~mdw / catacomb / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
base/asm-common.h: Implement the `r' decorator for `MEM' accesses.
[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 #if CPUFAM_X86
290 # define _DECOR_abcd_r(reg) e##reg##x
291 # define _DECOR_xp_r(reg) e##reg
292 # define _DECOR_ip_r(reg) e##reg
293 # define _DECOR_mem_r(addr) dword ptr addr
294 #endif
295 #if CPUFAM_AMD64
296 # define _DECOR_abcd_r(reg) r##reg##x
297 # define _DECOR_xp_r(reg) r##reg
298 # define _DECOR_ip_r(reg) r##reg
299 # define _DECOR_mem_r(addr) qword ptr addr
300 #endif
301
302 // R_r(decor) applies decoration decor to register r, which is an internal
303 // register name. The internal register names are: `ip', `a', `b', `c', `d',
304 // `si', `di', `bp', `sp', `r8'--`r15'.
305 #define R_ip(decor) _DECOR(ip, decor, ip)
306 #define R_a(decor) _DECOR(abcd, decor, a)
307 #define R_b(decor) _DECOR(abcd, decor, b)
308 #define R_c(decor) _DECOR(abcd, decor, c)
309 #define R_d(decor) _DECOR(abcd, decor, d)
310 #define R_si(decor) _DECOR(xp, decor, si)
311 #define R_di(decor) _DECOR(xp, decor, di)
312 #define R_bp(decor) _DECOR(xp, decor, bp)
313 #define R_sp(decor) _DECOR(xp, decor, sp)
314 #if CPUFAM_AMD64
315 # define R_r8(decor) _DECOR(rn, decor, r8)
316 # define R_r9(decor) _DECOR(rn, decor, r9)
317 # define R_r10(decor) _DECOR(rn, decor, r10)
318 # define R_r11(decor) _DECOR(rn, decor, r11)
319 # define R_r12(decor) _DECOR(rn, decor, r12)
320 # define R_r13(decor) _DECOR(rn, decor, r13)
321 # define R_r14(decor) _DECOR(rn, decor, r14)
322 # define R_r15(decor) _DECOR(rn, decor, r15)
323 #endif
324
325 // Refer to an in-memory datum of the type implied by decor residing at
326 // address addr (which should supply its own square-brackets).
327 #define MEM(decor, addr) _DECOR(mem, decor, addr)
328
329 // Applies decoration decor to assembler-level register name reg.
330 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
331
332 // Internal macros: _REGFORM_r(decor) applies decoration decor to an
333 // assembler-level register name, in place of any decoration that register
334 // name has already.
335
336 #define _REGFORM_ip(decor) R_ip(decor)
337 #define _REGFORM_eip(decor) R_ip(decor)
338
339 #define _REGFORM_a(decor) R_a(decor)
340 #define _REGFORM_al(decor) R_a(decor)
341 #define _REGFORM_ah(decor) R_a(decor)
342 #define _REGFORM_ax(decor) R_a(decor)
343 #define _REGFORM_eax(decor) R_a(decor)
344
345 #define _REGFORM_b(decor) R_b(decor)
346 #define _REGFORM_bl(decor) R_b(decor)
347 #define _REGFORM_bh(decor) R_b(decor)
348 #define _REGFORM_bx(decor) R_b(decor)
349 #define _REGFORM_ebx(decor) R_b(decor)
350
351 #define _REGFORM_c(decor) R_c(decor)
352 #define _REGFORM_cl(decor) R_c(decor)
353 #define _REGFORM_ch(decor) R_c(decor)
354 #define _REGFORM_cx(decor) R_c(decor)
355 #define _REGFORM_ecx(decor) R_c(decor)
356
357 #define _REGFORM_d(decor) R_d(decor)
358 #define _REGFORM_dl(decor) R_d(decor)
