1 /* SPDX-License-Identifier: GPL-2.0 */ 2 3 #ifndef _ASM_X86_NOSPEC_BRANCH_H_ 4 #define _ASM_X86_NOSPEC_BRANCH_H_ 5 6 #include <linux/static_key.h> 7 #include <linux/objtool.h> 8 #include <linux/linkage.h> 9 10 #include <asm/alternative.h> 11 #include <asm/cpufeatures.h> 12 #include <asm/msr-index.h> 13 #include <asm/unwind_hints.h> 14 #include <asm/percpu.h> 15 #include <asm/current.h> 16 17 /* 18 * Call depth tracking for Intel SKL CPUs to address the RSB underflow 19 * issue in software. 20 * 21 * The tracking does not use a counter. It uses uses arithmetic shift 22 * right on call entry and logical shift left on return. 23 * 24 * The depth tracking variable is initialized to 0x8000.... when the call 25 * depth is zero. The arithmetic shift right sign extends the MSB and 26 * saturates after the 12th call. The shift count is 5 for both directions 27 * so the tracking covers 12 nested calls. 28 * 29 * Call 30 * 0: 0x8000000000000000 0x0000000000000000 31 * 1: 0xfc00000000000000 0xf000000000000000 32 * ... 33 * 11: 0xfffffffffffffff8 0xfffffffffffffc00 34 * 12: 0xffffffffffffffff 0xffffffffffffffe0 35 * 36 * After a return buffer fill the depth is credited 12 calls before the 37 * next stuffing has to take place. 38 * 39 * There is a inaccuracy for situations like this: 40 * 41 * 10 calls 42 * 5 returns 43 * 3 calls 44 * 4 returns 45 * 3 calls 46 * .... 47 * 48 * The shift count might cause this to be off by one in either direction, 49 * but there is still a cushion vs. the RSB depth. The algorithm does not 50 * claim to be perfect and it can be speculated around by the CPU, but it 51 * is considered that it obfuscates the problem enough to make exploitation 52 * extremly difficult. 53 */ 54 #define RET_DEPTH_SHIFT 5 55 #define RSB_RET_STUFF_LOOPS 16 56 #define RET_DEPTH_INIT 0x8000000000000000ULL 57 #define RET_DEPTH_INIT_FROM_CALL 0xfc00000000000000ULL 58 #define RET_DEPTH_CREDIT 0xffffffffffffffffULL 59 60 #ifdef CONFIG_CALL_THUNKS_DEBUG 61 # define CALL_THUNKS_DEBUG_INC_CALLS \ 62 incq %gs:__x86_call_count; 63 # define CALL_THUNKS_DEBUG_INC_RETS \ 64 incq %gs:__x86_ret_count; 65 # define CALL_THUNKS_DEBUG_INC_STUFFS \ 66 incq %gs:__x86_stuffs_count; 67 # define CALL_THUNKS_DEBUG_INC_CTXSW \ 68 incq %gs:__x86_ctxsw_count; 69 #else 70 # define CALL_THUNKS_DEBUG_INC_CALLS 71 # define CALL_THUNKS_DEBUG_INC_RETS 72 # define CALL_THUNKS_DEBUG_INC_STUFFS 73 # define CALL_THUNKS_DEBUG_INC_CTXSW 74 #endif 75 76 #if defined(CONFIG_CALL_DEPTH_TRACKING) && !defined(COMPILE_OFFSETS) 77 78 #include <asm/asm-offsets.h> 79 80 #define CREDIT_CALL_DEPTH \ 81 movq $-1, PER_CPU_VAR(pcpu_hot + X86_call_depth); 82 83 #define ASM_CREDIT_CALL_DEPTH \ 84 movq $-1, PER_CPU_VAR(pcpu_hot + X86_call_depth); 85 86 #define RESET_CALL_DEPTH \ 87 xor %eax, %eax; \ 88 bts $63, %rax; \ 89 movq %rax, PER_CPU_VAR(pcpu_hot + X86_call_depth); 90 91 #define RESET_CALL_DEPTH_FROM_CALL \ 92 movb $0xfc, %al; \ 93 shl $56, %rax; \ 94 movq %rax, PER_CPU_VAR(pcpu_hot + X86_call_depth); \ 95 CALL_THUNKS_DEBUG_INC_CALLS 96 97 #define INCREMENT_CALL_DEPTH \ 98 sarq $5, %gs:pcpu_hot + X86_call_depth; \ 99 CALL_THUNKS_DEBUG_INC_CALLS 100 101 #define ASM_INCREMENT_CALL_DEPTH \ 102 sarq $5, PER_CPU_VAR(pcpu_hot + X86_call_depth); \ 103 CALL_THUNKS_DEBUG_INC_CALLS 104 105 #else 106 #define CREDIT_CALL_DEPTH 107 #define ASM_CREDIT_CALL_DEPTH 108 #define RESET_CALL_DEPTH 109 #define INCREMENT_CALL_DEPTH 110 #define ASM_INCREMENT_CALL_DEPTH 111 #define RESET_CALL_DEPTH_FROM_CALL 112 #endif 113 114 /* 115 * Fill the CPU return stack buffer. 116 * 117 * Each entry in the RSB, if used for a speculative 'ret', contains an 118 * infinite 'pause; lfence; jmp' loop to capture speculative execution. 119 * 120 * This is required in various cases for retpoline and IBRS-based 121 * mitigations for the Spectre variant 2 vulnerability. Sometimes to 122 * eliminate potentially bogus entries from the RSB, and sometimes 123 * purely to ensure that it doesn't get empty, which on some CPUs would 124 * allow predictions from other (unwanted!) sources to be used. 125 * 126 * We define a CPP macro such that it can be used from both .S files and 127 * inline assembly. It's possible to do a .macro and then include that 128 * from C via asm(".include <asm/nospec-branch.h>") but let's not go there. 129 */ 130 131 #define RETPOLINE_THUNK_SIZE 32 132 #define RSB_CLEAR_LOOPS 32 /* To forcibly overwrite all entries */ 133 134 /* 135 * Common helper for __FILL_RETURN_BUFFER and __FILL_ONE_RETURN. 136 */ 137 #define __FILL_RETURN_SLOT \ 138 ANNOTATE_INTRA_FUNCTION_CALL; \ 139 call 772f; \ 140 int3; \ 141 772: 142 143 /* 144 * Stuff the entire RSB. 145 * 146 * Google experimented with loop-unrolling and this turned out to be 147 * the optimal version - two calls, each with their own speculation 148 * trap should their return address end up getting used, in a loop. 149 */ 150 #ifdef CONFIG_X86_64 151 #define __FILL_RETURN_BUFFER(reg, nr) \ 152 mov $(nr/2), reg; \ 153 771: \ 154 __FILL_RETURN_SLOT \ 155 __FILL_RETURN_SLOT \ 156 add $(BITS_PER_LONG/8) * 2, %_ASM_SP; \ 157 dec reg; \ 158 jnz 771b; \ 159 /* barrier for jnz misprediction */ \ 160 lfence; \ 161 ASM_CREDIT_CALL_DEPTH \ 162 CALL_THUNKS_DEBUG_INC_CTXSW 163 #else 164 /* 165 * i386 doesn't unconditionally have LFENCE, as such it can't 166 * do a loop. 167 */ 168 #define __FILL_RETURN_BUFFER(reg, nr) \ 169 .rept nr; \ 170 __FILL_RETURN_SLOT; \ 171 .endr; \ 172 add $(BITS_PER_LONG/8) * nr, %_ASM_SP; 173 #endif 174 175 /* 176 * Stuff a single RSB slot. 177 * 178 * To mitigate Post-Barrier RSB speculation, one CALL instruction must be 179 * forced to retire before letting a RET instruction execute. 180 * 181 * On PBRSB-vulnerable CPUs, it is not safe for a RET to be executed 182 * before this point. 183 */ 184 #define __FILL_ONE_RETURN \ 185 __FILL_RETURN_SLOT \ 186 add $(BITS_PER_LONG/8), %_ASM_SP; \ 187 lfence; 188 189 #ifdef __ASSEMBLY__ 190 191 /* 192 * This should be used immediately before an indirect jump/call. It tells 193 * objtool the subsequent indirect jump/call is vouched safe for retpoline 194 * builds. 