~mdw / catacomb / blob
? search:


44c223da15a4e1614f8ddffcf4185bdc28542ec3
[catacomb] / base / asm-common.h
1 /// -*- mode: asm; asm-comment-char: ?/ -*-
2 ///
3 /// Common definitions for asesembler source files
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 #ifndef CATACOMB_ASM_COMMON_H
28 #define CATACOMB_ASM_COMMON_H
29
30 ///--------------------------------------------------------------------------
31 /// General definitions.
32
33 // Preprocessor hacks.
34 #define STRINGY(x) _STRINGY(x, y)
35 #define _STRINGY(x) #x
36 #define GLUE(x, y) _GLUE(x, y)
37 #define _GLUE(x, y) x##y
38 #define _EMPTY
39
40 // Some useful variables.
41 .L$_subsec = 0
42
43 // Literal pools done the hard way.
44 #define _LIT .text .L$_subsec + 1
45 #define _ENDLIT .text .L$_subsec
46 #define _LTORG .L$_subsec = .L$_subsec + 2; .text .L$_subsec
47
48 // ELF section types.
49 #if __ELF__
50 # if CPUFAM_ARMEL
51 # define _SECTTY(ty) %ty
52 # else
53 # define _SECTTY(ty) @ty
54 # endif
55 #endif
56
57 // Section selection.
58 #define TEXT .text .L$_subsec
59 #if ABI_WIN
60 # define RODATA .section .rdata, "dr"
61 #elif __ELF__
62 # define RODATA .section .rodata, "a", _SECTTY(progbits)
63 #else
64 # define RODATA TEXT
65 #endif
66 #define DATA .data
67
68 // Announcing an internal function.
69 #define INTFUNC(name) \
70 TYPE_FUNC(name); \
71 .macro ENDFUNC; _ENDFUNC(name); .endm; \
72 .L$_prologue_p = 0; .L$_frameptr_p = 0; \
73 FUNC_PREHOOK(name); \
74 name: \
75 FUNC_POSTHOOK(name)
76
77 // Announcing an external function.
78 #define FUNC(name) \
79 .globl F(name); \
80 INTFUNC(F(name))
81
82 // Marking the end of a function.
83 #define _ENDFUNC(name) \
84 .if ~ .L$_prologue_p; .error "Missing `endprologue'"; .endif; \
85 .if .L$_frameptr_p; .purgem dropfp; .endif; \
86 .purgem ENDFUNC; \
87 SIZE_OBJ(name); \
88 ENDFUNC_HOOK(name); \
89 _LTORG
90
91 // Make a helper function, if necessary.
92 #define AUXFN(name) \
93 .ifndef .L$_auxfn_def.name; \
94 .text 7128; \
95 .macro _ENDAUXFN; _ENDAUXFN_TAIL(name); .endm; \
96 FUNC_PREHOOK(name); \
97 name:
98 #define _ENDAUXFN_TAIL(name) \
99 .purgem _ENDAUXFN; \
100 .text .L$_subsec; \
101 .L$_auxfn_def.name = 1
102 #define ENDAUXFN _ENDAUXFN; .endif
103
104 ///--------------------------------------------------------------------------
105 /// ELF-specific hacking.
106
107 #if __ELF__
108
109 #if __PIC__ || __PIE__
110 # define WANT_PIC 1
111 #endif
112
113 #define TYPE_FUNC(name) .type name, STT_FUNC
114
115 #define SIZE_OBJ(name) .size name, . - name
116
117 #endif
118
119 ///--------------------------------------------------------------------------
120 /// Windows-specific hacking.
121
122 #if ABI_WIN
123
124 #if CPUFAM_X86
125 # define F(name) _##name
126 #endif
127
128 #endif
129
130 ///--------------------------------------------------------------------------
131 /// x86- and amd64-specific hacking.
132 ///
133 /// It's (slightly) easier to deal with both of these in one go.
134
135 #if CPUFAM_X86 || CPUFAM_AMD64
136
137 // Word size.
138 #if CPUFAM_X86
139 # define WORDSZ 4
140 #endif
141 #if CPUFAM_AMD64
142 # define WORDSZ 8
143 #endif
144
145 // Set the function hooks.
146 #define FUNC_PREHOOK(_) .balign 16
147
148 // On Windows, arrange to install stack-unwinding data.
149 #if CPUFAM_AMD64 && ABI_WIN
150 # define FUNC_POSTHOOK(name) .seh_proc name
151 # define ENDFUNC_HOOK(_) .seh_endproc
152 // Procedures are expected to invoke `.seh_setframe' if necessary, and
153 // `.seh_pushreg' and friends, and `.seh_endprologue'.
154 #endif
155
156 #if __ELF__
157 # define FUNC_POSTHOOK(_) .cfi_startproc
158 # define ENDFUNC_HOOK(_) .cfi_endproc
159 #endif
160
161 // Don't use the wretched AT&T syntax. It's festooned with pointless
162 // punctuation, and all of the data movement is backwards. Ugh!
163 .intel_syntax noprefix
164
165 // Call external subroutine at ADDR, possibly via PLT.
166 .macro callext addr
167 #if WANT_PIC
168 call \addr@PLT
169 #else
170 call \addr
171 #endif
172 .endm
173
174 // Do I need to arrange a spare GOT register?
175 #if WANT_PIC && CPUFAM_X86
176 # define NEED_GOT 1
177 #endif
178 #define GOTREG ebx // Not needed in AMD64 so don't care.
179
180 // Maybe load GOT address into GOT.
181 .macro ldgot got=GOTREG
182 #if WANT_PIC && CPUFAM_X86
183 AUXFN(_ldgot.\got)
184 mov \got, [esp]
185 ret
186 ENDAUXFN
187 call _ldgot.\got
188 add \got, offset _GLOBAL_OFFSET_TABLE_
189 #endif
190 .endm
191
192 // Load address of external symbol ADDR into REG, maybe using GOT.
