kris
/
gcm-poc
1
0
Atdalīts 0
Kods Problēmas 0 Izmaiņu pieprasījumi 0 Laidieni 0 Vikivietne Aktivitāte
Nevar pievienot vairāk kā 25 tēmas Tēmai ir jāsākas ar burtu vai ciparu, tā var saturēt domu zīmes ('-') un var būt līdz 35 simboliem gara.
1 Revīzija
1 Atzars
54 KiB
 
 
Koks: 40223b1828
Atzari Tagi
${ item.name }
Izveidot atzaru ${ searchTerm }
no '40223b1828'
${ noResults }
gcm-poc/asm-common.h

1181 rinda
30 KiB
Neapstrādāts Vainot Vēsture 

  1. ///--------------------------------------------------------------------------

  2. /// General definitions.

  3. 

  4. // Preprocessor hacks.

  5. #define STRINGY(x) _STRINGY(x, y)

  6. #define _STRINGY(x) #x

  7. #define GLUE(x, y) _GLUE(x, y)

  8. #define _GLUE(x, y) x##y

  9. #define _EMPTY

  10. 

  11. // Some useful variables.

  12. 	.L$_subsec = 0

  13. 

  14. // Literal pools done the hard way.

  15. #define _LIT .text .L$_subsec + 1

  16. #define _ENDLIT .text .L$_subsec

  17. #define _LTORG .L$_subsec = .L$_subsec + 2; .text .L$_subsec

  18. 

  19. // ELF section types.

  20. #if __ELF__

  21. #  if CPUFAM_ARMEL

  22. #    define _SECTTY(ty) %ty

  23. #  else

  24. #    define _SECTTY(ty) @ty

  25. #  endif

  26. #endif

  27. 

  28. // Section selection.

  29. #define TEXT .text .L$_subsec

  30. #if ABI_WIN

  31. #  define RODATA .section .rdata, "dr"

  32. #elif __ELF__

  33. #  define RODATA .section .rodata, "a", _SECTTY(progbits)

  34. #else

  35. #  define RODATA TEXT

  36. #endif

  37. #define DATA .data

  38. 

  39. // Announcing an internal function.

  40. #define INTFUNC(name)							\

  41. 	TYPE_FUNC(name);						\

  42. 	.macro ENDFUNC; _ENDFUNC(name); .endm;				\

  43. 	.L$_prologue_p = 0; .L$_frameptr_p = 0;				\

  44. 	FUNC_PREHOOK(name);						\

  45. name:									\

  46. 	FUNC_POSTHOOK(name)

  47. 

  48. // Announcing an external function.

  49. #define FUNC(name)							\

  50. 	.globl	F(name);						\

  51. INTFUNC(F(name))

  52. 

  53. // Marking the end of a function.

  54. #define _ENDFUNC(name)							\

  55. 	.if ~ .L$_prologue_p; .error "Missing `endprologue'"; .endif;	\

  56. 	.if .L$_frameptr_p; .purgem dropfp; .endif;			\

  57. 	.purgem	ENDFUNC;						\

  58. 	SIZE_OBJ(name);							\

  59. 	ENDFUNC_HOOK(name);						\

  60. 	_LTORG

  61. 

  62. // Make a helper function, if necessary.

  63. #define AUXFN(name)							\

  64.   .ifndef .L$_auxfn_def.name;						\

  65. 	.text 7128;							\

  66. 	.macro _ENDAUXFN; _ENDAUXFN_TAIL(name); .endm;			\

  67. 	FUNC_PREHOOK(name);						\

  68. name:

  69. #define _ENDAUXFN_TAIL(name)						\

  70. 	.purgem _ENDAUXFN;						\

  71. 	.text .L$_subsec;						\

  72. 	.L$_auxfn_def.name = 1

  73. #define ENDAUXFN _ENDAUXFN; .endif

  74. 

  75. ///--------------------------------------------------------------------------

  76. /// ELF-specific hacking.

  77. 

  78. #if __ELF__

  79. 

  80. #if __PIC__ || __PIE__

  81. #  define WANT_PIC 1

  82. #endif

  83. 

  84. #define TYPE_FUNC(name) .type name, STT_FUNC

  85. 

  86. #define SIZE_OBJ(name) .size name, . - name

  87. 

  88. #endif

  89. 

  90. ///--------------------------------------------------------------------------

  91. /// Windows-specific hacking.

  92. 

  93. #if ABI_WIN

  94. 

  95. #if CPUFAM_X86

  96. #  define F(name) _##name

  97. #endif

  98. 

  99. #endif

  100. 

  101. ///--------------------------------------------------------------------------

  102. /// x86- and amd64-specific hacking.

  103. ///

  104. /// It's (slightly) easier to deal with both of these in one go.

  105. 

  106. #if CPUFAM_X86 || CPUFAM_AMD64

  107. 

  108. // Word size.

  109. #if CPUFAM_X86

  110. #  define WORDSZ 4

  111. #endif

  112. #if CPUFAM_AMD64

  113. #  define WORDSZ 8

  114. #endif

  115. 

  116. // Set the function hooks.

  117. #define FUNC_PREHOOK(_) .balign 16

  118. 

  119. // On Windows, arrange to install stack-unwinding data.

  120. #if CPUFAM_AMD64 && ABI_WIN

  121. #  define FUNC_POSTHOOK(name) .seh_proc name

  122. #  define ENDFUNC_HOOK(_) .seh_endproc

  123. // Procedures are expected to invoke `.seh_setframe' if necessary, and

  124. // `.seh_pushreg' and friends, and `.seh_endprologue'.

