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 // Some useful variables.
  31         .L$_subsec = 0
  32
  33 // Literal pools done the hard way.
  34 #define _LIT .text .L$_subsec + 1
  35 #define _ENDLIT .text .L$_subsec
  36 #define _LTORG .L$_subsec = .L$_subsec + 2; .text .L$_subsec
  37
  38 // ELF section types.
  39 #if __ELF__
  40 #  if CPUFAM_ARMEL
  41 #    define _SECTTY(ty) %ty
  42 #  else
  43 #    define _SECTTY(ty) @ty
  44 #  endif
  45 #endif
  46
  47 // Section selection.
  48 #define TEXT .text .L$_subsec
  49 #if ABI_WIN
  50 #  define RODATA .section .rdata, "dr"
  51 #elif __ELF__
  52 #  define RODATA .section .rodata, "a", _SECTTY(progbits)
  53 #else
  54 #  define RODATA TEXT
  55 #endif
  56 #define DATA .data
  57
  58 // Announcing an external function.
  59 #define FUNC(name)                                                      \
  60         .globl  F(name);                                                \
  61         TYPE_FUNC(name);                                                \
  62         .macro ENDFUNC; _ENDFUNC(name); .endm;                          \
  63         FUNC_PREHOOK(name);                                             \
  64 F(name):                                                                \
  65         FUNC_POSTHOOK(name)
  66
  67 // Marking the end of a function.
  68 #define _ENDFUNC(name)                                                  \
  69         .purgem ENDFUNC;                                                \
  70         SIZE_OBJ(name);                                                 \
  71         ENDFUNC_HOOK(name);                                             \
  72         _LTORG
  73
  74 // Make a helper function, if necessary.
  75 #define AUXFN(name)                                                     \
  76   .ifndef .L$_auxfn_def.name;                                           \
  77         .text 7128;                                                     \
  78         .macro _ENDAUXFN; _ENDAUXFN_TAIL(name); .endm;                  \
  79         FUNC_PREHOOK(name);                                             \
  80 name:
  81 #define _ENDAUXFN_TAIL(name)                                            \
  82         .purgem _ENDAUXFN;                                              \
  83         .text .L$_subsec;                                               \
  84         .L$_auxfn_def.name = 1
  85 #define ENDAUXFN _ENDAUXFN; .endif
  86
  87 ///--------------------------------------------------------------------------
  88 /// ELF-specific hacking.
  89
  90 #if __ELF__
  91
  92 #if __PIC__ || __PIE__
  93 #  define WANT_PIC 1
  94 #endif
  95
  96 #define TYPE_FUNC(name) .type name, STT_FUNC
  97
  98 #define SIZE_OBJ(name) .size name, . - name
  99
 100 #endif
 101
 102 ///--------------------------------------------------------------------------
 103 /// Windows-specific hacking.
 104
 105 #if ABI_WIN
 106
 107 #if CPUFAM_X86
 108 #  define F(name) _##name
 109 #endif
 110
 111 #endif
 112
 113 ///--------------------------------------------------------------------------
 114 /// x86- and amd64-specific hacking.
 115 ///
 116 /// It's (slightly) easier to deal with both of these in one go.
 117
 118 #if CPUFAM_X86 || CPUFAM_AMD64
 119
 120 // Set the function hooks.
 121 #define FUNC_PREHOOK(_) .balign 16
 122
 123 // On Windows, arrange to install stack-unwinding data.
 124 #if CPUFAM_AMD64 && ABI_WIN
 125 #  define FUNC_POSTHOOK(name) .seh_proc name
 126 #  define ENDFUNC_HOOK(_) .seh_endproc
 127 // Procedures are expected to invoke `.seh_setframe' if necessary, and
 128 // `.seh_pushreg' and friends, and `.seh_endprologue'.
 129 #endif
 130
 131 // Don't use the wretched AT&T syntax.  It's festooned with pointless
 132 // punctuation, and all of the data movement is backwards.  Ugh!
 133         .intel_syntax noprefix
 134
 135 // Call external subroutine at ADDR, possibly via PLT.
 136 .macro  callext addr
 137 #if WANT_PIC
 138         call    \addr@PLT
 139 #else
 140         call    \addr
 141 #endif
 142 .endm
 143
 144 // Do I need to arrange a spare GOT register?
 145 #if WANT_PIC && CPUFAM_X86
 146 #  define NEED_GOT 1
 147 #endif
 148 #define GOTREG ebx                      // Not needed in AMD64 so don't care.
 149
 150 // Maybe load GOT address into GOT.
