~mdw / catacomb / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
math/pgen.c, math/pgen-simul.c: Add Baillie--PSW testers.
[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_nil(decor) nil
315 #define R_ip(decor) _DECOR(ip, decor, ip)
316 #define R_a(decor) _DECOR(abcd, decor, a)
317 #define R_b(decor) _DECOR(abcd, decor, b)
318 #define R_c(decor) _DECOR(abcd, decor, c)
319 #define R_d(decor) _DECOR(abcd, decor, d)
320 #define R_si(decor) _DECOR(xp, decor, si)
321 #define R_di(decor) _DECOR(xp, decor, di)
322 #define R_bp(decor) _DECOR(xp, decor, bp)
323 #define R_sp(decor) _DECOR(xp, decor, sp)
324 #if CPUFAM_AMD64
325 # define R_r8(decor) _DECOR(rn, decor, r8)
326 # define R_r9(decor) _DECOR(rn, decor, r9)
327 # define R_r10(decor) _DECOR(rn, decor, r10)
328 # define R_r11(decor) _DECOR(rn, decor, r11)
329 # define R_r12(decor) _DECOR(rn, decor, r12)
330 # define R_r13(decor) _DECOR(rn, decor, r13)
331 # define R_r14(decor) _DECOR(rn, decor, r14)
332 # define R_r15(decor) _DECOR(rn, decor, r15)
333 #endif
334
335 // Refer to an in-memory datum of the type implied by decor residing at
336 // address addr (which should supply its own square-brackets).
337 #define MEM(decor, addr) _DECOR(mem, decor, addr)
338
339 // Refer to an immediate datum of the type implied by decor.
340 #define IMM(decor, imm) _DECOR(mem, decor, imm)
341
342 // Applies decoration decor to assembler-level register name reg.
343 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
344
345 // Internal macros: _REGFORM_r(decor) applies decoration decor to an
346 // assembler-level register name, in place of any decoration that register
347 // name has already.
348
349 #define _REGFORM_nil(decor) R_nil(decor)
350
351 #define _REGFORM_ip(decor) R_ip(decor)
352 #define _REGFORM_eip(decor) R_ip(decor)
353
354 #define _REGFORM_a(decor) R_a(decor)
355 #define _REGFORM_al(decor) R_a(decor)
356 #define _REGFORM_ah(decor) R_a(decor)
357 #define _REGFORM_ax(decor) R_a(decor)
358 #define _REGFORM_eax(decor) R_a(decor)
359
360 #define _REGFORM_b(decor) R_b(decor)
361 #define _REGFORM_bl(decor) R_b(decor)
362 #define _REGFORM_bh(decor) R_b(decor)
363 #define _REGFORM_bx(decor) R_b(decor)
364 #define _REGFORM_ebx(decor) R_b(decor)
365
366 #define _REGFORM_c(decor) R_c(decor)
367 #define _REGFORM_cl(decor) R_c(decor)
368 #define _REGFORM_ch(decor) R_c(decor)
369 #define _REGFORM_cx(decor) R_c(decor)
370 #define _REGFORM_ecx(decor) R_c(decor)
371
372 #define _REGFORM_d(decor) R_d(decor)
373 #define _REGFORM_dl(decor) R_d(decor)
374 #define _REGFORM_dh(decor) R_d(decor)
375 #define _REGFORM_dx(decor) R_d(decor)
376 #define _REGFORM_edx(decor) R_d(decor)
377
378 #define _REGFORM_si(decor) R_si(decor)
379 #define _REGFORM_sil(decor) R_si(decor)
380 #define _REGFORM_esi(decor) R_si(decor)
381
382 #define _REGFORM_di(decor) R_di(decor)
383 #define _REGFORM_dil(decor) R_di(decor)
384 #define _REGFORM_edi(decor) R_di(decor)
385
386 #define _REGFORM_bp(decor) R_bp(decor)
387 #define _REGFORM_bpl(decor) R_bp(decor)
388 #define _REGFORM_ebp(decor) R_bp(decor)
389
390 #define _REGFORM_sp(decor) R_sp(decor)
391 #define _REGFORM_spl(decor) R_sp(decor)
392 #define _REGFORM_esp(decor) R_sp(decor)
393
394 #if CPUFAM_AMD64
395
396 # define _REGFORM_rip(decor) R_ip(decor)
397 # define _REGFORM_rsp(decor) R_sp(decor)
398 # define _REGFORM_rbp(decor) R_bp(decor)
399 # define _REGFORM_rdi(decor) R_di(decor)
400 # define _REGFORM_rsi(decor) R_si(decor)
401 # define _REGFORM_rdx(decor) R_d(decor)
402 # define _REGFORM_rcx(decor) R_c(decor)
403 # define _REGFORM_rbx(decor) R_b(decor)
404 # define _REGFORM_rax(decor) R_a(decor)
405
406 # define _REGFORM_r8(decor) R_r8(decor)
407 # define _REGFORM_r8b(decor) R_r8(decor)
408 # define _REGFORM_r8w(decor) R_r8(decor)
409 # define _REGFORM_r8d(decor) R_r8(decor)
410
411 # define _REGFORM_r9(decor) R_r9(decor)
412 # define _REGFORM_r9b(decor) R_r9(decor)
413 # define _REGFORM_r9w(decor) R_r9(decor)
414 # define _REGFORM_r9d(decor) R_r9(decor)
415
416 # define _REGFORM_r10(decor) R_r10(decor)
417 # define _REGFORM_r10b(decor) R_r10(decor)
418 # define _REGFORM_r10w(decor) R_r10(decor)
419 # define _REGFORM_r10d(decor) R_r10(decor)
420
421 # define _REGFORM_r11(decor) R_r11(decor)
422 # define _REGFORM_r11b(decor) R_r11(decor)
423 # define _REGFORM_r11w(decor) R_r11(decor)
424 # define _REGFORM_r11d(decor) R_r11(decor)
425
426 # define _REGFORM_r12(decor) R_r12(decor)
427 # define _REGFORM_r12b(decor) R_r12(decor)
428 # define _REGFORM_r12w(decor) R_r12(decor)
429 # define _REGFORM_r12d(decor) R_r12(decor)
430
431 # define _REGFORM_r13(decor) R_r13(decor)
432 # define _REGFORM_r13b(decor) R_r13(decor)
433 # define _REGFORM_r13w(decor) R_r13(decor)
434 # define _REGFORM_r13d(decor) R_r13(decor)
435
436 # define _REGFORM_r14(decor) R_r14(decor)
437 # define _REGFORM_r14b(decor) R_r14(decor)
438 # define _REGFORM_r14w(decor) R_r14(decor)
439 # define _REGFORM_r14d(decor) R_r14(decor)
440
441 # define _REGFORM_r15(decor) R_r15(decor)
442 # define _REGFORM_r15b(decor) R_r15(decor)
443 # define _REGFORM_r15w(decor) R_r15(decor)
444 # define _REGFORM_r15d(decor) R_r15(decor)
445
446 #endif
447
448 // Macros for converting register names.