  125. #endif

  126. 

  127. #if __ELF__

  128. #  define FUNC_POSTHOOK(_) .cfi_startproc

  129. #  define ENDFUNC_HOOK(_) .cfi_endproc

  130. #endif

  131. 

  132. // Don't use the wretched AT&T syntax.  It's festooned with pointless

  133. // punctuation, and all of the data movement is backwards.  Ugh!

  134. 	.intel_syntax noprefix

  135. 

  136. // Call external subroutine at ADDR, possibly via PLT.

  137. .macro	callext addr

  138. #if WANT_PIC

  139. 	call	\addr@PLT

  140. #else

  141. 	call	\addr

  142. #endif

  143. .endm

  144. 

  145. // Do I need to arrange a spare GOT register?

  146. #if WANT_PIC && CPUFAM_X86

  147. #  define NEED_GOT 1

  148. #endif

  149. #define GOTREG ebx			// Not needed in AMD64 so don't care.

  150. 

  151. // Maybe load GOT address into GOT.

  152. .macro	ldgot got=GOTREG

  153. #if WANT_PIC && CPUFAM_X86

  154.   AUXFN(_ldgot.\got)

  155. 	mov	\got, [esp]

  156. 	ret

  157.   ENDAUXFN

  158. 	call	_ldgot.\got

  159. 	add	\got, offset _GLOBAL_OFFSET_TABLE_

  160. #endif

  161. .endm

  162. 

  163. // Load address of external symbol ADDR into REG, maybe using GOT.

  164. .macro	leaext reg, addr, got=GOTREG

  165. #if WANT_PIC

  166. #  if CPUFAM_X86

  167. 	mov	\reg, [\got + \addr@GOT]

  168. #  endif

  169. #  if CPUFAM_AMD64

  170. 	mov	\reg, \addr@GOTPCREL[rip]

  171. #  endif

  172. #else

  173. #  if CPUFAM_X86

  174. 	mov	\reg, offset \addr

  175. #  endif

  176. #  if CPUFAM_AMD64

  177. 	lea	\reg, \addr[rip]

  178. #  endif

  179. #endif

  180. .endm

  181. 

  182. // Address expression (possibly using a base register, and a displacement)

  183. // referring to ADDR, which is within our module, maybe using GOT.

  184. #define INTADDR(...) INTADDR__0(__VA_ARGS__, GOTREG, dummy)

  185. #define INTADDR__0(addr, got, ...) INTADDR__1(addr, got)

  186. #if CPUFAM_AMD64

  187. #  define INTADDR__1(addr, got) addr + rip

  188. #elif WANT_PIC

  189. #  define INTADDR__1(addr, got) got + addr@GOTOFF

  190. #else

  191. #  define INTADDR__1(addr, got) addr

  192. #endif

  193. 

  194. // Permutations for SIMD instructions.  SHUF(A, B, C, D) is an immediate,

  195. // suitable for use in `pshufd' or `shufpd', which copies element A

  196. // (0 <= A < 4) of the source to element 0 of the destination, element B to

  197. // element 1, element C to element 2, and element D to element 3.

  198. #define SHUF(a, b, c, d) ((a) + 4*(b) + 16*(c) + 64*(d))

  199. 

  200. // Map register names to their individual pieces.

  201. 

  202. // Apply decoration decor to (internal) register name reg of type ty.

  203. //

  204. // See `R_...' for internal register names.  Decorations are as follows.

  205. //

  206. //	b	low byte (e.g., `al', `r8b')

  207. //	h	high byte (e.g., `ah')

  208. //	w	word (e.g., `ax', `r8w')

  209. //	d	doubleword (e.g., `eax', `r8d')

  210. //	q	quadword (e.g., `rax', `r8')

  211. //	r	whole register (doubleword on x86, quadword on amd64)

  212. //

  213. // And types are as follows.

  214. //

  215. //	abcd	the four traditional registers `a', `b', `c', `d'

  216. //	xp	the four pointer registers `si', `di', `bp', `sp'

  217. //	ip	the instruction pointer `ip'

  218. //	rn	the AMD64 numbered registers `r8'--`r15'

  219. #define _DECOR(ty, decor, reg) _DECOR_##ty##_##decor(reg)

  220. 

  221. // Internal macros: _DECOR_ty_decor(reg) applies decoration decor to

  222. // (internal) register name reg of type ty.

  223. 

  224. #define _DECOR_abcd_b(reg) reg##l

  225. #define _DECOR_abcd_h(reg) reg##h

  226. #define _DECOR_abcd_w(reg) reg##x

  227. #define _DECOR_abcd_d(reg) e##reg##x

  228. #if CPUFAM_AMD64

  229. #  define _DECOR_abcd_q(reg) r##reg##x

  230. #endif

  231. 

  232. #define _DECOR_xp_w(reg) reg

  233. #define _DECOR_xp_d(reg) e##reg

  234. #if CPUFAM_AMD64

  235. #  define _DECOR_xp_b(reg) reg##l

  236. #  define _DECOR_xp_q(reg) r##reg

  237. #endif

  238. 

  239. #define _DECOR_ip_w(reg) reg

  240. #define _DECOR_ip_d(reg) e##reg

  241. #if CPUFAM_AMD64

  242. #  define _DECOR_ip_q(reg) r##reg

  243. #endif

  244. 

  245. #if CPUFAM_AMD64

  246. #  define _DECOR_rn_b(reg) reg##b

  247. #  define _DECOR_rn_w(reg) reg##w

  248. #  define _DECOR_rn_d(reg) reg##d

  249. #  define _DECOR_rn_q(reg) reg

  250. #  define _DECOR_rn_r(reg) reg

  251. #endif

  252. 

