1 /* 2 * Tiny Code Generator for QEMU 3 * 4 * Copyright (c) 2008 Fabrice Bellard 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to deal 8 * in the Software without restriction, including without limitation the rights 9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 10 * copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 22 * THE SOFTWARE. 23 */ 24 25 #include "qemu/osdep.h" 26 27 /* Define to jump the ELF file used to communicate with GDB. */ 28 #undef DEBUG_JIT 29 30 #include "qemu/error-report.h" 31 #include "qemu/cutils.h" 32 #include "qemu/host-utils.h" 33 #include "qemu/qemu-print.h" 34 #include "qemu/cacheflush.h" 35 #include "qemu/cacheinfo.h" 36 #include "qemu/timer.h" 37 #include "exec/translation-block.h" 38 #include "exec/tlb-common.h" 39 #include "tcg/startup.h" 40 #include "tcg/tcg-op-common.h" 41 42 #if UINTPTR_MAX == UINT32_MAX 43 # define ELF_CLASS ELFCLASS32 44 #else 45 # define ELF_CLASS ELFCLASS64 46 #endif 47 #if HOST_BIG_ENDIAN 48 # define ELF_DATA ELFDATA2MSB 49 #else 50 # define ELF_DATA ELFDATA2LSB 51 #endif 52 53 #include "elf.h" 54 #include "exec/log.h" 55 #include "tcg/tcg-ldst.h" 56 #include "tcg/tcg-temp-internal.h" 57 #include "tcg-internal.h" 58 #include "tcg/perf.h" 59 #ifdef CONFIG_USER_ONLY 60 #include "user/guest-base.h" 61 #endif 62 63 /* Forward declarations for functions declared in tcg-target.c.inc and 64 used here. */ 65 static void tcg_target_init(TCGContext *s); 66 static void tcg_target_qemu_prologue(TCGContext *s); 67 static bool patch_reloc(tcg_insn_unit *code_ptr, int type, 68 intptr_t value, intptr_t addend); 69 70 /* The CIE and FDE header definitions will be common to all hosts. */ 71 typedef struct { 72 uint32_t len __attribute__((aligned((sizeof(void *))))); 73 uint32_t id; 74 uint8_t version; 75 char augmentation[1]; 76 uint8_t code_align; 77 uint8_t data_align; 78 uint8_t return_column; 79 } DebugFrameCIE; 80 81 typedef struct QEMU_PACKED { 82 uint32_t len __attribute__((aligned((sizeof(void *))))); 83 uint32_t cie_offset; 84 uintptr_t func_start; 85 uintptr_t func_len; 86 } DebugFrameFDEHeader; 87 88 typedef struct QEMU_PACKED { 89 DebugFrameCIE cie; 90 DebugFrameFDEHeader fde; 91 } DebugFrameHeader; 92 93 typedef struct TCGLabelQemuLdst { 94 bool is_ld; /* qemu_ld: true, qemu_st: false */ 95 MemOpIdx oi; 96 TCGType type; /* result type of a load */ 97 TCGReg addrlo_reg; /* reg index for low word of guest virtual addr */ 98 TCGReg addrhi_reg; /* reg index for high word of guest virtual addr */ 99 TCGReg datalo_reg; /* reg index for low word to be loaded or stored */ 100 TCGReg datahi_reg; /* reg index for high word to be loaded or stored */ 101 const tcg_insn_unit *raddr; /* addr of the next IR of qemu_ld/st IR */ 102 tcg_insn_unit *label_ptr[2]; /* label pointers to be updated */ 103 QSIMPLEQ_ENTRY(TCGLabelQemuLdst) next; 104 } TCGLabelQemuLdst; 105 106 static void tcg_register_jit_int(const void *buf, size_t size, 107 const void *debug_frame, 108 size_t debug_frame_size) 109 __attribute__((unused)); 110 111 /* Forward declarations for functions declared and used in tcg-target.c.inc. */ 112 static void tcg_out_tb_start(TCGContext *s); 113 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg1, 114 intptr_t arg2); 115 static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 116 static void tcg_out_movi(TCGContext *s, TCGType type, 117 TCGReg ret, tcg_target_long arg); 118 static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 119 static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg); 120 static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg); 121 static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg); 122 static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg); 123 static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg); 124 static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 125 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg); 126 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg); 127 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long); 128 static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2); 129 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg); 130 static void tcg_out_goto_tb(TCGContext *s, int which); 131 static void tcg_out_op(TCGContext *s, TCGOpcode opc, 132 const TCGArg args[TCG_MAX_OP_ARGS], 133 const int const_args[TCG_MAX_OP_ARGS]); 134 #if TCG_TARGET_MAYBE_vec 135 static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 136 TCGReg dst, TCGReg src); 137 static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 138 TCGReg dst, TCGReg base, intptr_t offset); 139 static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 140 TCGReg dst, int64_t arg); 141 static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 142 unsigned vecl, unsigned vece, 143 const TCGArg args[TCG_MAX_OP_ARGS], 144 const int const_args[TCG_MAX_OP_ARGS]); 145 #else 146 static inline bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece, 147 TCGReg dst, TCGReg src) 148 { 149 g_assert_not_reached(); 150 } 151 static inline bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece, 152 TCGReg dst, TCGReg base, intptr_t offset) 153 { 154 g_assert_not_reached(); 155 } 156 static inline void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece, 157 TCGReg dst, int64_t arg) 158 { 159 g_assert_not_reached(); 160 } 161 static inline void tcg_out_vec_op(TCGContext *s, TCGOpcode opc, 162 unsigned vecl, unsigned vece, 163 const TCGArg args[TCG_MAX_OP_ARGS], 164 const int const_args[TCG_MAX_OP_ARGS]) 165 { 166 g_assert_not_reached(); 167 } 168 #endif 169 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, TCGReg arg1, 170 intptr_t arg2); 171 static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, 172 TCGReg base, intptr_t ofs); 173 static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target, 174 const TCGHelperInfo *info); 175 static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot); 176 static bool tcg_target_const_match(int64_t val, int ct, 177 TCGType type, TCGCond cond, int vece); 178 #ifdef TCG_TARGET_NEED_LDST_LABELS 179 static int tcg_out_ldst_finalize(TCGContext *s); 180 #endif 181 182 #ifndef CONFIG_USER_ONLY 183 #define guest_base ({ qemu_build_not_reached(); (uintptr_t)0; }) 184 #endif 185 186 typedef struct TCGLdstHelperParam { 187 TCGReg (*ra_gen)(TCGContext *s, const TCGLabelQemuLdst *l, int arg_reg); 188 unsigned ntmp; 189 int tmp[3]; 190 } TCGLdstHelperParam; 191 192 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 193 const TCGLdstHelperParam *p) 194 __attribute__((unused)); 195 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *l, 196 bool load_sign, const TCGLdstHelperParam *p) 197 __attribute__((unused)); 198 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *l, 199 const TCGLdstHelperParam *p) 200 __attribute__((unused)); 201 202 static void * const qemu_ld_helpers[MO_SSIZE + 1] __attribute__((unused)) = { 203 [MO_UB] = helper_ldub_mmu, 204 [MO_SB] = helper_ldsb_mmu, 205 [MO_UW] = helper_lduw_mmu, 206 [MO_SW] = helper_ldsw_mmu, 207 [MO_UL] = helper_ldul_mmu, 208 [MO_UQ] = helper_ldq_mmu, 209 #if TCG_TARGET_REG_BITS == 64 210 [MO_SL] = helper_ldsl_mmu, 211 [MO_128] = helper_ld16_mmu, 212 #endif 213 }; 214 215 static void * const qemu_st_helpers[MO_SIZE + 1] __attribute__((unused)) = { 216 [MO_8] = helper_stb_mmu, 217 [MO_16] = helper_stw_mmu, 218 [MO_32] = helper_stl_mmu, 219 [MO_64] = helper_stq_mmu, 220 #if TCG_TARGET_REG_BITS == 64 221 [MO_128] = helper_st16_mmu, 222 #endif 223 }; 224 225 typedef struct { 226 MemOp atom; /* lg2 bits of atomicity required */ 227 MemOp align; /* lg2 bits of alignment to use */ 228 } TCGAtomAlign; 229 230 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 231 MemOp host_atom, bool allow_two_ops) 232 __attribute__((unused)); 233 234 #ifdef CONFIG_USER_ONLY 235 bool tcg_use_softmmu; 236 #endif 237 238 TCGContext tcg_init_ctx; 239 __thread TCGContext *tcg_ctx; 240 241 TCGContext **tcg_ctxs; 242 unsigned int tcg_cur_ctxs; 243 unsigned int tcg_max_ctxs; 244 TCGv_env tcg_env; 245 const void *tcg_code_gen_epilogue; 246 uintptr_t tcg_splitwx_diff; 247 248 #ifndef CONFIG_TCG_INTERPRETER 249 tcg_prologue_fn *tcg_qemu_tb_exec; 250 #endif 251 252 static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT]; 253 static TCGRegSet tcg_target_call_clobber_regs; 254 255 #if TCG_TARGET_INSN_UNIT_SIZE == 1 256 static __attribute__((unused)) inline void tcg_out8(TCGContext *s, uint8_t v) 257 { 258 *s->code_ptr++ = v; 259 } 260 261 static __attribute__((unused)) inline void tcg_patch8(tcg_insn_unit *p, 262 uint8_t v) 263 { 264 *p = v; 265 } 266 #endif 267 268 #if TCG_TARGET_INSN_UNIT_SIZE <= 2 269 static __attribute__((unused)) inline void tcg_out16(TCGContext *s, uint16_t v) 270 { 271 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 272 *s->code_ptr++ = v; 273 } else { 274 tcg_insn_unit *p = s->code_ptr; 275 memcpy(p, &v, sizeof(v)); 276 s->code_ptr = p + (2 / TCG_TARGET_INSN_UNIT_SIZE); 277 } 278 } 279 280 static __attribute__((unused)) inline void tcg_patch16(tcg_insn_unit *p, 281 uint16_t v) 282 { 283 if (TCG_TARGET_INSN_UNIT_SIZE == 2) { 284 *p = v; 285 } else { 286 memcpy(p, &v, sizeof(v)); 287 } 288 } 289 #endif 290 291 #if TCG_TARGET_INSN_UNIT_SIZE <= 4 292 static __attribute__((unused)) inline void tcg_out32(TCGContext *s, uint32_t v) 293 { 294 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 295 *s->code_ptr++ = v; 296 } else { 297 tcg_insn_unit *p = s->code_ptr; 298 memcpy(p, &v, sizeof(v)); 299 s->code_ptr = p + (4 / TCG_TARGET_INSN_UNIT_SIZE); 300 } 301 } 302 303 static __attribute__((unused)) inline void tcg_patch32(tcg_insn_unit *p, 304 uint32_t v) 305 { 306 if (TCG_TARGET_INSN_UNIT_SIZE == 4) { 307 *p = v; 308 } else { 309 memcpy(p, &v, sizeof(v)); 310 } 311 } 312 #endif 313 314 #if TCG_TARGET_INSN_UNIT_SIZE <= 8 315 static __attribute__((unused)) inline void tcg_out64(TCGContext *s, uint64_t v) 316 { 317 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 318 *s->code_ptr++ = v; 319 } else { 320 tcg_insn_unit *p = s->code_ptr; 321 memcpy(p, &v, sizeof(v)); 322 s->code_ptr = p + (8 / TCG_TARGET_INSN_UNIT_SIZE); 323 } 324 } 325 326 static __attribute__((unused)) inline void tcg_patch64(tcg_insn_unit *p, 327 uint64_t v) 328 { 329 if (TCG_TARGET_INSN_UNIT_SIZE == 8) { 330 *p = v; 331 } else { 332 memcpy(p, &v, sizeof(v)); 333 } 334 } 335 #endif 336 337 /* label relocation processing */ 338 339 static void tcg_out_reloc(TCGContext *s, tcg_insn_unit *code_ptr, int type, 340 TCGLabel *l, intptr_t addend) 341 { 342 TCGRelocation *r = tcg_malloc(sizeof(TCGRelocation)); 343 344 r->type = type; 345 r->ptr = code_ptr; 346 r->addend = addend; 347 QSIMPLEQ_INSERT_TAIL(&l->relocs, r, next); 348 } 349 350 static void tcg_out_label(TCGContext *s, TCGLabel *l) 351 { 352 tcg_debug_assert(!l->has_value); 353 l->has_value = 1; 354 l->u.value_ptr = tcg_splitwx_to_rx(s->code_ptr); 355 } 356 357 TCGLabel *gen_new_label(void) 358 { 359 TCGContext *s = tcg_ctx; 360 TCGLabel *l = tcg_malloc(sizeof(TCGLabel)); 361 362 memset(l, 0, sizeof(TCGLabel)); 363 l->id = s->nb_labels++; 364 QSIMPLEQ_INIT(&l->branches); 365 QSIMPLEQ_INIT(&l->relocs); 366 367 QSIMPLEQ_INSERT_TAIL(&s->labels, l, next); 368 369 return l; 370 } 371 372 static bool tcg_resolve_relocs(TCGContext *s) 373 { 374 TCGLabel *l; 375 376 QSIMPLEQ_FOREACH(l, &s->labels, next) { 377 TCGRelocation *r; 378 uintptr_t value = l->u.value; 379 380 QSIMPLEQ_FOREACH(r, &l->relocs, next) { 381 if (!patch_reloc(r->ptr, r->type, value, r->addend)) { 382 return false; 383 } 384 } 385 } 386 return true; 387 } 388 389 static void set_jmp_reset_offset(TCGContext *s, int which) 390 { 391 /* 392 * We will check for overflow at the end of the opcode loop in 393 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 394 */ 395 s->gen_tb->jmp_reset_offset[which] = tcg_current_code_size(s); 396 } 397 398 static void G_GNUC_UNUSED set_jmp_insn_offset(TCGContext *s, int which) 399 { 400 /* 401 * We will check for overflow at the end of the opcode loop in 402 * tcg_gen_code, where we bound tcg_current_code_size to UINT16_MAX. 403 */ 404 s->gen_tb->jmp_insn_offset[which] = tcg_current_code_size(s); 405 } 406 407 static uintptr_t G_GNUC_UNUSED get_jmp_target_addr(TCGContext *s, int which) 408 { 409 /* 410 * Return the read-execute version of the pointer, for the benefit 411 * of any pc-relative addressing mode. 412 */ 413 return (uintptr_t)tcg_splitwx_to_rx(&s->gen_tb->jmp_target_addr[which]); 414 } 415 416 static int __attribute__((unused)) 417 tlb_mask_table_ofs(TCGContext *s, int which) 418 { 419 return (offsetof(CPUNegativeOffsetState, tlb.f[which]) - 420 sizeof(CPUNegativeOffsetState)); 421 } 422 423 /* Signal overflow, starting over with fewer guest insns. */ 424 static G_NORETURN 425 void tcg_raise_tb_overflow(TCGContext *s) 426 { 427 siglongjmp(s->jmp_trans, -2); 428 } 429 430 /* 431 * Used by tcg_out_movext{1,2} to hold the arguments for tcg_out_movext. 432 * By the time we arrive at tcg_out_movext1, @dst is always a TCGReg. 433 * 434 * However, tcg_out_helper_load_slots reuses this field to hold an 435 * argument slot number (which may designate a argument register or an 436 * argument stack slot), converting to TCGReg once all arguments that 437 * are destined for the stack are processed. 438 */ 439 typedef struct TCGMovExtend { 440 unsigned dst; 441 TCGReg src; 442 TCGType dst_type; 443 TCGType src_type; 444 MemOp src_ext; 445 } TCGMovExtend; 446 447 /** 448 * tcg_out_movext -- move and extend 449 * @s: tcg context 450 * @dst_type: integral type for destination 451 * @dst: destination register 452 * @src_type: integral type for source 453 * @src_ext: extension to apply to source 454 * @src: source register 455 * 456 * Move or extend @src into @dst, depending on @src_ext and the types. 457 */ 458 static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst, 459 TCGType src_type, MemOp src_ext, TCGReg src) 460 { 461 switch (src_ext) { 462 case MO_UB: 463 tcg_out_ext8u(s, dst, src); 464 break; 465 case MO_SB: 466 tcg_out_ext8s(s, dst_type, dst, src); 467 break; 468 case MO_UW: 469 tcg_out_ext16u(s, dst, src); 470 break; 471 case MO_SW: 472 tcg_out_ext16s(s, dst_type, dst, src); 473 break; 474 case MO_UL: 475 case MO_SL: 476 if (dst_type == TCG_TYPE_I32) { 477 if (src_type == TCG_TYPE_I32) { 478 tcg_out_mov(s, TCG_TYPE_I32, dst, src); 479 } else { 480 tcg_out_extrl_i64_i32(s, dst, src); 481 } 482 } else if (src_type == TCG_TYPE_I32) { 483 if (src_ext & MO_SIGN) { 484 tcg_out_exts_i32_i64(s, dst, src); 485 } else { 486 tcg_out_extu_i32_i64(s, dst, src); 487 } 488 } else { 489 if (src_ext & MO_SIGN) { 490 tcg_out_ext32s(s, dst, src); 491 } else { 492 tcg_out_ext32u(s, dst, src); 493 } 494 } 495 break; 496 case MO_UQ: 497 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 498 if (dst_type == TCG_TYPE_I32) { 499 tcg_out_extrl_i64_i32(s, dst, src); 500 } else { 501 tcg_out_mov(s, TCG_TYPE_I64, dst, src); 502 } 503 break; 504 default: 505 g_assert_not_reached(); 506 } 507 } 508 509 /* Minor variations on a theme, using a structure. */ 510 static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i, 511 TCGReg src) 512 { 513 tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src); 514 } 515 516 static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i) 517 { 518 tcg_out_movext1_new_src(s, i, i->src); 519 } 520 521 /** 522 * tcg_out_movext2 -- move and extend two pair 523 * @s: tcg context 524 * @i1: first move description 525 * @i2: second move description 526 * @scratch: temporary register, or -1 for none 527 * 528 * As tcg_out_movext, for both @i1 and @i2, caring for overlap 529 * between the sources and destinations. 530 */ 531 532 static void tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1, 533 const TCGMovExtend *i2, int scratch) 534 { 535 TCGReg src1 = i1->src; 536 TCGReg src2 = i2->src; 537 538 if (i1->dst != src2) { 539 tcg_out_movext1(s, i1); 540 tcg_out_movext1(s, i2); 541 return; 542 } 543 if (i2->dst == src1) { 544 TCGType src1_type = i1->src_type; 545 TCGType src2_type = i2->src_type; 546 547 if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) { 548 /* The data is now in the correct registers, now extend. */ 549 src1 = i2->src; 550 src2 = i1->src; 551 } else { 552 tcg_debug_assert(scratch >= 0); 553 tcg_out_mov(s, src1_type, scratch, src1); 554 src1 = scratch; 555 } 556 } 557 tcg_out_movext1_new_src(s, i2, src2); 558 tcg_out_movext1_new_src(s, i1, src1); 559 } 560 561 /** 562 * tcg_out_movext3 -- move and extend three pair 563 * @s: tcg context 564 * @i1: first move description 565 * @i2: second move description 566 * @i3: third move description 567 * @scratch: temporary register, or -1 for none 568 * 569 * As tcg_out_movext, for all of @i1, @i2 and @i3, caring for overlap 570 * between the sources and destinations. 571 */ 572 573 static void tcg_out_movext3(TCGContext *s, const TCGMovExtend *i1, 574 const TCGMovExtend *i2, const TCGMovExtend *i3, 575 int scratch) 576 { 577 TCGReg src1 = i1->src; 578 TCGReg src2 = i2->src; 579 TCGReg src3 = i3->src; 580 581 if (i1->dst != src2 && i1->dst != src3) { 582 tcg_out_movext1(s, i1); 583 tcg_out_movext2(s, i2, i3, scratch); 584 return; 585 } 586 if (i2->dst != src1 && i2->dst != src3) { 587 tcg_out_movext1(s, i2); 588 tcg_out_movext2(s, i1, i3, scratch); 589 return; 590 } 591 if (i3->dst != src1 && i3->dst != src2) { 592 tcg_out_movext1(s, i3); 593 tcg_out_movext2(s, i1, i2, scratch); 594 return; 595 } 596 597 /* 598 * There is a cycle. Since there are only 3 nodes, the cycle is 599 * either "clockwise" or "anti-clockwise", and can be solved with 600 * a single scratch or two xchg. 601 */ 602 if (i1->dst == src2 && i2->dst == src3 && i3->dst == src1) { 603 /* "Clockwise" */ 604 if (tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2)) { 605 tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3); 606 /* The data is now in the correct registers, now extend. */ 607 tcg_out_movext1_new_src(s, i1, i1->dst); 608 tcg_out_movext1_new_src(s, i2, i2->dst); 609 tcg_out_movext1_new_src(s, i3, i3->dst); 610 } else { 611 tcg_debug_assert(scratch >= 0); 612 tcg_out_mov(s, i1->src_type, scratch, src1); 613 tcg_out_movext1(s, i3); 614 tcg_out_movext1(s, i2); 615 tcg_out_movext1_new_src(s, i1, scratch); 616 } 617 } else if (i1->dst == src3 && i2->dst == src1 && i3->dst == src2) { 618 /* "Anti-clockwise" */ 619 if (tcg_out_xchg(s, MAX(i2->src_type, i3->src_type), src2, src3)) { 620 tcg_out_xchg(s, MAX(i1->src_type, i2->src_type), src1, src2); 621 /* The data is now in the correct registers, now extend. */ 622 tcg_out_movext1_new_src(s, i1, i1->dst); 623 tcg_out_movext1_new_src(s, i2, i2->dst); 624 tcg_out_movext1_new_src(s, i3, i3->dst); 625 } else { 626 tcg_debug_assert(scratch >= 0); 627 tcg_out_mov(s, i1->src_type, scratch, src1); 628 tcg_out_movext1(s, i2); 629 tcg_out_movext1(s, i3); 630 tcg_out_movext1_new_src(s, i1, scratch); 631 } 632 } else { 633 g_assert_not_reached(); 634 } 635 } 636 637 #define C_PFX1(P, A) P##A 638 #define C_PFX2(P, A, B) P##A##_##B 639 #define C_PFX3(P, A, B, C) P##A##_##B##_##C 640 #define C_PFX4(P, A, B, C, D) P##A##_##B##_##C##_##D 641 #define C_PFX5(P, A, B, C, D, E) P##A##_##B##_##C##_##D##_##E 642 #define C_PFX6(P, A, B, C, D, E, F) P##A##_##B##_##C##_##D##_##E##_##F 643 644 /* Define an enumeration for the various combinations. */ 645 646 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1), 647 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2), 648 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3), 649 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4), 650 651 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1), 652 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2), 653 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3), 654 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4), 655 656 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2), 657 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1), 658 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1), 659 660 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1), 661 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2), 662 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3), 663 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4), 664 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4), 665 666 typedef enum { 667 #include "tcg-target-con-set.h" 668 } TCGConstraintSetIndex; 669 670 static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode); 671 672 #undef C_O0_I1 673 #undef C_O0_I2 674 #undef C_O0_I3 675 #undef C_O0_I4 676 #undef C_O1_I1 677 #undef C_O1_I2 678 #undef C_O1_I3 679 #undef C_O1_I4 680 #undef C_N1_I2 681 #undef C_N1O1_I1 682 #undef C_N2_I1 683 #undef C_O2_I1 684 #undef C_O2_I2 685 #undef C_O2_I3 686 #undef C_O2_I4 687 #undef C_N1_O1_I4 688 689 /* Put all of the constraint sets into an array, indexed by the enum. */ 690 691 #define C_O0_I1(I1) { .args_ct_str = { #I1 } }, 692 #define C_O0_I2(I1, I2) { .args_ct_str = { #I1, #I2 } }, 693 #define C_O0_I3(I1, I2, I3) { .args_ct_str = { #I1, #I2, #I3 } }, 694 #define C_O0_I4(I1, I2, I3, I4) { .args_ct_str = { #I1, #I2, #I3, #I4 } }, 695 696 #define C_O1_I1(O1, I1) { .args_ct_str = { #O1, #I1 } }, 697 #define C_O1_I2(O1, I1, I2) { .args_ct_str = { #O1, #I1, #I2 } }, 698 #define C_O1_I3(O1, I1, I2, I3) { .args_ct_str = { #O1, #I1, #I2, #I3 } }, 699 #define C_O1_I4(O1, I1, I2, I3, I4) { .args_ct_str = { #O1, #I1, #I2, #I3, #I4 } }, 700 701 #define C_N1_I2(O1, I1, I2) { .args_ct_str = { "&" #O1, #I1, #I2 } }, 702 #define C_N1O1_I1(O1, O2, I1) { .args_ct_str = { "&" #O1, #O2, #I1 } }, 703 #define C_N2_I1(O1, O2, I1) { .args_ct_str = { "&" #O1, "&" #O2, #I1 } }, 704 705 #define C_O2_I1(O1, O2, I1) { .args_ct_str = { #O1, #O2, #I1 } }, 706 #define C_O2_I2(O1, O2, I1, I2) { .args_ct_str = { #O1, #O2, #I1, #I2 } }, 707 #define C_O2_I3(O1, O2, I1, I2, I3) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3 } }, 708 #define C_O2_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { #O1, #O2, #I1, #I2, #I3, #I4 } }, 709 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) { .args_ct_str = { "&" #O1, #O2, #I1, #I2, #I3, #I4 } }, 710 711 static const TCGTargetOpDef constraint_sets[] = { 712 #include "tcg-target-con-set.h" 713 }; 714 715 716 #undef C_O0_I1 717 #undef C_O0_I2 718 #undef C_O0_I3 719 #undef C_O0_I4 720 #undef C_O1_I1 721 #undef C_O1_I2 722 #undef C_O1_I3 723 #undef C_O1_I4 724 #undef C_N1_I2 725 #undef C_N1O1_I1 726 #undef C_N2_I1 727 #undef C_O2_I1 728 #undef C_O2_I2 729 #undef C_O2_I3 730 #undef C_O2_I4 731 #undef C_N1_O1_I4 732 733 /* Expand the enumerator to be returned from tcg_target_op_def(). */ 734 735 #define C_O0_I1(I1) C_PFX1(c_o0_i1_, I1) 736 #define C_O0_I2(I1, I2) C_PFX2(c_o0_i2_, I1, I2) 737 #define C_O0_I3(I1, I2, I3) C_PFX3(c_o0_i3_, I1, I2, I3) 738 #define C_O0_I4(I1, I2, I3, I4) C_PFX4(c_o0_i4_, I1, I2, I3, I4) 739 740 #define C_O1_I1(O1, I1) C_PFX2(c_o1_i1_, O1, I1) 741 #define C_O1_I2(O1, I1, I2) C_PFX3(c_o1_i2_, O1, I1, I2) 742 #define C_O1_I3(O1, I1, I2, I3) C_PFX4(c_o1_i3_, O1, I1, I2, I3) 743 #define C_O1_I4(O1, I1, I2, I3, I4) C_PFX5(c_o1_i4_, O1, I1, I2, I3, I4) 744 745 #define C_N1_I2(O1, I1, I2) C_PFX3(c_n1_i2_, O1, I1, I2) 746 #define C_N1O1_I1(O1, O2, I1) C_PFX3(c_n1o1_i1_, O1, O2, I1) 747 #define C_N2_I1(O1, O2, I1) C_PFX3(c_n2_i1_, O1, O2, I1) 748 749 #define C_O2_I1(O1, O2, I1) C_PFX3(c_o2_i1_, O1, O2, I1) 750 #define C_O2_I2(O1, O2, I1, I2) C_PFX4(c_o2_i2_, O1, O2, I1, I2) 751 #define C_O2_I3(O1, O2, I1, I2, I3) C_PFX5(c_o2_i3_, O1, O2, I1, I2, I3) 752 #define C_O2_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_o2_i4_, O1, O2, I1, I2, I3, I4) 753 #define C_N1_O1_I4(O1, O2, I1, I2, I3, I4) C_PFX6(c_n1_o1_i4_, O1, O2, I1, I2, I3, I4) 754 755 #include "tcg-target.c.inc" 756 757 #ifndef CONFIG_TCG_INTERPRETER 758 /* Validate CPUTLBDescFast placement. */ 759 QEMU_BUILD_BUG_ON((int)(offsetof(CPUNegativeOffsetState, tlb.f[0]) - 760 sizeof(CPUNegativeOffsetState)) 761 < MIN_TLB_MASK_TABLE_OFS); 762 #endif 763 764 /* 765 * All TCG threads except the parent (i.e. the one that called tcg_context_init 766 * and registered the target's TCG globals) must register with this function 767 * before initiating translation. 