193 .macro leaext reg, addr, got=GOTREG
194 #if WANT_PIC
195 # if CPUFAM_X86
196 mov \reg, [\got + \addr@GOT]
197 # endif
198 # if CPUFAM_AMD64
199 mov \reg, \addr@GOTPCREL[rip]
200 # endif
201 #else
202 # if CPUFAM_X86
203 mov \reg, offset \addr
204 # endif
205 # if CPUFAM_AMD64
206 lea \reg, \addr[rip]
207 # endif
208 #endif
209 .endm
210
211 // Address expression (possibly using a base register, and a displacement)
212 // referring to ADDR, which is within our module, maybe using GOT.
213 #define INTADDR(...) INTADDR__0(__VA_ARGS__, GOTREG, dummy)
214 #define INTADDR__0(addr, got, ...) INTADDR__1(addr, got)
215 #if CPUFAM_AMD64
216 # define INTADDR__1(addr, got) addr + rip
217 #elif WANT_PIC
218 # define INTADDR__1(addr, got) got + addr@GOTOFF
219 #else
220 # define INTADDR__1(addr, got) addr
221 #endif
222
223 // Permutations for SIMD instructions. SHUF(A, B, C, D) is an immediate,
224 // suitable for use in `pshufd' or `shufpd', which copies element A
225 // (0 <= A < 4) of the source to element 0 of the destination, element B to
226 // element 1, element C to element 2, and element D to element 3.
227 #define SHUF(a, b, c, d) ((a) + 4*(b) + 16*(c) + 64*(d))
228
229 // Map register names to their individual pieces.
230
231 // Apply decoration decor to (internal) register name reg of type ty.
232 //
233 // See `R_...' for internal register names. Decorations are as follows.
234 //
235 // b low byte (e.g., `al', `r8b')
236 // h high byte (e.g., `ah')
237 // w word (e.g., `ax', `r8w')
238 // d doubleword (e.g., `eax', `r8d')
239 // q quadword (e.g., `rax', `r8')
240 // r whole register (doubleword on x86, quadword on amd64)
241 //
242 // And types are as follows.
243 //
244 // abcd the four traditional registers `a', `b', `c', `d'
245 // xp the four pointer registers `si', `di', `bp', `sp'
246 // ip the instruction pointer `ip'
247 // rn the AMD64 numbered registers `r8'--`r15'
248 #define _DECOR(ty, decor, reg) _DECOR_##ty##_##decor(reg)
249
250 // Internal macros: _DECOR_ty_decor(reg) applies decoration decor to
251 // (internal) register name reg of type ty.
252
253 #define _DECOR_abcd_b(reg) reg##l
254 #define _DECOR_abcd_h(reg) reg##h
255 #define _DECOR_abcd_w(reg) reg##x
256 #define _DECOR_abcd_d(reg) e##reg##x
257 #if CPUFAM_AMD64
258 # define _DECOR_abcd_q(reg) r##reg##x
259 #endif
260
261 #define _DECOR_xp_w(reg) reg
262 #define _DECOR_xp_d(reg) e##reg
263 #if CPUFAM_AMD64
264 # define _DECOR_xp_b(reg) reg##l
265 # define _DECOR_xp_q(reg) r##reg
266 #endif
267
268 #define _DECOR_ip_w(reg) reg
269 #define _DECOR_ip_d(reg) e##reg
270 #if CPUFAM_AMD64
271 # define _DECOR_ip_q(reg) r##reg
272 #endif
273
274 #if CPUFAM_AMD64
275 # define _DECOR_rn_b(reg) reg##b
276 # define _DECOR_rn_w(reg) reg##w
277 # define _DECOR_rn_d(reg) reg##d
278 # define _DECOR_rn_q(reg) reg
279 # define _DECOR_rn_r(reg) reg
280 #endif
281
282 #define _DECOR_mem_b(addr) byte ptr addr
283 #define _DECOR_mem_w(addr) word ptr addr
284 #define _DECOR_mem_d(addr) dword ptr addr
285 #if CPUFAM_AMD64
286 # define _DECOR_mem_q(addr) qword ptr addr
287 #endif
288
289 #define _DECOR_imm_b(imm) byte imm
290 #define _DECOR_imm_w(imm) word imm
291 #define _DECOR_imm_d(imm) dword imm
292 #if CPUFAM_AMD64
293 # define _DECOR_imm_q(imm) qword imm
294 #endif
295
296 #if CPUFAM_X86
297 # define _DECOR_abcd_r(reg) e##reg##x
298 # define _DECOR_xp_r(reg) e##reg
299 # define _DECOR_ip_r(reg) e##reg
300 # define _DECOR_mem_r(addr) dword ptr addr
301 # define _DECOR_imm_r(imm) dword imm
302 #endif
303 #if CPUFAM_AMD64
304 # define _DECOR_abcd_r(reg) r##reg##x
305 # define _DECOR_xp_r(reg) r##reg
306 # define _DECOR_ip_r(reg) r##reg
307 # define _DECOR_mem_r(addr) qword ptr addr
308 # define _DECOR_imm_r(imm) qword imm
309 #endif
310
311 // R_r(decor) applies decoration decor to register r, which is an internal
312 // register name. The internal register names are: `ip', `a', `b', `c', `d',
313 // `si', `di', `bp', `sp', `r8'--`r15'.
314 #define R_ip(decor) _DECOR(ip, decor, ip)
315 #define R_a(decor) _DECOR(abcd, decor, a)
316 #define R_b(decor) _DECOR(abcd, decor, b)
317 #define R_c(decor) _DECOR(abcd, decor, c)
318 #define R_d(decor) _DECOR(abcd, decor, d)
319 #define R_si(decor) _DECOR(xp, decor, si)
320 #define R_di(decor) _DECOR(xp, decor, di)
321 #define R_bp(decor) _DECOR(xp, decor, bp)
322 #define R_sp(decor) _DECOR(xp, decor, sp)
323 #if CPUFAM_AMD64
324 # define R_r8(decor) _DECOR(rn, decor, r8)
325 # define R_r9(decor) _DECOR(rn, decor, r9)
326 # define R_r10(decor) _DECOR(rn, decor, r10)
327 # define R_r11(decor) _DECOR(rn, decor, r11)
328 # define R_r12(decor) _DECOR(rn, decor, r12)
329 # define R_r13(decor) _DECOR(rn, decor, r13)
330 # define R_r14(decor) _DECOR(rn, decor, r14)
331 # define R_r15(decor) _DECOR(rn, decor, r15)
332 #endif
333
334 // Refer to an in-memory datum of the type implied by decor residing at
335 // address addr (which should supply its own square-brackets).