359 #define _REGFORM_dh(decor) R_d(decor)
360 #define _REGFORM_dx(decor) R_d(decor)
361 #define _REGFORM_edx(decor) R_d(decor)
362
363 #define _REGFORM_si(decor) R_si(decor)
364 #define _REGFORM_sil(decor) R_si(decor)
365 #define _REGFORM_esi(decor) R_si(decor)
366
367 #define _REGFORM_di(decor) R_di(decor)
368 #define _REGFORM_dil(decor) R_di(decor)
369 #define _REGFORM_edi(decor) R_di(decor)
370
371 #define _REGFORM_bp(decor) R_bp(decor)
372 #define _REGFORM_bpl(decor) R_bp(decor)
373 #define _REGFORM_ebp(decor) R_bp(decor)
374
375 #define _REGFORM_sp(decor) R_sp(decor)
376 #define _REGFORM_spl(decor) R_sp(decor)
377 #define _REGFORM_esp(decor) R_sp(decor)
378
379 #if CPUFAM_AMD64
380
381 # define _REGFORM_rip(decor) R_ip(decor)
382 # define _REGFORM_rsp(decor) R_sp(decor)
383 # define _REGFORM_rbp(decor) R_bp(decor)
384 # define _REGFORM_rdi(decor) R_di(decor)
385 # define _REGFORM_rsi(decor) R_si(decor)
386 # define _REGFORM_rdx(decor) R_d(decor)
387 # define _REGFORM_rcx(decor) R_c(decor)
388 # define _REGFORM_rbx(decor) R_b(decor)
389 # define _REGFORM_rax(decor) R_a(decor)
390
391 # define _REGFORM_r8(decor) R_r8(decor)
392 # define _REGFORM_r8b(decor) R_r8(decor)
393 # define _REGFORM_r8w(decor) R_r8(decor)
394 # define _REGFORM_r8d(decor) R_r8(decor)
395
396 # define _REGFORM_r9(decor) R_r9(decor)
397 # define _REGFORM_r9b(decor) R_r9(decor)
398 # define _REGFORM_r9w(decor) R_r9(decor)
399 # define _REGFORM_r9d(decor) R_r9(decor)
400
401 # define _REGFORM_r10(decor) R_r10(decor)
402 # define _REGFORM_r10b(decor) R_r10(decor)
403 # define _REGFORM_r10w(decor) R_r10(decor)
404 # define _REGFORM_r10d(decor) R_r10(decor)
405
406 # define _REGFORM_r11(decor) R_r11(decor)
407 # define _REGFORM_r11b(decor) R_r11(decor)
408 # define _REGFORM_r11w(decor) R_r11(decor)
409 # define _REGFORM_r11d(decor) R_r11(decor)
410
411 # define _REGFORM_r12(decor) R_r12(decor)
412 # define _REGFORM_r12b(decor) R_r12(decor)
413 # define _REGFORM_r12w(decor) R_r12(decor)
414 # define _REGFORM_r12d(decor) R_r12(decor)
415
416 # define _REGFORM_r13(decor) R_r13(decor)
417 # define _REGFORM_r13b(decor) R_r13(decor)
418 # define _REGFORM_r13w(decor) R_r13(decor)
419 # define _REGFORM_r13d(decor) R_r13(decor)
420
421 # define _REGFORM_r14(decor) R_r14(decor)
422 # define _REGFORM_r14b(decor) R_r14(decor)
423 # define _REGFORM_r14w(decor) R_r14(decor)
424 # define _REGFORM_r14d(decor) R_r14(decor)
425
426 # define _REGFORM_r15(decor) R_r15(decor)
427 # define _REGFORM_r15b(decor) R_r15(decor)
428 # define _REGFORM_r15w(decor) R_r15(decor)
429 # define _REGFORM_r15d(decor) R_r15(decor)
430
431 #endif
432
433 // Macros for converting register names.
434 #define BYTE(reg) _REGFORM(reg, b)
435 #define HIBYTE(reg) _REGFORM(reg, h)
436 #define WORD(reg) _REGFORM(reg, w)
437 #define DWORD(reg) _REGFORM(reg, d)
438 #if CPUFAM_AMD64
439 # define QWORD(reg) _REGFORM(reg, q)
440 #endif
441 #define WHOLE(reg) _REGFORM(reg, r)
442
443 // Stack management and unwinding.
