i = 0; i < NB_OPS; i++) { 5909 PROF_ADD(prof, orig, table_op_count[i]); 5910 } 5911 } 5912 } 5913 } 5914 5915 #undef PROF_ADD 5916 #undef PROF_MAX 5917 5918 static void tcg_profile_snapshot_counters(TCGProfile *prof) 5919 { 5920 tcg_profile_snapshot(prof, true, false); 5921 } 5922 5923 static void tcg_profile_snapshot_table(TCGProfile *prof) 5924 { 5925 tcg_profile_snapshot(prof, false, true); 5926 } 5927 5928 void tcg_dump_op_count(GString *buf) 5929 { 5930 TCGProfile prof = {}; 5931 int i; 5932 5933 tcg_profile_snapshot_table(&prof); 5934 for (i = 0; i < NB_OPS; i++) { 5935 g_string_append_printf(buf, "%s %" PRId64 "\n", tcg_op_defs[i].name, 5936 prof.table_op_count[i]); 5937 } 5938 } 5939 5940 int64_t tcg_cpu_exec_time(void) 5941 { 5942 unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); 5943 unsigned int i; 5944 int64_t ret = 0; 5945 5946 for (i = 0; i < n_ctxs; i++) { 5947 const TCGContext *s = qatomic_read(&tcg_ctxs[i]); 5948 const TCGProfile *prof = &s->prof; 5949 5950 ret += qatomic_read(&prof->cpu_exec_time); 5951 } 5952 return ret; 5953 } 5954 #else 5955 void tcg_dump_op_count(GString *buf) 5956 { 5957 g_string_append_printf(buf, "[TCG profiler not compiled]\n"); 5958 } 5959 5960 int64_t tcg_cpu_exec_time(void) 5961 { 5962 error_report("%s: TCG profiler not compiled", __func__); 5963 exit(EXIT_FAILURE); 5964 } 5965 #endif 5966 5967 5968 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start) 5969 { 5970 #ifdef CONFIG_PROFILER 5971 TCGProfile *prof = &s->prof; 5972 #endif 5973 int i, num_insns; 5974 TCGOp *op; 5975 5976 #ifdef CONFIG_PROFILER 5977 { 5978 int n = 0; 5979 5980 QTAILQ_FOREACH(op, &s->ops, link) { 5981 n++; 5982 } 5983 qatomic_set(&prof->op_count, prof->op_count + n); 5984 if (n > prof->op_count_max) { 5985 qatomic_set(&prof->op_count_max, n); 5986 } 5987 5988 n = s->nb_temps; 5989 qatomic_set(&prof->temp_count, prof->temp_count + n); 5990 if (n > prof->temp_count_max) { 5991 qatomic_set(&prof->temp_count_max, n); 5992 } 5993 } 5994 #endif 5995 5996 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP) 5997 && qemu_log_in_addr_range(pc_start))) { 5998 FILE *logfile = qemu_log_trylock(); 5999 if (logfile) { 6000 fprintf(logfile, "OP:\n"); 6001 tcg_dump_ops(s, logfile, false); 6002 fprintf(logfile, "\n"); 6003 qemu_log_unlock(logfile); 6004 } 6005 } 6006 6007 #ifdef CONFIG_DEBUG_TCG 6008 /* Ensure all labels referenced have been emitted. */ 6009 { 6010 TCGLabel *l; 6011 bool error = false; 6012 6013 QSIMPLEQ_FOREACH(l, &s->labels, next) { 6014 if (unlikely(!l->present) && !QSIMPLEQ_EMPTY(&l->branches)) { 6015 qemu_log_mask(CPU_LOG_TB_OP, 6016 "$L%d referenced but not present.\n", l->id); 6017 error = true; 6018 } 6019 } 6020 assert(!error); 6021 } 6022 #endif 6023 6024 #ifdef CONFIG_PROFILER 6025 qatomic_set(&prof->opt_time, prof->opt_time - profile_getclock()); 6026 #endif 6027 6028 tcg_optimize(s); 6029 6030 #ifdef CONFIG_PROFILER 6031 qatomic_set(&prof->opt_time, prof->opt_time + profile_getclock()); 6032 qatomic_set(&prof->la_time, prof->la_time - profile_getclock()); 6033 #endif 6034 6035 reachable_code_pass(s); 6036 liveness_pass_0(s); 6037 liveness_pass_1(s); 6038 6039 if (s->nb_indirects > 0) { 