336 #define MEM(decor, addr) _DECOR(mem, decor, addr)
337
338 // Refer to an immediate datum of the type implied by decor.
339 #define IMM(decor, imm) _DECOR(mem, decor, imm)
340
341 // Applies decoration decor to assembler-level register name reg.
342 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
343
344 // Internal macros: _REGFORM_r(decor) applies decoration decor to an
345 // assembler-level register name, in place of any decoration that register
346 // name has already.
347
348 #define _REGFORM_ip(decor) R_ip(decor)
349 #define _REGFORM_eip(decor) R_ip(decor)
350
351 #define _REGFORM_a(decor) R_a(decor)
352 #define _REGFORM_al(decor) R_a(decor)
353 #define _REGFORM_ah(decor) R_a(decor)
354 #define _REGFORM_ax(decor) R_a(decor)
355 #define _REGFORM_eax(decor) R_a(decor)
356
357 #define _REGFORM_b(decor) R_b(decor)
358 #define _REGFORM_bl(decor) R_b(decor)
359 #define _REGFORM_bh(decor) R_b(decor)
360 #define _REGFORM_bx(decor) R_b(decor)
361 #define _REGFORM_ebx(decor) R_b(decor)
362
363 #define _REGFORM_c(decor) R_c(decor)
364 #define _REGFORM_cl(decor) R_c(decor)
365 #define _REGFORM_ch(decor) R_c(decor)
366 #define _REGFORM_cx(decor) R_c(decor)
367 #define _REGFORM_ecx(decor) R_c(decor)
368
369 #define _REGFORM_d(decor) R_d(decor)
370 #define _REGFORM_dl(decor) R_d(decor)
371 #define _REGFORM_dh(decor) R_d(decor)
372 #define _REGFORM_dx(decor) R_d(decor)
373 #define _REGFORM_edx(decor) R_d(decor)
374
375 #define _REGFORM_si(decor) R_si(decor)
376 #define _REGFORM_sil(decor) R_si(decor)
377 #define _REGFORM_esi(decor) R_si(decor)
378
379 #define _REGFORM_di(decor) R_di(decor)
380 #define _REGFORM_dil(decor) R_di(decor)
381 #define _REGFORM_edi(decor) R_di(decor)
382
383 #define _REGFORM_bp(decor) R_bp(decor)
384 #define _REGFORM_bpl(decor) R_bp(decor)
385 #define _REGFORM_ebp(decor) R_bp(decor)
386
387 #define _REGFORM_sp(decor) R_sp(decor)
388 #define _REGFORM_spl(decor) R_sp(decor)
389 #define _REGFORM_esp(decor) R_sp(decor)
390
391 #if CPUFAM_AMD64
392
393 # define _REGFORM_rip(decor) R_ip(decor)
394 # define _REGFORM_rsp(decor) R_sp(decor)
395 # define _REGFORM_rbp(decor) R_bp(decor)
396 # define _REGFORM_rdi(decor) R_di(decor)
397 # define _REGFORM_rsi(decor) R_si(decor)
398 # define _REGFORM_rdx(decor) R_d(decor)
399 # define _REGFORM_rcx(decor) R_c(decor)
400 # define _REGFORM_rbx(decor) R_b(decor)
401 # define _REGFORM_rax(decor) R_a(decor)
402
403 # define _REGFORM_r8(decor) R_r8(decor)
404 # define _REGFORM_r8b(decor) R_r8(decor)
405 # define _REGFORM_r8w(decor) R_r8(decor)
406 # define _REGFORM_r8d(decor) R_r8(decor)
407
408 # define _REGFORM_r9(decor) R_r9(decor)
409 # define _REGFORM_r9b(decor) R_r9(decor)
410 # define _REGFORM_r9w(decor) R_r9(decor)
411 # define _REGFORM_r9d(decor) R_r9(decor)
412
413 # define _REGFORM_r10(decor) R_r10(decor)
414 # define _REGFORM_r10b(decor) R_r10(decor)
415 # define _REGFORM_r10w(decor) R_r10(decor)
416 # define _REGFORM_r10d(decor) R_r10(decor)
417
418 # define _REGFORM_r11(decor) R_r11(decor)
419 # define _REGFORM_r11b(decor) R_r11(decor)
420 # define _REGFORM_r11w(decor) R_r11(decor)
421 # define _REGFORM_r11d(decor) R_r11(decor)
422
423 # define _REGFORM_r12(decor) R_r12(decor)
424 # define _REGFORM_r12b(decor) R_r12(decor)
425 # define _REGFORM_r12w(decor) R_r12(decor)
426 # define _REGFORM_r12d(decor) R_r12(decor)
427
428 # define _REGFORM_r13(decor) R_r13(decor)
429 # define _REGFORM_r13b(decor) R_r13(decor)
430 # define _REGFORM_r13w(decor) R_r13(decor)
431 # define _REGFORM_r13d(decor) R_r13(decor)
432
433 # define _REGFORM_r14(decor) R_r14(decor)
434 # define _REGFORM_r14b(decor) R_r14(decor)
435 # define _REGFORM_r14w(decor) R_r14(decor)
436 # define _REGFORM_r14d(decor) R_r14(decor)
437
438 # define _REGFORM_r15(decor) R_r15(decor)
439 # define _REGFORM_r15b(decor) R_r15(decor)
440 # define _REGFORM_r15w(decor) R_r15(decor)
441 # define _REGFORM_r15d(decor) R_r15(decor)
442
443 #endif
444
445 // Macros for converting register names.
