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/tcg-op-common.h" 40 41 #if UINTPTR_MAX == UINT32_MAX 42 # define ELF_CLASS ELFCLASS32 43 #else 44 # define ELF_CLASS ELFCLASS64 45 #endif 46 #if HOST_BIG_ENDIAN 47 # define ELF_DATA ELFDATA2MSB 48 #else 49 # define ELF_DATA ELFDATA2LSB 50 #endif 51 52 #include "elf.h" 53 #include "exec/log.h" 54 #include "tcg/tcg-ldst.h" 55 #include "tcg/tcg-temp-internal.h" 56 #include "tcg-internal.h" 57 #include "accel/tcg/perf.h" 58 #ifdef CONFIG_USER_ONLY 59 #include "exec/user/guest-base.h" 60 #endif 61 62 /* Forward declarations for functions declared in tcg-target.c.inc and 63 used here. */ 64 static void tcg_target_init(TCGContext *s); 65 static void tcg_target_qemu_prologue(TCGContext *s); 66 static bool patch_reloc(tcg_insn_unit *code_ptr, int type, 67 intptr_t value, intptr_t addend); 68 69 /* The CIE and FDE header definitions will be common to all hosts. */ 70 typedef struct { 71 uint32_t len __attribute__((aligned((sizeof(void *))))); 72 uint32_t id; 73 uint8_t version; 74 char augmentation[1]; 75 uint8_t code_align; 76 uint8_t data_align; 77 uint8_t return_column; 78 } DebugFrameCIE; 79 80 typedef struct QEMU_PACKED { 81 uint32_t len __attribute__((aligned((sizeof(void *))))); 82 uint32_t cie_offset; 83 uintptr_t func_start; 84 uintptr_t func_len; 85 } DebugFrameFDEHeader; 86 87 typedef struct QEMU_PACKED { 88 DebugFrameCIE cie; 89 DebugFrameFDEHeader fde; 90 } DebugFrameHeader; 91 92 typedef struct TCGLabelQemuLdst { 93 bool is_ld; /* qemu_ld: true, qemu_st: false */ 94 MemOpIdx oi; 95 TCGType type; /* result type of a load */ 96 TCGReg addrlo_reg; /* reg index for low word of guest virtual addr */ 97 TCGReg addrhi_reg; /* reg index for high word of guest virtual addr */ 98 TCGReg datalo_reg; /* reg index for low word to be loaded or stored */ 99 TCGReg datahi_reg; /* reg index for high word to be loaded or stored */ 100 const tcg_insn_unit *raddr; /* addr of the next IR of qemu_ld/st IR */ 101 tcg_insn_unit *label_ptr[2]; /* label pointers to be updated */ 102 QSIMPLEQ_ENTRY(TCGLabelQemuLdst) next; 103 } TCGLabelQemuLdst; 104 105 static void tcg_register_jit_int(const void *buf, size_t size, 106 const void *debug_frame, 107 size_t debug_frame_size) 108 __attribute__((unused)); 109 110 /* Forward declarations for functions declared and used in tcg-target.c.inc. */ 111 static void tcg_out_tb_start(TCGContext *s); 112 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1, 113 intptr_t arg2); 114 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 115 static void tcg_out_movi(TCGContext *s, TCGType type, 116 TCGReg ret, tcg_target_long arg); 117 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 118 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 119 static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg); 120 static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg); 121 static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg); 122 static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg); 123 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 124 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 125 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg); 126 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long); 127 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2); 128 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg); 129 static void tcg_out_goto_tb(TCGContext *s, int which); 130 static void tcg_out_op(TCGContext *s, TCGOpcode opc, 131 const TCGArg args[TCG_MAX_OP_ARGS], 132 const int const_args[TCG_MAX_OP_ARGS]); 133 #if TCG_TARGET_MAYBE_vec 134 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 135 TCGReg dst, TCGReg src); 136 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 137 TCGReg dst, TCGReg base, intptr_t offset); 138 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 139 TCGReg dst, int64_t arg); 140 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 141 unsigned vecl, unsigned vece, 142 const TCGArg args[TCG_MAX_OP_ARGS], 143 const int const_args[TCG_MAX_OP_ARGS]); 144 #else 145 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 146 TCGReg dst, TCGReg src) 147 { 148 g_assert_not_reached(); 149 } 150 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 151 TCGReg dst, TCGReg base, intptr_t offset) 152 { 153 g_assert_not_reached(); 154 } 155 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 156 TCGReg dst, int64_t arg) 157 { 158 g_assert_not_reached(); 159 } 160 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 161 unsigned vecl, unsigned vece, 162 const TCGArg args[TCG_MAX_OP_ARGS], 163 const int const_args[TCG_MAX_OP_ARGS]) 164 { 165 g_assert_not_reached(); 166 } 167 #endif 168 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1, 169 intptr_t arg2); 170 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, 171 TCGReg base, intptr_t ofs); 172 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target, 173 const TCGHelperInfo *info); 174 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot); 175 static bool tcg_target_const_match(int64_t val, TCGType type, int ct, int vece); 176 #ifdef TCG_TARGET_NEED_LDST_LABELS 177 static int tcg_out_ldst_finalize(TCGContext *s); 178 #endif 179 180 typedef struct TCGLdstHelperParam { 181 TCGReg (*ra_gen)(TCGContext *s, const TCGLabelQemuLdst *l, int arg_reg); 182 unsigned ntmp; 183 int tmp[3]; 184 } TCGLdstHelperParam; 185 186 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 187 const TCGLdstHelperParam *p) 188 __attribute__((unused)); 189 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *l, 190 bool load_sign, const TCGLdstHelperParam *p) 191 __attribute__((unused)); 192 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 193 const TCGLdstHelperParam *p) 194 __attribute__((unused)); 195 196 static void * const qemu_ld_helpers[MO_SSIZE + 1] __attribute__((unused)) = { 197 [MO_UB] = helper_ldub_mmu, 198 [MO_SB] = helper_ldsb_mmu, 199 [MO_UW] = helper_lduw_mmu, 200 [MO_SW] = helper_ldsw_mmu, 201 [MO_UL] = helper_ldul_mmu, 202 [MO_UQ] = helper_ldq_mmu, 203 #if TCG_TARGET_REG_BITS == 64 204 [MO_SL] = helper_ldsl_mmu, 205 [MO_128] = helper_ld16_mmu, 206 #endif 207 }; 208 209 static void * const qemu_st_helpers[MO_SIZE + 1] __attribute__((unused)) = { 210 [MO_8] = helper_stb_mmu, 211 [MO_16] = helper_stw_mmu, 212 [MO_32] = helper_stl_mmu, 213 [MO_64] = helper_stq_mmu, 214 #if TCG_TARGET_REG_BITS == 64 215 [MO_128] = helper_st16_mmu, 216 #endif 217 }; 218 219 typedef struct { 220 MemOp atom; /* lg2 bits of atomicity required */ 221 MemOp align; /* lg2 bits of alignment to use */ 222 } TCGAtomAlign; 223 224 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 225 MemOp host_atom, bool allow_two_ops) 226 __attribute__((unused)); 227 228 TCGContext tcg_init_ctx; 229 __thread TCGContext *tcg_ctx; 230 231 TCGContext **tcg_ctxs; 232 unsigned int tcg_cur_ctxs; 233 unsigned int tcg_max_ctxs; 234 TCGv_env cpu_env = 0; 235 const void *tcg_code_gen_epilogue; 236 uintptr_t tcg_splitwx_diff; 237 238 #ifndef CONFIG_TCG_INTERPRETER 239 tcg_prologue_fn *tcg_qemu_tb_exec; 240 #endif 241 242 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT]; 243 static TCGRegSet tcg_target_call_clobber_regs; 244 245 #if TCG_TARGET_INSN_UNIT_SIZE == 1 246 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v) 247 { 248 *s->code_ptr++ = v; 249 } 250 251 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p, 252 uint8_t v) 253 { 254 *p = v; 255 } 256 #endif 257 258 #if TCG_TARGET_INSN_UNIT_SIZE <= 2 259 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v) 260 { 261 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 262 *s->code_ptr++ = v; 263 } else { 264 tcg_insn_unit *p = s->code_ptr; 265 memcpy(p, &v, sizeof(v)); 266 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE); 267 } 268 } 269 270 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p, 271 uint16_t v) 272 { 273 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 274 *p = v; 275 } else { 276 memcpy(p, &v, sizeof(v)); 277 } 278 } 279 #endif 280 281 #if TCG_TARGET_INSN_UNIT_SIZE <= 4 282 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v) 283 { 284 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 285 *s->code_ptr++ = v; 286 } else { 287 tcg_insn_unit *p = s->code_ptr; 288 memcpy(p, &v, sizeof(v)); 289 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE); 290 } 291 } 292 293 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p, 294 uint32_t v) 295 { 296 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 297 *p = v; 298 } else { 299 memcpy(p, &v, sizeof(v)); 300 } 301 } 302 #endif 303 304 #if TCG_TARGET_INSN_UNIT_SIZE <= 8 305 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v) 306 { 307 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 308 *s->code_ptr++ = v; 309 } else { 310 tcg_insn_unit *p = s->code_ptr; 311 memcpy(p, &v, sizeof(v)); 312 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE); 313 } 314 } 315 316 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p, 317 uint64_t v) 318 { 319 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 320 *p = v; 321 } else { 322 memcpy(p, &v, sizeof(v)); 323 } 324 } 325 #endif 326 327 /* label relocation processing */ 328 329 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type, 330 TCGLabel *l, intptr_t addend) 331 { 332 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation)); 333 334 r->type = type; 335 r->ptr = code_ptr; 336 r->addend = addend; 337 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next); 338 } 339 340 static void tcg_out_label(TCGContext *s, TCGLabel *l) 341 { 342 tcg_debug_assert(!l->has_value); 343 l->has_value = 1; 344 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr); 345 } 346 347 TCGLabel *gen_new_label(void) 348 { 349 TCGContext *s = tcg_ctx; 350 TCGLabel *l = tcg_malloc(sizeof(TCGLabel)); 351 352 memset(l, 0, sizeof(TCGLabel)); 353 l->id = s->nb_labels++; 354 QSIMPLEQ_INIT(&l->branches); 355 QSIMPLEQ_INIT(&l->relocs); 356 357 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next); 358 359 return l; 360 } 361 362 static bool tcg_resolve_relocs(TCGContext *s) 363 { 364 TCGLabel *l; 365 366 QSIMPLEQ_FOREACH(l, &s->labels, next) { 367 TCGRelocation *r; 368 uintptr_t value = l->u.value; 369 370 QSIMPLEQ_FOREACH(r, &l->relocs, next) { 371 if (!patch_reloc(r->ptr, r->type, value, r->addend)) { 372 return false; 373 } 374 } 375 } 376 return true; 377 } 378 379 static void set_jmp_reset_offset(TCGContext *s, int which) 380 { 381 /* 382 * We will check for overflow at the end of the opcode loop in 383 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 384 */ 385 s->gen_tb->jmp_reset_offset[which] = tcg_current_code_size(s); 386 } 387 388 static void G_GNUC_UNUSED set_jmp_insn_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_insn_offset[which] = tcg_current_code_size(s); 395 } 396 397 static uintptr_t G_GNUC_UNUSED get_jmp_target_addr(TCGContext *s, int which) 398 { 399 /* 400 * Return the read-execute version of the pointer, for the benefit 401 * of any pc-relative addressing mode. 402 */ 403 return (uintptr_t)tcg_splitwx_to_rx(&s->gen_tb->jmp_target_addr[which]); 404 } 405 406 #if defined(CONFIG_SOFTMMU) && !defined(CONFIG_TCG_INTERPRETER) 407 static int tlb_mask_table_ofs(TCGContext *s, int which) 408 { 409 return s->tlb_fast_offset + which * sizeof(CPUTLBDescFast); 410 } 411 #endif 412 413 /* Signal overflow, starting over with fewer guest insns. */ 414 static G_NORETURN 415 void tcg_raise_tb_overflow(TCGContext *s) 416 { 417 siglongjmp(s->jmp_trans, -2); 418 } 419 420 /* 421 * Used by tcg_out_movext{1,2} to hold the arguments for tcg_out_movext. 422 * By the time we arrive at tcg_out_movext1, @dst is always a TCGReg. 423 * 424 * However, tcg_out_helper_load_slots reuses this field to hold an 425 * argument slot number (which may designate a argument register or an 426 * argument stack slot), converting to TCGReg once all arguments that 427 * are destined for the stack are processed. 428 */ 429 typedef struct TCGMovExtend { 430 unsigned dst; 431 TCGReg src; 432 TCGType dst_type; 433 TCGType src_type; 434 MemOp src_ext; 435 } TCGMovExtend; 436 437 /** 438 * tcg_out_movext -- move and extend 439 * @s: tcg context 440 * @dst_type: integral type for destination 441 * @dst: destination register 442 * @src_type: integral type for source 443 * @src_ext: extension to apply to source 444 * @src: source register 445 * 446 * Move or extend @src into @dst, depending on @src_ext and the types. 447 */ 448 static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst, 449 TCGType src_type, MemOp src_ext, TCGReg src) 450 { 451 switch (src_ext) { 452 case MO_UB: 453 tcg_out_ext8u(s, dst, src); 454 break; 455 case MO_SB: 456 tcg_out_ext8s(s, dst_type, dst, src); 457 break; 458 case MO_UW: 459 tcg_out_ext16u(s, dst, src); 460 break; 461 case MO_SW: 462 tcg_out_ext16s(s, dst_type, dst, src); 463 break; 464 case MO_UL: 465 case MO_SL: 466 if (dst_type == TCG_TYPE_I32) { 467 if (src_type == TCG_TYPE_I32) { 468 tcg_out_mov(s, TCG_TYPE_I32, dst, src); 469 } else { 470 tcg_out_extrl_i64_i32(s, dst, src); 471 } 472 } else if (src_type == TCG_TYPE_I32) { 473 if (src_ext & MO_SIGN) { 474 tcg_out_exts_i32_i64(s, dst, src); 475 } else { 476 tcg_out_extu_i32_i64(s, dst, src); 477 } 478 } else { 479 if (src_ext & MO_SIGN) { 480 tcg_out_ext32s(s, dst, src); 481 } else { 482 tcg_out_ext32u(s, dst, src); 483 } 484 } 485 break; 486 case MO_UQ: 487 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 488 if (dst_type == TCG_TYPE_I32) { 489 tcg_out_extrl_i64_i32(s, dst, src); 490 } else { 491 tcg_out_mov(s, TCG_TYPE_I64, dst, src); 492 } 493 break; 494 default: 495 g_assert_not_reached(); 496 } 497 } 498 499 /* Minor variations on a theme, using a structure. */ 500 static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i, 501 TCGReg src) 502 { 503 tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src); 504 } 505 506 static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i) 507 { 508 tcg_out_movext1_new_src(s, i, i->src); 509 } 510 511 /** 512 * tcg_out_movext2 -- move and extend two pair 513 * @s: tcg context 514 * @i1: first move description 515 * @i2: second move description 516 * @scratch: temporary register, or -1 for none 517 * 518 * As tcg_out_movext, for both @i1 and @i2, caring for overlap 519 * between the sources and destinations. 520 */ 521 522 static void tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1, 523 const TCGMovExtend *i2, int scratch) 524 { 525 TCGReg src1 = i1->src; 526 TCGReg src2 = i2->src; 527 528 if (i1->dst != src2) { 529 tcg_out_movext1(s, i1); 530 tcg_out_movext1(s, i2); 531 return; 532 } 533 if (i2->dst == src1) { 534 TCGType src1_type = i1->src_type; 535 TCGType src2_type = i2->src_type; 536 537 if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) { 538 /* The data is now in the correct registers, now extend. */ 539 src1 = i2->src; 540 src2 = i1->src; 541 } else { 542 tcg_debug_assert(scratch >= 0); 543 tcg_out_mov(s, src1_type, scratch, src1); 544 src1 = scratch; 545 } 546 } 547 tcg_out_movext1_new_src(s, i2, src2); 548 tcg_out_movext1_new_src(s, i1, src1); 549 } 550 551 /** 552 * tcg_out_movext3 -- move and extend three pair 553 * @s: tcg context 554 * @i1: first move description 555 * @i2: second move description 556 * @i3: third move description 557 * @scratch: temporary register, or -1 for none 558 * 559 * As tcg_out_movext, for all of @i1, @i2 and @i3, caring for overlap 560 * between the sources and destinations. 561 */ 562 563 static void tcg_out_movext3(TCGContext *s, const TCGMovExtend *i1, 564 const TCGMovExtend *i2, const TCGMovExtend *i3, 565 int scratch) 566 { 567 TCGReg src1 = i1->src; 568 TCGReg src2 = i2->src; 569 TCGReg src3 = i3->src; 570 571 if (i1->dst != src2 && i1->dst != src3) { 572 tcg_out_movext1(s, i1); 573 tcg_out_movext2(s, i2, i3, scratch); 574 return; 575 } 576 if (i2->dst != src1 && i2->dst != src3) { 577 tcg_out_movext1(s, i2); 578 tcg_out_movext2(s, i1, i3, scratch); 579 return; 580 } 581 if (i3->dst != src1 && i3->dst != src2) { 582 tcg_out_movext1(s, i3); 583 tcg_out_movext2(s, i1, i2, scratch); 584 return; 585 } 586 587 /* 588 * There is a cycle. Since there are only 3 nodes, the cycle is 589 * either "clockwise" or "anti-clockwise", and can be solved with 590 * a single scratch or two xchg. 591 */ 592 if (i1->dst == src2 && i2->dst == src3 && i3->dst == src1) { 593 /* "Clockwise" */ 594 if (tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2)) { 595 tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3); 596 /* The data is now in the correct registers, now extend. */ 597 tcg_out_movext1_new_src(s, i1, i1->dst); 598 tcg_out_movext1_new_src(s, i2, i2->dst); 599 tcg_out_movext1_new_src(s, i3, i3->dst); 600 } else { 601 tcg_debug_assert(scratch >= 0); 602 tcg_out_mov(s, i1->src_type, scratch, src1); 603 tcg_out_movext1(s, i3); 604 tcg_out_movext1(s, i2); 605 tcg_out_movext1_new_src(s, i1, scratch); 606 } 607 } else if (i1->dst == src3 && i2->dst == src1 && i3->dst == src2) { 608 /* "Anti-clockwise" */ 609 if (tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3)) { 610 tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2); 611 /* The data is now in the correct registers, now extend. */ 612 tcg_out_movext1_new_src(s, i1, i1->dst); 613 tcg_out_movext1_new_src(s, i2, i2->dst); 614 tcg_out_movext1_new_src(s, i3, i3->dst); 615 } else { 616 tcg_debug_assert(scratch >= 0); 617 tcg_out_mov(s, i1->src_type, scratch, src1); 618 tcg_out_movext1(s, i2); 619 tcg_out_movext1(s, i3); 620 tcg_out_movext1_new_src(s, i1, scratch); 621 } 622 } else { 623 g_assert_not_reached(); 624 } 625 } 626 627 #define C_PFX1(P, A) P##A 628 #define C_PFX2(P, A, B) P##A##_##B 629 #define C_PFX3(P, A, B, C) P##A##_##B##_##C 630 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D 631 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E 632 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F 633 634 /* Define an enumeration for the various combinations. */ 635 636 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1), 637 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2), 638 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3), 639 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4), 640 641 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1), 642 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2), 643 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3), 644 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4), 645 646 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2), 647 648 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1), 649 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2), 650 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3), 651 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4), 652 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4), 653 654 typedef enum { 655 #include "tcg-target-con-set.h" 656 } TCGConstraintSetIndex; 657 658 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode); 659 660 #undef C_O0_I1 661 #undef C_O0_I2 662 #undef C_O0_I3 663 #undef C_O0_I4 664 #undef C_O1_I1 665 #undef C_O1_I2 666 #undef C_O1_I3 667 #undef C_O1_I4 668 #undef C_N1_I2 669 #undef C_O2_I1 670 #undef C_O2_I2 671 #undef C_O2_I3 672 #undef C_O2_I4 673 #undef C_N1_O1_I4 674 675 /* Put all of the constraint sets into an array, indexed by the enum. */ 676 677 #define C_O0_I1(I1) { .args_ct_str = { #I1 } }, 678 #define C_O0_I2(I1, I2) { .args_ct_str = { #I1, #I2 } }, 679 #define C_O0_I3(I1, I2, I3) { .args_ct_str = { #I1, #I2, #I3 } }, 680 #define C_O0_I4(I1, I2, I3, I4) { .args_ct_str = { #I1, #I2, #I3, #I4 } }, 681 682 #define C_O1_I1(O1, I1) { .args_ct_str = { #O1, #I1 } }, 683 #define C_O1_I2(O1, I1, I2) { .args_ct_str = { #O1, #I1, #I2 } }, 684 #define C_O1_I3(O1, I1, I2, I3) { .args_ct_str = { #O1, #I1, #I2, #I3 } }, 685 #define C_O1_I4(O1, I1, I2, I3, I4) { .args_ct_str = { #O1, #I1, #I2, #I3, #I4 } }, 686 687 #define C_N1_I2(O1, I1, I2) { .args_ct_str = { "&" #O1, #I1, #I2 } }, 688 689 #define C_O2_I1(O1, O2, I1) { .args_ct_str = { #O1, #O2, #I1 } }, 690 #define C_O2_I2(O1, O2, I1, I2) { .args_ct_str = { #O1, #O2, #I1, #I2 } }, 691 #define C_O2_I3(O1, O2, I1, I2, I3) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3 } }, 692 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3, #I4 } }, 693 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { "&" #O1, #O2, #I1, #I2, #I3, #I4 } }, 694 695 static const TCGTargetOpDef constraint_sets[] = { 696 #include "tcg-target-con-set.h" 697 }; 698 699 700 #undef C_O0_I1 701 #undef C_O0_I2 702 #undef C_O0_I3 703 #undef C_O0_I4 704 #undef C_O1_I1 705 #undef C_O1_I2 706 #undef C_O1_I3 707 #undef C_O1_I4 708 #undef C_N1_I2 709 #undef C_O2_I1 710 #undef C_O2_I2 711 #undef C_O2_I3 712 #undef C_O2_I4 713 #undef C_N1_O1_I4 714 715 /* Expand the enumerator to be returned from tcg_target_op_def(). */ 716 717 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1) 718 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2) 719 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3) 720 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4) 721 722 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1) 723 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2) 724 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3) 725 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4) 726 727 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2) 728 729 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1) 730 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2) 731 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3) 732 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4) 733 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4) 734 735 #include "tcg-target.c.inc" 736 737 static void alloc_tcg_plugin_context(TCGContext *s) 738 { 739 #ifdef CONFIG_PLUGIN 740 s->plugin_tb = g_new0(struct qemu_plugin_tb, 1); 741 s->plugin_tb->insns = 742 g_ptr_array_new_with_free_func(qemu_plugin_insn_cleanup_fn); 743 #endif 744 } 745 746 /* 747 * All TCG threads except the parent (i.e. the one that called tcg_context_init 748 * and registered the target's TCG globals) must register with this function 749 * before initiating translation. 750 * 751 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation 752 * of tcg_region_init() for the reasoning behind this. 753 * 754 * In softmmu each caller registers its context in tcg_ctxs[]. Note that in 755 * softmmu tcg_ctxs[] does not track tcg_ctx_init, since the initial context 756 * is not used anymore for translation once this function is called. 