 151 .macro  ldgot got=GOTREG
 152 #if WANT_PIC && CPUFAM_X86
 153   AUXFN(_ldgot.\got)
 154         mov     \got, [esp]
 155         ret
 156   ENDAUXFN
 157         call    _ldgot.\got
 158         add     \got, offset _GLOBAL_OFFSET_TABLE_
 159 #endif
 160 .endm
 161
 162 // Load address of external symbol ADDR into REG, maybe using GOT.
 163 .macro  leaext reg, addr, got=GOTREG
 164 #if WANT_PIC
 165 #  if CPUFAM_X86
 166         mov     \reg, [\got + \addr@GOT]
 167 #  endif
 168 #  if CPUFAM_AMD64
 169         mov     \reg, \addr@GOTPCREL[rip]
 170 #  endif
 171 #else
 172 #  if CPUFAM_X86
 173         mov     \reg, offset \addr
 174 #  endif
 175 #  if CPUFAM_AMD64
 176         lea     \reg, \addr[rip]
 177 #  endif
 178 #endif
 179 .endm
 180
 181 // Address expression (possibly using a base register, and a displacement)
 182 // referring to ADDR, which is within our module, maybe using GOT.
 183 #define INTADDR(...) INTADDR__0(__VA_ARGS__, GOTREG, dummy)
 184 #define INTADDR__0(addr, got, ...) INTADDR__1(addr, got)
 185 #if CPUFAM_AMD64
 186 #  define INTADDR__1(addr, got) addr + rip
 187 #elif WANT_PIC
 188 #  define INTADDR__1(addr, got) got + addr@GOTOFF
 189 #else
 190 #  define INTADDR__1(addr, got) addr
 191 #endif
 192
 193 // Permutations for SIMD instructions.  SHUF(D, C, B, A) is an immediate,
 194 // suitable for use in `pshufd' or `shufpd', which copies element D
 195 // (0 <= D < 4) of the source to element 3 of the destination, element C to
 196 // element 2, element B to element 1, and element A to element 0.
 197 #define SHUF(d, c, b, a) (64*(d) + 16*(c) + 4*(b) + (a))
 198
 199 #endif
 200
 201 #if CPUFAM_X86
 202
 203 .macro  _reg.0
 204         // Stash GP registers and establish temporary stack frame.
 205         pushfd
 206         push    eax
 207         push    ecx
 208         push    edx
 209         push    ebp
 210         mov     ebp, esp
 211         and     esp, ~15
 212         sub     esp, 512
 213         fxsave  [esp]
 214 .endm
 215
 216 .macro  _reg.1
 217 .endm
 218
 219 .macro  _reg.2
 220 .endm
 221
 222 .macro  _reg.3  fmt
 223         // Print FMT and the other established arguments.
 224         lea     eax, .L$_reg$msg.\@
 225         push    eax
 226         call    printf
 227         jmp     .L$_reg$cont.\@
 228 .L$_reg$msg.\@:
 229         .ascii  ";; \fmt\n\0"
 230 .L$_reg$cont.\@:
 231         mov     eax, ebp
 232         and     eax, ~15
 233         sub     eax, 512
 234         fxrstor [eax]
 235         mov     esp, ebp
 236         pop     ebp
 237         pop     edx
 238         pop     ecx
 239         pop     eax
 240         popfd
 241 .endm
 242
 243 .macro  msg     msg
 244         _reg.0
 245         _reg.1
 246         _reg.2
 247         _reg.3  "\msg"
 248 .endm
 249
 250 .macro  reg     r, msg
 251         _reg.0
 252   .ifeqs "\r", "esp"
 253         lea     eax, [ebp + 20]
 254         push    eax
 255   .else
 256     .ifeqs "\r", "ebp"
 257         push    [ebp]
 258     .else
 259         push    \r
 260     .endif
 261   .endif
 262         _reg.1
 263         _reg.2
 264         _reg.3  "\msg: \r = %08x"
 265 .endm
 266
 267 .macro  xmmreg  r, msg
 268         _reg.0
 269         _reg.1
 270         _reg.2
 271         movdqu  xmm0, \r
 272         pshufd  xmm0, xmm0, 0x1b
 273         sub     esp, 16
 274         movdqa  [esp], xmm0
 275         _reg.3  "\msg: \r = %08x %08x %08x %08x"
 276 .endm
 277
 278 .macro  mmreg   r, msg
 279         _reg.0
 280         _reg.1
 281         _reg.2
 282         pshufw  \r, \r, 0x4e
 283         sub     esp, 8
 284         movq    [esp], \r
 285         _reg.3  "\msg: \r = %08x %08x"
 286 .endm
 287
 288 .macro  freg    i, msg
 289         _reg.0
 290         _reg.1
 291         _reg.2
 292         finit
 293         fldt    [esp + 32 + 16*\i]
 294         sub     esp, 12
 295         fstpt   [esp]
 296         _reg.3  "\msg: st(\i) = %.20Lg"
 297 .endm
 298
 299 .macro  fxreg   i, msg
 300         _reg.0
 301         _reg.1
 302         _reg.2
 303         finit
 304         fldt    [esp + 32 + 16*\i]
 305         sub     esp, 12
 306         fstpt   [esp]
 307         _reg.3  "\msg: st(\i) = %La"
 308 .endm
 309
 310 #endif
 311
 312 ///--------------------------------------------------------------------------
 313 /// ARM-specific hacking.