768 * 769 * In user-mode we just point tcg_ctx to tcg_init_ctx. See the documentation 770 * of tcg_region_init() for the reasoning behind this. 771 * 772 * In system-mode each caller registers its context in tcg_ctxs[]. Note that in 773 * system-mode tcg_ctxs[] does not track tcg_ctx_init, since the initial context 774 * is not used anymore for translation once this function is called. 775 * 776 * Not tracking tcg_init_ctx in tcg_ctxs[] in system-mode keeps code that 777 * iterates over the array (e.g. tcg_code_size() the same for both system/user 778 * modes. 779 */ 780 #ifdef CONFIG_USER_ONLY 781 void tcg_register_thread(void) 782 { 783 tcg_ctx = &tcg_init_ctx; 784 } 785 #else 786 void tcg_register_thread(void) 787 { 788 TCGContext *s = g_malloc(sizeof(*s)); 789 unsigned int i, n; 790 791 *s = tcg_init_ctx; 792 793 /* Relink mem_base. */ 794 for (i = 0, n = tcg_init_ctx.nb_globals; i < n; ++i) { 795 if (tcg_init_ctx.temps[i].mem_base) { 796 ptrdiff_t b = tcg_init_ctx.temps[i].mem_base - tcg_init_ctx.temps; 797 tcg_debug_assert(b >= 0 && b < n); 798 s->temps[i].mem_base = &s->temps[b]; 799 } 800 } 801 802 /* Claim an entry in tcg_ctxs */ 803 n = qatomic_fetch_inc(&tcg_cur_ctxs); 804 g_assert(n < tcg_max_ctxs); 805 qatomic_set(&tcg_ctxs[n], s); 806 807 if (n > 0) { 808 tcg_region_initial_alloc(s); 809 } 810 811 tcg_ctx = s; 812 } 813 #endif /* !CONFIG_USER_ONLY */ 814 815 /* pool based memory allocation */ 816 void *tcg_malloc_internal(TCGContext *s, int size) 817 { 818 TCGPool *p; 819 int pool_size; 820 821 if (size > TCG_POOL_CHUNK_SIZE) { 822 /* big malloc: insert a new pool (XXX: could optimize) */ 823 p = g_malloc(sizeof(TCGPool) + size); 824 p->size = size; 825 p->next = s->pool_first_large; 826 s->pool_first_large = p; 827 return p->data; 828 } else { 829 p = s->pool_current; 830 if (!p) { 831 p = s->pool_first; 832 if (!p) 833 goto new_pool; 834 } else { 835 if (!p->next) { 836 new_pool: 837 pool_size = TCG_POOL_CHUNK_SIZE; 838 p = g_malloc(sizeof(TCGPool) + pool_size); 839 p->size = pool_size; 840 p->next = NULL; 841 if (s->pool_current) { 842 s->pool_current->next = p; 843 } else { 844 s->pool_first = p; 845 } 846 } else { 847 p = p->next; 848 } 849 } 850 } 851 s->pool_current = p; 852 s->pool_cur = p->data + size; 853 s->pool_end = p->data + p->size; 854 return p->data; 855 } 856 857 void tcg_pool_reset(TCGContext *s) 858 { 859 TCGPool *p, *t; 860 for (p = s->pool_first_large; p; p = t) { 861 t = p->next; 862 g_free(p); 863 } 864 s->pool_first_large = NULL; 865 s->pool_cur = s->pool_end = NULL; 866 s->pool_current = NULL; 867 } 868 869 /* 870 * Create TCGHelperInfo structures for "tcg/tcg-ldst.h" functions, 871 * akin to what "exec/helper-tcg.h" does with DEF_HELPER_FLAGS_N. 872 * We only use these for layout in tcg_out_ld_helper_ret and 873 * tcg_out_st_helper_args, and share them between several of 874 * the helpers, with the end result that it's easier to build manually. 875 */ 876 877 #if TCG_TARGET_REG_BITS == 32 878 # define dh_typecode_ttl dh_typecode_i32 879 #else 880 # define dh_typecode_ttl dh_typecode_i64 881 #endif 882 883 static TCGHelperInfo info_helper_ld32_mmu = { 884 .flags = TCG_CALL_NO_WG, 885 .typemask = dh_typemask(ttl, 0) /* return tcg_target_ulong */ 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_ld64_mmu = { 893 .flags = TCG_CALL_NO_WG, 894 .typemask = dh_typemask(i64, 0) /* return uint64_t */ 895 | dh_typemask(env, 1) 896 | dh_typemask(i64, 2) /* uint64_t addr */ 897 | dh_typemask(i32, 3) /* unsigned oi */ 898 | dh_typemask(ptr, 4) /* uintptr_t ra */ 899 }; 900 901 static TCGHelperInfo info_helper_ld128_mmu = { 902 .flags = TCG_CALL_NO_WG, 903 .typemask = dh_typemask(i128, 0) /* return Int128 */ 904 | dh_typemask(env, 1) 905 | dh_typemask(i64, 2) /* uint64_t addr */ 906 | dh_typemask(i32, 3) /* unsigned oi */ 907 | dh_typemask(ptr, 4) /* uintptr_t ra */ 908 }; 909 910 static TCGHelperInfo info_helper_st32_mmu = { 911 .flags = TCG_CALL_NO_WG, 912 .typemask = dh_typemask(void, 0) 913 | dh_typemask(env, 1) 914 | dh_typemask(i64, 2) /* uint64_t addr */ 915 | dh_typemask(i32, 3) /* uint32_t data */ 916 | dh_typemask(i32, 4) /* unsigned oi */ 917 | dh_typemask(ptr, 5) /* uintptr_t ra */ 918 }; 919 920 static TCGHelperInfo info_helper_st64_mmu = { 921 .flags = TCG_CALL_NO_WG, 922 .typemask = dh_typemask(void, 0) 923 | dh_typemask(env, 1) 924 | dh_typemask(i64, 2) /* uint64_t addr */ 925 | dh_typemask(i64, 3) /* uint64_t data */ 926 | dh_typemask(i32, 4) /* unsigned oi */ 927 | dh_typemask(ptr, 5) /* uintptr_t ra */ 928 }; 929 930 static TCGHelperInfo info_helper_st128_mmu = { 931 .flags = TCG_CALL_NO_WG, 932 .typemask = dh_typemask(void, 0) 933 | dh_typemask(env, 1) 934 | dh_typemask(i64, 2) /* uint64_t addr */ 935 | dh_typemask(i128, 3) /* Int128 data */ 936 | dh_typemask(i32, 4) /* unsigned oi */ 937 | dh_typemask(ptr, 5) /* uintptr_t ra */ 938 }; 939 940 #ifdef CONFIG_TCG_INTERPRETER 941 static ffi_type *typecode_to_ffi(int argmask) 942 { 943 /* 944 * libffi does not support __int128_t, so we have forced Int128 945 * to use the structure definition instead of the builtin type. 946 */ 947 static ffi_type *ffi_type_i128_elements[3] = { 948 &ffi_type_uint64, 949 &ffi_type_uint64, 950 NULL 951 }; 952 static ffi_type ffi_type_i128 = { 953 .size = 16, 954 .alignment = __alignof__(Int128), 955 .type = FFI_TYPE_STRUCT, 956 .elements = ffi_type_i128_elements, 957 }; 958 959 switch (argmask) { 960 case dh_typecode_void: 961 return &ffi_type_void; 962 case dh_typecode_i32: 963 return &ffi_type_uint32; 964 case dh_typecode_s32: 965 return &ffi_type_sint32; 966 case dh_typecode_i64: 967 return &ffi_type_uint64; 968 case dh_typecode_s64: 969 return &ffi_type_sint64; 970 case dh_typecode_ptr: 971 return &ffi_type_pointer; 972 case dh_typecode_i128: 973 return &ffi_type_i128; 974 } 975 g_assert_not_reached(); 976 } 977 978 static ffi_cif *init_ffi_layout(TCGHelperInfo *info) 979 { 980 unsigned typemask = info->typemask; 981 struct { 982 ffi_cif cif; 983 ffi_type *args[]; 984 } *ca; 985 ffi_status status; 986 int nargs; 987 988 /* Ignoring the return type, find the last non-zero field. */ 989 nargs = 32 - clz32(typemask >> 3); 990 nargs = DIV_ROUND_UP(nargs, 3); 991 assert(nargs <= MAX_CALL_IARGS); 992 993 ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *)); 994 ca->cif.rtype = typecode_to_ffi(typemask & 7); 995 ca->cif.nargs = nargs; 996 997 if (nargs != 0) { 998 ca->cif.arg_types = ca->args; 999 for (int j = 0; j < nargs; ++j) { 1000 int typecode = extract32(typemask, (j + 1) * 3, 3); 1001 ca->args[j] = typecode_to_ffi(typecode); 1002 } 1003 } 1004 1005 status = ffi_prep_cif(&ca->cif, FFI_DEFAULT_ABI, nargs, 1006 ca->cif.rtype, ca->cif.arg_types); 1007 assert(status == FFI_OK); 1008 1009 return &ca->cif; 1010 } 1011 1012 #define HELPER_INFO_INIT(I) (&(I)->cif) 1013 #define HELPER_INFO_INIT_VAL(I) init_ffi_layout(I) 1014 #else 1015 #define HELPER_INFO_INIT(I) (&(I)->init) 1016 #define HELPER_INFO_INIT_VAL(I) 1 1017 #endif /* CONFIG_TCG_INTERPRETER */ 1018 1019 static inline bool arg_slot_reg_p(unsigned arg_slot) 1020 { 1021 /* 1022 * Split the sizeof away from the comparison to avoid Werror from 1023 * "unsigned < 0 is always false", when iarg_regs is empty. 1024 */ 1025 unsigned nreg = ARRAY_SIZE(tcg_target_call_iarg_regs); 1026 return arg_slot < nreg; 1027 } 1028 1029 static inline int arg_slot_stk_ofs(unsigned arg_slot) 1030 { 1031 unsigned max = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1032 unsigned stk_slot = arg_slot - ARRAY_SIZE(tcg_target_call_iarg_regs); 1033 1034 tcg_debug_assert(stk_slot < max); 1035 return TCG_TARGET_CALL_STACK_OFFSET + stk_slot * sizeof(tcg_target_long); 1036 } 1037 1038 typedef struct TCGCumulativeArgs { 1039 int arg_idx; /* tcg_gen_callN args[] */ 1040 int info_in_idx; /* TCGHelperInfo in[] */ 1041 int arg_slot; /* regs+stack slot */ 1042 int ref_slot; /* stack slots for references */ 1043 } TCGCumulativeArgs; 1044 1045 static void layout_arg_even(TCGCumulativeArgs *cum) 1046 { 1047 cum->arg_slot += cum->arg_slot & 1; 1048 } 1049 1050 static void layout_arg_1(TCGCumulativeArgs *cum, TCGHelperInfo *info, 1051 TCGCallArgumentKind kind) 1052 { 1053 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1054 1055 *loc = (TCGCallArgumentLoc){ 1056 .kind = kind, 1057 .arg_idx = cum->arg_idx, 1058 .arg_slot = cum->arg_slot, 1059 }; 1060 cum->info_in_idx++; 1061 cum->arg_slot++; 1062 } 1063 1064 static void layout_arg_normal_n(TCGCumulativeArgs *cum, 1065 TCGHelperInfo *info, int n) 1066 { 1067 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1068 1069 for (int i = 0; i < n; ++i) { 1070 /* Layout all using the same arg_idx, adjusting the subindex. */ 1071 loc[i] = (TCGCallArgumentLoc){ 1072 .kind = TCG_CALL_ARG_NORMAL, 1073 .arg_idx = cum->arg_idx, 1074 .tmp_subindex = i, 1075 .arg_slot = cum->arg_slot + i, 1076 }; 1077 } 1078 cum->info_in_idx += n; 1079 cum->arg_slot += n; 1080 } 1081 1082 static void layout_arg_by_ref(TCGCumulativeArgs *cum, TCGHelperInfo *info) 1083 { 1084 TCGCallArgumentLoc *loc = &info->in[cum->info_in_idx]; 1085 int n = 128 / TCG_TARGET_REG_BITS; 1086 1087 /* The first subindex carries the pointer. */ 1088 layout_arg_1(cum, info, TCG_CALL_ARG_BY_REF); 1089 1090 /* 1091 * The callee is allowed to clobber memory associated with 1092 * structure pass by-reference. Therefore we must make copies. 1093 * Allocate space from "ref_slot", which will be adjusted to 1094 * follow the parameters on the stack. 1095 */ 1096 loc[0].ref_slot = cum->ref_slot; 1097 1098 /* 1099 * Subsequent words also go into the reference slot, but 1100 * do not accumulate into the regular arguments. 1101 */ 1102 for (int i = 1; i < n; ++i) { 1103 loc[i] = (TCGCallArgumentLoc){ 1104 .kind = TCG_CALL_ARG_BY_REF_N, 1105 .arg_idx = cum->arg_idx, 1106 .tmp_subindex = i, 1107 .ref_slot = cum->ref_slot + i, 1108 }; 1109 } 1110 cum->info_in_idx += n - 1; /* i=0 accounted for in layout_arg_1 */ 1111 cum->ref_slot += n; 1112 } 1113 1114 static void init_call_layout(TCGHelperInfo *info) 1115 { 1116 int max_reg_slots = ARRAY_SIZE(tcg_target_call_iarg_regs); 1117 int max_stk_slots = TCG_STATIC_CALL_ARGS_SIZE / sizeof(tcg_target_long); 1118 unsigned typemask = info->typemask; 1119 unsigned typecode; 1120 TCGCumulativeArgs cum = { }; 1121 1122 /* 1123 * Parse and place any function return value. 1124 */ 1125 typecode = typemask & 7; 1126 switch (typecode) { 1127 case dh_typecode_void: 1128 info->nr_out = 0; 1129 break; 1130 case dh_typecode_i32: 1131 case dh_typecode_s32: 1132 case dh_typecode_ptr: 1133 info->nr_out = 1; 1134 info->out_kind = TCG_CALL_RET_NORMAL; 1135 break; 1136 case dh_typecode_i64: 1137 case dh_typecode_s64: 1138 info->nr_out = 64 / TCG_TARGET_REG_BITS; 1139 info->out_kind = TCG_CALL_RET_NORMAL; 1140 /* Query the last register now to trigger any assert early. */ 1141 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1142 break; 1143 case dh_typecode_i128: 1144 info->nr_out = 128 / TCG_TARGET_REG_BITS; 1145 info->out_kind = TCG_TARGET_CALL_RET_I128; 1146 switch (TCG_TARGET_CALL_RET_I128) { 1147 case TCG_CALL_RET_NORMAL: 1148 /* Query the last register now to trigger any assert early. */ 1149 tcg_target_call_oarg_reg(info->out_kind, info->nr_out - 1); 1150 break; 1151 case TCG_CALL_RET_BY_VEC: 1152 /* Query the single register now to trigger any assert early. */ 1153 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0); 1154 break; 1155 case TCG_CALL_RET_BY_REF: 1156 /* 1157 * Allocate the first argument to the output. 1158 * We don't need to store this anywhere, just make it 1159 * unavailable for use in the input loop below. 1160 */ 1161 cum.arg_slot = 1; 1162 break; 1163 default: 1164 qemu_build_not_reached(); 1165 } 1166 break; 1167 default: 1168 g_assert_not_reached(); 1169 } 1170 1171 /* 1172 * Parse and place function arguments. 1173 */ 1174 for (typemask >>= 3; typemask; typemask >>= 3, cum.arg_idx++) { 1175 TCGCallArgumentKind kind; 1176 TCGType type; 1177 1178 typecode = typemask & 7; 1179 switch (typecode) { 1180 case dh_typecode_i32: 1181 case dh_typecode_s32: 1182 type = TCG_TYPE_I32; 1183 break; 1184 case dh_typecode_i64: 1185 case dh_typecode_s64: 1186 type = TCG_TYPE_I64; 1187 break; 1188 case dh_typecode_ptr: 1189 type = TCG_TYPE_PTR; 1190 break; 1191 case dh_typecode_i128: 1192 type = TCG_TYPE_I128; 1193 break; 1194 default: 1195 g_assert_not_reached(); 1196 } 1197 1198 switch (type) { 1199 case TCG_TYPE_I32: 1200 switch (TCG_TARGET_CALL_ARG_I32) { 1201 case TCG_CALL_ARG_EVEN: 1202 layout_arg_even(&cum); 1203 /* fall through */ 1204 case TCG_CALL_ARG_NORMAL: 1205 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1206 break; 1207 case TCG_CALL_ARG_EXTEND: 1208 kind = TCG_CALL_ARG_EXTEND_U + (typecode & 1); 1209 layout_arg_1(&cum, info, kind); 1210 break; 1211 default: 1212 qemu_build_not_reached(); 1213 } 1214 break; 1215 1216 case TCG_TYPE_I64: 1217 switch (TCG_TARGET_CALL_ARG_I64) { 1218 case TCG_CALL_ARG_EVEN: 1219 layout_arg_even(&cum); 1220 /* fall through */ 1221 case TCG_CALL_ARG_NORMAL: 1222 if (TCG_TARGET_REG_BITS == 32) { 1223 layout_arg_normal_n(&cum, info, 2); 1224 } else { 1225 layout_arg_1(&cum, info, TCG_CALL_ARG_NORMAL); 1226 } 1227 break; 1228 default: 1229 qemu_build_not_reached(); 1230 } 1231 break; 1232 1233 case TCG_TYPE_I128: 1234 switch (TCG_TARGET_CALL_ARG_I128) { 1235 case TCG_CALL_ARG_EVEN: 1236 layout_arg_even(&cum); 1237 /* fall through */ 1238 case TCG_CALL_ARG_NORMAL: 1239 layout_arg_normal_n(&cum, info, 128 / TCG_TARGET_REG_BITS); 1240 break; 1241 case TCG_CALL_ARG_BY_REF: 1242 layout_arg_by_ref(&cum, info); 1243 break; 1244 default: 1245 qemu_build_not_reached(); 1246 } 1247 break; 1248 1249 default: 1250 g_assert_not_reached(); 1251 } 1252 } 1253 info->nr_in = cum.info_in_idx; 1254 1255 /* Validate that we didn't overrun the input array. */ 1256 assert(cum.info_in_idx <= ARRAY_SIZE(info->in)); 1257 /* Validate the backend has enough argument space. */ 1258 assert(cum.arg_slot <= max_reg_slots + max_stk_slots); 1259 1260 /* 1261 * Relocate the "ref_slot" area to the end of the parameters. 1262 * Minimizing this stack offset helps code size for x86, 1263 * which has a signed 8-bit offset encoding. 1264 */ 1265 if (cum.ref_slot != 0) { 1266 int ref_base = 0; 1267 1268 if (cum.arg_slot > max_reg_slots) { 1269 int align = __alignof(Int128) / sizeof(tcg_target_long); 1270 1271 ref_base = cum.arg_slot - max_reg_slots; 1272 if (align > 1) { 1273 ref_base = ROUND_UP(ref_base, align); 1274 } 1275 } 1276 assert(ref_base + cum.ref_slot <= max_stk_slots); 1277 ref_base += max_reg_slots; 1278 1279 if (ref_base != 0) { 1280 for (int i = cum.info_in_idx - 1; i >= 0; --i) { 1281 TCGCallArgumentLoc *loc = &info->in[i]; 1282 switch (loc->kind) { 1283 case TCG_CALL_ARG_BY_REF: 1284 case TCG_CALL_ARG_BY_REF_N: 1285 loc->ref_slot += ref_base; 1286 break; 1287 default: 1288 break; 1289 } 1290 } 1291 } 1292 } 1293 } 1294 1295 static int indirect_reg_alloc_order[ARRAY_SIZE(tcg_target_reg_alloc_order)]; 1296 static void process_op_defs(TCGContext *s); 1297 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1298 TCGReg reg, const char *name); 1299 1300 static void tcg_context_init(unsigned max_cpus) 1301 { 1302 TCGContext *s = &tcg_init_ctx; 1303 int op, total_args, n, i; 1304 TCGOpDef *def; 1305 TCGArgConstraint *args_ct; 1306 TCGTemp *ts; 1307 1308 memset(s, 0, sizeof(*s)); 1309 s->nb_globals = 0; 1310 1311 /* Count total number of arguments and allocate the corresponding 1312 space */ 1313 total_args = 0; 1314 for(op = 0; op < NB_OPS; op++) { 1315 def = &tcg_op_defs[op]; 1316 n = def->nb_iargs + def->nb_oargs; 1317 total_args += n; 1318 } 1319 1320 args_ct = g_new0(TCGArgConstraint, total_args); 1321 1322 for(op = 0; op < NB_OPS; op++) { 1323 def = &tcg_op_defs[op]; 1324 def->args_ct = args_ct; 1325 n = def->nb_iargs + def->nb_oargs; 1326 args_ct += n; 1327 } 1328 1329 init_call_layout(&info_helper_ld32_mmu); 1330 init_call_layout(&info_helper_ld64_mmu); 1331 init_call_layout(&info_helper_ld128_mmu); 1332 init_call_layout(&info_helper_st32_mmu); 1333 init_call_layout(&info_helper_st64_mmu); 1334 init_call_layout(&info_helper_st128_mmu); 1335 1336 tcg_target_init(s); 1337 process_op_defs(s); 1338 1339 /* Reverse the order of the saved registers, assuming they're all at 1340 the start of tcg_target_reg_alloc_order. */ 1341 for (n = 0; n < ARRAY_SIZE(tcg_target_reg_alloc_order); ++n) { 1342 int r = tcg_target_reg_alloc_order[n]; 1343 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, r)) { 1344 break; 1345 } 1346 } 1347 for (i = 0; i < n; ++i) { 1348 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[n - 1 - i]; 1349 } 1350 for (; i < ARRAY_SIZE(tcg_target_reg_alloc_order); ++i) { 1351 indirect_reg_alloc_order[i] = tcg_target_reg_alloc_order[i]; 1352 } 1353 1354 tcg_ctx = s; 1355 /* 1356 * In user-mode we simply share the init context among threads, since we 1357 * use a single region. See the documentation tcg_region_init() for the 1358 * reasoning behind this. 1359 * In system-mode we will have at most max_cpus TCG threads. 1360 */ 1361 #ifdef CONFIG_USER_ONLY 1362 tcg_ctxs = &tcg_ctx; 1363 tcg_cur_ctxs = 1; 1364 tcg_max_ctxs = 1; 1365 #else 1366 tcg_max_ctxs = max_cpus; 1367 tcg_ctxs = g_new0(TCGContext *, max_cpus); 1368 #endif 1369 1370 tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0)); 1371 ts = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, TCG_AREG0, "env"); 1372 tcg_env = temp_tcgv_ptr(ts); 1373 } 1374 1375 void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus) 1376 { 1377 tcg_context_init(max_cpus); 1378 tcg_region_init(tb_size, splitwx, max_cpus); 1379 } 1380 1381 /* 1382 * Allocate TBs right before their corresponding translated code, making 1383 * sure that TBs and code are on different cache lines. 1384 */ 1385 TranslationBlock *tcg_tb_alloc(TCGContext *s) 1386 { 1387 uintptr_t align = qemu_icache_linesize; 1388 TranslationBlock *tb; 1389 void *next; 1390 1391 retry: 1392 tb = (void *)ROUND_UP((uintptr_t)s->code_gen_ptr, align); 1393 next = (void *)ROUND_UP((uintptr_t)(tb + 1), align); 1394 1395 if (unlikely(next > s->code_gen_highwater)) { 1396 if (tcg_region_alloc(s)) { 1397 return NULL; 1398 } 1399 goto retry; 1400 } 1401 qatomic_set(&s->code_gen_ptr, next); 1402 s->data_gen_ptr = NULL; 1403 return tb; 1404 } 1405 1406 void tcg_prologue_init(void) 1407 { 1408 TCGContext *s = tcg_ctx; 1409 size_t prologue_size; 1410 1411 s->code_ptr = s->code_gen_ptr; 1412 s->code_buf = s->code_gen_ptr; 1413 s->data_gen_ptr = NULL; 1414 1415 #ifndef CONFIG_TCG_INTERPRETER 1416 tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr); 1417 #endif 1418 1419 #ifdef TCG_TARGET_NEED_POOL_LABELS 1420 s->pool_labels = NULL; 1421 #endif 1422 1423 qemu_thread_jit_write(); 1424 /* Generate the prologue. */ 1425 tcg_target_qemu_prologue(s); 1426 1427 #ifdef TCG_TARGET_NEED_POOL_LABELS 1428 /* Allow the prologue to put e.g. guest_base into a pool entry. */ 1429 { 1430 int result = tcg_out_pool_finalize(s); 1431 tcg_debug_assert(result == 0); 1432 } 1433 #endif 1434 1435 prologue_size = tcg_current_code_size(s); 1436 perf_report_prologue(s->code_gen_ptr, prologue_size); 1437 1438 #ifndef CONFIG_TCG_INTERPRETER 1439 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 1440 (uintptr_t)s->code_buf, prologue_size); 1441 #endif 1442 1443 if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { 1444 FILE *logfile = qemu_log_trylock(); 1445 if (logfile) { 1446 fprintf(logfile, "PROLOGUE: [size=%zu]\n", prologue_size); 1447 if (s->data_gen_ptr) { 1448 size_t code_size = s->data_gen_ptr - s->code_gen_ptr; 1449 size_t data_size = prologue_size - code_size; 1450 size_t i; 1451 1452 disas(logfile, s->code_gen_ptr, code_size); 1453 1454 for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) { 1455 if (sizeof(tcg_target_ulong) == 8) { 1456 fprintf(logfile, 1457 "0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n", 1458 (uintptr_t)s->data_gen_ptr + i, 1459 *(uint64_t *)(s->data_gen_ptr + i)); 1460 } else { 1461 fprintf(logfile, 1462 "0x%08" PRIxPTR ": .long 0x%08x\n", 1463 (uintptr_t)s->data_gen_ptr + i, 1464 *(uint32_t *)(s->data_gen_ptr + i)); 1465 } 1466 } 1467 } else { 1468 disas(logfile, s->code_gen_ptr, prologue_size); 1469 } 1470 fprintf(logfile, "\n"); 1471 qemu_log_unlock(logfile); 1472 } 1473 } 1474 1475 #ifndef CONFIG_TCG_INTERPRETER 1476 /* 1477 * Assert that goto_ptr is implemented completely, setting an epilogue. 1478 * For tci, we use NULL as the signal to return from the interpreter, 1479 * so skip this check. 1480 */ 1481 tcg_debug_assert(tcg_code_gen_epilogue != NULL); 1482 #endif 1483 1484 tcg_region_prologue_set(s); 1485 } 1486 1487 void tcg_func_start(TCGContext *s) 1488 { 1489 tcg_pool_reset(s); 1490 s->nb_temps = s->nb_globals; 1491 1492 /* No temps have been previously allocated for size or locality. */ 1493 memset(s->free_temps, 0, sizeof(s->free_temps)); 1494 1495 /* No constant temps have been previously allocated. */ 1496 for (int i = 0; i < TCG_TYPE_COUNT; ++i) { 1497 if (s->const_table[i]) { 1498 g_hash_table_remove_all(s->const_table[i]); 1499 } 1500 } 1501 1502 s->nb_ops = 0; 1503 s->nb_labels = 0; 1504 s->current_frame_offset = s->frame_start; 1505 1506 #ifdef CONFIG_DEBUG_TCG 1507 s->goto_tb_issue_mask = 0; 1508 #endif 1509 1510 QTAILQ_INIT(&s->ops); 1511 QTAILQ_INIT(&s->free_ops); 1512 s->emit_before_op = NULL; 1513 QSIMPLEQ_INIT(&s->labels); 1514 1515 tcg_debug_assert(s->addr_type == TCG_TYPE_I32 || 1516 s->addr_type == TCG_TYPE_I64); 1517 1518 tcg_debug_assert(s->insn_start_words > 0); 1519 } 1520 1521 static TCGTemp *tcg_temp_alloc(TCGContext *s) 1522 { 1523 int n = s->nb_temps++; 1524 1525 if (n >= TCG_MAX_TEMPS) { 1526 tcg_raise_tb_overflow(s); 1527 } 1528 return memset(&s->temps[n], 0, sizeof(TCGTemp)); 1529 } 1530 1531 static TCGTemp *tcg_global_alloc(TCGContext *s) 1532 { 1533 TCGTemp *ts; 1534 1535 tcg_debug_assert(s->nb_globals == s->nb_temps); 1536 tcg_debug_assert(s->nb_globals < TCG_MAX_TEMPS); 1537 s->nb_globals++; 1538 ts = tcg_temp_alloc(s); 1539 ts->kind = TEMP_GLOBAL; 1540 1541 return ts; 1542 } 1543 1544 static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, 1545 TCGReg reg, const char *name) 1546 { 1547 TCGTemp *ts; 1548 1549 tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32); 1550 1551 ts = tcg_global_alloc(s); 1552 ts->base_type = type; 1553 ts->type = type; 1554 ts->kind = TEMP_FIXED; 1555 ts->reg = reg; 1556 ts->name = name; 1557 tcg_regset_set_reg(s->reserved_regs, reg); 1558 1559 return ts; 1560 } 1561 1562 void tcg_set_frame(TCGContext *s, TCGReg reg, intptr_t start, intptr_t size) 1563 { 1564 s->frame_start = start; 1565 s->frame_end = start + size; 1566 s->frame_temp 1567 = tcg_global_reg_new_internal(s, TCG_TYPE_PTR, reg, "_frame"); 1568 } 1569 1570 static TCGTemp *tcg_global_mem_new_internal(TCGv_ptr base, intptr_t offset, 1571 const char *name, TCGType type) 1572 { 1573 TCGContext *s = tcg_ctx; 1574 TCGTemp *base_ts = tcgv_ptr_temp(base); 1575 TCGTemp *ts = tcg_global_alloc(s); 1576 int indirect_reg = 0; 1577 1578 switch (base_ts->kind) { 1579 case TEMP_FIXED: 1580 break; 1581 case TEMP_GLOBAL: 1582 /* We do not support double-indirect registers. */ 1583 tcg_debug_assert(!base_ts->indirect_reg); 1584 base_ts->indirect_base = 1; 1585 s->nb_indirects += (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64 1586 ? 