6040 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND) 6041 && qemu_log_in_addr_range(pc_start))) { 6042 FILE *logfile = qemu_log_trylock(); 6043 if (logfile) { 6044 fprintf(logfile, "OP before indirect lowering:\n"); 6045 tcg_dump_ops(s, logfile, false); 6046 fprintf(logfile, "\n"); 6047 qemu_log_unlock(logfile); 6048 } 6049 } 6050 6051 /* Replace indirect temps with direct temps. */ 6052 if (liveness_pass_2(s)) { 6053 /* If changes were made, re-run liveness. */ 6054 liveness_pass_1(s); 6055 } 6056 } 6057 6058 #ifdef CONFIG_PROFILER 6059 qatomic_set(&prof->la_time, prof->la_time + profile_getclock()); 6060 #endif 6061 6062 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT) 6063 && qemu_log_in_addr_range(pc_start))) { 6064 FILE *logfile = qemu_log_trylock(); 6065 if (logfile) { 6066 fprintf(logfile, "OP after optimization and liveness analysis:\n"); 6067 tcg_dump_ops(s, logfile, true); 6068 fprintf(logfile, "\n"); 6069 qemu_log_unlock(logfile); 6070 } 6071 } 6072 6073 /* Initialize goto_tb jump offsets. */ 6074 tb->jmp_reset_offset[0] = TB_JMP_OFFSET_INVALID; 6075 tb->jmp_reset_offset[1] = TB_JMP_OFFSET_INVALID; 6076 tb->jmp_insn_offset[0] = TB_JMP_OFFSET_INVALID; 6077 tb->jmp_insn_offset[1] = TB_JMP_OFFSET_INVALID; 6078 6079 tcg_reg_alloc_start(s); 6080 6081 /* 6082 * Reset the buffer pointers when restarting after overflow. 6083 * TODO: Move this into translate-all.c with the rest of the 6084 * buffer management. Having only this done here is confusing. 6085 */ 6086 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr); 6087 s->code_ptr = s->code_buf; 6088 6089 #ifdef TCG_TARGET_NEED_LDST_LABELS 6090 QSIMPLEQ_INIT(&s->ldst_labels); 6091 #endif 6092 #ifdef TCG_TARGET_NEED_POOL_LABELS 6093 s->pool_labels = NULL; 6094 #endif 6095 6096 num_insns = -1; 6097 QTAILQ_FOREACH(op, &s->ops, link) { 6098 TCGOpcode opc = op->opc; 6099 6100 #ifdef CONFIG_PROFILER 6101 qatomic_set(&prof->table_op_count[opc], prof->table_op_count[opc] + 1); 6102 #endif 6103 6104 switch (opc) { 6105 case INDEX_op_mov_i32: 6106 case INDEX_op_mov_i64: 6107 case INDEX_op_mov_vec: 6108 tcg_reg_alloc_mov(s, op); 6109 break; 6110 case INDEX_op_dup_vec: 6111 tcg_reg_alloc_dup(s, op); 6112 break; 6113 case INDEX_op_insn_start: 6114 if (num_insns >= 0) { 6115 size_t off = tcg_current_code_size(s); 6116 s->gen_insn_end_off[num_insns] = off; 6117 /* Assert that we do not overflow our stored offset. */ 6118 assert(s->gen_insn_end_off[num_insns] == off); 6119 } 6120 num_insns++; 6121 for (i = 0; i < TARGET_INSN_START_WORDS; ++i) { 6122 s->gen_insn_data[num_insns][i] = 6123 tcg_get_insn_start_param(op, i); 6124 } 6125 break; 6126 case INDEX_op_discard: 6127 temp_dead(s, arg_temp(op->args[0])); 6128 break; 6129 case INDEX_op_set_label: 6130 tcg_reg_alloc_bb_end(s, s->reserved_regs); 6131 tcg_out_label(s, arg_label(op->args[0])); 6132 break; 6133 case INDEX_op_call: 6134 tcg_reg_alloc_call(s, op); 6135 break; 6136 case INDEX_op_exit_tb: 6137 tcg_out_exit_tb(s, op->args[0]); 6138 break; 6139 case INDEX_op_goto_tb: 6140 tcg_out_goto_tb(s, op->args[0]); 6141 break; 6142 case INDEX_op_dup2_vec: 6143 if (tcg_reg_alloc_dup2(s, op)) { 6144 break; 6145 } 6146 /* fall through */ 6147 default: 6148 /* Sanity check that we've not introduced any unhandled opcodes. */ 6149 tcg_debug_assert(tcg_op_supported(opc)); 6150 /* Note: in order to speed up the code, it would be much 6151 faster to have specialized register allocator functions for 6152 some common argument patterns */ 6153 tcg_reg_alloc_op(s, op); 6154 break; 6155 } 6156 /* Test for (pending) buffer overflow. The assumption is that any 6157 one operation beginning below the high water mark cannot overrun 6158 the buffer completely. Thus we can test for overflow after 6159 generating code without having to check during generation. */ 6160 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) { 6161 return -1; 6162 } 6163 /* Test for TB overflow, as seen by gen_insn_end_off. */ 6164 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) { 6165 return -2; 6166 } 6167 } 6168 tcg_debug_assert(num_insns >= 0); 6169 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s); 6170 6171 /* Generate TB finalization at the end of block */ 6172 #ifdef TCG_TARGET_NEED_LDST_LABELS 6173 i = tcg_out_ldst_finalize(s); 6174 if (i < 0) { 6175 return i; 6176 } 6177 #endif 6178 #ifdef TCG_TARGET_NEED_POOL_LABELS 6179 i = tcg_out_pool_finalize(s); 6180 if (i < 0) { 6181 return i; 6182 } 6183 #endif 6184 if (!tcg_resolve_relocs(s)) { 6185 return -2; 6186 } 6187 6188 #ifndef CONFIG_TCG_INTERPRETER 6189 /* flush instruction cache */ 6190 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 6191 (uintptr_t)s->code_buf, 6192 tcg_ptr_byte_diff(s->code_ptr, s->code_buf)); 6193 #endif 6194 6195 return tcg_current_code_size(s); 6196 } 6197 6198 #ifdef CONFIG_PROFILER 6199 void tcg_dump_info(GString *buf) 6200 { 6201 TCGProfile prof = {}; 6202 const TCGProfile *s; 6203 int64_t tb_count; 6204 int64_t tb_div_count; 6205 int64_t tot; 6206 6207 tcg_profile_snapshot_counters(&prof); 6208 s = &prof; 6209 tb_count = s->tb_count; 6210 tb_div_count = tb_count ? tb_count : 1; 6211 tot = s->interm_time + s->code_time; 6212 6213 g_string_append_printf(buf, "JIT cycles %" PRId64 6214 " (%0.3f s at 2.4 GHz)\n", 6215 tot, tot / 2.4e9); 6216 g_string_append_printf(buf, "translated TBs %" PRId64 6217 " (aborted=%" PRId64 " %0.1f%%)\n", 6218 tb_count, s->tb_count1 - tb_count, 6219 (double)(s->tb_count1 - s->tb_count) 6220 / (s->tb_count1 ? s->tb_count1 : 1) * 100.0); 6221 g_string_append_printf(buf, "avg ops/TB %0.1f max=%d\n", 6222 (double)s->op_count / tb_div_count, s->op_count_max); 6223 g_string_append_printf(buf, "deleted ops/TB %0.2f\n", 6224 (double)s->del_op_count / tb_div_count); 6225 g_string_append_printf(buf, "avg temps/TB %0.2f max=%d\n", 6226 (double)s->temp_count / tb_div_count, 6227 s->temp_count_max); 6228 g_string_append_printf(buf, "avg host code/TB %0.1f\n", 6229 (double)s->code_out_len / tb_div_count); 6230 g_string_append_printf(buf, "avg search data/TB %0.1f\n", 6231 (double)s->search_out_len / tb_div_count); 6232 6233 g_string_append_printf(buf, "cycles/op %0.1f\n", 6234 s->op_count ? (double)tot / s->op_count : 0); 6235 g_string_append_printf(buf, "cycles/in byte %0.1f\n", 6236 s->code_in_len ? (double)tot / s->code_in_len : 0); 6237 g_string_append_printf(buf, "cycles/out byte %0.1f\n", 6238 s->code_out_len ? (double)tot / s->code_out_len : 0); 6239 g_string_append_printf(buf, "cycles/search byte %0.1f\n", 6240 s->search_out_len ? 