444 .macro setfp fp, offset = 0
445 .if \offset == 0
446 mov \fp, R_sp(r)
447 #if __ELF__
448 .cfi_def_cfa_register \fp
449 #endif
450 #if ABI_WIN && CPUFAM_AMD64
451 .seh_setframe \fp, 0
452 #endif
453 .else
454 lea \fp, [R_sp(r) + \offset]
455 #if __ELF__
456 .cfi_def_cfa_register \fp
457 .cfi_adjust_cfa_offset -\offset
458 #endif
459 #if ABI_WIN && CPUFAM_AMD64
460 .seh_setframe \fp, \offset
461 #endif
462 .endif
463 .L$_frameptr_p = -1
464 .macro dropfp; _dropfp \fp, \offset; .endm
465 .endm
466
467 .macro _dropfp fp, offset = 0
468 .if \offset == 0
469 mov R_sp(r), \fp
470 #if __ELF__
471 .cfi_def_cfa_register R_sp(r)
472 #endif
473 .else
474 lea R_sp(r), [\fp - \offset]
475 #if __ELF__
476 .cfi_def_cfa_register R_sp(r)
477 .cfi_adjust_cfa_offset +\offset
478 #endif
479 .endif
480 .L$_frameptr_p = 0
481 .purgem dropfp
482 .endm
483
484 .macro stalloc n
485 sub R_sp(r), \n
486 #if __ELF__
487 .cfi_adjust_cfa_offset +\n
488 #endif
489 #if ABI_WIN && CPUFAM_AMD64
490 .seh_stackalloc \n
491 #endif
492 .endm
493
494 .macro stfree n
495 add R_sp(r), \n
496 #if __ELF__
497 .cfi_adjust_cfa_offset -\n
498 #endif
499 .endm
500
501 .macro pushreg r
502 push \r
503 #if __ELF__
504 .cfi_adjust_cfa_offset +WORDSZ
505 .cfi_rel_offset \r, 0
506 #endif
507 #if ABI_WIN && CPUFAM_AMD64
508 .seh_pushreg \r
509 #endif
510 .endm
511
512 .macro popreg r
513 pop \r
514 #if __ELF__
515 .cfi_adjust_cfa_offset -WORDSZ
516 .cfi_restore \r
517 #endif
518 .endm
519
520 .macro savexmm r, offset
521 movdqa [R_sp(r) + \offset], \r
522 #if ABI_WIN && CPUFAM_AMD64
523 .seh_savexmm \r, \offset
524 #endif
525 .endm
526
527 .macro rstrxmm r, offset
528 movdqa \r, [R_sp(r) + \offset]
529 .endm
530
531 .macro endprologue
532 #if ABI_WIN && CPUFAM_AMD64
533 .seh_endprologue
534 #endif
535 .L$_prologue_p = -1
536 .endm
537
538 #endif
539
540 #if CPUFAM_X86
541
542 .macro _reg.0
543 // Stash GP registers and establish temporary stack frame.
544 pushfd
545 push eax
546 push ecx
547 push edx
548 push ebp
549 mov ebp, esp
550 and esp, ~15
551 sub esp, 512
552 fxsave [esp]
553 .endm
554
555 .macro _reg.1
556 .endm
557
558 .macro _reg.2
559 .endm
560
561 .macro _reg.3 fmt
562 // Print FMT and the other established arguments.
563 lea eax, .L$_reg$msg.\@
564 push eax
565 call printf
566 jmp .L$_reg$cont.\@
567 .L$_reg$msg.\@:
568 .ascii ";; \fmt\n\0"
569 .L$_reg$cont.\@:
570 mov eax, ebp
571 and eax, ~15
572 sub eax, 512
573 fxrstor [eax]
574 mov esp, ebp
575 pop ebp
576 pop edx
577 pop ecx
578 pop eax
579 popfd
580 .endm
581
582 .macro msg msg
583 _reg.0
584 _reg.1
585 _reg.2
586 _reg.3 "\msg"
587 .endm
588
589 .macro reg r, msg
590 _reg.0
591 .ifeqs "\r", "esp"
592 lea eax, [ebp + 20]
593 push eax
594 .else
595 .ifeqs "\r", "ebp"
596 push [ebp]
597 .else
598 push \r
599 .endif
600 .endif
601 _reg.1
602 _reg.2
603 _reg.3 "\msg: \r = %08x"
604 .endm
605
606 .macro xmmreg r, msg
607 _reg.0
608 _reg.1
609 _reg.2
610 movdqu xmm0, \r
611 pshufd xmm0, xmm0, 0x1b
612 sub esp, 16
613 movdqa [esp], xmm0
614 _reg.3 "\msg: \r = %08x %08x %08x %08x"
615 .endm
616
617 .macro mmreg r, msg
618 _reg.0
619 _reg.1
620 _reg.2
621 pshufw \r, \r, 0x4e
622 sub esp, 8
623 movq [esp], \r
624 _reg.3 "\msg: \r = %08x %08x"
625 .endm
626
627 .macro freg i, msg
628 _reg.0
629 _reg.1
630 _reg.2
631 finit
632 fldt [esp + 32 + 16*\i]
633 sub esp, 12
634 fstpt [esp]
635 _reg.3 "\msg: st(\i) = %.20Lg"
636 .endm
637
638 .macro fxreg i, msg
639 _reg.0
640 _reg.1
641 _reg.2
642 finit
643 fldt [esp + 32 + 16*\i]
644 sub esp, 12
645 fstpt [esp]
646 _reg.3 "\msg: st(\i) = %La"
647 .endm
648
649 #endif
650
651 ///--------------------------------------------------------------------------
652 /// ARM-specific hacking.
653
654 #if CPUFAM_ARMEL
655
656 // ARM/Thumb mode things. Use ARM by default.
657 #define ARM .arm; .L$_pcoff = 8
658 #define THUMB .thumb; .L$_pcoff = 4
659 ARM
660
661 // Set the function hooks.
662 #define FUNC_PREHOOK(_) .balign 4; .fnstart
663 #define ENDFUNC_HOOK(_) .fnend; .ltorg
664
665 // Call external subroutine at ADDR, possibly via PLT.
