mdw@git.distorted.org.uk Git - catacomb/blob - base/asm-common.h

   1 /// -*- mode: asm; asm-comment-char: ?/ -*-
   2 ///
   3 /// Fancy SIMD implementation of Salsa20
   4 ///
   5 /// (c) 2015 Straylight/Edgeware
   6 ///
   7
   8 ///----- Licensing notice ---------------------------------------------------
   9 ///
  10 /// This file is part of Catacomb.
  11 ///
  12 /// Catacomb is free software; you can redistribute it and/or modify
  13 /// it under the terms of the GNU Library General Public License as
  14 /// published by the Free Software Foundation; either version 2 of the
  15 /// License, or (at your option) any later version.
  16 ///
  17 /// Catacomb is distributed in the hope that it will be useful,
  18 /// but WITHOUT ANY WARRANTY; without even the implied warranty of
  19 /// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20 /// GNU Library General Public License for more details.
  21 ///
  22 /// You should have received a copy of the GNU Library General Public
  23 /// License along with Catacomb; if not, write to the Free
  24 /// Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  25 /// MA 02111-1307, USA.
  26
  27 ///--------------------------------------------------------------------------
  28 /// General definitions.
  29
  30 // Preprocessor hacks.
  31 #define STRINGY(x) _STRINGY(x, y)
  32 #define _STRINGY(x) #x
  33 #define GLUE(x, y) _GLUE(x, y)
  34 #define _GLUE(x, y) x##y
  35 #define _EMPTY
  36
  37 // Some useful variables.
  38         .L$_subsec = 0
  39
  40 // Literal pools done the hard way.
  41 #define _LIT .text .L$_subsec + 1
  42 #define _ENDLIT .text .L$_subsec
  43 #define _LTORG .L$_subsec = .L$_subsec + 2; .text .L$_subsec
  44
  45 // ELF section types.
  46 #if __ELF__
  47 #  if CPUFAM_ARMEL
  48 #    define _SECTTY(ty) %ty
  49 #  else
  50 #    define _SECTTY(ty) @ty
  51 #  endif
  52 #endif
  53
  54 // Section selection.
  55 #define TEXT .text .L$_subsec
  56 #if ABI_WIN
  57 #  define RODATA .section .rdata, "dr"
  58 #elif __ELF__
  59 #  define RODATA .section .rodata, "a", _SECTTY(progbits)
  60 #else
  61 #  define RODATA TEXT
  62 #endif
  63 #define DATA .data
  64
  65 // Announcing an external function.
  66 #define FUNC(name)                                                      \
  67         .globl  F(name);                                                \
  68         TYPE_FUNC(name);                                                \
  69         .macro ENDFUNC; _ENDFUNC(name); .endm;                          \
  70         FUNC_PREHOOK(name);                                             \
  71 F(name):                                                                \
  72         FUNC_POSTHOOK(name)
  73
  74 // Marking the end of a function.
  75 #define _ENDFUNC(name)                                                  \
  76         .purgem ENDFUNC;                                                \
  77         SIZE_OBJ(name);                                                 \
  78         ENDFUNC_HOOK(name);                                             \
  79         _LTORG
  80
  81 // Make a helper function, if necessary.
  82 #define AUXFN(name)                                                     \
  83   .ifndef .L$_auxfn_def.name;                                           \
  84         .text 7128;                                                     \
  85         .macro _ENDAUXFN; _ENDAUXFN_TAIL(name); .endm;                  \
  86         FUNC_PREHOOK(name);                                             \
  87 name:
  88 #define _ENDAUXFN_TAIL(name)                                            \
  89         .purgem _ENDAUXFN;                                              \
  90         .text .L$_subsec;                                               \
  91         .L$_auxfn_def.name = 1
  92 #define ENDAUXFN _ENDAUXFN; .endif
  93
  94 ///--------------------------------------------------------------------------
  95 /// ELF-specific hacking.
  96
  97 #if __ELF__
  98
  99 #if __PIC__ || __PIE__
 100 #  define WANT_PIC 1
 101 #endif
 102
 103 #define TYPE_FUNC(name) .type name, STT_FUNC
 104
 105 #define SIZE_OBJ(name) .size name, . - name
 106
 107 #endif
 108
 109 ///--------------------------------------------------------------------------
 110 /// Windows-specific hacking.
 111
 112 #if ABI_WIN
 113
 114 #if CPUFAM_X86
 115 #  define F(name) _##name
 116 #endif
 117
 118 #endif
 119
 120 ///--------------------------------------------------------------------------
 121 /// x86- and amd64-specific hacking.
 122 ///
 123 /// It's (slightly) easier to deal with both of these in one go.
 124
 125 #if CPUFAM_X86 || CPUFAM_AMD64
 126
 127 // Word size.
 128 #if CPUFAM_X86
 129 #  define WORDSZ 4
 130 #endif
 131 #if CPUFAM_AMD64
 132 #  define WORDSZ 8
 133 #endif
 134
 135 // Set the function hooks.
 136 #define FUNC_PREHOOK(_) .balign 16
 137
 138 // On Windows, arrange to install stack-unwinding data.
 139 #if CPUFAM_AMD64 && ABI_WIN
 140 #  define FUNC_POSTHOOK(name) .seh_proc name
 141 #  define ENDFUNC_HOOK(_) .seh_endproc
 142 // Procedures are expected to invoke `.seh_setframe' if necessary, and
 143 // `.seh_pushreg' and friends, and `.seh_endprologue'.