446 #define BYTE(reg) _REGFORM(reg, b)
447 #define HIBYTE(reg) _REGFORM(reg, h)
448 #define WORD(reg) _REGFORM(reg, w)
449 #define DWORD(reg) _REGFORM(reg, d)
450 #if CPUFAM_AMD64
451 # define QWORD(reg) _REGFORM(reg, q)
452 #endif
453 #define WHOLE(reg) _REGFORM(reg, r)
454
455 // Stack management and unwinding.
456 .macro setfp fp=R_bp(r), offset=0
457 .if \offset == 0
458 mov \fp, R_sp(r)
459 #if __ELF__
460 .cfi_def_cfa_register \fp
461 #endif
462 #if ABI_WIN && CPUFAM_AMD64
463 .seh_setframe \fp, 0
464 #endif
465 .else
466 lea \fp, [R_sp(r) + \offset]
467 #if __ELF__
468 .cfi_def_cfa_register \fp
469 .cfi_adjust_cfa_offset -\offset
470 #endif
471 #if ABI_WIN && CPUFAM_AMD64
472 .seh_setframe \fp, \offset
473 #endif
474 .endif
475 .L$_frameptr_p = -1
476 .macro dropfp; _dropfp \fp, \offset; .endm
477 .endm
478
479 .macro _dropfp fp, offset=0
480 .if \offset == 0
481 mov R_sp(r), \fp
482 #if __ELF__
483 .cfi_def_cfa_register R_sp(r)
484 #endif
485 .else
486 lea R_sp(r), [\fp - \offset]
487 #if __ELF__
488 .cfi_def_cfa_register R_sp(r)
489 .cfi_adjust_cfa_offset +\offset
490 #endif
491 .endif
492 .L$_frameptr_p = 0
493 .purgem dropfp
494 .endm
495
496 .macro stalloc n
497 sub R_sp(r), \n
498 #if __ELF__
499 .cfi_adjust_cfa_offset +\n
500 #endif
501 #if ABI_WIN && CPUFAM_AMD64
502 .seh_stackalloc \n
503 #endif
504 .endm
505
506 .macro stfree n
507 add R_sp(r), \n
508 #if __ELF__
509 .cfi_adjust_cfa_offset -\n
510 #endif
511 .endm
512
513 .macro pushreg r
514 push \r
515 #if __ELF__
516 .cfi_adjust_cfa_offset +WORDSZ
517 .cfi_rel_offset \r, 0
518 #endif
519 #if ABI_WIN && CPUFAM_AMD64
520 .seh_pushreg \r
521 #endif
522 .endm
523
524 .macro popreg r
525 pop \r
526 #if __ELF__
527 .cfi_adjust_cfa_offset -WORDSZ
528 .cfi_restore \r
529 #endif
530 .endm
531
532 .macro savexmm r, offset
533 movdqa [R_sp(r) + \offset], \r
534 #if ABI_WIN && CPUFAM_AMD64
535 .seh_savexmm \r, \offset
536 #endif
537 .endm
538
539 .macro rstrxmm r, offset
540 movdqa \r, [R_sp(r) + \offset]
541 .endm
542
543 .macro endprologue
544 #if ABI_WIN && CPUFAM_AMD64
545 .seh_endprologue
546 #endif
547 .L$_prologue_p = -1
548 .endm
549
550 #endif
551
552 ///--------------------------------------------------------------------------
553 /// ARM-specific hacking.
554
555 #if CPUFAM_ARMEL
556
557 // ARM/Thumb mode things. Use ARM by default.
558 #define ARM .arm; .L$_pcoff = 8
559 #define THUMB .thumb; .L$_pcoff = 4
560 ARM
561
562 // Set the function hooks.
563 #define FUNC_PREHOOK(_) .balign 4; .fnstart
564 #define ENDFUNC_HOOK(_) .fnend; .ltorg
565
566 // Call external subroutine at ADDR, possibly via PLT.
567 .macro callext addr, cond=
568 #if WANT_PIC
569 bl\cond \addr(PLT)
570 #else
571 bl\cond \addr
572 #endif
573 .endm
574
575 // Do I need to arrange a spare GOT register?
576 #if WANT_PIC
577 # define NEED_GOT 1
578 #endif
579 #define GOTREG r9
580
581 // Maybe load GOT address into GOT.
582 .macro ldgot cond=, got=GOTREG
583 #if WANT_PIC
584 ldr\cond \got, .L$_ldgot$\@
585 .L$_ldgot_pc$\@:
586 add\cond \got, pc, \got
587 _LIT
588 .balign 4
589 .L$_ldgot$\@:
590 .word _GLOBAL_OFFSET_TABLE_ - .L$_ldgot_pc$\@ - .L$_pcoff
591 _ENDLIT
592 #endif
593 .endm
594
595 // Load address of external symbol ADDR into REG, maybe using GOT.
596 .macro leaext reg, addr, cond=, got=GOTREG
597 #if WANT_PIC
598 ldr\cond \reg, .L$_leaext$\@
599 ldr\cond \reg, [\got, \reg]
600 _LIT
601 .balign 4
602 .L$_leaext$\@:
603 .word \addr(GOT)
604 _ENDLIT
605 #else
606 ldr\cond \reg, =\addr
607 #endif
608 .endm
609
610 // Load address of external symbol ADDR into REG directly.
611 .macro leaextq reg, addr, cond=
612 #if WANT_PIC
613 ldr\cond \reg, .L$_leaextq$\@
614 .L$_leaextq_pc$\@:
615 .if .L$_pcoff == 8
616 ldr\cond \reg, [pc, \reg]
617 .else
618 add\cond \reg, pc
619 ldr\cond \reg, [\reg]
620 .endif
621 _LIT
622 .balign 4
623 .L$_leaextq$\@:
624 .word \addr(GOT_PREL) + (. - .L$_leaextq_pc$\@ - .L$_pcoff)
625 _ENDLIT
626 #else
627 ldr\cond \reg, =\addr
628 #endif
629 .endm
630
631 .macro vzero vz=q15
632 // Set VZ (default q15) to zero.