2 : 1); 1587 indirect_reg = 1; 1588 break; 1589 default: 1590 g_assert_not_reached(); 1591 } 1592 1593 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1594 TCGTemp *ts2 = tcg_global_alloc(s); 1595 char buf[64]; 1596 1597 ts->base_type = TCG_TYPE_I64; 1598 ts->type = TCG_TYPE_I32; 1599 ts->indirect_reg = indirect_reg; 1600 ts->mem_allocated = 1; 1601 ts->mem_base = base_ts; 1602 ts->mem_offset = offset; 1603 pstrcpy(buf, sizeof(buf), name); 1604 pstrcat(buf, sizeof(buf), "_0"); 1605 ts->name = strdup(buf); 1606 1607 tcg_debug_assert(ts2 == ts + 1); 1608 ts2->base_type = TCG_TYPE_I64; 1609 ts2->type = TCG_TYPE_I32; 1610 ts2->indirect_reg = indirect_reg; 1611 ts2->mem_allocated = 1; 1612 ts2->mem_base = base_ts; 1613 ts2->mem_offset = offset + 4; 1614 ts2->temp_subindex = 1; 1615 pstrcpy(buf, sizeof(buf), name); 1616 pstrcat(buf, sizeof(buf), "_1"); 1617 ts2->name = strdup(buf); 1618 } else { 1619 ts->base_type = type; 1620 ts->type = type; 1621 ts->indirect_reg = indirect_reg; 1622 ts->mem_allocated = 1; 1623 ts->mem_base = base_ts; 1624 ts->mem_offset = offset; 1625 ts->name = name; 1626 } 1627 return ts; 1628 } 1629 1630 TCGv_i32 tcg_global_mem_new_i32(TCGv_ptr reg, intptr_t off, const char *name) 1631 { 1632 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I32); 1633 return temp_tcgv_i32(ts); 1634 } 1635 1636 TCGv_i64 tcg_global_mem_new_i64(TCGv_ptr reg, intptr_t off, const char *name) 1637 { 1638 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_I64); 1639 return temp_tcgv_i64(ts); 1640 } 1641 1642 TCGv_ptr tcg_global_mem_new_ptr(TCGv_ptr reg, intptr_t off, const char *name) 1643 { 1644 TCGTemp *ts = tcg_global_mem_new_internal(reg, off, name, TCG_TYPE_PTR); 1645 return temp_tcgv_ptr(ts); 1646 } 1647 1648 TCGTemp *tcg_temp_new_internal(TCGType type, TCGTempKind kind) 1649 { 1650 TCGContext *s = tcg_ctx; 1651 TCGTemp *ts; 1652 int n; 1653 1654 if (kind == TEMP_EBB) { 1655 int idx = find_first_bit(s->free_temps[type].l, TCG_MAX_TEMPS); 1656 1657 if (idx < TCG_MAX_TEMPS) { 1658 /* There is already an available temp with the right type. */ 1659 clear_bit(idx, s->free_temps[type].l); 1660 1661 ts = &s->temps[idx]; 1662 ts->temp_allocated = 1; 1663 tcg_debug_assert(ts->base_type == type); 1664 tcg_debug_assert(ts->kind == kind); 1665 return ts; 1666 } 1667 } else { 1668 tcg_debug_assert(kind == TEMP_TB); 1669 } 1670 1671 switch (type) { 1672 case TCG_TYPE_I32: 1673 case TCG_TYPE_V64: 1674 case TCG_TYPE_V128: 1675 case TCG_TYPE_V256: 1676 n = 1; 1677 break; 1678 case TCG_TYPE_I64: 1679 n = 64 / TCG_TARGET_REG_BITS; 1680 break; 1681 case TCG_TYPE_I128: 1682 n = 128 / TCG_TARGET_REG_BITS; 1683 break; 1684 default: 1685 g_assert_not_reached(); 1686 } 1687 1688 ts = tcg_temp_alloc(s); 1689 ts->base_type = type; 1690 ts->temp_allocated = 1; 1691 ts->kind = kind; 1692 1693 if (n == 1) { 1694 ts->type = type; 1695 } else { 1696 ts->type = TCG_TYPE_REG; 1697 1698 for (int i = 1; i < n; ++i) { 1699 TCGTemp *ts2 = tcg_temp_alloc(s); 1700 1701 tcg_debug_assert(ts2 == ts + i); 1702 ts2->base_type = type; 1703 ts2->type = TCG_TYPE_REG; 1704 ts2->temp_allocated = 1; 1705 ts2->temp_subindex = i; 1706 ts2->kind = kind; 1707 } 1708 } 1709 return ts; 1710 } 1711 1712 TCGv_i32 tcg_temp_new_i32(void) 1713 { 1714 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_TB)); 1715 } 1716 1717 TCGv_i32 tcg_temp_ebb_new_i32(void) 1718 { 1719 return temp_tcgv_i32(tcg_temp_new_internal(TCG_TYPE_I32, TEMP_EBB)); 1720 } 1721 1722 TCGv_i64 tcg_temp_new_i64(void) 1723 { 1724 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_TB)); 1725 } 1726 1727 TCGv_i64 tcg_temp_ebb_new_i64(void) 1728 { 1729 return temp_tcgv_i64(tcg_temp_new_internal(TCG_TYPE_I64, TEMP_EBB)); 1730 } 1731 1732 TCGv_ptr tcg_temp_new_ptr(void) 1733 { 1734 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_TB)); 1735 } 1736 1737 TCGv_ptr tcg_temp_ebb_new_ptr(void) 1738 { 1739 return temp_tcgv_ptr(tcg_temp_new_internal(TCG_TYPE_PTR, TEMP_EBB)); 1740 } 1741 1742 TCGv_i128 tcg_temp_new_i128(void) 1743 { 1744 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_TB)); 1745 } 1746 1747 TCGv_i128 tcg_temp_ebb_new_i128(void) 1748 { 1749 return temp_tcgv_i128(tcg_temp_new_internal(TCG_TYPE_I128, TEMP_EBB)); 1750 } 1751 1752 TCGv_vec tcg_temp_new_vec(TCGType type) 1753 { 1754 TCGTemp *t; 1755 1756 #ifdef CONFIG_DEBUG_TCG 1757 switch (type) { 1758 case TCG_TYPE_V64: 1759 assert(TCG_TARGET_HAS_v64); 1760 break; 1761 case TCG_TYPE_V128: 1762 assert(TCG_TARGET_HAS_v128); 1763 break; 1764 case TCG_TYPE_V256: 1765 assert(TCG_TARGET_HAS_v256); 1766 break; 1767 default: 1768 g_assert_not_reached(); 1769 } 1770 #endif 1771 1772 t = tcg_temp_new_internal(type, TEMP_EBB); 1773 return temp_tcgv_vec(t); 1774 } 1775 1776 /* Create a new temp of the same type as an existing temp. */ 1777 TCGv_vec tcg_temp_new_vec_matching(TCGv_vec match) 1778 { 1779 TCGTemp *t = tcgv_vec_temp(match); 1780 1781 tcg_debug_assert(t->temp_allocated != 0); 1782 1783 t = tcg_temp_new_internal(t->base_type, TEMP_EBB); 1784 return temp_tcgv_vec(t); 1785 } 1786 1787 void tcg_temp_free_internal(TCGTemp *ts) 1788 { 1789 TCGContext *s = tcg_ctx; 1790 1791 switch (ts->kind) { 1792 case TEMP_CONST: 1793 case TEMP_TB: 1794 /* Silently ignore free. */ 1795 break; 1796 case TEMP_EBB: 1797 tcg_debug_assert(ts->temp_allocated != 0); 1798 ts->temp_allocated = 0; 1799 set_bit(temp_idx(ts), s->free_temps[ts->base_type].l); 1800 break; 1801 default: 1802 /* It never made sense to free TEMP_FIXED or TEMP_GLOBAL. */ 1803 g_assert_not_reached(); 1804 } 1805 } 1806 1807 void tcg_temp_free_i32(TCGv_i32 arg) 1808 { 1809 tcg_temp_free_internal(tcgv_i32_temp(arg)); 1810 } 1811 1812 void tcg_temp_free_i64(TCGv_i64 arg) 1813 { 1814 tcg_temp_free_internal(tcgv_i64_temp(arg)); 1815 } 1816 1817 void tcg_temp_free_i128(TCGv_i128 arg) 1818 { 1819 tcg_temp_free_internal(tcgv_i128_temp(arg)); 1820 } 1821 1822 void tcg_temp_free_ptr(TCGv_ptr arg) 1823 { 1824 tcg_temp_free_internal(tcgv_ptr_temp(arg)); 1825 } 1826 1827 void tcg_temp_free_vec(TCGv_vec arg) 1828 { 1829 tcg_temp_free_internal(tcgv_vec_temp(arg)); 1830 } 1831 1832 TCGTemp *tcg_constant_internal(TCGType type, int64_t val) 1833 { 1834 TCGContext *s = tcg_ctx; 1835 GHashTable *h = s->const_table[type]; 1836 TCGTemp *ts; 1837 1838 if (h == NULL) { 1839 h = g_hash_table_new(g_int64_hash, g_int64_equal); 1840 s->const_table[type] = h; 1841 } 1842 1843 ts = g_hash_table_lookup(h, &val); 1844 if (ts == NULL) { 1845 int64_t *val_ptr; 1846 1847 ts = tcg_temp_alloc(s); 1848 1849 if (TCG_TARGET_REG_BITS == 32 && type == TCG_TYPE_I64) { 1850 TCGTemp *ts2 = tcg_temp_alloc(s); 1851 1852 tcg_debug_assert(ts2 == ts + 1); 1853 1854 ts->base_type = TCG_TYPE_I64; 1855 ts->type = TCG_TYPE_I32; 1856 ts->kind = TEMP_CONST; 1857 ts->temp_allocated = 1; 1858 1859 ts2->base_type = TCG_TYPE_I64; 1860 ts2->type = TCG_TYPE_I32; 1861 ts2->kind = TEMP_CONST; 1862 ts2->temp_allocated = 1; 1863 ts2->temp_subindex = 1; 1864 1865 /* 1866 * Retain the full value of the 64-bit constant in the low 1867 * part, so that the hash table works. Actual uses will 1868 * truncate the value to the low part. 1869 */ 1870 ts[HOST_BIG_ENDIAN].val = val; 1871 ts[!HOST_BIG_ENDIAN].val = val >> 32; 1872 val_ptr = &ts[HOST_BIG_ENDIAN].val; 1873 } else { 1874 ts->base_type = type; 1875 ts->type = type; 1876 ts->kind = TEMP_CONST; 1877 ts->temp_allocated = 1; 1878 ts->val = val; 1879 val_ptr = &ts->val; 1880 } 1881 g_hash_table_insert(h, val_ptr, ts); 1882 } 1883 1884 return ts; 1885 } 1886 1887 TCGv_i32 tcg_constant_i32(int32_t val) 1888 { 1889 return temp_tcgv_i32(tcg_constant_internal(TCG_TYPE_I32, val)); 1890 } 1891 1892 TCGv_i64 tcg_constant_i64(int64_t val) 1893 { 1894 return temp_tcgv_i64(tcg_constant_internal(TCG_TYPE_I64, val)); 1895 } 1896 1897 TCGv_ptr tcg_constant_ptr_int(intptr_t val) 1898 { 1899 return temp_tcgv_ptr(tcg_constant_internal(TCG_TYPE_PTR, val)); 1900 } 1901 1902 TCGv_vec tcg_constant_vec(TCGType type, unsigned vece, int64_t val) 1903 { 1904 val = dup_const(vece, val); 1905 return temp_tcgv_vec(tcg_constant_internal(type, val)); 1906 } 1907 1908 TCGv_vec tcg_constant_vec_matching(TCGv_vec match, unsigned vece, int64_t val) 1909 { 1910 TCGTemp *t = tcgv_vec_temp(match); 1911 1912 tcg_debug_assert(t->temp_allocated != 0); 1913 return tcg_constant_vec(t->base_type, vece, val); 1914 } 1915 1916 #ifdef CONFIG_DEBUG_TCG 1917 size_t temp_idx(TCGTemp *ts) 1918 { 1919 ptrdiff_t n = ts - tcg_ctx->temps; 1920 assert(n >= 0 && n < tcg_ctx->nb_temps); 1921 return n; 1922 } 1923 1924 TCGTemp *tcgv_i32_temp(TCGv_i32 v) 1925 { 1926 uintptr_t o = (uintptr_t)v - offsetof(TCGContext, temps); 1927 1928 assert(o < sizeof(TCGTemp) * tcg_ctx->nb_temps); 1929 assert(o % sizeof(TCGTemp) == 0); 1930 1931 return (void *)tcg_ctx + (uintptr_t)v; 1932 } 1933 #endif /* CONFIG_DEBUG_TCG */ 1934 1935 /* Return true if OP may appear in the opcode stream. 1936 Test the runtime variable that controls each opcode. */ 1937 bool tcg_op_supported(TCGOpcode op) 1938 { 1939 const bool have_vec 1940 = TCG_TARGET_HAS_v64 | TCG_TARGET_HAS_v128 | TCG_TARGET_HAS_v256; 1941 1942 switch (op) { 1943 case INDEX_op_discard: 1944 case INDEX_op_set_label: 1945 case INDEX_op_call: 1946 case INDEX_op_br: 1947 case INDEX_op_mb: 1948 case INDEX_op_insn_start: 1949 case INDEX_op_exit_tb: 1950 case INDEX_op_goto_tb: 1951 case INDEX_op_goto_ptr: 1952 case INDEX_op_qemu_ld_a32_i32: 1953 case INDEX_op_qemu_ld_a64_i32: 1954 case INDEX_op_qemu_st_a32_i32: 1955 case INDEX_op_qemu_st_a64_i32: 1956 case INDEX_op_qemu_ld_a32_i64: 1957 case INDEX_op_qemu_ld_a64_i64: 1958 case INDEX_op_qemu_st_a32_i64: 1959 case INDEX_op_qemu_st_a64_i64: 1960 return true; 1961 1962 case INDEX_op_qemu_st8_a32_i32: 1963 case INDEX_op_qemu_st8_a64_i32: 1964 return TCG_TARGET_HAS_qemu_st8_i32; 1965 1966 case INDEX_op_qemu_ld_a32_i128: 1967 case INDEX_op_qemu_ld_a64_i128: 1968 case INDEX_op_qemu_st_a32_i128: 1969 case INDEX_op_qemu_st_a64_i128: 1970 return TCG_TARGET_HAS_qemu_ldst_i128; 1971 1972 case INDEX_op_mov_i32: 1973 case INDEX_op_setcond_i32: 1974 case INDEX_op_brcond_i32: 1975 case INDEX_op_movcond_i32: 1976 case INDEX_op_ld8u_i32: 1977 case INDEX_op_ld8s_i32: 1978 case INDEX_op_ld16u_i32: 1979 case INDEX_op_ld16s_i32: 1980 case INDEX_op_ld_i32: 1981 case INDEX_op_st8_i32: 1982 case INDEX_op_st16_i32: 1983 case INDEX_op_st_i32: 1984 case INDEX_op_add_i32: 1985 case INDEX_op_sub_i32: 1986 case INDEX_op_neg_i32: 1987 case INDEX_op_mul_i32: 1988 case INDEX_op_and_i32: 1989 case INDEX_op_or_i32: 1990 case INDEX_op_xor_i32: 1991 case INDEX_op_shl_i32: 1992 case INDEX_op_shr_i32: 1993 case INDEX_op_sar_i32: 1994 return true; 1995 1996 case INDEX_op_negsetcond_i32: 1997 return TCG_TARGET_HAS_negsetcond_i32; 1998 case INDEX_op_div_i32: 1999 case INDEX_op_divu_i32: 2000 return TCG_TARGET_HAS_div_i32; 2001 case INDEX_op_rem_i32: 2002 case INDEX_op_remu_i32: 2003 return TCG_TARGET_HAS_rem_i32; 2004 case INDEX_op_div2_i32: 2005 case INDEX_op_divu2_i32: 2006 return TCG_TARGET_HAS_div2_i32; 2007 case INDEX_op_rotl_i32: 2008 case INDEX_op_rotr_i32: 2009 return TCG_TARGET_HAS_rot_i32; 2010 case INDEX_op_deposit_i32: 2011 return TCG_TARGET_HAS_deposit_i32; 2012 case INDEX_op_extract_i32: 2013 return TCG_TARGET_HAS_extract_i32; 2014 case INDEX_op_sextract_i32: 2015 return TCG_TARGET_HAS_sextract_i32; 2016 case INDEX_op_extract2_i32: 2017 return TCG_TARGET_HAS_extract2_i32; 2018 case INDEX_op_add2_i32: 2019 return TCG_TARGET_HAS_add2_i32; 2020 case INDEX_op_sub2_i32: 2021 return TCG_TARGET_HAS_sub2_i32; 2022 case INDEX_op_mulu2_i32: 2023 return TCG_TARGET_HAS_mulu2_i32; 2024 case INDEX_op_muls2_i32: 2025 return TCG_TARGET_HAS_muls2_i32; 2026 case INDEX_op_muluh_i32: 2027 return TCG_TARGET_HAS_muluh_i32; 2028 case INDEX_op_mulsh_i32: 2029 return TCG_TARGET_HAS_mulsh_i32; 2030 case INDEX_op_ext8s_i32: 2031 return TCG_TARGET_HAS_ext8s_i32; 2032 case INDEX_op_ext16s_i32: 2033 return TCG_TARGET_HAS_ext16s_i32; 2034 case INDEX_op_ext8u_i32: 2035 return TCG_TARGET_HAS_ext8u_i32; 2036 case INDEX_op_ext16u_i32: 2037 return TCG_TARGET_HAS_ext16u_i32; 2038 case INDEX_op_bswap16_i32: 2039 return TCG_TARGET_HAS_bswap16_i32; 2040 case INDEX_op_bswap32_i32: 2041 return TCG_TARGET_HAS_bswap32_i32; 2042 case INDEX_op_not_i32: 2043 return TCG_TARGET_HAS_not_i32; 2044 case INDEX_op_andc_i32: 2045 return TCG_TARGET_HAS_andc_i32; 2046 case INDEX_op_orc_i32: 2047 return TCG_TARGET_HAS_orc_i32; 2048 case INDEX_op_eqv_i32: 2049 return TCG_TARGET_HAS_eqv_i32; 2050 case INDEX_op_nand_i32: 2051 return TCG_TARGET_HAS_nand_i32; 2052 case INDEX_op_nor_i32: 2053 return TCG_TARGET_HAS_nor_i32; 2054 case INDEX_op_clz_i32: 2055 return TCG_TARGET_HAS_clz_i32; 2056 case INDEX_op_ctz_i32: 2057 return TCG_TARGET_HAS_ctz_i32; 2058 case INDEX_op_ctpop_i32: 2059 return TCG_TARGET_HAS_ctpop_i32; 2060 2061 case INDEX_op_brcond2_i32: 2062 case INDEX_op_setcond2_i32: 2063 return TCG_TARGET_REG_BITS == 32; 2064 2065 case INDEX_op_mov_i64: 2066 case INDEX_op_setcond_i64: 2067 case INDEX_op_brcond_i64: 2068 case INDEX_op_movcond_i64: 2069 case INDEX_op_ld8u_i64: 2070 case INDEX_op_ld8s_i64: 2071 case INDEX_op_ld16u_i64: 2072 case INDEX_op_ld16s_i64: 2073 case INDEX_op_ld32u_i64: 2074 case INDEX_op_ld32s_i64: 2075 case INDEX_op_ld_i64: 2076 case INDEX_op_st8_i64: 2077 case INDEX_op_st16_i64: 2078 case INDEX_op_st32_i64: 2079 case INDEX_op_st_i64: 2080 case INDEX_op_add_i64: 2081 case INDEX_op_sub_i64: 2082 case INDEX_op_neg_i64: 2083 case INDEX_op_mul_i64: 2084 case INDEX_op_and_i64: 2085 case INDEX_op_or_i64: 2086 case INDEX_op_xor_i64: 2087 case INDEX_op_shl_i64: 2088 case INDEX_op_shr_i64: 2089 case INDEX_op_sar_i64: 2090 case INDEX_op_ext_i32_i64: 2091 case INDEX_op_extu_i32_i64: 2092 return TCG_TARGET_REG_BITS == 64; 2093 2094 case INDEX_op_negsetcond_i64: 2095 return TCG_TARGET_HAS_negsetcond_i64; 2096 case INDEX_op_div_i64: 2097 case INDEX_op_divu_i64: 2098 return TCG_TARGET_HAS_div_i64; 2099 case INDEX_op_rem_i64: 2100 case INDEX_op_remu_i64: 2101 return TCG_TARGET_HAS_rem_i64; 2102 case INDEX_op_div2_i64: 2103 case INDEX_op_divu2_i64: 2104 return TCG_TARGET_HAS_div2_i64; 2105 case INDEX_op_rotl_i64: 2106 case INDEX_op_rotr_i64: 2107 return TCG_TARGET_HAS_rot_i64; 2108 case INDEX_op_deposit_i64: 2109 return TCG_TARGET_HAS_deposit_i64; 2110 case INDEX_op_extract_i64: 2111 return TCG_TARGET_HAS_extract_i64; 2112 case INDEX_op_sextract_i64: 2113 return TCG_TARGET_HAS_sextract_i64; 2114 case INDEX_op_extract2_i64: 2115 return TCG_TARGET_HAS_extract2_i64; 2116 case INDEX_op_extrl_i64_i32: 2117 case INDEX_op_extrh_i64_i32: 2118 return TCG_TARGET_HAS_extr_i64_i32; 2119 case INDEX_op_ext8s_i64: 2120 return TCG_TARGET_HAS_ext8s_i64; 2121 case INDEX_op_ext16s_i64: 2122 return TCG_TARGET_HAS_ext16s_i64; 2123 case INDEX_op_ext32s_i64: 2124 return TCG_TARGET_HAS_ext32s_i64; 2125 case INDEX_op_ext8u_i64: 2126 return TCG_TARGET_HAS_ext8u_i64; 2127 case INDEX_op_ext16u_i64: 2128 return TCG_TARGET_HAS_ext16u_i64; 2129 case INDEX_op_ext32u_i64: 2130 return TCG_TARGET_HAS_ext32u_i64; 2131 case INDEX_op_bswap16_i64: 2132 return TCG_TARGET_HAS_bswap16_i64; 2133 case INDEX_op_bswap32_i64: 2134 return TCG_TARGET_HAS_bswap32_i64; 2135 case INDEX_op_bswap64_i64: 2136 return TCG_TARGET_HAS_bswap64_i64; 2137 case INDEX_op_not_i64: 2138 return TCG_TARGET_HAS_not_i64; 2139 case INDEX_op_andc_i64: 2140 return TCG_TARGET_HAS_andc_i64; 2141 case INDEX_op_orc_i64: 2142 return TCG_TARGET_HAS_orc_i64; 2143 case INDEX_op_eqv_i64: 2144 return TCG_TARGET_HAS_eqv_i64; 2145 case INDEX_op_nand_i64: 2146 return TCG_TARGET_HAS_nand_i64; 2147 case INDEX_op_nor_i64: 2148 return TCG_TARGET_HAS_nor_i64; 2149 case INDEX_op_clz_i64: 2150 return TCG_TARGET_HAS_clz_i64; 2151 case INDEX_op_ctz_i64: 2152 return TCG_TARGET_HAS_ctz_i64; 2153 case INDEX_op_ctpop_i64: 2154 return TCG_TARGET_HAS_ctpop_i64; 2155 case INDEX_op_add2_i64: 2156 return TCG_TARGET_HAS_add2_i64; 2157 case INDEX_op_sub2_i64: 2158 return TCG_TARGET_HAS_sub2_i64; 2159 case INDEX_op_mulu2_i64: 2160 return TCG_TARGET_HAS_mulu2_i64; 2161 case INDEX_op_muls2_i64: 2162 return TCG_TARGET_HAS_muls2_i64; 2163 case INDEX_op_muluh_i64: 2164 return TCG_TARGET_HAS_muluh_i64; 2165 case INDEX_op_mulsh_i64: 2166 return TCG_TARGET_HAS_mulsh_i64; 2167 2168 case INDEX_op_mov_vec: 2169 case INDEX_op_dup_vec: 2170 case INDEX_op_dupm_vec: 2171 case INDEX_op_ld_vec: 2172 case INDEX_op_st_vec: 2173 case INDEX_op_add_vec: 2174 case INDEX_op_sub_vec: 2175 case INDEX_op_and_vec: 2176 case INDEX_op_or_vec: 2177 case INDEX_op_xor_vec: 2178 case INDEX_op_cmp_vec: 2179 return have_vec; 2180 case INDEX_op_dup2_vec: 2181 return have_vec && TCG_TARGET_REG_BITS == 32; 2182 case INDEX_op_not_vec: 2183 return have_vec && TCG_TARGET_HAS_not_vec; 2184 case INDEX_op_neg_vec: 2185 return have_vec && TCG_TARGET_HAS_neg_vec; 2186 case INDEX_op_abs_vec: 2187 return have_vec && TCG_TARGET_HAS_abs_vec; 2188 case INDEX_op_andc_vec: 2189 return have_vec && TCG_TARGET_HAS_andc_vec; 2190 case INDEX_op_orc_vec: 2191 return have_vec && TCG_TARGET_HAS_orc_vec; 2192 case INDEX_op_nand_vec: 2193 return have_vec && TCG_TARGET_HAS_nand_vec; 2194 case INDEX_op_nor_vec: 2195 return have_vec && TCG_TARGET_HAS_nor_vec; 2196 case INDEX_op_eqv_vec: 2197 return have_vec && TCG_TARGET_HAS_eqv_vec; 2198 case INDEX_op_mul_vec: 2199 return have_vec && TCG_TARGET_HAS_mul_vec; 2200 case INDEX_op_shli_vec: 2201 case INDEX_op_shri_vec: 2202 case INDEX_op_sari_vec: 2203 return have_vec && TCG_TARGET_HAS_shi_vec; 2204 case INDEX_op_shls_vec: 2205 case INDEX_op_shrs_vec: 2206 case INDEX_op_sars_vec: 2207 return have_vec && TCG_TARGET_HAS_shs_vec; 2208 case INDEX_op_shlv_vec: 2209 case INDEX_op_shrv_vec: 2210 case INDEX_op_sarv_vec: 2211 return have_vec && TCG_TARGET_HAS_shv_vec; 2212 case INDEX_op_rotli_vec: 2213 return have_vec && TCG_TARGET_HAS_roti_vec; 2214 case INDEX_op_rotls_vec: 2215 return have_vec && TCG_TARGET_HAS_rots_vec; 2216 case INDEX_op_rotlv_vec: 2217 case INDEX_op_rotrv_vec: 2218 return have_vec && TCG_TARGET_HAS_rotv_vec; 2219 case INDEX_op_ssadd_vec: 2220 case INDEX_op_usadd_vec: 2221 case INDEX_op_sssub_vec: 2222 case INDEX_op_ussub_vec: 2223 return have_vec && TCG_TARGET_HAS_sat_vec; 2224 case INDEX_op_smin_vec: 2225 case INDEX_op_umin_vec: 2226 case INDEX_op_smax_vec: 2227 case INDEX_op_umax_vec: 2228 return have_vec && TCG_TARGET_HAS_minmax_vec; 2229 case INDEX_op_bitsel_vec: 2230 return have_vec && TCG_TARGET_HAS_bitsel_vec; 2231 case INDEX_op_cmpsel_vec: 2232 return have_vec && TCG_TARGET_HAS_cmpsel_vec; 2233 2234 default: 2235 tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS); 2236 return true; 2237 } 2238 } 2239 2240 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs); 2241 2242 static void tcg_gen_callN(void *func, TCGHelperInfo *info, 2243 TCGTemp *ret, TCGTemp **args) 2244 { 2245 TCGv_i64 extend_free[MAX_CALL_IARGS]; 2246 int n_extend = 0; 2247 TCGOp *op; 2248 int i, n, pi = 0, total_args; 2249 2250 if (unlikely(g_once_init_enter(HELPER_INFO_INIT(info)))) { 2251 init_call_layout(info); 2252 g_once_init_leave(HELPER_INFO_INIT(info), HELPER_INFO_INIT_VAL(info)); 2253 } 2254 2255 total_args = info->nr_out + info->nr_in + 2; 2256 op = tcg_op_alloc(INDEX_op_call, total_args); 2257 2258 #ifdef CONFIG_PLUGIN 2259 /* Flag helpers that may affect guest state */ 2260 if (tcg_ctx->plugin_insn && !(info->flags & TCG_CALL_NO_SIDE_EFFECTS)) { 2261 tcg_ctx->plugin_insn->calls_helpers = true; 2262 } 2263 #endif 2264 2265 TCGOP_CALLO(op) = n = info->nr_out; 2266 switch (n) { 2267 case 0: 2268 tcg_debug_assert(ret == NULL); 2269 break; 2270 case 1: 2271 tcg_debug_assert(ret != NULL); 2272 op->args[pi++] = temp_arg(ret); 2273 break; 2274 case 2: 2275 case 4: 2276 tcg_debug_assert(ret != NULL); 2277 tcg_debug_assert(ret->base_type == ret->type + ctz32(n)); 2278 tcg_debug_assert(ret->temp_subindex == 0); 2279 for (i = 0; i < n; ++i) { 2280 op->args[pi++] = temp_arg(ret + i); 2281 } 2282 break; 2283 default: 2284 g_assert_not_reached(); 2285 } 2286 2287 TCGOP_CALLI(op) = n = info->nr_in; 2288 for (i = 0; i < n; i++) { 2289 const TCGCallArgumentLoc *loc = &info->in[i]; 2290 TCGTemp *ts = args[loc->arg_idx] + loc->tmp_subindex; 2291 2292 switch (loc->kind) { 2293 case TCG_CALL_ARG_NORMAL: 2294 case TCG_CALL_ARG_BY_REF: 2295 case TCG_CALL_ARG_BY_REF_N: 2296 op->args[pi++] = temp_arg(ts); 2297 break; 2298 2299 case TCG_CALL_ARG_EXTEND_U: 2300 case TCG_CALL_ARG_EXTEND_S: 2301 { 2302 TCGv_i64 temp = tcg_temp_ebb_new_i64(); 2303 TCGv_i32 orig = temp_tcgv_i32(ts); 2304 2305 if (loc->kind == TCG_CALL_ARG_EXTEND_S) { 2306 tcg_gen_ext_i32_i64(temp, orig); 2307 } else { 2308 tcg_gen_extu_i32_i64(temp, orig); 2309 } 2310 op->args[pi++] = tcgv_i64_arg(temp); 2311 extend_free[n_extend++] = temp; 2312 } 2313 break; 2314 2315 default: 2316 g_assert_not_reached(); 2317 } 2318 } 2319 op->args[pi++] = (uintptr_t)func; 2320 op->args[pi++] = (uintptr_t)info; 2321 tcg_debug_assert(pi == total_args); 2322 2323 if (tcg_ctx->emit_before_op) { 2324 QTAILQ_INSERT_BEFORE(tcg_ctx->emit_before_op, op, link); 2325 } else { 2326 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 2327 } 2328 2329 tcg_debug_assert(n_extend < ARRAY_SIZE(extend_free)); 2330 for (i = 0; i < n_extend; ++i) { 2331 tcg_temp_free_i64(extend_free[i]); 2332 } 2333 } 2334 2335 void tcg_gen_call0(void *func, TCGHelperInfo *info, TCGTemp *ret) 2336 { 2337 tcg_gen_callN(func, info, ret, NULL); 2338 } 2339 2340 void tcg_gen_call1(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1) 2341 { 2342 tcg_gen_callN(func, info, ret, &t1); 2343 } 2344 2345 void tcg_gen_call2(void *func, TCGHelperInfo *info, TCGTemp *ret, 2346 TCGTemp *t1, TCGTemp *t2) 2347 { 2348 TCGTemp *args[2] = { t1, t2 }; 2349 tcg_gen_callN(func, info, ret, args); 2350 } 2351 2352 void tcg_gen_call3(void *func, TCGHelperInfo *info, TCGTemp *ret, 2353 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3) 2354 { 2355 TCGTemp *args[3] = { t1, t2, t3 }; 2356 tcg_gen_callN(func, info, ret, args); 2357 } 2358 2359 void tcg_gen_call4(void *func, TCGHelperInfo *info, TCGTemp *ret, 2360 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3, TCGTemp *t4) 2361 { 2362 TCGTemp *args[4] = { t1, t2, t3, t4 }; 2363 tcg_gen_callN(func, info, ret, args); 2364 } 2365 2366 void tcg_gen_call5(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2367 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, TCGTemp *t5) 2368 { 2369 TCGTemp *args[5] = { t1, t2, t3, t4, t5 }; 2370 tcg_gen_callN(func, info, ret, args); 2371 } 2372 2373 void tcg_gen_call6(void *func, TCGHelperInfo *info, TCGTemp *ret, 2374 TCGTemp *t1, TCGTemp *t2, TCGTemp *t3, 2375 TCGTemp *t4, TCGTemp *t5, TCGTemp *t6) 2376 { 2377 TCGTemp *args[6] = { t1, t2, t3, t4, t5, t6 }; 2378 tcg_gen_callN(func, info, ret, args); 2379 } 2380 2381 void tcg_gen_call7(void *func, TCGHelperInfo *info, TCGTemp *ret, TCGTemp *t1, 2382 TCGTemp *t2, TCGTemp *t3, TCGTemp *t4, 2383 TCGTemp *t5, TCGTemp *t6, TCGTemp *t7) 2384 { 2385 TCGTemp *args[7] = { t1, t2, t3, t4, t5, t6, t7 }; 2386 tcg_gen_callN(func, info, ret, args); 2387 } 2388 2389 static void tcg_reg_alloc_start(TCGContext *s) 2390 { 2391 int i, n; 2392 2393 for (i = 0, n = s->nb_temps; i < n; i++) { 2394 TCGTemp *ts = &s->temps[i]; 2395 TCGTempVal val = TEMP_VAL_MEM; 2396 2397 switch (ts->kind) { 2398 case TEMP_CONST: 2399 val = TEMP_VAL_CONST; 2400 break; 2401 case TEMP_FIXED: 2402 val = TEMP_VAL_REG; 2403 break; 2404 case TEMP_GLOBAL: 2405 break; 2406 case TEMP_EBB: 2407 val = TEMP_VAL_DEAD; 2408 /* fall through */ 2409 case TEMP_TB: 2410 ts->mem_allocated = 0; 2411 break; 2412 default: 2413 g_assert_not_reached(); 2414 } 2415 ts->val_type = val; 2416 } 2417 2418 memset(s->reg_to_temp, 0, sizeof(s->reg_to_temp)); 2419 } 2420 2421 static char *tcg_get_arg_str_ptr(TCGContext *s, char *buf, int buf_size, 2422 TCGTemp *ts) 2423 { 2424 int idx = temp_idx(ts); 2425 2426 switch (ts->kind) { 2427 case TEMP_FIXED: 2428 case TEMP_GLOBAL: 2429 pstrcpy(buf, buf_size, ts->name); 2430 break; 2431 case TEMP_TB: 2432 snprintf(buf, buf_size, "loc%d", idx - s->nb_globals); 2433 break; 2434 case TEMP_EBB: 2435 snprintf(buf, buf_size, "tmp%d", idx - s->nb_globals); 2436 break; 2437 case TEMP_CONST: 2438 switch (ts->type) { 2439 case TCG_TYPE_I32: 2440 snprintf(buf, buf_size, "$0x%x", (int32_t)ts->val); 2441 break; 