757 * 758 * Not tracking tcg_init_ctx in tcg_ctxs[] in softmmu keeps code that iterates 759 * over the array (e.g. tcg_code_size() the same for both softmmu and user-mode. 760 */ 761 #ifdef CONFIG_USER_ONLY 762 void tcg_register_thread(void) 763 { 764 tcg_ctx = &tcg_init_ctx; 765 } 766 #else 767 void tcg_register_thread(void) 768 { 769 TCGContext *s = g_malloc(sizeof(*s)); 770 unsigned int i, n; 771 772 *s = tcg_init_ctx; 773 774 /* Relink mem_base. */ 775 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) { 776 if (tcg_init_ctx.temps[i].mem_base) { 777 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps; 778 tcg_debug_assert(b >= 0 && b < n); 779 s->temps[i].mem_base = &s->temps[b]; 780 } 781 } 782 783 /* Claim an entry in tcg_ctxs */ 784 n = qatomic_fetch_inc(&tcg_cur_ctxs); 785 g_assert(n < tcg_max_ctxs); 786 qatomic_set(&tcg_ctxs[n], s); 787 788 if (n > 0) { 789 alloc_tcg_plugin_context(s); 790 tcg_region_initial_alloc(s); 791 } 792 793 tcg_ctx = s; 794 } 795 #endif /* !CONFIG_USER_ONLY */ 796 797 /* pool based memory allocation */ 798 void *tcg_malloc_internal(TCGContext *s, int size) 799 { 800 TCGPool *p; 801 int pool_size; 802 803 if (size > TCG_POOL_CHUNK_SIZE) { 804 /* big malloc: insert a new pool (XXX: could optimize) */ 805 p = g_malloc(sizeof(TCGPool) + size); 806 p->size = size; 807 p->next = s->pool_first_large; 808 s->pool_first_large = p; 809 return p->data; 810 } else { 811 p = s->pool_current; 812 if (!p) { 813 p = s->pool_first; 814 if (!p) 815 goto new_pool; 816 } else { 817 if (!p->next) { 818 new_pool: 819 pool_size = TCG_POOL_CHUNK_SIZE; 820 p = g_malloc(sizeof(TCGPool) + pool_size); 821 p->size = pool_size; 822 p->next = NULL; 823 if (s->pool_current) { 824 s->pool_current->next = p; 825 } else { 826 s->pool_first = p; 827 } 828 } else { 829 p = p->next; 830 } 831 } 832 } 833 s->pool_current = p; 834 s->pool_cur = p->data + size; 835 s->pool_end = p->data + p->size; 836 return p->data; 837 } 838 839 void tcg_pool_reset(TCGContext *s) 840 { 841 TCGPool *p, *t; 842 for (p = s->pool_first_large; p; p = t) { 843 t = p->next; 844 g_free(p); 845 } 846 s->pool_first_large = NULL; 847 s->pool_cur = s->pool_end = NULL; 848 s->pool_current = NULL; 849 } 850 851 /* 852 * Create TCGHelperInfo structures for "tcg/tcg-ldst.h" functions, 853 * akin to what "exec/helper-tcg.h" does with DEF_HELPER_FLAGS_N. 854 * We only use these for layout in tcg_out_ld_helper_ret and 855 * tcg_out_st_helper_args, and share them between several of 856 * the helpers, with the end result that it's easier to build manually. 857 */ 858 859 #if TCG_TARGET_REG_BITS == 32 860 # define dh_typecode_ttl dh_typecode_i32 861 #else 862 # define dh_typecode_ttl dh_typecode_i64 863 #endif 864 865 static TCGHelperInfo info_helper_ld32_mmu = { 866 .flags = TCG_CALL_NO_WG, 867 .typemask = dh_typemask(ttl, 0) /* return tcg_target_ulong */ 868 | dh_typemask(env, 1) 869 | dh_typemask(i64, 2) /* uint64_t addr */ 870 | dh_typemask(i32, 3) /* unsigned oi */ 871 | dh_typemask(ptr, 4) /* uintptr_t ra */ 872 }; 873 874 static TCGHelperInfo info_helper_ld64_mmu = { 875 .flags = TCG_CALL_NO_WG, 876 .typemask = dh_typemask(i64, 0) /* return uint64_t */ 877 | dh_typemask(env, 1) 878 | dh_typemask(i64, 2) /* uint64_t addr */ 879 | dh_typemask(i32, 3) /* unsigned oi */ 880 | dh_typemask(ptr, 4) /* uintptr_t ra */ 881 }; 882 883 static TCGHelperInfo info_helper_ld128_mmu = { 884 .flags = TCG_CALL_NO_WG, 885 .typemask = dh_typemask(i128, 0) /* return Int128 */ 886 | dh_typemask(env, 1) 887 | dh_typemask(i64, 2) /* uint64_t addr */ 888 | dh_typemask(i32, 3) /* unsigned oi */ 889 | dh_typemask(ptr, 4) /* uintptr_t ra */ 890 }; 891 892 static TCGHelperInfo info_helper_st32_mmu = { 893 .flags = TCG_CALL_NO_WG, 894 .typemask = dh_typemask(void, 0) 895 | dh_typemask(env, 1) 896 | dh_typemask(i64, 2) /* uint64_t addr */ 897 | dh_typemask(i32, 3) /* uint32_t data */ 898 | dh_typemask(i32, 4) /* unsigned oi */ 899 | dh_typemask(ptr, 5) /* uintptr_t ra */ 900 }; 901 902 static TCGHelperInfo info_helper_st64_mmu = { 903 .flags = TCG_CALL_NO_WG, 904 .typemask = dh_typemask(void, 0) 905 | dh_typemask(env, 1) 906 | dh_typemask(i64, 2) /* uint64_t addr */ 907 | dh_typemask(i64, 3) /* uint64_t data */ 908 | dh_typemask(i32, 4) /* unsigned oi */ 909 | dh_typemask(ptr, 5) /* uintptr_t ra */ 910 }; 911 912 static TCGHelperInfo info_helper_st128_mmu = { 913 .flags = TCG_CALL_NO_WG, 914 .typemask = dh_typemask(void, 0) 915 | dh_typemask(env, 1) 916 | dh_typemask(i64, 2) /* uint64_t addr */ 917 | dh_typemask(i128, 3) /* Int128 data */ 918 | dh_typemask(i32, 4) /* unsigned oi */ 919 | dh_typemask(ptr, 5) /* uintptr_t ra */ 920 }; 921 922 #ifdef CONFIG_TCG_INTERPRETER 923 static ffi_type *typecode_to_ffi(int argmask) 924 { 925 /* 926 * libffi does not support __int128_t, so we have forced Int128 927 * to use the structure definition instead of the builtin type. 928 */ 929 static ffi_type *ffi_type_i128_elements[3] = { 930 &ffi_type_uint64, 931 &ffi_type_uint64, 932 NULL 933 }; 934 static ffi_type ffi_type_i128 = { 935 .size = 16, 936 .alignment = __alignof__(Int128), 937 .type = FFI_TYPE_STRUCT, 938 .elements = ffi_type_i128_elements, 939 }; 940 941 switch (argmask) { 942 case dh_typecode_void: 943 return &ffi_type_void; 944 case dh_typecode_i32: 945 return &ffi_type_uint32; 946 case dh_typecode_s32: 947 return &ffi_type_sint32; 948 case dh_typecode_i64: 949 return &ffi_type_uint64; 950 case dh_typecode_s64: 951 return &ffi_type_sint64; 952 case dh_typecode_ptr: 953 return &ffi_type_pointer; 954 case dh_typecode_i128: 955 return &ffi_type_i128; 956 } 957 g_assert_not_reached(); 958 } 959 960 static ffi_cif *init_ffi_layout(TCGHelperInfo *info) 961 { 962 unsigned typemask = info->typemask; 963 struct { 964 ffi_cif cif; 965 ffi_type *args[]; 966 } *ca; 967 ffi_status status; 968 int nargs; 969 970 /* Ignoring the return type, find the last non-zero field. */ 971 nargs = 32 - clz32(typemask >> 3); 972 nargs = DIV_ROUND_UP(nargs, 3); 973 assert(nargs <= MAX_CALL_IARGS); 974 975 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *)); 976 ca->cif.rtype = typecode_to_ffi(typemask & 7); 977 ca->cif.nargs = nargs; 978 979 if (nargs != 0) { 980 ca->cif.arg_types = ca->args; 981 for (int j = 0; j < nargs; ++j) { 982 int typecode = extract32(typemask, (j + 1) * 3, 3); 983 ca->args[j] = typecode_to_ffi(typecode); 984 } 985 } 986 987 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs, 988 ca->cif.rtype, ca->cif.arg_types); 989 assert(status == FFI_OK); 990 991 return &ca->cif; 992 } 993 994 #define HELPER_INFO_INIT(I) (&(I)->cif) 995 #define HELPER_INFO_INIT_VAL(I) init_ffi_layout(I) 996 #else 997 #define HELPER_INFO_INIT(I) (&(I)->init) 998 #define HELPER_INFO_INIT_VAL(I) 1 999 #endif /* CONFIG_TCG_INTERPRETER */ 1000 1001 static inline bool arg_slot_reg_p(unsigned arg_slot) 1002 { 1003 /* 1004 * Split the sizeof away from the comparison to avoid Werror from 1005 * "unsigned < 0 is always false", when iarg_regs is empty. 1006 */ 1007 unsigned nreg = ARRAY_SIZE(tcg_target_call_iarg_regs); 1008 return arg_slot < nreg; 1009 } 1010 1011 static inline int arg_slot_stk_ofs(unsigned arg_slot) 1012 { 1013 unsigned max = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1014 unsigned stk_slot = arg_slot - ARRAY_SIZE(tcg_target_call_iarg_regs); 1015 1016 tcg_debug_assert(stk_slot < max); 1017 return TCG_TARGET_CALL_STACK_OFFSET + stk_slot * sizeof(tcg_target_long); 1018 } 1019 1020 typedef struct TCGCumulativeArgs { 1021 int arg_idx; /* tcg_gen_callN args[] */ 1022 int info_in_idx; /* TCGHelperInfo in[] */ 1023 int arg_slot; /* regs+stack slot */ 1024 int ref_slot; /* stack slots for references */ 1025 } TCGCumulativeArgs; 1026 1027 static void layout_arg_even(TCGCumulativeArgs *cum) 1028 { 1029 cum->arg_slot += cum->arg_slot & 1; 1030 } 1031 1032 static void layout_arg_1(TCGCumulativeArgs *cum, TCGHelperInfo *info, 1033 TCGCallArgumentKind kind) 1034 { 1035 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1036 1037 *loc = (TCGCallArgumentLoc){ 1038 .kind = kind, 1039 .arg_idx = cum->arg_idx, 1040 .arg_slot = cum->arg_slot, 1041 }; 1042 cum->info_in_idx++; 1043 cum->arg_slot++; 1044 } 1045 1046 static void layout_arg_normal_n(TCGCumulativeArgs *cum, 1047 TCGHelperInfo *info, int n) 1048 { 1049 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1050 1051 for (int i = 0; i < n; ++i) { 1052 /* Layout all using the same arg_idx, adjusting the subindex. */ 1053 loc[i] = (TCGCallArgumentLoc){ 1054 .kind = TCG_CALL_ARG_NORMAL, 1055 .arg_idx = cum->arg_idx, 1056 .tmp_subindex = i, 1057 .arg_slot = cum->arg_slot + i, 1058 }; 1059 } 1060 cum->info_in_idx += n; 1061 cum->arg_slot += n; 1062 } 1063 1064 static void layout_arg_by_ref(TCGCumulativeArgs *cum, TCGHelperInfo *info) 1065 { 1066 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1067 int n = 128 / TCG_TARGET_REG_BITS; 1068 1069 /* The first subindex carries the pointer. */ 1070 layout_arg_1(cum, info, TCG_CALL_ARG_BY_REF); 1071 1072 /* 1073 * The callee is allowed to clobber memory associated with 1074 * structure pass by-reference. Therefore we must make copies. 1075 * Allocate space from "ref_slot", which will be adjusted to 1076 * follow the parameters on the stack. 1077 */ 1078 loc[0].ref_slot = cum->ref_slot; 1079 1080 /* 1081 * Subsequent words also go into the reference slot, but 1082 * do not accumulate into the regular arguments. 1083 */ 1084 for (int i = 1; i < n; ++i) { 1085 loc[i] = (TCGCallArgumentLoc){ 1086 .kind = TCG_CALL_ARG_BY_REF_N, 1087 .arg_idx = cum->arg_idx, 1088 .tmp_subindex = i, 1089 .ref_slot = cum->ref_slot + i, 1090 }; 1091 } 1092 cum->info_in_idx += n - 1; /* i=0 accounted for in layout_arg_1 */ 1093 cum->ref_slot += n; 1094 } 1095 1096 static void init_call_layout(TCGHelperInfo *info) 1097 { 1098 int max_reg_slots = ARRAY_SIZE(tcg_target_call_iarg_regs); 1099 int max_stk_slots = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1100 unsigned typemask = info->typemask; 1101 unsigned typecode; 1102 TCGCumulativeArgs cum = { }; 1103 1104 /* 1105 * Parse and place any function return value. 1106 */ 1107 typecode = typemask & 7; 1108 switch (typecode) { 1109 case dh_typecode_void: 1110 info->nr_out = 0; 1111 break; 1112 case dh_typecode_i32: 1113 case dh_typecode_s32: 1114 case dh_typecode_ptr: 1115 info->nr_out = 1; 1116 info->out_kind = TCG_CALL_RET_NORMAL; 1117 break; 1118 case dh_typecode_i64: 1119 case dh_typecode_s64: 1120 info->nr_out = 64 / TCG_TARGET_REG_BITS; 1121 info->out_kind = TCG_CALL_RET_NORMAL; 1122 /* Query the last register now to trigger any assert early. */ 1123 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1124 break; 1125 case dh_typecode_i128: 1126 info->nr_out = 128 / TCG_TARGET_REG_BITS; 1127 info->out_kind = TCG_TARGET_CALL_RET_I128; 1128 switch (TCG_TARGET_CALL_RET_I128) { 1129 case TCG_CALL_RET_NORMAL: 1130 /* Query the last register now to trigger any assert early. */ 1131 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1132 break; 1133 case TCG_CALL_RET_BY_VEC: 1134 /* Query the single register now to trigger any assert early. */ 1135 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0); 1136 break; 1137 case TCG_CALL_RET_BY_REF: 1138 /* 1139 * Allocate the first argument to the output. 1140 * We don't need to store this anywhere, just make it 1141 * unavailable for use in the input loop below. 1142 */ 1143 cum.arg_slot = 1; 1144 break; 1145 default: 1146 qemu_build_not_reached(); 1147 } 1148 break; 1149 default: 1150 g_assert_not_reached(); 1151 } 1152 1153 /* 1154 * Parse and place function arguments. 1155 */ 1156 for (typemask >>= 3; typemask; typemask >>= 3, cum.arg_idx++) { 1157 TCGCallArgumentKind kind; 1158 TCGType type; 1159 1160 typecode = typemask & 7; 1161 switch (typecode) { 1162 case dh_typecode_i32: 1163 case dh_typecode_s32: 1164 type = TCG_TYPE_I32; 1165 break; 1166 case dh_typecode_i64: 1167 case dh_typecode_s64: 1168 type = TCG_TYPE_I64; 1169 break; 1170 case dh_typecode_ptr: 1171 type = TCG_TYPE_PTR; 1172 break; 1173 case dh_typecode_i128: 1174 type = TCG_TYPE_I128; 1175 break; 1176 default: 1177 g_assert_not_reached(); 1178 } 1179 1180 switch (type) { 1181 case TCG_TYPE_I32: 1182 switch (TCG_TARGET_CALL_ARG_I32) { 1183 case TCG_CALL_ARG_EVEN: 1184 layout_arg_even(&cum); 1185 /* fall through */ 1186 case TCG_CALL_ARG_NORMAL: 1187 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1188 break; 1189 case TCG_CALL_ARG_EXTEND: 1190 kind = TCG_CALL_ARG_EXTEND_U + (typecode & 1); 1191 layout_arg_1(&cum, info, kind); 1192 break; 1193 default: 1194 qemu_build_not_reached(); 1195 } 1196 break; 1197 1198 case TCG_TYPE_I64: 1199 switch (TCG_TARGET_CALL_ARG_I64) { 1200 case TCG_CALL_ARG_EVEN: 1201 layout_arg_even(&cum); 1202 /* fall through */ 1203 case TCG_CALL_ARG_NORMAL: 1204 if (TCG_TARGET_REG_BITS == 32) { 1205 layout_arg_normal_n(&cum, info, 2); 1206 } else { 1207 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1208 } 1209 break; 1210 default: 1211 qemu_build_not_reached(); 1212 } 1213 break; 1214 1215 case TCG_TYPE_I128: 1216 switch (TCG_TARGET_CALL_ARG_I128) { 1217 case TCG_CALL_ARG_EVEN: 1218 layout_arg_even(&cum); 1219 /* fall through */ 1220 case TCG_CALL_ARG_NORMAL: 1221 layout_arg_normal_n(&cum, info, 128 / TCG_TARGET_REG_BITS); 1222 break; 1223 case TCG_CALL_ARG_BY_REF: 1224 layout_arg_by_ref(&cum, info); 1225 break; 1226 default: 1227 qemu_build_not_reached(); 1228 } 1229 break; 1230 1231 default: 1232 g_assert_not_reached(); 1233 } 1234 } 1235 info->nr_in = cum.info_in_idx; 1236 1237 /* Validate that we didn't overrun the input array. */ 1238 assert(cum.info_in_idx <= ARRAY_SIZE(info->in)); 1239 /* Validate the backend has enough argument space. */ 1240 assert(cum.arg_slot <= max_reg_slots + max_stk_slots); 1241 1242 /* 1243 * Relocate the "ref_slot" area to the end of the parameters. 1244 * Minimizing this stack offset helps code size for x86, 1245 * which has a signed 8-bit offset encoding. 1246 */ 1247 if (cum.ref_slot != 0) { 1248 int ref_base = 0; 1249 1250 if (cum.arg_slot > max_reg_slots) { 1251 int align = __alignof(Int128) / sizeof(tcg_target_long); 1252 1253 ref_base = cum.arg_slot - max_reg_slots; 1254 if (align > 1) { 1255 ref_base = ROUND_UP(ref_base, align); 1256 } 1257 } 1258 assert(ref_base + cum.ref_slot <= max_stk_slots); 1259 ref_base += max_reg_slots; 1260 1261 if (ref_base != 0) { 1262 for (int i = cum.info_in_idx - 1; i >= 0; --i) { 1263 TCGCallArgumentLoc *loc = &info->in[i]; 1264 switch (loc->kind) { 1265 case TCG_CALL_ARG_BY_REF: 1266 case TCG_CALL_ARG_BY_REF_N: 1267 loc->ref_slot += ref_base; 1268 break; 1269 default: 1270 break; 1271 } 1272 } 1273 } 1274 } 1275 } 1276 1277 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)]; 1278 static void process_op_defs(TCGContext *s); 1279 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1280 TCGReg reg, const char *name); 1281 1282 static void tcg_context_init(unsigned max_cpus) 1283 { 1284 TCGContext *s = &tcg_init_ctx; 1285 int op, total_args, n, i; 1286 TCGOpDef *def; 1287 TCGArgConstraint *args_ct; 1288 TCGTemp *ts; 1289 1290 memset(s, 0, sizeof(*s)); 1291 s->nb_globals = 0; 1292 1293 /* Count total number of arguments and allocate the corresponding 1294 space */ 1295 total_args = 0; 1296 for(op = 0; op < NB_OPS; op++) { 1297 def = &tcg_op_defs[op]; 1298 n = def->nb_iargs + def->nb_oargs; 1299 total_args += n; 1300 } 1301 1302 args_ct = g_new0(TCGArgConstraint, total_args); 1303 1304 for(op = 0; op < NB_OPS; op++) { 1305 def = &tcg_op_defs[op]; 1306 def->args_ct = args_ct; 1307 n = def->nb_iargs + def->nb_oargs; 1308 args_ct += n; 1309 } 1310 1311 init_call_layout(&info_helper_ld32_mmu); 1312 init_call_layout(&info_helper_ld64_mmu); 1313 init_call_layout(&info_helper_ld128_mmu); 1314 init_call_layout(&info_helper_st32_mmu); 1315 init_call_layout(&info_helper_st64_mmu); 1316 init_call_layout(&info_helper_st128_mmu); 1317 1318 tcg_target_init(s); 1319 process_op_defs(s); 1320 1321 /* Reverse the order of the saved registers, assuming they're all at 1322 the start of tcg_target_reg_alloc_order. */ 1323 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) { 1324 int r = tcg_target_reg_alloc_order[n]; 1325 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) { 1326 break; 1327 } 1328 } 1329 for (i = 0; i < n; ++i) { 1330 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i]; 1331 } 1332 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) { 1333 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i]; 1334 } 1335 1336 alloc_tcg_plugin_context(s); 1337 1338 tcg_ctx = s; 1339 /* 1340 * In user-mode we simply share the init context among threads, since we 1341 * use a single region. See the documentation tcg_region_init() for the 1342 * reasoning behind this. 1343 * In softmmu we will have at most max_cpus TCG threads. 1344 */ 1345 #ifdef CONFIG_USER_ONLY 1346 tcg_ctxs = &tcg_ctx; 1347 tcg_cur_ctxs = 1; 1348 tcg_max_ctxs = 1; 1349 #else 1350 tcg_max_ctxs = max_cpus; 1351 tcg_ctxs = g_new0(TCGContext *, max_cpus); 1352 #endif 1353 1354 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0)); 1355 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env"); 1356 cpu_env = temp_tcgv_ptr(ts); 1357 } 1358 1359 void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus) 1360 { 1361 tcg_context_init(max_cpus); 1362 tcg_region_init(tb_size, splitwx, max_cpus); 1363 } 1364 1365 /* 1366 * Allocate TBs right before their corresponding translated code, making 1367 * sure that TBs and code are on different cache lines. 1368 */ 1369 TranslationBlock *tcg_tb_alloc(TCGContext *s) 1370 { 1371 uintptr_t align = qemu_icache_linesize; 1372 TranslationBlock *tb; 1373 void *next; 1374 1375 retry: 1376 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align); 1377 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align); 1378 1379 if (unlikely(next > s->code_gen_highwater)) { 1380 if (tcg_region_alloc(s)) { 1381 return NULL; 1382 } 1383 goto retry; 1384 } 1385 qatomic_set(&s->code_gen_ptr, next); 1386 s->data_gen_ptr = NULL; 1387 return tb; 1388 } 1389 1390 void tcg_prologue_init(TCGContext *s) 1391 { 1392 size_t prologue_size; 1393 1394 s->code_ptr = s->code_gen_ptr; 1395 s->code_buf = s->code_gen_ptr; 1396 s->data_gen_ptr = NULL; 1397 1398 #ifndef CONFIG_TCG_INTERPRETER 1399 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr); 1400 #endif 1401 1402 #ifdef TCG_TARGET_NEED_POOL_LABELS 1403 s->pool_labels = NULL; 1404 #endif 1405 1406 qemu_thread_jit_write(); 1407 /* Generate the prologue. */ 1408 tcg_target_qemu_prologue(s); 1409 1410 #ifdef TCG_TARGET_NEED_POOL_LABELS 1411 /* Allow the prologue to put e.g. guest_base into a pool entry. */ 1412 { 1413 int result = tcg_out_pool_finalize(s); 1414 tcg_debug_assert(result == 0); 1415 } 1416 #endif 1417 1418 prologue_size = tcg_current_code_size(s); 1419 perf_report_prologue(s->code_gen_ptr, prologue_size); 1420 1421 #ifndef CONFIG_TCG_INTERPRETER 1422 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 1423 (uintptr_t)s->code_buf, prologue_size); 1424 #endif 1425 1426 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { 1427 FILE *logfile = qemu_log_trylock(); 1428 if (logfile) { 1429 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size); 1430 if (s->data_gen_ptr) { 1431 size_t code_size = s->data_gen_ptr - s->code_gen_ptr; 1432 size_t data_size = prologue_size - code_size; 1433 size_t i; 1434 1435 disas(logfile, s->code_gen_ptr, code_size); 1436 1437 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) { 1438 if (sizeof(tcg_target_ulong) == 8) { 1439 fprintf(logfile, 1440 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n", 1441 (uintptr_t)s->data_gen_ptr + i, 1442 *(uint64_t *)(s->data_gen_ptr + i)); 1443 } else { 1444 fprintf(logfile, 1445 "0x%08" PRIxPTR ": .long 0x%08x\n", 1446 (uintptr_t)s->data_gen_ptr + i, 1447 *(uint32_t *)(s->data_gen_ptr + i)); 1448 } 1449 } 1450 } else { 1451 disas(logfile, s->code_gen_ptr, prologue_size); 1452 } 1453 fprintf(logfile, "\n"); 1454 qemu_log_unlock(logfile); 1455 } 1456 } 1457 1458 #ifndef CONFIG_TCG_INTERPRETER 1459 /* 1460 * Assert that goto_ptr is implemented completely, setting an epilogue. 1461 * For tci, we use NULL as the signal to return from the interpreter, 1462 * so skip this check. 1463 */ 1464 tcg_debug_assert(tcg_code_gen_epilogue != NULL); 1465 #endif 1466 1467 tcg_region_prologue_set(s); 1468 } 1469 1470 void tcg_func_start(TCGContext *s) 1471 { 1472 tcg_pool_reset(s); 1473 s->nb_temps = s->nb_globals; 1474 1475 /* No temps have been previously allocated for size or locality. */ 1476 memset(s->free_temps, 0, sizeof(s->free_temps)); 1477 1478 /* No constant temps have been previously allocated. */ 1479 for (int i = 0; i < TCG_TYPE_COUNT; ++i) { 1480 if (s->const_table[i]) { 1481 g_hash_table_remove_all(s->const_table[i]); 1482 } 1483 } 1484 1485 s->nb_ops = 0; 1486 s->nb_labels = 0; 1487 s->current_frame_offset = s->frame_start; 1488 1489 #ifdef CONFIG_DEBUG_TCG 1490 s->goto_tb_issue_mask = 0; 1491 #endif 1492 1493 QTAILQ_INIT(&s->ops); 1494 QTAILQ_INIT(&s->free_ops); 1495 QSIMPLEQ_INIT(&s->labels); 1496 1497 tcg_debug_assert(s->addr_type == TCG_TYPE_I32 || 1498 s->addr_type == TCG_TYPE_I64); 1499 1500 #if defined(CONFIG_SOFTMMU) && !defined(CONFIG_TCG_INTERPRETER) 1501 tcg_debug_assert(s->tlb_fast_offset < 0); 1502 tcg_debug_assert(s->tlb_fast_offset >= MIN_TLB_MASK_TABLE_OFS); 1503 #endif 1504 1505 tcg_debug_assert(s->insn_start_words > 0); 1506 } 1507 1508 static TCGTemp *tcg_temp_alloc(TCGContext *s) 1509 { 1510 int n = s->nb_temps++; 1511 1512 if (n >= TCG_MAX_TEMPS) { 1513 tcg_raise_tb_overflow(s); 1514 } 1515 return memset(&s->temps[n], 0, sizeof(TCGTemp)); 1516 } 1517 1518 static TCGTemp *tcg_global_alloc(TCGContext *s) 1519 { 1520 TCGTemp *ts; 1521 1522 tcg_debug_assert(s->nb_globals == s->nb_temps); 1523 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS); 1524 s->nb_globals++; 1525 ts = tcg_temp_alloc(s); 1526 ts->kind = TEMP_GLOBAL; 1527 1528 return ts; 1529 } 1530 1531 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1532 TCGReg reg, const char *name) 1533 { 1534 TCGTemp *ts; 1535 1536 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32); 1537 1538 ts = tcg_global_alloc(s); 1539 ts->base_type = type; 1540 ts->type = type; 1541 ts->kind = TEMP_FIXED; 1542 ts->reg = reg; 1543 ts->name = name; 1544 tcg_regset_set_reg(s->reserved_regs, reg); 1545 1546 return ts; 1547 } 1548 1549 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size) 1550 { 1551 s->frame_start = start; 1552 s->frame_end = start + size; 1553 s->frame_temp 1554 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame"); 1555 } 1556 1557 TCGTemp *tcg_global_mem_new_internal(TCGType type, TCGv_ptr base, 1558 intptr_t offset, const char *name) 1559 { 1560 TCGContext *s = tcg_ctx; 1561 TCGTemp *base_ts = tcgv_ptr_temp(base); 1562 TCGTemp *ts = tcg_global_alloc(s); 1563 int indirect_reg = 0; 1564 1565 switch (base_ts->kind) { 1566 case TEMP_FIXED: 1567 break; 1568 case TEMP_GLOBAL: 1569 /* We do not support double-indirect registers. */ 1570 tcg_debug_assert(!base_ts->indirect_reg); 1571 base_ts->indirect_base = 1; 1572 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64 1573 ? 