666 .macro callext addr, cond=
667 #if WANT_PIC
668 bl\cond \addr(PLT)
669 #else
670 bl\cond \addr
671 #endif
672 .endm
673
674 // Do I need to arrange a spare GOT register?
675 #if WANT_PIC
676 # define NEED_GOT 1
677 #endif
678 #define GOTREG r9
679
680 // Maybe load GOT address into GOT.
681 .macro ldgot cond=, got=GOTREG
682 #if WANT_PIC
683 ldr\cond \got, .L$_ldgot$\@
684 .L$_ldgot_pc$\@:
685 add\cond \got, pc, \got
686 _LIT
687 .balign 4
688 .L$_ldgot$\@:
689 .word _GLOBAL_OFFSET_TABLE_ - .L$_ldgot_pc$\@ - .L$_pcoff
690 _ENDLIT
691 #endif
692 .endm
693
694 // Load address of external symbol ADDR into REG, maybe using GOT.
695 .macro leaext reg, addr, cond=, got=GOTREG
696 #if WANT_PIC
697 ldr\cond \reg, .L$_leaext$\@
698 ldr\cond \reg, [\got, \reg]
699 _LIT
700 .balign 4
701 .L$_leaext$\@:
702 .word \addr(GOT)
703 _ENDLIT
704 #else
705 ldr\cond \reg, =\addr
706 #endif
707 .endm
708
709 // Load address of external symbol ADDR into REG directly.
710 .macro leaextq reg, addr, cond=
711 #if WANT_PIC
712 ldr\cond \reg, .L$_leaextq$\@
713 .L$_leaextq_pc$\@:
714 .if .L$_pcoff == 8
715 ldr\cond \reg, [pc, \reg]
716 .else
717 add\cond \reg, pc
718 ldr\cond \reg, [\reg]
719 .endif
720 _LIT
721 .balign 4
722 .L$_leaextq$\@:
723 .word \addr(GOT_PREL) + (. - .L$_leaextq_pc$\@ - .L$_pcoff)
724 _ENDLIT
725 #else
726 ldr\cond \reg, =\addr
727 #endif
728 .endm
729
730 // Apply decoration decor to register name reg.
731 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
732
733 // Internal macros: `_REGFORM_r(decor)' applies decoration decor to register
734 // name r.
735
736 #define _REGFORM_s0(decor) _DECOR(s, decor, 0)
737 #define _REGFORM_s1(decor) _DECOR(s, decor, 1)
738 #define _REGFORM_s2(decor) _DECOR(s, decor, 2)
739 #define _REGFORM_s3(decor) _DECOR(s, decor, 3)
740 #define _REGFORM_s4(decor) _DECOR(s, decor, 4)
741 #define _REGFORM_s5(decor) _DECOR(s, decor, 5)
742 #define _REGFORM_s6(decor) _DECOR(s, decor, 6)
743 #define _REGFORM_s7(decor) _DECOR(s, decor, 7)
744 #define _REGFORM_s8(decor) _DECOR(s, decor, 8)
745 #define _REGFORM_s9(decor) _DECOR(s, decor, 9)
746 #define _REGFORM_s10(decor) _DECOR(s, decor, 10)
747 #define _REGFORM_s11(decor) _DECOR(s, decor, 11)
748 #define _REGFORM_s12(decor) _DECOR(s, decor, 12)
749 #define _REGFORM_s13(decor) _DECOR(s, decor, 13)
750 #define _REGFORM_s14(decor) _DECOR(s, decor, 14)
751 #define _REGFORM_s15(decor) _DECOR(s, decor, 15)
752 #define _REGFORM_s16(decor) _DECOR(s, decor, 16)
753 #define _REGFORM_s17(decor) _DECOR(s, decor, 17)
754 #define _REGFORM_s18(decor) _DECOR(s, decor, 18)
755 #define _REGFORM_s19(decor) _DECOR(s, decor, 19)
756 #define _REGFORM_s20(decor) _DECOR(s, decor, 20)
757 #define _REGFORM_s21(decor) _DECOR(s, decor, 21)
758 #define _REGFORM_s22(decor) _DECOR(s, decor, 22)
759 #define _REGFORM_s23(decor) _DECOR(s, decor, 23)
760 #define _REGFORM_s24(decor) _DECOR(s, decor, 24)
761 #define _REGFORM_s25(decor) _DECOR(s, decor, 25)
762 #define _REGFORM_s26(decor) _DECOR(s, decor, 26)
763 #define _REGFORM_s27(decor) _DECOR(s, decor, 27)