2442 #if TCG_TARGET_REG_BITS > 32 2443 case TCG_TYPE_I64: 2444 snprintf(buf, buf_size, "$0x%" PRIx64, ts->val); 2445 break; 2446 #endif 2447 case TCG_TYPE_V64: 2448 case TCG_TYPE_V128: 2449 case TCG_TYPE_V256: 2450 snprintf(buf, buf_size, "v%d$0x%" PRIx64, 2451 64 << (ts->type - TCG_TYPE_V64), ts->val); 2452 break; 2453 default: 2454 g_assert_not_reached(); 2455 } 2456 break; 2457 } 2458 return buf; 2459 } 2460 2461 static char *tcg_get_arg_str(TCGContext *s, char *buf, 2462 int buf_size, TCGArg arg) 2463 { 2464 return tcg_get_arg_str_ptr(s, buf, buf_size, arg_temp(arg)); 2465 } 2466 2467 static const char * const cond_name[] = 2468 { 2469 [TCG_COND_NEVER] = "never", 2470 [TCG_COND_ALWAYS] = "always", 2471 [TCG_COND_EQ] = "eq", 2472 [TCG_COND_NE] = "ne", 2473 [TCG_COND_LT] = "lt", 2474 [TCG_COND_GE] = "ge", 2475 [TCG_COND_LE] = "le", 2476 [TCG_COND_GT] = "gt", 2477 [TCG_COND_LTU] = "ltu", 2478 [TCG_COND_GEU] = "geu", 2479 [TCG_COND_LEU] = "leu", 2480 [TCG_COND_GTU] = "gtu", 2481 [TCG_COND_TSTEQ] = "tsteq", 2482 [TCG_COND_TSTNE] = "tstne", 2483 }; 2484 2485 static const char * const ldst_name[(MO_BSWAP | MO_SSIZE) + 1] = 2486 { 2487 [MO_UB] = "ub", 2488 [MO_SB] = "sb", 2489 [MO_LEUW] = "leuw", 2490 [MO_LESW] = "lesw", 2491 [MO_LEUL] = "leul", 2492 [MO_LESL] = "lesl", 2493 [MO_LEUQ] = "leq", 2494 [MO_BEUW] = "beuw", 2495 [MO_BESW] = "besw", 2496 [MO_BEUL] = "beul", 2497 [MO_BESL] = "besl", 2498 [MO_BEUQ] = "beq", 2499 [MO_128 + MO_BE] = "beo", 2500 [MO_128 + MO_LE] = "leo", 2501 }; 2502 2503 static const char * const alignment_name[(MO_AMASK >> MO_ASHIFT) + 1] = { 2504 [MO_UNALN >> MO_ASHIFT] = "un+", 2505 [MO_ALIGN >> MO_ASHIFT] = "al+", 2506 [MO_ALIGN_2 >> MO_ASHIFT] = "al2+", 2507 [MO_ALIGN_4 >> MO_ASHIFT] = "al4+", 2508 [MO_ALIGN_8 >> MO_ASHIFT] = "al8+", 2509 [MO_ALIGN_16 >> MO_ASHIFT] = "al16+", 2510 [MO_ALIGN_32 >> MO_ASHIFT] = "al32+", 2511 [MO_ALIGN_64 >> MO_ASHIFT] = "al64+", 2512 }; 2513 2514 static const char * const atom_name[(MO_ATOM_MASK >> MO_ATOM_SHIFT) + 1] = { 2515 [MO_ATOM_IFALIGN >> MO_ATOM_SHIFT] = "", 2516 [MO_ATOM_IFALIGN_PAIR >> MO_ATOM_SHIFT] = "pair+", 2517 [MO_ATOM_WITHIN16 >> MO_ATOM_SHIFT] = "w16+", 2518 [MO_ATOM_WITHIN16_PAIR >> MO_ATOM_SHIFT] = "w16p+", 2519 [MO_ATOM_SUBALIGN >> MO_ATOM_SHIFT] = "sub+", 2520 [MO_ATOM_NONE >> MO_ATOM_SHIFT] = "noat+", 2521 }; 2522 2523 static const char bswap_flag_name[][6] = { 2524 [TCG_BSWAP_IZ] = "iz", 2525 [TCG_BSWAP_OZ] = "oz", 2526 [TCG_BSWAP_OS] = "os", 2527 [TCG_BSWAP_IZ | TCG_BSWAP_OZ] = "iz,oz", 2528 [TCG_BSWAP_IZ | TCG_BSWAP_OS] = "iz,os", 2529 }; 2530 2531 #ifdef CONFIG_PLUGIN 2532 static const char * const plugin_from_name[] = { 2533 "from-tb", 2534 "from-insn", 2535 "after-insn", 2536 "after-tb", 2537 }; 2538 #endif 2539 2540 static inline bool tcg_regset_single(TCGRegSet d) 2541 { 2542 return (d & (d - 1)) == 0; 2543 } 2544 2545 static inline TCGReg tcg_regset_first(TCGRegSet d) 2546 { 2547 if (TCG_TARGET_NB_REGS <= 32) { 2548 return ctz32(d); 2549 } else { 2550 return ctz64(d); 2551 } 2552 } 2553 2554 /* Return only the number of characters output -- no error return. */ 2555 #define ne_fprintf(...) \ 2556 ({ int ret_ = fprintf(__VA_ARGS__); ret_ >= 0 ? ret_ : 0; }) 2557 2558 void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs) 2559 { 2560 char buf[128]; 2561 TCGOp *op; 2562 2563 QTAILQ_FOREACH(op, &s->ops, link) { 2564 int i, k, nb_oargs, nb_iargs, nb_cargs; 2565 const TCGOpDef *def; 2566 TCGOpcode c; 2567 int col = 0; 2568 2569 c = op->opc; 2570 def = &tcg_op_defs[c]; 2571 2572 if (c == INDEX_op_insn_start) { 2573 nb_oargs = 0; 2574 col += ne_fprintf(f, "\n ----"); 2575 2576 for (i = 0, k = s->insn_start_words; i < k; ++i) { 2577 col += ne_fprintf(f, " %016" PRIx64, 2578 tcg_get_insn_start_param(op, i)); 2579 } 2580 } else if (c == INDEX_op_call) { 2581 const TCGHelperInfo *info = tcg_call_info(op); 2582 void *func = tcg_call_func(op); 2583 2584 /* variable number of arguments */ 2585 nb_oargs = TCGOP_CALLO(op); 2586 nb_iargs = TCGOP_CALLI(op); 2587 nb_cargs = def->nb_cargs; 2588 2589 col += ne_fprintf(f, " %s ", def->name); 2590 2591 /* 2592 * Print the function name from TCGHelperInfo, if available. 2593 * Note that plugins have a template function for the info, 2594 * but the actual function pointer comes from the plugin. 2595 */ 2596 if (func == info->func) { 2597 col += ne_fprintf(f, "%s", info->name); 2598 } else { 2599 col += ne_fprintf(f, "plugin(%p)", func); 2600 } 2601 2602 col += ne_fprintf(f, ",$0x%x,$%d", info->flags, nb_oargs); 2603 for (i = 0; i < nb_oargs; i++) { 2604 col += ne_fprintf(f, ",%s", tcg_get_arg_str(s, buf, sizeof(buf), 2605 op->args[i])); 2606 } 2607 for (i = 0; i < nb_iargs; i++) { 2608 TCGArg arg = op->args[nb_oargs + i]; 2609 const char *t = tcg_get_arg_str(s, buf, sizeof(buf), arg); 2610 col += ne_fprintf(f, ",%s", t); 2611 } 2612 } else { 2613 col += ne_fprintf(f, " %s ", def->name); 2614 2615 nb_oargs = def->nb_oargs; 2616 nb_iargs = def->nb_iargs; 2617 nb_cargs = def->nb_cargs; 2618 2619 if (def->flags & TCG_OPF_VECTOR) { 2620 col += ne_fprintf(f, "v%d,e%d,", 64 << TCGOP_VECL(op), 2621 8 << TCGOP_VECE(op)); 2622 } 2623 2624 k = 0; 2625 for (i = 0; i < nb_oargs; i++) { 2626 const char *sep = k ? "," : ""; 2627 col += ne_fprintf(f, "%s%s", sep, 2628 tcg_get_arg_str(s, buf, sizeof(buf), 2629 op->args[k++])); 2630 } 2631 for (i = 0; i < nb_iargs; i++) { 2632 const char *sep = k ? "," : ""; 2633 col += ne_fprintf(f, "%s%s", sep, 2634 tcg_get_arg_str(s, buf, sizeof(buf), 2635 op->args[k++])); 2636 } 2637 switch (c) { 2638 case INDEX_op_brcond_i32: 2639 case INDEX_op_setcond_i32: 2640 case INDEX_op_negsetcond_i32: 2641 case INDEX_op_movcond_i32: 2642 case INDEX_op_brcond2_i32: 2643 case INDEX_op_setcond2_i32: 2644 case INDEX_op_brcond_i64: 2645 case INDEX_op_setcond_i64: 2646 case INDEX_op_negsetcond_i64: 2647 case INDEX_op_movcond_i64: 2648 case INDEX_op_cmp_vec: 2649 case INDEX_op_cmpsel_vec: 2650 if (op->args[k] < ARRAY_SIZE(cond_name) 2651 && cond_name[op->args[k]]) { 2652 col += ne_fprintf(f, ",%s", cond_name[op->args[k++]]); 2653 } else { 2654 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, op->args[k++]); 2655 } 2656 i = 1; 2657 break; 2658 case INDEX_op_qemu_ld_a32_i32: 2659 case INDEX_op_qemu_ld_a64_i32: 2660 case INDEX_op_qemu_st_a32_i32: 2661 case INDEX_op_qemu_st_a64_i32: 2662 case INDEX_op_qemu_st8_a32_i32: 2663 case INDEX_op_qemu_st8_a64_i32: 2664 case INDEX_op_qemu_ld_a32_i64: 2665 case INDEX_op_qemu_ld_a64_i64: 2666 case INDEX_op_qemu_st_a32_i64: 2667 case INDEX_op_qemu_st_a64_i64: 2668 case INDEX_op_qemu_ld_a32_i128: 2669 case INDEX_op_qemu_ld_a64_i128: 2670 case INDEX_op_qemu_st_a32_i128: 2671 case INDEX_op_qemu_st_a64_i128: 2672 { 2673 const char *s_al, *s_op, *s_at; 2674 MemOpIdx oi = op->args[k++]; 2675 MemOp mop = get_memop(oi); 2676 unsigned ix = get_mmuidx(oi); 2677 2678 s_al = alignment_name[(mop & MO_AMASK) >> MO_ASHIFT]; 2679 s_op = ldst_name[mop & (MO_BSWAP | MO_SSIZE)]; 2680 s_at = atom_name[(mop & MO_ATOM_MASK) >> MO_ATOM_SHIFT]; 2681 mop &= ~(MO_AMASK | MO_BSWAP | MO_SSIZE | MO_ATOM_MASK); 2682 2683 /* If all fields are accounted for, print symbolically. */ 2684 if (!mop && s_al && s_op && s_at) { 2685 col += ne_fprintf(f, ",%s%s%s,%u", 2686 s_at, s_al, s_op, ix); 2687 } else { 2688 mop = get_memop(oi); 2689 col += ne_fprintf(f, ",$0x%x,%u", mop, ix); 2690 } 2691 i = 1; 2692 } 2693 break; 2694 case INDEX_op_bswap16_i32: 2695 case INDEX_op_bswap16_i64: 2696 case INDEX_op_bswap32_i32: 2697 case INDEX_op_bswap32_i64: 2698 case INDEX_op_bswap64_i64: 2699 { 2700 TCGArg flags = op->args[k]; 2701 const char *name = NULL; 2702 2703 if (flags < ARRAY_SIZE(bswap_flag_name)) { 2704 name = bswap_flag_name[flags]; 2705 } 2706 if (name) { 2707 col += ne_fprintf(f, ",%s", name); 2708 } else { 2709 col += ne_fprintf(f, ",$0x%" TCG_PRIlx, flags); 2710 } 2711 i = k = 1; 2712 } 2713 break; 2714 #ifdef CONFIG_PLUGIN 2715 case INDEX_op_plugin_cb: 2716 { 2717 TCGArg from = op->args[k++]; 2718 const char *name = NULL; 2719 2720 if (from < ARRAY_SIZE(plugin_from_name)) { 2721 name = plugin_from_name[from]; 2722 } 2723 if (name) { 2724 col += ne_fprintf(f, "%s", name); 2725 } else { 2726 col += ne_fprintf(f, "$0x%" TCG_PRIlx, from); 2727 } 2728 i = 1; 2729 } 2730 break; 2731 #endif 2732 default: 2733 i = 0; 2734 break; 2735 } 2736 switch (c) { 2737 case INDEX_op_set_label: 2738 case INDEX_op_br: 2739 case INDEX_op_brcond_i32: 2740 case INDEX_op_brcond_i64: 2741 case INDEX_op_brcond2_i32: 2742 col += ne_fprintf(f, "%s$L%d", k ? "," : "", 2743 arg_label(op->args[k])->id); 2744 i++, k++; 2745 break; 2746 case INDEX_op_mb: 2747 { 2748 TCGBar membar = op->args[k]; 2749 const char *b_op, *m_op; 2750 2751 switch (membar & TCG_BAR_SC) { 2752 case 0: 2753 b_op = "none"; 2754 break; 2755 case TCG_BAR_LDAQ: 2756 b_op = "acq"; 2757 break; 2758 case TCG_BAR_STRL: 2759 b_op = "rel"; 2760 break; 2761 case TCG_BAR_SC: 2762 b_op = "seq"; 2763 break; 2764 default: 2765 g_assert_not_reached(); 2766 } 2767 2768 switch (membar & TCG_MO_ALL) { 2769 case 0: 2770 m_op = "none"; 2771 break; 2772 case TCG_MO_LD_LD: 2773 m_op = "rr"; 2774 break; 2775 case TCG_MO_LD_ST: 2776 m_op = "rw"; 2777 break; 2778 case TCG_MO_ST_LD: 2779 m_op = "wr"; 2780 break; 2781 case TCG_MO_ST_ST: 2782 m_op = "ww"; 2783 break; 2784 case TCG_MO_LD_LD | TCG_MO_LD_ST: 2785 m_op = "rr+rw"; 2786 break; 2787 case TCG_MO_LD_LD | TCG_MO_ST_LD: 2788 m_op = "rr+wr"; 2789 break; 2790 case TCG_MO_LD_LD | TCG_MO_ST_ST: 2791 m_op = "rr+ww"; 2792 break; 2793 case TCG_MO_LD_ST | TCG_MO_ST_LD: 2794 m_op = "rw+wr"; 2795 break; 2796 case TCG_MO_LD_ST | TCG_MO_ST_ST: 2797 m_op = "rw+ww"; 2798 break; 2799 case TCG_MO_ST_LD | TCG_MO_ST_ST: 2800 m_op = "wr+ww"; 2801 break; 2802 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_LD: 2803 m_op = "rr+rw+wr"; 2804 break; 2805 case TCG_MO_LD_LD | TCG_MO_LD_ST | TCG_MO_ST_ST: 2806 m_op = "rr+rw+ww"; 2807 break; 2808 case TCG_MO_LD_LD | TCG_MO_ST_LD | TCG_MO_ST_ST: 2809 m_op = "rr+wr+ww"; 2810 break; 2811 case TCG_MO_LD_ST | TCG_MO_ST_LD | TCG_MO_ST_ST: 2812 m_op = "rw+wr+ww"; 2813 break; 2814 case TCG_MO_ALL: 2815 m_op = "all"; 2816 break; 2817 default: 2818 g_assert_not_reached(); 2819 } 2820 2821 col += ne_fprintf(f, "%s%s:%s", (k ? "," : ""), b_op, m_op); 2822 i++, k++; 2823 } 2824 break; 2825 default: 2826 break; 2827 } 2828 for (; i < nb_cargs; i++, k++) { 2829 col += ne_fprintf(f, "%s$0x%" TCG_PRIlx, k ? "," : "", 2830 op->args[k]); 2831 } 2832 } 2833 2834 if (have_prefs || op->life) { 2835 for (; col < 40; ++col) { 2836 putc(' ', f); 2837 } 2838 } 2839 2840 if (op->life) { 2841 unsigned life = op->life; 2842 2843 if (life & (SYNC_ARG * 3)) { 2844 ne_fprintf(f, " sync:"); 2845 for (i = 0; i < 2; ++i) { 2846 if (life & (SYNC_ARG << i)) { 2847 ne_fprintf(f, " %d", i); 2848 } 2849 } 2850 } 2851 life /= DEAD_ARG; 2852 if (life) { 2853 ne_fprintf(f, " dead:"); 2854 for (i = 0; life; ++i, life >>= 1) { 2855 if (life & 1) { 2856 ne_fprintf(f, " %d", i); 2857 } 2858 } 2859 } 2860 } 2861 2862 if (have_prefs) { 2863 for (i = 0; i < nb_oargs; ++i) { 2864 TCGRegSet set = output_pref(op, i); 2865 2866 if (i == 0) { 2867 ne_fprintf(f, " pref="); 2868 } else { 2869 ne_fprintf(f, ","); 2870 } 2871 if (set == 0) { 2872 ne_fprintf(f, "none"); 2873 } else if (set == MAKE_64BIT_MASK(0, TCG_TARGET_NB_REGS)) { 2874 ne_fprintf(f, "all"); 2875 #ifdef CONFIG_DEBUG_TCG 2876 } else if (tcg_regset_single(set)) { 2877 TCGReg reg = tcg_regset_first(set); 2878 ne_fprintf(f, "%s", tcg_target_reg_names[reg]); 2879 #endif 2880 } else if (TCG_TARGET_NB_REGS <= 32) { 2881 ne_fprintf(f, "0x%x", (uint32_t)set); 2882 } else { 2883 ne_fprintf(f, "0x%" PRIx64, (uint64_t)set); 2884 } 2885 } 2886 } 2887 2888 putc('\n', f); 2889 } 2890 } 2891 2892 /* we give more priority to constraints with less registers */ 2893 static int get_constraint_priority(const TCGOpDef *def, int k) 2894 { 2895 const TCGArgConstraint *arg_ct = &def->args_ct[k]; 2896 int n = ctpop64(arg_ct->regs); 2897 2898 /* 2899 * Sort constraints of a single register first, which includes output 2900 * aliases (which must exactly match the input already allocated). 2901 */ 2902 if (n == 1 || arg_ct->oalias) { 2903 return INT_MAX; 2904 } 2905 2906 /* 2907 * Sort register pairs next, first then second immediately after. 2908 * Arbitrarily sort multiple pairs by the index of the first reg; 2909 * there shouldn't be many pairs. 2910 */ 2911 switch (arg_ct->pair) { 2912 case 1: 2913 case 3: 2914 return (k + 1) * 2; 2915 case 2: 2916 return (arg_ct->pair_index + 1) * 2 - 1; 2917 } 2918 2919 /* Finally, sort by decreasing register count. */ 2920 assert(n > 1); 2921 return -n; 2922 } 2923 2924 /* sort from highest priority to lowest */ 2925 static void sort_constraints(TCGOpDef *def, int start, int n) 2926 { 2927 int i, j; 2928 TCGArgConstraint *a = def->args_ct; 2929 2930 for (i = 0; i < n; i++) { 2931 a[start + i].sort_index = start + i; 2932 } 2933 if (n <= 1) { 2934 return; 2935 } 2936 for (i = 0; i < n - 1; i++) { 2937 for (j = i + 1; j < n; j++) { 2938 int p1 = get_constraint_priority(def, a[start + i].sort_index); 2939 int p2 = get_constraint_priority(def, a[start + j].sort_index); 2940 if (p1 < p2) { 2941 int tmp = a[start + i].sort_index; 2942 a[start + i].sort_index = a[start + j].sort_index; 2943 a[start + j].sort_index = tmp; 2944 } 2945 } 2946 } 2947 } 2948 2949 static void process_op_defs(TCGContext *s) 2950 { 2951 TCGOpcode op; 2952 2953 for (op = 0; op < NB_OPS; op++) { 2954 TCGOpDef *def = &tcg_op_defs[op]; 2955 const TCGTargetOpDef *tdefs; 2956 bool saw_alias_pair = false; 2957 int i, o, i2, o2, nb_args; 2958 2959 if (def->flags & TCG_OPF_NOT_PRESENT) { 2960 continue; 2961 } 2962 2963 nb_args = def->nb_iargs + def->nb_oargs; 2964 if (nb_args == 0) { 2965 continue; 2966 } 2967 2968 /* 2969 * Macro magic should make it impossible, but double-check that 2970 * the array index is in range. Since the signness of an enum 2971 * is implementation defined, force the result to unsigned. 2972 */ 2973 unsigned con_set = tcg_target_op_def(op); 2974 tcg_debug_assert(con_set < ARRAY_SIZE(constraint_sets)); 2975 tdefs = &constraint_sets[con_set]; 2976 2977 for (i = 0; i < nb_args; i++) { 2978 const char *ct_str = tdefs->args_ct_str[i]; 2979 bool input_p = i >= def->nb_oargs; 2980 2981 /* Incomplete TCGTargetOpDef entry. */ 2982 tcg_debug_assert(ct_str != NULL); 2983 2984 switch (*ct_str) { 2985 case '0' ... '9': 2986 o = *ct_str - '0'; 2987 tcg_debug_assert(input_p); 2988 tcg_debug_assert(o < def->nb_oargs); 2989 tcg_debug_assert(def->args_ct[o].regs != 0); 2990 tcg_debug_assert(!def->args_ct[o].oalias); 2991 def->args_ct[i] = def->args_ct[o]; 2992 /* The output sets oalias. */ 2993 def->args_ct[o].oalias = 1; 2994 def->args_ct[o].alias_index = i; 2995 /* The input sets ialias. */ 2996 def->args_ct[i].ialias = 1; 2997 def->args_ct[i].alias_index = o; 2998 if (def->args_ct[i].pair) { 2999 saw_alias_pair = true; 3000 } 3001 tcg_debug_assert(ct_str[1] == '\0'); 3002 continue; 3003 3004 case '&': 3005 tcg_debug_assert(!input_p); 3006 def->args_ct[i].newreg = true; 3007 ct_str++; 3008 break; 3009 3010 case 'p': /* plus */ 3011 /* Allocate to the register after the previous. */ 3012 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0)); 3013 o = i - 1; 3014 tcg_debug_assert(!def->args_ct[o].pair); 3015 tcg_debug_assert(!def->args_ct[o].ct); 3016 def->args_ct[i] = (TCGArgConstraint){ 3017 .pair = 2, 3018 .pair_index = o, 3019 .regs = def->args_ct[o].regs << 1, 3020 .newreg = def->args_ct[o].newreg, 3021 }; 3022 def->args_ct[o].pair = 1; 3023 def->args_ct[o].pair_index = i; 3024 tcg_debug_assert(ct_str[1] == '\0'); 3025 continue; 3026 3027 case 'm': /* minus */ 3028 /* Allocate to the register before the previous. */ 3029 tcg_debug_assert(i > (input_p ? def->nb_oargs : 0)); 3030 o = i - 1; 3031 tcg_debug_assert(!def->args_ct[o].pair); 3032 tcg_debug_assert(!def->args_ct[o].ct); 3033 def->args_ct[i] = (TCGArgConstraint){ 3034 .pair = 1, 3035 .pair_index = o, 3036 .regs = def->args_ct[o].regs >> 1, 3037 .newreg = def->args_ct[o].newreg, 3038 }; 3039 def->args_ct[o].pair = 2; 3040 def->args_ct[o].pair_index = i; 3041 tcg_debug_assert(ct_str[1] == '\0'); 3042 continue; 3043 } 3044 3045 do { 3046 switch (*ct_str) { 3047 case 'i': 3048 def->args_ct[i].ct |= TCG_CT_CONST; 3049 break; 3050 3051 /* Include all of the target-specific constraints. */ 3052 3053 #undef CONST 3054 #define CONST(CASE, MASK) \ 3055 case CASE: def->args_ct[i].ct |= MASK; break; 3056 #define REGS(CASE, MASK) \ 3057 case CASE: def->args_ct[i].regs |= MASK; break; 3058 3059 #include "tcg-target-con-str.h" 3060 3061 #undef REGS 3062 #undef CONST 3063 default: 3064 case '0' ... '9': 3065 case '&': 3066 case 'p': 3067 case 'm': 3068 /* Typo in TCGTargetOpDef constraint. */ 3069 g_assert_not_reached(); 3070 } 3071 } while (*++ct_str != '\0'); 3072 } 3073 3074 /* TCGTargetOpDef entry with too much information? */ 3075 tcg_debug_assert(i == TCG_MAX_OP_ARGS || tdefs->args_ct_str[i] == NULL); 3076 3077 /* 3078 * Fix up output pairs that are aliased with inputs. 3079 * When we created the alias, we copied pair from the output. 3080 * There are three cases: 3081 * (1a) Pairs of inputs alias pairs of outputs. 3082 * (1b) One input aliases the first of a pair of outputs. 3083 * (2) One input aliases the second of a pair of outputs. 3084 * 3085 * Case 1a is handled by making sure that the pair_index'es are 3086 * properly updated so that they appear the same as a pair of inputs. 3087 * 3088 * Case 1b is handled by setting the pair_index of the input to 3089 * itself, simply so it doesn't point to an unrelated argument. 3090 * Since we don't encounter the "second" during the input allocation 3091 * phase, nothing happens with the second half of the input pair. 3092 * 3093 * Case 2 is handled by setting the second input to pair=3, the 3094 * first output to pair=3, and the pair_index'es to match. 3095 */ 3096 if (saw_alias_pair) { 3097 for (i = def->nb_oargs; i < nb_args; i++) { 3098 /* 3099 * Since [0-9pm] must be alone in the constraint string, 3100 * the only way they can both be set is if the pair comes 3101 * from the output alias. 3102 */ 3103 if (!def->args_ct[i].ialias) { 3104 continue; 3105 } 3106 switch (def->args_ct[i].pair) { 3107 case 0: 3108 break; 3109 case 1: 3110 o = def->args_ct[i].alias_index; 3111 o2 = def->args_ct[o].pair_index; 3112 tcg_debug_assert(def->args_ct[o].pair == 1); 3113 tcg_debug_assert(def->args_ct[o2].pair == 2); 3114 if (def->args_ct[o2].oalias) { 3115 /* Case 1a */ 3116 i2 = def->args_ct[o2].alias_index; 3117 tcg_debug_assert(def->args_ct[i2].pair == 2); 3118 def->args_ct[i2].pair_index = i; 3119 def->args_ct[i].pair_index = i2; 3120 } else { 3121 /* Case 1b */ 3122 def->args_ct[i].pair_index = i; 3123 } 3124 break; 3125 case 2: 3126 o = def->args_ct[i].alias_index; 3127 o2 = def->args_ct[o].pair_index; 3128 tcg_debug_assert(def->args_ct[o].pair == 2); 3129 tcg_debug_assert(def->args_ct[o2].pair == 1); 3130 if (def->args_ct[o2].oalias) { 3131 /* Case 1a */ 3132 i2 = def->args_ct[o2].alias_index; 3133 tcg_debug_assert(def->args_ct[i2].pair == 1); 3134 def->args_ct[i2].pair_index = i; 3135 def->args_ct[i].pair_index = i2; 3136 } else { 3137 /* Case 2 */ 3138 def->args_ct[i].pair = 3; 3139 def->args_ct[o2].pair = 3; 3140 def->args_ct[i].pair_index = o2; 3141 def->args_ct[o2].pair_index = i; 3142 } 3143 break; 3144 default: 3145 g_assert_not_reached(); 3146 } 3147 } 3148 } 3149 3150 /* sort the constraints (XXX: this is just an heuristic) */ 3151 sort_constraints(def, 0, def->nb_oargs); 3152 sort_constraints(def, def->nb_oargs, def->nb_iargs); 3153 } 3154 } 3155 3156 static void remove_label_use(TCGOp *op, int idx) 3157 { 3158 TCGLabel *label = arg_label(op->args[idx]); 3159 TCGLabelUse *use; 3160 3161 QSIMPLEQ_FOREACH(use, &label->branches, next) { 3162 if (use->op == op) { 3163 QSIMPLEQ_REMOVE(&label->branches, use, TCGLabelUse, next); 3164 return; 3165 } 3166 } 3167 g_assert_not_reached(); 3168 } 3169 3170 void tcg_op_remove(TCGContext *s, TCGOp *op) 3171 { 3172 switch (op->opc) { 3173 case INDEX_op_br: 3174 remove_label_use(op, 0); 3175 break; 3176 case INDEX_op_brcond_i32: 3177 case INDEX_op_brcond_i64: 3178 remove_label_use(op, 3); 3179 break; 3180 case INDEX_op_brcond2_i32: 3181 remove_label_use(op, 5); 3182 break; 3183 default: 3184 break; 3185 } 3186 3187 QTAILQ_REMOVE(&s->ops, op, link); 3188 QTAILQ_INSERT_TAIL(&s->free_ops, op, link); 3189 s->nb_ops--; 3190 } 3191 3192 void tcg_remove_ops_after(TCGOp *op) 3193 { 3194 TCGContext *s = tcg_ctx; 3195 3196 while (true) { 3197 TCGOp *last = tcg_last_op(); 3198 if (last == op) { 3199 return; 3200 } 3201 tcg_op_remove(s, last); 3202 } 3203 } 3204 3205 static TCGOp *tcg_op_alloc(TCGOpcode opc, unsigned nargs) 3206 { 3207 TCGContext *s = tcg_ctx; 3208 TCGOp *op = NULL; 3209 3210 if (unlikely(!QTAILQ_EMPTY(&s->free_ops))) { 3211 QTAILQ_FOREACH(op, &s->free_ops, link) { 3212 if (nargs <= op->nargs) { 3213 QTAILQ_REMOVE(&s->free_ops, op, link); 3214 nargs = op->nargs; 3215 goto found; 3216 } 3217 } 3218 } 3219 3220 /* Most opcodes have 3 or 4 operands: reduce fragmentation. */ 3221 nargs = MAX(4, nargs); 3222 op = tcg_malloc(sizeof(TCGOp) + sizeof(TCGArg) * nargs); 3223 3224 found: 3225 memset(op, 0, offsetof(TCGOp, link)); 3226 op->opc = opc; 3227 op->nargs = nargs; 3228 3229 /* Check for bitfield overflow. */ 3230 tcg_debug_assert(op->nargs == nargs); 3231 3232 s->nb_ops++; 3233 return op; 3234 } 3235 3236 TCGOp *tcg_emit_op(TCGOpcode opc, unsigned nargs) 3237 { 3238 TCGOp *op = tcg_op_alloc(opc, nargs); 3239 3240 if (tcg_ctx->emit_before_op) { 3241 QTAILQ_INSERT_BEFORE(tcg_ctx->emit_before_op, op, link); 3242 } else { 3243 QTAILQ_INSERT_TAIL(&tcg_ctx->ops, op, link); 3244 } 3245 return op; 3246 } 3247 3248 TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *old_op, 3249 TCGOpcode opc, unsigned nargs) 3250 { 3251 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3252 QTAILQ_INSERT_BEFORE(old_op, new_op, link); 3253 return new_op; 3254 } 3255 3256 TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *old_op, 3257 TCGOpcode opc, unsigned nargs) 3258 { 3259 TCGOp *new_op = tcg_op_alloc(opc, nargs); 3260 QTAILQ_INSERT_AFTER(&s->ops, old_op, new_op, link); 3261 return new_op; 3262 } 3263 3264 static void move_label_uses(TCGLabel *to, TCGLabel *from) 3265 { 3266 TCGLabelUse *u; 3267 3268 QSIMPLEQ_FOREACH(u, &from->branches, next) { 3269 TCGOp *op = u->op; 3270 switch (op->opc) { 3271 case INDEX_op_br: 3272 op->args[0] = label_arg(to); 3273 break; 3274 case INDEX_op_brcond_i32: 3275 case INDEX_op_brcond_i64: 3276 op->args[3] = label_arg(to); 3277 break; 3278 case INDEX_op_brcond2_i32: 3279 op->args[5] = label_arg(to); 3280 break; 3281 default: 3282 g_assert_not_reached(); 3283 } 3284 } 3285 3286 QSIMPLEQ_CONCAT(&to->branches, &from->branches); 3287 } 3288 3289 /* Reachable analysis : remove unreachable code. */ 3290 static void __attribute__((noinline)) 3291 reachable_code_pass(TCGContext *s) 3292 { 3293 TCGOp *op, *op_next, *op_prev; 3294 bool dead = false; 3295 3296 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3297 bool remove = dead; 3298 TCGLabel *label; 3299 3300 switch (op->opc) { 3301 case INDEX_op_set_label: 3302 label = arg_label(op->args[0]); 3303 3304 /* 3305 * Note that the first op in the TB is always a load, 3306 * so there is always something before a label. 3307 */ 3308 op_prev = QTAILQ_PREV(op, link); 3309 3310 /* 3311 * If we find two sequential labels, move all branches to 3312 * reference the second label and remove the first label. 3313 * Do this before branch to next optimization, so that the 3314 * middle label is out of the way. 3315 */ 3316 if (op_prev->opc == INDEX_op_set_label) { 3317 move_label_uses(label, arg_label(op_prev->args[0])); 3318 tcg_op_remove(s, op_prev); 3319 op_prev = QTAILQ_PREV(op, link); 3320 } 3321 3322 /* 3323 * Optimization can fold conditional branches to unconditional. 