 144 #endif
 145
 146 // Don't use the wretched AT&T syntax.  It's festooned with pointless
 147 // punctuation, and all of the data movement is backwards.  Ugh!
 148         .intel_syntax noprefix
 149
 150 // Call external subroutine at ADDR, possibly via PLT.
 151 .macro  callext addr
 152 #if WANT_PIC
 153         call    \addr@PLT
 154 #else
 155         call    \addr
 156 #endif
 157 .endm
 158
 159 // Do I need to arrange a spare GOT register?
 160 #if WANT_PIC && CPUFAM_X86
 161 #  define NEED_GOT 1
 162 #endif
 163 #define GOTREG ebx                      // Not needed in AMD64 so don't care.
 164
 165 // Maybe load GOT address into GOT.
 166 .macro  ldgot got=GOTREG
 167 #if WANT_PIC && CPUFAM_X86
 168   AUXFN(_ldgot.\got)
 169         mov     \got, [esp]
 170         ret
 171   ENDAUXFN
 172         call    _ldgot.\got
 173         add     \got, offset _GLOBAL_OFFSET_TABLE_
 174 #endif
 175 .endm
 176
 177 // Load address of external symbol ADDR into REG, maybe using GOT.
 178 .macro  leaext reg, addr, got=GOTREG
 179 #if WANT_PIC
 180 #  if CPUFAM_X86
 181         mov     \reg, [\got + \addr@GOT]
 182 #  endif
 183 #  if CPUFAM_AMD64
 184         mov     \reg, \addr@GOTPCREL[rip]
 185 #  endif
 186 #else
 187 #  if CPUFAM_X86
 188         mov     \reg, offset \addr
 189 #  endif
 190 #  if CPUFAM_AMD64
 191         lea     \reg, \addr[rip]
 192 #  endif
 193 #endif
 194 .endm
 195
 196 // Address expression (possibly using a base register, and a displacement)
 197 // referring to ADDR, which is within our module, maybe using GOT.
 198 #define INTADDR(...) INTADDR__0(__VA_ARGS__, GOTREG, dummy)
 199 #define INTADDR__0(addr, got, ...) INTADDR__1(addr, got)
 200 #if CPUFAM_AMD64
 201 #  define INTADDR__1(addr, got) addr + rip
 202 #elif WANT_PIC
 203 #  define INTADDR__1(addr, got) got + addr@GOTOFF
 204 #else
 205 #  define INTADDR__1(addr, got) addr
 206 #endif
 207
 208 // Permutations for SIMD instructions.  SHUF(D, C, B, A) is an immediate,
 209 // suitable for use in `pshufd' or `shufpd', which copies element D
 210 // (0 <= D < 4) of the source to element 3 of the destination, element C to
 211 // element 2, element B to element 1, and element A to element 0.
 212 #define SHUF(d, c, b, a) (64*(d) + 16*(c) + 4*(b) + (a))
 213
 214 // Map register names to their individual pieces.
 215
 216 // Apply decoration decor to (internal) register name reg of type ty.
 217 //
 218 // See `R_...' for internal register names.  Decorations are as follows.
 219 //
 220 //      b       low byte (e.g., `al', `r8b')
 221 //      h       high byte (e.g., `ah')
 222 //      w       word (e.g., `ax', `r8w')
 223 //      d       doubleword (e.g., `eax', `r8d')
 224 //      q       quadword (e.g., `rax', `r8')
 225 //      r       whole register (doubleword on x86, quadword on amd64)
 226 //
 227 // And types are as follows.
 228 //
 229 //      abcd    the four traditional registers `a', `b', `c', `d'
 230 //      xp      the four pointer registers `si', `di', `bp', `sp'
 231 //      ip      the instruction pointer `ip'
 232 //      rn      the AMD64 numbered registers `r8'--`r15'
 233 #define _DECOR(ty, decor, reg) _DECOR_##ty##_##decor(reg)
 234
 235 // Internal macros: _DECOR_ty_decor(reg) applies decoration decor to
 236 // (internal) register name reg of type ty.