764 #define _REGFORM_s28(decor) _DECOR(s, decor, 28)
765 #define _REGFORM_s29(decor) _DECOR(s, decor, 29)
766 #define _REGFORM_s30(decor) _DECOR(s, decor, 30)
767 #define _REGFORM_s31(decor) _DECOR(s, decor, 31)
768
769 #define _REGFORM_d0(decor) _DECOR(d, decor, 0)
770 #define _REGFORM_d1(decor) _DECOR(d, decor, 1)
771 #define _REGFORM_d2(decor) _DECOR(d, decor, 2)
772 #define _REGFORM_d3(decor) _DECOR(d, decor, 3)
773 #define _REGFORM_d4(decor) _DECOR(d, decor, 4)
774 #define _REGFORM_d5(decor) _DECOR(d, decor, 5)
775 #define _REGFORM_d6(decor) _DECOR(d, decor, 6)
776 #define _REGFORM_d7(decor) _DECOR(d, decor, 7)
777 #define _REGFORM_d8(decor) _DECOR(d, decor, 8)
778 #define _REGFORM_d9(decor) _DECOR(d, decor, 9)
779 #define _REGFORM_d10(decor) _DECOR(d, decor, 10)
780 #define _REGFORM_d11(decor) _DECOR(d, decor, 11)
781 #define _REGFORM_d12(decor) _DECOR(d, decor, 12)
782 #define _REGFORM_d13(decor) _DECOR(d, decor, 13)
783 #define _REGFORM_d14(decor) _DECOR(d, decor, 14)
784 #define _REGFORM_d15(decor) _DECOR(d, decor, 15)
785 #define _REGFORM_d16(decor) _DECOR(d, decor, 16)
786 #define _REGFORM_d17(decor) _DECOR(d, decor, 17)
787 #define _REGFORM_d18(decor) _DECOR(d, decor, 18)
788 #define _REGFORM_d19(decor) _DECOR(d, decor, 19)
789 #define _REGFORM_d20(decor) _DECOR(d, decor, 20)
790 #define _REGFORM_d21(decor) _DECOR(d, decor, 21)
791 #define _REGFORM_d22(decor) _DECOR(d, decor, 22)
792 #define _REGFORM_d23(decor) _DECOR(d, decor, 23)
793 #define _REGFORM_d24(decor) _DECOR(d, decor, 24)
794 #define _REGFORM_d25(decor) _DECOR(d, decor, 25)
795 #define _REGFORM_d26(decor) _DECOR(d, decor, 26)
796 #define _REGFORM_d27(decor) _DECOR(d, decor, 27)
797 #define _REGFORM_d28(decor) _DECOR(d, decor, 28)
798 #define _REGFORM_d29(decor) _DECOR(d, decor, 29)
799 #define _REGFORM_d30(decor) _DECOR(d, decor, 30)
800 #define _REGFORM_d31(decor) _DECOR(d, decor, 31)
801
802 #define _REGFORM_q0(decor) _DECOR(q, decor, 0)
803 #define _REGFORM_q1(decor) _DECOR(q, decor, 1)
804 #define _REGFORM_q2(decor) _DECOR(q, decor, 2)
805 #define _REGFORM_q3(decor) _DECOR(q, decor, 3)
806 #define _REGFORM_q4(decor) _DECOR(q, decor, 4)
807 #define _REGFORM_q5(decor) _DECOR(q, decor, 5)
808 #define _REGFORM_q6(decor) _DECOR(q, decor, 6)
809 #define _REGFORM_q7(decor) _DECOR(q, decor, 7)
810 #define _REGFORM_q8(decor) _DECOR(q, decor, 8)
811 #define _REGFORM_q9(decor) _DECOR(q, decor, 9)
812 #define _REGFORM_q10(decor) _DECOR(q, decor, 10)
813 #define _REGFORM_q11(decor) _DECOR(q, decor, 11)
814 #define _REGFORM_q12(decor) _DECOR(q, decor, 12)
815 #define _REGFORM_q13(decor) _DECOR(q, decor, 13)
816 #define _REGFORM_q14(decor) _DECOR(q, decor, 14)
817 #define _REGFORM_q15(decor) _DECOR(q, decor, 15)
818
819 // `_LOPART(n)' and `_HIPART(n)' return the numbers of the register halves of
820 // register n, i.e., 2*n and 2*n + 1 respectively.
821 #define _LOPART(n) _GLUE(_LOPART_, n)
822 #define _HIPART(n) _GLUE(_HIPART_, n)
823
824 // Internal macros: `_LOPART_n' and `_HIPART_n' return the numbers of the
825 // register halves of register n, i.e., 2*n and 2*n + 1 respectively.