449 #define BYTE(reg) _REGFORM(reg, b)
450 #define HIBYTE(reg) _REGFORM(reg, h)
451 #define WORD(reg) _REGFORM(reg, w)
452 #define DWORD(reg) _REGFORM(reg, d)
453 #if CPUFAM_AMD64
454 # define QWORD(reg) _REGFORM(reg, q)
455 #endif
456 #define WHOLE(reg) _REGFORM(reg, r)
457
458 // Macros for some common registers.
459 #define AX R_a(r)
460 #define BX R_b(r)
461 #define CX R_c(r)
462 #define DX R_d(r)
463 #define SI R_si(r)
464 #define DI R_di(r)
465 #define BP R_bp(r)
466 #define SP R_sp(r)
467
468 // Stack management and unwinding.
469 .macro setfp fp=BP, offset=0
470 .if \offset == 0
471 mov \fp, SP
472 #if __ELF__
473 .cfi_def_cfa_register \fp
474 #endif
475 #if ABI_WIN && CPUFAM_AMD64
476 .seh_setframe \fp, 0
477 #endif
478 .else
479 lea \fp, [SP + \offset]
480 #if __ELF__
481 .cfi_def_cfa_register \fp
482 .cfi_adjust_cfa_offset -\offset
483 #endif
484 #if ABI_WIN && CPUFAM_AMD64
485 .seh_setframe \fp, \offset
486 #endif
487 .endif
488 .L$_frameptr_p = -1
489 .macro dropfp; _dropfp \fp, \offset; .endm
490 .endm
491
492 .macro _dropfp fp, offset=0
493 .if \offset == 0
494 mov SP, \fp
495 #if __ELF__
496 .cfi_def_cfa_register SP
497 #endif
498 .else
499 lea SP, [\fp - \offset]
500 #if __ELF__
501 .cfi_def_cfa_register SP
502 .cfi_adjust_cfa_offset +\offset
503 #endif
504 .endif
505 .L$_frameptr_p = 0
506 .purgem dropfp
507 .endm
508
509 .macro stalloc n
510 sub SP, \n
511 #if __ELF__
512 .cfi_adjust_cfa_offset +\n
513 #endif
514 #if ABI_WIN && CPUFAM_AMD64
515 .seh_stackalloc \n
516 #endif
517 .endm
518
519 .macro stfree n
520 add SP, \n
521 #if __ELF__
522 .cfi_adjust_cfa_offset -\n
523 #endif
524 .endm
525
526 .macro pushreg r
527 push \r
528 #if __ELF__
529 .cfi_adjust_cfa_offset +WORDSZ
530 .cfi_rel_offset \r, 0
531 #endif
532 #if ABI_WIN && CPUFAM_AMD64
533 .seh_pushreg \r
534 #endif
535 .endm
536
537 .macro popreg r
538 pop \r
539 #if __ELF__
540 .cfi_adjust_cfa_offset -WORDSZ
541 .cfi_restore \r
542 #endif
543 .endm
544
545 .macro savexmm r, offset
546 movdqa [SP + \offset], \r
547 #if ABI_WIN && CPUFAM_AMD64
548 .seh_savexmm \r, \offset
549 #endif
550 .endm
551
552 .macro rstrxmm r, offset
553 movdqa \r, [SP + \offset]
554 .endm
555
556 .macro endprologue
557 #if ABI_WIN && CPUFAM_AMD64
558 .seh_endprologue
559 #endif
560 .L$_prologue_p = -1
561 .endm
562
563 #endif
564
565 ///--------------------------------------------------------------------------
566 /// ARM-specific hacking.
567
568 #if CPUFAM_ARMEL
569
570 // ARM/Thumb mode things. Use ARM by default.
571 #define ARM .arm; .L$_pcoff = 8
572 #define THUMB .thumb; .L$_pcoff = 4
573 ARM
574
575 // Set the function hooks.
576 #define FUNC_PREHOOK(_) .balign 4; .fnstart
577 #define ENDFUNC_HOOK(_) .fnend; .ltorg
578
579 // Call external subroutine at ADDR, possibly via PLT.
580 .macro callext addr, cond=
581 #if WANT_PIC
582 bl\cond \addr(PLT)
583 #else
584 bl\cond \addr
585 #endif
586 .endm
587
588 // Do I need to arrange a spare GOT register?
589 #if WANT_PIC
590 # define NEED_GOT 1
591 #endif
592 #define GOTREG r9
593
594 // Maybe load GOT address into GOT.
595 .macro ldgot cond=, got=GOTREG
596 #if WANT_PIC
597 ldr\cond \got, .L$_ldgot$\@
598 .L$_ldgot_pc$\@:
599 add\cond \got, pc, \got
600 _LIT
601 .balign 4
602 .L$_ldgot$\@:
603 .word _GLOBAL_OFFSET_TABLE_ - .L$_ldgot_pc$\@ - .L$_pcoff
604 _ENDLIT
605 #endif
606 .endm
607
608 // Load address of external symbol ADDR into REG, maybe using GOT.
609 .macro leaext reg, addr, cond=, got=GOTREG
610 #if WANT_PIC
611 ldr\cond \reg, .L$_leaext$\@
612 ldr\cond \reg, [\got, \reg]
613 _LIT
614 .balign 4
615 .L$_leaext$\@:
616 .word \addr(GOT)
617 _ENDLIT
618 #else
619 ldr\cond \reg, =\addr
620 #endif
621 .endm
622
623 // Load address of external symbol ADDR into REG directly.
624 .macro leaextq reg, addr, cond=
625 #if WANT_PIC
626 ldr\cond \reg, .L$_leaextq$\@
627 .L$_leaextq_pc$\@:
628 .if .L$_pcoff == 8
629 ldr\cond \reg, [pc, \reg]
630 .else
631 add\cond \reg, pc
632 ldr\cond \reg, [\reg]
633 .endif
634 _LIT
635 .balign 4
636 .L$_leaextq$\@:
637 .word \addr(GOT_PREL) + (. - .L$_leaextq_pc$\@ - .L$_pcoff)
638 _ENDLIT
639 #else
640 ldr\cond \reg, =\addr
641 #endif
642 .endm
643
644 .macro vzero vz=q15
645 // Set VZ (default q15) to zero.