2 : 1); 1574 indirect_reg = 1; 1575 break; 1576 default: 1577 g_assert_not_reached(); 1578 } 1579 1580 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1581 TCGTemp *ts2 = tcg_global_alloc(s); 1582 char buf[64]; 1583 1584 ts->base_type = TCG_TYPE_I64; 1585 ts->type = TCG_TYPE_I32; 1586 ts->indirect_reg = indirect_reg; 1587 ts->mem_allocated = 1; 1588 ts->mem_base = base_ts; 1589 ts->mem_offset = offset; 1590 pstrcpy(buf, sizeof(buf), name); 1591 pstrcat(buf, sizeof(buf), "_0"); 1592 ts->name = strdup(buf); 1593 1594 tcg_debug_assert(ts2 == ts + 1); 1595 ts2->base_type = TCG_TYPE_I64; 1596 ts2->type = TCG_TYPE_I32; 1597 ts2->indirect_reg = indirect_reg; 1598 ts2->mem_allocated = 1; 1599 ts2->mem_base = base_ts; 1600 ts2->mem_offset = offset + 4; 1601 ts2->temp_subindex = 1; 1602 pstrcpy(buf, sizeof(buf), name); 1603 pstrcat(buf, sizeof(buf), "_1"); 1604 ts2->name = strdup(buf); 1605 } else { 1606 ts->base_type = type; 1607 ts->type = type; 1608 ts->indirect_reg = indirect_reg; 1609 ts->mem_allocated = 1; 1610 ts->mem_base = base_ts; 1611 ts->mem_offset = offset; 1612 ts->name = name; 1613 } 1614 return ts; 1615 } 1616 1617 TCGTemp *tcg_temp_new_internal(TCGType type, TCGTempKind kind) 1618 { 1619 TCGContext *s = tcg_ctx; 1620 TCGTemp *ts; 1621 int n; 1622 1623 if (kind == TEMP_EBB) { 1624 int idx = find_first_bit(s->free_temps[type].l, TCG_MAX_TEMPS); 1625 1626 if (idx < TCG_MAX_TEMPS) { 1627 /* There is already an available temp with the right type. */ 1628 clear_bit(idx, s->free_temps[type].l); 1629 1630 ts = &s->temps[idx]; 1631 ts->temp_allocated = 1; 1632 tcg_debug_assert(ts->base_type == type); 1633 tcg_debug_assert(ts->kind == kind); 1634 return ts; 1635 } 1636 } else { 1637 tcg_debug_assert(kind == TEMP_TB); 1638 } 1639 1640 switch (type) { 1641 case TCG_TYPE_I32: 1642 case TCG_TYPE_V64: 1643 case TCG_TYPE_V128: 1644 case TCG_TYPE_V256: 1645 n = 1; 1646 break; 1647 case TCG_TYPE_I64: 1648 n = 64 / TCG_TARGET_REG_BITS; 1649 break; 1650 case TCG_TYPE_I128: 1651 n = 128 / TCG_TARGET_REG_BITS; 1652 break; 1653 default: 1654 g_assert_not_reached(); 1655 } 1656 1657 ts = tcg_temp_alloc(s); 1658 ts->base_type = type; 1659 ts->temp_allocated = 1; 1660 ts->kind = kind; 1661 1662 if (n == 1) { 1663 ts->type = type; 1664 } else { 1665 ts->type = TCG_TYPE_REG; 1666 1667 for (int i = 1; i < n; ++i) { 1668 TCGTemp *ts2 = tcg_temp_alloc(s); 1669 1670 tcg_debug_assert(ts2 == ts + i); 1671 ts2->base_type = type; 1672 ts2->type = TCG_TYPE_REG; 1673 ts2->temp_allocated = 1; 1674 ts2->temp_subindex = i; 1675 ts2->kind = kind; 1676 } 1677 } 1678 return ts; 1679 } 1680 1681 TCGv_vec tcg_temp_new_vec(TCGType type) 1682 { 1683 TCGTemp *t; 1684 1685 #ifdef CONFIG_DEBUG_TCG 1686 switch (type) { 1687 case TCG_TYPE_V64: 1688 assert(TCG_TARGET_HAS_v64); 1689 break; 1690 case TCG_TYPE_V128: 1691 assert(TCG_TARGET_HAS_v128); 1692 break; 1693 case TCG_TYPE_V256: 1694 assert(TCG_TARGET_HAS_v256); 1695 break; 1696 default: 1697 g_assert_not_reached(); 1698 } 1699 #endif 1700 1701 t = tcg_temp_new_internal(type, TEMP_EBB); 1702 return temp_tcgv_vec(t); 1703 } 1704 1705 /* Create a new temp of the same type as an existing temp. */ 1706 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match) 1707 { 1708 TCGTemp *t = tcgv_vec_temp(match); 1709 1710 tcg_debug_assert(t->temp_allocated != 0); 1711 1712 t = tcg_temp_new_internal(t->base_type, TEMP_EBB); 1713 return temp_tcgv_vec(t); 1714 } 1715 1716 void tcg_temp_free_internal(TCGTemp *ts) 1717 { 1718 TCGContext *s = tcg_ctx; 1719 1720 switch (ts->kind) { 1721 case TEMP_CONST: 1722 case TEMP_TB: 1723 /* Silently ignore free. */ 1724 break; 1725 case TEMP_EBB: 1726 tcg_debug_assert(ts->temp_allocated != 0); 1727 ts->temp_allocated = 0; 1728 set_bit(temp_idx(ts), s->free_temps[ts->base_type].l); 1729 break; 1730 default: 1731 /* It never made sense to free TEMP_FIXED or TEMP_GLOBAL. */ 1732 g_assert_not_reached(); 1733 } 1734 } 1735 1736 TCGTemp *tcg_constant_internal(TCGType type, int64_t val) 1737 { 1738 TCGContext *s = tcg_ctx; 1739 GHashTable *h = s->const_table[type]; 1740 TCGTemp *ts; 1741 1742 if (h == NULL) { 1743 h = g_hash_table_new(g_int64_hash, g_int64_equal); 1744 s->const_table[type] = h; 1745 } 1746 1747 ts = g_hash_table_lookup(h, &val); 1748 if (ts == NULL) { 1749 int64_t *val_ptr; 1750 1751 ts = tcg_temp_alloc(s); 1752 1753 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1754 TCGTemp *ts2 = tcg_temp_alloc(s); 1755 1756 tcg_debug_assert(ts2 == ts + 1); 1757 1758 ts->base_type = TCG_TYPE_I64; 1759 ts->type = TCG_TYPE_I32; 1760 ts->kind = TEMP_CONST; 1761 ts->temp_allocated = 1; 1762 1763 ts2->base_type = TCG_TYPE_I64; 1764 ts2->type = TCG_TYPE_I32; 1765 ts2->kind = TEMP_CONST; 1766 ts2->temp_allocated = 1; 1767 ts2->temp_subindex = 1; 1768 1769 /* 1770 * Retain the full value of the 64-bit constant in the low 1771 * part, so that the hash table works. Actual uses will 1772 * truncate the value to the low part. 1773 */ 1774 ts[HOST_BIG_ENDIAN].val = val; 1775 ts[!HOST_BIG_ENDIAN].val = val >> 32; 1776 val_ptr = &ts[HOST_BIG_ENDIAN].val; 1777 } else { 1778 ts->base_type = type; 1779 ts->type = type; 1780 ts->kind = TEMP_CONST; 1781 ts->temp_allocated = 1; 1782 ts->val = val; 1783 val_ptr = &ts->val; 1784 } 1785 g_hash_table_insert(h, val_ptr, ts); 1786 } 1787 1788 return ts; 1789 } 1790 1791 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val) 1792 { 1793 val = dup_const(vece, val); 1794 return temp_tcgv_vec(tcg_constant_internal(type, val)); 1795 } 1796 1797 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val) 1798 { 1799 TCGTemp *t = tcgv_vec_temp(match); 1800 1801 tcg_debug_assert(t->temp_allocated != 0); 1802 return tcg_constant_vec(t->base_type, vece, val); 1803 } 1804 1805 #ifdef CONFIG_DEBUG_TCG 1806 size_t temp_idx(TCGTemp *ts) 1807 { 1808 ptrdiff_t n = ts - tcg_ctx->temps; 1809 assert(n >= 0 && n < tcg_ctx->nb_temps); 1810 return n; 1811 } 1812 1813 TCGTemp *tcgv_i32_temp(TCGv_i32 v) 1814 { 1815 uintptr_t o = (uintptr_t)v - offsetof(TCGContext, temps); 1816 1817 assert(o < sizeof(TCGTemp) * tcg_ctx->nb_temps); 1818 assert(o % sizeof(TCGTemp) == 0); 1819 1820 return (void *)tcg_ctx + (uintptr_t)v; 1821 } 1822 #endif /* CONFIG_DEBUG_TCG */ 1823 1824 /* Return true if OP may appear in the opcode stream. 1825 Test the runtime variable that controls each opcode. */ 1826 bool tcg_op_supported(TCGOpcode op) 1827 { 1828 const bool have_vec 1829 = TCG_TARGET_HAS_v64 | TCG_TARGET_HAS_v128 | TCG_TARGET_HAS_v256; 1830 1831 switch (op) { 1832 case INDEX_op_discard: 1833 case INDEX_op_set_label: 1834 case INDEX_op_call: 1835 case INDEX_op_br: 1836 case INDEX_op_mb: 1837 case INDEX_op_insn_start: 1838 case INDEX_op_exit_tb: 1839 case INDEX_op_goto_tb: 1840 case INDEX_op_goto_ptr: 1841 case INDEX_op_qemu_ld_a32_i32: 1842 case INDEX_op_qemu_ld_a64_i32: 1843 case INDEX_op_qemu_st_a32_i32: 1844 case INDEX_op_qemu_st_a64_i32: 1845 case INDEX_op_qemu_ld_a32_i64: 1846 case INDEX_op_qemu_ld_a64_i64: 1847 case INDEX_op_qemu_st_a32_i64: 1848 case INDEX_op_qemu_st_a64_i64: 1849 return true; 1850 1851 case INDEX_op_qemu_st8_a32_i32: 1852 case INDEX_op_qemu_st8_a64_i32: 1853 return TCG_TARGET_HAS_qemu_st8_i32; 1854 1855 case INDEX_op_qemu_ld_a32_i128: 1856 case INDEX_op_qemu_ld_a64_i128: 1857 case INDEX_op_qemu_st_a32_i128: 1858 case INDEX_op_qemu_st_a64_i128: 1859 return TCG_TARGET_HAS_qemu_ldst_i128; 1860 1861 case INDEX_op_mov_i32: 1862 case INDEX_op_setcond_i32: 1863 case INDEX_op_brcond_i32: 1864 case INDEX_op_ld8u_i32: 1865 case INDEX_op_ld8s_i32: 1866 case INDEX_op_ld16u_i32: 1867 case INDEX_op_ld16s_i32: 1868 case INDEX_op_ld_i32: 1869 case INDEX_op_st8_i32: 1870 case INDEX_op_st16_i32: 1871 case INDEX_op_st_i32: 1872 case INDEX_op_add_i32: 1873 case INDEX_op_sub_i32: 1874 case INDEX_op_mul_i32: 1875 case INDEX_op_and_i32: 1876 case INDEX_op_or_i32: 1877 case INDEX_op_xor_i32: 1878 case INDEX_op_shl_i32: 1879 case INDEX_op_shr_i32: 1880 case INDEX_op_sar_i32: 1881 return true; 1882 1883 case INDEX_op_negsetcond_i32: 1884 return TCG_TARGET_HAS_negsetcond_i32; 1885 case INDEX_op_movcond_i32: 1886 return TCG_TARGET_HAS_movcond_i32; 1887 case INDEX_op_div_i32: 1888 case INDEX_op_divu_i32: 1889 return TCG_TARGET_HAS_div_i32; 1890 case INDEX_op_rem_i32: 1891 case INDEX_op_remu_i32: 1892 return TCG_TARGET_HAS_rem_i32; 1893 case INDEX_op_div2_i32: 1894 case INDEX_op_divu2_i32: 1895 return TCG_TARGET_HAS_div2_i32; 1896 case INDEX_op_rotl_i32: 1897 case INDEX_op_rotr_i32: 1898 return TCG_TARGET_HAS_rot_i32; 1899 case INDEX_op_deposit_i32: 1900 return TCG_TARGET_HAS_deposit_i32; 1901 case INDEX_op_extract_i32: 1902 return TCG_TARGET_HAS_extract_i32; 1903 case INDEX_op_sextract_i32: 1904 return TCG_TARGET_HAS_sextract_i32; 1905 case INDEX_op_extract2_i32: 1906 return TCG_TARGET_HAS_extract2_i32; 1907 case INDEX_op_add2_i32: 1908 return TCG_TARGET_HAS_add2_i32; 1909 case INDEX_op_sub2_i32: 1910 return TCG_TARGET_HAS_sub2_i32; 1911 case INDEX_op_mulu2_i32: 1912 return TCG_TARGET_HAS_mulu2_i32; 1913 case INDEX_op_muls2_i32: 1914 return TCG_TARGET_HAS_muls2_i32; 1915 case INDEX_op_muluh_i32: 1916 return TCG_TARGET_HAS_muluh_i32; 1917 case INDEX_op_mulsh_i32: 1918 return TCG_TARGET_HAS_mulsh_i32; 1919 case INDEX_op_ext8s_i32: 1920 return TCG_TARGET_HAS_ext8s_i32; 1921 case INDEX_op_ext16s_i32: 1922 return TCG_TARGET_HAS_ext16s_i32; 1923 case INDEX_op_ext8u_i32: 1924 return TCG_TARGET_HAS_ext8u_i32; 1925 case INDEX_op_ext16u_i32: 1926 return TCG_TARGET_HAS_ext16u_i32; 1927 case INDEX_op_bswap16_i32: 1928 return TCG_TARGET_HAS_bswap16_i32; 1929 case INDEX_op_bswap32_i32: 1930 return TCG_TARGET_HAS_bswap32_i32; 1931 case INDEX_op_not_i32: 1932 return TCG_TARGET_HAS_not_i32; 1933 case INDEX_op_neg_i32: 1934 return TCG_TARGET_HAS_neg_i32; 1935 case INDEX_op_andc_i32: 1936 return TCG_TARGET_HAS_andc_i32; 1937 case INDEX_op_orc_i32: 1938 return TCG_TARGET_HAS_orc_i32; 1939 case INDEX_op_eqv_i32: 1940 return TCG_TARGET_HAS_eqv_i32; 1941 case INDEX_op_nand_i32: 1942 return TCG_TARGET_HAS_nand_i32; 1943 case INDEX_op_nor_i32: 1944 return TCG_TARGET_HAS_nor_i32; 1945 case INDEX_op_clz_i32: 1946 return TCG_TARGET_HAS_clz_i32; 1947 case INDEX_op_ctz_i32: 1948 return TCG_TARGET_HAS_ctz_i32; 1949 case INDEX_op_ctpop_i32: 1950 return TCG_TARGET_HAS_ctpop_i32; 1951 1952 case INDEX_op_brcond2_i32: 1953 case INDEX_op_setcond2_i32: 1954 return TCG_TARGET_REG_BITS == 32; 1955 1956 case INDEX_op_mov_i64: 1957 case INDEX_op_setcond_i64: 1958 case INDEX_op_brcond_i64: 1959 case INDEX_op_ld8u_i64: 1960 case INDEX_op_ld8s_i64: 1961 case INDEX_op_ld16u_i64: 1962 case INDEX_op_ld16s_i64: 1963 case INDEX_op_ld32u_i64: 1964 case INDEX_op_ld32s_i64: 1965 case INDEX_op_ld_i64: 1966 case INDEX_op_st8_i64: 1967 case INDEX_op_st16_i64: 1968 case INDEX_op_st32_i64: 1969 case INDEX_op_st_i64: 1970 case INDEX_op_add_i64: 1971 case INDEX_op_sub_i64: 1972 case INDEX_op_mul_i64: 1973 case INDEX_op_and_i64: 1974 case INDEX_op_or_i64: 1975 case INDEX_op_xor_i64: 1976 case INDEX_op_shl_i64: 1977 case INDEX_op_shr_i64: 1978 case INDEX_op_sar_i64: 1979 case INDEX_op_ext_i32_i64: 1980 case INDEX_op_extu_i32_i64: 1981 return TCG_TARGET_REG_BITS == 64; 1982 1983 case INDEX_op_negsetcond_i64: 1984 return TCG_TARGET_HAS_negsetcond_i64; 1985 case INDEX_op_movcond_i64: 1986 return TCG_TARGET_HAS_movcond_i64; 1987 case INDEX_op_div_i64: 1988 case INDEX_op_divu_i64: 1989 return TCG_TARGET_HAS_div_i64; 1990 case INDEX_op_rem_i64: 1991 case INDEX_op_remu_i64: 1992 return TCG_TARGET_HAS_rem_i64; 1993 case INDEX_op_div2_i64: 1994 case INDEX_op_divu2_i64: 1995 return TCG_TARGET_HAS_div2_i64; 1996 case INDEX_op_rotl_i64: 1997 case INDEX_op_rotr_i64: 1998 return TCG_TARGET_HAS_rot_i64; 1999 case INDEX_op_deposit_i64: 2000 return TCG_TARGET_HAS_deposit_i64; 2001 case INDEX_op_extract_i64: 2002 return TCG_TARGET_HAS_extract_i64; 2003 case INDEX_op_sextract_i64: 2004 return TCG_TARGET_HAS_sextract_i64; 2005 case INDEX_op_extract2_i64: 2006 return TCG_TARGET_HAS_extract2_i64; 2007 case INDEX_op_extrl_i64_i32: 2008 case INDEX_op_extrh_i64_i32: 2009 return TCG_TARGET_HAS_extr_i64_i32; 2010 case INDEX_op_ext8s_i64: 2011 return TCG_TARGET_HAS_ext8s_i64; 2012 case INDEX_op_ext16s_i64: 2013 return TCG_TARGET_HAS_ext16s_i64; 2014 case INDEX_op_ext32s_i64: 2015 return TCG_TARGET_HAS_ext32s_i64; 2016 case INDEX_op_ext8u_i64: 2017 return TCG_TARGET_HAS_ext8u_i64; 2018 case INDEX_op_ext16u_i64: 2019 return TCG_TARGET_HAS_ext16u_i64; 2020 case INDEX_op_ext32u_i64: 2021 return TCG_TARGET_HAS_ext32u_i64; 2022 case INDEX_op_bswap16_i64: 2023 return TCG_TARGET_HAS_bswap16_i64; 2024 case INDEX_op_bswap32_i64: 2025 return TCG_TARGET_HAS_bswap32_i64; 2026 case INDEX_op_bswap64_i64: 2027 return TCG_TARGET_HAS_bswap64_i64; 2028 case INDEX_op_not_i64: 2029 return TCG_TARGET_HAS_not_i64; 2030 case INDEX_op_neg_i64: 2031 return TCG_TARGET_HAS_neg_i64; 2032 case INDEX_op_andc_i64: 2033 return TCG_TARGET_HAS_andc_i64; 2034 case INDEX_op_orc_i64: 2035 return TCG_TARGET_HAS_orc_i64; 2036 case INDEX_op_eqv_i64: 2037 return TCG_TARGET_HAS_eqv_i64; 2038 case INDEX_op_nand_i64: 2039 return TCG_TARGET_HAS_nand_i64; 2040 case INDEX_op_nor_i64: 2041 return TCG_TARGET_HAS_nor_i64; 2042 case INDEX_op_clz_i64: 2043 return TCG_TARGET_HAS_clz_i64; 2044 case INDEX_op_ctz_i64: 2045 return TCG_TARGET_HAS_ctz_i64; 2046 case INDEX_op_ctpop_i64: 2047 return TCG_TARGET_HAS_ctpop_i64; 2048 case INDEX_op_add2_i64: 2049 return TCG_TARGET_HAS_add2_i64; 2050 case INDEX_op_sub2_i64: 2051 return TCG_TARGET_HAS_sub2_i64; 2052 case INDEX_op_mulu2_i64: 2053 return TCG_TARGET_HAS_mulu2_i64; 2054 case INDEX_op_muls2_i64: 2055 return TCG_TARGET_HAS_muls2_i64; 2056 case INDEX_op_muluh_i64: 2057 return TCG_TARGET_HAS_muluh_i64; 2058 case INDEX_op_mulsh_i64: 2059 return TCG_TARGET_HAS_mulsh_i64; 2060 2061 case INDEX_op_mov_vec: 2062 case INDEX_op_dup_vec: 2063 case INDEX_op_dupm_vec: 2064 case INDEX_op_ld_vec: 2065 case INDEX_op_st_vec: 2066 case INDEX_op_add_vec: 2067 case INDEX_op_sub_vec: 2068 case INDEX_op_and_vec: 2069 case INDEX_op_or_vec: 2070 case INDEX_op_xor_vec: 2071 case INDEX_op_cmp_vec: 2072 return have_vec; 2073 case INDEX_op_dup2_vec: 2074 return have_vec && TCG_TARGET_REG_BITS == 32; 2075 case INDEX_op_not_vec: 2076 return have_vec && TCG_TARGET_HAS_not_vec; 2077 case INDEX_op_neg_vec: 2078 return have_vec && TCG_TARGET_HAS_neg_vec; 2079 case INDEX_op_abs_vec: 2080 return have_vec && TCG_TARGET_HAS_abs_vec; 2081 case INDEX_op_andc_vec: 2082 return have_vec && TCG_TARGET_HAS_andc_vec; 2083 case INDEX_op_orc_vec: 2084 return have_vec && TCG_TARGET_HAS_orc_vec; 2085 case INDEX_op_nand_vec: 2086 return have_vec && TCG_TARGET_HAS_nand_vec; 2087 case INDEX_op_nor_vec: 2088 return have_vec && TCG_TARGET_HAS_nor_vec; 2089 case INDEX_op_eqv_vec: 2090 return have_vec && TCG_TARGET_HAS_eqv_vec; 2091 case INDEX_op_mul_vec: 2092 return have_vec && TCG_TARGET_HAS_mul_vec; 2093 case INDEX_op_shli_vec: 2094 case INDEX_op_shri_vec: 2095 case INDEX_op_sari_vec: 2096 return have_vec && TCG_TARGET_HAS_shi_vec; 2097 case INDEX_op_shls_vec: 2098 case INDEX_op_shrs_vec: 2099 case INDEX_op_sars_vec: 2100 return have_vec && TCG_TARGET_HAS_shs_vec; 2101 case INDEX_op_shlv_vec: 2102 case INDEX_op_shrv_vec: 2103 case INDEX_op_sarv_vec: 2104 return have_vec && TCG_TARGET_HAS_shv_vec; 2105 case INDEX_op_rotli_vec: 2106 return have_vec && TCG_TARGET_HAS_roti_vec; 2107 case INDEX_op_rotls_vec: 2108 return have_vec && TCG_TARGET_HAS_rots_vec; 2109 case INDEX_op_rotlv_vec: 2110 case INDEX_op_rotrv_vec: 2111 return have_vec && TCG_TARGET_HAS_rotv_vec; 2112 case INDEX_op_ssadd_vec: 2113 case INDEX_op_usadd_vec: 2114 case INDEX_op_sssub_vec: 2115 case INDEX_op_ussub_vec: 2116 return have_vec && TCG_TARGET_HAS_sat_vec; 2117 case INDEX_op_smin_vec: 2118 case INDEX_op_umin_vec: 2119 case INDEX_op_smax_vec: 2120 case INDEX_op_umax_vec: 2121 return have_vec && TCG_TARGET_HAS_minmax_vec; 2122 case INDEX_op_bitsel_vec: 2123 return have_vec && TCG_TARGET_HAS_bitsel_vec; 2124 case INDEX_op_cmpsel_vec: 2125 return have_vec && TCG_TARGET_HAS_cmpsel_vec; 2126 2127 default: 2128 tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS); 2129 return true; 2130 } 2131 } 2132 2133 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs); 2134 2135 static void tcg_gen_callN(TCGHelperInfo *info, TCGTemp *ret, TCGTemp **args) 2136 { 2137 TCGv_i64 extend_free[MAX_CALL_IARGS]; 2138 int n_extend = 0; 2139 TCGOp *op; 2140 int i, n, pi = 0, total_args; 2141 2142 if (unlikely(g_once_init_enter(HELPER_INFO_INIT(info)))) { 2143 init_call_layout(info); 2144 g_once_init_leave(HELPER_INFO_INIT(info), HELPER_INFO_INIT_VAL(info)); 2145 } 2146 2147 total_args = info->nr_out + info->nr_in + 2; 2148 op = tcg_op_alloc(INDEX_op_call, total_args); 2149 2150 #ifdef CONFIG_PLUGIN 2151 /* Flag helpers that may affect guest state */ 2152 if (tcg_ctx->plugin_insn && 2153 !(info->flags & TCG_CALL_PLUGIN) && 2154 !(info->flags & TCG_CALL_NO_SIDE_EFFECTS)) { 2155 tcg_ctx->plugin_insn->calls_helpers = true; 2156 } 2157 #endif 2158 2159 TCGOP_CALLO(op) = n = info->nr_out; 2160 switch (n) { 2161 case 0: 2162 tcg_debug_assert(ret == NULL); 2163 break; 2164 case 1: 2165 tcg_debug_assert(ret != NULL); 2166 op->args[pi++] = temp_arg(ret); 2167 break; 2168 case 2: 2169 case 4: 2170 tcg_debug_assert(ret != NULL); 2171 tcg_debug_assert(ret->base_type == ret->type + ctz32(n)); 2172 tcg_debug_assert(ret->temp_subindex == 0); 2173 for (i = 0; i < n; ++i) { 2174 op->args[pi++] = temp_arg(ret + i); 2175 } 2176 break; 2177 default: 2178 g_assert_not_reached(); 2179 } 2180 2181 TCGOP_CALLI(op) = n = info->nr_in; 2182 for (i = 0; i < n; i++) { 2183 const TCGCallArgumentLoc *loc = &info->in[i]; 2184 TCGTemp *ts = args[loc->arg_idx] + loc->tmp_subindex; 2185 2186 switch (loc->kind) { 2187 case TCG_CALL_ARG_NORMAL: 2188 case TCG_CALL_ARG_BY_REF: 2189 case TCG_CALL_ARG_BY_REF_N: 2190 op->args[pi++] = temp_arg(ts); 2191 break; 2192 2193 case TCG_CALL_ARG_EXTEND_U: 2194 case TCG_CALL_ARG_EXTEND_S: 2195 { 2196 TCGv_i64 temp = tcg_temp_ebb_new_i64(); 2197 TCGv_i32 orig = temp_tcgv_i32(ts); 2198 2199 if (loc->kind == TCG_CALL_ARG_EXTEND_S) { 2200 tcg_gen_ext_i32_i64(temp, orig); 2201 } else { 2202 tcg_gen_extu_i32_i64(temp, orig); 2203 } 2204 op->args[pi++] = tcgv_i64_arg(temp); 2205 extend_free[n_extend++] = temp; 2206 } 2207 break; 2208 2209 default: 2210 g_assert_not_reached(); 2211 } 2212 } 2213 op->args[pi++] = (uintptr_t)info->func; 2214 op->args[pi++] = (uintptr_t)info; 2215 tcg_debug_assert(pi == total_args); 2216 2217 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 2218 2219 tcg_debug_assert(n_extend < ARRAY_SIZE(extend_free)); 2220 for (i = 0; i < n_extend; ++i) { 2221 tcg_temp_free_i64(extend_free[i]); 2222 } 2223 } 2224 2225 void tcg_gen_call0(TCGHelperInfo *info, TCGTemp *ret) 2226 { 2227 tcg_gen_callN(info, ret, NULL); 2228 } 2229 2230 void tcg_gen_call1(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1) 2231 { 2232 tcg_gen_callN(info, ret, &t1); 2233 } 2234 2235 void tcg_gen_call2(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, TCGTemp *t2) 2236 { 2237 TCGTemp *args[2] = { t1, t2 }; 2238 tcg_gen_callN(info, ret, args); 2239 } 2240 2241 void tcg_gen_call3(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2242 TCGTemp *t2, TCGTemp *t3) 2243 { 2244 TCGTemp *args[3] = { t1, t2, t3 }; 2245 tcg_gen_callN(info, ret, args); 2246 } 2247 2248 void tcg_gen_call4(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2249 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4) 2250 { 2251 TCGTemp *args[4] = { t1, t2, t3, t4 }; 2252 tcg_gen_callN(info, ret, args); 2253 } 2254 2255 void tcg_gen_call5(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2256 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, TCGTemp *t5) 2257 { 2258 TCGTemp *args[5] = { t1, t2, t3, t4, t5 }; 2259 tcg_gen_callN(info, ret, args); 2260 } 2261 2262 void tcg_gen_call6(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, TCGTemp *t2, 2263 TCGTemp *t3, TCGTemp *t4, TCGTemp *t5, TCGTemp *t6) 2264 { 2265 TCGTemp *args[6] = { t1, t2, t3, t4, t5, t6 }; 2266 tcg_gen_callN(info, ret, args); 2267 } 2268 2269 void tcg_gen_call7(TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2270 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, 2271 TCGTemp *t5, TCGTemp *t6, TCGTemp *t7) 2272 { 2273 TCGTemp *args[7] = { t1, t2, t3, t4, t5, t6, t7 }; 2274 tcg_gen_callN(info, ret, args); 2275 } 2276 2277 static void tcg_reg_alloc_start(TCGContext *s) 2278 { 2279 int i, n; 2280 2281 for (i = 0, n = s->nb_temps; i < n; i++) { 2282 TCGTemp *ts = &s->temps[i]; 2283 TCGTempVal val = TEMP_VAL_MEM; 2284 2285 switch (ts->kind) { 2286 case TEMP_CONST: 2287 val = TEMP_VAL_CONST; 2288 break; 2289 case TEMP_FIXED: 2290 val = TEMP_VAL_REG; 2291 break; 2292 case TEMP_GLOBAL: 2293 break; 2294 case TEMP_EBB: 2295 val = TEMP_VAL_DEAD; 2296 /* fall through */ 2297 case TEMP_TB: 2298 ts->mem_allocated = 0; 2299 break; 2300 default: 2301 g_assert_not_reached(); 2302 } 2303 ts->val_type = val; 2304 } 2305 2306 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp)); 2307 } 2308 2309 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size, 2310 TCGTemp *ts) 2311 { 2312 int idx = temp_idx(ts); 2313 2314 switch (ts->kind) { 2315 case TEMP_FIXED: 2316 case TEMP_GLOBAL: 2317 pstrcpy(buf, buf_size, ts->name); 2318 break; 2319 case TEMP_TB: 2320 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals); 2321 break; 2322 case TEMP_EBB: 2323 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals); 2324 break; 2325 case TEMP_CONST: 2326 switch (ts->type) { 2327 case TCG_TYPE_I32: 2328 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val); 2329 break; 2330 #if TCG_TARGET_REG_BITS > 32 2331 case TCG_TYPE_I64: 2332 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val); 2333 break; 2334 #endif 2335 case TCG_TYPE_V64: 2336 case TCG_TYPE_V128: 2337 case TCG_TYPE_V256: 2338 snprintf(buf, buf_size, "v%d$0x%" PRIx64, 2339 64 << (ts->type - TCG_TYPE_V64), ts->val); 2340 break; 2341 default: 2342 g_assert_not_reached(); 2343 } 2344 break; 2345 } 2346 return buf; 2347 } 2348 2349 static char *tcg_get_arg_str(TCGContext *s, char *buf, 2350 int buf_size, TCGArg arg) 2351 { 2352 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg)); 2353 } 2354 2355 static const char * const cond_name[] = 2356 { 2357 [TCG_COND_NEVER] = "never", 2358 [TCG_COND_ALWAYS] = "always", 2359 [TCG_COND_EQ] = "eq", 2360 [TCG_COND_NE] = "ne", 2361 [TCG_COND_LT] = "lt", 2362 [TCG_COND_GE] = "ge", 2363 [TCG_COND_LE] = "le", 2364 [TCG_COND_GT] = "gt", 2365 [TCG_COND_LTU] = "ltu", 2366 [TCG_COND_GEU] = "geu", 2367 [TCG_COND_LEU] = "leu", 2368 [TCG_COND_GTU] = "gtu" 2369 }; 2370 2371 static const char * const ldst_name[(MO_BSWAP | MO_SSIZE) + 1] = 2372 { 2373 [MO_UB] = "ub", 2374 [MO_SB] = "sb", 2375 [MO_LEUW] = "leuw", 2376 [MO_LESW] = "lesw", 2377 [MO_LEUL] = "leul", 2378 [MO_LESL] = "lesl", 2379 [MO_LEUQ] = "leq", 2380 [MO_BEUW] = "beuw", 2381 [MO_BESW] = "besw", 2382 [MO_BEUL] = "beul", 2383 [MO_BESL] = "besl", 2384 [MO_BEUQ] = "beq", 2385 [MO_128 + MO_BE] = "beo", 2386 [MO_128 + MO_LE] = "leo", 2387 }; 2388 2389 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = { 2390 [MO_UNALN >> MO_ASHIFT] = "un+", 2391 [MO_ALIGN >> MO_ASHIFT] = "al+", 2392 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+", 2393 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+", 2394 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+", 2395 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+", 2396 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+", 2397 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+", 2398 }; 2399 2400 static const char * const atom_name[(MO_ATOM_MASK >> MO_ATOM_SHIFT) + 1] = { 2401 [MO_ATOM_IFALIGN >> MO_ATOM_SHIFT] = "", 2402 [MO_ATOM_IFALIGN_PAIR >> MO_ATOM_SHIFT] = "pair+", 2403 [MO_ATOM_WITHIN16 >> MO_ATOM_SHIFT] = "w16+", 2404 [MO_ATOM_WITHIN16_PAIR >> MO_ATOM_SHIFT] = "w16p+", 2405 [MO_ATOM_SUBALIGN >> MO_ATOM_SHIFT] = "sub+", 2406 [MO_ATOM_NONE >> MO_ATOM_SHIFT] = "noat+", 2407 }; 2408 2409 static const char bswap_flag_name[][6] = { 2410 [TCG_BSWAP_IZ] = "iz", 2411 [TCG_BSWAP_OZ] = "oz", 2412 [TCG_BSWAP_OS] = "os", 2413 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz", 2414 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os", 2415 }; 2416 2417 static inline bool tcg_regset_single(TCGRegSet d) 2418 { 2419 return (d & (d - 1)) == 0; 2420 } 2421 2422 static inline TCGReg tcg_regset_first(TCGRegSet d) 2423 { 2424 if (TCG_TARGET_NB_REGS <= 32) { 2425 return ctz32(d); 2426 } else { 2427 return ctz64(d); 2428 } 2429 } 2430 2431 /* Return only the number of characters output -- no error return. */ 2432 #define ne_fprintf(...) \ 2433 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; }) 2434 2435 static void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs) 2436 { 2437 char buf[128]; 2438 TCGOp *op; 2439 2440 QTAILQ_FOREACH(op, &s->ops, link) { 2441 int i, k, nb_oargs, nb_iargs, nb_cargs; 2442 const TCGOpDef *def; 2443 TCGOpcode c; 2444 int col = 0; 2445 2446 c = op->opc; 2447 def = &tcg_op_defs[c]; 2448 2449 if (c == INDEX_op_insn_start) { 2450 nb_oargs = 0; 2451 col += ne_fprintf(f, "\n ----"); 2452 2453 for (i = 0, k = s->insn_start_words; i < k; ++i) { 2454 col += ne_fprintf(f, " %016" PRIx64, 2455 tcg_get_insn_start_param(op, i)); 2456 } 2457 } else if (c == INDEX_op_call) { 2458 const TCGHelperInfo *info = tcg_call_info(op); 2459 void *func = tcg_call_func(op); 2460 2461 /* variable number of arguments */ 2462 nb_oargs = TCGOP_CALLO(op); 2463 nb_iargs = TCGOP_CALLI(op); 2464 nb_cargs = def->nb_cargs; 2465 2466 col += ne_fprintf(f, " %s ", def->name); 2467 2468 /* 2469 * Print the function name from TCGHelperInfo, if available. 