 237
 238 #define _DECOR_abcd_b(reg) reg##l
 239 #define _DECOR_abcd_h(reg) reg##h
 240 #define _DECOR_abcd_w(reg) reg##x
 241 #define _DECOR_abcd_d(reg) e##reg##x
 242 #if CPUFAM_AMD64
 243 #  define _DECOR_abcd_q(reg) r##reg##x
 244 #endif
 245
 246 #define _DECOR_xp_b(reg) reg##l
 247 #define _DECOR_xp_w(reg) reg
 248 #define _DECOR_xp_d(reg) e##reg
 249 #if CPUFAM_AMD64
 250 #  define _DECOR_xp_q(reg) r##reg
 251 #endif
 252
 253 #define _DECOR_ip_w(reg) reg
 254 #define _DECOR_ip_d(reg) e##reg
 255 #if CPUFAM_AMD64
 256 #  define _DECOR_ip_q(reg) r##reg
 257 #endif
 258
 259 #if CPUFAM_AMD64
 260 #  define _DECOR_rn_b(reg) reg##b
 261 #  define _DECOR_rn_w(reg) reg##w
 262 #  define _DECOR_rn_d(reg) reg##d
 263 #  define _DECOR_rn_q(reg) reg
 264 #  define _DECOR_rn_r(reg) reg
 265 #endif
 266
 267 #if CPUFAM_X86
 268 #  define _DECOR_abcd_r(reg) e##reg##x
 269 #  define _DECOR_xp_r(reg) e##reg
 270 #  define _DECOR_ip_r(reg) e##reg
 271 #endif
 272 #if CPUFAM_AMD64
 273 #  define _DECOR_abcd_r(reg) r##reg##x
 274 #  define _DECOR_xp_r(reg) r##reg
 275 #  define _DECOR_ip_r(reg) r##reg
 276 #endif
 277
 278 #define _DECOR_mem_b(addr) byte ptr addr
 279 #define _DECOR_mem_w(addr) word ptr addr
 280 #define _DECOR_mem_d(addr) dword ptr addr
 281 #if CPUFAM_AMD64
 282 #  define _DECOR_mem_q(addr) qword ptr addr
 283 #endif
 284
 285 // R_r(decor) applies decoration decor to register r, which is an internal
 286 // register name.  The internal register names are: `ip', `a', `b', `c', `d',
 287 // `si', `di', `bp', `sp', `r8'--`r15'.
 288 #define R_ip(decor) _DECOR(ip, decor, ip)
 289 #define R_a(decor) _DECOR(abcd, decor, a)
 290 #define R_b(decor) _DECOR(abcd, decor, b)
 291 #define R_c(decor) _DECOR(abcd, decor, c)
 292 #define R_d(decor) _DECOR(abcd, decor, d)
 293 #define R_si(decor) _DECOR(xp, decor, si)
 294 #define R_di(decor) _DECOR(xp, decor, di)
 295 #define R_bp(decor) _DECOR(xp, decor, bp)
 296 #define R_sp(decor) _DECOR(xp, decor, sp)
 297 #if CPUFAM_AMD64
 298 #  define R_r8(decor) _DECOR(rn, decor, r8)
 299 #  define R_r9(decor) _DECOR(rn, decor, r9)
 300 #  define R_r10(decor) _DECOR(rn, decor, r10)
 301 #  define R_r11(decor) _DECOR(rn, decor, r11)
 302 #  define R_r12(decor) _DECOR(rn, decor, r12)
 303 #  define R_r13(decor) _DECOR(rn, decor, r13)
 304 #  define R_r14(decor) _DECOR(rn, decor, r14)
 305 #  define R_r15(decor) _DECOR(rn, decor, r15)
 306 #endif
 307
 308 // Refer to an in-memory datum of the type implied by decor residing at
 309 // address addr (which should supply its own square-brackets).
 310 #define MEM(decor, addr) _DECOR(mem, decor, addr)
 311
 312 // Applies decoration decor to assembler-level register name reg.
 313 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
 314
 315 // Internal macros: _REGFORM_r(decor) applies decoration decor to an
 316 // assembler-level register name, in place of any decoration that register
 317 // name has already.
 318
 319 #define _REGFORM_ip(decor) R_ip(decor)
 320 #define _REGFORM_eip(decor) R_ip(decor)
 321
 322 #define _REGFORM_a(decor) R_a(decor)
 323 #define _REGFORM_al(decor) R_a(decor)
 324 #define _REGFORM_ah(decor) R_a(decor)
 325 #define _REGFORM_ax(decor) R_a(decor)
 326 #define _REGFORM_eax(decor) R_a(decor)
 327
 328 #define _REGFORM_b(decor) R_b(decor)
 329 #define _REGFORM_bl(decor) R_b(decor)
 330 #define _REGFORM_bh(decor) R_b(decor)
 331 #define _REGFORM_bx(decor) R_b(decor)
 332 #define _REGFORM_ebx(decor) R_b(decor)
 333
 334 #define _REGFORM_c(decor) R_c(decor)
 335 #define _REGFORM_cl(decor) R_c(decor)
 336 #define _REGFORM_ch(decor) R_c(decor)
 337 #define _REGFORM_cx(decor) R_c(decor)
 338 #define _REGFORM_ecx(decor) R_c(decor)
 339