646 vmov.u32 \vz, #0
647 .endm
648
649 .macro vshl128 vd, vn, nbit, vz=q15
650 // Set VD to VN shifted left by NBIT. Assume VZ (default q15) is
651 // all-bits-zero. NBIT must be a multiple of 8.
652 .if \nbit&3 != 0
653 .error "shift quantity must be whole number of bytes"
654 .endif
655 vext.8 \vd, \vz, \vn, #16 - (\nbit >> 3)
656 .endm
657
658 .macro vshr128 vd, vn, nbit, vz=q15
659 // Set VD to VN shifted right by NBIT. Assume VZ (default q15) is
660 // all-bits-zero. NBIT must be a multiple of 8.
661 .if \nbit&3 != 0
662 .error "shift quantity must be whole number of bytes"
663 .endif
664 vext.8 \vd, \vn, \vz, #\nbit >> 3
665 .endm
666
667 // Apply decoration decor to register name reg.
668 #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)
669
670 // Internal macros: `_REGFORM_r(decor)' applies decoration decor to register
671 // name r.
672
673 #define _REGFORM_nil(decor) nil
674
675 #define _REGFORM_s0(decor) _DECOR(s, decor, 0)
676 #define _REGFORM_s1(decor) _DECOR(s, decor, 1)
677 #define _REGFORM_s2(decor) _DECOR(s, decor, 2)
678 #define _REGFORM_s3(decor) _DECOR(s, decor, 3)
679 #define _REGFORM_s4(decor) _DECOR(s, decor, 4)
680 #define _REGFORM_s5(decor) _DECOR(s, decor, 5)
681 #define _REGFORM_s6(decor) _DECOR(s, decor, 6)
682 #define _REGFORM_s7(decor) _DECOR(s, decor, 7)
683 #define _REGFORM_s8(decor) _DECOR(s, decor, 8)
684 #define _REGFORM_s9(decor) _DECOR(s, decor, 9)
685 #define _REGFORM_s10(decor) _DECOR(s, decor, 10)
686 #define _REGFORM_s11(decor) _DECOR(s, decor, 11)
687 #define _REGFORM_s12(decor) _DECOR(s, decor, 12)
688 #define _REGFORM_s13(decor) _DECOR(s, decor, 13)
689 #define _REGFORM_s14(decor) _DECOR(s, decor, 14)
690 #define _REGFORM_s15(decor) _DECOR(s, decor, 15)
691 #define _REGFORM_s16(decor) _DECOR(s, decor, 16)
692 #define _REGFORM_s17(decor) _DECOR(s, decor, 17)
693 #define _REGFORM_s18(decor) _DECOR(s, decor, 18)
694 #define _REGFORM_s19(decor) _DECOR(s, decor, 19)
695 #define _REGFORM_s20(decor) _DECOR(s, decor, 20)
696 #define _REGFORM_s21(decor) _DECOR(s, decor, 21)
697 #define _REGFORM_s22(decor) _DECOR(s, decor, 22)
698 #define _REGFORM_s23(decor) _DECOR(s, decor, 23)
699 #define _REGFORM_s24(decor) _DECOR(s, decor, 24)
700 #define _REGFORM_s25(decor) _DECOR(s, decor, 25)
701 #define _REGFORM_s26(decor) _DECOR(s, decor, 26)
702 #define _REGFORM_s27(decor) _DECOR(s, decor, 27)
703 #define _REGFORM_s28(decor) _DECOR(s, decor, 28)
704 #define _REGFORM_s29(decor) _DECOR(s, decor, 29)
705 #define _REGFORM_s30(decor) _DECOR(s, decor, 30)
706 #define _REGFORM_s31(decor) _DECOR(s, decor, 31)
707
708 #define _REGFORM_d0(decor) _DECOR(d, decor, 0)
709 #define _REGFORM_d1(decor) _DECOR(d, decor, 1)
710 #define _REGFORM_d2(decor) _DECOR(d, decor, 2)
711 #define _REGFORM_d3(decor) _DECOR(d, decor, 3)
712 #define _REGFORM_d4(decor) _DECOR(d, decor, 4)
713 #define _REGFORM_d5(decor) _DECOR(d, decor, 5)
714 #define _REGFORM_d6(decor) _DECOR(d, decor, 6)
715 #define _REGFORM_d7(decor) _DECOR(d, decor, 7)
716 #define _REGFORM_d8(decor) _DECOR(d, decor, 8)
717 #define _REGFORM_d9(decor) _DECOR(d, decor, 9)
718 #define _REGFORM_d10(decor) _DECOR(d, decor, 10)
719 #define _REGFORM_d11(decor) _DECOR(d, decor, 11)
720 #define _REGFORM_d12(decor) _DECOR(d, decor, 12)
721 #define _REGFORM_d13(decor) _DECOR(d, decor, 13)
722 #define _REGFORM_d14(decor) _DECOR(d, decor, 14)
723 #define _REGFORM_d15(decor) _DECOR(d, decor, 15)
724 #define _REGFORM_d16(decor) _DECOR(d, decor, 16)
725 #define _REGFORM_d17(decor) _DECOR(d, decor, 17)
726 #define _REGFORM_d18(decor) _DECOR(d, decor, 18)
727 #define _REGFORM_d19(decor) _DECOR(d, decor, 19)
728 #define _REGFORM_d20(decor) _DECOR(d, decor, 20)
729 #define _REGFORM_d21(decor) _DECOR(d, decor, 21)
730 #define _REGFORM_d22(decor) _DECOR(d, decor, 22)
731 #define _REGFORM_d23(decor) _DECOR(d, decor, 23)
732 #define _REGFORM_d24(decor) _DECOR(d, decor, 24)
733 #define _REGFORM_d25(decor) _DECOR(d, decor, 25)
734 #define _REGFORM_d26(decor) _DECOR(d, decor, 26)
735 #define _REGFORM_d27(decor) _DECOR(d, decor, 27)
736 #define _REGFORM_d28(decor) _DECOR(d, decor, 28)
737 #define _REGFORM_d29(decor) _DECOR(d, decor, 29)
738 #define _REGFORM_d30(decor) _DECOR(d, decor, 30)
739 #define _REGFORM_d31(decor) _DECOR(d, decor, 31)
740
741 #define _REGFORM_q0(decor) _DECOR(q, decor, 0)
742 #define _REGFORM_q1(decor) _DECOR(q, decor, 1)
743 #define _REGFORM_q2(decor) _DECOR(q, decor, 2)
744 #define _REGFORM_q3(decor) _DECOR(q, decor, 3)
745 #define _REGFORM_q4(decor) _DECOR(q, decor, 4)
746 #define _REGFORM_q5(decor) _DECOR(q, decor, 5)
747 #define _REGFORM_q6(decor) _DECOR(q, decor, 6)
748 #define _REGFORM_q7(decor) _DECOR(q, decor, 7)
749 #define _REGFORM_q8(decor) _DECOR(q, decor, 8)
750 #define _REGFORM_q9(decor) _DECOR(q, decor, 9)
751 #define _REGFORM_q10(decor) _DECOR(q, decor, 10)
752 #define _REGFORM_q11(decor) _DECOR(q, decor, 11)
753 #define _REGFORM_q12(decor) _DECOR(q, decor, 12)
754 #define _REGFORM_q13(decor) _DECOR(q, decor, 13)
755 #define _REGFORM_q14(decor) _DECOR(q, decor, 14)
756 #define _REGFORM_q15(decor) _DECOR(q, decor, 15)
757
758 // `_LOPART(n)' and `_HIPART(n)' return the numbers of the register halves of
759 // register n, i.e., 2*n and 2*n + 1 respectively.