2470 * Note that plugins have a template function for the info, 2471 * but the actual function pointer comes from the plugin. 2472 */ 2473 if (func == info->func) { 2474 col += ne_fprintf(f, "%s", info->name); 2475 } else { 2476 col += ne_fprintf(f, "plugin(%p)", func); 2477 } 2478 2479 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs); 2480 for (i = 0; i < nb_oargs; i++) { 2481 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf), 2482 op->args[i])); 2483 } 2484 for (i = 0; i < nb_iargs; i++) { 2485 TCGArg arg = op->args[nb_oargs + i]; 2486 const char *t = tcg_get_arg_str(s, buf, sizeof(buf), arg); 2487 col += ne_fprintf(f, ",%s", t); 2488 } 2489 } else { 2490 col += ne_fprintf(f, " %s ", def->name); 2491 2492 nb_oargs = def->nb_oargs; 2493 nb_iargs = def->nb_iargs; 2494 nb_cargs = def->nb_cargs; 2495 2496 if (def->flags & TCG_OPF_VECTOR) { 2497 col += ne_fprintf(f, "v%d,e%d,", 64 << TCGOP_VECL(op), 2498 8 << TCGOP_VECE(op)); 2499 } 2500 2501 k = 0; 2502 for (i = 0; i < nb_oargs; i++) { 2503 const char *sep = k ? "," : ""; 2504 col += ne_fprintf(f, "%s%s", sep, 2505 tcg_get_arg_str(s, buf, sizeof(buf), 2506 op->args[k++])); 2507 } 2508 for (i = 0; i < nb_iargs; i++) { 2509 const char *sep = k ? "," : ""; 2510 col += ne_fprintf(f, "%s%s", sep, 2511 tcg_get_arg_str(s, buf, sizeof(buf), 2512 op->args[k++])); 2513 } 2514 switch (c) { 2515 case INDEX_op_brcond_i32: 2516 case INDEX_op_setcond_i32: 2517 case INDEX_op_negsetcond_i32: 2518 case INDEX_op_movcond_i32: 2519 case INDEX_op_brcond2_i32: 2520 case INDEX_op_setcond2_i32: 2521 case INDEX_op_brcond_i64: 2522 case INDEX_op_setcond_i64: 2523 case INDEX_op_negsetcond_i64: 2524 case INDEX_op_movcond_i64: 2525 case INDEX_op_cmp_vec: 2526 case INDEX_op_cmpsel_vec: 2527 if (op->args[k] < ARRAY_SIZE(cond_name) 2528 && cond_name[op->args[k]]) { 2529 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]); 2530 } else { 2531 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]); 2532 } 2533 i = 1; 2534 break; 2535 case INDEX_op_qemu_ld_a32_i32: 2536 case INDEX_op_qemu_ld_a64_i32: 2537 case INDEX_op_qemu_st_a32_i32: 2538 case INDEX_op_qemu_st_a64_i32: 2539 case INDEX_op_qemu_st8_a32_i32: 2540 case INDEX_op_qemu_st8_a64_i32: 2541 case INDEX_op_qemu_ld_a32_i64: 2542 case INDEX_op_qemu_ld_a64_i64: 2543 case INDEX_op_qemu_st_a32_i64: 2544 case INDEX_op_qemu_st_a64_i64: 2545 case INDEX_op_qemu_ld_a32_i128: 2546 case INDEX_op_qemu_ld_a64_i128: 2547 case INDEX_op_qemu_st_a32_i128: 2548 case INDEX_op_qemu_st_a64_i128: 2549 { 2550 const char *s_al, *s_op, *s_at; 2551 MemOpIdx oi = op->args[k++]; 2552 MemOp op = get_memop(oi); 2553 unsigned ix = get_mmuidx(oi); 2554 2555 s_al = alignment_name[(op & MO_AMASK) >> MO_ASHIFT]; 2556 s_op = ldst_name[op & (MO_BSWAP | MO_SSIZE)]; 2557 s_at = atom_name[(op & MO_ATOM_MASK) >> MO_ATOM_SHIFT]; 2558 op &= ~(MO_AMASK | MO_BSWAP | MO_SSIZE | MO_ATOM_MASK); 2559 2560 /* If all fields are accounted for, print symbolically. */ 2561 if (!op && s_al && s_op && s_at) { 2562 col += ne_fprintf(f, ",%s%s%s,%u", 2563 s_at, s_al, s_op, ix); 2564 } else { 2565 op = get_memop(oi); 2566 col += ne_fprintf(f, ",$0x%x,%u", op, ix); 2567 } 2568 i = 1; 2569 } 2570 break; 2571 case INDEX_op_bswap16_i32: 2572 case INDEX_op_bswap16_i64: 2573 case INDEX_op_bswap32_i32: 2574 case INDEX_op_bswap32_i64: 2575 case INDEX_op_bswap64_i64: 2576 { 2577 TCGArg flags = op->args[k]; 2578 const char *name = NULL; 2579 2580 if (flags < ARRAY_SIZE(bswap_flag_name)) { 2581 name = bswap_flag_name[flags]; 2582 } 2583 if (name) { 2584 col += ne_fprintf(f, ",%s", name); 2585 } else { 2586 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags); 2587 } 2588 i = k = 1; 2589 } 2590 break; 2591 default: 2592 i = 0; 2593 break; 2594 } 2595 switch (c) { 2596 case INDEX_op_set_label: 2597 case INDEX_op_br: 2598 case INDEX_op_brcond_i32: 2599 case INDEX_op_brcond_i64: 2600 case INDEX_op_brcond2_i32: 2601 col += ne_fprintf(f, "%s$L%d", k ? "," : "", 2602 arg_label(op->args[k])->id); 2603 i++, k++; 2604 break; 2605 case INDEX_op_mb: 2606 { 2607 TCGBar membar = op->args[k]; 2608 const char *b_op, *m_op; 2609 2610 switch (membar & TCG_BAR_SC) { 2611 case 0: 2612 b_op = "none"; 2613 break; 2614 case TCG_BAR_LDAQ: 2615 b_op = "acq"; 2616 break; 2617 case TCG_BAR_STRL: 2618 b_op = "rel"; 2619 break; 2620 case TCG_BAR_SC: 2621 b_op = "seq"; 2622 break; 2623 default: 2624 g_assert_not_reached(); 2625 } 2626 2627 switch (membar & TCG_MO_ALL) { 2628 case 0: 2629 m_op = "none"; 2630 break; 2631 case TCG_MO_LD_LD: 2632 m_op = "rr"; 2633 break; 2634 case TCG_MO_LD_ST: 2635 m_op = "rw"; 2636 break; 2637 case TCG_MO_ST_LD: 2638 m_op = "wr"; 2639 break; 2640 case TCG_MO_ST_ST: 2641 m_op = "ww"; 2642 break; 2643 case TCG_MO_LD_LD | TCG_MO_LD_ST: 2644 m_op = "rr+rw"; 2645 break; 2646 case TCG_MO_LD_LD | TCG_MO_ST_LD: 2647 m_op = "rr+wr"; 2648 break; 2649 case TCG_MO_LD_LD | TCG_MO_ST_ST: 2650 m_op = "rr+ww"; 2651 break; 2652 case TCG_MO_LD_ST | TCG_MO_ST_LD: 2653 m_op = "rw+wr"; 2654 break; 2655 case TCG_MO_LD_ST | TCG_MO_ST_ST: 2656 m_op = "rw+ww"; 2657 break; 2658 case TCG_MO_ST_LD | TCG_MO_ST_ST: 2659 m_op = "wr+ww"; 2660 break; 2661 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_LD: 2662 m_op = "rr+rw+wr"; 2663 break; 2664 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_ST: 2665 m_op = "rr+rw+ww"; 2666 break; 2667 case TCG_MO_LD_LD | TCG_MO_ST_LD | TCG_MO_ST_ST: 2668 m_op = "rr+wr+ww"; 2669 break; 2670 case TCG_MO_LD_ST | TCG_MO_ST_LD | TCG_MO_ST_ST: 2671 m_op = "rw+wr+ww"; 2672 break; 2673 case TCG_MO_ALL: 2674 m_op = "all"; 2675 break; 2676 default: 2677 g_assert_not_reached(); 2678 } 2679 2680 col += ne_fprintf(f, "%s%s:%s", (k ? "," : ""), b_op, m_op); 2681 i++, k++; 2682 } 2683 break; 2684 default: 2685 break; 2686 } 2687 for (; i < nb_cargs; i++, k++) { 2688 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "", 2689 op->args[k]); 2690 } 2691 } 2692 2693 if (have_prefs || op->life) { 2694 for (; col < 40; ++col) { 2695 putc(' ', f); 2696 } 2697 } 2698 2699 if (op->life) { 2700 unsigned life = op->life; 2701 2702 if (life & (SYNC_ARG * 3)) { 2703 ne_fprintf(f, " sync:"); 2704 for (i = 0; i < 2; ++i) { 2705 if (life & (SYNC_ARG << i)) { 2706 ne_fprintf(f, " %d", i); 2707 } 2708 } 2709 } 2710 life /= DEAD_ARG; 2711 if (life) { 2712 ne_fprintf(f, " dead:"); 2713 for (i = 0; life; ++i, life >>= 1) { 2714 if (life & 1) { 2715 ne_fprintf(f, " %d", i); 2716 } 2717 } 2718 } 2719 } 2720 2721 if (have_prefs) { 2722 for (i = 0; i < nb_oargs; ++i) { 2723 TCGRegSet set = output_pref(op, i); 2724 2725 if (i == 0) { 2726 ne_fprintf(f, " pref="); 2727 } else { 2728 ne_fprintf(f, ","); 2729 } 2730 if (set == 0) { 2731 ne_fprintf(f, "none"); 2732 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) { 2733 ne_fprintf(f, "all"); 2734 #ifdef CONFIG_DEBUG_TCG 2735 } else if (tcg_regset_single(set)) { 2736 TCGReg reg = tcg_regset_first(set); 2737 ne_fprintf(f, "%s", tcg_target_reg_names[reg]); 2738 #endif 2739 } else if (TCG_TARGET_NB_REGS <= 32) { 2740 ne_fprintf(f, "0x%x", (uint32_t)set); 2741 } else { 2742 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set); 2743 } 2744 } 2745 } 2746 2747 putc('\n', f); 2748 } 2749 } 2750 2751 /* we give more priority to constraints with less registers */ 2752 static int get_constraint_priority(const TCGOpDef *def, int k) 2753 { 2754 const TCGArgConstraint *arg_ct = &def->args_ct[k]; 2755 int n = ctpop64(arg_ct->regs); 2756 2757 /* 2758 * Sort constraints of a single register first, which includes output 2759 * aliases (which must exactly match the input already allocated). 2760 */ 2761 if (n == 1 || arg_ct->oalias) { 2762 return INT_MAX; 2763 } 2764 2765 /* 2766 * Sort register pairs next, first then second immediately after. 2767 * Arbitrarily sort multiple pairs by the index of the first reg; 2768 * there shouldn't be many pairs. 2769 */ 2770 switch (arg_ct->pair) { 2771 case 1: 2772 case 3: 2773 return (k + 1) * 2; 2774 case 2: 2775 return (arg_ct->pair_index + 1) * 2 - 1; 2776 } 2777 2778 /* Finally, sort by decreasing register count. */ 2779 assert(n > 1); 2780 return -n; 2781 } 2782 2783 /* sort from highest priority to lowest */ 2784 static void sort_constraints(TCGOpDef *def, int start, int n) 2785 { 2786 int i, j; 2787 TCGArgConstraint *a = def->args_ct; 2788 2789 for (i = 0; i < n; i++) { 2790 a[start + i].sort_index = start + i; 2791 } 2792 if (n <= 1) { 2793 return; 2794 } 2795 for (i = 0; i < n - 1; i++) { 2796 for (j = i + 1; j < n; j++) { 2797 int p1 = get_constraint_priority(def, a[start + i].sort_index); 2798 int p2 = get_constraint_priority(def, a[start + j].sort_index); 2799 if (p1 < p2) { 2800 int tmp = a[start + i].sort_index; 2801 a[start + i].sort_index = a[start + j].sort_index; 2802 a[start + j].sort_index = tmp; 2803 } 2804 } 2805 } 2806 } 2807 2808 static void process_op_defs(TCGContext *s) 2809 { 2810 TCGOpcode op; 2811 2812 for (op = 0; op < NB_OPS; op++) { 2813 TCGOpDef *def = &tcg_op_defs[op]; 2814 const TCGTargetOpDef *tdefs; 2815 bool saw_alias_pair = false; 2816 int i, o, i2, o2, nb_args; 2817 2818 if (def->flags & TCG_OPF_NOT_PRESENT) { 2819 continue; 2820 } 2821 2822 nb_args = def->nb_iargs + def->nb_oargs; 2823 if (nb_args == 0) { 2824 continue; 2825 } 2826 2827 /* 2828 * Macro magic should make it impossible, but double-check that 2829 * the array index is in range. Since the signness of an enum 2830 * is implementation defined, force the result to unsigned. 2831 */ 2832 unsigned con_set = tcg_target_op_def(op); 2833 tcg_debug_assert(con_set < ARRAY_SIZE(constraint_sets)); 2834 tdefs = &constraint_sets[con_set]; 2835 2836 for (i = 0; i < nb_args; i++) { 2837 const char *ct_str = tdefs->args_ct_str[i]; 2838 bool input_p = i >= def->nb_oargs; 2839 2840 /* Incomplete TCGTargetOpDef entry. */ 2841 tcg_debug_assert(ct_str != NULL); 2842 2843 switch (*ct_str) { 2844 case '0' ... '9': 2845 o = *ct_str - '0'; 2846 tcg_debug_assert(input_p); 2847 tcg_debug_assert(o < def->nb_oargs); 2848 tcg_debug_assert(def->args_ct[o].regs != 0); 2849 tcg_debug_assert(!def->args_ct[o].oalias); 2850 def->args_ct[i] = def->args_ct[o]; 2851 /* The output sets oalias. */ 2852 def->args_ct[o].oalias = 1; 2853 def->args_ct[o].alias_index = i; 2854 /* The input sets ialias. */ 2855 def->args_ct[i].ialias = 1; 2856 def->args_ct[i].alias_index = o; 2857 if (def->args_ct[i].pair) { 2858 saw_alias_pair = true; 2859 } 2860 tcg_debug_assert(ct_str[1] == '\0'); 2861 continue; 2862 2863 case '&': 2864 tcg_debug_assert(!input_p); 2865 def->args_ct[i].newreg = true; 2866 ct_str++; 2867 break; 2868 2869 case 'p': /* plus */ 2870 /* Allocate to the register after the previous. */ 2871 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0)); 2872 o = i - 1; 2873 tcg_debug_assert(!def->args_ct[o].pair); 2874 tcg_debug_assert(!def->args_ct[o].ct); 2875 def->args_ct[i] = (TCGArgConstraint){ 2876 .pair = 2, 2877 .pair_index = o, 2878 .regs = def->args_ct[o].regs << 1, 2879 }; 2880 def->args_ct[o].pair = 1; 2881 def->args_ct[o].pair_index = i; 2882 tcg_debug_assert(ct_str[1] == '\0'); 2883 continue; 2884 2885 case 'm': /* minus */ 2886 /* Allocate to the register before the previous. */ 2887 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0)); 2888 o = i - 1; 2889 tcg_debug_assert(!def->args_ct[o].pair); 2890 tcg_debug_assert(!def->args_ct[o].ct); 2891 def->args_ct[i] = (TCGArgConstraint){ 2892 .pair = 1, 2893 .pair_index = o, 2894 .regs = def->args_ct[o].regs >> 1, 2895 }; 2896 def->args_ct[o].pair = 2; 2897 def->args_ct[o].pair_index = i; 2898 tcg_debug_assert(ct_str[1] == '\0'); 2899 continue; 2900 } 2901 2902 do { 2903 switch (*ct_str) { 2904 case 'i': 2905 def->args_ct[i].ct |= TCG_CT_CONST; 2906 break; 2907 2908 /* Include all of the target-specific constraints. */ 2909 2910 #undef CONST 2911 #define CONST(CASE, MASK) \ 2912 case CASE: def->args_ct[i].ct |= MASK; break; 2913 #define REGS(CASE, MASK) \ 2914 case CASE: def->args_ct[i].regs |= MASK; break; 2915 2916 #include "tcg-target-con-str.h" 2917 2918 #undef REGS 2919 #undef CONST 2920 default: 2921 case '0' ... '9': 2922 case '&': 2923 case 'p': 2924 case 'm': 2925 /* Typo in TCGTargetOpDef constraint. */ 2926 g_assert_not_reached(); 2927 } 2928 } while (*++ct_str != '\0'); 2929 } 2930 2931 /* TCGTargetOpDef entry with too much information? */ 2932 tcg_debug_assert(i == TCG_MAX_OP_ARGS || tdefs->args_ct_str[i] == NULL); 2933 2934 /* 2935 * Fix up output pairs that are aliased with inputs. 2936 * When we created the alias, we copied pair from the output. 2937 * There are three cases: 2938 * (1a) Pairs of inputs alias pairs of outputs. 2939 * (1b) One input aliases the first of a pair of outputs. 2940 * (2) One input aliases the second of a pair of outputs. 2941 * 2942 * Case 1a is handled by making sure that the pair_index'es are 2943 * properly updated so that they appear the same as a pair of inputs. 2944 * 2945 * Case 1b is handled by setting the pair_index of the input to 2946 * itself, simply so it doesn't point to an unrelated argument. 2947 * Since we don't encounter the "second" during the input allocation 2948 * phase, nothing happens with the second half of the input pair. 2949 * 2950 * Case 2 is handled by setting the second input to pair=3, the 2951 * first output to pair=3, and the pair_index'es to match. 2952 */ 2953 if (saw_alias_pair) { 2954 for (i = def->nb_oargs; i < nb_args; i++) { 2955 /* 2956 * Since [0-9pm] must be alone in the constraint string, 2957 * the only way they can both be set is if the pair comes 2958 * from the output alias. 2959 */ 2960 if (!def->args_ct[i].ialias) { 2961 continue; 2962 } 2963 switch (def->args_ct[i].pair) { 2964 case 0: 2965 break; 2966 case 1: 2967 o = def->args_ct[i].alias_index; 2968 o2 = def->args_ct[o].pair_index; 2969 tcg_debug_assert(def->args_ct[o].pair == 1); 2970 tcg_debug_assert(def->args_ct[o2].pair == 2); 2971 if (def->args_ct[o2].oalias) { 2972 /* Case 1a */ 2973 i2 = def->args_ct[o2].alias_index; 2974 tcg_debug_assert(def->args_ct[i2].pair == 2); 2975 def->args_ct[i2].pair_index = i; 2976 def->args_ct[i].pair_index = i2; 2977 } else { 2978 /* Case 1b */ 2979 def->args_ct[i].pair_index = i; 2980 } 2981 break; 2982 case 2: 2983 o = def->args_ct[i].alias_index; 2984 o2 = def->args_ct[o].pair_index; 2985 tcg_debug_assert(def->args_ct[o].pair == 2); 2986 tcg_debug_assert(def->args_ct[o2].pair == 1); 2987 if (def->args_ct[o2].oalias) { 2988 /* Case 1a */ 2989 i2 = def->args_ct[o2].alias_index; 2990 tcg_debug_assert(def->args_ct[i2].pair == 1); 2991 def->args_ct[i2].pair_index = i; 2992 def->args_ct[i].pair_index = i2; 2993 } else { 2994 /* Case 2 */ 2995 def->args_ct[i].pair = 3; 2996 def->args_ct[o2].pair = 3; 2997 def->args_ct[i].pair_index = o2; 2998 def->args_ct[o2].pair_index = i; 2999 } 3000 break; 3001 default: 3002 g_assert_not_reached(); 3003 } 3004 } 3005 } 3006 3007 /* sort the constraints (XXX: this is just an heuristic) */ 3008 sort_constraints(def, 0, def->nb_oargs); 3009 sort_constraints(def, def->nb_oargs, def->nb_iargs); 3010 } 3011 } 3012 3013 static void remove_label_use(TCGOp *op, int idx) 3014 { 3015 TCGLabel *label = arg_label(op->args[idx]); 3016 TCGLabelUse *use; 3017 3018 QSIMPLEQ_FOREACH(use, &label->branches, next) { 3019 if (use->op == op) { 3020 QSIMPLEQ_REMOVE(&label->branches, use, TCGLabelUse, next); 3021 return; 3022 } 3023 } 3024 g_assert_not_reached(); 3025 } 3026 3027 void tcg_op_remove(TCGContext *s, TCGOp *op) 3028 { 3029 switch (op->opc) { 3030 case INDEX_op_br: 3031 remove_label_use(op, 0); 3032 break; 3033 case INDEX_op_brcond_i32: 3034 case INDEX_op_brcond_i64: 3035 remove_label_use(op, 3); 3036 break; 3037 case INDEX_op_brcond2_i32: 3038 remove_label_use(op, 5); 3039 break; 3040 default: 3041 break; 3042 } 3043 3044 QTAILQ_REMOVE(&s->ops, op, link); 3045 QTAILQ_INSERT_TAIL(&s->free_ops, op, link); 3046 s->nb_ops--; 3047 } 3048 3049 void tcg_remove_ops_after(TCGOp *op) 3050 { 3051 TCGContext *s = tcg_ctx; 3052 3053 while (true) { 3054 TCGOp *last = tcg_last_op(); 3055 if (last == op) { 3056 return; 3057 } 3058 tcg_op_remove(s, last); 3059 } 3060 } 3061 3062 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs) 3063 { 3064 TCGContext *s = tcg_ctx; 3065 TCGOp *op = NULL; 3066 3067 if (unlikely(!QTAILQ_EMPTY(&s->free_ops))) { 3068 QTAILQ_FOREACH(op, &s->free_ops, link) { 3069 if (nargs <= op->nargs) { 3070 QTAILQ_REMOVE(&s->free_ops, op, link); 3071 nargs = op->nargs; 3072 goto found; 3073 } 3074 } 3075 } 3076 3077 /* Most opcodes have 3 or 4 operands: reduce fragmentation. */ 3078 nargs = MAX(4, nargs); 3079 op = tcg_malloc(sizeof(TCGOp) + sizeof(TCGArg) * nargs); 3080 3081 found: 3082 memset(op, 0, offsetof(TCGOp, link)); 3083 op->opc = opc; 3084 op->nargs = nargs; 3085 3086 /* Check for bitfield overflow. */ 3087 tcg_debug_assert(op->nargs == nargs); 3088 3089 s->nb_ops++; 3090 return op; 3091 } 3092 3093 TCGOp *tcg_emit_op(TCGOpcode opc, unsigned nargs) 3094 { 3095 TCGOp *op = tcg_op_alloc(opc, nargs); 3096 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 3097 return op; 3098 } 3099 3100 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op, 3101 TCGOpcode opc, unsigned nargs) 3102 { 3103 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3104 QTAILQ_INSERT_BEFORE(old_op, new_op, link); 3105 return new_op; 3106 } 3107 3108 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op, 3109 TCGOpcode opc, unsigned nargs) 3110 { 3111 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3112 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link); 3113 return new_op; 3114 } 3115 3116 static void move_label_uses(TCGLabel *to, TCGLabel *from) 3117 { 3118 TCGLabelUse *u; 3119 3120 QSIMPLEQ_FOREACH(u, &from->branches, next) { 3121 TCGOp *op = u->op; 3122 switch (op->opc) { 3123 case INDEX_op_br: 3124 op->args[0] = label_arg(to); 3125 break; 3126 case INDEX_op_brcond_i32: 3127 case INDEX_op_brcond_i64: 3128 op->args[3] = label_arg(to); 3129 break; 3130 case INDEX_op_brcond2_i32: 3131 op->args[5] = label_arg(to); 3132 break; 3133 default: 3134 g_assert_not_reached(); 3135 } 3136 } 3137 3138 QSIMPLEQ_CONCAT(&to->branches, &from->branches); 3139 } 3140 3141 /* Reachable analysis : remove unreachable code. */ 3142 static void __attribute__((noinline)) 3143 reachable_code_pass(TCGContext *s) 3144 { 3145 TCGOp *op, *op_next, *op_prev; 3146 bool dead = false; 3147 3148 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3149 bool remove = dead; 3150 TCGLabel *label; 3151 3152 switch (op->opc) { 3153 case INDEX_op_set_label: 3154 label = arg_label(op->args[0]); 3155 3156 /* 3157 * Note that the first op in the TB is always a load, 3158 * so there is always something before a label. 3159 */ 3160 op_prev = QTAILQ_PREV(op, link); 3161 3162 /* 3163 * If we find two sequential labels, move all branches to 3164 * reference the second label and remove the first label. 3165 * Do this before branch to next optimization, so that the 3166 * middle label is out of the way. 3167 */ 3168 if (op_prev->opc == INDEX_op_set_label) { 3169 move_label_uses(label, arg_label(op_prev->args[0])); 3170 tcg_op_remove(s, op_prev); 3171 op_prev = QTAILQ_PREV(op, link); 3172 } 3173 3174 /* 3175 * Optimization can fold conditional branches to unconditional. 3176 * If we find a label which is preceded by an unconditional 3177 * branch to next, remove the branch. We couldn't do this when 3178 * processing the branch because any dead code between the branch 3179 * and label had not yet been removed. 3180 */ 3181 if (op_prev->opc == INDEX_op_br && 3182 label == arg_label(op_prev->args[0])) { 3183 tcg_op_remove(s, op_prev); 3184 /* Fall through means insns become live again. */ 3185 dead = false; 3186 } 3187 3188 if (QSIMPLEQ_EMPTY(&label->branches)) { 3189 /* 3190 * While there is an occasional backward branch, virtually 3191 * all branches generated by the translators are forward. 3192 * Which means that generally we will have already removed 3193 * all references to the label that will be, and there is 3194 * little to be gained by iterating. 3195 */ 3196 remove = true; 3197 } else { 3198 /* Once we see a label, insns become live again. */ 3199 dead = false; 3200 remove = false; 3201 } 3202 break; 3203 3204 case INDEX_op_br: 3205 case INDEX_op_exit_tb: 3206 case INDEX_op_goto_ptr: 3207 /* Unconditional branches; everything following is dead. */ 3208 dead = true; 3209 break; 3210 3211 case INDEX_op_call: 3212 /* Notice noreturn helper calls, raising exceptions. */ 3213 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) { 3214 dead = true; 3215 } 3216 break; 3217 3218 case INDEX_op_insn_start: 3219 /* Never remove -- we need to keep these for unwind. */ 3220 remove = false; 3221 break; 3222 3223 default: 3224 break; 3225 } 3226 3227 if (remove) { 3228 tcg_op_remove(s, op); 3229 } 3230 } 3231 } 3232 3233 #define TS_DEAD 1 3234 #define TS_MEM 2 3235 3236 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n))) 3237 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n))) 3238 3239 /* For liveness_pass_1, the register preferences for a given temp. */ 3240 static inline TCGRegSet *la_temp_pref(TCGTemp *ts) 3241 { 3242 return ts->state_ptr; 3243 } 3244 3245 /* For liveness_pass_1, reset the preferences for a given temp to the 3246 * maximal regset for its type. 3247 */ 3248 static inline void la_reset_pref(TCGTemp *ts) 3249 { 3250 *la_temp_pref(ts) 3251 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]); 3252 } 3253 3254 /* liveness analysis: end of function: all temps are dead, and globals 3255 should be in memory. */ 3256 static void la_func_end(TCGContext *s, int ng, int nt) 3257 { 3258 int i; 3259 3260 for (i = 0; i < ng; ++i) { 3261 s->temps[i].state = TS_DEAD | TS_MEM; 3262 la_reset_pref(&s->temps[i]); 3263 } 3264 for (i = ng; i < nt; ++i) { 3265 s->temps[i].state = TS_DEAD; 3266 la_reset_pref(&s->temps[i]); 3267 } 3268 } 3269 3270 /* liveness analysis: end of basic block: all temps are dead, globals 3271 and local temps should be in memory. */ 3272 static void la_bb_end(TCGContext *s, int ng, int nt) 3273 { 3274 int i; 3275 3276 for (i = 0; i < nt; ++i) { 3277 TCGTemp *ts = &s->temps[i]; 3278 int state; 3279 3280 switch (ts->kind) { 3281 case TEMP_FIXED: 3282 case TEMP_GLOBAL: 3283 case TEMP_TB: 3284 state = TS_DEAD | TS_MEM; 3285 break; 3286 case TEMP_EBB: 3287 case TEMP_CONST: 3288 state = TS_DEAD; 3289 break; 3290 default: 3291 g_assert_not_reached(); 3292 } 3293 ts->state = state; 3294 la_reset_pref(ts); 3295 } 3296 } 3297 3298 /* liveness analysis: sync globals back to memory. */ 3299 static void la_global_sync(TCGContext *s, int ng) 3300 { 3301 int i; 3302 3303 for (i = 0; i < ng; ++i) { 3304 int state = s->temps[i].state; 3305 s->temps[i].state = state | TS_MEM; 3306 if (state == TS_DEAD) { 3307 /* If the global was previously dead, reset prefs. */ 3308 la_reset_pref(&s->temps[i]); 3309 } 3310 } 3311 } 3312 3313 /* 3314 * liveness analysis: conditional branch: all temps are dead unless 3315 * explicitly live-across-conditional-branch, globals and local temps 3316 * should be synced. 