195 */ 196 .macro ANNOTATE_RETPOLINE_SAFE 197 .Lhere_\@: 198 .pushsection .discard.retpoline_safe 199 .long .Lhere_\@ 200 .popsection 201 .endm 202 203 /* 204 * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions 205 * vs RETBleed validation. 206 */ 207 #define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE 208 209 /* 210 * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should 211 * eventually turn into it's own annotation. 212 */ 213 .macro VALIDATE_UNRET_END 214 #if defined(CONFIG_NOINSTR_VALIDATION) && \ 215 (defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO)) 216 ANNOTATE_RETPOLINE_SAFE 217 nop 218 #endif 219 .endm 220 221 /* 222 * Equivalent to -mindirect-branch-cs-prefix; emit the 5 byte jmp/call 223 * to the retpoline thunk with a CS prefix when the register requires 224 * a RAX prefix byte to encode. Also see apply_retpolines(). 225 */ 226 .macro __CS_PREFIX reg:req 227 .irp rs,r8,r9,r10,r11,r12,r13,r14,r15 228 .ifc \reg,\rs 229 .byte 0x2e 230 .endif 231 .endr 232 .endm 233 234 /* 235 * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple 236 * indirect jmp/call which may be susceptible to the Spectre variant 2 237 * attack. 238 * 239 * NOTE: these do not take kCFI into account and are thus not comparable to C 240 * indirect calls, take care when using. The target of these should be an ENDBR 241 * instruction irrespective of kCFI. 242 */ 243 .macro JMP_NOSPEC reg:req 244 #ifdef CONFIG_RETPOLINE 245 __CS_PREFIX \reg 246 jmp __x86_indirect_thunk_\reg 247 #else 248 jmp *%\reg 249 int3 250 #endif 251 .endm 252 253 .macro CALL_NOSPEC reg:req 254 #ifdef CONFIG_RETPOLINE 255 __CS_PREFIX \reg 256 call __x86_indirect_thunk_\reg 257 #else 258 call *%\reg 259 #endif 260 .endm 261 262 /* 263 * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP 264 * monstrosity above, manually. 265 */ 266 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req ftr2=ALT_NOT(X86_FEATURE_ALWAYS) 267 ALTERNATIVE_2 "jmp .Lskip_rsb_\@", \ 268 __stringify(__FILL_RETURN_BUFFER(\reg,\nr)), \ftr, \ 269 __stringify(nop;nop;__FILL_ONE_RETURN), \ftr2 270 271 .Lskip_rsb_\@: 272 .endm 273 274 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO) 275 #define CALL_UNTRAIN_RET "call entry_untrain_ret" 276 #else 277 #define CALL_UNTRAIN_RET "" 278 #endif 279 280 /* 281 * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the 282 * return thunk isn't mapped into the userspace tables (then again, AMD 283 * typically has NO_MELTDOWN). 284 * 285 * While retbleed_untrain_ret() doesn't clobber anything but requires stack, 286 * entry_ibpb() will clobber AX, CX, DX. 287 * 288 * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point 289 * where we have a stack but before any RET instruction. 