633 vmov.u32 \vz, #0
634 .endm
635
636 .macro vshl128 vd, vn, nbit, vz=q15
637 // Set VD to VN shifted left by NBIT. Assume VZ (default q15) is
638 // all-bits-zero. NBIT must be a multiple of 8.
639 .if \nbit&3 != 0
640 .error "shift quantity must be whole number of bytes"
641 .endif
642 vext.8 \vd, \vz, \vn, #16 - (\nbit >> 3)
643 .endm
644
645 .macro vshr128 vd, vn, nbit, vz=q15
646 // Set VD to VN shifted right by NBIT. Assume VZ (default q15) is
647 // all-bits-zero. NBIT must be a multiple of 8.
648 .if \nbit&3 != 0
649 .error "shift quantity must be whole number of bytes"
650 .endif
651 vext.8 \vd, \vn, \vz, #\nbit >> 3
652 .endm
653
654 // Apply decoration decor to register name reg.
655 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
656
657 // Internal macros: `_REGFORM_r(decor)' applies decoration decor to register
658 // name r.
659
660 #define _REGFORM_s0(decor) _DECOR(s, decor, 0)
661 #define _REGFORM_s1(decor) _DECOR(s, decor, 1)
662 #define _REGFORM_s2(decor) _DECOR(s, decor, 2)
663 #define _REGFORM_s3(decor) _DECOR(s, decor, 3)
664 #define _REGFORM_s4(decor) _DECOR(s, decor, 4)
665 #define _REGFORM_s5(decor) _DECOR(s, decor, 5)
666 #define _REGFORM_s6(decor) _DECOR(s, decor, 6)
667 #define _REGFORM_s7(decor) _DECOR(s, decor, 7)
668 #define _REGFORM_s8(decor) _DECOR(s, decor, 8)
669 #define _REGFORM_s9(decor) _DECOR(s, decor, 9)
670 #define _REGFORM_s10(decor) _DECOR(s, decor, 10)
671 #define _REGFORM_s11(decor) _DECOR(s, decor, 11)
672 #define _REGFORM_s12(decor) _DECOR(s, decor, 12)
673 #define _REGFORM_s13(decor) _DECOR(s, decor, 13)
674 #define _REGFORM_s14(decor) _DECOR(s, decor, 14)
675 #define _REGFORM_s15(decor) _DECOR(s, decor, 15)
676 #define _REGFORM_s16(decor) _DECOR(s, decor, 16)
677 #define _REGFORM_s17(decor) _DECOR(s, decor, 17)
678 #define _REGFORM_s18(decor) _DECOR(s, decor, 18)
679 #define _REGFORM_s19(decor) _DECOR(s, decor, 19)
680 #define _REGFORM_s20(decor) _DECOR(s, decor, 20)
681 #define _REGFORM_s21(decor) _DECOR(s, decor, 21)
682 #define _REGFORM_s22(decor) _DECOR(s, decor, 22)
683 #define _REGFORM_s23(decor) _DECOR(s, decor, 23)
684 #define _REGFORM_s24(decor) _DECOR(s, decor, 24)
685 #define _REGFORM_s25(decor) _DECOR(s, decor, 25)
686 #define _REGFORM_s26(decor) _DECOR(s, decor, 26)
687 #define _REGFORM_s27(decor) _DECOR(s, decor, 27)
688 #define _REGFORM_s28(decor) _DECOR(s, decor, 28)
689 #define _REGFORM_s29(decor) _DECOR(s, decor, 29)
690 #define _REGFORM_s30(decor) _DECOR(s, decor, 30)
691 #define _REGFORM_s31(decor) _DECOR(s, decor, 31)
692
693 #define _REGFORM_d0(decor) _DECOR(d, decor, 0)
694 #define _REGFORM_d1(decor) _DECOR(d, decor, 1)
695 #define _REGFORM_d2(decor) _DECOR(d, decor, 2)
696 #define _REGFORM_d3(decor) _DECOR(d, decor, 3)
697 #define _REGFORM_d4(decor) _DECOR(d, decor, 4)
698 #define _REGFORM_d5(decor) _DECOR(d, decor, 5)
699 #define _REGFORM_d6(decor) _DECOR(d, decor, 6)
700 #define _REGFORM_d7(decor) _DECOR(d, decor, 7)
701 #define _REGFORM_d8(decor) _DECOR(d, decor, 8)
702 #define _REGFORM_d9(decor) _DECOR(d, decor, 9)
703 #define _REGFORM_d10(decor) _DECOR(d, decor, 10)
704 #define _REGFORM_d11(decor) _DECOR(d, decor, 11)
705 #define _REGFORM_d12(decor) _DECOR(d, decor, 12)
706 #define _REGFORM_d13(decor) _DECOR(d, decor, 13)
707 #define _REGFORM_d14(decor) _DECOR(d, decor, 14)
708 #define _REGFORM_d15(decor) _DECOR(d, decor, 15)
709 #define _REGFORM_d16(decor) _DECOR(d, decor, 16)
710 #define _REGFORM_d17(decor) _DECOR(d, decor, 17)
711 #define _REGFORM_d18(decor) _DECOR(d, decor, 18)
712 #define _REGFORM_d19(decor) _DECOR(d, decor, 19)
713 #define _REGFORM_d20(decor) _DECOR(d, decor, 20)
714 #define _REGFORM_d21(decor) _DECOR(d, decor, 21)
715 #define _REGFORM_d22(decor) _DECOR(d, decor, 22)
716 #define _REGFORM_d23(decor) _DECOR(d, decor, 23)
717 #define _REGFORM_d24(decor) _DECOR(d, decor, 24)
718 #define _REGFORM_d25(decor) _DECOR(d, decor, 25)
719 #define _REGFORM_d26(decor) _DECOR(d, decor, 26)
720 #define _REGFORM_d27(decor) _DECOR(d, decor, 27)
721 #define _REGFORM_d28(decor) _DECOR(d, decor, 28)