826
827 #define _LOPART_0 0
828 #define _HIPART_0 1
829 #define _LOPART_1 2
830 #define _HIPART_1 3
831 #define _LOPART_2 4
832 #define _HIPART_2 5
833 #define _LOPART_3 6
834 #define _HIPART_3 7
835 #define _LOPART_4 8
836 #define _HIPART_4 9
837 #define _LOPART_5 10
838 #define _HIPART_5 11
839 #define _LOPART_6 12
840 #define _HIPART_6 13
841 #define _LOPART_7 14
842 #define _HIPART_7 15
843 #define _LOPART_8 16
844 #define _HIPART_8 17
845 #define _LOPART_9 18
846 #define _HIPART_9 19
847 #define _LOPART_10 20
848 #define _HIPART_10 21
849 #define _LOPART_11 22
850 #define _HIPART_11 23
851 #define _LOPART_12 24
852 #define _HIPART_12 25
853 #define _LOPART_13 26
854 #define _HIPART_13 27
855 #define _LOPART_14 28
856 #define _HIPART_14 29
857 #define _LOPART_15 30
858 #define _HIPART_15 31
859
860 // Return the register number of the pair containing register n, i.e.,
861 // floor(n/2).
862 #define _PAIR(n) _GLUE(_PAIR_, n)
863
864 // Internal macros: `_PAIR_n' returns the register number of the pair
865 // containing register n, i.e., floor(n/2).
866 #define _PAIR_0 0
867 #define _PAIR_1 0
868 #define _PAIR_2 1
869 #define _PAIR_3 1
870 #define _PAIR_4 2
871 #define _PAIR_5 2
872 #define _PAIR_6 3
873 #define _PAIR_7 3
874 #define _PAIR_8 4
875 #define _PAIR_9 4
876 #define _PAIR_10 5
877 #define _PAIR_11 5
878 #define _PAIR_12 6
879 #define _PAIR_13 6
880 #define _PAIR_14 7
881 #define _PAIR_15 7
882 #define _PAIR_16 8
883 #define _PAIR_17 8
884 #define _PAIR_18 9
885 #define _PAIR_19 9
886 #define _PAIR_20 10
887 #define _PAIR_21 10
888 #define _PAIR_22 11
889 #define _PAIR_23 11
890 #define _PAIR_24 12
891 #define _PAIR_25 12
892 #define _PAIR_26 13
893 #define _PAIR_27 13
894 #define _PAIR_28 14
895 #define _PAIR_29 14
896 #define _PAIR_30 15
897 #define _PAIR_31 15
898
899 // Apply decoration decor to register number n of type ty. Decorations are
900 // as follows.
901 //
902 // decor types meaning
903 // Q s, d the NEON qN register containing this one
904 // D s the NEON dN register containing this one
905 // D0 q the low 64-bit half of this one
906 // D1 q the high 64-bit half of this one
907 // S0 d, q the first 32-bit piece of this one
908 // S1 d, q the second 32-bit piece of this one
909 // S2 q the third 32-bit piece of this one
910 // S3 q the fourth 32-bit piece of this one
911 // Bn q the nth byte of this register, as a scalar
912 // Hn q the nth halfword of this register, as a scalar
913 // Wn q the nth word of this register, as a scalar
914 #define _DECOR(ty, decor, n) _DECOR_##ty##_##decor(n)
915
916 // Internal macros: `_DECOR_ty_decor(n)' applies decoration decor to register
917 // number n of type ty.
918
919 #define _DECOR_s_Q(n) GLUE(q, _PAIR(_PAIR(n)))
920 #define _DECOR_s_D(n) GLUE(d, _PAIR(n))
921
922 #define _DECOR_d_Q(n) GLUE(q, _PAIR(n))
923 #define _DECOR_d_S0(n) GLUE(s, _LOPART(n))
924 #define _DECOR_d_S1(n) GLUE(s, _LOPART(n))
925
926 #define _DECOR_q_D0(n) GLUE(d, _LOPART(n))
927 #define _DECOR_q_D1(n) GLUE(d, _HIPART(n))
928 #define _DECOR_q_S0(n) GLUE(s, _LOPART(_LOPART(n)))
929 #define _DECOR_q_S1(n) GLUE(s, _HIPART(_LOPART(n)))
930 #define _DECOR_q_S2(n) GLUE(s, _LOPART(_HIPART(n)))
931 #define _DECOR_q_S3(n) GLUE(s, _HIPART(_HIPART(n)))
932 #define _DECOR_q_W0(n) GLUE(d, _LOPART(n))[0]
933 #define _DECOR_q_W1(n) GLUE(d, _LOPART(n))[1]
934 #define _DECOR_q_W2(n) GLUE(d, _HIPART(n))[0]