760 #define _LOPART(n) _GLUE(_LOPART_, n)
761 #define _HIPART(n) _GLUE(_HIPART_, n)
762
763 // Internal macros: `_LOPART_n' and `_HIPART_n' return the numbers of the
764 // register halves of register n, i.e., 2*n and 2*n + 1 respectively.
765
766 #define _LOPART_0 0
767 #define _HIPART_0 1
768 #define _LOPART_1 2
769 #define _HIPART_1 3
770 #define _LOPART_2 4
771 #define _HIPART_2 5
772 #define _LOPART_3 6
773 #define _HIPART_3 7
774 #define _LOPART_4 8
775 #define _HIPART_4 9
776 #define _LOPART_5 10
777 #define _HIPART_5 11
778 #define _LOPART_6 12
779 #define _HIPART_6 13
780 #define _LOPART_7 14
781 #define _HIPART_7 15
782 #define _LOPART_8 16
783 #define _HIPART_8 17
784 #define _LOPART_9 18
785 #define _HIPART_9 19
786 #define _LOPART_10 20
787 #define _HIPART_10 21
788 #define _LOPART_11 22
789 #define _HIPART_11 23
790 #define _LOPART_12 24
791 #define _HIPART_12 25
792 #define _LOPART_13 26
793 #define _HIPART_13 27
794 #define _LOPART_14 28
795 #define _HIPART_14 29
796 #define _LOPART_15 30
797 #define _HIPART_15 31
798
799 // Return the register number of the pair containing register n, i.e.,
800 // floor(n/2).
801 #define _PAIR(n) _GLUE(_PAIR_, n)
802
803 // Internal macros: `_PAIR_n' returns the register number of the pair
804 // containing register n, i.e., floor(n/2).
805 #define _PAIR_0 0
806 #define _PAIR_1 0
807 #define _PAIR_2 1
808 #define _PAIR_3 1
809 #define _PAIR_4 2
810 #define _PAIR_5 2
811 #define _PAIR_6 3
812 #define _PAIR_7 3
813 #define _PAIR_8 4
814 #define _PAIR_9 4
815 #define _PAIR_10 5
816 #define _PAIR_11 5
817 #define _PAIR_12 6
818 #define _PAIR_13 6
819 #define _PAIR_14 7
820 #define _PAIR_15 7
821 #define _PAIR_16 8
822 #define _PAIR_17 8
823 #define _PAIR_18 9
824 #define _PAIR_19 9
825 #define _PAIR_20 10
826 #define _PAIR_21 10
827 #define _PAIR_22 11
828 #define _PAIR_23 11
829 #define _PAIR_24 12
830 #define _PAIR_25 12
831 #define _PAIR_26 13
832 #define _PAIR_27 13
833 #define _PAIR_28 14
834 #define _PAIR_29 14
835 #define _PAIR_30 15
836 #define _PAIR_31 15
837
838 // Apply decoration decor to register number n of type ty. Decorations are
839 // as follows.
840 //
841 // decor types meaning
842 // Q s, d the NEON qN register containing this one
843 // D s the NEON dN register containing this one
844 // D0 q the low 64-bit half of this one
845 // D1 q the high 64-bit half of this one
846 // S0 d, q the first 32-bit piece of this one
847 // S1 d, q the second 32-bit piece of this one
848 // S2 q the third 32-bit piece of this one
849 // S3 q the fourth 32-bit piece of this one
850 // Bn q the nth byte of this register, as a scalar
851 // Hn q the nth halfword of this register, as a scalar
852 // Wn q the nth word of this register, as a scalar
853 #define _DECOR(ty, decor, n) _DECOR_##ty##_##decor(n)
854
855 // Internal macros: `_DECOR_ty_decor(n)' applies decoration decor to register
856 // number n of type ty.