3324 * If we find a label which is preceded by an unconditional 3325 * branch to next, remove the branch. We couldn't do this when 3326 * processing the branch because any dead code between the branch 3327 * and label had not yet been removed. 3328 */ 3329 if (op_prev->opc == INDEX_op_br && 3330 label == arg_label(op_prev->args[0])) { 3331 tcg_op_remove(s, op_prev); 3332 /* Fall through means insns become live again. */ 3333 dead = false; 3334 } 3335 3336 if (QSIMPLEQ_EMPTY(&label->branches)) { 3337 /* 3338 * While there is an occasional backward branch, virtually 3339 * all branches generated by the translators are forward. 3340 * Which means that generally we will have already removed 3341 * all references to the label that will be, and there is 3342 * little to be gained by iterating. 3343 */ 3344 remove = true; 3345 } else { 3346 /* Once we see a label, insns become live again. */ 3347 dead = false; 3348 remove = false; 3349 } 3350 break; 3351 3352 case INDEX_op_br: 3353 case INDEX_op_exit_tb: 3354 case INDEX_op_goto_ptr: 3355 /* Unconditional branches; everything following is dead. */ 3356 dead = true; 3357 break; 3358 3359 case INDEX_op_call: 3360 /* Notice noreturn helper calls, raising exceptions. */ 3361 if (tcg_call_flags(op) & TCG_CALL_NO_RETURN) { 3362 dead = true; 3363 } 3364 break; 3365 3366 case INDEX_op_insn_start: 3367 /* Never remove -- we need to keep these for unwind. */ 3368 remove = false; 3369 break; 3370 3371 default: 3372 break; 3373 } 3374 3375 if (remove) { 3376 tcg_op_remove(s, op); 3377 } 3378 } 3379 } 3380 3381 #define TS_DEAD 1 3382 #define TS_MEM 2 3383 3384 #define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n))) 3385 #define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n))) 3386 3387 /* For liveness_pass_1, the register preferences for a given temp. */ 3388 static inline TCGRegSet *la_temp_pref(TCGTemp *ts) 3389 { 3390 return ts->state_ptr; 3391 } 3392 3393 /* For liveness_pass_1, reset the preferences for a given temp to the 3394 * maximal regset for its type. 3395 */ 3396 static inline void la_reset_pref(TCGTemp *ts) 3397 { 3398 *la_temp_pref(ts) 3399 = (ts->state == TS_DEAD ? 0 : tcg_target_available_regs[ts->type]); 3400 } 3401 3402 /* liveness analysis: end of function: all temps are dead, and globals 3403 should be in memory. */ 3404 static void la_func_end(TCGContext *s, int ng, int nt) 3405 { 3406 int i; 3407 3408 for (i = 0; i < ng; ++i) { 3409 s->temps[i].state = TS_DEAD | TS_MEM; 3410 la_reset_pref(&s->temps[i]); 3411 } 3412 for (i = ng; i < nt; ++i) { 3413 s->temps[i].state = TS_DEAD; 3414 la_reset_pref(&s->temps[i]); 3415 } 3416 } 3417 3418 /* liveness analysis: end of basic block: all temps are dead, globals 3419 and local temps should be in memory. */ 3420 static void la_bb_end(TCGContext *s, int ng, int nt) 3421 { 3422 int i; 3423 3424 for (i = 0; i < nt; ++i) { 3425 TCGTemp *ts = &s->temps[i]; 3426 int state; 3427 3428 switch (ts->kind) { 3429 case TEMP_FIXED: 3430 case TEMP_GLOBAL: 3431 case TEMP_TB: 3432 state = TS_DEAD | TS_MEM; 3433 break; 3434 case TEMP_EBB: 3435 case TEMP_CONST: 3436 state = TS_DEAD; 3437 break; 3438 default: 3439 g_assert_not_reached(); 3440 } 3441 ts->state = state; 3442 la_reset_pref(ts); 3443 } 3444 } 3445 3446 /* liveness analysis: sync globals back to memory. */ 3447 static void la_global_sync(TCGContext *s, int ng) 3448 { 3449 int i; 3450 3451 for (i = 0; i < ng; ++i) { 3452 int state = s->temps[i].state; 3453 s->temps[i].state = state | TS_MEM; 3454 if (state == TS_DEAD) { 3455 /* If the global was previously dead, reset prefs. */ 3456 la_reset_pref(&s->temps[i]); 3457 } 3458 } 3459 } 3460 3461 /* 3462 * liveness analysis: conditional branch: all temps are dead unless 3463 * explicitly live-across-conditional-branch, globals and local temps 3464 * should be synced. 3465 */ 3466 static void la_bb_sync(TCGContext *s, int ng, int nt) 3467 { 3468 la_global_sync(s, ng); 3469 3470 for (int i = ng; i < nt; ++i) { 3471 TCGTemp *ts = &s->temps[i]; 3472 int state; 3473 3474 switch (ts->kind) { 3475 case TEMP_TB: 3476 state = ts->state; 3477 ts->state = state | TS_MEM; 3478 if (state != TS_DEAD) { 3479 continue; 3480 } 3481 break; 3482 case TEMP_EBB: 3483 case TEMP_CONST: 3484 continue; 3485 default: 3486 g_assert_not_reached(); 3487 } 3488 la_reset_pref(&s->temps[i]); 3489 } 3490 } 3491 3492 /* liveness analysis: sync globals back to memory and kill. */ 3493 static void la_global_kill(TCGContext *s, int ng) 3494 { 3495 int i; 3496 3497 for (i = 0; i < ng; i++) { 3498 s->temps[i].state = TS_DEAD | TS_MEM; 3499 la_reset_pref(&s->temps[i]); 3500 } 3501 } 3502 3503 /* liveness analysis: note live globals crossing calls. */ 3504 static void la_cross_call(TCGContext *s, int nt) 3505 { 3506 TCGRegSet mask = ~tcg_target_call_clobber_regs; 3507 int i; 3508 3509 for (i = 0; i < nt; i++) { 3510 TCGTemp *ts = &s->temps[i]; 3511 if (!(ts->state & TS_DEAD)) { 3512 TCGRegSet *pset = la_temp_pref(ts); 3513 TCGRegSet set = *pset; 3514 3515 set &= mask; 3516 /* If the combination is not possible, restart. */ 3517 if (set == 0) { 3518 set = tcg_target_available_regs[ts->type] & mask; 3519 } 3520 *pset = set; 3521 } 3522 } 3523 } 3524 3525 /* 3526 * Liveness analysis: Verify the lifetime of TEMP_TB, and reduce 3527 * to TEMP_EBB, if possible. 3528 */ 3529 static void __attribute__((noinline)) 3530 liveness_pass_0(TCGContext *s) 3531 { 3532 void * const multiple_ebb = (void *)(uintptr_t)-1; 3533 int nb_temps = s->nb_temps; 3534 TCGOp *op, *ebb; 3535 3536 for (int i = s->nb_globals; i < nb_temps; ++i) { 3537 s->temps[i].state_ptr = NULL; 3538 } 3539 3540 /* 3541 * Represent each EBB by the op at which it begins. In the case of 3542 * the first EBB, this is the first op, otherwise it is a label. 3543 * Collect the uses of each TEMP_TB: NULL for unused, EBB for use 3544 * within a single EBB, else MULTIPLE_EBB. 3545 */ 3546 ebb = QTAILQ_FIRST(&s->ops); 3547 QTAILQ_FOREACH(op, &s->ops, link) { 3548 const TCGOpDef *def; 3549 int nb_oargs, nb_iargs; 3550 3551 switch (op->opc) { 3552 case INDEX_op_set_label: 3553 ebb = op; 3554 continue; 3555 case INDEX_op_discard: 3556 continue; 3557 case INDEX_op_call: 3558 nb_oargs = TCGOP_CALLO(op); 3559 nb_iargs = TCGOP_CALLI(op); 3560 break; 3561 default: 3562 def = &tcg_op_defs[op->opc]; 3563 nb_oargs = def->nb_oargs; 3564 nb_iargs = def->nb_iargs; 3565 break; 3566 } 3567 3568 for (int i = 0; i < nb_oargs + nb_iargs; ++i) { 3569 TCGTemp *ts = arg_temp(op->args[i]); 3570 3571 if (ts->kind != TEMP_TB) { 3572 continue; 3573 } 3574 if (ts->state_ptr == NULL) { 3575 ts->state_ptr = ebb; 3576 } else if (ts->state_ptr != ebb) { 3577 ts->state_ptr = multiple_ebb; 3578 } 3579 } 3580 } 3581 3582 /* 3583 * For TEMP_TB that turned out not to be used beyond one EBB, 3584 * reduce the liveness to TEMP_EBB. 3585 */ 3586 for (int i = s->nb_globals; i < nb_temps; ++i) { 3587 TCGTemp *ts = &s->temps[i]; 3588 if (ts->kind == TEMP_TB && ts->state_ptr != multiple_ebb) { 3589 ts->kind = TEMP_EBB; 3590 } 3591 } 3592 } 3593 3594 /* Liveness analysis : update the opc_arg_life array to tell if a 3595 given input arguments is dead. Instructions updating dead 3596 temporaries are removed. */ 3597 static void __attribute__((noinline)) 3598 liveness_pass_1(TCGContext *s) 3599 { 3600 int nb_globals = s->nb_globals; 3601 int nb_temps = s->nb_temps; 3602 TCGOp *op, *op_prev; 3603 TCGRegSet *prefs; 3604 int i; 3605 3606 prefs = tcg_malloc(sizeof(TCGRegSet) * nb_temps); 3607 for (i = 0; i < nb_temps; ++i) { 3608 s->temps[i].state_ptr = prefs + i; 3609 } 3610 3611 /* ??? Should be redundant with the exit_tb that ends the TB. */ 3612 la_func_end(s, nb_globals, nb_temps); 3613 3614 QTAILQ_FOREACH_REVERSE_SAFE(op, &s->ops, link, op_prev) { 3615 int nb_iargs, nb_oargs; 3616 TCGOpcode opc_new, opc_new2; 3617 bool have_opc_new2; 3618 TCGLifeData arg_life = 0; 3619 TCGTemp *ts; 3620 TCGOpcode opc = op->opc; 3621 const TCGOpDef *def = &tcg_op_defs[opc]; 3622 3623 switch (opc) { 3624 case INDEX_op_call: 3625 { 3626 const TCGHelperInfo *info = tcg_call_info(op); 3627 int call_flags = tcg_call_flags(op); 3628 3629 nb_oargs = TCGOP_CALLO(op); 3630 nb_iargs = TCGOP_CALLI(op); 3631 3632 /* pure functions can be removed if their result is unused */ 3633 if (call_flags & TCG_CALL_NO_SIDE_EFFECTS) { 3634 for (i = 0; i < nb_oargs; i++) { 3635 ts = arg_temp(op->args[i]); 3636 if (ts->state != TS_DEAD) { 3637 goto do_not_remove_call; 3638 } 3639 } 3640 goto do_remove; 3641 } 3642 do_not_remove_call: 3643 3644 /* Output args are dead. */ 3645 for (i = 0; i < nb_oargs; i++) { 3646 ts = arg_temp(op->args[i]); 3647 if (ts->state & TS_DEAD) { 3648 arg_life |= DEAD_ARG << i; 3649 } 3650 if (ts->state & TS_MEM) { 3651 arg_life |= SYNC_ARG << i; 3652 } 3653 ts->state = TS_DEAD; 3654 la_reset_pref(ts); 3655 } 3656 3657 /* Not used -- it will be tcg_target_call_oarg_reg(). */ 3658 memset(op->output_pref, 0, sizeof(op->output_pref)); 3659 3660 if (!(call_flags & (TCG_CALL_NO_WRITE_GLOBALS | 3661 TCG_CALL_NO_READ_GLOBALS))) { 3662 la_global_kill(s, nb_globals); 3663 } else if (!(call_flags & TCG_CALL_NO_READ_GLOBALS)) { 3664 la_global_sync(s, nb_globals); 3665 } 3666 3667 /* Record arguments that die in this helper. */ 3668 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 3669 ts = arg_temp(op->args[i]); 3670 if (ts->state & TS_DEAD) { 3671 arg_life |= DEAD_ARG << i; 3672 } 3673 } 3674 3675 /* For all live registers, remove call-clobbered prefs. */ 3676 la_cross_call(s, nb_temps); 3677 3678 /* 3679 * Input arguments are live for preceding opcodes. 3680 * 3681 * For those arguments that die, and will be allocated in 3682 * registers, clear the register set for that arg, to be 3683 * filled in below. For args that will be on the stack, 3684 * reset to any available reg. Process arguments in reverse 3685 * order so that if a temp is used more than once, the stack 3686 * reset to max happens before the register reset to 0. 3687 */ 3688 for (i = nb_iargs - 1; i >= 0; i--) { 3689 const TCGCallArgumentLoc *loc = &info->in[i]; 3690 ts = arg_temp(op->args[nb_oargs + i]); 3691 3692 if (ts->state & TS_DEAD) { 3693 switch (loc->kind) { 3694 case TCG_CALL_ARG_NORMAL: 3695 case TCG_CALL_ARG_EXTEND_U: 3696 case TCG_CALL_ARG_EXTEND_S: 3697 if (arg_slot_reg_p(loc->arg_slot)) { 3698 *la_temp_pref(ts) = 0; 3699 break; 3700 } 3701 /* fall through */ 3702 default: 3703 *la_temp_pref(ts) = 3704 tcg_target_available_regs[ts->type]; 3705 break; 3706 } 3707 ts->state &= ~TS_DEAD; 3708 } 3709 } 3710 3711 /* 3712 * For each input argument, add its input register to prefs. 3713 * If a temp is used once, this produces a single set bit; 3714 * if a temp is used multiple times, this produces a set. 3715 */ 3716 for (i = 0; i < nb_iargs; i++) { 3717 const TCGCallArgumentLoc *loc = &info->in[i]; 3718 ts = arg_temp(op->args[nb_oargs + i]); 3719 3720 switch (loc->kind) { 3721 case TCG_CALL_ARG_NORMAL: 3722 case TCG_CALL_ARG_EXTEND_U: 3723 case TCG_CALL_ARG_EXTEND_S: 3724 if (arg_slot_reg_p(loc->arg_slot)) { 3725 tcg_regset_set_reg(*la_temp_pref(ts), 3726 tcg_target_call_iarg_regs[loc->arg_slot]); 3727 } 3728 break; 3729 default: 3730 break; 3731 } 3732 } 3733 } 3734 break; 3735 case INDEX_op_insn_start: 3736 break; 3737 case INDEX_op_discard: 3738 /* mark the temporary as dead */ 3739 ts = arg_temp(op->args[0]); 3740 ts->state = TS_DEAD; 3741 la_reset_pref(ts); 3742 break; 3743 3744 case INDEX_op_add2_i32: 3745 opc_new = INDEX_op_add_i32; 3746 goto do_addsub2; 3747 case INDEX_op_sub2_i32: 3748 opc_new = INDEX_op_sub_i32; 3749 goto do_addsub2; 3750 case INDEX_op_add2_i64: 3751 opc_new = INDEX_op_add_i64; 3752 goto do_addsub2; 3753 case INDEX_op_sub2_i64: 3754 opc_new = INDEX_op_sub_i64; 3755 do_addsub2: 3756 nb_iargs = 4; 3757 nb_oargs = 2; 3758 /* Test if the high part of the operation is dead, but not 3759 the low part. The result can be optimized to a simple 3760 add or sub. This happens often for x86_64 guest when the 3761 cpu mode is set to 32 bit. */ 3762 if (arg_temp(op->args[1])->state == TS_DEAD) { 3763 if (arg_temp(op->args[0])->state == TS_DEAD) { 3764 goto do_remove; 3765 } 3766 /* Replace the opcode and adjust the args in place, 3767 leaving 3 unused args at the end. */ 3768 op->opc = opc = opc_new; 3769 op->args[1] = op->args[2]; 3770 op->args[2] = op->args[4]; 3771 /* Fall through and mark the single-word operation live. */ 3772 nb_iargs = 2; 3773 nb_oargs = 1; 3774 } 3775 goto do_not_remove; 3776 3777 case INDEX_op_mulu2_i32: 3778 opc_new = INDEX_op_mul_i32; 3779 opc_new2 = INDEX_op_muluh_i32; 3780 have_opc_new2 = TCG_TARGET_HAS_muluh_i32; 3781 goto do_mul2; 3782 case INDEX_op_muls2_i32: 3783 opc_new = INDEX_op_mul_i32; 3784 opc_new2 = INDEX_op_mulsh_i32; 3785 have_opc_new2 = TCG_TARGET_HAS_mulsh_i32; 3786 goto do_mul2; 3787 case INDEX_op_mulu2_i64: 3788 opc_new = INDEX_op_mul_i64; 3789 opc_new2 = INDEX_op_muluh_i64; 3790 have_opc_new2 = TCG_TARGET_HAS_muluh_i64; 3791 goto do_mul2; 3792 case INDEX_op_muls2_i64: 3793 opc_new = INDEX_op_mul_i64; 3794 opc_new2 = INDEX_op_mulsh_i64; 3795 have_opc_new2 = TCG_TARGET_HAS_mulsh_i64; 3796 goto do_mul2; 3797 do_mul2: 3798 nb_iargs = 2; 3799 nb_oargs = 2; 3800 if (arg_temp(op->args[1])->state == TS_DEAD) { 3801 if (arg_temp(op->args[0])->state == TS_DEAD) { 3802 /* Both parts of the operation are dead. */ 3803 goto do_remove; 3804 } 3805 /* The high part of the operation is dead; generate the low. */ 3806 op->opc = opc = opc_new; 3807 op->args[1] = op->args[2]; 3808 op->args[2] = op->args[3]; 3809 } else if (arg_temp(op->args[0])->state == TS_DEAD && have_opc_new2) { 3810 /* The low part of the operation is dead; generate the high. */ 3811 op->opc = opc = opc_new2; 3812 op->args[0] = op->args[1]; 3813 op->args[1] = op->args[2]; 3814 op->args[2] = op->args[3]; 3815 } else { 3816 goto do_not_remove; 3817 } 3818 /* Mark the single-word operation live. */ 3819 nb_oargs = 1; 3820 goto do_not_remove; 3821 3822 default: 3823 /* XXX: optimize by hardcoding common cases (e.g. triadic ops) */ 3824 nb_iargs = def->nb_iargs; 3825 nb_oargs = def->nb_oargs; 3826 3827 /* Test if the operation can be removed because all 3828 its outputs are dead. We assume that nb_oargs == 0 3829 implies side effects */ 3830 if (!(def->flags & TCG_OPF_SIDE_EFFECTS) && nb_oargs != 0) { 3831 for (i = 0; i < nb_oargs; i++) { 3832 if (arg_temp(op->args[i])->state != TS_DEAD) { 3833 goto do_not_remove; 3834 } 3835 } 3836 goto do_remove; 3837 } 3838 goto do_not_remove; 3839 3840 do_remove: 3841 tcg_op_remove(s, op); 3842 break; 3843 3844 do_not_remove: 3845 for (i = 0; i < nb_oargs; i++) { 3846 ts = arg_temp(op->args[i]); 3847 3848 /* Remember the preference of the uses that followed. */ 3849 if (i < ARRAY_SIZE(op->output_pref)) { 3850 op->output_pref[i] = *la_temp_pref(ts); 3851 } 3852 3853 /* Output args are dead. */ 3854 if (ts->state & TS_DEAD) { 3855 arg_life |= DEAD_ARG << i; 3856 } 3857 if (ts->state & TS_MEM) { 3858 arg_life |= SYNC_ARG << i; 3859 } 3860 ts->state = TS_DEAD; 3861 la_reset_pref(ts); 3862 } 3863 3864 /* If end of basic block, update. */ 3865 if (def->flags & TCG_OPF_BB_EXIT) { 3866 la_func_end(s, nb_globals, nb_temps); 3867 } else if (def->flags & TCG_OPF_COND_BRANCH) { 3868 la_bb_sync(s, nb_globals, nb_temps); 3869 } else if (def->flags & TCG_OPF_BB_END) { 3870 la_bb_end(s, nb_globals, nb_temps); 3871 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3872 la_global_sync(s, nb_globals); 3873 if (def->flags & TCG_OPF_CALL_CLOBBER) { 3874 la_cross_call(s, nb_temps); 3875 } 3876 } 3877 3878 /* Record arguments that die in this opcode. */ 3879 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3880 ts = arg_temp(op->args[i]); 3881 if (ts->state & TS_DEAD) { 3882 arg_life |= DEAD_ARG << i; 3883 } 3884 } 3885 3886 /* Input arguments are live for preceding opcodes. */ 3887 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3888 ts = arg_temp(op->args[i]); 3889 if (ts->state & TS_DEAD) { 3890 /* For operands that were dead, initially allow 3891 all regs for the type. */ 3892 *la_temp_pref(ts) = tcg_target_available_regs[ts->type]; 3893 ts->state &= ~TS_DEAD; 3894 } 3895 } 3896 3897 /* Incorporate constraints for this operand. */ 3898 switch (opc) { 3899 case INDEX_op_mov_i32: 3900 case INDEX_op_mov_i64: 3901 /* Note that these are TCG_OPF_NOT_PRESENT and do not 3902 have proper constraints. That said, special case 3903 moves to propagate preferences backward. */ 3904 if (IS_DEAD_ARG(1)) { 3905 *la_temp_pref(arg_temp(op->args[0])) 3906 = *la_temp_pref(arg_temp(op->args[1])); 3907 } 3908 break; 3909 3910 default: 3911 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 3912 const TCGArgConstraint *ct = &def->args_ct[i]; 3913 TCGRegSet set, *pset; 3914 3915 ts = arg_temp(op->args[i]); 3916 pset = la_temp_pref(ts); 3917 set = *pset; 3918 3919 set &= ct->regs; 3920 if (ct->ialias) { 3921 set &= output_pref(op, ct->alias_index); 3922 } 3923 /* If the combination is not possible, restart. */ 3924 if (set == 0) { 3925 set = ct->regs; 3926 } 3927 *pset = set; 3928 } 3929 break; 3930 } 3931 break; 3932 } 3933 op->life = arg_life; 3934 } 3935 } 3936 3937 /* Liveness analysis: Convert indirect regs to direct temporaries. */ 3938 static bool __attribute__((noinline)) 3939 liveness_pass_2(TCGContext *s) 3940 { 3941 int nb_globals = s->nb_globals; 3942 int nb_temps, i; 3943 bool changes = false; 3944 TCGOp *op, *op_next; 3945 3946 /* Create a temporary for each indirect global. */ 3947 for (i = 0; i < nb_globals; ++i) { 3948 TCGTemp *its = &s->temps[i]; 3949 if (its->indirect_reg) { 3950 TCGTemp *dts = tcg_temp_alloc(s); 3951 dts->type = its->type; 3952 dts->base_type = its->base_type; 3953 dts->temp_subindex = its->temp_subindex; 3954 dts->kind = TEMP_EBB; 3955 its->state_ptr = dts; 3956 } else { 3957 its->state_ptr = NULL; 3958 } 3959 /* All globals begin dead. */ 3960 its->state = TS_DEAD; 3961 } 3962 for (nb_temps = s->nb_temps; i < nb_temps; ++i) { 3963 TCGTemp *its = &s->temps[i]; 3964 its->state_ptr = NULL; 3965 its->state = TS_DEAD; 3966 } 3967 3968 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 3969 TCGOpcode opc = op->opc; 3970 const TCGOpDef *def = &tcg_op_defs[opc]; 3971 TCGLifeData arg_life = op->life; 3972 int nb_iargs, nb_oargs, call_flags; 3973 TCGTemp *arg_ts, *dir_ts; 3974 3975 if (opc == INDEX_op_call) { 3976 nb_oargs = TCGOP_CALLO(op); 3977 nb_iargs = TCGOP_CALLI(op); 3978 call_flags = tcg_call_flags(op); 3979 } else { 3980 nb_iargs = def->nb_iargs; 3981 nb_oargs = def->nb_oargs; 3982 3983 /* Set flags similar to how calls require. */ 3984 if (def->flags & TCG_OPF_COND_BRANCH) { 3985 /* Like reading globals: sync_globals */ 3986 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 3987 } else if (def->flags & TCG_OPF_BB_END) { 3988 /* Like writing globals: save_globals */ 3989 call_flags = 0; 3990 } else if (def->flags & TCG_OPF_SIDE_EFFECTS) { 3991 /* Like reading globals: sync_globals */ 3992 call_flags = TCG_CALL_NO_WRITE_GLOBALS; 3993 } else { 3994 /* No effect on globals. */ 3995 call_flags = (TCG_CALL_NO_READ_GLOBALS | 3996 TCG_CALL_NO_WRITE_GLOBALS); 3997 } 3998 } 3999 4000 /* Make sure that input arguments are available. */ 4001 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 4002 arg_ts = arg_temp(op->args[i]); 4003 dir_ts = arg_ts->state_ptr; 4004 if (dir_ts && arg_ts->state == TS_DEAD) { 4005 TCGOpcode lopc = (arg_ts->type == TCG_TYPE_I32 4006 ? INDEX_op_ld_i32 4007 : INDEX_op_ld_i64); 4008 TCGOp *lop = tcg_op_insert_before(s, op, lopc, 3); 4009 4010 lop->args[0] = temp_arg(dir_ts); 4011 lop->args[1] = temp_arg(arg_ts->mem_base); 4012 lop->args[2] = arg_ts->mem_offset; 4013 4014 /* Loaded, but synced with memory. */ 4015 arg_ts->state = TS_MEM; 4016 } 4017 } 4018 4019 /* Perform input replacement, and mark inputs that became dead. 4020 No action is required except keeping temp_state up to date 4021 so that we reload when needed. */ 4022 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 4023 arg_ts = arg_temp(op->args[i]); 4024 dir_ts = arg_ts->state_ptr; 4025 if (dir_ts) { 4026 op->args[i] = temp_arg(dir_ts); 4027 changes = true; 4028 if (IS_DEAD_ARG(i)) { 4029 arg_ts->state = TS_DEAD; 4030 } 4031 } 4032 } 4033 4034 /* Liveness analysis should ensure that the following are 4035 all correct, for call sites and basic block end points. */ 4036 if (call_flags & TCG_CALL_NO_READ_GLOBALS) { 4037 /* Nothing to do */ 4038 } else if (call_flags & TCG_CALL_NO_WRITE_GLOBALS) { 4039 for (i = 0; i < nb_globals; ++i) { 4040 /* Liveness should see that globals are synced back, 4041 that is, either TS_DEAD or TS_MEM. */ 4042 arg_ts = &s->temps[i]; 4043 tcg_debug_assert(arg_ts->state_ptr == 0 4044 || arg_ts->state != 0); 4045 } 4046 } else { 4047 for (i = 0; i < nb_globals; ++i) { 4048 /* Liveness should see that globals are saved back, 4049 that is, TS_DEAD, waiting to be reloaded. */ 4050 arg_ts = &s->temps[i]; 4051 tcg_debug_assert(arg_ts->state_ptr == 0 4052 || arg_ts->state == TS_DEAD); 4053 } 4054 } 4055 4056 /* Outputs become available. */ 4057 if (opc == INDEX_op_mov_i32 || opc == INDEX_op_mov_i64) { 4058 arg_ts = arg_temp(op->args[0]); 4059 dir_ts = arg_ts->state_ptr; 4060 if (dir_ts) { 4061 op->args[0] = temp_arg(dir_ts); 4062 changes = true; 4063 4064 /* The output is now live and modified. */ 4065 arg_ts->state = 0; 4066 4067 if (NEED_SYNC_ARG(0)) { 4068 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 4069 ? INDEX_op_st_i32 4070 : INDEX_op_st_i64); 4071 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3); 4072 TCGTemp *out_ts = dir_ts; 4073 4074 if (IS_DEAD_ARG(0)) { 4075 out_ts = arg_temp(op->args[1]); 4076 arg_ts->state = TS_DEAD; 4077 tcg_op_remove(s, op); 4078 } else { 4079 arg_ts->state = TS_MEM; 4080 } 4081 4082 sop->args[0] = temp_arg(out_ts); 4083 sop->args[1] = temp_arg(arg_ts->mem_base); 4084 sop->args[2] = arg_ts->mem_offset; 4085 } else { 4086 tcg_debug_assert(!IS_DEAD_ARG(0)); 4087 } 4088 } 4089 } else { 4090 for (i = 0; i < nb_oargs; i++) { 4091 arg_ts = arg_temp(op->args[i]); 4092 dir_ts = arg_ts->state_ptr; 4093 if (!dir_ts) { 4094 continue; 4095 } 4096 op->args[i] = temp_arg(dir_ts); 4097 changes = true; 4098 4099 /* The output is now live and modified. */ 4100 arg_ts->state = 0; 4101 4102 /* Sync outputs upon their last write. */ 4103 if (NEED_SYNC_ARG(i)) { 4104 TCGOpcode sopc = (arg_ts->type == TCG_TYPE_I32 4105 ? INDEX_op_st_i32 4106 : INDEX_op_st_i64); 4107 TCGOp *sop = tcg_op_insert_after(s, op, sopc, 3); 4108 4109 sop->args[0] = temp_arg(dir_ts); 4110 sop->args[1] = temp_arg(arg_ts->mem_base); 4111 sop->args[2] = arg_ts->mem_offset; 4112 4113 arg_ts->state = TS_MEM; 4114 } 4115 /* Drop outputs that are dead. */ 4116 if (IS_DEAD_ARG(i)) { 4117 arg_ts->state = TS_DEAD; 4118 } 4119 } 4120 } 4121 } 4122 4123 return changes; 4124 } 4125 4126 static void temp_allocate_frame(TCGContext *s, TCGTemp *ts) 4127 { 4128 intptr_t off; 4129 int size, align; 4130 4131 /* When allocating an object, look at the full type. */ 4132 size = tcg_type_size(ts->base_type); 4133 switch (ts->base_type) { 4134 case TCG_TYPE_I32: 4135 align = 4; 4136 break; 4137 case TCG_TYPE_I64: 4138 case TCG_TYPE_V64: 4139 align = 8; 4140 break; 4141 case TCG_TYPE_I128: 4142 case TCG_TYPE_V128: 4143 case TCG_TYPE_V256: 4144 /* 4145 * Note that we do not require aligned storage for V256, 4146 * and that we provide alignment for I128 to match V128, 4147 * even if that's above what the host ABI requires. 4148 */ 4149 align = 16; 4150 break; 4151 default: 4152 g_assert_not_reached(); 4153 } 4154 4155 /* 4156 * Assume the stack is sufficiently aligned. 4157 * This affects e.g. ARM NEON, where we have 8 byte stack alignment 4158 * and do not require 16 byte vector alignment. This seems slightly 4159 * easier than fully parameterizing the above switch statement. 