722 #define _REGFORM_d29(decor) _DECOR(d, decor, 29)
723 #define _REGFORM_d30(decor) _DECOR(d, decor, 30)
724 #define _REGFORM_d31(decor) _DECOR(d, decor, 31)
725
726 #define _REGFORM_q0(decor) _DECOR(q, decor, 0)
727 #define _REGFORM_q1(decor) _DECOR(q, decor, 1)
728 #define _REGFORM_q2(decor) _DECOR(q, decor, 2)
729 #define _REGFORM_q3(decor) _DECOR(q, decor, 3)
730 #define _REGFORM_q4(decor) _DECOR(q, decor, 4)
731 #define _REGFORM_q5(decor) _DECOR(q, decor, 5)
732 #define _REGFORM_q6(decor) _DECOR(q, decor, 6)
733 #define _REGFORM_q7(decor) _DECOR(q, decor, 7)
734 #define _REGFORM_q8(decor) _DECOR(q, decor, 8)
735 #define _REGFORM_q9(decor) _DECOR(q, decor, 9)
736 #define _REGFORM_q10(decor) _DECOR(q, decor, 10)
737 #define _REGFORM_q11(decor) _DECOR(q, decor, 11)
738 #define _REGFORM_q12(decor) _DECOR(q, decor, 12)
739 #define _REGFORM_q13(decor) _DECOR(q, decor, 13)
740 #define _REGFORM_q14(decor) _DECOR(q, decor, 14)
741 #define _REGFORM_q15(decor) _DECOR(q, decor, 15)
742
743 // `_LOPART(n)' and `_HIPART(n)' return the numbers of the register halves of
744 // register n, i.e., 2*n and 2*n + 1 respectively.
745 #define _LOPART(n) _GLUE(_LOPART_, n)
746 #define _HIPART(n) _GLUE(_HIPART_, n)
747
748 // Internal macros: `_LOPART_n' and `_HIPART_n' return the numbers of the
749 // register halves of register n, i.e., 2*n and 2*n + 1 respectively.
750
751 #define _LOPART_0 0
752 #define _HIPART_0 1
753 #define _LOPART_1 2
754 #define _HIPART_1 3
755 #define _LOPART_2 4
756 #define _HIPART_2 5
757 #define _LOPART_3 6
758 #define _HIPART_3 7
759 #define _LOPART_4 8
760 #define _HIPART_4 9
761 #define _LOPART_5 10
762 #define _HIPART_5 11
763 #define _LOPART_6 12
764 #define _HIPART_6 13
765 #define _LOPART_7 14
766 #define _HIPART_7 15
767 #define _LOPART_8 16
768 #define _HIPART_8 17
769 #define _LOPART_9 18
770 #define _HIPART_9 19
771 #define _LOPART_10 20
772 #define _HIPART_10 21
773 #define _LOPART_11 22
774 #define _HIPART_11 23
775 #define _LOPART_12 24
776 #define _HIPART_12 25
777 #define _LOPART_13 26
778 #define _HIPART_13 27
779 #define _LOPART_14 28
780 #define _HIPART_14 29
781 #define _LOPART_15 30
782 #define _HIPART_15 31
783
784 // Return the register number of the pair containing register n, i.e.,
785 // floor(n/2).
786 #define _PAIR(n) _GLUE(_PAIR_, n)
787
788 // Internal macros: `_PAIR_n' returns the register number of the pair
789 // containing register n, i.e., floor(n/2).
790 #define _PAIR_0 0
791 #define _PAIR_1 0
792 #define _PAIR_2 1
793 #define _PAIR_3 1
794 #define _PAIR_4 2
795 #define _PAIR_5 2
796 #define _PAIR_6 3
797 #define _PAIR_7 3
798 #define _PAIR_8 4
799 #define _PAIR_9 4
800 #define _PAIR_10 5
801 #define _PAIR_11 5
802 #define _PAIR_12 6
803 #define _PAIR_13 6
804 #define _PAIR_14 7
805 #define _PAIR_15 7
806 #define _PAIR_16 8
807 #define _PAIR_17 8
808 #define _PAIR_18 9
809 #define _PAIR_19 9
810 #define _PAIR_20 10
811 #define _PAIR_21 10
812 #define _PAIR_22 11
813 #define _PAIR_23 11
814 #define _PAIR_24 12
815 #define _PAIR_25 12
816 #define _PAIR_26 13
817 #define _PAIR_27 13
818 #define _PAIR_28 14
819 #define _PAIR_29 14
820 #define _PAIR_30 15
821 #define _PAIR_31 15
822
823 // Apply decoration decor to register number n of type ty. Decorations are
824 // as follows.
825 //
826 // decor types meaning
827 // Q s, d the NEON qN register containing this one
828 // D s the NEON dN register containing this one
829 // D0 q the low 64-bit half of this one
830 // D1 q the high 64-bit half of this one
831 // S0 d, q the first 32-bit piece of this one
832 // S1 d, q the second 32-bit piece of this one
833 // S2 q the third 32-bit piece of this one
834 // S3 q the fourth 32-bit piece of this one
835 // Bn q the nth byte of this register, as a scalar
836 // Hn q the nth halfword of this register, as a scalar
837 // Wn q the nth word of this register, as a scalar
838 #define _DECOR(ty, decor, n) _DECOR_##ty##_##decor(n)
839
840 // Internal macros: `_DECOR_ty_decor(n)' applies decoration decor to register
841 // number n of type ty.