 340 #define _REGFORM_d(decor) R_d(decor)
 341 #define _REGFORM_dl(decor) R_d(decor)
 342 #define _REGFORM_dh(decor) R_d(decor)
 343 #define _REGFORM_dx(decor) R_d(decor)
 344 #define _REGFORM_edx(decor) R_d(decor)
 345
 346 #define _REGFORM_si(decor) R_si(decor)
 347 #define _REGFORM_sil(decor) R_si(decor)
 348 #define _REGFORM_esi(decor) R_si(decor)
 349
 350 #define _REGFORM_di(decor) R_di(decor)
 351 #define _REGFORM_dil(decor) R_di(decor)
 352 #define _REGFORM_edi(decor) R_di(decor)
 353
 354 #define _REGFORM_bp(decor) R_bp(decor)
 355 #define _REGFORM_bpl(decor) R_bp(decor)
 356 #define _REGFORM_ebp(decor) R_bp(decor)
 357
 358 #define _REGFORM_sp(decor) R_sp(decor)
 359 #define _REGFORM_spl(decor) R_sp(decor)
 360 #define _REGFORM_esp(decor) R_sp(decor)
 361
 362 #if CPUFAM_AMD64
 363
 364 #  define _REGFORM_rip(decor) R_ip(decor)
 365 #  define _REGFORM_rsp(decor) R_sp(decor)
 366 #  define _REGFORM_rbp(decor) R_bp(decor)
 367 #  define _REGFORM_rdi(decor) R_di(decor)
 368 #  define _REGFORM_rsi(decor) R_si(decor)
 369 #  define _REGFORM_rdx(decor) R_d(decor)
 370 #  define _REGFORM_rcx(decor) R_c(decor)
 371 #  define _REGFORM_rbx(decor) R_b(decor)
 372 #  define _REGFORM_rax(decor) R_a(decor)
 373
 374 #  define _REGFORM_r8(decor) R_r8(decor)
 375 #  define _REGFORM_r8b(decor) R_r8(decor)
 376 #  define _REGFORM_r8w(decor) R_r8(decor)
 377 #  define _REGFORM_r8d(decor) R_r8(decor)
 378
 379 #  define _REGFORM_r9(decor) R_r9(decor)
 380 #  define _REGFORM_r9b(decor) R_r9(decor)
 381 #  define _REGFORM_r9w(decor) R_r9(decor)
 382 #  define _REGFORM_r9d(decor) R_r9(decor)
 383
 384 #  define _REGFORM_r10(decor) R_r10(decor)
 385 #  define _REGFORM_r10b(decor) R_r10(decor)
 386 #  define _REGFORM_r10w(decor) R_r10(decor)
 387 #  define _REGFORM_r10d(decor) R_r10(decor)
 388
 389 #  define _REGFORM_r11(decor) R_r11(decor)
 390 #  define _REGFORM_r11b(decor) R_r11(decor)
 391 #  define _REGFORM_r11w(decor) R_r11(decor)
 392 #  define _REGFORM_r11d(decor) R_r11(decor)
 393
 394 #  define _REGFORM_r12(decor) R_r12(decor)
 395 #  define _REGFORM_r12b(decor) R_r12(decor)
 396 #  define _REGFORM_r12w(decor) R_r12(decor)
 397 #  define _REGFORM_r12d(decor) R_r12(decor)
 398
 399 #  define _REGFORM_r13(decor) R_r13(decor)
 400 #  define _REGFORM_r13b(decor) R_r13(decor)
 401 #  define _REGFORM_r13w(decor) R_r13(decor)
 402 #  define _REGFORM_r13d(decor) R_r13(decor)
 403
 404 #  define _REGFORM_r14(decor) R_r14(decor)
 405 #  define _REGFORM_r14b(decor) R_r14(decor)
 406 #  define _REGFORM_r14w(decor) R_r14(decor)
 407 #  define _REGFORM_r14d(decor) R_r14(decor)
 408
 409 #  define _REGFORM_r15(decor) R_r15(decor)
 410 #  define _REGFORM_r15b(decor) R_r15(decor)
 411 #  define _REGFORM_r15w(decor) R_r15(decor)
 412 #  define _REGFORM_r15d(decor) R_r15(decor)
 413
 414 #endif
 415
 416 // Macros for converting register names.
 417 #define BYTE(reg) _REGFORM(reg, b)
 418 #define HIBYTE(reg) _REGFORM(reg, h)
 419 #define WORD(reg) _REGFORM(reg, w)
 420 #define DWORD(reg) _REGFORM(reg, d)
 421 #if CPUFAM_AMD64
 422 #  define QWORD(reg) _REGFORM(reg, q)
 423 #endif
 424 #define WHOLE(reg) _REGFORM(reg, r)
 425
 426 #endif
 427
 428 #if CPUFAM_X86
 429
 430 .macro  _reg.0
 431         // Stash GP registers and establish temporary stack frame.
 432         pushfd
 433         push    eax
 434         push    ecx
 435         push    edx
 436         push    ebp
 437         mov     ebp, esp
 438         and     esp, ~15
 439         sub     esp, 512
 440         fxsave  [esp]
 441 .endm
 442
 443 .macro  _reg.1
 444 .endm
 445
 446 .macro  _reg.2
 447 .endm
 448
 449 .macro  _reg.3  fmt
 450         // Print FMT and the other established arguments.