935 #define _DECOR_q_W3(n) GLUE(d, _HIPART(n))[1]
936 #define _DECOR_q_H0(n) GLUE(d, _LOPART(n))[0]
937 #define _DECOR_q_H1(n) GLUE(d, _LOPART(n))[1]
938 #define _DECOR_q_H2(n) GLUE(d, _LOPART(n))[2]
939 #define _DECOR_q_H3(n) GLUE(d, _LOPART(n))[3]
940 #define _DECOR_q_H4(n) GLUE(d, _HIPART(n))[0]
941 #define _DECOR_q_H5(n) GLUE(d, _HIPART(n))[1]
942 #define _DECOR_q_H6(n) GLUE(d, _HIPART(n))[2]
943 #define _DECOR_q_H7(n) GLUE(d, _HIPART(n))[3]
944 #define _DECOR_q_B0(n) GLUE(d, _LOPART(n))[0]
945 #define _DECOR_q_B1(n) GLUE(d, _LOPART(n))[1]
946 #define _DECOR_q_B2(n) GLUE(d, _LOPART(n))[2]
947 #define _DECOR_q_B3(n) GLUE(d, _LOPART(n))[3]
948 #define _DECOR_q_B4(n) GLUE(d, _LOPART(n))[4]
949 #define _DECOR_q_B5(n) GLUE(d, _LOPART(n))[5]
950 #define _DECOR_q_B6(n) GLUE(d, _LOPART(n))[6]
951 #define _DECOR_q_B7(n) GLUE(d, _LOPART(n))[7]
952 #define _DECOR_q_B8(n) GLUE(d, _HIPART(n))[0]
953 #define _DECOR_q_B9(n) GLUE(d, _HIPART(n))[1]
954 #define _DECOR_q_B10(n) GLUE(d, _HIPART(n))[2]
955 #define _DECOR_q_B11(n) GLUE(d, _HIPART(n))[3]
956 #define _DECOR_q_B12(n) GLUE(d, _HIPART(n))[4]
957 #define _DECOR_q_B13(n) GLUE(d, _HIPART(n))[5]
958 #define _DECOR_q_B14(n) GLUE(d, _HIPART(n))[6]
959 #define _DECOR_q_B15(n) GLUE(d, _HIPART(n))[7]
960
961 // Macros for navigating the NEON register hierarchy.
962 #define S0(reg) _REGFORM(reg, S0)
963 #define S1(reg) _REGFORM(reg, S1)
964 #define S2(reg) _REGFORM(reg, S2)
965 #define S3(reg) _REGFORM(reg, S3)
966 #define D(reg) _REGFORM(reg, D)
967 #define D0(reg) _REGFORM(reg, D0)
968 #define D1(reg) _REGFORM(reg, D1)
969 #define Q(reg) _REGFORM(reg, Q)
970
971 // Macros for indexing quadword registers.
972 #define QB(reg, i) _REGFORM(reg, B##i)
973 #define QH(reg, i) _REGFORM(reg, H##i)
974 #define QW(reg, i) _REGFORM(reg, W##i)
975
976 // Macros for converting vldm/vstm ranges.
977 #define QQ(qlo, qhi) D0(qlo)-D1(qhi)
978
979 // Stack management and unwinding.
980 .macro setfp fp, offset = 0
981 .if \offset == 0
982 mov \fp, sp
983 .setfp \fp, sp
984 .else
985 add \fp, sp, #\offset
986 .setfp \fp, sp, #\offset
987 .endif
988 .macro dropfp; _dropfp \fp, \offset; .endm
989 .L$_frameptr_p = -1
990 .endm
991
992 .macro _dropfp fp, offset = 0
993 .if \offset == 0
994 mov sp, \fp
995 .else
996 sub sp, \fp, #\offset
997 .endif
998 .purgem dropfp
999 .L$_frameptr_p = 0
1000 .endm
1001
1002 .macro stalloc n
1003 sub sp, sp, #\n
1004 .pad #\n
1005 .endm
1006
1007 .macro stfree n
1008 add sp, sp, #\n
1009 .pad #-\n
1010 .endm
1011
1012 .macro pushreg rr:vararg
1013 stmfd sp!, {\rr}
1014 .save {\rr}
1015 .endm
1016
1017 .macro popreg rr:vararg
1018 ldmfd sp!, {\rr}
1019 .endm
1020
1021 .macro pushvfp rr:vararg
1022 vstmdb sp!, {\rr}
1023 .vsave {\rr}
1024 .endm
1025
1026 .macro popvfp rr:vararg
1027 vldmia sp!, {\rr}
1028 .endm
1029
1030 .macro endprologue
1031 .endm
1032
1033 // No need for prologue markers on ARM.
1034 #define FUNC_POSTHOOK(_) .L$_prologue_p = -1
1035
1036 #endif
1037
1038 ///--------------------------------------------------------------------------
1039 /// AArch64-specific hacking.
1040
1041 #if CPUFAM_ARM64
1042
1043 // Set the function hooks.
1044 #define FUNC_PREHOOK(_) .balign 4
1045 #define FUNC_POSTHOOK(_) .cfi_startproc; .L$_prologue_p = -1
1046 #define ENDFUNC_HOOK(_) .cfi_endproc
1047
1048 // Call external subroutine at ADDR, possibly via PLT.