290 */ 291 .macro UNTRAIN_RET 292 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \ 293 defined(CONFIG_CALL_DEPTH_TRACKING) || defined(CONFIG_CPU_SRSO) 294 VALIDATE_UNRET_END 295 ALTERNATIVE_3 "", \ 296 CALL_UNTRAIN_RET, X86_FEATURE_UNRET, \ 297 "call entry_ibpb", X86_FEATURE_ENTRY_IBPB, \ 298 __stringify(RESET_CALL_DEPTH), X86_FEATURE_CALL_DEPTH 299 #endif 300 .endm 301 302 .macro UNTRAIN_RET_VM 303 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \ 304 defined(CONFIG_CALL_DEPTH_TRACKING) || defined(CONFIG_CPU_SRSO) 305 VALIDATE_UNRET_END 306 ALTERNATIVE_3 "", \ 307 CALL_UNTRAIN_RET, X86_FEATURE_UNRET, \ 308 "call entry_ibpb", X86_FEATURE_IBPB_ON_VMEXIT, \ 309 __stringify(RESET_CALL_DEPTH), X86_FEATURE_CALL_DEPTH 310 #endif 311 .endm 312 313 .macro UNTRAIN_RET_FROM_CALL 314 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \ 315 defined(CONFIG_CALL_DEPTH_TRACKING) || defined(CONFIG_CPU_SRSO) 316 VALIDATE_UNRET_END 317 ALTERNATIVE_3 "", \ 318 CALL_UNTRAIN_RET, X86_FEATURE_UNRET, \ 319 "call entry_ibpb", X86_FEATURE_ENTRY_IBPB, \ 320 __stringify(RESET_CALL_DEPTH_FROM_CALL), X86_FEATURE_CALL_DEPTH 321 #endif 322 .endm 323 324 325 .macro CALL_DEPTH_ACCOUNT 326 #ifdef CONFIG_CALL_DEPTH_TRACKING 327 ALTERNATIVE "", \ 328 __stringify(ASM_INCREMENT_CALL_DEPTH), X86_FEATURE_CALL_DEPTH 329 #endif 330 .endm 331 332 #else /* __ASSEMBLY__ */ 333 334 #define ANNOTATE_RETPOLINE_SAFE \ 335 "999:\n\t" \ 336 ".pushsection .discard.retpoline_safe\n\t" \ 337 ".long 999b\n\t" \ 338 ".popsection\n\t" 339 340 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE]; 341 extern retpoline_thunk_t __x86_indirect_thunk_array[]; 342 extern retpoline_thunk_t __x86_indirect_call_thunk_array[]; 343 extern retpoline_thunk_t __x86_indirect_jump_thunk_array[]; 344 345 #ifdef CONFIG_RETHUNK 346 extern void __x86_return_thunk(void); 347 #else 348 static inline void __x86_return_thunk(void) {} 349 #endif 350 351 extern void retbleed_return_thunk(void); 352 extern void srso_return_thunk(void); 353 extern void srso_alias_return_thunk(void); 354 355 extern void retbleed_untrain_ret(void); 356 extern void srso_untrain_ret(void); 357 extern void srso_alias_untrain_ret(void); 358 359 extern void entry_untrain_ret(void); 360 extern void entry_ibpb(void); 361 362 extern void (*x86_return_thunk)(void); 363 364 #ifdef CONFIG_CALL_DEPTH_TRACKING 365 extern void __x86_return_skl(void); 366 367 static inline void x86_set_skl_return_thunk(void) 368 { 369 x86_return_thunk = &__x86_return_skl; 370 } 371 372 #define CALL_DEPTH_ACCOUNT \ 373 ALTERNATIVE("", \ 374 __stringify(INCREMENT_CALL_DEPTH), \ 375 X86_FEATURE_CALL_DEPTH) 376 377 #ifdef CONFIG_CALL_THUNKS_DEBUG 378 DECLARE_PER_CPU(u64, __x86_call_count); 379 DECLARE_PER_CPU(u64, __x86_ret_count); 380 DECLARE_PER_CPU(u64, __x86_stuffs_count); 381 DECLARE_PER_CPU(u64, __x86_ctxsw_count); 382 #endif 383 #else 384 static inline void x86_set_skl_return_thunk(void) {} 385 386 #define CALL_DEPTH_ACCOUNT "" 387 388 #endif 389 390 #ifdef CONFIG_RETPOLINE 391 392 #define GEN(reg) \ 393 extern retpoline_thunk_t __x86_indirect_thunk_ ## reg; 394 #include <asm/GEN-for-each-reg.h> 