 451         lea     eax, .L$_reg$msg.\@
 452         push    eax
 453         call    printf
 454         jmp     .L$_reg$cont.\@
 455 .L$_reg$msg.\@:
 456         .ascii  ";; \fmt\n\0"
 457 .L$_reg$cont.\@:
 458         mov     eax, ebp
 459         and     eax, ~15
 460         sub     eax, 512
 461         fxrstor [eax]
 462         mov     esp, ebp
 463         pop     ebp
 464         pop     edx
 465         pop     ecx
 466         pop     eax
 467         popfd
 468 .endm
 469
 470 .macro  msg     msg
 471         _reg.0
 472         _reg.1
 473         _reg.2
 474         _reg.3  "\msg"
 475 .endm
 476
 477 .macro  reg     r, msg
 478         _reg.0
 479   .ifeqs "\r", "esp"
 480         lea     eax, [ebp + 20]
 481         push    eax
 482   .else
 483     .ifeqs "\r", "ebp"
 484         push    [ebp]
 485     .else
 486         push    \r
 487     .endif
 488   .endif
 489         _reg.1
 490         _reg.2
 491         _reg.3  "\msg: \r = %08x"
 492 .endm
 493
 494 .macro  xmmreg  r, msg
 495         _reg.0
 496         _reg.1
 497         _reg.2
 498         movdqu  xmm0, \r
 499         pshufd  xmm0, xmm0, 0x1b
 500         sub     esp, 16
 501         movdqa  [esp], xmm0
 502         _reg.3  "\msg: \r = %08x %08x %08x %08x"
 503 .endm
 504
 505 .macro  mmreg   r, msg
 506         _reg.0
 507         _reg.1
 508         _reg.2
 509         pshufw  \r, \r, 0x4e
 510         sub     esp, 8
 511         movq    [esp], \r
 512         _reg.3  "\msg: \r = %08x %08x"
 513 .endm
 514
 515 .macro  freg    i, msg
 516         _reg.0
 517         _reg.1
 518         _reg.2
 519         finit
 520         fldt    [esp + 32 + 16*\i]
 521         sub     esp, 12
 522         fstpt   [esp]
 523         _reg.3  "\msg: st(\i) = %.20Lg"
 524 .endm
 525
 526 .macro  fxreg   i, msg
 527         _reg.0
 528         _reg.1
 529         _reg.2
 530         finit
 531         fldt    [esp + 32 + 16*\i]
 532         sub     esp, 12
 533         fstpt   [esp]
 534         _reg.3  "\msg: st(\i) = %La"
 535 .endm
 536
 537 #endif
 538
 539 ///--------------------------------------------------------------------------
 540 /// ARM-specific hacking.
 541
 542 #if CPUFAM_ARMEL
 543
 544 // ARM/Thumb mode things.  Use ARM by default.
 545 #define ARM .arm; .L$_pcoff = 8
 546 #define THUMB .thumb; .L$_pcoff = 4
 547         ARM
 548
 549 // Set the function hooks.
 550 #define FUNC_PREHOOK(_) .balign 4
 551 #define ENDFUNC_HOOK(name) .ltorg
 552
 553 // Call external subroutine at ADDR, possibly via PLT.
 554 .macro  callext addr, cond=
 555 #if WANT_PIC
 556         bl\cond \addr(PLT)
 557 #else
 558         bl\cond \addr
 559 #endif
 560 .endm
 561
 562 // Do I need to arrange a spare GOT register?
 563 #if WANT_PIC
 564 #  define NEED_GOT 1
 565 #endif
 566 #define GOTREG r9
 567
 568 // Maybe load GOT address into GOT.
 569 .macro  ldgot   cond=, got=GOTREG
 570 #if WANT_PIC
 571         ldr\cond \got, .L$_ldgot$\@
 572 .L$_ldgot_pc$\@:
 573         add\cond \got, pc, \got
 574   _LIT
 575         .balign 4
 576 .L$_ldgot$\@:
 577         .word   _GLOBAL_OFFSET_TABLE_ - .L$_ldgot_pc$\@ - .L$_pcoff
 578   _ENDLIT
 579 #endif
 580 .endm
 581
 582 // Load address of external symbol ADDR into REG, maybe using GOT.
 583 .macro  leaext  reg, addr, cond=, got=GOTREG
 584 #if WANT_PIC
 585         ldr\cond \reg, .L$_leaext$\@
 586         ldr\cond \reg, [\got, \reg]
 587   _LIT
 588         .balign 4
 589 .L$_leaext$\@:
 590         .word   \addr(GOT)
 591   _ENDLIT
 592 #else
 593         ldr\cond \reg, =\addr
 594 #endif
 595 .endm
 596
 597 // Load address of external symbol ADDR into REG directly.
 598 .macro  leaextq reg, addr, cond=
 599 #if WANT_PIC
 600         ldr\cond \reg, .L$_leaextq$\@
 601 .L$_leaextq_pc$\@:
 602         .if     .L$_pcoff == 8
 603         ldr\cond \reg, [pc, \reg]
 604         .else
 605         add\cond \reg, pc
 606         ldr\cond \reg, [\reg]
 607         .endif
 608   _LIT
 609         .balign 4
 610 .L$_leaextq$\@:
 611         .word   \addr(GOT_PREL) + (. - .L$_leaextq_pc$\@ - .L$_pcoff)
 612   _ENDLIT
 613 #else
 614         ldr\cond \reg, =\addr
 615 #endif
 616 .endm
 617
 618 // Apply decoration decor to register name reg.
 619 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
 620
 621 // Internal macros: `_REGFORM_r(decor)' applies decoration decor to register
 622 // name r.