  253. #define _DECOR_mem_b(addr) byte ptr addr

  254. #define _DECOR_mem_w(addr) word ptr addr

  255. #define _DECOR_mem_d(addr) dword ptr addr

  256. #if CPUFAM_AMD64

  257. #  define _DECOR_mem_q(addr) qword ptr addr

  258. #endif

  259. 

  260. #define _DECOR_imm_b(imm) byte imm

  261. #define _DECOR_imm_w(imm) word imm

  262. #define _DECOR_imm_d(imm) dword imm

  263. #if CPUFAM_AMD64

  264. #  define _DECOR_imm_q(imm) qword imm

  265. #endif

  266. 

  267. #if CPUFAM_X86

  268. #  define _DECOR_abcd_r(reg) e##reg##x

  269. #  define _DECOR_xp_r(reg) e##reg

  270. #  define _DECOR_ip_r(reg) e##reg

  271. #  define _DECOR_mem_r(addr) dword ptr addr

  272. #  define _DECOR_imm_r(imm) dword imm

  273. #endif

  274. #if CPUFAM_AMD64

  275. #  define _DECOR_abcd_r(reg) r##reg##x

  276. #  define _DECOR_xp_r(reg) r##reg

  277. #  define _DECOR_ip_r(reg) r##reg

  278. #  define _DECOR_mem_r(addr) qword ptr addr

  279. #  define _DECOR_imm_r(imm) qword imm

  280. #endif

  281. 

  282. // R_r(decor) applies decoration decor to register r, which is an internal

  283. // register name.  The internal register names are: `ip', `a', `b', `c', `d',

  284. // `si', `di', `bp', `sp', `r8'--`r15'.

  285. #define R_nil(decor) nil

  286. #define R_ip(decor) _DECOR(ip, decor, ip)

  287. #define R_a(decor) _DECOR(abcd, decor, a)

  288. #define R_b(decor) _DECOR(abcd, decor, b)

  289. #define R_c(decor) _DECOR(abcd, decor, c)

  290. #define R_d(decor) _DECOR(abcd, decor, d)

  291. #define R_si(decor) _DECOR(xp, decor, si)

  292. #define R_di(decor) _DECOR(xp, decor, di)

  293. #define R_bp(decor) _DECOR(xp, decor, bp)

  294. #define R_sp(decor) _DECOR(xp, decor, sp)

  295. #if CPUFAM_AMD64

  296. #  define R_r8(decor) _DECOR(rn, decor, r8)

  297. #  define R_r9(decor) _DECOR(rn, decor, r9)

  298. #  define R_r10(decor) _DECOR(rn, decor, r10)

  299. #  define R_r11(decor) _DECOR(rn, decor, r11)

  300. #  define R_r12(decor) _DECOR(rn, decor, r12)

  301. #  define R_r13(decor) _DECOR(rn, decor, r13)

  302. #  define R_r14(decor) _DECOR(rn, decor, r14)

  303. #  define R_r15(decor) _DECOR(rn, decor, r15)

  304. #endif

  305. 

  306. // Refer to an in-memory datum of the type implied by decor residing at

  307. // address addr (which should supply its own square-brackets).

  308. #define MEM(decor, addr) _DECOR(mem, decor, addr)

  309. 

  310. // Refer to an immediate datum of the type implied by decor.

  311. #define IMM(decor, imm) _DECOR(mem, decor, imm)

  312. 

  313. // Applies decoration decor to assembler-level register name reg.

  314. #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)

  315. 

  316. // Internal macros: _REGFORM_r(decor) applies decoration decor to an

  317. // assembler-level register name, in place of any decoration that register

  318. // name has already.

  319. 

  320. #define _REGFORM_nil(decor) R_nil(decor)

  321. 

  322. #define _REGFORM_ip(decor) R_ip(decor)

  323. #define _REGFORM_eip(decor) R_ip(decor)

  324. 

  325. #define _REGFORM_a(decor) R_a(decor)

  326. #define _REGFORM_al(decor) R_a(decor)

  327. #define _REGFORM_ah(decor) R_a(decor)

  328. #define _REGFORM_ax(decor) R_a(decor)

  329. #define _REGFORM_eax(decor) R_a(decor)

  330. 

  331. #define _REGFORM_b(decor) R_b(decor)

  332. #define _REGFORM_bl(decor) R_b(decor)

  333. #define _REGFORM_bh(decor) R_b(decor)

  334. #define _REGFORM_bx(decor) R_b(decor)

  335. #define _REGFORM_ebx(decor) R_b(decor)

  336. 

  337. #define _REGFORM_c(decor) R_c(decor)

  338. #define _REGFORM_cl(decor) R_c(decor)

  339. #define _REGFORM_ch(decor) R_c(decor)

  340. #define _REGFORM_cx(decor) R_c(decor)

  341. #define _REGFORM_ecx(decor) R_c(decor)

  342. 

  343. #define _REGFORM_d(decor) R_d(decor)

  344. #define _REGFORM_dl(decor) R_d(decor)

  345. #define _REGFORM_dh(decor) R_d(decor)

  346. #define _REGFORM_dx(decor) R_d(decor)

  347. #define _REGFORM_edx(decor) R_d(decor)

  348. 

  349. #define _REGFORM_si(decor) R_si(decor)

  350. #define _REGFORM_sil(decor) R_si(decor)

  351. #define _REGFORM_esi(decor) R_si(decor)

  352. 