857
858 #define _DECOR_s_Q(n) GLUE(q, _PAIR(_PAIR(n)))
859 #define _DECOR_s_D(n) GLUE(d, _PAIR(n))
860
861 #define _DECOR_d_Q(n) GLUE(q, _PAIR(n))
862 #define _DECOR_d_S0(n) GLUE(s, _LOPART(n))
863 #define _DECOR_d_S1(n) GLUE(s, _LOPART(n))
864
865 #define _DECOR_q_D0(n) GLUE(d, _LOPART(n))
866 #define _DECOR_q_D1(n) GLUE(d, _HIPART(n))
867 #define _DECOR_q_S0(n) GLUE(s, _LOPART(_LOPART(n)))
868 #define _DECOR_q_S1(n) GLUE(s, _HIPART(_LOPART(n)))
869 #define _DECOR_q_S2(n) GLUE(s, _LOPART(_HIPART(n)))
870 #define _DECOR_q_S3(n) GLUE(s, _HIPART(_HIPART(n)))
871 #define _DECOR_q_W0(n) GLUE(d, _LOPART(n))[0]
872 #define _DECOR_q_W1(n) GLUE(d, _LOPART(n))[1]
873 #define _DECOR_q_W2(n) GLUE(d, _HIPART(n))[0]
874 #define _DECOR_q_W3(n) GLUE(d, _HIPART(n))[1]
875 #define _DECOR_q_H0(n) GLUE(d, _LOPART(n))[0]
876 #define _DECOR_q_H1(n) GLUE(d, _LOPART(n))[1]
877 #define _DECOR_q_H2(n) GLUE(d, _LOPART(n))[2]
878 #define _DECOR_q_H3(n) GLUE(d, _LOPART(n))[3]
879 #define _DECOR_q_H4(n) GLUE(d, _HIPART(n))[0]
880 #define _DECOR_q_H5(n) GLUE(d, _HIPART(n))[1]
881 #define _DECOR_q_H6(n) GLUE(d, _HIPART(n))[2]
882 #define _DECOR_q_H7(n) GLUE(d, _HIPART(n))[3]
883 #define _DECOR_q_B0(n) GLUE(d, _LOPART(n))[0]
884 #define _DECOR_q_B1(n) GLUE(d, _LOPART(n))[1]
885 #define _DECOR_q_B2(n) GLUE(d, _LOPART(n))[2]
886 #define _DECOR_q_B3(n) GLUE(d, _LOPART(n))[3]
887 #define _DECOR_q_B4(n) GLUE(d, _LOPART(n))[4]
888 #define _DECOR_q_B5(n) GLUE(d, _LOPART(n))[5]
889 #define _DECOR_q_B6(n) GLUE(d, _LOPART(n))[6]
890 #define _DECOR_q_B7(n) GLUE(d, _LOPART(n))[7]
891 #define _DECOR_q_B8(n) GLUE(d, _HIPART(n))[0]
892 #define _DECOR_q_B9(n) GLUE(d, _HIPART(n))[1]
893 #define _DECOR_q_B10(n) GLUE(d, _HIPART(n))[2]
894 #define _DECOR_q_B11(n) GLUE(d, _HIPART(n))[3]
895 #define _DECOR_q_B12(n) GLUE(d, _HIPART(n))[4]
896 #define _DECOR_q_B13(n) GLUE(d, _HIPART(n))[5]
897 #define _DECOR_q_B14(n) GLUE(d, _HIPART(n))[6]
898 #define _DECOR_q_B15(n) GLUE(d, _HIPART(n))[7]
899
900 // Macros for navigating the NEON register hierarchy.
901 #define S0(reg) _REGFORM(reg, S0)
902 #define S1(reg) _REGFORM(reg, S1)
903 #define S2(reg) _REGFORM(reg, S2)
904 #define S3(reg) _REGFORM(reg, S3)
905 #define D(reg) _REGFORM(reg, D)
906 #define D0(reg) _REGFORM(reg, D0)
907 #define D1(reg) _REGFORM(reg, D1)
908 #define Q(reg) _REGFORM(reg, Q)
909
910 // Macros for indexing quadword registers.
911 #define QB(reg, i) _REGFORM(reg, B##i)
912 #define QH(reg, i) _REGFORM(reg, H##i)
913 #define QW(reg, i) _REGFORM(reg, W##i)
914
915 // Macros for converting vldm/vstm ranges.
916 #define QQ(qlo, qhi) D0(qlo)-D1(qhi)
917
918 // Stack management and unwinding.
919 .macro setfp fp=r11, offset=0
920 .if \offset == 0
921 mov \fp, sp
922 .setfp \fp, sp
923 .else
924 add \fp, sp, #\offset
925 .setfp \fp, sp, #\offset
926 .endif
927 .macro dropfp; _dropfp \fp, \offset; .endm
928 .L$_frameptr_p = -1
929 .endm
930
931 .macro _dropfp fp, offset=0
932 .if \offset == 0
933 mov sp, \fp
934 .else
935 sub sp, \fp, #\offset
936 .endif
937 .purgem dropfp
938 .L$_frameptr_p = 0
939 .endm
940
941 .macro stalloc n
942 sub sp, sp, #\n
943 .pad #\n
944 .endm
945
946 .macro stfree n
947 add sp, sp, #\n
948 .pad #-\n
949 .endm
950
951 .macro pushreg rr:vararg
952 push {\rr}
953 .save {\rr}
954 .endm
955
956 .macro popreg rr:vararg
957 pop {\rr}
958 .endm
959
960 .macro pushvfp rr:vararg
961 vstmdb sp!, {\rr}
962 .vsave {\rr}
963 .endm
964
965 .macro popvfp rr:vararg
966 vldmia sp!, {\rr}
967 .endm
968
969 .macro endprologue
970 .endm
971
972 // No need for prologue markers on ARM.
973 #define FUNC_POSTHOOK(_) .L$_prologue_p = -1
974
975 #endif
976
977 ///--------------------------------------------------------------------------
978 /// AArch64-specific hacking.
979
980 #if CPUFAM_ARM64
981
982 // Set the function hooks.
983 #define FUNC_PREHOOK(_) .balign 4
984 #define FUNC_POSTHOOK(_) .cfi_startproc; .L$_prologue_p = -1
985 #define ENDFUNC_HOOK(_) .cfi_endproc
986
987 // Call external subroutine at ADDR, possibly via PLT.
988 .macro callext addr
989 bl \addr
990 .endm
991
992 // Load address of external symbol ADDR into REG.
993 .macro leaext reg, addr
994 #if WANT_PIC
995 adrp \reg, :got:\addr
996 ldr \reg, [\reg, #:got_lo12:\addr]
997 #else
998 adrp \reg, \addr
999 add \reg, \reg, #:lo12:\addr
1000 #endif
1001 .endm
1002
1003 .macro vzero vz=v31
1004 // Set VZ (default v31) to zero.
1005 dup \vz\().4s, wzr
1006 .endm
1007
1008 .macro vshl128 vd, vn, nbit, vz=v31
1009 // Set VD to VN shifted left by NBIT. Assume VZ (default v31) is
1010 // all-bits-zero. NBIT must be a multiple of 8.
1011 .if \nbit&3 != 0
1012 .error "shift quantity must be whole number of bytes"
1013 .endif
1014 ext \vd\().16b, \vz\().16b, \vn\().16b, #16 - (\nbit >> 3)
1015 .endm
1016
1017 .macro vshr128 vd, vn, nbit, vz=v31
1018 // Set VD to VN shifted right by NBIT. Assume VZ (default v31) is
1019 // all-bits-zero. NBIT must be a multiple of 8.