 623
 624 #define _REGFORM_s0(decor) _DECOR(s, decor, 0)
 625 #define _REGFORM_s1(decor) _DECOR(s, decor, 1)
 626 #define _REGFORM_s2(decor) _DECOR(s, decor, 2)
 627 #define _REGFORM_s3(decor) _DECOR(s, decor, 3)
 628 #define _REGFORM_s4(decor) _DECOR(s, decor, 4)
 629 #define _REGFORM_s5(decor) _DECOR(s, decor, 5)
 630 #define _REGFORM_s6(decor) _DECOR(s, decor, 6)
 631 #define _REGFORM_s7(decor) _DECOR(s, decor, 7)
 632 #define _REGFORM_s8(decor) _DECOR(s, decor, 8)
 633 #define _REGFORM_s9(decor) _DECOR(s, decor, 9)
 634 #define _REGFORM_s10(decor) _DECOR(s, decor, 10)
 635 #define _REGFORM_s11(decor) _DECOR(s, decor, 11)
 636 #define _REGFORM_s12(decor) _DECOR(s, decor, 12)
 637 #define _REGFORM_s13(decor) _DECOR(s, decor, 13)
 638 #define _REGFORM_s14(decor) _DECOR(s, decor, 14)
 639 #define _REGFORM_s15(decor) _DECOR(s, decor, 15)
 640 #define _REGFORM_s16(decor) _DECOR(s, decor, 16)
 641 #define _REGFORM_s17(decor) _DECOR(s, decor, 17)
 642 #define _REGFORM_s18(decor) _DECOR(s, decor, 18)
 643 #define _REGFORM_s19(decor) _DECOR(s, decor, 19)
 644 #define _REGFORM_s20(decor) _DECOR(s, decor, 20)
 645 #define _REGFORM_s21(decor) _DECOR(s, decor, 21)
 646 #define _REGFORM_s22(decor) _DECOR(s, decor, 22)
 647 #define _REGFORM_s23(decor) _DECOR(s, decor, 23)
 648 #define _REGFORM_s24(decor) _DECOR(s, decor, 24)
 649 #define _REGFORM_s25(decor) _DECOR(s, decor, 25)
 650 #define _REGFORM_s26(decor) _DECOR(s, decor, 26)
 651 #define _REGFORM_s27(decor) _DECOR(s, decor, 27)
 652 #define _REGFORM_s28(decor) _DECOR(s, decor, 28)
 653 #define _REGFORM_s29(decor) _DECOR(s, decor, 29)
 654 #define _REGFORM_s30(decor) _DECOR(s, decor, 30)
 655 #define _REGFORM_s31(decor) _DECOR(s, decor, 31)
 656
 657 #define _REGFORM_d0(decor) _DECOR(d, decor, 0)
 658 #define _REGFORM_d1(decor) _DECOR(d, decor, 1)
 659 #define _REGFORM_d2(decor) _DECOR(d, decor, 2)
 660 #define _REGFORM_d3(decor) _DECOR(d, decor, 3)
 661 #define _REGFORM_d4(decor) _DECOR(d, decor, 4)
 662 #define _REGFORM_d5(decor) _DECOR(d, decor, 5)
 663 #define _REGFORM_d6(decor) _DECOR(d, decor, 6)
 664 #define _REGFORM_d7(decor) _DECOR(d, decor, 7)
 665 #define _REGFORM_d8(decor) _DECOR(d, decor, 8)
 666 #define _REGFORM_d9(decor) _DECOR(d, decor, 9)
 667 #define _REGFORM_d10(decor) _DECOR(d, decor, 10)
 668 #define _REGFORM_d11(decor) _DECOR(d, decor, 11)
 669 #define _REGFORM_d12(decor) _DECOR(d, decor, 12)
 670 #define _REGFORM_d13(decor) _DECOR(d, decor, 13)
 671 #define _REGFORM_d14(decor) _DECOR(d, decor, 14)
 672 #define _REGFORM_d15(decor) _DECOR(d, decor, 15)
 673 #define _REGFORM_d16(decor) _DECOR(d, decor, 16)
 674 #define _REGFORM_d17(decor) _DECOR(d, decor, 17)
 675 #define _REGFORM_d18(decor) _DECOR(d, decor, 18)
 676 #define _REGFORM_d19(decor) _DECOR(d, decor, 19)
 677 #define _REGFORM_d20(decor) _DECOR(d, decor, 20)
 678 #define _REGFORM_d21(decor) _DECOR(d, decor, 21)
 679 #define _REGFORM_d22(decor) _DECOR(d, decor, 22)
 680 #define _REGFORM_d23(decor) _DECOR(d, decor, 23)
 681 #define _REGFORM_d24(decor) _DECOR(d, decor, 24)
 682 #define _REGFORM_d25(decor) _DECOR(d, decor, 25)
 683 #define _REGFORM_d26(decor) _DECOR(d, decor, 26)
 684 #define _REGFORM_d27(decor) _DECOR(d, decor, 27)
 685 #define _REGFORM_d28(decor) _DECOR(d, decor, 28)
 686 #define _REGFORM_d29(decor) _DECOR(d, decor, 29)
 687 #define _REGFORM_d30(decor) _DECOR(d, decor, 30)
 688 #define _REGFORM_d31(decor) _DECOR(d, decor, 31)
 689
 690 #define _REGFORM_q0(decor) _DECOR(q, decor, 0)
 691 #define _REGFORM_q1(decor) _DECOR(q, decor, 1)
 692 #define _REGFORM_q2(decor) _DECOR(q, decor, 2)
 693 #define _REGFORM_q3(decor) _DECOR(q, decor, 3)
 694 #define _REGFORM_q4(decor) _DECOR(q, decor, 4)
 695 #define _REGFORM_q5(decor) _DECOR(q, decor, 5)
 696 #define _REGFORM_q6(decor) _DECOR(q, decor, 6)
 697 #define _REGFORM_q7(decor) _DECOR(q, decor, 7)
 698 #define _REGFORM_q8(decor) _DECOR(q, decor, 8)
 699 #define _REGFORM_q9(decor) _DECOR(q, decor, 9)
 700 #define _REGFORM_q10(decor) _DECOR(q, decor, 10)
 701 #define _REGFORM_q11(decor) _DECOR(q, decor, 11)
 702 #define _REGFORM_q12(decor) _DECOR(q, decor, 12)
 703 #define _REGFORM_q13(decor) _DECOR(q, decor, 13)
 704 #define _REGFORM_q14(decor) _DECOR(q, decor, 14)
 705 #define _REGFORM_q15(decor) _DECOR(q, decor, 15)
 706
 707 // `_LOPART(n)' and `_HIPART(n)' return the numbers of the register halves of
 708 // register n, i.e., 2*n and 2*n + 1 respectively.