395 #undef GEN 396 397 #define GEN(reg) \ 398 extern retpoline_thunk_t __x86_indirect_call_thunk_ ## reg; 399 #include <asm/GEN-for-each-reg.h> 400 #undef GEN 401 402 #define GEN(reg) \ 403 extern retpoline_thunk_t __x86_indirect_jump_thunk_ ## reg; 404 #include <asm/GEN-for-each-reg.h> 405 #undef GEN 406 407 #ifdef CONFIG_X86_64 408 409 /* 410 * Inline asm uses the %V modifier which is only in newer GCC 411 * which is ensured when CONFIG_RETPOLINE is defined. 412 */ 413 # define CALL_NOSPEC \ 414 ALTERNATIVE_2( \ 415 ANNOTATE_RETPOLINE_SAFE \ 416 "call *%[thunk_target]\n", \ 417 "call __x86_indirect_thunk_%V[thunk_target]\n", \ 418 X86_FEATURE_RETPOLINE, \ 419 "lfence;\n" \ 420 ANNOTATE_RETPOLINE_SAFE \ 421 "call *%[thunk_target]\n", \ 422 X86_FEATURE_RETPOLINE_LFENCE) 423 424 # define THUNK_TARGET(addr) [thunk_target] "r" (addr) 425 426 #else /* CONFIG_X86_32 */ 427 /* 428 * For i386 we use the original ret-equivalent retpoline, because 429 * otherwise we'll run out of registers. We don't care about CET 430 * here, anyway. 431 */ 432 # define CALL_NOSPEC \ 433 ALTERNATIVE_2( \ 434 ANNOTATE_RETPOLINE_SAFE \ 435 "call *%[thunk_target]\n", \ 436 " jmp 904f;\n" \ 437 " .align 16\n" \ 438 "901: call 903f;\n" \ 439 "902: pause;\n" \ 440 " lfence;\n" \ 441 " jmp 902b;\n" \ 442 " .align 16\n" \ 443 "903: lea 4(%%esp), %%esp;\n" \ 444 " pushl %[thunk_target];\n" \ 445 " ret;\n" \ 446 " .align 16\n" \ 447 "904: call 901b;\n", \ 448 X86_FEATURE_RETPOLINE, \ 449 "lfence;\n" \ 450 ANNOTATE_RETPOLINE_SAFE \ 451 "call *%[thunk_target]\n", \ 452 X86_FEATURE_RETPOLINE_LFENCE) 453 454 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr) 455 #endif 456 #else /* No retpoline for C / inline asm */ 457 # define CALL_NOSPEC "call *%[thunk_target]\n" 458 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr) 459 #endif 460 461 /* The Spectre V2 mitigation variants */ 462 enum spectre_v2_mitigation { 463 SPECTRE_V2_NONE, 464 SPECTRE_V2_RETPOLINE, 465 SPECTRE_V2_LFENCE, 466 SPECTRE_V2_EIBRS, 467 SPECTRE_V2_EIBRS_RETPOLINE, 468 SPECTRE_V2_EIBRS_LFENCE, 469 SPECTRE_V2_IBRS, 470 }; 471 472 /* The indirect branch speculation control variants */ 473 enum spectre_v2_user_mitigation { 474 SPECTRE_V2_USER_NONE, 475 SPECTRE_V2_USER_STRICT, 476 SPECTRE_V2_USER_STRICT_PREFERRED, 477 SPECTRE_V2_USER_PRCTL, 478 SPECTRE_V2_USER_SECCOMP, 479 }; 480 481 /* The Speculative Store Bypass disable variants */ 482 enum ssb_mitigation { 483 SPEC_STORE_BYPASS_NONE, 484 SPEC_STORE_BYPASS_DISABLE, 485 SPEC_STORE_BYPASS_PRCTL, 486 SPEC_STORE_BYPASS_SECCOMP, 487 }; 488 489 static __always_inline 490 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature) 491 { 492 asm volatile(ALTERNATIVE("", "wrmsr", %c[feature]) 493 : : "c" (msr), 494 "a" ((u32)val), 495 "d" ((u32)(val >> 32)), 496 [feature] "i" (feature) 497 : "memory"); 498 } 499 500 extern u64 x86_pred_cmd; 501 502 static inline void indirect_branch_prediction_barrier(void) 503 { 504 alternative_msr_write(MSR_IA32_PRED_CMD, x86_pred_cmd, X86_FEATURE_USE_IBPB); 505 } 506 507 /* The Intel SPEC CTRL MSR base value cache */ 508 extern u64 x86_spec_ctrl_base; 509 DECLARE_PER_CPU(u64, x86_spec_ctrl_current); 510 extern void update_spec_ctrl_cond(u64 val); 511 extern u64 spec_ctrl_current(void); 512 513 /* 514 * With retpoline, we must use IBRS to restrict branch prediction 515 * before calling into firmware. 