1020 .if \nbit&3 != 0
1021 .error "shift quantity must be whole number of bytes"
1022 .endif
1023 ext \vd\().16b, \vn\().16b, \vz\().16b, #\nbit >> 3
1024 .endm
1025
1026 // Register class conversions.
1027 #define _GPNUM_w0 0
1028 #define _GPNUM_w1 1
1029 #define _GPNUM_w2 2
1030 #define _GPNUM_w3 3
1031 #define _GPNUM_w4 4
1032 #define _GPNUM_w5 5
1033 #define _GPNUM_w6 6
1034 #define _GPNUM_w7 7
1035 #define _GPNUM_w8 8
1036 #define _GPNUM_w9 9
1037 #define _GPNUM_w10 10
1038 #define _GPNUM_w11 11
1039 #define _GPNUM_w12 12
1040 #define _GPNUM_w13 13
1041 #define _GPNUM_w14 14
1042 #define _GPNUM_w15 15
1043 #define _GPNUM_w16 16
1044 #define _GPNUM_w17 17
1045 #define _GPNUM_w18 18
1046 #define _GPNUM_w19 19
1047 #define _GPNUM_w20 20
1048 #define _GPNUM_w21 21
1049 #define _GPNUM_w22 22
1050 #define _GPNUM_w23 23
1051 #define _GPNUM_w24 24
1052 #define _GPNUM_w25 25
1053 #define _GPNUM_w26 26
1054 #define _GPNUM_w27 27
1055 #define _GPNUM_w28 28
1056 #define _GPNUM_w29 29
1057 #define _GPNUM_w30 30
1058 #define _GPNUM_wzr zr
1059 #define _GPNUM_wsp sp
1060
1061 #define _GPNUM_x0 0
1062 #define _GPNUM_x1 1
1063 #define _GPNUM_x2 2
1064 #define _GPNUM_x3 3
1065 #define _GPNUM_x4 4
1066 #define _GPNUM_x5 5
1067 #define _GPNUM_x6 6
1068 #define _GPNUM_x7 7
1069 #define _GPNUM_x8 8
1070 #define _GPNUM_x9 9
1071 #define _GPNUM_x10 10
1072 #define _GPNUM_x11 11
1073 #define _GPNUM_x12 12
1074 #define _GPNUM_x13 13
1075 #define _GPNUM_x14 14
1076 #define _GPNUM_x15 15
1077 #define _GPNUM_x16 16
1078 #define _GPNUM_x17 17
1079 #define _GPNUM_x18 18
1080 #define _GPNUM_x19 19
1081 #define _GPNUM_x20 20
1082 #define _GPNUM_x21 21
1083 #define _GPNUM_x22 22
1084 #define _GPNUM_x23 23
1085 #define _GPNUM_x24 24
1086 #define _GPNUM_x25 25
1087 #define _GPNUM_x26 26
1088 #define _GPNUM_x27 27
1089 #define _GPNUM_x28 28
1090 #define _GPNUM_x29 29
1091 #define _GPNUM_x30 30
1092 #define _GPNUM_xzr zr
1093 #define _GPNUM_sp sp
1094 #define _GPNUM_xsp sp
1095 #define xsp sp
1096
1097 #define _VNUM_b0 0
1098 #define _VNUM_b1 1
1099 #define _VNUM_b2 2
1100 #define _VNUM_b3 3
1101 #define _VNUM_b4 4
1102 #define _VNUM_b5 5
1103 #define _VNUM_b6 6
1104 #define _VNUM_b7 7
1105 #define _VNUM_b8 8
1106 #define _VNUM_b9 9
1107 #define _VNUM_b10 10
1108 #define _VNUM_b11 11
1109 #define _VNUM_b12 12
1110 #define _VNUM_b13 13
1111 #define _VNUM_b14 14
1112 #define _VNUM_b15 15
1113 #define _VNUM_b16 16
1114 #define _VNUM_b17 17
1115 #define _VNUM_b18 18
1116 #define _VNUM_b19 19
1117 #define _VNUM_b20 20
1118 #define _VNUM_b21 21
1119 #define _VNUM_b22 22
1120 #define _VNUM_b23 23
1121 #define _VNUM_b24 24
1122 #define _VNUM_b25 25
1123 #define _VNUM_b26 26
1124 #define _VNUM_b27 27
1125 #define _VNUM_b28 28
1126 #define _VNUM_b29 29
1127 #define _VNUM_b30 30
1128 #define _VNUM_b31 31
1129
1130 #define _VNUM_h0 0
1131 #define _VNUM_h1 1
1132 #define _VNUM_h2 2
1133 #define _VNUM_h3 3
1134 #define _VNUM_h4 4
1135 #define _VNUM_h5 5
1136 #define _VNUM_h6 6
1137 #define _VNUM_h7 7
1138 #define _VNUM_h8 8
1139 #define _VNUM_h9 9
1140 #define _VNUM_h10 10
1141 #define _VNUM_h11 11
1142 #define _VNUM_h12 12
1143 #define _VNUM_h13 13
1144 #define _VNUM_h14 14
1145 #define _VNUM_h15 15
1146 #define _VNUM_h16 16
1147 #define _VNUM_h17 17
1148 #define _VNUM_h18 18
1149 #define _VNUM_h19 19
1150 #define _VNUM_h20 20
1151 #define _VNUM_h21 21
1152 #define _VNUM_h22 22
1153 #define _VNUM_h23 23
1154 #define _VNUM_h24 24
1155 #define _VNUM_h25 25
1156 #define _VNUM_h26 26
1157 #define _VNUM_h27 27
1158 #define _VNUM_h28 28
1159 #define _VNUM_h29 29
1160 #define _VNUM_h30 30
1161 #define _VNUM_h31 31
1162
1163 #define _VNUM_s0 0
1164 #define _VNUM_s1 1
1165 #define _VNUM_s2 2
1166 #define _VNUM_s3 3
1167 #define _VNUM_s4 4
1168 #define _VNUM_s5 5
1169 #define _VNUM_s6 6
1170 #define _VNUM_s7 7
1171 #define _VNUM_s8 8
1172 #define _VNUM_s9 9
1173 #define _VNUM_s10 10
1174 #define _VNUM_s11 11
1175 #define _VNUM_s12 12
1176 #define _VNUM_s13 13
1177 #define _VNUM_s14 14
1178 #define _VNUM_s15 15
1179 #define _VNUM_s16 16
1180 #define _VNUM_s17 17
1181 #define _VNUM_s18 18
1182 #define _VNUM_s19 19
1183 #define _VNUM_s20 20
1184 #define _VNUM_s21 21