4160 */ 4161 align = MIN(TCG_TARGET_STACK_ALIGN, align); 4162 off = ROUND_UP(s->current_frame_offset, align); 4163 4164 /* If we've exhausted the stack frame, restart with a smaller TB. */ 4165 if (off + size > s->frame_end) { 4166 tcg_raise_tb_overflow(s); 4167 } 4168 s->current_frame_offset = off + size; 4169 #if defined(__sparc__) 4170 off += TCG_TARGET_STACK_BIAS; 4171 #endif 4172 4173 /* If the object was subdivided, assign memory to all the parts. */ 4174 if (ts->base_type != ts->type) { 4175 int part_size = tcg_type_size(ts->type); 4176 int part_count = size / part_size; 4177 4178 /* 4179 * Each part is allocated sequentially in tcg_temp_new_internal. 4180 * Jump back to the first part by subtracting the current index. 4181 */ 4182 ts -= ts->temp_subindex; 4183 for (int i = 0; i < part_count; ++i) { 4184 ts[i].mem_offset = off + i * part_size; 4185 ts[i].mem_base = s->frame_temp; 4186 ts[i].mem_allocated = 1; 4187 } 4188 } else { 4189 ts->mem_offset = off; 4190 ts->mem_base = s->frame_temp; 4191 ts->mem_allocated = 1; 4192 } 4193 } 4194 4195 /* Assign @reg to @ts, and update reg_to_temp[]. */ 4196 static void set_temp_val_reg(TCGContext *s, TCGTemp *ts, TCGReg reg) 4197 { 4198 if (ts->val_type == TEMP_VAL_REG) { 4199 TCGReg old = ts->reg; 4200 tcg_debug_assert(s->reg_to_temp[old] == ts); 4201 if (old == reg) { 4202 return; 4203 } 4204 s->reg_to_temp[old] = NULL; 4205 } 4206 tcg_debug_assert(s->reg_to_temp[reg] == NULL); 4207 s->reg_to_temp[reg] = ts; 4208 ts->val_type = TEMP_VAL_REG; 4209 ts->reg = reg; 4210 } 4211 4212 /* Assign a non-register value type to @ts, and update reg_to_temp[]. */ 4213 static void set_temp_val_nonreg(TCGContext *s, TCGTemp *ts, TCGTempVal type) 4214 { 4215 tcg_debug_assert(type != TEMP_VAL_REG); 4216 if (ts->val_type == TEMP_VAL_REG) { 4217 TCGReg reg = ts->reg; 4218 tcg_debug_assert(s->reg_to_temp[reg] == ts); 4219 s->reg_to_temp[reg] = NULL; 4220 } 4221 ts->val_type = type; 4222 } 4223 4224 static void temp_load(TCGContext *, TCGTemp *, TCGRegSet, TCGRegSet, TCGRegSet); 4225 4226 /* Mark a temporary as free or dead. If 'free_or_dead' is negative, 4227 mark it free; otherwise mark it dead. */ 4228 static void temp_free_or_dead(TCGContext *s, TCGTemp *ts, int free_or_dead) 4229 { 4230 TCGTempVal new_type; 4231 4232 switch (ts->kind) { 4233 case TEMP_FIXED: 4234 return; 4235 case TEMP_GLOBAL: 4236 case TEMP_TB: 4237 new_type = TEMP_VAL_MEM; 4238 break; 4239 case TEMP_EBB: 4240 new_type = free_or_dead < 0 ? TEMP_VAL_MEM : TEMP_VAL_DEAD; 4241 break; 4242 case TEMP_CONST: 4243 new_type = TEMP_VAL_CONST; 4244 break; 4245 default: 4246 g_assert_not_reached(); 4247 } 4248 set_temp_val_nonreg(s, ts, new_type); 4249 } 4250 4251 /* Mark a temporary as dead. */ 4252 static inline void temp_dead(TCGContext *s, TCGTemp *ts) 4253 { 4254 temp_free_or_dead(s, ts, 1); 4255 } 4256 4257 /* Sync a temporary to memory. 'allocated_regs' is used in case a temporary 4258 registers needs to be allocated to store a constant. If 'free_or_dead' 4259 is non-zero, subsequently release the temporary; if it is positive, the 4260 temp is dead; if it is negative, the temp is free. */ 4261 static void temp_sync(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs, 4262 TCGRegSet preferred_regs, int free_or_dead) 4263 { 4264 if (!temp_readonly(ts) && !ts->mem_coherent) { 4265 if (!ts->mem_allocated) { 4266 temp_allocate_frame(s, ts); 4267 } 4268 switch (ts->val_type) { 4269 case TEMP_VAL_CONST: 4270 /* If we're going to free the temp immediately, then we won't 4271 require it later in a register, so attempt to store the 4272 constant to memory directly. */ 4273 if (free_or_dead 4274 && tcg_out_sti(s, ts->type, ts->val, 4275 ts->mem_base->reg, ts->mem_offset)) { 4276 break; 4277 } 4278 temp_load(s, ts, tcg_target_available_regs[ts->type], 4279 allocated_regs, preferred_regs); 4280 /* fallthrough */ 4281 4282 case TEMP_VAL_REG: 4283 tcg_out_st(s, ts->type, ts->reg, 4284 ts->mem_base->reg, ts->mem_offset); 4285 break; 4286 4287 case TEMP_VAL_MEM: 4288 break; 4289 4290 case TEMP_VAL_DEAD: 4291 default: 4292 g_assert_not_reached(); 4293 } 4294 ts->mem_coherent = 1; 4295 } 4296 if (free_or_dead) { 4297 temp_free_or_dead(s, ts, free_or_dead); 4298 } 4299 } 4300 4301 /* free register 'reg' by spilling the corresponding temporary if necessary */ 4302 static void tcg_reg_free(TCGContext *s, TCGReg reg, TCGRegSet allocated_regs) 4303 { 4304 TCGTemp *ts = s->reg_to_temp[reg]; 4305 if (ts != NULL) { 4306 temp_sync(s, ts, allocated_regs, 0, -1); 4307 } 4308 } 4309 4310 /** 4311 * tcg_reg_alloc: 4312 * @required_regs: Set of registers in which we must allocate. 4313 * @allocated_regs: Set of registers which must be avoided. 4314 * @preferred_regs: Set of registers we should prefer. 4315 * @rev: True if we search the registers in "indirect" order. 4316 * 4317 * The allocated register must be in @required_regs & ~@allocated_regs, 4318 * but if we can put it in @preferred_regs we may save a move later. 4319 */ 4320 static TCGReg tcg_reg_alloc(TCGContext *s, TCGRegSet required_regs, 4321 TCGRegSet allocated_regs, 4322 TCGRegSet preferred_regs, bool rev) 4323 { 4324 int i, j, f, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4325 TCGRegSet reg_ct[2]; 4326 const int *order; 4327 4328 reg_ct[1] = required_regs & ~allocated_regs; 4329 tcg_debug_assert(reg_ct[1] != 0); 4330 reg_ct[0] = reg_ct[1] & preferred_regs; 4331 4332 /* Skip the preferred_regs option if it cannot be satisfied, 4333 or if the preference made no difference. */ 4334 f = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4335 4336 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4337 4338 /* Try free registers, preferences first. */ 4339 for (j = f; j < 2; j++) { 4340 TCGRegSet set = reg_ct[j]; 4341 4342 if (tcg_regset_single(set)) { 4343 /* One register in the set. */ 4344 TCGReg reg = tcg_regset_first(set); 4345 if (s->reg_to_temp[reg] == NULL) { 4346 return reg; 4347 } 4348 } else { 4349 for (i = 0; i < n; i++) { 4350 TCGReg reg = order[i]; 4351 if (s->reg_to_temp[reg] == NULL && 4352 tcg_regset_test_reg(set, reg)) { 4353 return reg; 4354 } 4355 } 4356 } 4357 } 4358 4359 /* We must spill something. */ 4360 for (j = f; j < 2; j++) { 4361 TCGRegSet set = reg_ct[j]; 4362 4363 if (tcg_regset_single(set)) { 4364 /* One register in the set. */ 4365 TCGReg reg = tcg_regset_first(set); 4366 tcg_reg_free(s, reg, allocated_regs); 4367 return reg; 4368 } else { 4369 for (i = 0; i < n; i++) { 4370 TCGReg reg = order[i]; 4371 if (tcg_regset_test_reg(set, reg)) { 4372 tcg_reg_free(s, reg, allocated_regs); 4373 return reg; 4374 } 4375 } 4376 } 4377 } 4378 4379 g_assert_not_reached(); 4380 } 4381 4382 static TCGReg tcg_reg_alloc_pair(TCGContext *s, TCGRegSet required_regs, 4383 TCGRegSet allocated_regs, 4384 TCGRegSet preferred_regs, bool rev) 4385 { 4386 int i, j, k, fmin, n = ARRAY_SIZE(tcg_target_reg_alloc_order); 4387 TCGRegSet reg_ct[2]; 4388 const int *order; 4389 4390 /* Ensure that if I is not in allocated_regs, I+1 is not either. */ 4391 reg_ct[1] = required_regs & ~(allocated_regs | (allocated_regs >> 1)); 4392 tcg_debug_assert(reg_ct[1] != 0); 4393 reg_ct[0] = reg_ct[1] & preferred_regs; 4394 4395 order = rev ? indirect_reg_alloc_order : tcg_target_reg_alloc_order; 4396 4397 /* 4398 * Skip the preferred_regs option if it cannot be satisfied, 4399 * or if the preference made no difference. 4400 */ 4401 k = reg_ct[0] == 0 || reg_ct[0] == reg_ct[1]; 4402 4403 /* 4404 * Minimize the number of flushes by looking for 2 free registers first, 4405 * then a single flush, then two flushes. 4406 */ 4407 for (fmin = 2; fmin >= 0; fmin--) { 4408 for (j = k; j < 2; j++) { 4409 TCGRegSet set = reg_ct[j]; 4410 4411 for (i = 0; i < n; i++) { 4412 TCGReg reg = order[i]; 4413 4414 if (tcg_regset_test_reg(set, reg)) { 4415 int f = !s->reg_to_temp[reg] + !s->reg_to_temp[reg + 1]; 4416 if (f >= fmin) { 4417 tcg_reg_free(s, reg, allocated_regs); 4418 tcg_reg_free(s, reg + 1, allocated_regs); 4419 return reg; 4420 } 4421 } 4422 } 4423 } 4424 } 4425 g_assert_not_reached(); 4426 } 4427 4428 /* Make sure the temporary is in a register. If needed, allocate the register 4429 from DESIRED while avoiding ALLOCATED. */ 4430 static void temp_load(TCGContext *s, TCGTemp *ts, TCGRegSet desired_regs, 4431 TCGRegSet allocated_regs, TCGRegSet preferred_regs) 4432 { 4433 TCGReg reg; 4434 4435 switch (ts->val_type) { 4436 case TEMP_VAL_REG: 4437 return; 4438 case TEMP_VAL_CONST: 4439 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4440 preferred_regs, ts->indirect_base); 4441 if (ts->type <= TCG_TYPE_I64) { 4442 tcg_out_movi(s, ts->type, reg, ts->val); 4443 } else { 4444 uint64_t val = ts->val; 4445 MemOp vece = MO_64; 4446 4447 /* 4448 * Find the minimal vector element that matches the constant. 4449 * The targets will, in general, have to do this search anyway, 4450 * do this generically. 4451 */ 4452 if (val == dup_const(MO_8, val)) { 4453 vece = MO_8; 4454 } else if (val == dup_const(MO_16, val)) { 4455 vece = MO_16; 4456 } else if (val == dup_const(MO_32, val)) { 4457 vece = MO_32; 4458 } 4459 4460 tcg_out_dupi_vec(s, ts->type, vece, reg, ts->val); 4461 } 4462 ts->mem_coherent = 0; 4463 break; 4464 case TEMP_VAL_MEM: 4465 reg = tcg_reg_alloc(s, desired_regs, allocated_regs, 4466 preferred_regs, ts->indirect_base); 4467 tcg_out_ld(s, ts->type, reg, ts->mem_base->reg, ts->mem_offset); 4468 ts->mem_coherent = 1; 4469 break; 4470 case TEMP_VAL_DEAD: 4471 default: 4472 g_assert_not_reached(); 4473 } 4474 set_temp_val_reg(s, ts, reg); 4475 } 4476 4477 /* Save a temporary to memory. 'allocated_regs' is used in case a 4478 temporary registers needs to be allocated to store a constant. */ 4479 static void temp_save(TCGContext *s, TCGTemp *ts, TCGRegSet allocated_regs) 4480 { 4481 /* The liveness analysis already ensures that globals are back 4482 in memory. Keep an tcg_debug_assert for safety. */ 4483 tcg_debug_assert(ts->val_type == TEMP_VAL_MEM || temp_readonly(ts)); 4484 } 4485 4486 /* save globals to their canonical location and assume they can be 4487 modified be the following code. 'allocated_regs' is used in case a 4488 temporary registers needs to be allocated to store a constant. */ 4489 static void save_globals(TCGContext *s, TCGRegSet allocated_regs) 4490 { 4491 int i, n; 4492 4493 for (i = 0, n = s->nb_globals; i < n; i++) { 4494 temp_save(s, &s->temps[i], allocated_regs); 4495 } 4496 } 4497 4498 /* sync globals to their canonical location and assume they can be 4499 read by the following code. 'allocated_regs' is used in case a 4500 temporary registers needs to be allocated to store a constant. */ 4501 static void sync_globals(TCGContext *s, TCGRegSet allocated_regs) 4502 { 4503 int i, n; 4504 4505 for (i = 0, n = s->nb_globals; i < n; i++) { 4506 TCGTemp *ts = &s->temps[i]; 4507 tcg_debug_assert(ts->val_type != TEMP_VAL_REG 4508 || ts->kind == TEMP_FIXED 4509 || ts->mem_coherent); 4510 } 4511 } 4512 4513 /* at the end of a basic block, we assume all temporaries are dead and 4514 all globals are stored at their canonical location. */ 4515 static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) 4516 { 4517 int i; 4518 4519 for (i = s->nb_globals; i < s->nb_temps; i++) { 4520 TCGTemp *ts = &s->temps[i]; 4521 4522 switch (ts->kind) { 4523 case TEMP_TB: 4524 temp_save(s, ts, allocated_regs); 4525 break; 4526 case TEMP_EBB: 4527 /* The liveness analysis already ensures that temps are dead. 4528 Keep an tcg_debug_assert for safety. */ 4529 tcg_debug_assert(ts->val_type == TEMP_VAL_DEAD); 4530 break; 4531 case TEMP_CONST: 4532 /* Similarly, we should have freed any allocated register. */ 4533 tcg_debug_assert(ts->val_type == TEMP_VAL_CONST); 4534 break; 4535 default: 4536 g_assert_not_reached(); 4537 } 4538 } 4539 4540 save_globals(s, allocated_regs); 4541 } 4542 4543 /* 4544 * At a conditional branch, we assume all temporaries are dead unless 4545 * explicitly live-across-conditional-branch; all globals and local 4546 * temps are synced to their location. 4547 */ 4548 static void tcg_reg_alloc_cbranch(TCGContext *s, TCGRegSet allocated_regs) 4549 { 4550 sync_globals(s, allocated_regs); 4551 4552 for (int i = s->nb_globals; i < s->nb_temps; i++) { 4553 TCGTemp *ts = &s->temps[i]; 4554 /* 4555 * The liveness analysis already ensures that temps are dead. 4556 * Keep tcg_debug_asserts for safety. 4557 */ 4558 switch (ts->kind) { 4559 case TEMP_TB: 4560 tcg_debug_assert(ts->val_type != TEMP_VAL_REG || ts->mem_coherent); 4561 break; 4562 case TEMP_EBB: 4563 case TEMP_CONST: 4564 break; 4565 default: 4566 g_assert_not_reached(); 4567 } 4568 } 4569 } 4570 4571 /* 4572 * Specialized code generation for INDEX_op_mov_* with a constant. 4573 */ 4574 static void tcg_reg_alloc_do_movi(TCGContext *s, TCGTemp *ots, 4575 tcg_target_ulong val, TCGLifeData arg_life, 4576 TCGRegSet preferred_regs) 4577 { 4578 /* ENV should not be modified. */ 4579 tcg_debug_assert(!temp_readonly(ots)); 4580 4581 /* The movi is not explicitly generated here. */ 4582 set_temp_val_nonreg(s, ots, TEMP_VAL_CONST); 4583 ots->val = val; 4584 ots->mem_coherent = 0; 4585 if (NEED_SYNC_ARG(0)) { 4586 temp_sync(s, ots, s->reserved_regs, preferred_regs, IS_DEAD_ARG(0)); 4587 } else if (IS_DEAD_ARG(0)) { 4588 temp_dead(s, ots); 4589 } 4590 } 4591 4592 /* 4593 * Specialized code generation for INDEX_op_mov_*. 4594 */ 4595 static void tcg_reg_alloc_mov(TCGContext *s, const TCGOp *op) 4596 { 4597 const TCGLifeData arg_life = op->life; 4598 TCGRegSet allocated_regs, preferred_regs; 4599 TCGTemp *ts, *ots; 4600 TCGType otype, itype; 4601 TCGReg oreg, ireg; 4602 4603 allocated_regs = s->reserved_regs; 4604 preferred_regs = output_pref(op, 0); 4605 ots = arg_temp(op->args[0]); 4606 ts = arg_temp(op->args[1]); 4607 4608 /* ENV should not be modified. */ 4609 tcg_debug_assert(!temp_readonly(ots)); 4610 4611 /* Note that otype != itype for no-op truncation. */ 4612 otype = ots->type; 4613 itype = ts->type; 4614 4615 if (ts->val_type == TEMP_VAL_CONST) { 4616 /* propagate constant or generate sti */ 4617 tcg_target_ulong val = ts->val; 4618 if (IS_DEAD_ARG(1)) { 4619 temp_dead(s, ts); 4620 } 4621 tcg_reg_alloc_do_movi(s, ots, val, arg_life, preferred_regs); 4622 return; 4623 } 4624 4625 /* If the source value is in memory we're going to be forced 4626 to have it in a register in order to perform the copy. Copy 4627 the SOURCE value into its own register first, that way we 4628 don't have to reload SOURCE the next time it is used. */ 4629 if (ts->val_type == TEMP_VAL_MEM) { 4630 temp_load(s, ts, tcg_target_available_regs[itype], 4631 allocated_regs, preferred_regs); 4632 } 4633 tcg_debug_assert(ts->val_type == TEMP_VAL_REG); 4634 ireg = ts->reg; 4635 4636 if (IS_DEAD_ARG(0)) { 4637 /* mov to a non-saved dead register makes no sense (even with 4638 liveness analysis disabled). */ 4639 tcg_debug_assert(NEED_SYNC_ARG(0)); 4640 if (!ots->mem_allocated) { 4641 temp_allocate_frame(s, ots); 4642 } 4643 tcg_out_st(s, otype, ireg, ots->mem_base->reg, ots->mem_offset); 4644 if (IS_DEAD_ARG(1)) { 4645 temp_dead(s, ts); 4646 } 4647 temp_dead(s, ots); 4648 return; 4649 } 4650 4651 if (IS_DEAD_ARG(1) && ts->kind != TEMP_FIXED) { 4652 /* 4653 * The mov can be suppressed. Kill input first, so that it 4654 * is unlinked from reg_to_temp, then set the output to the 4655 * reg that we saved from the input. 4656 */ 4657 temp_dead(s, ts); 4658 oreg = ireg; 4659 } else { 4660 if (ots->val_type == TEMP_VAL_REG) { 4661 oreg = ots->reg; 4662 } else { 4663 /* Make sure to not spill the input register during allocation. */ 4664 oreg = tcg_reg_alloc(s, tcg_target_available_regs[otype], 4665 allocated_regs | ((TCGRegSet)1 << ireg), 4666 preferred_regs, ots->indirect_base); 4667 } 4668 if (!tcg_out_mov(s, otype, oreg, ireg)) { 4669 /* 4670 * Cross register class move not supported. 4671 * Store the source register into the destination slot 4672 * and leave the destination temp as TEMP_VAL_MEM. 4673 */ 4674 assert(!temp_readonly(ots)); 4675 if (!ts->mem_allocated) { 4676 temp_allocate_frame(s, ots); 4677 } 4678 tcg_out_st(s, ts->type, ireg, ots->mem_base->reg, ots->mem_offset); 4679 set_temp_val_nonreg(s, ts, TEMP_VAL_MEM); 4680 ots->mem_coherent = 1; 4681 return; 4682 } 4683 } 4684 set_temp_val_reg(s, ots, oreg); 4685 ots->mem_coherent = 0; 4686 4687 if (NEED_SYNC_ARG(0)) { 4688 temp_sync(s, ots, allocated_regs, 0, 0); 4689 } 4690 } 4691 4692 /* 4693 * Specialized code generation for INDEX_op_dup_vec. 4694 */ 4695 static void tcg_reg_alloc_dup(TCGContext *s, const TCGOp *op) 4696 { 4697 const TCGLifeData arg_life = op->life; 4698 TCGRegSet dup_out_regs, dup_in_regs; 4699 TCGTemp *its, *ots; 4700 TCGType itype, vtype; 4701 unsigned vece; 4702 int lowpart_ofs; 4703 bool ok; 4704 4705 ots = arg_temp(op->args[0]); 4706 its = arg_temp(op->args[1]); 4707 4708 /* ENV should not be modified. */ 4709 tcg_debug_assert(!temp_readonly(ots)); 4710 4711 itype = its->type; 4712 vece = TCGOP_VECE(op); 4713 vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 4714 4715 if (its->val_type == TEMP_VAL_CONST) { 4716 /* Propagate constant via movi -> dupi. */ 4717 tcg_target_ulong val = its->val; 4718 if (IS_DEAD_ARG(1)) { 4719 temp_dead(s, its); 4720 } 4721 tcg_reg_alloc_do_movi(s, ots, val, arg_life, output_pref(op, 0)); 4722 return; 4723 } 4724 4725 dup_out_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 4726 dup_in_regs = tcg_op_defs[INDEX_op_dup_vec].args_ct[1].regs; 4727 4728 /* Allocate the output register now. */ 4729 if (ots->val_type != TEMP_VAL_REG) { 4730 TCGRegSet allocated_regs = s->reserved_regs; 4731 TCGReg oreg; 4732 4733 if (!IS_DEAD_ARG(1) && its->val_type == TEMP_VAL_REG) { 4734 /* Make sure to not spill the input register. */ 4735 tcg_regset_set_reg(allocated_regs, its->reg); 4736 } 4737 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 4738 output_pref(op, 0), ots->indirect_base); 4739 set_temp_val_reg(s, ots, oreg); 4740 } 4741 4742 switch (its->val_type) { 4743 case TEMP_VAL_REG: 4744 /* 4745 * The dup constriaints must be broad, covering all possible VECE. 4746 * However, tcg_op_dup_vec() gets to see the VECE and we allow it 4747 * to fail, indicating that extra moves are required for that case. 4748 */ 4749 if (tcg_regset_test_reg(dup_in_regs, its->reg)) { 4750 if (tcg_out_dup_vec(s, vtype, vece, ots->reg, its->reg)) { 4751 goto done; 4752 } 4753 /* Try again from memory or a vector input register. */ 4754 } 4755 if (!its->mem_coherent) { 4756 /* 4757 * The input register is not synced, and so an extra store 4758 * would be required to use memory. Attempt an integer-vector 4759 * register move first. We do not have a TCGRegSet for this. 4760 */ 4761 if (tcg_out_mov(s, itype, ots->reg, its->reg)) { 4762 break; 4763 } 4764 /* Sync the temp back to its slot and load from there. */ 4765 temp_sync(s, its, s->reserved_regs, 0, 0); 4766 } 4767 /* fall through */ 4768 4769 case TEMP_VAL_MEM: 4770 lowpart_ofs = 0; 4771 if (HOST_BIG_ENDIAN) { 4772 lowpart_ofs = tcg_type_size(itype) - (1 << vece); 4773 } 4774 if (tcg_out_dupm_vec(s, vtype, vece, ots->reg, its->mem_base->reg, 4775 its->mem_offset + lowpart_ofs)) { 4776 goto done; 4777 } 4778 /* Load the input into the destination vector register. */ 4779 tcg_out_ld(s, itype, ots->reg, its->mem_base->reg, its->mem_offset); 4780 break; 4781 4782 default: 4783 g_assert_not_reached(); 4784 } 4785 4786 /* We now have a vector input register, so dup must succeed. */ 4787 ok = tcg_out_dup_vec(s, vtype, vece, ots->reg, ots->reg); 4788 tcg_debug_assert(ok); 4789 4790 done: 4791 ots->mem_coherent = 0; 4792 if (IS_DEAD_ARG(1)) { 4793 temp_dead(s, its); 4794 } 4795 if (NEED_SYNC_ARG(0)) { 4796 temp_sync(s, ots, s->reserved_regs, 0, 0); 4797 } 4798 if (IS_DEAD_ARG(0)) { 4799 temp_dead(s, ots); 4800 } 4801 } 4802 4803 static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op) 4804 { 4805 const TCGLifeData arg_life = op->life; 4806 const TCGOpDef * const def = &tcg_op_defs[op->opc]; 4807 TCGRegSet i_allocated_regs; 4808 TCGRegSet o_allocated_regs; 4809 int i, k, nb_iargs, nb_oargs; 4810 TCGReg reg; 4811 TCGArg arg; 4812 const TCGArgConstraint *arg_ct; 4813 TCGTemp *ts; 4814 TCGArg new_args[TCG_MAX_OP_ARGS]; 4815 int const_args[TCG_MAX_OP_ARGS]; 4816 TCGCond op_cond; 4817 4818 nb_oargs = def->nb_oargs; 4819 nb_iargs = def->nb_iargs; 4820 4821 /* copy constants */ 4822 memcpy(new_args + nb_oargs + nb_iargs, 4823 op->args + nb_oargs + nb_iargs, 4824 sizeof(TCGArg) * def->nb_cargs); 4825 4826 i_allocated_regs = s->reserved_regs; 4827 o_allocated_regs = s->reserved_regs; 4828 4829 switch (op->opc) { 4830 case INDEX_op_brcond_i32: 4831 case INDEX_op_brcond_i64: 4832 op_cond = op->args[2]; 4833 break; 4834 case INDEX_op_setcond_i32: 4835 case INDEX_op_setcond_i64: 4836 case INDEX_op_negsetcond_i32: 4837 case INDEX_op_negsetcond_i64: 4838 case INDEX_op_cmp_vec: 4839 op_cond = op->args[3]; 4840 break; 4841 case INDEX_op_brcond2_i32: 4842 op_cond = op->args[4]; 4843 break; 4844 case INDEX_op_movcond_i32: 4845 case INDEX_op_movcond_i64: 4846 case INDEX_op_setcond2_i32: 4847 case INDEX_op_cmpsel_vec: 4848 op_cond = op->args[5]; 4849 break; 4850 default: 4851 /* No condition within opcode. */ 4852 op_cond = TCG_COND_ALWAYS; 4853 break; 4854 } 4855 4856 /* satisfy input constraints */ 4857 for (k = 0; k < nb_iargs; k++) { 4858 TCGRegSet i_preferred_regs, i_required_regs; 4859 bool allocate_new_reg, copyto_new_reg; 4860 TCGTemp *ts2; 4861 int i1, i2; 4862 4863 i = def->args_ct[nb_oargs + k].sort_index; 4864 arg = op->args[i]; 4865 arg_ct = &def->args_ct[i]; 4866 ts = arg_temp(arg); 4867 4868 if (ts->val_type == TEMP_VAL_CONST 4869 && tcg_target_const_match(ts->val, arg_ct->ct, ts->type, 4870 op_cond, TCGOP_VECE(op))) { 4871 /* constant is OK for instruction */ 4872 const_args[i] = 1; 4873 new_args[i] = ts->val; 4874 continue; 4875 } 4876 4877 reg = ts->reg; 4878 i_preferred_regs = 0; 4879 i_required_regs = arg_ct->regs; 4880 allocate_new_reg = false; 4881 copyto_new_reg = false; 4882 4883 switch (arg_ct->pair) { 4884 case 0: /* not paired */ 4885 if (arg_ct->ialias) { 4886 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4887 4888 /* 4889 * If the input is readonly, then it cannot also be an 4890 * output and aliased to itself. If the input is not 4891 * dead after the instruction, we must allocate a new 4892 * register and move it. 4893 */ 4894 if (temp_readonly(ts) || !IS_DEAD_ARG(i) 4895 || def->args_ct[arg_ct->alias_index].newreg) { 4896 allocate_new_reg = true; 4897 } else if (ts->val_type == TEMP_VAL_REG) { 4898 /* 4899 * Check if the current register has already been 4900 * allocated for another input. 4901 */ 4902 allocate_new_reg = 4903 tcg_regset_test_reg(i_allocated_regs, reg); 4904 } 4905 } 4906 if (!allocate_new_reg) { 4907 temp_load(s, ts, i_required_regs, i_allocated_regs, 4908 i_preferred_regs); 4909 reg = ts->reg; 4910 allocate_new_reg = !tcg_regset_test_reg(i_required_regs, reg); 4911 } 4912 if (allocate_new_reg) { 4913 /* 4914 * Allocate a new register matching the constraint 4915 * and move the temporary register into it. 4916 */ 4917 temp_load(s, ts, tcg_target_available_regs[ts->type], 4918 i_allocated_regs, 0); 4919 reg = tcg_reg_alloc(s, i_required_regs, i_allocated_regs, 4920 i_preferred_regs, ts->indirect_base); 4921 copyto_new_reg = true; 4922 } 4923 break; 4924 4925 case 1: 4926 /* First of an input pair; if i1 == i2, the second is an output. */ 4927 i1 = i; 4928 i2 = arg_ct->pair_index; 4929 ts2 = i1 != i2 ? arg_temp(op->args[i2]) : NULL; 4930 4931 /* 4932 * It is easier to default to allocating a new pair 4933 * and to identify a few cases where it's not required. 