6241 (double)tot / s->search_out_len : 0); 6242 if (tot == 0) { 6243 tot = 1; 6244 } 6245 g_string_append_printf(buf, " gen_interm time %0.1f%%\n", 6246 (double)s->interm_time / tot * 100.0); 6247 g_string_append_printf(buf, " gen_code time %0.1f%%\n", 6248 (double)s->code_time / tot * 100.0); 6249 g_string_append_printf(buf, "optim./code time %0.1f%%\n", 6250 (double)s->opt_time / (s->code_time ? 6251 s->code_time : 1) 6252 * 100.0); 6253 g_string_append_printf(buf, "liveness/code time %0.1f%%\n", 6254 (double)s->la_time / (s->code_time ? 6255 s->code_time : 1) * 100.0); 6256 g_string_append_printf(buf, "cpu_restore count %" PRId64 "\n", 6257 s->restore_count); 6258 g_string_append_printf(buf, " avg cycles %0.1f\n", 6259 s->restore_count ? 6260 (double)s->restore_time / s->restore_count : 0); 6261 } 6262 #else 6263 void tcg_dump_info(GString *buf) 6264 { 6265 g_string_append_printf(buf, "[TCG profiler not compiled]\n"); 6266 } 6267 #endif 6268 6269 #ifdef ELF_HOST_MACHINE 6270 /* In order to use this feature, the backend needs to do three things: 6271 6272 (1) Define ELF_HOST_MACHINE to indicate both what value to 6273 put into the ELF image and to indicate support for the feature. 6274 6275 (2) Define tcg_register_jit. This should create a buffer containing 6276 the contents of a .debug_frame section that describes the post- 6277 prologue unwind info for the tcg machine. 6278 6279 (3) Call tcg_register_jit_int, with the constructed .debug_frame. 6280 */ 6281 6282 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */ 6283 typedef enum { 6284 JIT_NOACTION = 0, 6285 JIT_REGISTER_FN, 6286 JIT_UNREGISTER_FN 6287 } jit_actions_t; 6288 6289 struct jit_code_entry { 6290 struct jit_code_entry *next_entry; 6291 struct jit_code_entry *prev_entry; 6292 const void *symfile_addr; 6293 uint64_t symfile_size; 6294 }; 6295 6296 struct jit_descriptor { 6297 uint32_t version; 6298 uint32_t action_flag; 6299 struct jit_code_entry *relevant_entry; 6300 struct jit_code_entry *first_entry; 6301 }; 6302 6303 void __jit_debug_register_code(void) __attribute__((noinline)); 6304 void __jit_debug_register_code(void) 6305 { 6306 asm(""); 6307 } 6308 6309 /* Must statically initialize the version, because GDB may check 6310 the version before we can set it. */ 6311 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 }; 6312 6313 /* End GDB interface. */ 6314 6315 static int find_string(const char *strtab, const char *str) 6316 { 6317 const char *p = strtab + 1; 6318 6319 while (1) { 6320 if (strcmp(p, str) == 0) { 6321 return p - strtab; 6322 } 6323 p += strlen(p) + 1; 6324 } 6325 } 6326 6327 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size, 6328 const void *debug_frame, 6329 size_t debug_frame_size) 6330 { 6331 struct __attribute__((packed)) DebugInfo { 6332 uint32_t len; 6333 uint16_t version; 6334 uint32_t abbrev; 6335 uint8_t ptr_size; 6336 uint8_t cu_die; 6337 uint16_t cu_lang; 6338 uintptr_t cu_low_pc; 6339 uintptr_t cu_high_pc; 6340 uint8_t fn_die; 6341 char fn_name[16]; 6342 uintptr_t