3317 */ 3318 static void la_bb_sync(TCGContext *s, int ng, int nt) 3319 { 3320 la_global_sync(s, ng); 3321 3322 for (int i = ng; i < nt; ++i) { 3323 TCGTemp *ts = &s->temps[i]; 3324 int state; 3325 3326 switch (ts->kind) { 3327 case TEMP_TB: 3328 state = ts->state; 3329 ts->state = state | TS_MEM; 3330 if (state != TS_DEAD) { 3331 continue; 3332 } 3333 break; 3334 case TEMP_EBB: 3335 case TEMP_CONST: 3336 continue; 3337 default: 3338 g_assert_not_reached(); 3339 } 3340 la_reset_pref(&s->temps[i]); 3341 } 3342 } 3343 3344 /* liveness analysis: sync globals back to memory and kill. */ 3345 static void la_global_kill(TCGContext *s, int ng) 3346 { 3347 int i; 3348 3349 for (i = 0; i < ng; i++) { 3350 s->temps[i].state = TS_DEAD | TS_MEM; 3351 la_reset_pref(&s->temps[i]); 3352 } 3353 } 3354 3355 /* liveness analysis: note live globals crossing calls. */ 3356 static void la_cross_call(TCGContext *s, int nt) 3357 { 3358 TCGRegSet mask = ~tcg_target_call_clobber_regs; 3359 int i; 3360 3361 for (i = 0; i < nt; i++) { 3362 TCGTemp *ts = &s->temps[i]; 3363 if (!(ts->state & TS_DEAD)) { 3364 TCGRegSet *pset = la_temp_pref(ts); 3365 TCGRegSet set = *pset; 3366 3367 set &= mask; 3368 /* If the combination is not possible, restart. */ 3369 if (set == 0) { 3370 set = tcg_target_available_regs[ts->type] & mask; 3371 } 3372 *pset = set; 3373 } 3374 } 3375 } 3376 3377 /* 3378 * Liveness analysis: Verify the lifetime of TEMP_TB, and reduce 3379 * to TEMP_EBB, if possible. 3380 */ 3381 static void __attribute__((noinline)) 3382 liveness_pass_0(TCGContext *s) 3383 { 3384 void * const multiple_ebb = (void *)(uintptr_t)-1; 3385 int nb_temps = s->nb_temps; 3386 TCGOp *op, *ebb; 3387 3388 for (int i = s->nb_globals; i < nb_temps; ++i) { 3389 s->temps[i].state_ptr = NULL; 3390 } 3391 3392 /* 3393 * Represent each EBB by the op at which it begins. In the case of 3394 * the first EBB, this is the first op, otherwise it is a label. 3395 * Collect the uses of each TEMP_TB: NULL for unused, EBB for use 3396 * within a single EBB, else MULTIPLE_EBB. 3397 */ 3398 ebb = QTAILQ_FIRST(&s->ops); 3399 QTAILQ_FOREACH(op, &s->ops, link) { 3400 const TCGOpDef *def; 3401 int nb_oargs, nb_iargs; 3402 3403 switch (op->opc) { 3404 case INDEX_op_set_label: 3405 ebb = op; 3406 continue; 3407 case INDEX_op_discard: 3408 continue; 3409 case INDEX_op_call: 3410 nb_oargs = TCGOP_CALLO(op); 3411 nb_iargs = TCGOP_CALLI(op); 3412 break; 3413 default: 3414 def = &tcg_op_defs[op->opc]; 3415 nb_oargs = def->nb_oargs; 3416 nb_iargs = def->nb_iargs; 3417 break; 3418 } 3419 3420 for (int i = 0; i < nb_oargs + nb_iargs; ++i) { 3421 TCGTemp *ts = arg_temp(op->args[i]); 3422 3423 if (ts->kind != TEMP_TB) { 3424 continue; 3425 } 3426 if (ts->state_ptr == NULL) { 3427 ts->state_ptr = ebb; 3428 } else if (ts->state_ptr != ebb) { 3429 ts->state_ptr = multiple_ebb; 3430 } 3431 } 3432 } 3433 3434 /* 3435 * For TEMP_TB that turned out not to be used beyond one EBB, 3436 * reduce the liveness to TEMP_EBB. 3437 */ 3438 for (int i = s->nb_globals; i < nb_temps; ++i) { 3439 TCGTemp *ts = &s->temps[i]; 3440 if (ts->kind == TEMP_TB && ts->state_ptr != multiple_ebb) { 3441 ts->kind = TEMP_EBB; 3442 } 3443 } 3444 } 3445 3446 /* Liveness analysis : update the opc_arg_life array to tell if a 3447 given input arguments is dead. Instructions updating dead 3448 temporaries are removed. */ 3449 static void __attribute__((noinline)) 3450 liveness_pass_1(TCGContext *s) 3451 { 3452 int nb_globals = s->nb_globals; 3453 int nb_temps = s->nb_temps; 3454 TCGOp *op, *op_prev; 3455 TCGRegSet *prefs; 3456 int i; 3457 3458 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps); 3459 for (i = 0; i < nb_temps; ++i) { 3460 s->temps[i].state_ptr = prefs + i; 3461 } 3462 3463 /* ??? Should be redundant with the exit_tb that ends the TB. */ 3464 la_func_end(s, nb_globals, nb_temps); 3465 3466 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) { 3467 int nb_iargs, nb_oargs; 3468 TCGOpcode opc_new, opc_new2; 3469 bool have_opc_new2; 3470 TCGLifeData arg_life = 0; 3471 TCGTemp *ts; 3472 TCGOpcode opc = op->opc; 3473 const TCGOpDef *def = &tcg_op_defs[opc]; 3474 3475 switch (opc) { 3476 case INDEX_op_call: 3477 { 3478 const TCGHelperInfo *info = tcg_call_info(op); 3479 int call_flags = tcg_call_flags(op); 3480 3481 nb_oargs = TCGOP_CALLO(op); 3482 nb_iargs = TCGOP_CALLI(op); 3483 3484 /* pure functions can be removed if their result is unused */ 3485 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) { 3486 for (i = 0; i < nb_oargs; i++) { 3487 ts = arg_temp(op->args[i]); 3488 if (ts->state != TS_DEAD) { 3489 goto do_not_remove_call; 3490 } 3491 } 3492 goto do_remove; 3493 } 3494 do_not_remove_call: 3495 3496 /* Output args are dead. */ 3497 for (i = 0; i < nb_oargs; i++) { 3498 ts = arg_temp(op->args[i]); 3499 if (ts->state & TS_DEAD) { 3500 arg_life |= DEAD_ARG << i; 3501 } 3502 if (ts->state & TS_MEM) { 3503 arg_life |= SYNC_ARG << i; 3504 } 3505 ts->state = TS_DEAD; 3506 la_reset_pref(ts); 3507 } 3508 3509 /* Not used -- it will be tcg_target_call_oarg_reg(). */ 3510 memset(op->output_pref, 0, sizeof(op->output_pref)); 3511 3512 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS | 3513 TCG_CALL_NO_READ_GLOBALS))) { 3514 la_global_kill(s, nb_globals); 3515 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) { 3516 la_global_sync(s, nb_globals); 3517 } 3518 3519 /* Record arguments that die in this helper. */ 3520 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3521 ts = arg_temp(op->args[i]); 3522 if (ts->state & TS_DEAD) { 3523 arg_life |= DEAD_ARG << i; 3524 } 3525 } 3526 3527 /* For all live registers, remove call-clobbered prefs. */ 3528 la_cross_call(s, nb_temps); 3529 3530 /* 3531 * Input arguments are live for preceding opcodes. 3532 * 3533 * For those arguments that die, and will be allocated in 3534 * registers, clear the register set for that arg, to be 3535 * filled in below. For args that will be on the stack, 3536 * reset to any available reg. Process arguments in reverse 3537 * order so that if a temp is used more than once, the stack 3538 * reset to max happens before the register reset to 0. 3539 */ 3540 for (i = nb_iargs - 1; i >= 0; i--) { 3541 const TCGCallArgumentLoc *loc = &info->in[i]; 3542 ts = arg_temp(op->args[nb_oargs + i]); 3543 3544 if (ts->state & TS_DEAD) { 3545 switch (loc->kind) { 3546 case TCG_CALL_ARG_NORMAL: 3547 case TCG_CALL_ARG_EXTEND_U: 3548 case TCG_CALL_ARG_EXTEND_S: 3549 if (arg_slot_reg_p(loc->arg_slot)) { 3550 *la_temp_pref(ts) = 0; 3551 break; 3552 } 3553 /* fall through */ 3554 default: 3555 *la_temp_pref(ts) = 3556 tcg_target_available_regs[ts->type]; 3557 break; 3558 } 3559 ts->state &= ~TS_DEAD; 3560 } 3561 } 3562 3563 /* 3564 * For each input argument, add its input register to prefs. 3565 * If a temp is used once, this produces a single set bit; 3566 * if a temp is used multiple times, this produces a set. 3567 */ 3568 for (i = 0; i < nb_iargs; i++) { 3569 const TCGCallArgumentLoc *loc = &info->in[i]; 3570 ts = arg_temp(op->args[nb_oargs + i]); 3571 3572 switch (loc->kind) { 3573 case TCG_CALL_ARG_NORMAL: 3574 case TCG_CALL_ARG_EXTEND_U: 3575 case TCG_CALL_ARG_EXTEND_S: 3576 if (arg_slot_reg_p(loc->arg_slot)) { 3577 tcg_regset_set_reg(*la_temp_pref(ts), 3578 tcg_target_call_iarg_regs[loc->arg_slot]); 3579 } 3580 break; 3581 default: 3582 break; 3583 } 3584 } 3585 } 3586 break; 3587 case INDEX_op_insn_start: 3588 break; 3589 case INDEX_op_discard: 3590 /* mark the temporary as dead */ 3591 ts = arg_temp(op->args[0]); 3592 ts->state = TS_DEAD; 3593 la_reset_pref(ts); 3594 break; 3595 3596 case INDEX_op_add2_i32: 3597 opc_new = INDEX_op_add_i32; 3598 goto do_addsub2; 3599 case INDEX_op_sub2_i32: 3600 opc_new = INDEX_op_sub_i32; 3601 goto do_addsub2; 3602 case INDEX_op_add2_i64: 3603 opc_new = INDEX_op_add_i64; 3604 goto do_addsub2; 3605 case INDEX_op_sub2_i64: 3606 opc_new = INDEX_op_sub_i64; 3607 do_addsub2: 3608 nb_iargs = 4; 3609 nb_oargs = 2; 3610 /* Test if the high part of the operation is dead, but not 3611 the low part. The result can be optimized to a simple 3612 add or sub. This happens often for x86_64 guest when the 3613 cpu mode is set to 32 bit. */ 3614 if (arg_temp(op->args[1])->state == TS_DEAD) { 3615 if (arg_temp(op->args[0])->state == TS_DEAD) { 3616 goto do_remove; 3617 } 3618 /* Replace the opcode and adjust the args in place, 3619 leaving 3 unused args at the end. */ 3620 op->opc = opc = opc_new; 3621 op->args[1] = op->args[2]; 3622 op->args[2] = op->args[4]; 3623 /* Fall through and mark the single-word operation live. */ 3624 nb_iargs = 2; 3625 nb_oargs = 1; 3626 } 3627 goto do_not_remove; 3628 3629 case INDEX_op_mulu2_i32: 3630 opc_new = INDEX_op_mul_i32; 3631 opc_new2 = INDEX_op_muluh_i32; 3632 have_opc_new2 = TCG_TARGET_HAS_muluh_i32; 3633 goto do_mul2; 3634 case INDEX_op_muls2_i32: 3635 opc_new = INDEX_op_mul_i32; 3636 opc_new2 = INDEX_op_mulsh_i32; 3637 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32; 3638 goto do_mul2; 3639 case INDEX_op_mulu2_i64: 3640 opc_new = INDEX_op_mul_i64; 3641 opc_new2 = INDEX_op_muluh_i64; 3642 have_opc_new2 = TCG_TARGET_HAS_muluh_i64; 3643 goto do_mul2; 3644 case INDEX_op_muls2_i64: 3645 opc_new = INDEX_op_mul_i64; 3646 opc_new2 = INDEX_op_mulsh_i64; 3647 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64; 3648 goto do_mul2; 3649 do_mul2: 3650 nb_iargs = 2; 3651 nb_oargs = 2; 3652 if (arg_temp(op->args[1])->state == TS_DEAD) { 3653 if (arg_temp(op->args[0])->state == TS_DEAD) { 3654 /* Both parts of the operation are dead. */ 3655 goto do_remove; 3656 } 3657 /* The high part of the operation is dead; generate the low. */ 3658 op->opc = opc = opc_new; 3659 op->args[1] = op->args[2]; 3660 op->args[2] = op->args[3]; 3661 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) { 3662 /* The low part of the operation is dead; generate the high. */ 3663 op->opc = opc = opc_new2; 3664 op->args[0] = op->args[1]; 3665 op->args[1] = op->args[2]; 3666 op->args[2] = op->args[3]; 3667 } else { 3668 goto do_not_remove; 3669 } 3670 /* Mark the single-word operation live. */ 3671 nb_oargs = 1; 3672 goto do_not_remove; 3673 3674 default: 3675 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */ 3676 nb_iargs = def->nb_iargs; 3677 nb_oargs = def->nb_oargs; 3678 3679 /* Test if the operation can be removed because all 3680 its outputs are dead. We assume that nb_oargs == 0 3681 implies side effects */ 3682 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { 3683 for (i = 0; i < nb_oargs; i++) { 3684 if (arg_temp(op->args[i])->state != TS_DEAD) { 3685 goto do_not_remove; 3686 } 3687 } 3688 goto do_remove; 3689 } 3690 goto do_not_remove; 3691 3692 do_remove: 3693 tcg_op_remove(s, op); 3694 break; 3695 3696 do_not_remove: 3697 for (i = 0; i < nb_oargs; i++) { 3698 ts = arg_temp(op->args[i]); 3699 3700 /* Remember the preference of the uses that followed. */ 3701 if (i < ARRAY_SIZE(op->output_pref)) { 3702 op->output_pref[i] = *la_temp_pref(ts); 3703 } 3704 3705 /* Output args are dead. */ 3706 if (ts->state & TS_DEAD) { 3707 arg_life |= DEAD_ARG << i; 3708 } 3709 if (ts->state & TS_MEM) { 3710 arg_life |= SYNC_ARG << i; 3711 } 3712 ts->state = TS_DEAD; 3713 la_reset_pref(ts); 3714 } 3715 3716 /* If end of basic block, update. */ 3717 if (def->flags & TCG_OPF_BB_EXIT) { 3718 la_func_end(s, nb_globals, nb_temps); 3719 } else if (def->flags & TCG_OPF_COND_BRANCH) { 3720 la_bb_sync(s, nb_globals, nb_temps); 3721 } else if (def->flags & TCG_OPF_BB_END) { 3722 la_bb_end(s, nb_globals, nb_temps); 3723 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3724 la_global_sync(s, nb_globals); 3725 if (def->flags & TCG_OPF_CALL_CLOBBER) { 3726 la_cross_call(s, nb_temps); 3727 } 3728 } 3729 3730 /* Record arguments that die in this opcode. */ 3731 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3732 ts = arg_temp(op->args[i]); 3733 if (ts->state & TS_DEAD) { 3734 arg_life |= DEAD_ARG << i; 3735 } 3736 } 3737 3738 /* Input arguments are live for preceding opcodes. */ 3739 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3740 ts = arg_temp(op->args[i]); 3741 if (ts->state & TS_DEAD) { 3742 /* For operands that were dead, initially allow 3743 all regs for the type. */ 3744 *la_temp_pref(ts) = tcg_target_available_regs[ts->type]; 3745 ts->state &= ~TS_DEAD; 3746 } 3747 } 3748 3749 /* Incorporate constraints for this operand. */ 3750 switch (opc) { 3751 case INDEX_op_mov_i32: 3752 case INDEX_op_mov_i64: 3753 /* Note that these are TCG_OPF_NOT_PRESENT and do not 3754 have proper constraints. That said, special case 3755 moves to propagate preferences backward. */ 3756 if (IS_DEAD_ARG(1)) { 3757 *la_temp_pref(arg_temp(op->args[0])) 3758 = *la_temp_pref(arg_temp(op->args[1])); 3759 } 3760 break; 3761 3762 default: 3763 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3764 const TCGArgConstraint *ct = &def->args_ct[i]; 3765 TCGRegSet set, *pset; 3766 3767 ts = arg_temp(op->args[i]); 3768 pset = la_temp_pref(ts); 3769 set = *pset; 3770 3771 set &= ct->regs; 3772 if (ct->ialias) { 3773 set &= output_pref(op, ct->alias_index); 3774 } 3775 /* If the combination is not possible, restart. */ 3776 if (set == 0) { 3777 set = ct->regs; 3778 } 3779 *pset = set; 3780 } 3781 break; 3782 } 3783 break; 3784 } 3785 op->life = arg_life; 3786 } 3787 } 3788 3789 /* Liveness analysis: Convert indirect regs to direct temporaries. */ 3790 static bool __attribute__((noinline)) 3791 liveness_pass_2(TCGContext *s) 3792 { 3793 int nb_globals = s->nb_globals; 3794 int nb_temps, i; 3795 bool changes = false; 3796 TCGOp *op, *op_next; 3797 3798 /* Create a temporary for each indirect global. */ 3799 for (i = 0; i < nb_globals; ++i) { 3800 TCGTemp *its = &s->temps[i]; 3801 if (its->indirect_reg) { 3802 TCGTemp *dts = tcg_temp_alloc(s); 3803 dts->type = its->type; 3804 dts->base_type = its->base_type; 3805 dts->temp_subindex = its->temp_subindex; 3806 dts->kind = TEMP_EBB; 3807 its->state_ptr = dts; 3808 } else { 3809 its->state_ptr = NULL; 3810 } 3811 /* All globals begin dead. */ 3812 its->state = TS_DEAD; 3813 } 3814 for (nb_temps = s->nb_temps; i < nb_temps; ++i) { 3815 TCGTemp *its = &s->temps[i]; 3816 its->state_ptr = NULL; 3817 its->state = TS_DEAD; 3818 } 3819 3820 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3821 TCGOpcode opc = op->opc; 3822 const TCGOpDef *def = &tcg_op_defs[opc]; 3823 TCGLifeData arg_life = op->life; 3824 int nb_iargs, nb_oargs, call_flags; 3825 TCGTemp *arg_ts, *dir_ts; 3826 3827 if (opc == INDEX_op_call) { 3828 nb_oargs = TCGOP_CALLO(op); 3829 nb_iargs = TCGOP_CALLI(op); 3830 call_flags = tcg_call_flags(op); 3831 } else { 3832 nb_iargs = def->nb_iargs; 3833 nb_oargs = def->nb_oargs; 3834 3835 /* Set flags similar to how calls require. */ 3836 if (def->flags & TCG_OPF_COND_BRANCH) { 3837 /* Like reading globals: sync_globals */ 3838 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 3839 } else if (def->flags & TCG_OPF_BB_END) { 3840 /* Like writing globals: save_globals */ 3841 call_flags = 0; 3842 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3843 /* Like reading globals: sync_globals */ 3844 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 3845 } else { 3846 /* No effect on globals. */ 3847 call_flags = (TCG_CALL_NO_READ_GLOBALS | 3848 TCG_CALL_NO_WRITE_GLOBALS); 3849 } 3850 } 3851 3852 /* Make sure that input arguments are available. */ 3853 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3854 arg_ts = arg_temp(op->args[i]); 3855 dir_ts = arg_ts->state_ptr; 3856 if (dir_ts && arg_ts->state == TS_DEAD) { 3857 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32 3858 ? INDEX_op_ld_i32 3859 : INDEX_op_ld_i64); 3860 TCGOp *lop = tcg_op_insert_before(s, op, lopc, 3); 3861 3862 lop->args[0] = temp_arg(dir_ts); 3863 lop->args[1] = temp_arg(arg_ts->mem_base); 3864 lop->args[2] = arg_ts->mem_offset; 3865 3866 /* Loaded, but synced with memory. */ 3867 arg_ts->state = TS_MEM; 3868 } 3869 } 3870 3871 /* Perform input replacement, and mark inputs that became dead. 3872 No action is required except keeping temp_state up to date 3873 so that we reload when needed. */ 3874 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3875 arg_ts = arg_temp(op->args[i]); 3876 dir_ts = arg_ts->state_ptr; 3877 if (dir_ts) { 3878 op->args[i] = temp_arg(dir_ts); 3879 changes = true; 3880 if (IS_DEAD_ARG(i)) { 3881 arg_ts->state = TS_DEAD; 3882 } 3883 } 3884 } 3885 3886 /* Liveness analysis should ensure that the following are 3887 all correct, for call sites and basic block end points. */ 3888 if (call_flags & TCG_CALL_NO_READ_GLOBALS) { 3889 /* Nothing to do */ 3890 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) { 3891 for (i = 0; i < nb_globals; ++i) { 3892 /* Liveness should see that globals are synced back, 3893 that is, either TS_DEAD or TS_MEM. */ 3894 arg_ts = &s->temps[i]; 3895 tcg_debug_assert(arg_ts->state_ptr == 0 3896 || arg_ts->state != 0); 3897 } 3898 } else { 3899 for (i = 0; i < nb_globals; ++i) { 3900 /* Liveness should see that globals are saved back, 3901 that is, TS_DEAD, waiting to be reloaded. */ 3902 arg_ts = &s->temps[i]; 3903 tcg_debug_assert(arg_ts->state_ptr == 0 3904 || arg_ts->state == TS_DEAD); 3905 } 3906 } 3907 3908 /* Outputs become available. */ 3909 if (opc == INDEX_op_mov_i32 || opc == INDEX_op_mov_i64) { 3910 arg_ts = arg_temp(op->args[0]); 3911 dir_ts = arg_ts->state_ptr; 3912 if (dir_ts) { 3913 op->args[0] = temp_arg(dir_ts); 3914 changes = true; 3915 3916 /* The output is now live and modified. */ 3917 arg_ts->state = 0; 3918 3919 if (NEED_SYNC_ARG(0)) { 3920 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 3921 ? INDEX_op_st_i32 3922 : INDEX_op_st_i64); 3923 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3); 3924 TCGTemp *out_ts = dir_ts; 3925 3926 if (IS_DEAD_ARG(0)) { 3927 out_ts = arg_temp(op->args[1]); 3928 arg_ts->state = TS_DEAD; 3929 tcg_op_remove(s, op); 3930 } else { 3931 arg_ts->state = TS_MEM; 3932 } 3933 3934 sop->args[0] = temp_arg(out_ts); 3935 sop->args[1] = temp_arg(arg_ts->mem_base); 3936 sop->args[2] = arg_ts->mem_offset; 3937 } else { 3938 tcg_debug_assert(!IS_DEAD_ARG(0)); 3939 } 3940 } 3941 } else { 3942 for (i = 0; i < nb_oargs; i++) { 3943 arg_ts = arg_temp(op->args[i]); 3944 dir_ts = arg_ts->state_ptr; 3945 if (!dir_ts) { 3946 continue; 3947 } 3948 op->args[i] = temp_arg(dir_ts); 3949 changes = true; 3950 3951 /* The output is now live and modified. */ 3952 arg_ts->state = 0; 3953 3954 /* Sync outputs upon their last write. */ 3955 if (NEED_SYNC_ARG(i)) { 3956 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 3957 ? INDEX_op_st_i32 3958 : INDEX_op_st_i64); 3959 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3); 3960 3961 sop->args[0] = temp_arg(dir_ts); 3962 sop->args[1] = temp_arg(arg_ts->mem_base); 3963 sop->args[2] = arg_ts->mem_offset; 3964 3965 arg_ts->state = TS_MEM; 3966 } 3967 /* Drop outputs that are dead. */ 3968 if (IS_DEAD_ARG(i)) { 3969 arg_ts->state = TS_DEAD; 3970 } 3971 } 3972 } 3973 } 3974 3975 return changes; 3976 } 3977 3978 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts) 3979 { 3980 intptr_t off; 3981 int size, align; 3982 3983 /* When allocating an object, look at the full type. */ 3984 size = tcg_type_size(ts->base_type); 3985 switch (ts->base_type) { 3986 case TCG_TYPE_I32: 3987 align = 4; 3988 break; 3989 case TCG_TYPE_I64: 3990 case TCG_TYPE_V64: 3991 align = 8; 3992 break; 3993 case TCG_TYPE_I128: 3994 case TCG_TYPE_V128: 3995 case TCG_TYPE_V256: 3996 /* 3997 * Note that we do not require aligned storage for V256, 3998 * and that we provide alignment for I128 to match V128, 3999 * even if that's above what the host ABI requires. 4000 */ 4001 align = 16; 4002 break; 4003 default: 4004 g_assert_not_reached(); 4005 } 4006 4007 /* 4008 * Assume the stack is sufficiently aligned. 4009 * This affects e.g. ARM NEON, where we have 8 byte stack alignment 4010 * and do not require 16 byte vector alignment. This seems slightly 4011 * easier than fully parameterizing the above switch statement. 4012 */ 4013 align = MIN(TCG_TARGET_STACK_ALIGN, align); 4014 off = ROUND_UP(s->current_frame_offset, align); 4015 4016 /* If we've exhausted the stack frame, restart with a smaller TB. */ 4017 if (off + size > s->frame_end) { 4018 tcg_raise_tb_overflow(s); 4019 } 4020 s->current_frame_offset = off + size; 4021 #if defined(__sparc__) 4022 off += TCG_TARGET_STACK_BIAS; 4023 #endif 4024 4025 /* If the object was subdivided, assign memory to all the parts. */ 4026 if (ts->base_type != ts->type) { 4027 int part_size = tcg_type_size(ts->type); 4028 int part_count = size / part_size; 4029 4030 /* 4031 * Each part is allocated sequentially in tcg_temp_new_internal. 4032 * Jump back to the first part by subtracting the current index. 