  353. #define _REGFORM_di(decor) R_di(decor)

  354. #define _REGFORM_dil(decor) R_di(decor)

  355. #define _REGFORM_edi(decor) R_di(decor)

  356. 

  357. #define _REGFORM_bp(decor) R_bp(decor)

  358. #define _REGFORM_bpl(decor) R_bp(decor)

  359. #define _REGFORM_ebp(decor) R_bp(decor)

  360. 

  361. #define _REGFORM_sp(decor) R_sp(decor)

  362. #define _REGFORM_spl(decor) R_sp(decor)

  363. #define _REGFORM_esp(decor) R_sp(decor)

  364. 

  365. #if CPUFAM_AMD64

  366. 

  367. #  define _REGFORM_rip(decor) R_ip(decor)

  368. #  define _REGFORM_rsp(decor) R_sp(decor)

  369. #  define _REGFORM_rbp(decor) R_bp(decor)

  370. #  define _REGFORM_rdi(decor) R_di(decor)

  371. #  define _REGFORM_rsi(decor) R_si(decor)

  372. #  define _REGFORM_rdx(decor) R_d(decor)

  373. #  define _REGFORM_rcx(decor) R_c(decor)

  374. #  define _REGFORM_rbx(decor) R_b(decor)

  375. #  define _REGFORM_rax(decor) R_a(decor)

  376. 

  377. #  define _REGFORM_r8(decor) R_r8(decor)

  378. #  define _REGFORM_r8b(decor) R_r8(decor)

  379. #  define _REGFORM_r8w(decor) R_r8(decor)

  380. #  define _REGFORM_r8d(decor) R_r8(decor)

  381. 

  382. #  define _REGFORM_r9(decor) R_r9(decor)

  383. #  define _REGFORM_r9b(decor) R_r9(decor)

  384. #  define _REGFORM_r9w(decor) R_r9(decor)

  385. #  define _REGFORM_r9d(decor) R_r9(decor)

  386. 

  387. #  define _REGFORM_r10(decor) R_r10(decor)

  388. #  define _REGFORM_r10b(decor) R_r10(decor)

  389. #  define _REGFORM_r10w(decor) R_r10(decor)

  390. #  define _REGFORM_r10d(decor) R_r10(decor)

  391. 

  392. #  define _REGFORM_r11(decor) R_r11(decor)

  393. #  define _REGFORM_r11b(decor) R_r11(decor)

  394. #  define _REGFORM_r11w(decor) R_r11(decor)

  395. #  define _REGFORM_r11d(decor) R_r11(decor)

  396. 

  397. #  define _REGFORM_r12(decor) R_r12(decor)

  398. #  define _REGFORM_r12b(decor) R_r12(decor)

  399. #  define _REGFORM_r12w(decor) R_r12(decor)

  400. #  define _REGFORM_r12d(decor) R_r12(decor)

  401. 

  402. #  define _REGFORM_r13(decor) R_r13(decor)

  403. #  define _REGFORM_r13b(decor) R_r13(decor)

  404. #  define _REGFORM_r13w(decor) R_r13(decor)

  405. #  define _REGFORM_r13d(decor) R_r13(decor)

  406. 

  407. #  define _REGFORM_r14(decor) R_r14(decor)

  408. #  define _REGFORM_r14b(decor) R_r14(decor)

  409. #  define _REGFORM_r14w(decor) R_r14(decor)

  410. #  define _REGFORM_r14d(decor) R_r14(decor)

  411. 

  412. #  define _REGFORM_r15(decor) R_r15(decor)

  413. #  define _REGFORM_r15b(decor) R_r15(decor)

  414. #  define _REGFORM_r15w(decor) R_r15(decor)

  415. #  define _REGFORM_r15d(decor) R_r15(decor)

  416. 

  417. #endif

  418. 

  419. // Macros for converting register names.

  420. #define BYTE(reg) _REGFORM(reg, b)

  421. #define HIBYTE(reg) _REGFORM(reg, h)

  422. #define WORD(reg) _REGFORM(reg, w)

  423. #define DWORD(reg) _REGFORM(reg, d)

  424. #if CPUFAM_AMD64

  425. #  define QWORD(reg) _REGFORM(reg, q)

  426. #endif

  427. #define WHOLE(reg) _REGFORM(reg, r)

  428. 

  429. // Macros for some common registers.

  430. #define AX R_a(r)

  431. #define BX R_b(r)

  432. #define CX R_c(r)

  433. #define DX R_d(r)

  434. #define SI R_si(r)

  435. #define DI R_di(r)

  436. #define BP R_bp(r)

  437. #define SP R_sp(r)

  438. 

  439. // Stack management and unwinding.

  440. .macro	setfp	fp=BP, offset=0

  441.   .if \offset == 0

  442. 	mov	\fp, SP

  443. #if __ELF__

  444. 	  .cfi_def_cfa_register \fp

  445. #endif

  446. #if ABI_WIN && CPUFAM_AMD64

  447. 	  .seh_setframe \fp, 0

  448. #endif

  449.   .else

  450. 	lea	\fp, [SP + \offset]

  451. #if __ELF__

  452. 	  .cfi_def_cfa_register \fp

  453. 	  .cfi_adjust_cfa_offset -\offset

  454. #endif

  455. #if ABI_WIN && CPUFAM_AMD64

  456. 	  .seh_setframe \fp, \offset

  457. #endif

  458.   .endif

  459. 	.L$_frameptr_p = -1

  460. 	.macro dropfp; _dropfp	\fp, \offset; .endm

  461. .endm

  462. 