 709 #define _LOPART(n) _GLUE(_LOPART_, n)
 710 #define _HIPART(n) _GLUE(_HIPART_, n)
 711
 712 // Internal macros: `_LOPART_n' and `_HIPART_n' return the numbers of the
 713 // register halves of register n, i.e., 2*n and 2*n + 1 respectively.
 714
 715 #define _LOPART_0 0
 716 #define _HIPART_0 1
 717 #define _LOPART_1 2
 718 #define _HIPART_1 3
 719 #define _LOPART_2 4
 720 #define _HIPART_2 5
 721 #define _LOPART_3 6
 722 #define _HIPART_3 7
 723 #define _LOPART_4 8
 724 #define _HIPART_4 9
 725 #define _LOPART_5 10
 726 #define _HIPART_5 11
 727 #define _LOPART_6 12
 728 #define _HIPART_6 13
 729 #define _LOPART_7 14
 730 #define _HIPART_7 15
 731 #define _LOPART_8 16
 732 #define _HIPART_8 17
 733 #define _LOPART_9 18
 734 #define _HIPART_9 19
 735 #define _LOPART_10 20
 736 #define _HIPART_10 21
 737 #define _LOPART_11 22
 738 #define _HIPART_11 23
 739 #define _LOPART_12 24
 740 #define _HIPART_12 25
 741 #define _LOPART_13 26
 742 #define _HIPART_13 27
 743 #define _LOPART_14 28
 744 #define _HIPART_14 29
 745 #define _LOPART_15 30
 746 #define _HIPART_15 31
 747
 748 // Return the register number of the pair containing register n, i.e.,
 749 // floor(n/2).
 750 #define _PAIR(n) _GLUE(_PAIR_, n)
 751
 752 // Internal macros: `_PAIR_n' returns the register number of the pair
 753 // containing register n, i.e., floor(n/2).
 754 #define _PAIR_0 0
 755 #define _PAIR_1 0
 756 #define _PAIR_2 1
 757 #define _PAIR_3 1
 758 #define _PAIR_4 2
 759 #define _PAIR_5 2
 760 #define _PAIR_6 3
 761 #define _PAIR_7 3
 762 #define _PAIR_8 4
 763 #define _PAIR_9 4
 764 #define _PAIR_10 5
 765 #define _PAIR_11 5
 766 #define _PAIR_12 6
 767 #define _PAIR_13 6
 768 #define _PAIR_14 7
 769 #define _PAIR_15 7
 770 #define _PAIR_16 8
 771 #define _PAIR_17 8
 772 #define _PAIR_18 9
 773 #define _PAIR_19 9
 774 #define _PAIR_20 10
 775 #define _PAIR_21 10
 776 #define _PAIR_22 11
 777 #define _PAIR_23 11
 778 #define _PAIR_24 12
 779 #define _PAIR_25 12
 780 #define _PAIR_26 13
 781 #define _PAIR_27 13
 782 #define _PAIR_28 14
 783 #define _PAIR_29 14
 784 #define _PAIR_30 15
 785 #define _PAIR_31 15
 786
 787 // Apply decoration decor to register number n of type ty.  Decorations are
 788 // as follows.
 789 //
 790 //      decor   types   meaning
 791 //      Q       s, d    the NEON qN register containing this one
 792 //      D       s       the NEON dN register containing this one
 793 //      D0      q       the low 64-bit half of this one
 794 //      D1      q       the high 64-bit half of this one
 795 //      S0      d, q    the first 32-bit piece of this one
 796 //      S1      d, q    the second 32-bit piece of this one
 797 //      S2      q       the third 32-bit piece of this one
 798 //      S3      q       the fourth 32-bit piece of this one
 799 //      Bn      q       the nth byte of this register, as a scalar
 800 //      Hn      q       the nth halfword of this register, as a scalar
 801 //      Wn      q       the nth word of this register, as a scalar
 802 #define _DECOR(ty, decor, n) _DECOR_##ty##_##decor(n)
 803
 804 // Internal macros: `_DECOR_ty_decor(n)' applies decoration decor to register
 805 // number n of type ty.