fn_low_pc; 6343 uintptr_t fn_high_pc; 6344 uint8_t cu_eoc; 6345 }; 6346 6347 struct ElfImage { 6348 ElfW(Ehdr) ehdr; 6349 ElfW(Phdr) phdr; 6350 ElfW(Shdr) shdr[7]; 6351 ElfW(Sym) sym[2]; 6352 struct DebugInfo di; 6353 uint8_t da[24]; 6354 char str[80]; 6355 }; 6356 6357 struct ElfImage *img; 6358 6359 static const struct ElfImage img_template = { 6360 .ehdr = { 6361 .e_ident[EI_MAG0] = ELFMAG0, 6362 .e_ident[EI_MAG1] = ELFMAG1, 6363 .e_ident[EI_MAG2] = ELFMAG2, 6364 .e_ident[EI_MAG3] = ELFMAG3, 6365 .e_ident[EI_CLASS] = ELF_CLASS, 6366 .e_ident[EI_DATA] = ELF_DATA, 6367 .e_ident[EI_VERSION] = EV_CURRENT, 6368 .e_type = ET_EXEC, 6369 .e_machine = ELF_HOST_MACHINE, 6370 .e_version = EV_CURRENT, 6371 .e_phoff = offsetof(struct ElfImage, phdr), 6372 .e_shoff = offsetof(struct ElfImage, shdr), 6373 .e_ehsize = sizeof(ElfW(Shdr)), 6374 .e_phentsize = sizeof(ElfW(Phdr)), 6375 .e_phnum = 1, 6376 .e_shentsize = sizeof(ElfW(Shdr)), 6377 .e_shnum = ARRAY_SIZE(img->shdr), 6378 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1, 6379 #ifdef ELF_HOST_FLAGS 6380 .e_flags = ELF_HOST_FLAGS, 6381 #endif 6382 #ifdef ELF_OSABI 6383 .e_ident[EI_OSABI] = ELF_OSABI, 6384 #endif 6385 }, 6386 .phdr = { 6387 .p_type = PT_LOAD, 6388 .p_flags = PF_X, 6389 }, 6390 .shdr = { 6391 [0] = { .sh_type = SHT_NULL }, 6392 /* Trick: The contents of code_gen_buffer are not present in 6393 this fake ELF file; that got allocated elsewhere. Therefore 6394 we mark .text as SHT_NOBITS (similar to .bss) so that readers 6395 will not look for contents. We can record any address. */ 6396 [1] = { /* .text */ 6397 .sh_type = SHT_NOBITS, 6398 .sh_flags = SHF_EXECINSTR | SHF_ALLOC, 6399 }, 6400 [2] = { /* .debug_info */ 6401 .sh_type = SHT_PROGBITS, 6402 .sh_offset = offsetof(struct ElfImage, di), 6403 .sh_size = sizeof(struct DebugInfo), 6404 }, 6405 [3] = { /* .debug_abbrev */ 6406 .sh_type = SHT_PROGBITS, 6407 .sh_offset = offsetof(struct ElfImage, da), 6408 .sh_size = sizeof(img->da), 6409 }, 6410 [4] = { /* .debug_frame */ 6411 .sh_type = SHT_PROGBITS, 6412 .sh_offset = sizeof(struct ElfImage), 6413 }, 6414 [5] = { /* .symtab */ 6415 .sh_type = SHT_SYMTAB, 6416 .sh_offset = offsetof(struct ElfImage, sym), 6417 .sh_size = sizeof(img->sym), 6418 .sh_info = 1, 6419 .sh_link = ARRAY_SIZE(img->shdr) - 1, 6420 .sh_entsize = sizeof(ElfW(Sym)), 6421 }, 6422 [6] = { /* .strtab */ 6423 .sh_type = SHT_STRTAB, 6424 .sh_offset = offsetof(struct ElfImage, str), 6425 .sh_size = sizeof(img->str), 6426 } 6427 }, 6428 .sym = { 6429 [1] = { /* code_gen_buffer */ 6430 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC), 6431 .st_shndx = 1, 6432 } 6433 }, 6434 .di = { 6435 .len = sizeof(struct DebugInfo) - 4, 6436 .version = 2, 6437 .ptr_size = sizeof(void *), 6438 .cu_die = 1, 6439 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */ 6440 .fn_die = 2, 6441 .fn_name = "code_gen_buffer" 6442 }, 6443 .da = { 6444 1, /* abbrev number (the cu) */ 6445 0x11, 1, /* DW_TAG_compile_unit, has children */ 6446 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */ 6447 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6448 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6449 0, 0, /* end