516 * 517 * (Implemented as CPP macros due to header hell.) 518 */ 519 #define firmware_restrict_branch_speculation_start() \ 520 do { \ 521 preempt_disable(); \ 522 alternative_msr_write(MSR_IA32_SPEC_CTRL, \ 523 spec_ctrl_current() | SPEC_CTRL_IBRS, \ 524 X86_FEATURE_USE_IBRS_FW); \ 525 alternative_msr_write(MSR_IA32_PRED_CMD, PRED_CMD_IBPB, \ 526 X86_FEATURE_USE_IBPB_FW); \ 527 } while (0) 528 529 #define firmware_restrict_branch_speculation_end() \ 530 do { \ 531 alternative_msr_write(MSR_IA32_SPEC_CTRL, \ 532 spec_ctrl_current(), \ 533 X86_FEATURE_USE_IBRS_FW); \ 534 preempt_enable(); \ 535 } while (0) 536 537 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp); 538 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb); 539 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb); 540 541 DECLARE_STATIC_KEY_FALSE(mds_user_clear); 542 DECLARE_STATIC_KEY_FALSE(mds_idle_clear); 543 544 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush); 545 546 DECLARE_STATIC_KEY_FALSE(mmio_stale_data_clear); 547 548 #include <asm/segment.h> 549 550 /** 551 * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability 552 * 553 * This uses the otherwise unused and obsolete VERW instruction in 554 * combination with microcode which triggers a CPU buffer flush when the 555 * instruction is executed. 556 */ 557 static __always_inline void mds_clear_cpu_buffers(void) 558 { 559 static const u16 ds = __KERNEL_DS; 560 561 /* 562 * Has to be the memory-operand variant because only that 563 * guarantees the CPU buffer flush functionality according to 564 * documentation. The register-operand variant does not. 565 * Works with any segment selector, but a valid writable 566 * data segment is the fastest variant. 567 * 568 * "cc" clobber is required because VERW modifies ZF. 569 */ 570 asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc"); 571 } 572 573 /** 574 * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability 575 * 576 * Clear CPU buffers if the corresponding static key is enabled 577 */ 578 static __always_inline void mds_user_clear_cpu_buffers(void) 579 { 580 if (static_branch_likely(&mds_user_clear)) 581 mds_clear_cpu_buffers(); 582 } 583 584 /** 585 * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability 586 * 587 * Clear CPU buffers if the corresponding static key is enabled 588 */ 589 static __always_inline void mds_idle_clear_cpu_buffers(void) 590 { 591 if (static_branch_likely(&mds_idle_clear)) 592 mds_clear_cpu_buffers(); 593 } 594 595 #endif /* __ASSEMBLY__ */ 596 597 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */ 598