 806
 807 #define _DECOR_s_Q(n) GLUE(q, _PAIR(_PAIR(n)))
 808 #define _DECOR_s_D(n) GLUE(d, _PAIR(n))
 809
 810 #define _DECOR_d_Q(n) GLUE(q, _PAIR(n))
 811 #define _DECOR_d_S0(n) GLUE(s, _LOPART(n))
 812 #define _DECOR_d_S1(n) GLUE(s, _LOPART(n))
 813
 814 #define _DECOR_q_D0(n) GLUE(d, _LOPART(n))
 815 #define _DECOR_q_D1(n) GLUE(d, _HIPART(n))
 816 #define _DECOR_q_S0(n) GLUE(s, _LOPART(_LOPART(n)))
 817 #define _DECOR_q_S1(n) GLUE(s, _HIPART(_LOPART(n)))
 818 #define _DECOR_q_S2(n) GLUE(s, _LOPART(_HIPART(n)))
 819 #define _DECOR_q_S3(n) GLUE(s, _HIPART(_HIPART(n)))
 820 #define _DECOR_q_W0(n) GLUE(d, _LOPART(n))[0]
 821 #define _DECOR_q_W1(n) GLUE(d, _LOPART(n))[1]
 822 #define _DECOR_q_W2(n) GLUE(d, _HIPART(n))[0]
 823 #define _DECOR_q_W3(n) GLUE(d, _HIPART(n))[1]
 824 #define _DECOR_q_H0(n) GLUE(d, _LOPART(n))[0]
 825 #define _DECOR_q_H1(n) GLUE(d, _LOPART(n))[1]
 826 #define _DECOR_q_H2(n) GLUE(d, _LOPART(n))[2]
 827 #define _DECOR_q_H3(n) GLUE(d, _LOPART(n))[3]
 828 #define _DECOR_q_H4(n) GLUE(d, _HIPART(n))[0]
 829 #define _DECOR_q_H5(n) GLUE(d, _HIPART(n))[1]
 830 #define _DECOR_q_H6(n) GLUE(d, _HIPART(n))[2]
 831 #define _DECOR_q_H7(n) GLUE(d, _HIPART(n))[3]
 832 #define _DECOR_q_B0(n) GLUE(d, _LOPART(n))[0]
 833 #define _DECOR_q_B1(n) GLUE(d, _LOPART(n))[1]
 834 #define _DECOR_q_B2(n) GLUE(d, _LOPART(n))[2]
 835 #define _DECOR_q_B3(n) GLUE(d, _LOPART(n))[3]
 836 #define _DECOR_q_B4(n) GLUE(d, _LOPART(n))[4]
 837 #define _DECOR_q_B5(n) GLUE(d, _LOPART(n))[5]
 838 #define _DECOR_q_B6(n) GLUE(d, _LOPART(n))[6]
 839 #define _DECOR_q_B7(n) GLUE(d, _LOPART(n))[7]
 840 #define _DECOR_q_B8(n) GLUE(d, _HIPART(n))[0]
 841 #define _DECOR_q_B9(n) GLUE(d, _HIPART(n))[1]
 842 #define _DECOR_q_B10(n) GLUE(d, _HIPART(n))[2]
 843 #define _DECOR_q_B11(n) GLUE(d, _HIPART(n))[3]
 844 #define _DECOR_q_B12(n) GLUE(d, _HIPART(n))[4]
 845 #define _DECOR_q_B13(n) GLUE(d, _HIPART(n))[5]
 846 #define _DECOR_q_B14(n) GLUE(d, _HIPART(n))[6]
 847 #define _DECOR_q_B15(n) GLUE(d, _HIPART(n))[7]
 848
 849 // Macros for navigating the NEON register hierarchy.
 850 #define S0(reg) _REGFORM(reg, S0)
 851 #define S1(reg) _REGFORM(reg, S1)
 852 #define S2(reg) _REGFORM(reg, S2)
 853 #define S3(reg) _REGFORM(reg, S3)
 854 #define D(reg) _REGFORM(reg, D)
 855 #define D0(reg) _REGFORM(reg, D0)
 856 #define D1(reg) _REGFORM(reg, D1)
 857 #define Q(reg) _REGFORM(reg, Q)
 858
 859 // Macros for indexing quadword registers.
 860 #define QB(reg, i) _REGFORM(reg, B##i)
 861 #define QH(reg, i) _REGFORM(reg, H##i)
 862 #define QW(reg, i) _REGFORM(reg, W##i)
 863
 864 // Macros for converting vldm/vstm ranges.
 865 #define QQ(qlo, qhi) D0(qlo)-D1(qhi)
 866
 867 #endif
 868
 869 ///--------------------------------------------------------------------------
 870 /// Final stuff.
 871
 872 // Default values for the various hooks.
 873 #ifndef FUNC_PREHOOK
 874 #  define FUNC_PREHOOK(_)
 875 #endif
 876 #ifndef FUNC_POSTHOOK
 877 #  define FUNC_POSTHOOK(_)
 878 #endif
 879 #ifndef ENDFUNC_HOOK
 880 #  define ENDFUNC_HOOK(_)
 881 #endif
 882
 883 #ifndef F
 884 #  define F(name) name
 885 #endif
 886
 887 #ifndef TYPE_FUNC
 888 #  define TYPE_FUNC(name)
 889 #endif
 890
 891 #ifndef SIZE_OBJ
 892 #  define SIZE_OBJ(name)
 893 #endif
 894
 895 #if __ELF__ && defined(WANT_EXECUTABLE_STACK)
 896         .pushsection .note.GNU-stack, "", _SECTTY(progbits)
 897         .popsection
 898 #endif
 899
 900 ///----- That's all, folks --------------------------------------------------