  463. .macro	_dropfp	fp, offset=0

  464.   .if \offset == 0

  465. 	mov	SP, \fp

  466. #if __ELF__

  467. 	  .cfi_def_cfa_register SP

  468. #endif

  469.   .else

  470. 	lea	SP, [\fp - \offset]

  471. #if __ELF__

  472. 	  .cfi_def_cfa_register SP

  473. 	  .cfi_adjust_cfa_offset +\offset

  474. #endif

  475.   .endif

  476. 	.L$_frameptr_p = 0

  477. 	.purgem	dropfp

  478. .endm

  479. 

  480. .macro	stalloc	n

  481. 	sub	SP, \n

  482. #if __ELF__

  483. 	  .cfi_adjust_cfa_offset +\n

  484. #endif

  485. #if ABI_WIN && CPUFAM_AMD64

  486. 	  .seh_stackalloc \n

  487. #endif

  488. .endm

  489. 

  490. .macro	stfree	n

  491. 	add	SP, \n

  492. #if __ELF__

  493. 	  .cfi_adjust_cfa_offset -\n

  494. #endif

  495. .endm

  496. 

  497. .macro	pushreg	r

  498. 	push	\r

  499. #if __ELF__

  500. 	  .cfi_adjust_cfa_offset +WORDSZ

  501. 	  .cfi_rel_offset \r, 0

  502. #endif

  503. #if ABI_WIN && CPUFAM_AMD64

  504. 	  .seh_pushreg \r

  505. #endif

  506. .endm

  507. 

  508. .macro	popreg	r

  509. 	pop	\r

  510. #if __ELF__

  511. 	  .cfi_adjust_cfa_offset -WORDSZ

  512. 	  .cfi_restore \r

  513. #endif

  514. .endm

  515. 

  516. .macro	savexmm	r, offset

  517. 	movdqa	[SP + \offset], \r

  518. #if ABI_WIN && CPUFAM_AMD64

  519. 	  .seh_savexmm \r, \offset

  520. #endif

  521. .endm

  522. 

  523. .macro	rstrxmm	r, offset

  524. 	movdqa	\r, [SP + \offset]

  525. .endm

  526. 

  527. .macro	endprologue

  528. #if ABI_WIN && CPUFAM_AMD64

  529. 	  .seh_endprologue

  530. #endif

  531. 	.L$_prologue_p = -1

  532. .endm

  533. 

  534. #endif

  535. 

  536. ///--------------------------------------------------------------------------

  537. /// ARM-specific hacking.

  538. 

  539. #if CPUFAM_ARMEL

  540. 

  541. // ARM/Thumb mode things.  Use ARM by default.

  542. #define ARM .arm; .L$_pcoff = 8

  543. #define THUMB .thumb; .L$_pcoff = 4

  544. 	ARM

  545. 

  546. // Set the function hooks.

  547. #define FUNC_PREHOOK(_) .balign 4; .fnstart

  548. #define ENDFUNC_HOOK(_) .fnend; .ltorg

  549. 

  550. // Call external subroutine at ADDR, possibly via PLT.

  551. .macro	callext addr, cond=

  552. #if WANT_PIC

  553. 	bl\cond	\addr(PLT)

  554. #else

  555. 	bl\cond	\addr

  556. #endif

  557. .endm

  558. 

  559. // Do I need to arrange a spare GOT register?

  560. #if WANT_PIC

  561. #  define NEED_GOT 1

  562. #endif

  563. #define GOTREG r9

  564. 

  565. // Maybe load GOT address into GOT.

  566. .macro	ldgot	cond=, got=GOTREG

  567. #if WANT_PIC

  568. 	ldr\cond \got, .L$_ldgot$\@

  569. .L$_ldgot_pc$\@:

  570. 	add\cond \got, pc, \got

  571.   _LIT

  572. 	.balign	4

  573. .L$_ldgot$\@:

  574. 	.word	_GLOBAL_OFFSET_TABLE_ - .L$_ldgot_pc$\@ - .L$_pcoff

  575.   _ENDLIT

  576. #endif

  577. .endm

  578. 

  579. // Load address of external symbol ADDR into REG, maybe using GOT.

  580. .macro	leaext	reg, addr, cond=, got=GOTREG

  581. #if WANT_PIC

  582. 	ldr\cond \reg, .L$_leaext$\@

  583. 	ldr\cond \reg, [\got, \reg]

  584.   _LIT

  585. 	.balign	4

  586. .L$_leaext$\@:

  587. 	.word	\addr(GOT)

  588.   _ENDLIT

  589. #else

  590. 	ldr\cond \reg, =\addr

  591. #endif

  592. .endm

  593. 

  594. // Load address of external symbol ADDR into REG directly.

  595. .macro	leaextq	reg, addr, cond=

  596. #if WANT_PIC

  597. 	ldr\cond \reg, .L$_leaextq$\@

  598. .L$_leaextq_pc$\@:

  599.   .if .L$_pcoff == 8

  600. 	ldr\cond \reg, [pc, \reg]

  601.   .else

  602. 	add\cond \reg, pc

  603. 	ldr\cond \reg, [\reg]

  604.   .endif

  605.   _LIT

  606. 	.balign	4

  607. .L$_leaextq$\@:

  608. 	.word	\addr(GOT_PREL) + (. - .L$_leaextq_pc$\@ - .L$_pcoff)

  609.   _ENDLIT

  610. #else

  611. 	ldr\cond \reg, =\addr

  612. #endif

  613. .endm

  614. 

  615. .macro	vzero	vz=q15

  616. 	// Set VZ (default q15) to zero.

  617. 	vmov.u32 \vz, #0

  618. .endm

  619. 