4033 */ 4034 ts -= ts->temp_subindex; 4035 for (int i = 0; i < part_count; ++i) { 4036 ts[i].mem_offset = off + i * part_size; 4037 ts[i].mem_base = s->frame_temp; 4038 ts[i].mem_allocated = 1; 4039 } 4040 } else { 4041 ts->mem_offset = off; 4042 ts->mem_base = s->frame_temp; 4043 ts->mem_allocated = 1; 4044 } 4045 } 4046 4047 /* Assign @reg to @ts, and update reg_to_temp[]. */ 4048 static void set_temp_val_reg(TCGContext *s, TCGTemp *ts, TCGReg reg) 4049 { 4050 if (ts->val_type == TEMP_VAL_REG) { 4051 TCGReg old = ts->reg; 4052 tcg_debug_assert(s->reg_to_temp[old] == ts); 4053 if (old == reg) { 4054 return; 4055 } 4056 s->reg_to_temp[old] = NULL; 4057 } 4058 tcg_debug_assert(s->reg_to_temp[reg] == NULL); 4059 s->reg_to_temp[reg] = ts; 4060 ts->val_type = TEMP_VAL_REG; 4061 ts->reg = reg; 4062 } 4063 4064 /* Assign a non-register value type to @ts, and update reg_to_temp[]. */ 4065 static void set_temp_val_nonreg(TCGContext *s, TCGTemp *ts, TCGTempVal type) 4066 { 4067 tcg_debug_assert(type != TEMP_VAL_REG); 4068 if (ts->val_type == TEMP_VAL_REG) { 4069 TCGReg reg = ts->reg; 4070 tcg_debug_assert(s->reg_to_temp[reg] == ts); 4071 s->reg_to_temp[reg] = NULL; 4072 } 4073 ts->val_type = type; 4074 } 4075 4076 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet); 4077 4078 /* Mark a temporary as free or dead. If 'free_or_dead' is negative, 4079 mark it free; otherwise mark it dead. */ 4080 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead) 4081 { 4082 TCGTempVal new_type; 4083 4084 switch (ts->kind) { 4085 case TEMP_FIXED: 4086 return; 4087 case TEMP_GLOBAL: 4088 case TEMP_TB: 4089 new_type = TEMP_VAL_MEM; 4090 break; 4091 case TEMP_EBB: 4092 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD; 4093 break; 4094 case TEMP_CONST: 4095 new_type = TEMP_VAL_CONST; 4096 break; 4097 default: 4098 g_assert_not_reached(); 4099 } 4100 set_temp_val_nonreg(s, ts, new_type); 4101 } 4102 4103 /* Mark a temporary as dead. */ 4104 static inline void temp_dead(TCGContext *s, TCGTemp *ts) 4105 { 4106 temp_free_or_dead(s, ts, 1); 4107 } 4108 4109 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary 4110 registers needs to be allocated to store a constant. If 'free_or_dead' 4111 is non-zero, subsequently release the temporary; if it is positive, the 4112 temp is dead; if it is negative, the temp is free. */ 4113 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs, 4114 TCGRegSet preferred_regs, int free_or_dead) 4115 { 4116 if (!temp_readonly(ts) && !ts->mem_coherent) { 4117 if (!ts->mem_allocated) { 4118 temp_allocate_frame(s, ts); 4119 } 4120 switch (ts->val_type) { 4121 case TEMP_VAL_CONST: 4122 /* If we're going to free the temp immediately, then we won't 4123 require it later in a register, so attempt to store the 4124 constant to memory directly. */ 4125 if (free_or_dead 4126 && tcg_out_sti(s, ts->type, ts->val, 4127 ts->mem_base->reg, ts->mem_offset)) { 4128 break; 4129 } 4130 temp_load(s, ts, tcg_target_available_regs[ts->type], 4131 allocated_regs, preferred_regs); 4132 /* fallthrough */ 4133 4134 case TEMP_VAL_REG: 4135 tcg_out_st(s, ts->type, ts->reg, 4136 ts->mem_base->reg, ts->mem_offset); 4137 break; 4138 4139 case TEMP_VAL_MEM: 4140 break; 4141 4142 case TEMP_VAL_DEAD: 4143 default: 4144 g_assert_not_reached(); 4145 } 4146 ts->mem_coherent = 1; 4147 } 4148 if (free_or_dead) { 4149 temp_free_or_dead(s, ts, free_or_dead); 4150 } 4151 } 4152 4153 /* free register 'reg' by spilling the corresponding temporary if necessary */ 4154 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs) 4155 { 4156 TCGTemp *ts = s->reg_to_temp[reg]; 4157 if (ts != NULL) { 4158 temp_sync(s, ts, allocated_regs, 0, -1); 4159 } 4160 } 4161 4162 /** 4163 * tcg_reg_alloc: 4164 * @required_regs: Set of registers in which we must allocate. 4165 * @allocated_regs: Set of registers which must be avoided. 4166 * @preferred_regs: Set of registers we should prefer. 4167 * @rev: True if we search the registers in "indirect" order. 4168 * 4169 * The allocated register must be in @required_regs & ~@allocated_regs, 4170 * but if we can put it in @preferred_regs we may save a move later. 4171 */ 4172 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs, 4173 TCGRegSet allocated_regs, 4174 TCGRegSet preferred_regs, bool rev) 4175 { 4176 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4177 TCGRegSet reg_ct[2]; 4178 const int *order; 4179 4180 reg_ct[1] = required_regs & ~allocated_regs; 4181 tcg_debug_assert(reg_ct[1] != 0); 4182 reg_ct[0] = reg_ct[1] & preferred_regs; 4183 4184 /* Skip the preferred_regs option if it cannot be satisfied, 4185 or if the preference made no difference. */ 4186 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4187 4188 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4189 4190 /* Try free registers, preferences first. */ 4191 for (j = f; j < 2; j++) { 4192 TCGRegSet set = reg_ct[j]; 4193 4194 if (tcg_regset_single(set)) { 4195 /* One register in the set. */ 4196 TCGReg reg = tcg_regset_first(set); 4197 if (s->reg_to_temp[reg] == NULL) { 4198 return reg; 4199 } 4200 } else { 4201 for (i = 0; i < n; i++) { 4202 TCGReg reg = order[i]; 4203 if (s->reg_to_temp[reg] == NULL && 4204 tcg_regset_test_reg(set, reg)) { 4205 return reg; 4206 } 4207 } 4208 } 4209 } 4210 4211 /* We must spill something. */ 4212 for (j = f; j < 2; j++) { 4213 TCGRegSet set = reg_ct[j]; 4214 4215 if (tcg_regset_single(set)) { 4216 /* One register in the set. */ 4217 TCGReg reg = tcg_regset_first(set); 4218 tcg_reg_free(s, reg, allocated_regs); 4219 return reg; 4220 } else { 4221 for (i = 0; i < n; i++) { 4222 TCGReg reg = order[i]; 4223 if (tcg_regset_test_reg(set, reg)) { 4224 tcg_reg_free(s, reg, allocated_regs); 4225 return reg; 4226 } 4227 } 4228 } 4229 } 4230 4231 g_assert_not_reached(); 4232 } 4233 4234 static TCGReg tcg_reg_alloc_pair(TCGContext *s, TCGRegSet required_regs, 4235 TCGRegSet allocated_regs, 4236 TCGRegSet preferred_regs, bool rev) 4237 { 4238 int i, j, k, fmin, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4239 TCGRegSet reg_ct[2]; 4240 const int *order; 4241 4242 /* Ensure that if I is not in allocated_regs, I+1 is not either. */ 4243 reg_ct[1] = required_regs & ~(allocated_regs | (allocated_regs >> 1)); 4244 tcg_debug_assert(reg_ct[1] != 0); 4245 reg_ct[0] = reg_ct[1] & preferred_regs; 4246 4247 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4248 4249 /* 4250 * Skip the preferred_regs option if it cannot be satisfied, 4251 * or if the preference made no difference. 4252 */ 4253 k = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4254 4255 /* 4256 * Minimize the number of flushes by looking for 2 free registers first, 4257 * then a single flush, then two flushes. 4258 */ 4259 for (fmin = 2; fmin >= 0; fmin--) { 4260 for (j = k; j < 2; j++) { 4261 TCGRegSet set = reg_ct[j]; 4262 4263 for (i = 0; i < n; i++) { 4264 TCGReg reg = order[i]; 4265 4266 if (tcg_regset_test_reg(set, reg)) { 4267 int f = !s->reg_to_temp[reg] + !s->reg_to_temp[reg + 1]; 4268 if (f >= fmin) { 4269 tcg_reg_free(s, reg, allocated_regs); 4270 tcg_reg_free(s, reg + 1, allocated_regs); 4271 return reg; 4272 } 4273 } 4274 } 4275 } 4276 } 4277 g_assert_not_reached(); 4278 } 4279 4280 /* Make sure the temporary is in a register. If needed, allocate the register 4281 from DESIRED while avoiding ALLOCATED. */ 4282 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs, 4283 TCGRegSet allocated_regs, TCGRegSet preferred_regs) 4284 { 4285 TCGReg reg; 4286 4287 switch (ts->val_type) { 4288 case TEMP_VAL_REG: 4289 return; 4290 case TEMP_VAL_CONST: 4291 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4292 preferred_regs, ts->indirect_base); 4293 if (ts->type <= TCG_TYPE_I64) { 4294 tcg_out_movi(s, ts->type, reg, ts->val); 4295 } else { 4296 uint64_t val = ts->val; 4297 MemOp vece = MO_64; 4298 4299 /* 4300 * Find the minimal vector element that matches the constant. 4301 * The targets will, in general, have to do this search anyway, 4302 * do this generically. 4303 */ 4304 if (val == dup_const(MO_8, val)) { 4305 vece = MO_8; 4306 } else if (val == dup_const(MO_16, val)) { 4307 vece = MO_16; 4308 } else if (val == dup_const(MO_32, val)) { 4309 vece = MO_32; 4310 } 4311 4312 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val); 4313 } 4314 ts->mem_coherent = 0; 4315 break; 4316 case TEMP_VAL_MEM: 4317 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4318 preferred_regs, ts->indirect_base); 4319 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset); 4320 ts->mem_coherent = 1; 4321 break; 4322 case TEMP_VAL_DEAD: 4323 default: 4324 g_assert_not_reached(); 4325 } 4326 set_temp_val_reg(s, ts, reg); 4327 } 4328 4329 /* Save a temporary to memory. 'allocated_regs' is used in case a 4330 temporary registers needs to be allocated to store a constant. */ 4331 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs) 4332 { 4333 /* The liveness analysis already ensures that globals are back 4334 in memory. Keep an tcg_debug_assert for safety. */ 4335 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts)); 4336 } 4337 4338 /* save globals to their canonical location and assume they can be 4339 modified be the following code. 'allocated_regs' is used in case a 4340 temporary registers needs to be allocated to store a constant. */ 4341 static void save_globals(TCGContext *s, TCGRegSet allocated_regs) 4342 { 4343 int i, n; 4344 4345 for (i = 0, n = s->nb_globals; i < n; i++) { 4346 temp_save(s, &s->temps[i], allocated_regs); 4347 } 4348 } 4349 4350 /* sync globals to their canonical location and assume they can be 4351 read by the following code. 'allocated_regs' is used in case a 4352 temporary registers needs to be allocated to store a constant. */ 4353 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs) 4354 { 4355 int i, n; 4356 4357 for (i = 0, n = s->nb_globals; i < n; i++) { 4358 TCGTemp *ts = &s->temps[i]; 4359 tcg_debug_assert(ts->val_type != TEMP_VAL_REG 4360 || ts->kind == TEMP_FIXED 4361 || ts->mem_coherent); 4362 } 4363 } 4364 4365 /* at the end of a basic block, we assume all temporaries are dead and 4366 all globals are stored at their canonical location. */ 4367 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) 4368 { 4369 int i; 4370 4371 for (i = s->nb_globals; i < s->nb_temps; i++) { 4372 TCGTemp *ts = &s->temps[i]; 4373 4374 switch (ts->kind) { 4375 case TEMP_TB: 4376 temp_save(s, ts, allocated_regs); 4377 break; 4378 case TEMP_EBB: 4379 /* The liveness analysis already ensures that temps are dead. 4380 Keep an tcg_debug_assert for safety. */ 4381 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD); 4382 break; 4383 case TEMP_CONST: 4384 /* Similarly, we should have freed any allocated register. */ 4385 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST); 4386 break; 4387 default: 4388 g_assert_not_reached(); 4389 } 4390 } 4391 4392 save_globals(s, allocated_regs); 4393 } 4394 4395 /* 4396 * At a conditional branch, we assume all temporaries are dead unless 4397 * explicitly live-across-conditional-branch; all globals and local 4398 * temps are synced to their location. 4399 */ 4400 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs) 4401 { 4402 sync_globals(s, allocated_regs); 4403 4404 for (int i = s->nb_globals; i < s->nb_temps; i++) { 4405 TCGTemp *ts = &s->temps[i]; 4406 /* 4407 * The liveness analysis already ensures that temps are dead. 4408 * Keep tcg_debug_asserts for safety. 4409 */ 4410 switch (ts->kind) { 4411 case TEMP_TB: 4412 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent); 4413 break; 4414 case TEMP_EBB: 4415 case TEMP_CONST: 4416 break; 4417 default: 4418 g_assert_not_reached(); 4419 } 4420 } 4421 } 4422 4423 /* 4424 * Specialized code generation for INDEX_op_mov_* with a constant. 4425 */ 4426 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots, 4427 tcg_target_ulong val, TCGLifeData arg_life, 4428 TCGRegSet preferred_regs) 4429 { 4430 /* ENV should not be modified. */ 4431 tcg_debug_assert(!temp_readonly(ots)); 4432 4433 /* The movi is not explicitly generated here. */ 4434 set_temp_val_nonreg(s, ots, TEMP_VAL_CONST); 4435 ots->val = val; 4436 ots->mem_coherent = 0; 4437 if (NEED_SYNC_ARG(0)) { 4438 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0)); 4439 } else if (IS_DEAD_ARG(0)) { 4440 temp_dead(s, ots); 4441 } 4442 } 4443 4444 /* 4445 * Specialized code generation for INDEX_op_mov_*. 4446 */ 4447 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op) 4448 { 4449 const TCGLifeData arg_life = op->life; 4450 TCGRegSet allocated_regs, preferred_regs; 4451 TCGTemp *ts, *ots; 4452 TCGType otype, itype; 4453 TCGReg oreg, ireg; 4454 4455 allocated_regs = s->reserved_regs; 4456 preferred_regs = output_pref(op, 0); 4457 ots = arg_temp(op->args[0]); 4458 ts = arg_temp(op->args[1]); 4459 4460 /* ENV should not be modified. */ 4461 tcg_debug_assert(!temp_readonly(ots)); 4462 4463 /* Note that otype != itype for no-op truncation. */ 4464 otype = ots->type; 4465 itype = ts->type; 4466 4467 if (ts->val_type == TEMP_VAL_CONST) { 4468 /* propagate constant or generate sti */ 4469 tcg_target_ulong val = ts->val; 4470 if (IS_DEAD_ARG(1)) { 4471 temp_dead(s, ts); 4472 } 4473 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs); 4474 return; 4475 } 4476 4477 /* If the source value is in memory we're going to be forced 4478 to have it in a register in order to perform the copy. Copy 4479 the SOURCE value into its own register first, that way we 4480 don't have to reload SOURCE the next time it is used. */ 4481 if (ts->val_type == TEMP_VAL_MEM) { 4482 temp_load(s, ts, tcg_target_available_regs[itype], 4483 allocated_regs, preferred_regs); 4484 } 4485 tcg_debug_assert(ts->val_type == TEMP_VAL_REG); 4486 ireg = ts->reg; 4487 4488 if (IS_DEAD_ARG(0)) { 4489 /* mov to a non-saved dead register makes no sense (even with 4490 liveness analysis disabled). */ 4491 tcg_debug_assert(NEED_SYNC_ARG(0)); 4492 if (!ots->mem_allocated) { 4493 temp_allocate_frame(s, ots); 4494 } 4495 tcg_out_st(s, otype, ireg, ots->mem_base->reg, ots->mem_offset); 4496 if (IS_DEAD_ARG(1)) { 4497 temp_dead(s, ts); 4498 } 4499 temp_dead(s, ots); 4500 return; 4501 } 4502 4503 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) { 4504 /* 4505 * The mov can be suppressed. Kill input first, so that it 4506 * is unlinked from reg_to_temp, then set the output to the 4507 * reg that we saved from the input. 4508 */ 4509 temp_dead(s, ts); 4510 oreg = ireg; 4511 } else { 4512 if (ots->val_type == TEMP_VAL_REG) { 4513 oreg = ots->reg; 4514 } else { 4515 /* Make sure to not spill the input register during allocation. */ 4516 oreg = tcg_reg_alloc(s, tcg_target_available_regs[otype], 4517 allocated_regs | ((TCGRegSet)1 << ireg), 4518 preferred_regs, ots->indirect_base); 4519 } 4520 if (!tcg_out_mov(s, otype, oreg, ireg)) { 4521 /* 4522 * Cross register class move not supported. 4523 * Store the source register into the destination slot 4524 * and leave the destination temp as TEMP_VAL_MEM. 4525 */ 4526 assert(!temp_readonly(ots)); 4527 if (!ts->mem_allocated) { 4528 temp_allocate_frame(s, ots); 4529 } 4530 tcg_out_st(s, ts->type, ireg, ots->mem_base->reg, ots->mem_offset); 4531 set_temp_val_nonreg(s, ts, TEMP_VAL_MEM); 4532 ots->mem_coherent = 1; 4533 return; 4534 } 4535 } 4536 set_temp_val_reg(s, ots, oreg); 4537 ots->mem_coherent = 0; 4538 4539 if (NEED_SYNC_ARG(0)) { 4540 temp_sync(s, ots, allocated_regs, 0, 0); 4541 } 4542 } 4543 4544 /* 4545 * Specialized code generation for INDEX_op_dup_vec. 4546 */ 4547 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op) 4548 { 4549 const TCGLifeData arg_life = op->life; 4550 TCGRegSet dup_out_regs, dup_in_regs; 4551 TCGTemp *its, *ots; 4552 TCGType itype, vtype; 4553 unsigned vece; 4554 int lowpart_ofs; 4555 bool ok; 4556 4557 ots = arg_temp(op->args[0]); 4558 its = arg_temp(op->args[1]); 4559 4560 /* ENV should not be modified. */ 4561 tcg_debug_assert(!temp_readonly(ots)); 4562 4563 itype = its->type; 4564 vece = TCGOP_VECE(op); 4565 vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 4566 4567 if (its->val_type == TEMP_VAL_CONST) { 4568 /* Propagate constant via movi -> dupi. */ 4569 tcg_target_ulong val = its->val; 4570 if (IS_DEAD_ARG(1)) { 4571 temp_dead(s, its); 4572 } 4573 tcg_reg_alloc_do_movi(s, ots, val, arg_life, output_pref(op, 0)); 4574 return; 4575 } 4576 4577 dup_out_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 4578 dup_in_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[1].regs; 4579 4580 /* Allocate the output register now. */ 4581 if (ots->val_type != TEMP_VAL_REG) { 4582 TCGRegSet allocated_regs = s->reserved_regs; 4583 TCGReg oreg; 4584 4585 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) { 4586 /* Make sure to not spill the input register. */ 4587 tcg_regset_set_reg(allocated_regs, its->reg); 4588 } 4589 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 4590 output_pref(op, 0), ots->indirect_base); 4591 set_temp_val_reg(s, ots, oreg); 4592 } 4593 4594 switch (its->val_type) { 4595 case TEMP_VAL_REG: 4596 /* 4597 * The dup constriaints must be broad, covering all possible VECE. 4598 * However, tcg_op_dup_vec() gets to see the VECE and we allow it 4599 * to fail, indicating that extra moves are required for that case. 4600 */ 4601 if (tcg_regset_test_reg(dup_in_regs, its->reg)) { 4602 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) { 4603 goto done; 4604 } 4605 /* Try again from memory or a vector input register. */ 4606 } 4607 if (!its->mem_coherent) { 4608 /* 4609 * The input register is not synced, and so an extra store 4610 * would be required to use memory. Attempt an integer-vector 4611 * register move first. We do not have a TCGRegSet for this. 4612 */ 4613 if (tcg_out_mov(s, itype, ots->reg, its->reg)) { 4614 break; 4615 } 4616 /* Sync the temp back to its slot and load from there. */ 4617 temp_sync(s, its, s->reserved_regs, 0, 0); 4618 } 4619 /* fall through */ 4620 4621 case TEMP_VAL_MEM: 4622 lowpart_ofs = 0; 4623 if (HOST_BIG_ENDIAN) { 4624 lowpart_ofs = tcg_type_size(itype) - (1 << vece); 4625 } 4626 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg, 4627 its->mem_offset + lowpart_ofs)) { 4628 goto done; 4629 } 4630 /* Load the input into the destination vector register. */ 4631 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset); 4632 break; 4633 4634 default: 4635 g_assert_not_reached(); 4636 } 4637 4638 /* We now have a vector input register, so dup must succeed. */ 4639 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg); 4640 tcg_debug_assert(ok); 4641 4642 done: 4643 ots->mem_coherent = 0; 4644 if (IS_DEAD_ARG(1)) { 4645 temp_dead(s, its); 4646 } 4647 if (NEED_SYNC_ARG(0)) { 4648 temp_sync(s, ots, s->reserved_regs, 0, 0); 4649 } 4650 if (IS_DEAD_ARG(0)) { 4651 temp_dead(s, ots); 4652 } 4653 } 4654 4655 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op) 4656 { 4657 const TCGLifeData arg_life = op->life; 4658 const TCGOpDef * const def = &tcg_op_defs[op->opc]; 4659 TCGRegSet i_allocated_regs; 4660 TCGRegSet o_allocated_regs; 4661 int i, k, nb_iargs, nb_oargs; 4662 TCGReg reg; 4663 TCGArg arg; 4664 const TCGArgConstraint *arg_ct; 4665 TCGTemp *ts; 4666 TCGArg new_args[TCG_MAX_OP_ARGS]; 4667 int const_args[TCG_MAX_OP_ARGS]; 4668 4669 nb_oargs = def->nb_oargs; 4670 nb_iargs = def->nb_iargs; 4671 4672 /* copy constants */ 4673 memcpy(new_args + nb_oargs + nb_iargs, 4674 op->args + nb_oargs + nb_iargs, 4675 sizeof(TCGArg) * def->nb_cargs); 4676 4677 i_allocated_regs = s->reserved_regs; 4678 o_allocated_regs = s->reserved_regs; 4679 4680 /* satisfy input constraints */ 4681 for (k = 0; k < nb_iargs; k++) { 4682 TCGRegSet i_preferred_regs, i_required_regs; 4683 bool allocate_new_reg, copyto_new_reg; 4684 TCGTemp *ts2; 4685 int i1, i2; 4686 4687 i = def->args_ct[nb_oargs + k].sort_index; 4688 arg = op->args[i]; 4689 arg_ct = &def->args_ct[i]; 4690 ts = arg_temp(arg); 4691 4692 if (ts->val_type == TEMP_VAL_CONST 4693 && tcg_target_const_match(ts->val, ts->type, arg_ct->ct, TCGOP_VECE(op))) { 4694 /* constant is OK for instruction */ 4695 const_args[i] = 1; 4696 new_args[i] = ts->val; 4697 continue; 4698 } 4699 4700 reg = ts->reg; 4701 i_preferred_regs = 0; 4702 i_required_regs = arg_ct->regs; 4703 allocate_new_reg = false; 4704 copyto_new_reg = false; 4705 4706 switch (arg_ct->pair) { 4707 case 0: /* not paired */ 4708 if (arg_ct->ialias) { 4709 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4710 4711 /* 4712 * If the input is readonly, then it cannot also be an 4713 * output and aliased to itself. If the input is not 4714 * dead after the instruction, we must allocate a new 4715 * register and move it. 4716 */ 4717 if (temp_readonly(ts) || !IS_DEAD_ARG(i) 4718 || def->args_ct[arg_ct->alias_index].newreg) { 4719 allocate_new_reg = true; 4720 } else if (ts->val_type == TEMP_VAL_REG) { 4721 /* 4722 * Check if the current register has already been 4723 * allocated for another input. 4724 */ 4725 allocate_new_reg = 4726 tcg_regset_test_reg(i_allocated_regs, reg); 4727 } 4728 } 4729 if (!allocate_new_reg) { 4730 temp_load(s, ts, i_required_regs, i_allocated_regs, 4731 i_preferred_regs); 4732 reg = ts->reg; 4733 allocate_new_reg = !tcg_regset_test_reg(i_required_regs, reg); 4734 } 4735 if (allocate_new_reg) { 4736 /* 4737 * Allocate a new register matching the constraint 4738 * and move the temporary register into it. 4739 */ 4740 temp_load(s, ts, tcg_target_available_regs[ts->type], 4741 i_allocated_regs, 0); 4742 reg = tcg_reg_alloc(s, i_required_regs, i_allocated_regs, 4743 i_preferred_regs, ts->indirect_base); 4744 copyto_new_reg = true; 4745 } 4746 break; 4747 4748 case 1: 4749 /* First of an input pair; if i1 == i2, the second is an output. */ 4750 i1 = i; 4751 i2 = arg_ct->pair_index; 4752 ts2 = i1 != i2 ? arg_temp(op->args[i2]) : NULL; 4753 4754 /* 4755 * It is easier to default to allocating a new pair 4756 * and to identify a few cases where it's not required. 4757 */ 4758 if (arg_ct->ialias) { 4759 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4760 if (IS_DEAD_ARG(i1) && 4761 IS_DEAD_ARG(i2) && 4762 !temp_readonly(ts) && 4763 ts->val_type == TEMP_VAL_REG && 4764 ts->reg < TCG_TARGET_NB_REGS - 1 && 4765 tcg_regset_test_reg(i_required_regs, reg) && 4766 !tcg_regset_test_reg(i_allocated_regs, reg) && 4767 !tcg_regset_test_reg(i_allocated_regs, reg + 1) && 4768 (ts2 4769 ? ts2->val_type == TEMP_VAL_REG && 4770 ts2->reg == reg + 1 && 4771 !temp_readonly(ts2) 4772 : s->reg_to_temp[reg + 1] == NULL)) { 4773 break; 4774 } 4775 } else { 4776 /* Without aliasing, the pair must also be an input. */ 4777 tcg_debug_assert(ts2); 4778 if (ts->val_type == TEMP_VAL_REG && 4779 ts2->val_type == TEMP_VAL_REG && 4780 ts2->reg == reg + 1 && 4781 tcg_regset_test_reg(i_required_regs, reg)) { 4782 break; 4783 } 4784 } 4785 reg = tcg_reg_alloc_pair(s, i_required_regs, i_allocated_regs, 4786 0, ts->indirect_base); 4787 goto do_pair; 4788 4789 case 2: /* pair second */ 4790 reg = new_args[arg_ct->pair_index] + 1; 4791 goto do_pair; 4792 4793 case 3: /* ialias with second output, no first input */ 4794 tcg_debug_assert(arg_ct->ialias); 4795 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4796 4797 if (IS_DEAD_ARG(i) && 4798 !temp_readonly(ts) && 4799 ts->val_type == TEMP_VAL_REG && 4800 reg > 0 && 4801 s->reg_to_temp[reg - 1] == NULL && 4802 tcg_regset_test_reg(i_required_regs, reg) && 4803 !tcg_regset_test_reg(i_allocated_regs, reg) && 4804 !tcg_regset_test_reg(i_allocated_regs, reg - 1)) { 4805 tcg_regset_set_reg(i_allocated_regs, reg - 1); 4806 break; 4807 } 4808 reg = tcg_reg_alloc_pair(s, i_required_regs >> 1, 4809 i_allocated_regs, 0, 4810 ts->indirect_base); 4811 tcg_regset_set_reg(i_allocated_regs, reg); 4812 reg += 1; 4813 goto do_pair; 4814 4815 do_pair: 4816 /* 4817 * If an aliased input is not dead after the instruction, 4818 * we must allocate a new register and move it. 4819 */ 4820 if (arg_ct->ialias && (!IS_DEAD_ARG(i) || temp_readonly(ts))) { 4821 TCGRegSet t_allocated_regs = i_allocated_regs; 4822 4823 /* 4824 * Because of the alias, and the continued life, make sure 4825 * that the temp is somewhere *other* than the reg pair, 4826 * and we get a copy in reg. 4827 */ 4828 tcg_regset_set_reg(t_allocated_regs, reg); 4829 tcg_regset_set_reg(t_allocated_regs, reg + 1); 4830 if (ts->val_type == TEMP_VAL_REG && ts->reg == reg) { 4831 /* If ts was already in reg, copy it somewhere else. */ 4832 TCGReg nr; 4833 bool ok; 4834 4835 tcg_debug_assert(ts->kind != TEMP_FIXED); 4836 nr = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], 4837 t_allocated_regs, 0, ts->indirect_base); 4838 ok = tcg_out_mov(s, ts->type, nr, reg); 4839 tcg_debug_assert(ok); 4840 4841 set_temp_val_reg(s, ts, nr); 4842 } else { 4843 temp_load(s, ts, tcg_target_available_regs[ts->type], 4844 t_allocated_regs, 0); 4845 copyto_new_reg = true; 4846 } 4847 } else { 4848 /* Preferably allocate to reg, otherwise copy. */ 4849 i_required_regs = (TCGRegSet)1 << reg; 4850 temp_load(s, ts, i_required_regs, i_allocated_regs, 4851 i_preferred_regs); 4852 copyto_new_reg = ts->reg != reg; 4853 } 4854 break; 4855 4856 default: 4857 g_assert_not_reached(); 4858 } 4859 4860 if (copyto_new_reg) { 4861 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 4862 /* 4863 * Cross register class move not supported. Sync the 4864 * temp back to its slot and load from there. 