4934 */ 4935 if (arg_ct->ialias) { 4936 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4937 if (IS_DEAD_ARG(i1) && 4938 IS_DEAD_ARG(i2) && 4939 !temp_readonly(ts) && 4940 ts->val_type == TEMP_VAL_REG && 4941 ts->reg < TCG_TARGET_NB_REGS - 1 && 4942 tcg_regset_test_reg(i_required_regs, reg) && 4943 !tcg_regset_test_reg(i_allocated_regs, reg) && 4944 !tcg_regset_test_reg(i_allocated_regs, reg + 1) && 4945 (ts2 4946 ? ts2->val_type == TEMP_VAL_REG && 4947 ts2->reg == reg + 1 && 4948 !temp_readonly(ts2) 4949 : s->reg_to_temp[reg + 1] == NULL)) { 4950 break; 4951 } 4952 } else { 4953 /* Without aliasing, the pair must also be an input. */ 4954 tcg_debug_assert(ts2); 4955 if (ts->val_type == TEMP_VAL_REG && 4956 ts2->val_type == TEMP_VAL_REG && 4957 ts2->reg == reg + 1 && 4958 tcg_regset_test_reg(i_required_regs, reg)) { 4959 break; 4960 } 4961 } 4962 reg = tcg_reg_alloc_pair(s, i_required_regs, i_allocated_regs, 4963 0, ts->indirect_base); 4964 goto do_pair; 4965 4966 case 2: /* pair second */ 4967 reg = new_args[arg_ct->pair_index] + 1; 4968 goto do_pair; 4969 4970 case 3: /* ialias with second output, no first input */ 4971 tcg_debug_assert(arg_ct->ialias); 4972 i_preferred_regs = output_pref(op, arg_ct->alias_index); 4973 4974 if (IS_DEAD_ARG(i) && 4975 !temp_readonly(ts) && 4976 ts->val_type == TEMP_VAL_REG && 4977 reg > 0 && 4978 s->reg_to_temp[reg - 1] == NULL && 4979 tcg_regset_test_reg(i_required_regs, reg) && 4980 !tcg_regset_test_reg(i_allocated_regs, reg) && 4981 !tcg_regset_test_reg(i_allocated_regs, reg - 1)) { 4982 tcg_regset_set_reg(i_allocated_regs, reg - 1); 4983 break; 4984 } 4985 reg = tcg_reg_alloc_pair(s, i_required_regs >> 1, 4986 i_allocated_regs, 0, 4987 ts->indirect_base); 4988 tcg_regset_set_reg(i_allocated_regs, reg); 4989 reg += 1; 4990 goto do_pair; 4991 4992 do_pair: 4993 /* 4994 * If an aliased input is not dead after the instruction, 4995 * we must allocate a new register and move it. 4996 */ 4997 if (arg_ct->ialias && (!IS_DEAD_ARG(i) || temp_readonly(ts))) { 4998 TCGRegSet t_allocated_regs = i_allocated_regs; 4999 5000 /* 5001 * Because of the alias, and the continued life, make sure 5002 * that the temp is somewhere *other* than the reg pair, 5003 * and we get a copy in reg. 5004 */ 5005 tcg_regset_set_reg(t_allocated_regs, reg); 5006 tcg_regset_set_reg(t_allocated_regs, reg + 1); 5007 if (ts->val_type == TEMP_VAL_REG && ts->reg == reg) { 5008 /* If ts was already in reg, copy it somewhere else. */ 5009 TCGReg nr; 5010 bool ok; 5011 5012 tcg_debug_assert(ts->kind != TEMP_FIXED); 5013 nr = tcg_reg_alloc(s, tcg_target_available_regs[ts->type], 5014 t_allocated_regs, 0, ts->indirect_base); 5015 ok = tcg_out_mov(s, ts->type, nr, reg); 5016 tcg_debug_assert(ok); 5017 5018 set_temp_val_reg(s, ts, nr); 5019 } else { 5020 temp_load(s, ts, tcg_target_available_regs[ts->type], 5021 t_allocated_regs, 0); 5022 copyto_new_reg = true; 5023 } 5024 } else { 5025 /* Preferably allocate to reg, otherwise copy. */ 5026 i_required_regs = (TCGRegSet)1 << reg; 5027 temp_load(s, ts, i_required_regs, i_allocated_regs, 5028 i_preferred_regs); 5029 copyto_new_reg = ts->reg != reg; 5030 } 5031 break; 5032 5033 default: 5034 g_assert_not_reached(); 5035 } 5036 5037 if (copyto_new_reg) { 5038 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 5039 /* 5040 * Cross register class move not supported. Sync the 5041 * temp back to its slot and load from there. 5042 */ 5043 temp_sync(s, ts, i_allocated_regs, 0, 0); 5044 tcg_out_ld(s, ts->type, reg, 5045 ts->mem_base->reg, ts->mem_offset); 5046 } 5047 } 5048 new_args[i] = reg; 5049 const_args[i] = 0; 5050 tcg_regset_set_reg(i_allocated_regs, reg); 5051 } 5052 5053 /* mark dead temporaries and free the associated registers */ 5054 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 5055 if (IS_DEAD_ARG(i)) { 5056 temp_dead(s, arg_temp(op->args[i])); 5057 } 5058 } 5059 5060 if (def->flags & TCG_OPF_COND_BRANCH) { 5061 tcg_reg_alloc_cbranch(s, i_allocated_regs); 5062 } else if (def->flags & TCG_OPF_BB_END) { 5063 tcg_reg_alloc_bb_end(s, i_allocated_regs); 5064 } else { 5065 if (def->flags & TCG_OPF_CALL_CLOBBER) { 5066 /* XXX: permit generic clobber register list ? */ 5067 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 5068 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 5069 tcg_reg_free(s, i, i_allocated_regs); 5070 } 5071 } 5072 } 5073 if (def->flags & TCG_OPF_SIDE_EFFECTS) { 5074 /* sync globals if the op has side effects and might trigger 5075 an exception. */ 5076 sync_globals(s, i_allocated_regs); 5077 } 5078 5079 /* satisfy the output constraints */ 5080 for(k = 0; k < nb_oargs; k++) { 5081 i = def->args_ct[k].sort_index; 5082 arg = op->args[i]; 5083 arg_ct = &def->args_ct[i]; 5084 ts = arg_temp(arg); 5085 5086 /* ENV should not be modified. */ 5087 tcg_debug_assert(!temp_readonly(ts)); 5088 5089 switch (arg_ct->pair) { 5090 case 0: /* not paired */ 5091 if (arg_ct->oalias && !const_args[arg_ct->alias_index]) { 5092 reg = new_args[arg_ct->alias_index]; 5093 } else if (arg_ct->newreg) { 5094 reg = tcg_reg_alloc(s, arg_ct->regs, 5095 i_allocated_regs | o_allocated_regs, 5096 output_pref(op, k), ts->indirect_base); 5097 } else { 5098 reg = tcg_reg_alloc(s, arg_ct->regs, o_allocated_regs, 5099 output_pref(op, k), ts->indirect_base); 5100 } 5101 break; 5102 5103 case 1: /* first of pair */ 5104 if (arg_ct->oalias) { 5105 reg = new_args[arg_ct->alias_index]; 5106 } else if (arg_ct->newreg) { 5107 reg = tcg_reg_alloc_pair(s, arg_ct->regs, 5108 i_allocated_regs | o_allocated_regs, 5109 output_pref(op, k), 5110 ts->indirect_base); 5111 } else { 5112 reg = tcg_reg_alloc_pair(s, arg_ct->regs, o_allocated_regs, 5113 output_pref(op, k), 5114 ts->indirect_base); 5115 } 5116 break; 5117 5118 case 2: /* second of pair */ 5119 if (arg_ct->oalias) { 5120 reg = new_args[arg_ct->alias_index]; 5121 } else { 5122 reg = new_args[arg_ct->pair_index] + 1; 5123 } 5124 break; 5125 5126 case 3: /* first of pair, aliasing with a second input */ 5127 tcg_debug_assert(!arg_ct->newreg); 5128 reg = new_args[arg_ct->pair_index] - 1; 5129 break; 5130 5131 default: 5132 g_assert_not_reached(); 5133 } 5134 tcg_regset_set_reg(o_allocated_regs, reg); 5135 set_temp_val_reg(s, ts, reg); 5136 ts->mem_coherent = 0; 5137 new_args[i] = reg; 5138 } 5139 } 5140 5141 /* emit instruction */ 5142 switch (op->opc) { 5143 case INDEX_op_ext8s_i32: 5144 tcg_out_ext8s(s, TCG_TYPE_I32, new_args[0], new_args[1]); 5145 break; 5146 case INDEX_op_ext8s_i64: 5147 tcg_out_ext8s(s, TCG_TYPE_I64, new_args[0], new_args[1]); 5148 break; 5149 case INDEX_op_ext8u_i32: 5150 case INDEX_op_ext8u_i64: 5151 tcg_out_ext8u(s, new_args[0], new_args[1]); 5152 break; 5153 case INDEX_op_ext16s_i32: 5154 tcg_out_ext16s(s, TCG_TYPE_I32, new_args[0], new_args[1]); 5155 break; 5156 case INDEX_op_ext16s_i64: 5157 tcg_out_ext16s(s, TCG_TYPE_I64, new_args[0], new_args[1]); 5158 break; 5159 case INDEX_op_ext16u_i32: 5160 case INDEX_op_ext16u_i64: 5161 tcg_out_ext16u(s, new_args[0], new_args[1]); 5162 break; 5163 case INDEX_op_ext32s_i64: 5164 tcg_out_ext32s(s, new_args[0], new_args[1]); 5165 break; 5166 case INDEX_op_ext32u_i64: 5167 tcg_out_ext32u(s, new_args[0], new_args[1]); 5168 break; 5169 case INDEX_op_ext_i32_i64: 5170 tcg_out_exts_i32_i64(s, new_args[0], new_args[1]); 5171 break; 5172 case INDEX_op_extu_i32_i64: 5173 tcg_out_extu_i32_i64(s, new_args[0], new_args[1]); 5174 break; 5175 case INDEX_op_extrl_i64_i32: 5176 tcg_out_extrl_i64_i32(s, new_args[0], new_args[1]); 5177 break; 5178 default: 5179 if (def->flags & TCG_OPF_VECTOR) { 5180 tcg_out_vec_op(s, op->opc, TCGOP_VECL(op), TCGOP_VECE(op), 5181 new_args, const_args); 5182 } else { 5183 tcg_out_op(s, op->opc, new_args, const_args); 5184 } 5185 break; 5186 } 5187 5188 /* move the outputs in the correct register if needed */ 5189 for(i = 0; i < nb_oargs; i++) { 5190 ts = arg_temp(op->args[i]); 5191 5192 /* ENV should not be modified. */ 5193 tcg_debug_assert(!temp_readonly(ts)); 5194 5195 if (NEED_SYNC_ARG(i)) { 5196 temp_sync(s, ts, o_allocated_regs, 0, IS_DEAD_ARG(i)); 5197 } else if (IS_DEAD_ARG(i)) { 5198 temp_dead(s, ts); 5199 } 5200 } 5201 } 5202 5203 static bool tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op) 5204 { 5205 const TCGLifeData arg_life = op->life; 5206 TCGTemp *ots, *itsl, *itsh; 5207 TCGType vtype = TCGOP_VECL(op) + TCG_TYPE_V64; 5208 5209 /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */ 5210 tcg_debug_assert(TCG_TARGET_REG_BITS == 32); 5211 tcg_debug_assert(TCGOP_VECE(op) == MO_64); 5212 5213 ots = arg_temp(op->args[0]); 5214 itsl = arg_temp(op->args[1]); 5215 itsh = arg_temp(op->args[2]); 5216 5217 /* ENV should not be modified. */ 5218 tcg_debug_assert(!temp_readonly(ots)); 5219 5220 /* Allocate the output register now. */ 5221 if (ots->val_type != TEMP_VAL_REG) { 5222 TCGRegSet allocated_regs = s->reserved_regs; 5223 TCGRegSet dup_out_regs = 5224 tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs; 5225 TCGReg oreg; 5226 5227 /* Make sure to not spill the input registers. */ 5228 if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) { 5229 tcg_regset_set_reg(allocated_regs, itsl->reg); 5230 } 5231 if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) { 5232 tcg_regset_set_reg(allocated_regs, itsh->reg); 5233 } 5234 5235 oreg = tcg_reg_alloc(s, dup_out_regs, allocated_regs, 5236 output_pref(op, 0), ots->indirect_base); 5237 set_temp_val_reg(s, ots, oreg); 5238 } 5239 5240 /* Promote dup2 of immediates to dupi_vec. */ 5241 if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) { 5242 uint64_t val = deposit64(itsl->val, 32, 32, itsh->val); 5243 MemOp vece = MO_64; 5244 5245 if (val == dup_const(MO_8, val)) { 5246 vece = MO_8; 5247 } else if (val == dup_const(MO_16, val)) { 5248 vece = MO_16; 5249 } else if (val == dup_const(MO_32, val)) { 5250 vece = MO_32; 5251 } 5252 5253 tcg_out_dupi_vec(s, vtype, vece, ots->reg, val); 5254 goto done; 5255 } 5256 5257 /* If the two inputs form one 64-bit value, try dupm_vec. */ 5258 if (itsl->temp_subindex == HOST_BIG_ENDIAN && 5259 itsh->temp_subindex == !HOST_BIG_ENDIAN && 5260 itsl == itsh + (HOST_BIG_ENDIAN ? 1 : -1)) { 5261 TCGTemp *its = itsl - HOST_BIG_ENDIAN; 5262 5263 temp_sync(s, its + 0, s->reserved_regs, 0, 0); 5264 temp_sync(s, its + 1, s->reserved_regs, 0, 0); 5265 5266 if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg, 5267 its->mem_base->reg, its->mem_offset)) { 5268 goto done; 5269 } 5270 } 5271 5272 /* Fall back to generic expansion. */ 5273 return false; 5274 5275 done: 5276 ots->mem_coherent = 0; 5277 if (IS_DEAD_ARG(1)) { 5278 temp_dead(s, itsl); 5279 } 5280 if (IS_DEAD_ARG(2)) { 5281 temp_dead(s, itsh); 5282 } 5283 if (NEED_SYNC_ARG(0)) { 5284 temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0)); 5285 } else if (IS_DEAD_ARG(0)) { 5286 temp_dead(s, ots); 5287 } 5288 return true; 5289 } 5290 5291 static void load_arg_reg(TCGContext *s, TCGReg reg, TCGTemp *ts, 5292 TCGRegSet allocated_regs) 5293 { 5294 if (ts->val_type == TEMP_VAL_REG) { 5295 if (ts->reg != reg) { 5296 tcg_reg_free(s, reg, allocated_regs); 5297 if (!tcg_out_mov(s, ts->type, reg, ts->reg)) { 5298 /* 5299 * Cross register class move not supported. Sync the 5300 * temp back to its slot and load from there. 5301 */ 5302 temp_sync(s, ts, allocated_regs, 0, 0); 5303 tcg_out_ld(s, ts->type, reg, 5304 ts->mem_base->reg, ts->mem_offset); 5305 } 5306 } 5307 } else { 5308 TCGRegSet arg_set = 0; 5309 5310 tcg_reg_free(s, reg, allocated_regs); 5311 tcg_regset_set_reg(arg_set, reg); 5312 temp_load(s, ts, arg_set, allocated_regs, 0); 5313 } 5314 } 5315 5316 static void load_arg_stk(TCGContext *s, unsigned arg_slot, TCGTemp *ts, 5317 TCGRegSet allocated_regs) 5318 { 5319 /* 5320 * When the destination is on the stack, load up the temp and store. 5321 * If there are many call-saved registers, the temp might live to 5322 * see another use; otherwise it'll be discarded. 5323 */ 5324 temp_load(s, ts, tcg_target_available_regs[ts->type], allocated_regs, 0); 5325 tcg_out_st(s, ts->type, ts->reg, TCG_REG_CALL_STACK, 5326 arg_slot_stk_ofs(arg_slot)); 5327 } 5328 5329 static void load_arg_normal(TCGContext *s, const TCGCallArgumentLoc *l, 5330 TCGTemp *ts, TCGRegSet *allocated_regs) 5331 { 5332 if (arg_slot_reg_p(l->arg_slot)) { 5333 TCGReg reg = tcg_target_call_iarg_regs[l->arg_slot]; 5334 load_arg_reg(s, reg, ts, *allocated_regs); 5335 tcg_regset_set_reg(*allocated_regs, reg); 5336 } else { 5337 load_arg_stk(s, l->arg_slot, ts, *allocated_regs); 5338 } 5339 } 5340 5341 static void load_arg_ref(TCGContext *s, unsigned arg_slot, TCGReg ref_base, 5342 intptr_t ref_off, TCGRegSet *allocated_regs) 5343 { 5344 TCGReg reg; 5345 5346 if (arg_slot_reg_p(arg_slot)) { 5347 reg = tcg_target_call_iarg_regs[arg_slot]; 5348 tcg_reg_free(s, reg, *allocated_regs); 5349 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5350 tcg_regset_set_reg(*allocated_regs, reg); 5351 } else { 5352 reg = tcg_reg_alloc(s, tcg_target_available_regs[TCG_TYPE_PTR], 5353 *allocated_regs, 0, false); 5354 tcg_out_addi_ptr(s, reg, ref_base, ref_off); 5355 tcg_out_st(s, TCG_TYPE_PTR, reg, TCG_REG_CALL_STACK, 5356 arg_slot_stk_ofs(arg_slot)); 5357 } 5358 } 5359 5360 static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op) 5361 { 5362 const int nb_oargs = TCGOP_CALLO(op); 5363 const int nb_iargs = TCGOP_CALLI(op); 5364 const TCGLifeData arg_life = op->life; 5365 const TCGHelperInfo *info = tcg_call_info(op); 5366 TCGRegSet allocated_regs = s->reserved_regs; 5367 int i; 5368 5369 /* 5370 * Move inputs into place in reverse order, 5371 * so that we place stacked arguments first. 5372 */ 5373 for (i = nb_iargs - 1; i >= 0; --i) { 5374 const TCGCallArgumentLoc *loc = &info->in[i]; 5375 TCGTemp *ts = arg_temp(op->args[nb_oargs + i]); 5376 5377 switch (loc->kind) { 5378 case TCG_CALL_ARG_NORMAL: 5379 case TCG_CALL_ARG_EXTEND_U: 5380 case TCG_CALL_ARG_EXTEND_S: 5381 load_arg_normal(s, loc, ts, &allocated_regs); 5382 break; 5383 case TCG_CALL_ARG_BY_REF: 5384 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5385 load_arg_ref(s, loc->arg_slot, TCG_REG_CALL_STACK, 5386 arg_slot_stk_ofs(loc->ref_slot), 5387 &allocated_regs); 5388 break; 5389 case TCG_CALL_ARG_BY_REF_N: 5390 load_arg_stk(s, loc->ref_slot, ts, allocated_regs); 5391 break; 5392 default: 5393 g_assert_not_reached(); 5394 } 5395 } 5396 5397 /* Mark dead temporaries and free the associated registers. */ 5398 for (i = nb_oargs; i < nb_iargs + nb_oargs; i++) { 5399 if (IS_DEAD_ARG(i)) { 5400 temp_dead(s, arg_temp(op->args[i])); 5401 } 5402 } 5403 5404 /* Clobber call registers. */ 5405 for (i = 0; i < TCG_TARGET_NB_REGS; i++) { 5406 if (tcg_regset_test_reg(tcg_target_call_clobber_regs, i)) { 5407 tcg_reg_free(s, i, allocated_regs); 5408 } 5409 } 5410 5411 /* 5412 * Save globals if they might be written by the helper, 5413 * sync them if they might be read. 5414 */ 5415 if (info->flags & TCG_CALL_NO_READ_GLOBALS) { 5416 /* Nothing to do */ 5417 } else if (info->flags & TCG_CALL_NO_WRITE_GLOBALS) { 5418 sync_globals(s, allocated_regs); 5419 } else { 5420 save_globals(s, allocated_regs); 5421 } 5422 5423 /* 5424 * If the ABI passes a pointer to the returned struct as the first 5425 * argument, load that now. Pass a pointer to the output home slot. 5426 */ 5427 if (info->out_kind == TCG_CALL_RET_BY_REF) { 5428 TCGTemp *ts = arg_temp(op->args[0]); 5429 5430 if (!ts->mem_allocated) { 5431 temp_allocate_frame(s, ts); 5432 } 5433 load_arg_ref(s, 0, ts->mem_base->reg, ts->mem_offset, &allocated_regs); 5434 } 5435 5436 tcg_out_call(s, tcg_call_func(op), info); 5437 5438 /* Assign output registers and emit moves if needed. */ 5439 switch (info->out_kind) { 5440 case TCG_CALL_RET_NORMAL: 5441 for (i = 0; i < nb_oargs; i++) { 5442 TCGTemp *ts = arg_temp(op->args[i]); 5443 TCGReg reg = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, i); 5444 5445 /* ENV should not be modified. */ 5446 tcg_debug_assert(!temp_readonly(ts)); 5447 5448 set_temp_val_reg(s, ts, reg); 5449 ts->mem_coherent = 0; 5450 } 5451 break; 5452 5453 case TCG_CALL_RET_BY_VEC: 5454 { 5455 TCGTemp *ts = arg_temp(op->args[0]); 5456 5457 tcg_debug_assert(ts->base_type == TCG_TYPE_I128); 5458 tcg_debug_assert(ts->temp_subindex == 0); 5459 if (!ts->mem_allocated) { 5460 temp_allocate_frame(s, ts); 5461 } 5462 tcg_out_st(s, TCG_TYPE_V128, 5463 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5464 ts->mem_base->reg, ts->mem_offset); 5465 } 5466 /* fall through to mark all parts in memory */ 5467 5468 case TCG_CALL_RET_BY_REF: 5469 /* The callee has performed a write through the reference. */ 5470 for (i = 0; i < nb_oargs; i++) { 5471 TCGTemp *ts = arg_temp(op->args[i]); 5472 ts->val_type = TEMP_VAL_MEM; 5473 } 5474 break; 5475 5476 default: 5477 g_assert_not_reached(); 5478 } 5479 5480 /* Flush or discard output registers as needed. */ 5481 for (i = 0; i < nb_oargs; i++) { 5482 TCGTemp *ts = arg_temp(op->args[i]); 5483 if (NEED_SYNC_ARG(i)) { 5484 temp_sync(s, ts, s->reserved_regs, 0, IS_DEAD_ARG(i)); 5485 } else if (IS_DEAD_ARG(i)) { 5486 temp_dead(s, ts); 5487 } 5488 } 5489 } 5490 5491 /** 5492 * atom_and_align_for_opc: 5493 * @s: tcg context 5494 * @opc: memory operation code 5495 * @host_atom: MO_ATOM_{IFALIGN,WITHIN16,SUBALIGN} for host operations 5496 * @allow_two_ops: true if we are prepared to issue two operations 5497 * 5498 * Return the alignment and atomicity to use for the inline fast path 5499 * for the given memory operation. The alignment may be larger than 5500 * that specified in @opc, and the correct alignment will be diagnosed 5501 * by the slow path helper. 5502 * 5503 * If @allow_two_ops, the host is prepared to test for 2x alignment, 5504 * and issue two loads or stores for subalignment. 5505 */ 5506 static TCGAtomAlign atom_and_align_for_opc(TCGContext *s, MemOp opc, 5507 MemOp host_atom, bool allow_two_ops) 5508 { 5509 MemOp align = memop_alignment_bits(opc); 5510 MemOp size = opc & MO_SIZE; 5511 MemOp half = size ? size - 1 : 0; 5512 MemOp atom = opc & MO_ATOM_MASK; 5513 MemOp atmax; 5514 5515 switch (atom) { 5516 case MO_ATOM_NONE: 5517 /* The operation requires no specific atomicity. */ 5518 atmax = MO_8; 5519 break; 5520 5521 case MO_ATOM_IFALIGN: 5522 atmax = size; 5523 break; 5524 5525 case MO_ATOM_IFALIGN_PAIR: 5526 atmax = half; 5527 break; 5528 5529 case MO_ATOM_WITHIN16: 5530 atmax = size; 5531 if (size == MO_128) { 5532 /* Misalignment implies !within16, and therefore no atomicity. */ 5533 } else if (host_atom != MO_ATOM_WITHIN16) { 5534 /* The host does not implement within16, so require alignment. */ 5535 align = MAX(align, size); 5536 } 5537 break; 5538 5539 case MO_ATOM_WITHIN16_PAIR: 5540 atmax = size; 5541 /* 5542 * Misalignment implies !within16, and therefore half atomicity. 5543 * Any host prepared for two operations can implement this with 5544 * half alignment. 5545 */ 5546 if (host_atom != MO_ATOM_WITHIN16 && allow_two_ops) { 5547 align = MAX(align, half); 5548 } 5549 break; 5550 5551 case MO_ATOM_SUBALIGN: 5552 atmax = size; 5553 if (host_atom != MO_ATOM_SUBALIGN) { 5554 /* If unaligned but not odd, there are subobjects up to half. */ 5555 if (allow_two_ops) { 5556 align = MAX(align, half); 5557 } else { 5558 align = MAX(align, size); 5559 } 5560 } 5561 break; 5562 5563 default: 5564 g_assert_not_reached(); 5565 } 5566 5567 return (TCGAtomAlign){ .atom = atmax, .align = align }; 5568 } 5569 5570 /* 5571 * Similarly for qemu_ld/st slow path helpers. 5572 * We must re-implement tcg_gen_callN and tcg_reg_alloc_call simultaneously, 5573 * using only the provided backend tcg_out_* functions. 5574 */ 5575 5576 static int tcg_out_helper_stk_ofs(TCGType type, unsigned slot) 5577 { 5578 int ofs = arg_slot_stk_ofs(slot); 5579 5580 /* 5581 * Each stack slot is TCG_TARGET_LONG_BITS. If the host does not 5582 * require extension to uint64_t, adjust the address for uint32_t. 5583 */ 5584 if (HOST_BIG_ENDIAN && 5585 TCG_TARGET_REG_BITS == 64 && 5586 type == TCG_TYPE_I32) { 5587 ofs += 4; 5588 } 5589 return ofs; 5590 } 5591 5592 static void tcg_out_helper_load_slots(TCGContext *s, 5593 unsigned nmov, TCGMovExtend *mov, 5594 const TCGLdstHelperParam *parm) 5595 { 5596 unsigned i; 5597 TCGReg dst3; 5598 5599 /* 5600 * Start from the end, storing to the stack first. 5601 * This frees those registers, so we need not consider overlap. 5602 */ 5603 for (i = nmov; i-- > 0; ) { 5604 unsigned slot = mov[i].dst; 5605 5606 if (arg_slot_reg_p(slot)) { 5607 goto found_reg; 5608 } 5609 5610 TCGReg src = mov[i].src; 5611 TCGType dst_type = mov[i].dst_type; 5612 MemOp dst_mo = dst_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5613 5614 /* The argument is going onto the stack; extend into scratch. */ 5615 if ((mov[i].src_ext & MO_SIZE) != dst_mo) { 5616 tcg_debug_assert(parm->ntmp != 0); 5617 mov[i].dst = src = parm->tmp[0]; 5618 tcg_out_movext1(s, &mov[i]); 5619 } 5620 5621 tcg_out_st(s, dst_type, src, TCG_REG_CALL_STACK, 5622 tcg_out_helper_stk_ofs(dst_type, slot)); 5623 } 5624 return; 5625 5626 found_reg: 5627 /* 5628 * The remaining arguments are in registers. 5629 * Convert slot numbers to argument registers. 5630 */ 5631 nmov = i + 1; 5632 for (i = 0; i < nmov; ++i) { 5633 mov[i].dst = tcg_target_call_iarg_regs[mov[i].dst]; 5634 } 5635 5636 switch (nmov) { 5637 case 4: 5638 /* The backend must have provided enough temps for the worst case. */ 5639 tcg_debug_assert(parm->ntmp >= 2); 5640 5641 dst3 = mov[3].dst; 5642 for (unsigned j = 0; j < 3; ++j) { 5643 if (dst3 == mov[j].src) { 5644 /* 5645 * Conflict. Copy the source to a temporary, perform the 5646 * remaining moves, then the extension from our scratch 5647 * on the way out. 5648 */ 5649 TCGReg scratch = parm->tmp[1]; 5650 5651 tcg_out_mov(s, mov[3].src_type, scratch, mov[3].src); 5652 tcg_out_movext3(s, mov, mov + 1, mov + 2, parm->tmp[0]); 5653 tcg_out_movext1_new_src(s, &mov[3], scratch); 5654 break; 5655 } 5656 } 5657 5658 /* No conflicts: perform this move and continue. */ 5659 tcg_out_movext1(s, &mov[3]); 5660 /* fall through */ 5661 5662 case 3: 5663 tcg_out_movext3(s, mov, mov + 1, mov + 2, 5664 parm->ntmp ? parm->tmp[0] : -1); 5665 break; 5666 case 2: 5667 tcg_out_movext2(s, mov, mov + 1, 5668 parm->ntmp ? parm->tmp[0] : -1); 5669 break; 5670 case 1: 5671 tcg_out_movext1(s, mov); 5672 break; 5673 default: 5674 g_assert_not_reached(); 5675 } 5676 } 5677 5678 static void tcg_out_helper_load_imm(TCGContext *s, unsigned slot, 5679 TCGType type, tcg_target_long imm, 5680 const TCGLdstHelperParam *parm) 5681 { 5682 if (arg_slot_reg_p(slot)) { 5683 tcg_out_movi(s, type, tcg_target_call_iarg_regs[slot], imm); 5684 } else { 5685 int ofs = tcg_out_helper_stk_ofs(type, slot); 5686 if (!tcg_out_sti(s, type, imm, TCG_REG_CALL_STACK, ofs)) { 5687 tcg_debug_assert(parm->ntmp != 0); 5688 tcg_out_movi(s, type, parm->tmp[0], imm); 5689 tcg_out_st(s, type, parm->tmp[0], TCG_REG_CALL_STACK, ofs); 5690 } 5691 } 5692 } 5693 5694 static void tcg_out_helper_load_common_args(TCGContext *s, 5695 const TCGLabelQemuLdst *ldst, 5696 const TCGLdstHelperParam *parm, 5697 const TCGHelperInfo *info, 5698 unsigned next_arg) 5699 { 5700 TCGMovExtend ptr_mov = { 5701 .dst_type = TCG_TYPE_PTR, 5702 .src_type = TCG_TYPE_PTR, 5703 .src_ext = sizeof(void *) == 4 ? MO_32 : MO_64 5704 }; 5705 const TCGCallArgumentLoc *loc = &info->in[0]; 5706 TCGType type; 5707 unsigned slot; 5708 tcg_target_ulong imm; 5709 5710 /* 5711 * Handle env, which is always first. 