1049 .macro callext addr
1050 bl \addr
1051 .endm
1052
1053 // Load address of external symbol ADDR into REG.
1054 .macro leaext reg, addr
1055 #if WANT_PIC
1056 adrp \reg, :got:\addr
1057 ldr \reg, [\reg, #:got_lo12:\addr]
1058 #else
1059 adrp \reg, \addr
1060 add \reg, \reg, #:lo12:\addr
1061 #endif
1062 .endm
1063
1064 // Stack management and unwinding.
1065 .macro setfp fp, offset = 0
1066 // If you're just going through the motions with a fixed-size stack frame,
1067 // then you want to say `add x29, sp, #OFFSET' directly, which will avoid
1068 // pointlessly restoring sp later.
1069 .if \offset == 0
1070 mov \fp, sp
1071 .cfi_def_cfa_register \fp
1072 .else
1073 add \fp, sp, #\offset
1074 .cfi_def_cfa_register \fp
1075 .cfi_adjust_cfa_offset -\offset
1076 .endif
1077 .macro dropfp; _dropfp \fp, \offset; .endm
1078 .L$_frameptr_p = -1
1079 .endm
1080
1081 .macro _dropfp fp, offset = 0
1082 .if \offset == 0
1083 mov sp, \fp
1084 .cfi_def_cfa_register sp
1085 .else
1086 sub sp, \fp, #\offset
1087 .cfi_def_cfa_register sp
1088 .cfi_adjust_cfa_offset +\offset
1089 .endif
1090 .purgem dropfp
1091 .L$_frameptr_p = 0
1092 .endm
1093
1094 .macro stalloc n
1095 sub sp, sp, #\n
1096 .cfi_adjust_cfa_offset +\n
1097 .endm
1098
1099 .macro stfree n
1100 add sp, sp, #\n
1101 .cfi_adjust_cfa_offset -\n
1102 .endm
1103
1104 .macro pushreg x, y=
1105 .ifeqs "\y", ""
1106 str \x, [sp, #-16]!
1107 .cfi_adjust_cfa_offset +16
1108 .cfi_rel_offset \x, 0
1109 .else
1110 stp \x, \y, [sp, #-16]!
1111 .cfi_adjust_cfa_offset +16
1112 .cfi_rel_offset \x, 0
1113 .cfi_rel_offset \y, 8
1114 .endif
1115 .endm
1116
1117 .macro popreg x, y=
1118 .ifeqs "\y", ""
1119 ldr \x, [sp], #16
1120 .cfi_restore \x
1121 .cfi_adjust_cfa_offset -16
1122 .else
1123 ldp \x, \y, [sp], #16
1124 .cfi_restore \x
1125 .cfi_restore \y
1126 .cfi_adjust_cfa_offset -16
1127 .endif
1128 .endm
1129
1130 .macro savereg x, y, z=
1131 .ifeqs "\z", ""
1132 str \x, [sp, #\y]
1133 .cfi_rel_offset \x, \y
1134 .else
1135 stp \x, \y, [sp, #\z]
1136 .cfi_rel_offset \x, \z
1137 .cfi_rel_offset \y, \z + 8
1138 .endif
1139 .endm
1140
1141 .macro rstrreg x, y, z=
1142 .ifeqs "\z", ""
1143 ldr \x, [sp, #\y]
1144 .cfi_restore \x
1145 .else
1146 ldp \x, \y, [sp, #\z]
1147 .cfi_restore \x
1148 .cfi_restore \y
1149 .endif
1150 .endm
1151
1152 .macro endprologue
1153 .endm
1154
1155 #endif
1156
1157 ///--------------------------------------------------------------------------
1158 /// Final stuff.
1159
1160 // Default values for the various hooks.
1161 #ifndef FUNC_PREHOOK
1162 # define FUNC_PREHOOK(_)
1163 #endif
1164 #ifndef FUNC_POSTHOOK
1165 # define FUNC_POSTHOOK(_)
1166 #endif
1167 #ifndef ENDFUNC_HOOK
1168 # define ENDFUNC_HOOK(_)
1169 #endif
1170
1171 #ifndef F
1172 # define F(name) name
1173 #endif
1174
1175 #ifndef TYPE_FUNC
1176 # define TYPE_FUNC(name)
1177 #endif
1178
1179 #ifndef SIZE_OBJ
1180 # define SIZE_OBJ(name)
1181 #endif
1182
1183 #if __ELF__ && !defined(WANT_EXECUTABLE_STACK)
1184 .pushsection .note.GNU-stack, "", _SECTTY(progbits)
1185 .popsection
1186 #endif
1187
1188 ///----- That's all, folks --------------------------------------------------
1189
1190 #endif
Catacomb cryptographic library