  620. .macro	vshl128	vd, vn, nbit, vz=q15

  621. 	// Set VD to VN shifted left by NBIT.  Assume VZ (default q15) is

  622. 	// all-bits-zero.  NBIT must be a multiple of 8.

  623.   .if \nbit&3 != 0

  624. 	.error	"shift quantity must be whole number of bytes"

  625.   .endif

  626. 	vext.8	\vd, \vz, \vn, #16 - (\nbit >> 3)

  627. .endm

  628. 

  629. .macro	vshr128	vd, vn, nbit, vz=q15

  630. 	// Set VD to VN shifted right by NBIT.  Assume VZ (default q15) is

  631. 	// all-bits-zero.  NBIT must be a multiple of 8.

  632.   .if \nbit&3 != 0

  633. 	.error	"shift quantity must be whole number of bytes"

  634.   .endif

  635. 	vext.8	\vd, \vn, \vz, #\nbit >> 3

  636. .endm

  637. 

  638. // Apply decoration decor to register name reg.

  639. #define _REGFORM(reg, decor) _GLUE(_REGFORM_, reg)(decor)

  640. 

  641. // Internal macros: `_REGFORM_r(decor)' applies decoration decor to register

  642. // name r.

  643. 

  644. #define _REGFORM_nil(decor) nil

  645. 

  646. #define _REGFORM_s0(decor) _DECOR(s, decor, 0)

  647. #define _REGFORM_s1(decor) _DECOR(s, decor, 1)

  648. #define _REGFORM_s2(decor) _DECOR(s, decor, 2)

  649. #define _REGFORM_s3(decor) _DECOR(s, decor, 3)

  650. #define _REGFORM_s4(decor) _DECOR(s, decor, 4)

  651. #define _REGFORM_s5(decor) _DECOR(s, decor, 5)

  652. #define _REGFORM_s6(decor) _DECOR(s, decor, 6)

  653. #define _REGFORM_s7(decor) _DECOR(s, decor, 7)

  654. #define _REGFORM_s8(decor) _DECOR(s, decor, 8)

  655. #define _REGFORM_s9(decor) _DECOR(s, decor, 9)

  656. #define _REGFORM_s10(decor) _DECOR(s, decor, 10)

  657. #define _REGFORM_s11(decor) _DECOR(s, decor, 11)

  658. #define _REGFORM_s12(decor) _DECOR(s, decor, 12)

  659. #define _REGFORM_s13(decor) _DECOR(s, decor, 13)

  660. #define _REGFORM_s14(decor) _DECOR(s, decor, 14)

  661. #define _REGFORM_s15(decor) _DECOR(s, decor, 15)

  662. #define _REGFORM_s16(decor) _DECOR(s, decor, 16)

  663. #define _REGFORM_s17(decor) _DECOR(s, decor, 17)

  664. #define _REGFORM_s18(decor) _DECOR(s, decor, 18)

  665. #define _REGFORM_s19(decor) _DECOR(s, decor, 19)

  666. #define _REGFORM_s20(decor) _DECOR(s, decor, 20)

  667. #define _REGFORM_s21(decor) _DECOR(s, decor, 21)

  668. #define _REGFORM_s22(decor) _DECOR(s, decor, 22)

  669. #define _REGFORM_s23(decor) _DECOR(s, decor, 23)

  670. #define _REGFORM_s24(decor) _DECOR(s, decor, 24)

  671. #define _REGFORM_s25(decor) _DECOR(s, decor, 25)

  672. #define _REGFORM_s26(decor) _DECOR(s, decor, 26)

  673. #define _REGFORM_s27(decor) _DECOR(s, decor, 27)

  674. #define _REGFORM_s28(decor) _DECOR(s, decor, 28)

  675. #define _REGFORM_s29(decor) _DECOR(s, decor, 29)

  676. #define _REGFORM_s30(decor) _DECOR(s, decor, 30)

  677. #define _REGFORM_s31(decor) _DECOR(s, decor, 31)

  678. 

  679. #define _REGFORM_d0(decor) _DECOR(d, decor, 0)

  680. #define _REGFORM_d1(decor) _DECOR(d, decor, 1)

  681. #define _REGFORM_d2(decor) _DECOR(d, decor, 2)

  682. #define _REGFORM_d3(decor) _DECOR(d, decor, 3)

  683. #define _REGFORM_d4(decor) _DECOR(d, decor, 4)

  684. #define _REGFORM_d5(decor) _DECOR(d, decor, 5)

  685. #define _REGFORM_d6(decor) _DECOR(d, decor, 6)

  686. #define _REGFORM_d7(decor) _DECOR(d, decor, 7)

  687. #define _REGFORM_d8(decor) _DECOR(d, decor, 8)

  688. #define _REGFORM_d9(decor) _DECOR(d, decor, 9)

  689. #define _REGFORM_d10(decor) _DECOR(d, decor, 10)

  690. #define _REGFORM_d11(decor) _DECOR(d, decor, 11)

  691. #define _REGFORM_d12(decor) _DECOR(d, decor, 12)

  692. #define _REGFORM_d13(decor) _DECOR(d, decor, 13)

  693. #define _REGFORM_d14(decor) _DECOR(d, decor, 14)

  694. #define _REGFORM_d15(decor) _DECOR(d, decor, 15)

  695. #define _REGFORM_d16(decor) _DECOR(d, decor, 16)

  696. #define _REGFORM_d17(decor) _DECOR(d, decor, 17)

  697. #define _REGFORM_d18(decor) _DECOR(d, decor, 18)