5712 */ 5713 ptr_mov.dst = loc->arg_slot; 5714 ptr_mov.src = TCG_AREG0; 5715 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5716 5717 /* 5718 * Handle oi. 5719 */ 5720 imm = ldst->oi; 5721 loc = &info->in[next_arg]; 5722 type = TCG_TYPE_I32; 5723 switch (loc->kind) { 5724 case TCG_CALL_ARG_NORMAL: 5725 break; 5726 case TCG_CALL_ARG_EXTEND_U: 5727 case TCG_CALL_ARG_EXTEND_S: 5728 /* No extension required for MemOpIdx. */ 5729 tcg_debug_assert(imm <= INT32_MAX); 5730 type = TCG_TYPE_REG; 5731 break; 5732 default: 5733 g_assert_not_reached(); 5734 } 5735 tcg_out_helper_load_imm(s, loc->arg_slot, type, imm, parm); 5736 next_arg++; 5737 5738 /* 5739 * Handle ra. 5740 */ 5741 loc = &info->in[next_arg]; 5742 slot = loc->arg_slot; 5743 if (parm->ra_gen) { 5744 int arg_reg = -1; 5745 TCGReg ra_reg; 5746 5747 if (arg_slot_reg_p(slot)) { 5748 arg_reg = tcg_target_call_iarg_regs[slot]; 5749 } 5750 ra_reg = parm->ra_gen(s, ldst, arg_reg); 5751 5752 ptr_mov.dst = slot; 5753 ptr_mov.src = ra_reg; 5754 tcg_out_helper_load_slots(s, 1, &ptr_mov, parm); 5755 } else { 5756 imm = (uintptr_t)ldst->raddr; 5757 tcg_out_helper_load_imm(s, slot, TCG_TYPE_PTR, imm, parm); 5758 } 5759 } 5760 5761 static unsigned tcg_out_helper_add_mov(TCGMovExtend *mov, 5762 const TCGCallArgumentLoc *loc, 5763 TCGType dst_type, TCGType src_type, 5764 TCGReg lo, TCGReg hi) 5765 { 5766 MemOp reg_mo; 5767 5768 if (dst_type <= TCG_TYPE_REG) { 5769 MemOp src_ext; 5770 5771 switch (loc->kind) { 5772 case TCG_CALL_ARG_NORMAL: 5773 src_ext = src_type == TCG_TYPE_I32 ? MO_32 : MO_64; 5774 break; 5775 case TCG_CALL_ARG_EXTEND_U: 5776 dst_type = TCG_TYPE_REG; 5777 src_ext = MO_UL; 5778 break; 5779 case TCG_CALL_ARG_EXTEND_S: 5780 dst_type = TCG_TYPE_REG; 5781 src_ext = MO_SL; 5782 break; 5783 default: 5784 g_assert_not_reached(); 5785 } 5786 5787 mov[0].dst = loc->arg_slot; 5788 mov[0].dst_type = dst_type; 5789 mov[0].src = lo; 5790 mov[0].src_type = src_type; 5791 mov[0].src_ext = src_ext; 5792 return 1; 5793 } 5794 5795 if (TCG_TARGET_REG_BITS == 32) { 5796 assert(dst_type == TCG_TYPE_I64); 5797 reg_mo = MO_32; 5798 } else { 5799 assert(dst_type == TCG_TYPE_I128); 5800 reg_mo = MO_64; 5801 } 5802 5803 mov[0].dst = loc[HOST_BIG_ENDIAN].arg_slot; 5804 mov[0].src = lo; 5805 mov[0].dst_type = TCG_TYPE_REG; 5806 mov[0].src_type = TCG_TYPE_REG; 5807 mov[0].src_ext = reg_mo; 5808 5809 mov[1].dst = loc[!HOST_BIG_ENDIAN].arg_slot; 5810 mov[1].src = hi; 5811 mov[1].dst_type = TCG_TYPE_REG; 5812 mov[1].src_type = TCG_TYPE_REG; 5813 mov[1].src_ext = reg_mo; 5814 5815 return 2; 5816 } 5817 5818 static void tcg_out_ld_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 5819 const TCGLdstHelperParam *parm) 5820 { 5821 const TCGHelperInfo *info; 5822 const TCGCallArgumentLoc *loc; 5823 TCGMovExtend mov[2]; 5824 unsigned next_arg, nmov; 5825 MemOp mop = get_memop(ldst->oi); 5826 5827 switch (mop & MO_SIZE) { 5828 case MO_8: 5829 case MO_16: 5830 case MO_32: 5831 info = &info_helper_ld32_mmu; 5832 break; 5833 case MO_64: 5834 info = &info_helper_ld64_mmu; 5835 break; 5836 case MO_128: 5837 info = &info_helper_ld128_mmu; 5838 break; 5839 default: 5840 g_assert_not_reached(); 5841 } 5842 5843 /* Defer env argument. */ 5844 next_arg = 1; 5845 5846 loc = &info->in[next_arg]; 5847 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 5848 /* 5849 * 32-bit host with 32-bit guest: zero-extend the guest address 5850 * to 64-bits for the helper by storing the low part, then 5851 * load a 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 tcg_out_helper_load_slots(s, 1, mov, parm); 5857 5858 tcg_out_helper_load_imm(s, loc[!HOST_BIG_ENDIAN].arg_slot, 5859 TCG_TYPE_I32, 0, parm); 5860 next_arg += 2; 5861 } else { 5862 nmov = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 5863 ldst->addrlo_reg, ldst->addrhi_reg); 5864 tcg_out_helper_load_slots(s, nmov, mov, parm); 5865 next_arg += nmov; 5866 } 5867 5868 switch (info->out_kind) { 5869 case TCG_CALL_RET_NORMAL: 5870 case TCG_CALL_RET_BY_VEC: 5871 break; 5872 case TCG_CALL_RET_BY_REF: 5873 /* 5874 * The return reference is in the first argument slot. 5875 * We need memory in which to return: re-use the top of stack. 5876 */ 5877 { 5878 int ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 5879 5880 if (arg_slot_reg_p(0)) { 5881 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[0], 5882 TCG_REG_CALL_STACK, ofs_slot0); 5883 } else { 5884 tcg_debug_assert(parm->ntmp != 0); 5885 tcg_out_addi_ptr(s, parm->tmp[0], 5886 TCG_REG_CALL_STACK, ofs_slot0); 5887 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 5888 TCG_REG_CALL_STACK, ofs_slot0); 5889 } 5890 } 5891 break; 5892 default: 5893 g_assert_not_reached(); 5894 } 5895 5896 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 5897 } 5898 5899 static void tcg_out_ld_helper_ret(TCGContext *s, const TCGLabelQemuLdst *ldst, 5900 bool load_sign, 5901 const TCGLdstHelperParam *parm) 5902 { 5903 MemOp mop = get_memop(ldst->oi); 5904 TCGMovExtend mov[2]; 5905 int ofs_slot0; 5906 5907 switch (ldst->type) { 5908 case TCG_TYPE_I64: 5909 if (TCG_TARGET_REG_BITS == 32) { 5910 break; 5911 } 5912 /* fall through */ 5913 5914 case TCG_TYPE_I32: 5915 mov[0].dst = ldst->datalo_reg; 5916 mov[0].src = tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, 0); 5917 mov[0].dst_type = ldst->type; 5918 mov[0].src_type = TCG_TYPE_REG; 5919 5920 /* 5921 * If load_sign, then we allowed the helper to perform the 5922 * appropriate sign extension to tcg_target_ulong, and all 5923 * we need now is a plain move. 5924 * 5925 * If they do not, then we expect the relevant extension 5926 * instruction to be no more expensive than a move, and 5927 * we thus save the icache etc by only using one of two 5928 * helper functions. 5929 */ 5930 if (load_sign || !(mop & MO_SIGN)) { 5931 if (TCG_TARGET_REG_BITS == 32 || ldst->type == TCG_TYPE_I32) { 5932 mov[0].src_ext = MO_32; 5933 } else { 5934 mov[0].src_ext = MO_64; 5935 } 5936 } else { 5937 mov[0].src_ext = mop & MO_SSIZE; 5938 } 5939 tcg_out_movext1(s, mov); 5940 return; 5941 5942 case TCG_TYPE_I128: 5943 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 5944 ofs_slot0 = TCG_TARGET_CALL_STACK_OFFSET; 5945 switch (TCG_TARGET_CALL_RET_I128) { 5946 case TCG_CALL_RET_NORMAL: 5947 break; 5948 case TCG_CALL_RET_BY_VEC: 5949 tcg_out_st(s, TCG_TYPE_V128, 5950 tcg_target_call_oarg_reg(TCG_CALL_RET_BY_VEC, 0), 5951 TCG_REG_CALL_STACK, ofs_slot0); 5952 /* fall through */ 5953 case TCG_CALL_RET_BY_REF: 5954 tcg_out_ld(s, TCG_TYPE_I64, ldst->datalo_reg, 5955 TCG_REG_CALL_STACK, ofs_slot0 + 8 * HOST_BIG_ENDIAN); 5956 tcg_out_ld(s, TCG_TYPE_I64, ldst->datahi_reg, 5957 TCG_REG_CALL_STACK, ofs_slot0 + 8 * !HOST_BIG_ENDIAN); 5958 return; 5959 default: 5960 g_assert_not_reached(); 5961 } 5962 break; 5963 5964 default: 5965 g_assert_not_reached(); 5966 } 5967 5968 mov[0].dst = ldst->datalo_reg; 5969 mov[0].src = 5970 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, HOST_BIG_ENDIAN); 5971 mov[0].dst_type = TCG_TYPE_REG; 5972 mov[0].src_type = TCG_TYPE_REG; 5973 mov[0].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 5974 5975 mov[1].dst = ldst->datahi_reg; 5976 mov[1].src = 5977 tcg_target_call_oarg_reg(TCG_CALL_RET_NORMAL, !HOST_BIG_ENDIAN); 5978 mov[1].dst_type = TCG_TYPE_REG; 5979 mov[1].src_type = TCG_TYPE_REG; 5980 mov[1].src_ext = TCG_TARGET_REG_BITS == 32 ? MO_32 : MO_64; 5981 5982 tcg_out_movext2(s, mov, mov + 1, parm->ntmp ? parm->tmp[0] : -1); 5983 } 5984 5985 static void tcg_out_st_helper_args(TCGContext *s, const TCGLabelQemuLdst *ldst, 5986 const TCGLdstHelperParam *parm) 5987 { 5988 const TCGHelperInfo *info; 5989 const TCGCallArgumentLoc *loc; 5990 TCGMovExtend mov[4]; 5991 TCGType data_type; 5992 unsigned next_arg, nmov, n; 5993 MemOp mop = get_memop(ldst->oi); 5994 5995 switch (mop & MO_SIZE) { 5996 case MO_8: 5997 case MO_16: 5998 case MO_32: 5999 info = &info_helper_st32_mmu; 6000 data_type = TCG_TYPE_I32; 6001 break; 6002 case MO_64: 6003 info = &info_helper_st64_mmu; 6004 data_type = TCG_TYPE_I64; 6005 break; 6006 case MO_128: 6007 info = &info_helper_st128_mmu; 6008 data_type = TCG_TYPE_I128; 6009 break; 6010 default: 6011 g_assert_not_reached(); 6012 } 6013 6014 /* Defer env argument. */ 6015 next_arg = 1; 6016 nmov = 0; 6017 6018 /* Handle addr argument. */ 6019 loc = &info->in[next_arg]; 6020 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 6021 /* 6022 * 32-bit host with 32-bit guest: zero-extend the guest address 6023 * to 64-bits for the helper by storing the low part. Later, 6024 * after we have processed the register inputs, we will load a 6025 * zero for the high part. 6026 */ 6027 tcg_out_helper_add_mov(mov, loc + HOST_BIG_ENDIAN, 6028 TCG_TYPE_I32, TCG_TYPE_I32, 6029 ldst->addrlo_reg, -1); 6030 next_arg += 2; 6031 nmov += 1; 6032 } else { 6033 n = tcg_out_helper_add_mov(mov, loc, TCG_TYPE_I64, s->addr_type, 6034 ldst->addrlo_reg, ldst->addrhi_reg); 6035 next_arg += n; 6036 nmov += n; 6037 } 6038 6039 /* Handle data argument. */ 6040 loc = &info->in[next_arg]; 6041 switch (loc->kind) { 6042 case TCG_CALL_ARG_NORMAL: 6043 case TCG_CALL_ARG_EXTEND_U: 6044 case TCG_CALL_ARG_EXTEND_S: 6045 n = tcg_out_helper_add_mov(mov + nmov, loc, data_type, ldst->type, 6046 ldst->datalo_reg, ldst->datahi_reg); 6047 next_arg += n; 6048 nmov += n; 6049 tcg_out_helper_load_slots(s, nmov, mov, parm); 6050 break; 6051 6052 case TCG_CALL_ARG_BY_REF: 6053 tcg_debug_assert(TCG_TARGET_REG_BITS == 64); 6054 tcg_debug_assert(data_type == TCG_TYPE_I128); 6055 tcg_out_st(s, TCG_TYPE_I64, 6056 HOST_BIG_ENDIAN ? ldst->datahi_reg : ldst->datalo_reg, 6057 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[0].ref_slot)); 6058 tcg_out_st(s, TCG_TYPE_I64, 6059 HOST_BIG_ENDIAN ? ldst->datalo_reg : ldst->datahi_reg, 6060 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc[1].ref_slot)); 6061 6062 tcg_out_helper_load_slots(s, nmov, mov, parm); 6063 6064 if (arg_slot_reg_p(loc->arg_slot)) { 6065 tcg_out_addi_ptr(s, tcg_target_call_iarg_regs[loc->arg_slot], 6066 TCG_REG_CALL_STACK, 6067 arg_slot_stk_ofs(loc->ref_slot)); 6068 } else { 6069 tcg_debug_assert(parm->ntmp != 0); 6070 tcg_out_addi_ptr(s, parm->tmp[0], TCG_REG_CALL_STACK, 6071 arg_slot_stk_ofs(loc->ref_slot)); 6072 tcg_out_st(s, TCG_TYPE_PTR, parm->tmp[0], 6073 TCG_REG_CALL_STACK, arg_slot_stk_ofs(loc->arg_slot)); 6074 } 6075 next_arg += 2; 6076 break; 6077 6078 default: 6079 g_assert_not_reached(); 6080 } 6081 6082 if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I32) { 6083 /* Zero extend the address by loading a zero for the high part. */ 6084 loc = &info->in[1 + !HOST_BIG_ENDIAN]; 6085 tcg_out_helper_load_imm(s, loc->arg_slot, TCG_TYPE_I32, 0, parm); 6086 } 6087 6088 tcg_out_helper_load_common_args(s, ldst, parm, info, next_arg); 6089 } 6090 6091 int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start) 6092 { 6093 int i, start_words, num_insns; 6094 TCGOp *op; 6095 6096 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP) 6097 && qemu_log_in_addr_range(pc_start))) { 6098 FILE *logfile = qemu_log_trylock(); 6099 if (logfile) { 6100 fprintf(logfile, "OP:\n"); 6101 tcg_dump_ops(s, logfile, false); 6102 fprintf(logfile, "\n"); 6103 qemu_log_unlock(logfile); 6104 } 6105 } 6106 6107 #ifdef CONFIG_DEBUG_TCG 6108 /* Ensure all labels referenced have been emitted. */ 6109 { 6110 TCGLabel *l; 6111 bool error = false; 6112 6113 QSIMPLEQ_FOREACH(l, &s->labels, next) { 6114 if (unlikely(!l->present) && !QSIMPLEQ_EMPTY(&l->branches)) { 6115 qemu_log_mask(CPU_LOG_TB_OP, 6116 "$L%d referenced but not present.\n", l->id); 6117 error = true; 6118 } 6119 } 6120 assert(!error); 6121 } 6122 #endif 6123 6124 tcg_optimize(s); 6125 6126 reachable_code_pass(s); 6127 liveness_pass_0(s); 6128 liveness_pass_1(s); 6129 6130 if (s->nb_indirects > 0) { 6131 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_IND) 6132 && qemu_log_in_addr_range(pc_start))) { 6133 FILE *logfile = qemu_log_trylock(); 6134 if (logfile) { 6135 fprintf(logfile, "OP before indirect lowering:\n"); 6136 tcg_dump_ops(s, logfile, false); 6137 fprintf(logfile, "\n"); 6138 qemu_log_unlock(logfile); 6139 } 6140 } 6141 6142 /* Replace indirect temps with direct temps. */ 6143 if (liveness_pass_2(s)) { 6144 /* If changes were made, re-run liveness. */ 6145 liveness_pass_1(s); 6146 } 6147 } 6148 6149 if (unlikely(qemu_loglevel_mask(CPU_LOG_TB_OP_OPT) 6150 && qemu_log_in_addr_range(pc_start))) { 6151 FILE *logfile = qemu_log_trylock(); 6152 if (logfile) { 6153 fprintf(logfile, "OP after optimization and liveness analysis:\n"); 6154 tcg_dump_ops(s, logfile, true); 6155 fprintf(logfile, "\n"); 6156 qemu_log_unlock(logfile); 6157 } 6158 } 6159 6160 /* Initialize goto_tb jump offsets. */ 6161 tb->jmp_reset_offset[0] = TB_JMP_OFFSET_INVALID; 6162 tb->jmp_reset_offset[1] = TB_JMP_OFFSET_INVALID; 6163 tb->jmp_insn_offset[0] = TB_JMP_OFFSET_INVALID; 6164 tb->jmp_insn_offset[1] = TB_JMP_OFFSET_INVALID; 6165 6166 tcg_reg_alloc_start(s); 6167 6168 /* 6169 * Reset the buffer pointers when restarting after overflow. 6170 * TODO: Move this into translate-all.c with the rest of the 6171 * buffer management. Having only this done here is confusing. 6172 */ 6173 s->code_buf = tcg_splitwx_to_rw(tb->tc.ptr); 6174 s->code_ptr = s->code_buf; 6175 6176 #ifdef TCG_TARGET_NEED_LDST_LABELS 6177 QSIMPLEQ_INIT(&s->ldst_labels); 6178 #endif 6179 #ifdef TCG_TARGET_NEED_POOL_LABELS 6180 s->pool_labels = NULL; 6181 #endif 6182 6183 start_words = s->insn_start_words; 6184 s->gen_insn_data = 6185 tcg_malloc(sizeof(uint64_t) * s->gen_tb->icount * start_words); 6186 6187 tcg_out_tb_start(s); 6188 6189 num_insns = -1; 6190 QTAILQ_FOREACH(op, &s->ops, link) { 6191 TCGOpcode opc = op->opc; 6192 6193 switch (opc) { 6194 case INDEX_op_mov_i32: 6195 case INDEX_op_mov_i64: 6196 case INDEX_op_mov_vec: 6197 tcg_reg_alloc_mov(s, op); 6198 break; 6199 case INDEX_op_dup_vec: 6200 tcg_reg_alloc_dup(s, op); 6201 break; 6202 case INDEX_op_insn_start: 6203 if (num_insns >= 0) { 6204 size_t off = tcg_current_code_size(s); 6205 s->gen_insn_end_off[num_insns] = off; 6206 /* Assert that we do not overflow our stored offset. */ 6207 assert(s->gen_insn_end_off[num_insns] == off); 6208 } 6209 num_insns++; 6210 for (i = 0; i < start_words; ++i) { 6211 s->gen_insn_data[num_insns * start_words + i] = 6212 tcg_get_insn_start_param(op, i); 6213 } 6214 break; 6215 case INDEX_op_discard: 6216 temp_dead(s, arg_temp(op->args[0])); 6217 break; 6218 case INDEX_op_set_label: 6219 tcg_reg_alloc_bb_end(s, s->reserved_regs); 6220 tcg_out_label(s, arg_label(op->args[0])); 6221 break; 6222 case INDEX_op_call: 6223 tcg_reg_alloc_call(s, op); 6224 break; 6225 case INDEX_op_exit_tb: 6226 tcg_out_exit_tb(s, op->args[0]); 6227 break; 6228 case INDEX_op_goto_tb: 6229 tcg_out_goto_tb(s, op->args[0]); 6230 break; 6231 case INDEX_op_dup2_vec: 6232 if (tcg_reg_alloc_dup2(s, op)) { 6233 break; 6234 } 6235 /* fall through */ 6236 default: 6237 /* Sanity check that we've not introduced any unhandled opcodes. */ 6238 tcg_debug_assert(tcg_op_supported(opc)); 6239 /* Note: in order to speed up the code, it would be much 6240 faster to have specialized register allocator functions for 6241 some common argument patterns */ 6242 tcg_reg_alloc_op(s, op); 6243 break; 6244 } 6245 /* Test for (pending) buffer overflow. The assumption is that any 6246 one operation beginning below the high water mark cannot overrun 6247 the buffer completely. Thus we can test for overflow after 6248 generating code without having to check during generation. */ 6249 if (unlikely((void *)s->code_ptr > s->code_gen_highwater)) { 6250 return -1; 6251 } 6252 /* Test for TB overflow, as seen by gen_insn_end_off. */ 6253 if (unlikely(tcg_current_code_size(s) > UINT16_MAX)) { 6254 return -2; 6255 } 6256 } 6257 tcg_debug_assert(num_insns + 1 == s->gen_tb->icount); 6258 s->gen_insn_end_off[num_insns] = tcg_current_code_size(s); 6259 6260 /* Generate TB finalization at the end of block */ 6261 #ifdef TCG_TARGET_NEED_LDST_LABELS 6262 i = tcg_out_ldst_finalize(s); 6263 if (i < 0) { 6264 return i; 6265 } 6266 #endif 6267 #ifdef TCG_TARGET_NEED_POOL_LABELS 6268 i = tcg_out_pool_finalize(s); 6269 if (i < 0) { 6270 return i; 6271 } 6272 #endif 6273 if (!tcg_resolve_relocs(s)) { 6274 return -2; 6275 } 6276 6277 #ifndef CONFIG_TCG_INTERPRETER 6278 /* flush instruction cache */ 6279 flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), 6280 (uintptr_t)s->code_buf, 6281 tcg_ptr_byte_diff(s->code_ptr, s->code_buf)); 6282 #endif 6283 6284 return tcg_current_code_size(s); 6285 } 6286 6287 #ifdef ELF_HOST_MACHINE 6288 /* In order to use this feature, the backend needs to do three things: 6289 6290 (1) Define ELF_HOST_MACHINE to indicate both what value to 6291 put into the ELF image and to indicate support for the feature. 6292 6293 (2) Define tcg_register_jit. This should create a buffer containing 6294 the contents of a .debug_frame section that describes the post- 6295 prologue unwind info for the tcg machine. 6296 6297 (3) Call tcg_register_jit_int, with the constructed .debug_frame. 6298 */ 6299 6300 /* Begin GDB interface. THE FOLLOWING MUST MATCH GDB DOCS. */ 6301 typedef enum { 6302 JIT_NOACTION = 0, 6303 JIT_REGISTER_FN, 6304 JIT_UNREGISTER_FN 6305 } jit_actions_t; 6306 6307 struct jit_code_entry { 6308 struct jit_code_entry *next_entry; 6309 struct jit_code_entry *prev_entry; 6310 const void *symfile_addr; 6311 uint64_t symfile_size; 6312 }; 6313 6314 struct jit_descriptor { 6315 uint32_t version; 6316 uint32_t action_flag; 6317 struct jit_code_entry *relevant_entry; 6318 struct jit_code_entry *first_entry; 6319 }; 6320 6321 void __jit_debug_register_code(void) __attribute__((noinline)); 6322 void __jit_debug_register_code(void) 6323 { 6324 asm(""); 6325 } 6326 6327 /* Must statically initialize the version, because GDB may check 6328 the version before we can set it. */ 6329 struct jit_descriptor __jit_debug_descriptor = { 1, 0, 0, 0 }; 6330 6331 /* End GDB interface. */ 6332 6333 static int find_string(const char *strtab, const char *str) 6334 { 6335 const char *p = strtab + 1; 6336 6337 while (1) { 6338 if (strcmp(p, str) == 0) { 6339 return p - strtab; 6340 } 6341 p += strlen(p) + 1; 6342 } 6343 } 6344 6345 static void tcg_register_jit_int(const void *buf_ptr, size_t buf_size, 6346 const void *debug_frame, 6347 size_t debug_frame_size) 6348 { 6349 struct __attribute__((packed)) DebugInfo { 6350 uint32_t len; 6351 uint16_t version; 6352 uint32_t abbrev; 6353 uint8_t ptr_size; 6354 uint8_t cu_die; 6355 uint16_t cu_lang; 6356 uintptr_t cu_low_pc; 6357 uintptr_t cu_high_pc; 6358 uint8_t fn_die; 6359 char fn_name[16]; 6360 uintptr_t fn_low_pc; 6361 uintptr_t fn_high_pc; 6362 uint8_t cu_eoc; 6363 }; 6364 6365 struct ElfImage { 6366 ElfW(Ehdr) ehdr; 6367 ElfW(Phdr) phdr; 6368 ElfW(Shdr) shdr[7]; 6369 ElfW(Sym) sym[2]; 6370 struct DebugInfo di; 6371 uint8_t da[24]; 6372 char str[80]; 6373 }; 6374 6375 struct ElfImage *img; 6376 6377 static const struct ElfImage img_template = { 6378 .ehdr = { 6379 .e_ident[EI_MAG0] = ELFMAG0, 6380 .e_ident[EI_MAG1] = ELFMAG1, 6381 .e_ident[EI_MAG2] = ELFMAG2, 6382 .e_ident[EI_MAG3] = ELFMAG3, 6383 .e_ident[EI_CLASS] = ELF_CLASS, 6384 .e_ident[EI_DATA] = ELF_DATA, 6385 .e_ident[EI_VERSION] = EV_CURRENT, 6386 .e_type = ET_EXEC, 6387 .e_machine = ELF_HOST_MACHINE, 6388 .e_version = EV_CURRENT, 6389 .e_phoff = offsetof(struct ElfImage, phdr), 6390 .e_shoff = offsetof(struct ElfImage, shdr), 6391 .e_ehsize = sizeof(ElfW(Shdr)), 6392 .e_phentsize = sizeof(ElfW(Phdr)), 6393 .e_phnum = 1, 6394 .e_shentsize = sizeof(ElfW(Shdr)), 6395 .e_shnum = ARRAY_SIZE(img->shdr), 6396 .e_shstrndx = ARRAY_SIZE(img->shdr) - 1, 6397 #ifdef ELF_HOST_FLAGS 6398 .e_flags = ELF_HOST_FLAGS, 6399 #endif 6400 #ifdef ELF_OSABI 6401 .e_ident[EI_OSABI] = ELF_OSABI, 6402 #endif 6403 }, 6404 .phdr = { 6405 .p_type = PT_LOAD, 6406 .p_flags = PF_X, 6407 }, 6408 .shdr = { 6409 [0] = { .sh_type = SHT_NULL }, 6410 /* Trick: The contents of code_gen_buffer are not present in 6411 this fake ELF file; that got allocated elsewhere. Therefore 6412 we mark .text as SHT_NOBITS (similar to .bss) so that readers 6413 will not look for contents. We can record any address. */ 6414 [1] = { /* .text */ 6415 .sh_type = SHT_NOBITS, 6416 .sh_flags = SHF_EXECINSTR | SHF_ALLOC, 6417 }, 6418 [2] = { /* .debug_info */ 6419 .sh_type = SHT_PROGBITS, 6420 .sh_offset = offsetof(struct ElfImage, di), 6421 .sh_size = sizeof(struct DebugInfo), 6422 }, 6423 [3] = { /* .debug_abbrev */ 6424 .sh_type = SHT_PROGBITS, 6425 .sh_offset = offsetof(struct ElfImage, da), 6426 .sh_size = sizeof(img->da), 6427 }, 6428 [4] = { /* .debug_frame */ 6429 .sh_type = SHT_PROGBITS, 6430 .sh_offset = sizeof(struct ElfImage), 6431 }, 6432 [5] = { /* .symtab */ 6433 .sh_type = SHT_SYMTAB, 6434 .sh_offset = offsetof(struct ElfImage, sym), 6435 .sh_size = sizeof(img->sym), 6436 .sh_info = 1, 6437 .sh_link = ARRAY_SIZE(img->shdr) - 1, 6438 .sh_entsize = sizeof(ElfW(Sym)), 6439 }, 6440 [6] = { /* .strtab */ 6441 .sh_type = SHT_STRTAB, 6442 .sh_offset = offsetof(struct ElfImage, str), 6443 .sh_size = sizeof(img->str), 6444 } 6445 }, 6446 .sym = { 6447 [1] = { /* code_gen_buffer */ 6448 .st_info = ELF_ST_INFO(STB_GLOBAL, STT_FUNC), 6449 .st_shndx = 1, 6450 } 6451 }, 6452 .di = { 6453 .len = sizeof(struct DebugInfo) - 4, 6454 .version = 2, 6455 .ptr_size = sizeof(void *), 6456 .cu_die = 1, 6457 .cu_lang = 0x8001, /* DW_LANG_Mips_Assembler */ 6458 .fn_die = 2, 6459 .fn_name = "code_gen_buffer" 6460 }, 6461 .da = { 6462 1, /* abbrev number (the cu) */ 6463 0x11, 1, /* DW_TAG_compile_unit, has children */ 6464 0x13, 0x5, /* DW_AT_language, DW_FORM_data2 */ 6465 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6466 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6467 0, 0, /* end of abbrev */ 6468 2, /* abbrev number (the fn) */ 6469 0x2e, 0, /* DW_TAG_subprogram, no children */ 6470 0x3, 0x8, /* DW_AT_name, DW_FORM_string */ 6471 0x11, 0x1, /* DW_AT_low_pc, DW_FORM_addr */ 6472 0x12, 0x1, /* DW_AT_high_pc, DW_FORM_addr */ 6473 0, 0, /* end of abbrev */ 6474 0 /* no more abbrev */ 6475 }, 6476 .str = "\0" ".text\0" ".debug_info\0" ".debug_abbrev\0" 6477 ".debug_frame\0" ".symtab\0" ".strtab\0" "code_gen_buffer", 6478 }; 6479 6480 /* We only need a single jit entry; statically allocate it. */ 6481 static struct jit_code_entry one_entry; 6482 6483 uintptr_t buf = (uintptr_t)buf_ptr; 6484 size_t img_size = sizeof(struct ElfImage) + debug_frame_size; 6485 DebugFrameHeader *dfh; 6486 6487 img = g_malloc(img_size); 6488 *img = img_template; 6489 6490 img->phdr.p_vaddr = buf; 6491 img->phdr.p_paddr = buf; 6492 img->phdr.p_memsz = buf_size; 6493 6494 img->shdr[1].sh_name = find_string(img->str, ".text"); 6495 img->shdr[1].sh_addr = buf; 6496 img->shdr[1].sh_size = buf_size; 6497 6498 img->shdr[2].sh_name = find_string(img->str, ".debug_info"); 6499 img->shdr[3].sh_name = find_string(img->str, ".debug_abbrev"); 6500 6501 img->shdr[4].sh_name = find_string(img->str, ".debug_frame"); 6502 img->shdr[4].sh_size = debug_frame_size; 6503 6504 img->shdr[5].sh_name = find_string(img->str, ".symtab"); 6505 img->shdr[6].sh_name = find_string(img->str, ".strtab"); 6506 6507 img->sym[1].st_name = find_string(img->str, "code_gen_buffer"); 6508 img->sym[1].st_value = buf; 6509 img->sym[1].st_size = buf_size; 6510 6511 img->di.cu_low_pc = buf; 6512 img->di.cu_high_pc = buf + buf_size; 6513 img->di.fn_low_pc = buf; 6514 img->di.fn_high_pc = buf + buf_size; 6515 6516 dfh = (DebugFrameHeader *)(img + 1); 6517 memcpy(dfh, debug_frame, debug_frame_size); 6518 dfh->fde.func_start = buf; 6519 dfh->fde.func_len = buf_size; 6520 6521 #ifdef DEBUG_JIT 6522 /* Enable this block to be able to debug the ELF image file creation. 6523 One can use readelf, objdump, or other inspection utilities. */ 6524 { 6525 g_autofree char *jit = g_strdup_printf("%s/qemu.jit", g_get_tmp_dir()); 6526 FILE *f = fopen(jit, "w+b"); 6527 if (f) { 6528 if (fwrite(img, img_size, 1, f) != img_size) { 6529 /* Avoid stupid unused return value warning for fwrite. */ 6530 } 6531 fclose(f); 6532 } 6533 } 6534 #endif 6535 6536 one_entry.symfile_addr = img; 6537 one_entry.symfile_size = img_size; 6538 6539 __jit_debug_descriptor.action_flag = JIT_REGISTER_FN; 6540 __jit_debug_descriptor.relevant_entry = &one_entry; 6541 __jit_debug_descriptor.first_entry = &one_entry; 6542 __jit_debug_register_code(); 6543 } 6544 #else 6545 /* No support for the feature. Provide the entry point expected by exec.c, 6546 and implement the internal function we declared earlier. */ 6547 6548 static void tcg_register_jit_int(const void *buf, size_t size, 6549 const void *debug_frame, 6550 size_t debug_frame_size) 6551 { 6552 } 6553 6554 void tcg_register_jit(const void *buf, size_t buf_size) 6555 { 6556 } 6557 #endif /* ELF_HOST_MACHINE */ 6558 6559 #if !TCG_TARGET_MAYBE_vec 6560 void tcg_expand_vec_op(TCGOpcode o, TCGType t, unsigned e, TCGArg a0, ...) 6561 { 6562 g_assert_not_reached(); 6563 } 6564 #endif 6565