842
843 #define _DECOR_s_Q(n) GLUE(q, _PAIR(_PAIR(n)))
844 #define _DECOR_s_D(n) GLUE(d, _PAIR(n))
845
846 #define _DECOR_d_Q(n) GLUE(q, _PAIR(n))
847 #define _DECOR_d_S0(n) GLUE(s, _LOPART(n))
848 #define _DECOR_d_S1(n) GLUE(s, _LOPART(n))
849
850 #define _DECOR_q_D0(n) GLUE(d, _LOPART(n))
851 #define _DECOR_q_D1(n) GLUE(d, _HIPART(n))
852 #define _DECOR_q_S0(n) GLUE(s, _LOPART(_LOPART(n)))
853 #define _DECOR_q_S1(n) GLUE(s, _HIPART(_LOPART(n)))
854 #define _DECOR_q_S2(n) GLUE(s, _LOPART(_HIPART(n)))
855 #define _DECOR_q_S3(n) GLUE(s, _HIPART(_HIPART(n)))
856 #define _DECOR_q_W0(n) GLUE(d, _LOPART(n))[0]
857 #define _DECOR_q_W1(n) GLUE(d, _LOPART(n))[1]
858 #define _DECOR_q_W2(n) GLUE(d, _HIPART(n))[0]
859 #define _DECOR_q_W3(n) GLUE(d, _HIPART(n))[1]
860 #define _DECOR_q_H0(n) GLUE(d, _LOPART(n))[0]
861 #define _DECOR_q_H1(n) GLUE(d, _LOPART(n))[1]
862 #define _DECOR_q_H2(n) GLUE(d, _LOPART(n))[2]
863 #define _DECOR_q_H3(n) GLUE(d, _LOPART(n))[3]
864 #define _DECOR_q_H4(n) GLUE(d, _HIPART(n))[0]
865 #define _DECOR_q_H5(n) GLUE(d, _HIPART(n))[1]
866 #define _DECOR_q_H6(n) GLUE(d, _HIPART(n))[2]
867 #define _DECOR_q_H7(n) GLUE(d, _HIPART(n))[3]
868 #define _DECOR_q_B0(n) GLUE(d, _LOPART(n))[0]
869 #define _DECOR_q_B1(n) GLUE(d, _LOPART(n))[1]
870 #define _DECOR_q_B2(n) GLUE(d, _LOPART(n))[2]
871 #define _DECOR_q_B3(n) GLUE(d, _LOPART(n))[3]
872 #define _DECOR_q_B4(n) GLUE(d, _LOPART(n))[4]
873 #define _DECOR_q_B5(n) GLUE(d, _LOPART(n))[5]
874 #define _DECOR_q_B6(n) GLUE(d, _LOPART(n))[6]
875 #define _DECOR_q_B7(n) GLUE(d, _LOPART(n))[7]
876 #define _DECOR_q_B8(n) GLUE(d, _HIPART(n))[0]
877 #define _DECOR_q_B9(n) GLUE(d, _HIPART(n))[1]
878 #define _DECOR_q_B10(n) GLUE(d, _HIPART(n))[2]
879 #define _DECOR_q_B11(n) GLUE(d, _HIPART(n))[3]
880 #define _DECOR_q_B12(n) GLUE(d, _HIPART(n))[4]
881 #define _DECOR_q_B13(n) GLUE(d, _HIPART(n))[5]
882 #define _DECOR_q_B14(n) GLUE(d, _HIPART(n))[6]
883 #define _DECOR_q_B15(n) GLUE(d, _HIPART(n))[7]
884
885 // Macros for navigating the NEON register hierarchy.
886 #define S0(reg) _REGFORM(reg, S0)
887 #define S1(reg) _REGFORM(reg, S1)
888 #define S2(reg) _REGFORM(reg, S2)
889 #define S3(reg) _REGFORM(reg, S3)
890 #define D(reg) _REGFORM(reg, D)
891 #define D0(reg) _REGFORM(reg, D0)
892 #define D1(reg) _REGFORM(reg, D1)
893 #define Q(reg) _REGFORM(reg, Q)
894
895 // Macros for indexing quadword registers.
896 #define QB(reg, i) _REGFORM(reg, B##i)
897 #define QH(reg, i) _REGFORM(reg, H##i)
898 #define QW(reg, i) _REGFORM(reg, W##i)
899
900 // Macros for converting vldm/vstm ranges.
901 #define QQ(qlo, qhi) D0(qlo)-D1(qhi)
902
903 // Stack management and unwinding.
904 .macro setfp fp=r11, offset=0
905 .if \offset == 0
906 mov \fp, sp
907 .setfp \fp, sp
908 .else
909 add \fp, sp, #\offset
910 .setfp \fp, sp, #\offset
911 .endif
912 .macro dropfp; _dropfp \fp, \offset; .endm
913 .L$_frameptr_p = -1
914 .endm
915
916 .macro _dropfp fp, offset=0
917 .if \offset == 0
918 mov sp, \fp
919 .else
920 sub sp, \fp, #\offset
921 .endif
922 .purgem dropfp
923 .L$_frameptr_p = 0
924 .endm
925
926 .macro stalloc n
927 sub sp, sp, #\n
928 .pad #\n
929 .endm
930
931 .macro stfree n
932 add sp, sp, #\n
933 .pad #-\n
934 .endm
935
936 .macro pushreg rr:vararg
937 push {\rr}
938 .save {\rr}
939 .endm
940
941 .macro popreg rr:vararg
942 pop {\rr}
943 .endm
944
945 .macro pushvfp rr:vararg
946 vstmdb sp!, {\rr}
947 .vsave {\rr}
948 .endm
949
950 .macro popvfp rr:vararg
951 vldmia sp!, {\rr}
952 .endm
953
954 .macro endprologue
955 .endm
956
957 // No need for prologue markers on ARM.
958 #define FUNC_POSTHOOK(_) .L$_prologue_p = -1
959
960 #endif
961
962 ///--------------------------------------------------------------------------
963 /// AArch64-specific hacking.
964
965 #if CPUFAM_ARM64
966
967 // Set the function hooks.
968 #define FUNC_PREHOOK(_) .balign 4
969 #define FUNC_POSTHOOK(_) .cfi_startproc; .L$_prologue_p = -1
970 #define ENDFUNC_HOOK(_) .cfi_endproc
971
972 // Call external subroutine at ADDR, possibly via PLT.