  698. #define _REGFORM_d19(decor) _DECOR(d, decor, 19)

  699. #define _REGFORM_d20(decor) _DECOR(d, decor, 20)

  700. #define _REGFORM_d21(decor) _DECOR(d, decor, 21)

  701. #define _REGFORM_d22(decor) _DECOR(d, decor, 22)

  702. #define _REGFORM_d23(decor) _DECOR(d, decor, 23)

  703. #define _REGFORM_d24(decor) _DECOR(d, decor, 24)

  704. #define _REGFORM_d25(decor) _DECOR(d, decor, 25)

  705. #define _REGFORM_d26(decor) _DECOR(d, decor, 26)

  706. #define _REGFORM_d27(decor) _DECOR(d, decor, 27)

  707. #define _REGFORM_d28(decor) _DECOR(d, decor, 28)

  708. #define _REGFORM_d29(decor) _DECOR(d, decor, 29)

  709. #define _REGFORM_d30(decor) _DECOR(d, decor, 30)

  710. #define _REGFORM_d31(decor) _DECOR(d, decor, 31)

  711. 

  712. #define _REGFORM_q0(decor) _DECOR(q, decor, 0)

  713. #define _REGFORM_q1(decor) _DECOR(q, decor, 1)

  714. #define _REGFORM_q2(decor) _DECOR(q, decor, 2)

  715. #define _REGFORM_q3(decor) _DECOR(q, decor, 3)

  716. #define _REGFORM_q4(decor) _DECOR(q, decor, 4)

  717. #define _REGFORM_q5(decor) _DECOR(q, decor, 5)

  718. #define _REGFORM_q6(decor) _DECOR(q, decor, 6)

  719. #define _REGFORM_q7(decor) _DECOR(q, decor, 7)

  720. #define _REGFORM_q8(decor) _DECOR(q, decor, 8)

  721. #define _REGFORM_q9(decor) _DECOR(q, decor, 9)

  722. #define _REGFORM_q10(decor) _DECOR(q, decor, 10)

  723. #define _REGFORM_q11(decor) _DECOR(q, decor, 11)

  724. #define _REGFORM_q12(decor) _DECOR(q, decor, 12)

  725. #define _REGFORM_q13(decor) _DECOR(q, decor, 13)

  726. #define _REGFORM_q14(decor) _DECOR(q, decor, 14)

  727. #define _REGFORM_q15(decor) _DECOR(q, decor, 15)

  728. 

  729. // `_LOPART(n)' and `_HIPART(n)' return the numbers of the register halves of

  730. // register n, i.e., 2*n and 2*n + 1 respectively.

  731. #define _LOPART(n) _GLUE(_LOPART_, n)

  732. #define _HIPART(n) _GLUE(_HIPART_, n)

  733. 

  734. // Internal macros: `_LOPART_n' and `_HIPART_n' return the numbers of the

  735. // register halves of register n, i.e., 2*n and 2*n + 1 respectively.

  736. 

  737. #define _LOPART_0 0

  738. #define _HIPART_0 1

  739. #define _LOPART_1 2

  740. #define _HIPART_1 3

  741. #define _LOPART_2 4

  742. #define _HIPART_2 5

  743. #define _LOPART_3 6

  744. #define _HIPART_3 7

  745. #define _LOPART_4 8

  746. #define _HIPART_4 9

  747. #define _LOPART_5 10

  748. #define _HIPART_5 11

  749. #define _LOPART_6 12

  750. #define _HIPART_6 13

  751. #define _LOPART_7 14

  752. #define _HIPART_7 15

  753. #define _LOPART_8 16

  754. #define _HIPART_8 17

  755. #define _LOPART_9 18

  756. #define _HIPART_9 19

  757. #define _LOPART_10 20

  758. #define _HIPART_10 21

  759. #define _LOPART_11 22

  760. #define _HIPART_11 23

  761. #define _LOPART_12 24

  762. #define _HIPART_12 25

  763. #define _LOPART_13 26

  764. #define _HIPART_13 27

  765. #define _LOPART_14 28

  766. #define _HIPART_14 29

  767. #define _LOPART_15 30

  768. #define _HIPART_15 31

  769. 

  770. // Return the register number of the pair containing register n, i.e.,

  771. // floor(n/2).

  772. #define _PAIR(n) _GLUE(_PAIR_, n)

  773. 

  774. // Internal macros: `_PAIR_n' returns the register number of the pair

  775. // containing register n, i.e., floor(n/2).

  776. #define _PAIR_0 0

  777. #define _PAIR_1 0

  778. #define _PAIR_2 1

  779. #define _PAIR_3 1

  780. #define _PAIR_4 2

  781. #define _PAIR_5 2

  782. #define _PAIR_6 3

  783. #define _PAIR_7 3

  784. #define _PAIR_8 4

  785. #define _PAIR_9 4

  786. #define _PAIR_10 5

  787. #define _PAIR_11 5

  788. #define _PAIR_12 6

  789. #define _PAIR_13 6

  790. #define _PAIR_14 7

  791. #define _PAIR_15 7

  792. #define _PAIR_16 8

  793. #define _PAIR_17 8

  794. #define _PAIR_18 9

  795. #define _PAIR_19 9

  796. #define _PAIR_20 10

  797. #define _PAIR_21 10

  798. #define _PAIR_22 11

  799. #define _PAIR_23 11

  800. #define _PAIR_24 12

  801. #define _PAIR_25 12

  802. #define _PAIR_26 13

  803. #define _PAIR_27 13

  804. #define _PAIR_28 14

  805. #define _PAIR_29 14

  806. #define _PAIR_30 15

  807. #define _PAIR_31 15

  808. 