of abbrev */ 6450 2, /* abbrev number (the fn) */ 6451 0x2e, 0, /* DW_TAG_subprogram, no children */ 6452 0x3, 0x8, /* DW_AT_name, DW_FORM_string */ 6453 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6454 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6455 0, 0, /* end of abbrev */ 6456 0 /* no more abbrev */ 6457 }, 6458 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0" 6459 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer", 6460 }; 6461 6462 /* We only need a single jit entry; statically allocate it. */ 6463 static struct jit_code_entry one_entry; 6464 6465 uintptr_t buf = (uintptr_t)buf_ptr; 6466 size_t img_size = sizeof(struct ElfImage) + debug_frame_size; 6467 DebugFrameHeader *dfh; 6468 6469 img = g_malloc(img_size); 6470 *img = img_template; 6471 6472 img->phdr.p_vaddr = buf; 6473 img->phdr.p_paddr = buf; 6474 img->phdr.p_memsz = buf_size; 6475 6476 img->shdr[1].sh_name = find_string(img->str, ".text"); 6477 img->shdr[1].sh_addr = buf; 6478 img->shdr[1].sh_size = buf_size; 6479 6480 img->shdr[2].sh_name = find_string(img->str, ".debug_info"); 6481 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev"); 6482 6483 img->shdr[4].sh_name = find_string(img->str, ".debug_frame"); 6484 img->shdr[4].sh_size = debug_frame_size; 6485 6486 img->shdr[5].sh_name = find_string(img->str, ".symtab"); 6487 img->shdr[6].sh_name = find_string(img->str, ".strtab"); 6488 6489 img->sym[1].st_name = find_string(img->str, "code_gen_buffer"); 6490 img->sym[1].st_value = buf; 6491 img->sym[1].st_size = buf_size; 6492 6493 img->di.cu_low_pc = buf; 6494 img->di.cu_high_pc = buf + buf_size; 6495 img->di.fn_low_pc = buf; 6496 img->di.fn_high_pc = buf + buf_size; 6497 6498 dfh = (DebugFrameHeader *)(img + 1); 6499 memcpy(dfh, debug_frame, debug_frame_size); 6500 dfh->fde.func_start = buf; 6501 dfh->fde.func_len = buf_size; 6502 6503 #ifdef DEBUG_JIT 6504 /* Enable this block to be able to debug the ELF image file creation. 6505 One can use readelf, objdump, or other inspection utilities. */ 6506 { 6507 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir()); 6508 FILE *f = fopen(jit, "w+b"); 6509 if (f) { 6510 if (fwrite(img, img_size, 1, f) != img_size) { 6511 /* Avoid stupid unused return value warning for fwrite. */ 6512 } 6513 fclose(f); 6514 } 6515 } 6516 #endif 6517 6518 one_entry.symfile_addr = img; 6519 one_entry.symfile_size = img_size; 6520 6521 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN; 6522 __jit_debug_descriptor.relevant_entry = &one_entry; 6523 __jit_debug_descriptor.first_entry = &one_entry; 6524 __jit_debug_register_code(); 6525 } 6526 #else 6527 /* No support for the feature. Provide the entry point expected by exec.c, 6528 and implement the internal function we declared earlier. */ 6529 6530 static void tcg_register_jit_int(const void *buf, size_t size, 6531 const void *debug_frame, 6532 size_t debug_frame_size) 6533 { 6534 } 6535 6536 void tcg_register_jit(const void *buf, size_t buf_size) 6537 { 6538 } 6539 #endif /* ELF_HOST_MACHINE */ 6540 6541 #if !TCG_TARGET_MAYBE_vec 6542 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...) 6543 { 6544 g_assert_not_reached(); 6545 } 6546 #endif 6547