973 .macro callext addr
974 bl \addr
975 .endm
976
977 // Load address of external symbol ADDR into REG.
978 .macro leaext reg, addr
979 #if WANT_PIC
980 adrp \reg, :got:\addr
981 ldr \reg, [\reg, #:got_lo12:\addr]
982 #else
983 adrp \reg, \addr
984 add \reg, \reg, #:lo12:\addr
985 #endif
986 .endm
987
988 .macro vzero vz=v31
989 // Set VZ (default v31) to zero.
990 dup \vz\().4s, wzr
991 .endm
992
993 .macro vshl128 vd, vn, nbit, vz=v31
994 // Set VD to VN shifted left by NBIT. Assume VZ (default v31) is
995 // all-bits-zero. NBIT must be a multiple of 8.
996 .if \nbit&3 != 0
997 .error "shift quantity must be whole number of bytes"
998 .endif
999 ext \vd\().16b, \vz\().16b, \vn\().16b, #16 - (\nbit >> 3)
1000 .endm
1001
1002 .macro vshr128 vd, vn, nbit, vz=v31
1003 // Set VD to VN shifted right by NBIT. Assume VZ (default v31) is
1004 // all-bits-zero. NBIT must be a multiple of 8.
1005 .if \nbit&3 != 0
1006 .error "shift quantity must be whole number of bytes"
1007 .endif
1008 ext \vd\().16b, \vn\().16b, \vz\().16b, #\nbit >> 3
1009 .endm
1010
1011 // Stack management and unwinding.
1012 .macro setfp fp=x29, offset=0
1013 // If you're just going through the motions with a fixed-size stack frame,
1014 // then you want to say `add x29, sp, #OFFSET' directly, which will avoid
1015 // pointlessly restoring sp later.
1016 .if \offset == 0
1017 mov \fp, sp
1018 .cfi_def_cfa_register \fp
1019 .else
1020 add \fp, sp, #\offset
1021 .cfi_def_cfa_register \fp
1022 .cfi_adjust_cfa_offset -\offset
1023 .endif
1024 .macro dropfp; _dropfp \fp, \offset; .endm
1025 .L$_frameptr_p = -1
1026 .endm
1027
1028 .macro _dropfp fp, offset=0
1029 .if \offset == 0
1030 mov sp, \fp
1031 .cfi_def_cfa_register sp
1032 .else
1033 sub sp, \fp, #\offset
1034 .cfi_def_cfa_register sp
1035 .cfi_adjust_cfa_offset +\offset
1036 .endif
1037 .purgem dropfp
1038 .L$_frameptr_p = 0
1039 .endm
1040
1041 .macro stalloc n
1042 sub sp, sp, #\n
1043 .cfi_adjust_cfa_offset +\n
1044 .endm
1045
1046 .macro stfree n
1047 add sp, sp, #\n
1048 .cfi_adjust_cfa_offset -\n
1049 .endm
1050
1051 .macro pushreg x, y=nil
1052 .ifeqs "\y", "nil"
1053 str \x, [sp, #-16]!
1054 .cfi_adjust_cfa_offset +16
1055 .cfi_rel_offset \x, 0
1056 .else
1057 stp \x, \y, [sp, #-16]!
1058 .cfi_adjust_cfa_offset +16
1059 .cfi_rel_offset \x, 0
1060 .cfi_rel_offset \y, 8
1061 .endif
1062 .endm
1063
1064 .macro popreg x, y=nil
1065 .ifeqs "\y", "nil"
1066 ldr \x, [sp], #16
1067 .cfi_restore \x
1068 .cfi_adjust_cfa_offset -16
1069 .else
1070 ldp \x, \y, [sp], #16
1071 .cfi_restore \x
1072 .cfi_restore \y
1073 .cfi_adjust_cfa_offset -16
1074 .endif
1075 .endm
1076
1077 .macro savereg x, y, z=nil
1078 .ifeqs "\z", "nil"
1079 str \x, [sp, \y]
1080 .cfi_rel_offset \x, \y
1081 .else
1082 stp \x, \y, [sp, #\z]
1083 .cfi_rel_offset \x, \z
1084 .cfi_rel_offset \y, \z + 8
1085 .endif
1086 .endm
1087
1088 .macro rstrreg x, y, z=nil
1089 .ifeqs "\z", "nil"
1090 ldr \x, [sp, \y]
1091 .cfi_restore \x
1092 .else
1093 ldp \x, \y, [sp, #\z]
1094 .cfi_restore \x
1095 .cfi_restore \y
1096 .endif
1097 .endm
1098
1099 .macro endprologue
1100 .endm
1101
1102 #endif
1103
1104 ///--------------------------------------------------------------------------
1105 /// Final stuff.
1106
1107 // Default values for the various hooks.
1108 #ifndef FUNC_PREHOOK
1109 # define FUNC_PREHOOK(_)
1110 #endif
1111 #ifndef FUNC_POSTHOOK
1112 # define FUNC_POSTHOOK(_)
1113 #endif
1114 #ifndef ENDFUNC_HOOK
1115 # define ENDFUNC_HOOK(_)
1116 #endif
1117
1118 #ifndef F
1119 # ifdef SYM_USCORE
1120 # define F(name) _##name
1121 # else
1122 # define F(name) name
1123 # endif
1124 #endif
1125
1126 #ifndef TYPE_FUNC
1127 # define TYPE_FUNC(name)
1128 #endif
1129
1130 #ifndef SIZE_OBJ
1131 # define SIZE_OBJ(name)
1132 #endif
1133
1134 #if __ELF__ && !defined(WANT_EXECUTABLE_STACK)
1135 .pushsection .note.GNU-stack, "", _SECTTY(progbits)
1136 .popsection
1137 #endif
1138
1139 ///----- That's all, folks --------------------------------------------------
1140
1141 #endif
Catacomb cryptographic library