 314
 315 #if CPUFAM_ARMEL
 316
 317 // ARM/Thumb mode things.  Use ARM by default.
 318 #define ARM .arm; .L$_pcoff = 8
 319 #define THUMB .thumb; .L$_pcoff = 4
 320         ARM
 321
 322 // Set the function hooks.
 323 #define FUNC_PREHOOK(_) .balign 4
 324 #define ENDFUNC_HOOK(name) .ltorg
 325
 326 // Call external subroutine at ADDR, possibly via PLT.
 327 .macro  callext addr, cond=
 328 #if WANT_PIC
 329         bl\cond \addr(PLT)
 330 #else
 331         bl\cond \addr
 332 #endif
 333 .endm
 334
 335 // Do I need to arrange a spare GOT register?
 336 #if WANT_PIC
 337 #  define NEED_GOT 1
 338 #endif
 339 #define GOTREG r9
 340
 341 // Maybe load GOT address into GOT.
 342 .macro  ldgot   cond=, got=GOTREG
 343 #if WANT_PIC
 344         ldr\cond \got, .L$_ldgot$\@
 345 .L$_ldgot_pc$\@:
 346         add\cond \got, pc, \got
 347   _LIT
 348         .balign 4
 349 .L$_ldgot$\@:
 350         .word   _GLOBAL_OFFSET_TABLE_ - .L$_ldgot_pc$\@ - .L$_pcoff
 351   _ENDLIT
 352 #endif
 353 .endm
 354
 355 // Load address of external symbol ADDR into REG, maybe using GOT.
 356 .macro  leaext  reg, addr, cond=, got=GOTREG
 357 #if WANT_PIC
 358         ldr\cond \reg, .L$_leaext$\@
 359         ldr\cond \reg, [\got, \reg]
 360   _LIT
 361         .balign 4
 362 .L$_leaext$\@:
 363         .word   \addr(GOT)
 364   _ENDLIT
 365 #else
 366         ldr\cond \reg, =\addr
 367 #endif
 368 .endm
 369
 370 // Load address of external symbol ADDR into REG directly.
 371 .macro  leaextq reg, addr, cond=
 372 #if WANT_PIC
 373         ldr\cond \reg, .L$_leaextq$\@
 374 .L$_leaextq_pc$\@:
 375         .if     .L$_pcoff == 8
 376         ldr\cond \reg, [pc, \reg]
 377         .else
 378         add\cond \reg, pc
 379         ldr\cond \reg, [\reg]
 380         .endif
 381   _LIT
 382         .balign 4
 383 .L$_leaextq$\@:
 384         .word   \addr(GOT_PREL) + (. - .L$_leaextq_pc$\@ - .L$_pcoff)
 385   _ENDLIT
 386 #else
 387         ldr\cond \reg, =\addr
 388 #endif
 389 .endm
 390
 391 #endif
 392
 393 ///--------------------------------------------------------------------------
 394 /// Final stuff.
 395
 396 // Default values for the various hooks.
 397 #ifndef FUNC_PREHOOK
 398 #  define FUNC_PREHOOK(name)
 399 #endif
 400 #ifndef FUNC_POSTHOOK
 401 #  define FUNC_POSTHOOK(name)
 402 #endif
 403 #ifndef ENDFUNC_HOOK
 404 #  define ENDFUNC_HOOK(name)
 405 #endif
 406
 407 #ifndef F
 408 #  define F(name) name
 409 #endif
 410
 411 #ifndef TYPE_FUNC
 412 #  define TYPE_FUNC(name)
 413 #endif
 414
 415 #ifndef SIZE_OBJ
 416 #  define SIZE_OBJ(name)
 417 #endif
 418
 419 #if __ELF__ && defined(WANT_EXECUTABLE_STACK)
 420         .pushsection .note.GNU-stack, "", _SECTTY(progbits)
 421         .popsection
 422 #endif
 423
 424 ///----- That's all, folks --------------------------------------------------