4865 */ 4866 temp_sync(s, ts, i_allocated_regs, 0, 0); 4867 tcg_out_ld(s, ts->type, reg, 4868 ts->mem_base->reg, ts->mem_offset); 4869 } 4870 } 4871 new_args[i] = reg; 4872 const_args[i] = 0; 4873 tcg_regset_set_reg(i_allocated_regs, reg); 4874 } 4875 4876 /* mark dead temporaries and free the associated registers */ 4877 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 4878 if (IS_DEAD_ARG(i)) { 4879 temp_dead(s, arg_temp(op->args[i])); 4880 } 4881 } 4882 4883 if (def->flags & TCG_OPF_COND_BRANCH) { 4884 tcg_reg_alloc_cbranch(s, i_allocated_regs); 4885 } else if (def->flags & TCG_OPF_BB_END) { 4886 tcg_reg_alloc_bb_end(s, i_allocated_regs); 4887 } else { 4888 if (def->flags & TCG_OPF_CALL_CLOBBER) { 4889 /* XXX: permit generic clobber register list ? */ 4890 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 4891 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 4892 tcg_reg_free(s, i, i_allocated_regs); 4893 } 4894 } 4895 } 4896 if (def->flags & TCG_OPF_SIDE_EFFECTS) { 4897 /* sync globals if the op has side effects and might trigger 4898 an exception. */ 4899 sync_globals(s, i_allocated_regs); 4900 } 4901 4902 /* satisfy the output constraints */ 4903 for(k = 0; k < nb_oargs; k++) { 4904 i = def->args_ct[k].sort_index; 4905 arg = op->args[i]; 4906 arg_ct = &def->args_ct[i]; 4907 ts = arg_temp(arg); 4908 4909 /* ENV should not be modified. */ 4910 tcg_debug_assert(!temp_readonly(ts)); 4911 4912 switch (arg_ct->pair) { 4913 case 0: /* not paired */ 4914 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) { 4915 reg = new_args[arg_ct->alias_index]; 4916 } else if (arg_ct->newreg) { 4917 reg = tcg_reg_alloc(s, arg_ct->regs, 4918 i_allocated_regs | o_allocated_regs, 4919 output_pref(op, k), ts->indirect_base); 4920 } else { 4921 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs, 4922 output_pref(op, k), ts->indirect_base); 4923 } 4924 break; 4925 4926 case 1: /* first of pair */ 4927 tcg_debug_assert(!arg_ct->newreg); 4928 if (arg_ct->oalias) { 4929 reg = new_args[arg_ct->alias_index]; 4930 break; 4931 } 4932 reg = tcg_reg_alloc_pair(s, arg_ct->regs, o_allocated_regs, 4933 output_pref(op, k), ts->indirect_base); 4934 break; 4935 4936 case 2: /* second of pair */ 4937 tcg_debug_assert(!arg_ct->newreg); 4938 if (arg_ct->oalias) { 4939 reg = new_args[arg_ct->alias_index]; 4940 } else { 4941 reg = new_args[arg_ct->pair_index] + 1; 4942 } 4943 break; 4944 4945 case 3: /* first of pair, aliasing with a second input */ 4946 tcg_debug_assert(!arg_ct->newreg); 4947 reg = new_args[arg_ct->pair_index] - 1; 4948 break; 4949 4950 default: 4951 g_assert_not_reached(); 4952 } 4953 tcg_regset_set_reg(o_allocated_regs, reg); 4954 set_temp_val_reg(s, ts, reg); 4955 ts->mem_coherent = 0; 4956 new_args[i] = reg; 4957 } 4958 } 4959 4960 /* emit instruction */ 4961 switch (op->opc) { 4962 case INDEX_op_ext8s_i32: 4963 tcg_out_ext8s(s, TCG_TYPE_I32, new_args[0], new_args[1]); 4964 break; 4965 case INDEX_op_ext8s_i64: 4966 tcg_out_ext8s(s, TCG_TYPE_I64, new_args[0], new_args[1]); 4967 break; 4968 case INDEX_op_ext8u_i32: 4969 case INDEX_op_ext8u_i64: 4970 tcg_out_ext8u(s, new_args[0], new_args[1]); 4971 break; 4972 case INDEX_op_ext16s_i32: 4973 tcg_out_ext16s(s, TCG_TYPE_I32, new_args[0], new_args[1]); 4974 break; 4975 case INDEX_op_ext16s_i64: 4976 tcg_out_ext16s(s, TCG_TYPE_I64, new_args[0], new_args[1]); 4977 break; 4978 case INDEX_op_ext16u_i32: 4979 case INDEX_op_ext16u_i64: 4980 tcg_out_ext16u(s, new_args[0], new_args[1]); 4981 break; 4982 case INDEX_op_ext32s_i64: 4983 tcg_out_ext32s(s, new_args[0], new_args[1]); 4984 break; 4985 case INDEX_op_ext32u_i64: 4986 tcg_out_ext32u(s, new_args[0], new_args[1]); 4987 break; 4988 case INDEX_op_ext_i32_i64: 4989 tcg_out_exts_i32_i64(s, new_args[0], new_args[1]); 4990 break; 4991 case INDEX_op_extu_i32_i64: 4992 tcg_out_extu_i32_i64(s, new_args[0], new_args[1]); 4993 break; 4994 case INDEX_op_extrl_i64_i32: 4995 tcg_out_extrl_i64_i32(s, new_args[0], new_args[1]); 4996 break; 4997 default: 4998 if (def->flags & TCG_OPF_VECTOR) { 4999 tcg_out_vec_op(s, op->opc, TCGOP_VECL(op), TCGOP_VECE(op), 5000 new_args, const_args); 5001 } else { 5002 tcg_out_op(s, op->opc, new_args, const_args); 5003 } 5004 break; 5005 } 5006 5007 /* move the outputs in the correct register if needed */ 5008 for(i = 0; i < nb_oargs; i++) { 5009 ts = arg_temp(op->args[i]); 5010 5011 /* ENV should not be modified. */ 5012 tcg_debug_assert(!temp_readonly(ts)); 5013 5014 if (NEED_SYNC_ARG(i)) { 5015 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i)); 5016 } else if (IS_DEAD_ARG(i)) { 5017 temp_dead(s, ts); 5018 } 5019 } 5020 } 5021 5022 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op) 5023 { 5024 const TCGLifeData arg_life = op->life; 5025 TCGTemp *ots, *itsl, *itsh; 5026 TCGType vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 5027 5028 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */ 5029 tcg_debug_assert(TCG_TARGET_REG_BITS == 32); 5030 tcg_debug_assert(TCGOP_VECE(op) == MO_64); 5031 5032 ots = arg_temp(op->args[0]); 5033 itsl = arg_temp(op->args[1]); 5034 itsh = arg_temp(op->args[2]); 5035 5036 /* ENV should not be modified. */ 5037 tcg_debug_assert(!temp_readonly(ots)); 5038 5039 /* Allocate the output register now. */ 5040 if (ots->val_type != TEMP_VAL_REG) { 5041 TCGRegSet allocated_regs = s->reserved_regs; 5042 TCGRegSet dup_out_regs = 5043 tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 5044 TCGReg oreg; 5045 5046 /* Make sure to not spill the input registers. */ 5047 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) { 5048 tcg_regset_set_reg(allocated_regs, itsl->reg); 5049 } 5050 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) { 5051 tcg_regset_set_reg(allocated_regs, itsh->reg); 5052 } 5053 5054 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 5055 output_pref(op, 0), ots->indirect_base); 5056 set_temp_val_reg(s, ots, oreg); 5057 } 5058 5059 /* Promote dup2 of immediates to dupi_vec. */ 5060 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) { 5061 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val); 5062 MemOp vece = MO_64; 5063 5064 if (val == dup_const(MO_8, val)) { 5065 vece = MO_8; 5066 } else if (val == dup_const(MO_16, val)) { 5067 vece = MO_16; 5068 } else if (val == dup_const(MO_32, val)) { 5069 vece = MO_32; 5070 } 5071 5072 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val); 5073 goto done; 5074 } 5075 5076 /* If the two inputs form one 64-bit value, try dupm_vec. */ 5077 if (itsl->temp_subindex == HOST_BIG_ENDIAN && 5078 itsh->temp_subindex == !HOST_BIG_ENDIAN && 5079 itsl == itsh + (HOST_BIG_ENDIAN ? 1 : -1)) { 5080 TCGTemp *its = itsl - HOST_BIG_ENDIAN; 5081 5082 temp_sync(s, its + 0, s->reserved_regs, 0, 0); 5083 temp_sync(s, its + 1, s->reserved_regs, 0, 0); 5084 5085 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg, 5086 its->mem_base->reg, its->mem_offset)) { 5087 goto done; 5088 } 5089 } 5090 5091 /* Fall back to generic expansion. */ 5092 return false; 5093 5094 done: 5095 ots->mem_coherent = 0; 5096 if (IS_DEAD_ARG(1)) { 5097 temp_dead(s, itsl); 5098 } 5099 if (IS_DEAD_ARG(2)) { 5100 temp_dead(s, itsh); 5101 } 5102 if (NEED_SYNC_ARG(0)) { 5103 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0)); 5104 } else if (IS_DEAD_ARG(0)) { 5105 temp_dead(s, ots); 5106 } 5107 return true; 5108 } 5109 5110 static void load_arg_reg(TCGContext *s, TCGReg reg, TCGTemp *ts, 5111 TCGRegSet allocated_regs) 5112 { 5113 if (ts->val_type == TEMP_VAL_REG) { 5114 if (ts->reg != reg) { 5115 tcg_reg_free(s, reg, allocated_regs); 5116 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 5117 /* 5118 * Cross register class move not supported. Sync the 5119 * temp back to its slot and load from there. 5120 */ 5121 temp_sync(s, ts, allocated_regs, 0, 0); 5122 tcg_out_ld(s, ts->type, reg, 5123 ts->mem_base->reg, ts->mem_offset); 5124 } 5125 } 5126 } else { 5127 TCGRegSet arg_set = 0; 5128 5129 tcg_reg_free(s, reg, allocated_regs); 5130 tcg_regset_set_reg(arg_set, reg); 5131 temp_load(s, ts, arg_set, allocated_regs, 0); 5132 } 5133 } 5134 5135 static void load_arg_stk(TCGContext *s, unsigned arg_slot, TCGTemp *ts, 5136 TCGRegSet allocated_regs) 5137 { 5138 /* 5139 * When the destination is on the stack, load up the temp and store. 5140 * If there are many call-saved registers, the temp might live to 5141 * see another use; otherwise it'll be discarded. 5142 */ 5143 temp_load(s, ts, tcg_target_available_regs[ts->type], allocated_regs, 0); 5144 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, 5145 arg_slot_stk_ofs(arg_slot)); 5146 } 5147 5148 static void load_arg_normal(TCGContext *s, const TCGCallArgumentLoc *l, 5149 TCGTemp *ts, TCGRegSet *allocated_regs) 5150 { 5151 if (arg_slot_reg_p(l->arg_slot)) { 5152 TCGReg reg = tcg_target_call_iarg_regs[l->arg_slot]; 5153 load_arg_reg(s, reg, ts, *allocated_regs); 5154 tcg_regset_set_reg(*allocated_regs, reg); 5155 } else { 5156 load_arg_stk(s, l->arg_slot, ts, *allocated_regs); 5157 } 5158 } 5159 5160 static void load_arg_ref(TCGContext *s, unsigned arg_slot, TCGReg ref_base, 5161 intptr_t ref_off, TCGRegSet *allocated_regs) 5162 { 5163 TCGReg reg; 5164 5165 if (arg_slot_reg_p(arg_slot)) { 5166 reg = tcg_target_call_iarg_regs[arg_slot]; 5167 tcg_reg_free(s, reg, *allocated_regs); 5168 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5169 tcg_regset_set_reg(*allocated_regs, reg); 5170 } else { 5171 reg = tcg_reg_alloc(s, tcg_target_available_regs[TCG_TYPE_PTR], 5172 *allocated_regs, 0, false); 5173 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5174 tcg_out_st(s, TCG_TYPE_PTR, reg, TCG_REG_CALL_STACK, 5175 arg_slot_stk_ofs(arg_slot)); 5176 } 5177 } 5178 5179 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op) 5180 { 5181 const int nb_oargs = TCGOP_CALLO(op); 5182 const int nb_iargs = TCGOP_CALLI(op); 5183 const TCGLifeData arg_life = op->life; 5184 const TCGHelperInfo *info = tcg_call_info(op); 5185 TCGRegSet allocated_regs = s->reserved_regs; 5186 int i; 5187 5188 /* 5189 * Move inputs into place in reverse order, 5190 * so that we place stacked arguments first. 5191 */ 5192 for (i = nb_iargs - 1; i >= 0; --i) { 5193 const TCGCallArgumentLoc *loc = &info->in[i]; 5194 TCGTemp *ts = arg_temp(op->args[nb_oargs + i]); 5195 5196 switch (loc->kind) { 5197 case TCG_CALL_ARG_NORMAL: 5198 case TCG_CALL_ARG_EXTEND_U: 5199 case TCG_CALL_ARG_EXTEND_S: 5200 load_arg_normal(s, loc, ts, &allocated_regs); 5201 break; 5202 case TCG_CALL_ARG_BY_REF: 5203 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5204 load_arg_ref(s, loc->arg_slot, TCG_REG_CALL_STACK, 5205 arg_slot_stk_ofs(loc->ref_slot), 5206 &allocated_regs); 5207 break; 5208 case TCG_CALL_ARG_BY_REF_N: 5209 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5210 break; 5211 default: 5212 g_assert_not_reached(); 5213 } 5214 } 5215 5216 /* Mark dead temporaries and free the associated registers. */ 5217 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 5218 if (IS_DEAD_ARG(i)) { 5219 temp_dead(s, arg_temp(op->args[i])); 5220 } 5221 } 5222 5223 /* Clobber call registers. */ 5224 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 5225 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 5226 tcg_reg_free(s, i, allocated_regs); 5227 } 5228 } 5229 5230 /* 5231 * Save globals if they might be written by the helper, 5232 * sync them if they might be read. 5233 */ 5234 if (info->flags & TCG_CALL_NO_READ_GLOBALS) { 5235 /* Nothing to do */ 5236 } else if (info->flags & TCG_CALL_NO_WRITE_GLOBALS) { 5237 sync_globals(s, allocated_regs); 5238 } else { 5239 save_globals(s, allocated_regs); 5240 } 5241 5242 /* 5243 * If the ABI passes a pointer to the returned struct as the first 5244 * argument, load that now. Pass a pointer to the output home slot. 5245 */ 5246 if (info->out_kind == TCG_CALL_RET_BY_REF) { 5247 TCGTemp *ts = arg_temp(op->args[0]); 5248 5249 if (!ts->mem_allocated) { 5250 temp_allocate_frame(s, ts); 5251 } 5252 load_arg_ref(s, 0, ts->mem_base->reg, ts->mem_offset, &allocated_regs); 5253 } 5254 5255 tcg_out_call(s, tcg_call_func(op), info); 5256 5257 /* Assign output registers and emit moves if needed. */ 5258 switch (info->out_kind) { 5259 case TCG_CALL_RET_NORMAL: 5260 for (i = 0; i < nb_oargs; i++) { 5261 TCGTemp *ts = arg_temp(op->args[i]); 5262 TCGReg reg = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, i); 5263 5264 /* ENV should not be modified. */ 5265 tcg_debug_assert(!temp_readonly(ts)); 5266 5267 set_temp_val_reg(s, ts, reg); 5268 ts->mem_coherent = 0; 5269 } 5270 break; 5271 5272 case TCG_CALL_RET_BY_VEC: 5273 { 5274 TCGTemp *ts = arg_temp(op->args[0]); 5275 5276 tcg_debug_assert(ts->base_type == TCG_TYPE_I128); 5277 tcg_debug_assert(ts->temp_subindex == 0); 5278 if (!ts->mem_allocated) { 5279 temp_allocate_frame(s, ts); 5280 } 5281 tcg_out_st(s, TCG_TYPE_V128, 5282 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5283 ts->mem_base->reg, ts->mem_offset); 5284 } 5285 /* fall through to mark all parts in memory */ 5286 5287 case TCG_CALL_RET_BY_REF: 5288 /* The callee has performed a write through the reference. */ 5289 for (i = 0; i < nb_oargs; i++) { 5290 TCGTemp *ts = arg_temp(op->args[i]); 5291 ts->val_type = TEMP_VAL_MEM; 5292 } 5293 break; 5294 5295 default: 5296 g_assert_not_reached(); 5297 } 5298 5299 /* Flush or discard output registers as needed. */ 5300 for (i = 0; i < nb_oargs; i++) { 5301 TCGTemp *ts = arg_temp(op->args[i]); 5302 if (NEED_SYNC_ARG(i)) { 5303 temp_sync(s, ts, s->reserved_regs, 0, IS_DEAD_ARG(i)); 5304 } else if (IS_DEAD_ARG(i)) { 5305 temp_dead(s, ts); 5306 } 5307 } 5308 } 5309 5310 /** 5311 * atom_and_align_for_opc: 5312 * @s: tcg context 5313 * @opc: memory operation code 5314 * @host_atom: MO_ATOM_{IFALIGN,WITHIN16,SUBALIGN} for host operations 5315 * @allow_two_ops: true if we are prepared to issue two operations 5316 * 5317 * Return the alignment and atomicity to use for the inline fast path 5318 * for the given memory operation. The alignment may be larger than 5319 * that specified in @opc, and the correct alignment will be diagnosed 5320 * by the slow path helper. 5321 * 5322 * If @allow_two_ops, the host is prepared to test for 2x alignment, 5323 * and issue two loads or stores for subalignment. 5324 */ 5325 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 5326 MemOp host_atom, bool allow_two_ops) 5327 { 5328 MemOp align = get_alignment_bits(opc); 5329 MemOp size = opc & MO_SIZE; 5330 MemOp half = size ? size - 1 : 0; 5331 MemOp atmax; 5332 MemOp atom; 5333 5334 /* When serialized, no further atomicity required. */ 5335 if (s->gen_tb->cflags & CF_PARALLEL) { 5336 atom = opc & MO_ATOM_MASK; 5337 } else { 5338 atom = MO_ATOM_NONE; 5339 } 5340 5341 switch (atom) { 5342 case MO_ATOM_NONE: 5343 /* The operation requires no specific atomicity. */ 5344 atmax = MO_8; 5345 break; 5346 5347 case MO_ATOM_IFALIGN: 5348 atmax = size; 5349 break; 5350 5351 case MO_ATOM_IFALIGN_PAIR: 5352 atmax = half; 5353 break; 5354 5355 case MO_ATOM_WITHIN16: 5356 atmax = size; 5357 if (size == MO_128) { 5358 /* Misalignment implies !within16, and therefore no atomicity. */ 5359 } else if (host_atom != MO_ATOM_WITHIN16) { 5360 /* The host does not implement within16, so require alignment. */ 5361 align = MAX(align, size); 5362 } 5363 break; 5364 5365 case MO_ATOM_WITHIN16_PAIR: 5366 atmax = size; 5367 /* 5368 * Misalignment implies !within16, and therefore half atomicity. 5369 * Any host prepared for two operations can implement this with 5370 * half alignment. 5371 */ 5372 if (host_atom != MO_ATOM_WITHIN16 && allow_two_ops) { 5373 align = MAX(align, half); 5374 } 5375 break; 5376 5377 case MO_ATOM_SUBALIGN: 5378 atmax = size; 5379 if (host_atom != MO_ATOM_SUBALIGN) { 5380 /* If unaligned but not odd, there are subobjects up to half. */ 5381 if (allow_two_ops) { 5382 align = MAX(align, half); 5383 } else { 5384 align = MAX(align, size); 5385 } 5386 } 5387 break; 5388 5389 default: 5390 g_assert_not_reached(); 5391 } 5392 5393 return (TCGAtomAlign){ .atom = atmax, .align = align }; 5394 } 5395 5396 /* 5397 * Similarly for qemu_ld/st slow path helpers. 5398 * We must re-implement tcg_gen_callN and tcg_reg_alloc_call simultaneously, 5399 * using only the provided backend tcg_out_* functions. 5400 */ 5401 5402 static int tcg_out_helper_stk_ofs(TCGType type, unsigned slot) 5403 { 5404 int ofs = arg_slot_stk_ofs(slot); 5405 5406 /* 5407 * Each stack slot is TCG_TARGET_LONG_BITS. If the host does not 5408 * require extension to uint64_t, adjust the address for uint32_t. 5409 */ 5410 if (HOST_BIG_ENDIAN && 5411 TCG_TARGET_REG_BITS == 64 && 5412 type == TCG_TYPE_I32) { 5413 ofs += 4; 5414 } 5415 return ofs; 5416 } 5417 5418 static void tcg_out_helper_load_slots(TCGContext *s, 5419 unsigned nmov, TCGMovExtend *mov, 5420 const TCGLdstHelperParam *parm) 5421 { 5422 unsigned i; 5423 TCGReg dst3; 5424 5425 /* 5426 * Start from the end, storing to the stack first. 5427 * This frees those registers, so we need not consider overlap. 5428 */ 5429 for (i = nmov; i-- > 0; ) { 5430 unsigned slot = mov[i].dst; 5431 5432 if (arg_slot_reg_p(slot)) { 5433 goto found_reg; 5434 } 5435 5436 TCGReg src = mov[i].src; 5437 TCGType dst_type = mov[i].dst_type; 5438 MemOp dst_mo = dst_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5439 5440 /* The argument is going onto the stack; extend into scratch. */ 5441 if ((mov[i].src_ext & MO_SIZE) != dst_mo) { 5442 tcg_debug_assert(parm->ntmp != 0); 5443 mov[i].dst = src = parm->tmp[0]; 5444 tcg_out_movext1(s, &mov[i]); 5445 } 5446 5447 tcg_out_st(s, dst_type, src, TCG_REG_CALL_STACK, 5448 tcg_out_helper_stk_ofs(dst_type, slot)); 5449 } 5450 return; 5451 5452 found_reg: 5453 /* 5454 * The remaining arguments are in registers. 5455 * Convert slot numbers to argument registers. 5456 */ 5457 nmov = i + 1; 5458 for (i = 0; i < nmov; ++i) { 5459 mov[i].dst = tcg_target_call_iarg_regs[mov[i].dst]; 5460 } 5461 5462 switch (nmov) { 5463 case 4: 5464 /* The backend must have provided enough temps for the worst case. */ 5465 tcg_debug_assert(parm->ntmp >= 2); 5466 5467 dst3 = mov[3].dst; 5468 for (unsigned j = 0; j < 3; ++j) { 5469 if (dst3 == mov[j].src) { 5470 /* 5471 * Conflict. Copy the source to a temporary, perform the 5472 * remaining moves, then the extension from our scratch 5473 * on the way out. 5474 */ 5475 TCGReg scratch = parm->tmp[1]; 5476 5477 tcg_out_mov(s, mov[3].src_type, scratch, mov[3].src); 5478 tcg_out_movext3(s, mov, mov + 1, mov + 2, parm->tmp[0]); 5479 tcg_out_movext1_new_src(s, &mov[3], scratch); 5480 break; 5481 } 5482 } 5483 5484 /* No conflicts: perform this move and continue. */ 5485 tcg_out_movext1(s, &mov[3]); 5486 /* fall through */ 5487 5488 case 3: 5489 tcg_out_movext3(s, mov, mov + 1, mov + 2, 5490 parm->ntmp ? parm->tmp[0] : -1); 5491 break; 5492 case 2: 5493 tcg_out_movext2(s, mov, mov + 1, 5494 parm->ntmp ? parm->tmp[0] : -1); 5495 break; 5496 case 1: 5497 tcg_out_movext1(s, mov); 5498 break; 5499 default: 5500 g_assert_not_reached(); 5501 } 5502 } 5503 5504 static void tcg_out_helper_load_imm(TCGContext *s, unsigned slot, 5505 TCGType type, tcg_target_long imm, 5506 const TCGLdstHelperParam *parm) 5507 { 5508 if (arg_slot_reg_p(slot)) { 5509 tcg_out_movi(s, type, tcg_target_call_iarg_regs[slot], imm); 5510 } else { 5511 int ofs = tcg_out_helper_stk_ofs(type, slot); 5512 if (!tcg_out_sti(s, type, imm, TCG_REG_CALL_STACK, ofs)) { 5513 tcg_debug_assert(parm->ntmp != 0); 5514 tcg_out_movi(s, type, parm->tmp[0], imm); 5515 tcg_out_st(s, type, parm->tmp[0], TCG_REG_CALL_STACK, ofs); 5516 } 5517 } 5518 } 5519 5520 static void tcg_out_helper_load_common_args(TCGContext *s, 5521 const TCGLabelQemuLdst *ldst, 5522 const TCGLdstHelperParam *parm, 5523 const TCGHelperInfo *info, 5524 unsigned next_arg) 5525 { 5526 TCGMovExtend ptr_mov = { 5527 .dst_type = TCG_TYPE_PTR, 5528 .src_type = TCG_TYPE_PTR, 5529 .src_ext = sizeof(void *) == 4 ? MO_32 : MO_64 5530 }; 5531 const TCGCallArgumentLoc *loc = &info->in[0]; 5532 TCGType type; 5533 unsigned slot; 5534 tcg_target_ulong imm; 5535 5536 /* 5537 * Handle env, which is always first. 