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