1185 #define _VNUM_s22 22
1186 #define _VNUM_s23 23
1187 #define _VNUM_s24 24
1188 #define _VNUM_s25 25
1189 #define _VNUM_s26 26
1190 #define _VNUM_s27 27
1191 #define _VNUM_s28 28
1192 #define _VNUM_s29 29
1193 #define _VNUM_s30 30
1194 #define _VNUM_s31 31
1195
1196 #define _VNUM_d0 0
1197 #define _VNUM_d1 1
1198 #define _VNUM_d2 2
1199 #define _VNUM_d3 3
1200 #define _VNUM_d4 4
1201 #define _VNUM_d5 5
1202 #define _VNUM_d6 6
1203 #define _VNUM_d7 7
1204 #define _VNUM_d8 8
1205 #define _VNUM_d9 9
1206 #define _VNUM_d10 10
1207 #define _VNUM_d11 11
1208 #define _VNUM_d12 12
1209 #define _VNUM_d13 13
1210 #define _VNUM_d14 14
1211 #define _VNUM_d15 15
1212 #define _VNUM_d16 16
1213 #define _VNUM_d17 17
1214 #define _VNUM_d18 18
1215 #define _VNUM_d19 19
1216 #define _VNUM_d20 20
1217 #define _VNUM_d21 21
1218 #define _VNUM_d22 22
1219 #define _VNUM_d23 23
1220 #define _VNUM_d24 24
1221 #define _VNUM_d25 25
1222 #define _VNUM_d26 26
1223 #define _VNUM_d27 27
1224 #define _VNUM_d28 28
1225 #define _VNUM_d29 29
1226 #define _VNUM_d30 30
1227 #define _VNUM_d31 31
1228
1229 #define _VNUM_q0 0
1230 #define _VNUM_q1 1
1231 #define _VNUM_q2 2
1232 #define _VNUM_q3 3
1233 #define _VNUM_q4 4
1234 #define _VNUM_q5 5
1235 #define _VNUM_q6 6
1236 #define _VNUM_q7 7
1237 #define _VNUM_q8 8
1238 #define _VNUM_q9 9
1239 #define _VNUM_q10 10
1240 #define _VNUM_q11 11
1241 #define _VNUM_q12 12
1242 #define _VNUM_q13 13
1243 #define _VNUM_q14 14
1244 #define _VNUM_q15 15
1245 #define _VNUM_q16 16
1246 #define _VNUM_q17 17
1247 #define _VNUM_q18 18
1248 #define _VNUM_q19 19
1249 #define _VNUM_q20 20
1250 #define _VNUM_q21 21
1251 #define _VNUM_q22 22
1252 #define _VNUM_q23 23
1253 #define _VNUM_q24 24
1254 #define _VNUM_q25 25
1255 #define _VNUM_q26 26
1256 #define _VNUM_q27 27
1257 #define _VNUM_q28 28
1258 #define _VNUM_q29 29
1259 #define _VNUM_q30 30
1260 #define _VNUM_q31 31
1261
1262 #define _VNUM_v0 0
1263 #define _VNUM_v1 1
1264 #define _VNUM_v2 2
1265 #define _VNUM_v3 3
1266 #define _VNUM_v4 4
1267 #define _VNUM_v5 5
1268 #define _VNUM_v6 6
1269 #define _VNUM_v7 7
1270 #define _VNUM_v8 8
1271 #define _VNUM_v9 9
1272 #define _VNUM_v10 10
1273 #define _VNUM_v11 11
1274 #define _VNUM_v12 12
1275 #define _VNUM_v13 13
1276 #define _VNUM_v14 14
1277 #define _VNUM_v15 15
1278 #define _VNUM_v16 16
1279 #define _VNUM_v17 17
1280 #define _VNUM_v18 18
1281 #define _VNUM_v19 19
1282 #define _VNUM_v20 20
1283 #define _VNUM_v21 21
1284 #define _VNUM_v22 22
1285 #define _VNUM_v23 23
1286 #define _VNUM_v24 24
1287 #define _VNUM_v25 25
1288 #define _VNUM_v26 26
1289 #define _VNUM_v27 27
1290 #define _VNUM_v28 28
1291 #define _VNUM_v29 29
1292 #define _VNUM_v30 30
1293 #define _VNUM_v31 31
1294
1295 #define _VNUM_z0 0
1296 #define _VNUM_z1 1
1297 #define _VNUM_z2 2
1298 #define _VNUM_z3 3
1299 #define _VNUM_z4 4
1300 #define _VNUM_z5 5
1301 #define _VNUM_z6 6
1302 #define _VNUM_z7 7
1303 #define _VNUM_z8 8
1304 #define _VNUM_z9 9
1305 #define _VNUM_z10 10
1306 #define _VNUM_z11 11
1307 #define _VNUM_z12 12
1308 #define _VNUM_z13 13
1309 #define _VNUM_z14 14
1310 #define _VNUM_z15 15
1311 #define _VNUM_z16 16
1312 #define _VNUM_z17 17
1313 #define _VNUM_z18 18
1314 #define _VNUM_z19 19
1315 #define _VNUM_z20 20
1316 #define _VNUM_z21 21
1317 #define _VNUM_z22 22
1318 #define _VNUM_z23 23
1319 #define _VNUM_z24 24
1320 #define _VNUM_z25 25
1321 #define _VNUM_z26 26
1322 #define _VNUM_z27 27
1323 #define _VNUM_z28 28
1324 #define _VNUM_z29 29
1325 #define _VNUM_z30 30
1326 #define _VNUM_z31 31
1327
1328 #define _RDECOR(cls, pre, r) GLUE(pre, _##cls##NUM_##r)
1329 #define W(r) _RDECOR(GP, w, r)
1330 #define X(r) _RDECOR(GP, x, r)
1331 #define B(r) _RDECOR(V, b, r)
1332 #define H(r) _RDECOR(V, h, r)
1333 #define S(r) _RDECOR(V, s, r)
1334 #define D(r) _RDECOR(V, d, r)
1335 #define Q(r) _RDECOR(V, q, r)
1336 #define V(r) _RDECOR(V, v, r)
1337 #define Z(r) _RDECOR(V, z, r)
1338
1339 // Stack management and unwinding.
1340 .macro setfp fp=x29, offset=0
1341 // If you're just going through the motions with a fixed-size stack frame,
1342 // then you want to say `add x29, sp, #OFFSET' directly, which will avoid
1343 // pointlessly restoring sp later.