5538 */ 5539 ptr_mov.dst = loc->arg_slot; 5540 ptr_mov.src = TCG_AREG0; 5541 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5542 5543 /* 5544 * Handle oi. 5545 */ 5546 imm = ldst->oi; 5547 loc = &info->in[next_arg]; 5548 type = TCG_TYPE_I32; 5549 switch (loc->kind) { 5550 case TCG_CALL_ARG_NORMAL: 5551 break; 5552 case TCG_CALL_ARG_EXTEND_U: 5553 case TCG_CALL_ARG_EXTEND_S: 5554 /* No extension required for MemOpIdx. */ 5555 tcg_debug_assert(imm <= INT32_MAX); 5556 type = TCG_TYPE_REG; 5557 break; 5558 default: 5559 g_assert_not_reached(); 5560 } 5561 tcg_out_helper_load_imm(s, loc->arg_slot, type, imm, parm); 5562 next_arg++; 5563 5564 /* 5565 * Handle ra. 5566 */ 5567 loc = &info->in[next_arg]; 5568 slot = loc->arg_slot; 5569 if (parm->ra_gen) { 5570 int arg_reg = -1; 5571 TCGReg ra_reg; 5572 5573 if (arg_slot_reg_p(slot)) { 5574 arg_reg = tcg_target_call_iarg_regs[slot]; 5575 } 5576 ra_reg = parm->ra_gen(s, ldst, arg_reg); 5577 5578 ptr_mov.dst = slot; 5579 ptr_mov.src = ra_reg; 5580 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5581 } else { 5582 imm = (uintptr_t)ldst->raddr; 5583 tcg_out_helper_load_imm(s, slot, TCG_TYPE_PTR, imm, parm); 5584 } 5585 } 5586 5587 static unsigned tcg_out_helper_add_mov(TCGMovExtend *mov, 5588 const TCGCallArgumentLoc *loc, 5589 TCGType dst_type, TCGType src_type, 5590 TCGReg lo, TCGReg hi) 5591 { 5592 MemOp reg_mo; 5593 5594 if (dst_type <= TCG_TYPE_REG) { 5595 MemOp src_ext; 5596 5597 switch (loc->kind) { 5598 case TCG_CALL_ARG_NORMAL: 5599 src_ext = src_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5600 break; 5601 case TCG_CALL_ARG_EXTEND_U: 5602 dst_type = TCG_TYPE_REG; 5603 src_ext = MO_UL; 5604 break; 5605 case TCG_CALL_ARG_EXTEND_S: 5606 dst_type = TCG_TYPE_REG; 5607 src_ext = MO_SL; 5608 break; 5609 default: 5610 g_assert_not_reached(); 5611 } 5612 5613 mov[0].dst = loc->arg_slot; 5614 mov[0].dst_type = dst_type; 5615 mov[0].src = lo; 5616 mov[0].src_type = src_type; 5617 mov[0].src_ext = src_ext; 5618 return 1; 5619 } 5620 5621 if (TCG_TARGET_REG_BITS == 32) { 5622 assert(dst_type == TCG_TYPE_I64); 5623 reg_mo = MO_32; 5624 } else { 5625 assert(dst_type == TCG_TYPE_I128); 5626 reg_mo = MO_64; 5627 } 5628 5629 mov[0].dst = loc[HOST_BIG_ENDIAN].arg_slot; 5630 mov[0].src = lo; 5631 mov[0].dst_type = TCG_TYPE_REG; 5632 mov[0].src_type = TCG_TYPE_REG; 5633 mov[0].src_ext = reg_mo; 5634 5635 mov[1].dst = loc[!HOST_BIG_ENDIAN].arg_slot; 5636 mov[1].src = hi; 5637 mov[1].dst_type = TCG_TYPE_REG; 5638 mov[1].src_type = TCG_TYPE_REG; 5639 mov[1].src_ext = reg_mo; 5640 5641 return 2; 5642 } 5643 5644 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 5645 const TCGLdstHelperParam *parm) 5646 { 5647 const TCGHelperInfo *info; 5648 const TCGCallArgumentLoc *loc; 5649 TCGMovExtend mov[2]; 5650 unsigned next_arg, nmov; 5651 MemOp mop = get_memop(ldst->oi); 5652 5653 switch (mop & MO_SIZE) { 5654 case MO_8: 5655 case MO_16: 5656 case MO_32: 5657 info = &info_helper_ld32_mmu; 5658 break; 5659 case MO_64: 5660 info = &info_helper_ld64_mmu; 5661 break; 5662 case MO_128: 5663 info = &info_helper_ld128_mmu; 5664 break; 5665 default: 5666 g_assert_not_reached(); 5667 } 5668 5669 /* Defer env argument. */ 5670 next_arg = 1; 5671 5672 loc = &info->in[next_arg]; 5673 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5674 /* 5675 * 32-bit host with 32-bit guest: zero-extend the guest address 5676 * to 64-bits for the helper by storing the low part, then 5677 * load a zero for the high part. 5678 */ 5679 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 5680 TCG_TYPE_I32, TCG_TYPE_I32, 5681 ldst->addrlo_reg, -1); 5682 tcg_out_helper_load_slots(s, 1, mov, parm); 5683 5684 tcg_out_helper_load_imm(s, loc[!HOST_BIG_ENDIAN].arg_slot, 5685 TCG_TYPE_I32, 0, parm); 5686 next_arg += 2; 5687 } else { 5688 nmov = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 5689 ldst->addrlo_reg, ldst->addrhi_reg); 5690 tcg_out_helper_load_slots(s, nmov, mov, parm); 5691 next_arg += nmov; 5692 } 5693 5694 switch (info->out_kind) { 5695 case TCG_CALL_RET_NORMAL: 5696 case TCG_CALL_RET_BY_VEC: 5697 break; 5698 case TCG_CALL_RET_BY_REF: 5699 /* 5700 * The return reference is in the first argument slot. 5701 * We need memory in which to return: re-use the top of stack. 5702 */ 5703 { 5704 int ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 5705 5706 if (arg_slot_reg_p(0)) { 5707 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[0], 5708 TCG_REG_CALL_STACK, ofs_slot0); 5709 } else { 5710 tcg_debug_assert(parm->ntmp != 0); 5711 tcg_out_addi_ptr(s, parm->tmp[0], 5712 TCG_REG_CALL_STACK, ofs_slot0); 5713 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 5714 TCG_REG_CALL_STACK, ofs_slot0); 5715 } 5716 } 5717 break; 5718 default: 5719 g_assert_not_reached(); 5720 } 5721 5722 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 5723 } 5724 5725 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *ldst, 5726 bool load_sign, 5727 const TCGLdstHelperParam *parm) 5728 { 5729 MemOp mop = get_memop(ldst->oi); 5730 TCGMovExtend mov[2]; 5731 int ofs_slot0; 5732 5733 switch (ldst->type) { 5734 case TCG_TYPE_I64: 5735 if (TCG_TARGET_REG_BITS == 32) { 5736 break; 5737 } 5738 /* fall through */ 5739 5740 case TCG_TYPE_I32: 5741 mov[0].dst = ldst->datalo_reg; 5742 mov[0].src = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, 0); 5743 mov[0].dst_type = ldst->type; 5744 mov[0].src_type = TCG_TYPE_REG; 5745 5746 /* 5747 * If load_sign, then we allowed the helper to perform the 5748 * appropriate sign extension to tcg_target_ulong, and all 5749 * we need now is a plain move. 5750 * 5751 * If they do not, then we expect the relevant extension 5752 * instruction to be no more expensive than a move, and 5753 * we thus save the icache etc by only using one of two 5754 * helper functions. 5755 */ 5756 if (load_sign || !(mop & MO_SIGN)) { 5757 if (TCG_TARGET_REG_BITS == 32 || ldst->type == TCG_TYPE_I32) { 5758 mov[0].src_ext = MO_32; 5759 } else { 5760 mov[0].src_ext = MO_64; 5761 } 5762 } else { 5763 mov[0].src_ext = mop & MO_SSIZE; 5764 } 5765 tcg_out_movext1(s, mov); 5766 return; 5767 5768 case TCG_TYPE_I128: 5769 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 5770 ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 5771 switch (TCG_TARGET_CALL_RET_I128) { 5772 case TCG_CALL_RET_NORMAL: 5773 break; 5774 case TCG_CALL_RET_BY_VEC: 5775 tcg_out_st(s, TCG_TYPE_V128, 5776 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5777 TCG_REG_CALL_STACK, ofs_slot0); 5778 /* fall through */ 5779 case TCG_CALL_RET_BY_REF: 5780 tcg_out_ld(s, TCG_TYPE_I64, ldst->datalo_reg, 5781 TCG_REG_CALL_STACK, ofs_slot0 + 8 * HOST_BIG_ENDIAN); 5782 tcg_out_ld(s, TCG_TYPE_I64, ldst->datahi_reg, 5783 TCG_REG_CALL_STACK, ofs_slot0 + 8 * !HOST_BIG_ENDIAN); 5784 return; 5785 default: 5786 g_assert_not_reached(); 5787 } 5788 break; 5789 5790 default: 5791 g_assert_not_reached(); 5792 } 5793 5794 mov[0].dst = ldst->datalo_reg; 5795 mov[0].src = 5796 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, HOST_BIG_ENDIAN); 5797 mov[0].dst_type = TCG_TYPE_REG; 5798 mov[0].src_type = TCG_TYPE_REG; 5799 mov[0].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 5800 5801 mov[1].dst = ldst->datahi_reg; 5802 mov[1].src = 5803 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, !HOST_BIG_ENDIAN); 5804 mov[1].dst_type = TCG_TYPE_REG; 5805 mov[1].src_type = TCG_TYPE_REG; 5806 mov[1].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 5807 5808 tcg_out_movext2(s, mov, mov + 1, parm->ntmp ? parm->tmp[0] : -1); 5809 } 5810 5811 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 5812 const TCGLdstHelperParam *parm) 5813 { 5814 const TCGHelperInfo *info; 5815 const TCGCallArgumentLoc *loc; 5816 TCGMovExtend mov[4]; 5817 TCGType data_type; 5818 unsigned next_arg, nmov, n; 5819 MemOp mop = get_memop(ldst->oi); 5820 5821 switch (mop & MO_SIZE) { 5822 case MO_8: 5823 case MO_16: 5824 case MO_32: 5825 info = &info_helper_st32_mmu; 5826 data_type = TCG_TYPE_I32; 5827 break; 5828 case MO_64: 5829 info = &info_helper_st64_mmu; 5830 data_type = TCG_TYPE_I64; 5831 break; 5832 case MO_128: 5833 info = &info_helper_st128_mmu; 5834 data_type = TCG_TYPE_I128; 5835 break; 5836 default: 5837 g_assert_not_reached(); 5838 } 5839 5840 /* Defer env argument. */ 5841 next_arg = 1; 5842 nmov = 0; 5843 5844 /* Handle addr argument. */ 5845 loc = &info->in[next_arg]; 5846 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5847 /* 5848 * 32-bit host with 32-bit guest: zero-extend the guest address 5849 * to 64-bits for the helper by storing the low part. Later, 5850 * after we have processed the register inputs, we will load a 5851 * zero for the high part. 5852 */ 5853 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 5854 TCG_TYPE_I32, TCG_TYPE_I32, 5855 ldst->addrlo_reg, -1); 5856 next_arg += 2; 5857 nmov += 1; 5858 } else { 5859 n = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 5860 ldst->addrlo_reg, ldst->addrhi_reg); 5861 next_arg += n; 5862 nmov += n; 5863 } 5864 5865 /* Handle data argument. */ 5866 loc = &info->in[next_arg]; 5867 switch (loc->kind) { 5868 case TCG_CALL_ARG_NORMAL: 5869 case TCG_CALL_ARG_EXTEND_U: 5870 case TCG_CALL_ARG_EXTEND_S: 5871 n = tcg_out_helper_add_mov(mov + nmov, loc, data_type, ldst->type, 5872 ldst->datalo_reg, ldst->datahi_reg); 5873 next_arg += n; 5874 nmov += n; 5875 tcg_out_helper_load_slots(s, nmov, mov, parm); 5876 break; 5877 5878 case TCG_CALL_ARG_BY_REF: 5879 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 5880 tcg_debug_assert(data_type == TCG_TYPE_I128); 5881 tcg_out_st(s, TCG_TYPE_I64, 5882 HOST_BIG_ENDIAN ? ldst->datahi_reg : ldst->datalo_reg, 5883 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[0].ref_slot)); 5884 tcg_out_st(s, TCG_TYPE_I64, 5885 HOST_BIG_ENDIAN ? ldst->datalo_reg : ldst->datahi_reg, 5886 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[1].ref_slot)); 5887 5888 tcg_out_helper_load_slots(s, nmov, mov, parm); 5889 5890 if (arg_slot_reg_p(loc->arg_slot)) { 5891 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[loc->arg_slot], 5892 TCG_REG_CALL_STACK, 5893 arg_slot_stk_ofs(loc->ref_slot)); 5894 } else { 5895 tcg_debug_assert(parm->ntmp != 0); 5896 tcg_out_addi_ptr(s, parm->tmp[0], TCG_REG_CALL_STACK, 5897 arg_slot_stk_ofs(loc->ref_slot)); 5898 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 5899 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc->arg_slot)); 5900 } 5901 next_arg += 2; 5902 break; 5903 5904 default: 5905 g_assert_not_reached(); 5906 } 5907 5908 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5909 /* Zero extend the address by loading a zero for the high part. */ 5910 loc = &info->in[1 + !HOST_BIG_ENDIAN]; 5911 tcg_out_helper_load_imm(s, loc->arg_slot, TCG_TYPE_I32, 0, parm); 5912 } 5913 5914 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 5915 } 5916 5917 void tcg_dump_op_count(GString *buf) 5918 { 5919 g_string_append_printf(buf, "[TCG profiler not compiled]\n"); 5920 } 5921 5922 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start) 5923 { 5924 int i, start_words, num_insns; 5925 TCGOp *op; 5926 5927 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP) 5928 && qemu_log_in_addr_range(pc_start))) { 5929 FILE *logfile = qemu_log_trylock(); 5930 if (logfile) { 5931 fprintf(logfile, "OP:\n"); 5932 tcg_dump_ops(s, logfile, false); 5933 fprintf(logfile, "\n"); 5934 qemu_log_unlock(logfile); 5935 } 5936 } 5937 5938 #ifdef CONFIG_DEBUG_TCG 5939 /* Ensure all labels referenced have been emitted. */ 5940 { 5941 TCGLabel *l; 5942 bool error = false; 5943 5944 QSIMPLEQ_FOREACH(l, &s->labels, next) { 5945 if (unlikely(!l->present) && !QSIMPLEQ_EMPTY(&l->branches)) { 5946 qemu_log_mask(CPU_LOG_TB_OP, 5947 "$L%d referenced but not present.\n", l->id); 5948 error = true; 5949 } 5950 } 5951 assert(!error); 5952 } 5953 #endif 5954 5955 tcg_optimize(s); 5956 5957 reachable_code_pass(s); 5958 liveness_pass_0(s); 5959 liveness_pass_1(s); 5960 5961 if (s->nb_indirects > 0) { 5962 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND) 5963 && qemu_log_in_addr_range(pc_start))) { 5964 FILE *logfile = qemu_log_trylock(); 5965 if (logfile) { 5966 fprintf(logfile, "OP before indirect lowering:\n"); 5967 tcg_dump_ops(s, logfile, false); 5968 fprintf(logfile, "\n"); 5969 qemu_log_unlock(logfile); 5970 } 5971 } 5972 5973 /* Replace indirect temps with direct temps. */ 5974 if (liveness_pass_2(s)) { 5975 /* If changes were made, re-run liveness. */ 5976 liveness_pass_1(s); 5977 } 5978 } 5979 5980 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT) 5981 && qemu_log_in_addr_range(pc_start))) { 5982 FILE *logfile = qemu_log_trylock(); 5983 if (logfile) { 5984 fprintf(logfile, "OP after optimization and liveness analysis:\n"); 5985 tcg_dump_ops(s, logfile, true); 5986 fprintf(logfile, "\n"); 5987 qemu_log_unlock(logfile); 5988 } 5989 } 5990 5991 /* Initialize goto_tb jump offsets. */ 5992 tb->jmp_reset_offset[0] = TB_JMP_OFFSET_INVALID; 5993 tb->jmp_reset_offset[1] = TB_JMP_OFFSET_INVALID; 5994 tb->jmp_insn_offset[0] = TB_JMP_OFFSET_INVALID; 5995 tb->jmp_insn_offset[1] = TB_JMP_OFFSET_INVALID; 5996 5997 tcg_reg_alloc_start(s); 5998 5999 /* 6000 * Reset the buffer pointers when restarting after overflow. 6001 * TODO: Move this into translate-all.c with the rest of the 6002 * buffer management. Having only this done here is confusing. 6003 */ 6004 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr); 6005 s->code_ptr = s->code_buf; 6006 6007 #ifdef TCG_TARGET_NEED_LDST_LABELS 6008 QSIMPLEQ_INIT(&s->ldst_labels); 6009 #endif 6010 #ifdef TCG_TARGET_NEED_POOL_LABELS 6011 s->pool_labels = NULL; 6012 #endif 6013 6014 start_words = s->insn_start_words; 6015 s->gen_insn_data = 6016 tcg_malloc(sizeof(uint64_t) * s->gen_tb->icount * start_words); 6017 6018 tcg_out_tb_start(s); 6019 6020 num_insns = -1; 6021 QTAILQ_FOREACH(op, &s->ops, link) { 6022 TCGOpcode opc = op->opc; 6023 6024 switch (opc) { 6025 case INDEX_op_mov_i32: 6026 case INDEX_op_mov_i64: 6027 case INDEX_op_mov_vec: 6028 tcg_reg_alloc_mov(s, op); 6029 break; 6030 case INDEX_op_dup_vec: 6031 tcg_reg_alloc_dup(s, op); 6032 break; 6033 case INDEX_op_insn_start: 6034 if (num_insns >= 0) { 6035 size_t off = tcg_current_code_size(s); 6036 s->gen_insn_end_off[num_insns] = off; 6037 /* Assert that we do not overflow our stored offset. */ 6038 assert(s->gen_insn_end_off[num_insns] == off); 6039 } 6040 num_insns++; 6041 for (i = 0; i < start_words; ++i) { 6042 s->gen_insn_data[num_insns * start_words + i] = 6043 tcg_get_insn_start_param(op, i); 6044 } 6045 break; 6046 case INDEX_op_discard: 6047 temp_dead(s, arg_temp(op->args[0])); 6048 break; 6049 case INDEX_op_set_label: 6050 tcg_reg_alloc_bb_end(s, s->reserved_regs); 6051 tcg_out_label(s, arg_label(op->args[0])); 6052 break; 6053 case INDEX_op_call: 6054 tcg_reg_alloc_call(s, op); 6055 break; 6056 case INDEX_op_exit_tb: 6057 tcg_out_exit_tb(s, op->args[0]); 6058 break; 6059 case INDEX_op_goto_tb: 6060 tcg_out_goto_tb(s, op->args[0]); 6061 break; 6062 case INDEX_op_dup2_vec: 6063 if (tcg_reg_alloc_dup2(s, op)) { 6064 break; 6065 } 6066 /* fall through */ 6067 default: 6068 /* Sanity check that we've not introduced any unhandled opcodes. */ 6069 tcg_debug_assert(tcg_op_supported(opc)); 6070 /* Note: in order to speed up the code, it would be much 6071 faster to have specialized register allocator functions for 6072 some common argument patterns */ 6073 tcg_reg_alloc_op(s, op); 6074 break; 6075 } 6076 /* Test for (pending) buffer overflow. The assumption is that any 6077 one operation beginning below the high water mark cannot overrun 6078 the buffer completely. Thus we can test for overflow after 6079 generating code without having to check during generation. */ 6080 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) { 6081 return -1; 6082 } 6083 /* Test for TB overflow, as seen by gen_insn_end_off. */ 6084 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) { 6085 return -2; 6086 } 6087 } 6088 tcg_debug_assert(num_insns + 1 == s->gen_tb->icount); 6089 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s); 6090 6091 /* Generate TB finalization at the end of block */ 6092 #ifdef TCG_TARGET_NEED_LDST_LABELS 6093 i = tcg_out_ldst_finalize(s); 6094 if (i < 0) { 6095 return i; 6096 } 6097 #endif 6098 #ifdef TCG_TARGET_NEED_POOL_LABELS 6099 i = tcg_out_pool_finalize(s); 6100 if (i < 0) { 6101 return i; 6102 } 6103 #endif 6104 if (!tcg_resolve_relocs(s)) { 6105 return -2; 6106 } 6107 6108 #ifndef CONFIG_TCG_INTERPRETER 6109 /* flush instruction cache */ 6110 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 6111 (uintptr_t)s->code_buf, 6112 tcg_ptr_byte_diff(s->code_ptr, s->code_buf)); 6113 #endif 6114 6115 return tcg_current_code_size(s); 6116 } 6117 6118 void tcg_dump_info(GString *buf) 6119 { 6120 g_string_append_printf(buf, "[TCG profiler not compiled]\n"); 6121 } 6122 6123 #ifdef ELF_HOST_MACHINE 6124 /* In order to use this feature, the backend needs to do three things: 6125 6126 (1) Define ELF_HOST_MACHINE to indicate both what value to 6127 put into the ELF image and to indicate support for the feature. 6128 6129 (2) Define tcg_register_jit. This should create a buffer containing 6130 the contents of a .debug_frame section that describes the post- 6131 prologue unwind info for the tcg machine. 6132 6133 (3) Call tcg_register_jit_int, with the constructed .debug_frame. 6134 */ 6135 6136 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */ 6137 typedef enum { 6138 JIT_NOACTION = 0, 6139 JIT_REGISTER_FN, 6140 JIT_UNREGISTER_FN 6141 } jit_actions_t; 6142 6143 struct jit_code_entry { 6144 struct jit_code_entry *next_entry; 6145 struct jit_code_entry *prev_entry; 6146 const void *symfile_addr; 6147 uint64_t symfile_size; 6148 }; 6149 6150 struct jit_descriptor { 6151 uint32_t version; 6152 uint32_t action_flag; 6153 struct jit_code_entry *relevant_entry; 6154 struct jit_code_entry *first_entry; 6155 }; 6156 6157 void __jit_debug_register_code(void) __attribute__((noinline)); 6158 void __jit_debug_register_code(void) 6159 { 6160 asm(""); 6161 } 6162 6163 /* Must statically initialize the version, because GDB may check 6164 the version before we can set it. */ 6165 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 }; 6166 6167 /* End GDB interface. */ 6168 6169 static int find_string(const char *strtab, const char *str) 6170 { 6171 const char *p = strtab + 1; 6172 6173 while (1) { 6174 if (strcmp(p, str) == 0) { 6175 return p - strtab; 6176 } 6177 p += strlen(p) + 1; 6178 } 6179 } 6180 6181 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size, 6182 const void *debug_frame, 6183 size_t debug_frame_size) 6184 { 6185 struct __attribute__((packed)) DebugInfo { 6186 uint32_t len; 6187 uint16_t version; 6188 uint32_t abbrev; 6189 uint8_t ptr_size; 6190 uint8_t cu_die; 6191 uint16_t cu_lang; 6192 uintptr_t cu_low_pc; 6193 uintptr_t cu_high_pc; 6194 uint8_t fn_die; 6195 char fn_name[16]; 6196 uintptr_t fn_low_pc; 6197 uintptr_t fn_high_pc; 6198 uint8_t cu_eoc; 6199 }; 6200 6201 struct ElfImage { 6202 ElfW(Ehdr) ehdr; 6203 ElfW(Phdr) phdr; 6204 ElfW(Shdr) shdr[7]; 6205 ElfW(Sym) sym[2]; 6206 struct DebugInfo di; 6207 uint8_t da[24]; 6208 char str[80]; 6209 }; 6210 6211 struct ElfImage *img; 6212 6213 static const struct ElfImage img_template = { 6214 .ehdr = { 6215 .e_ident[EI_MAG0] = ELFMAG0, 6216 .e_ident[EI_MAG1] = ELFMAG1, 6217 .e_ident[EI_MAG2] = ELFMAG2, 6218 .e_ident[EI_MAG3] = ELFMAG3, 6219 .e_ident[EI_CLASS] = ELF_CLASS, 6220 .e_ident[EI_DATA] = ELF_DATA, 6221 .e_ident[EI_VERSION] = EV_CURRENT, 6222 .e_type = ET_EXEC, 6223 .e_machine = ELF_HOST_MACHINE, 6224 .e_version = EV_CURRENT, 6225 .e_phoff = offsetof(struct ElfImage, phdr), 6226 .e_shoff = offsetof(struct ElfImage, shdr), 6227 .e_ehsize = sizeof(ElfW(Shdr)), 6228 .e_phentsize = sizeof(ElfW(Phdr)), 6229 .e_phnum = 1, 6230 .e_shentsize = sizeof(ElfW(Shdr)), 6231 .e_shnum = ARRAY_SIZE(img->shdr), 6232 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1, 6233 #ifdef ELF_HOST_FLAGS 6234 .e_flags = ELF_HOST_FLAGS, 6235 #endif 6236 #ifdef ELF_OSABI 6237 .e_ident[EI_OSABI] = ELF_OSABI, 6238 #endif 6239 }, 6240 .phdr = { 6241 .p_type = PT_LOAD, 6242 .p_flags = PF_X, 6243 }, 6244 .shdr = { 6245 [0] = { .sh_type = SHT_NULL }, 6246 /* Trick: The contents of code_gen_buffer are not present in 6247 this fake ELF file; that got allocated elsewhere. Therefore 6248 we mark .text as SHT_NOBITS (similar to .bss) so that readers 6249 will not look for contents. We can record any address. */ 6250 [1] = { /* .text */ 6251 .sh_type = SHT_NOBITS, 6252 .sh_flags = SHF_EXECINSTR | SHF_ALLOC, 6253 }, 6254 [2] = { /* .debug_info */ 6255 .sh_type = SHT_PROGBITS, 6256 .sh_offset = offsetof(struct ElfImage, di), 6257 .sh_size = sizeof(struct DebugInfo), 6258 }, 6259 [3] = { /* .debug_abbrev */ 6260 .sh_type = SHT_PROGBITS, 6261 .sh_offset = offsetof(struct ElfImage, da), 6262 .sh_size = sizeof(img->da), 6263 }, 6264 [4] = { /* .debug_frame */ 6265 .sh_type = SHT_PROGBITS, 6266 .sh_offset = sizeof(struct ElfImage), 6267 }, 6268 [5] = { /* .symtab */ 6269 .sh_type = SHT_SYMTAB, 6270 .sh_offset = offsetof(struct ElfImage, sym), 6271 .sh_size = sizeof(img->sym), 6272 .sh_info = 1, 6273 .sh_link = ARRAY_SIZE(img->shdr) - 1, 6274 .sh_entsize = sizeof(ElfW(Sym)), 6275 }, 6276 [6] = { /* .strtab */ 6277 .sh_type = SHT_STRTAB, 6278 .sh_offset = offsetof(struct ElfImage, str), 6279 .sh_size = sizeof(img->str), 6280 } 6281 }, 6282 .sym = { 6283 [1] = { /* code_gen_buffer */ 6284 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC), 6285 .st_shndx = 1, 6286 } 6287 }, 6288 .di = { 6289 .len = sizeof(struct DebugInfo) - 4, 6290 .version = 2, 6291 .ptr_size = sizeof(void *), 6292 .cu_die = 1, 6293 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */ 6294 .fn_die = 2, 6295 .fn_name = "code_gen_buffer" 6296 }, 6297 .da = { 6298 1, /* abbrev number (the cu) */ 6299 0x11, 1, /* DW_TAG_compile_unit, has children */ 6300 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */ 6301 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6302 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6303 0, 0, /* end of abbrev */ 6304 2, /* abbrev number (the fn) */ 6305 0x2e, 0, /* DW_TAG_subprogram, no children */ 6306 0x3, 0x8, /* DW_AT_name, DW_FORM_string */ 6307 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6308 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6309 0, 0, /* end of abbrev */ 6310 0 /* no more abbrev */ 6311 }, 6312 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0" 6313 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer", 6314 }; 6315 6316 /* We only need a single jit entry; statically allocate it. */ 6317 static struct jit_code_entry one_entry; 6318 6319 uintptr_t buf = (uintptr_t)buf_ptr; 6320 size_t img_size = sizeof(struct ElfImage) + debug_frame_size; 6321 DebugFrameHeader *dfh; 6322 6323 img = g_malloc(img_size); 6324 *img = img_template; 6325 6326 img->phdr.p_vaddr = buf; 6327 img->phdr.p_paddr = buf; 6328 img->phdr.p_memsz = buf_size; 6329 6330 img->shdr[1].sh_name = find_string(img->str, ".text"); 6331 img->shdr[1].sh_addr = buf; 6332 img->shdr[1].sh_size = buf_size; 6333 6334 img->shdr[2].sh_name = find_string(img->str, ".debug_info"); 6335 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev"); 6336 6337 img->shdr[4].sh_name = find_string(img->str, ".debug_frame"); 6338 img->shdr[4].sh_size = debug_frame_size; 6339 6340 img->shdr[5].sh_name = find_string(img->str, ".symtab"); 6341 img->shdr[6].sh_name = find_string(img->str, ".strtab"); 6342 6343 img->sym[1].st_name = find_string(img->str, "code_gen_buffer"); 6344 img->sym[1].st_value = buf; 6345 img->sym[1].st_size = buf_size; 6346 6347 img->di.cu_low_pc = buf; 6348 img->di.cu_high_pc = buf + buf_size; 6349 img->di.fn_low_pc = buf; 6350 img->di.fn_high_pc = buf + buf_size; 6351 6352 dfh = (DebugFrameHeader *)(img + 1); 6353 memcpy(dfh, debug_frame, debug_frame_size); 6354 dfh->fde.func_start = buf; 6355 dfh->fde.func_len = buf_size; 6356 6357 #ifdef DEBUG_JIT 6358 /* Enable this block to be able to debug the ELF image file creation. 6359 One can use readelf, objdump, or other inspection utilities. */ 6360 { 6361 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir()); 6362 FILE *f = fopen(jit, "w+b"); 6363 if (f) { 6364 if (fwrite(img, img_size, 1, f) != img_size) { 6365 /* Avoid stupid unused return value warning for fwrite. */ 6366 } 6367 fclose(f); 6368 } 6369 } 6370 #endif 6371 6372 one_entry.symfile_addr = img; 6373 one_entry.symfile_size = img_size; 6374 6375 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN; 6376 __jit_debug_descriptor.relevant_entry = &one_entry; 6377 __jit_debug_descriptor.first_entry = &one_entry; 6378 __jit_debug_register_code(); 6379 } 6380 #else 6381 /* No support for the feature. Provide the entry point expected by exec.c, 6382 and implement the internal function we declared earlier. */ 6383 6384 static void tcg_register_jit_int(const void *buf, size_t size, 6385 const void *debug_frame, 6386 size_t debug_frame_size) 6387 { 6388 } 6389 6390 void tcg_register_jit(const void *buf, size_t buf_size) 6391 { 6392 } 6393 #endif /* ELF_HOST_MACHINE */ 6394 6395 #if !TCG_TARGET_MAYBE_vec 6396 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...) 6397 { 6398 g_assert_not_reached(); 6399 } 6400 #endif 6401