1344 .if \offset == 0
1345 mov \fp, sp
1346 .cfi_def_cfa_register \fp
1347 .else
1348 add \fp, sp, #\offset
1349 .cfi_def_cfa_register \fp
1350 .cfi_adjust_cfa_offset -\offset
1351 .endif
1352 .macro dropfp; _dropfp \fp, \offset; .endm
1353 .L$_frameptr_p = -1
1354 .endm
1355
1356 .macro _dropfp fp, offset=0
1357 .if \offset == 0
1358 mov sp, \fp
1359 .cfi_def_cfa_register sp
1360 .else
1361 sub sp, \fp, #\offset
1362 .cfi_def_cfa_register sp
1363 .cfi_adjust_cfa_offset +\offset
1364 .endif
1365 .purgem dropfp
1366 .L$_frameptr_p = 0
1367 .endm
1368
1369 .macro stalloc n
1370 sub sp, sp, #\n
1371 .cfi_adjust_cfa_offset +\n
1372 .endm
1373
1374 .macro stfree n
1375 add sp, sp, #\n
1376 .cfi_adjust_cfa_offset -\n
1377 .endm
1378
1379 .macro pushreg x, y=nil
1380 .ifeqs "\y", "nil"
1381 str \x, [sp, #-16]!
1382 .cfi_adjust_cfa_offset +16
1383 .cfi_rel_offset \x, 0
1384 .else
1385 stp \x, \y, [sp, #-16]!
1386 .cfi_adjust_cfa_offset +16
1387 .cfi_rel_offset \x, 0
1388 .cfi_rel_offset \y, 8
1389 .endif
1390 .endm
1391
1392 .macro popreg x, y=nil
1393 .ifeqs "\y", "nil"
1394 ldr \x, [sp], #16
1395 .cfi_restore \x
1396 .cfi_adjust_cfa_offset -16
1397 .else
1398 ldp \x, \y, [sp], #16
1399 .cfi_restore \x
1400 .cfi_restore \y
1401 .cfi_adjust_cfa_offset -16
1402 .endif
1403 .endm
1404
1405 .macro savereg x, y, z=nil
1406 .ifeqs "\z", "nil"
1407 str \x, [sp, \y]
1408 .cfi_rel_offset \x, \y
1409 .else
1410 stp \x, \y, [sp, #\z]
1411 .cfi_rel_offset \x, \z
1412 .cfi_rel_offset \y, \z + 8
1413 .endif
1414 .endm
1415
1416 .macro rstrreg x, y, z=nil
1417 .ifeqs "\z", "nil"
1418 ldr \x, [sp, \y]
1419 .cfi_restore \x
1420 .else
1421 ldp \x, \y, [sp, #\z]
1422 .cfi_restore \x
1423 .cfi_restore \y
1424 .endif
1425 .endm
1426
1427 .macro endprologue
1428 .endm
1429
1430 // cmov RD, RN, CC: set RD to RN if CC is satisfied, otherwise do nothing
1431 .macro cmov rd, rn, cc
1432 csel \rd, \rn, \rd, \cc
1433 .endm
1434
1435 // Notational improvement: write `csel.CC' etc., rather than `csel ..., CC'.
1436 #define _COND(_) \
1437 _(eq) _(ne) _(cs) _(cc) _(vs) _(vc) _(mi) _(pl) \
1438 _(ge) _(lt) _(gt) _(le) _(hi) _(ls) _(al) _(nv) \
1439 _(hs) _(lo)
1440 #define _INST(_) \
1441 _(ccmp) _(ccmn) \
1442 _(csel) _(cmov) \
1443 _(csinc) _(cinc) _(cset) \
1444 _(csneg) _(cneg) \
1445 _(csinv) _(cinv) _(csetm)
1446 #define _CONDVAR(cc) _definstvar cc;
1447 #define _INSTVARS(inst) \
1448 .macro _definstvar cc; \
1449 .macro inst.\cc args:vararg; inst \args, \cc; .endm; \
1450 .endm; \
1451 _COND(_CONDVAR); \
1452 .purgem _definstvar;
1453 _INST(_INSTVARS)
1454 #undef _COND
1455 #undef _INST
1456 #undef _CONDVAR
1457 #undef _INSTVARS
1458
1459 // Flag bits for `ccmp' and friends.
1460 #define CCMP_N 8
1461 #define CCMP_Z 4
1462 #define CCMP_C 2
1463 #define CCMP_V 1
1464
1465 // Flag settings for satisfying conditions.
1466 #define CCMP_MI CCMP_N
1467 #define CCMP_PL 0
1468 #define CCMP_EQ CCMP_Z
1469 #define CCMP_NE 0
1470 #define CCMP_CS CCMP_C
1471 #define CCMP_HS CCMP_C
1472 #define CCMP_CC 0
1473 #define CCMP_LO 0
1474 #define CCMP_VS CCMP_V
1475 #define CCMP_VC 0
1476 #define CCMP_HI CCMP_C
1477 #define CCMP_LS 0
1478 #define CCMP_LT CCMP_N
1479 #define CCMP_GE 0
1480 #define CCMP_LE CCMP_N
1481 #define CCMP_GT 0
1482
1483 #endif
1484
1485 ///--------------------------------------------------------------------------
1486 /// Final stuff.
1487
1488 // Default values for the various hooks.
1489 #ifndef FUNC_PREHOOK
1490 # define FUNC_PREHOOK(_)
1491 #endif
1492 #ifndef FUNC_POSTHOOK
1493 # define FUNC_POSTHOOK(_)
1494 #endif
1495 #ifndef ENDFUNC_HOOK
1496 # define ENDFUNC_HOOK(_)
1497 #endif
1498
1499 #ifndef F
1500 # ifdef SYM_USCORE
1501 # define F(name) _##name
1502 # else
1503 # define F(name) name
1504 # endif
1505 #endif
1506
1507 #ifndef TYPE_FUNC
1508 # define TYPE_FUNC(name)
1509 #endif
1510
1511 #ifndef SIZE_OBJ
1512 # define SIZE_OBJ(name)
1513 #endif
1514
1515 #if __ELF__ && !defined(WANT_EXECUTABLE_STACK)
1516 .pushsection .note.GNU-stack, "", _SECTTY(progbits)
1517 .popsection
1518 #endif
1519
1520 ///----- That's all, folks --------------------------------------------------
1521
1522 #endif
Catacomb cryptographic library