  809. // Apply decoration decor to register number n of type ty.  Decorations are

  810. // as follows.

  811. //

  812. //	decor	types	meaning

  813. //	Q	s, d	the NEON qN register containing this one

  814. //	D	s	the NEON dN register containing this one

  815. //	D0	q	the low 64-bit half of this one

  816. //	D1	q	the high 64-bit half of this one

  817. //	S0	d, q	the first 32-bit piece of this one

  818. //	S1	d, q	the second 32-bit piece of this one

  819. //	S2	q	the third 32-bit piece of this one

  820. //	S3	q	the fourth 32-bit piece of this one

  821. //	Bn	q	the nth byte of this register, as a scalar

  822. //	Hn	q	the nth halfword of this register, as a scalar

  823. //	Wn	q	the nth word of this register, as a scalar

  824. #define _DECOR(ty, decor, n) _DECOR_##ty##_##decor(n)

  825. 

  826. // Internal macros: `_DECOR_ty_decor(n)' applies decoration decor to register

  827. // number n of type ty.

  828. 

  829. #define _DECOR_s_Q(n) GLUE(q, _PAIR(_PAIR(n)))

  830. #define _DECOR_s_D(n) GLUE(d, _PAIR(n))

  831. 

  832. #define _DECOR_d_Q(n) GLUE(q, _PAIR(n))

  833. #define _DECOR_d_S0(n) GLUE(s, _LOPART(n))

  834. #define _DECOR_d_S1(n) GLUE(s, _LOPART(n))

  835. 

  836. #define _DECOR_q_D0(n) GLUE(d, _LOPART(n))

  837. #define _DECOR_q_D1(n) GLUE(d, _HIPART(n))

  838. #define _DECOR_q_S0(n) GLUE(s, _LOPART(_LOPART(n)))

  839. #define _DECOR_q_S1(n) GLUE(s, _HIPART(_LOPART(n)))

  840. #define _DECOR_q_S2(n) GLUE(s, _LOPART(_HIPART(n)))

  841. #define _DECOR_q_S3(n) GLUE(s, _HIPART(_HIPART(n)))

  842. #define _DECOR_q_W0(n) GLUE(d, _LOPART(n))[0]

  843. #define _DECOR_q_W1(n) GLUE(d, _LOPART(n))[1]

  844. #define _DECOR_q_W2(n) GLUE(d, _HIPART(n))[0]

  845. #define _DECOR_q_W3(n) GLUE(d, _HIPART(n))[1]

  846. #define _DECOR_q_H0(n) GLUE(d, _LOPART(n))[0]

  847. #define _DECOR_q_H1(n) GLUE(d, _LOPART(n))[1]

  848. #define _DECOR_q_H2(n) GLUE(d, _LOPART(n))[2]

  849. #define _DECOR_q_H3(n) GLUE(d, _LOPART(n))[3]

  850. #define _DECOR_q_H4(n) GLUE(d, _HIPART(n))[0]

  851. #define _DECOR_q_H5(n) GLUE(d, _HIPART(n))[1]

  852. #define _DECOR_q_H6(n) GLUE(d, _HIPART(n))[2]

  853. #define _DECOR_q_H7(n) GLUE(d, _HIPART(n))[3]

  854. #define _DECOR_q_B0(n) GLUE(d, _LOPART(n))[0]

  855. #define _DECOR_q_B1(n) GLUE(d, _LOPART(n))[1]

  856. #define _DECOR_q_B2(n) GLUE(d, _LOPART(n))[2]

  857. #define _DECOR_q_B3(n) GLUE(d, _LOPART(n))[3]

  858. #define _DECOR_q_B4(n) GLUE(d, _LOPART(n))[4]

  859. #define _DECOR_q_B5(n) GLUE(d, _LOPART(n))[5]

  860. #define _DECOR_q_B6(n) GLUE(d, _LOPART(n))[6]

  861. #define _DECOR_q_B7(n) GLUE(d, _LOPART(n))[7]

  862. #define _DECOR_q_B8(n) GLUE(d, _HIPART(n))[0]

  863. #define _DECOR_q_B9(n) GLUE(d, _HIPART(n))[1]

  864. #define _DECOR_q_B10(n) GLUE(d, _HIPART(n))[2]

  865. #define _DECOR_q_B11(n) GLUE(d, _HIPART(n))[3]

  866. #define _DECOR_q_B12(n) GLUE(d, _HIPART(n))[4]

  867. #define _DECOR_q_B13(n) GLUE(d, _HIPART(n))[5]

  868. #define _DECOR_q_B14(n) GLUE(d, _HIPART(n))[6]

  869. #define _DECOR_q_B15(n) GLUE(d, _HIPART(n))[7]

  870. 

  871. // Macros for navigating the NEON register hierarchy.

  872. #define S0(reg) _REGFORM(reg, S0)

  873. #define S1(reg) _REGFORM(reg, S1)

  874. #define S2(reg) _REGFORM(reg, S2)

  875. #define S3(reg) _REGFORM(reg, S3)

  876. #define D(reg) _REGFORM(reg, D)

  877. #define D0(reg) _REGFORM(reg, D0)

  878. #define D1(reg) _REGFORM(reg, D1)

  879. #define Q(reg) _REGFORM(reg, Q)

  880. 

  881. // Macros for indexing quadword registers.

  882. #define QB(reg, i) _REGFORM(reg, B##i)

  883. #define QH(reg, i) _REGFORM(reg, H##i)

  884. #define QW(reg, i) _REGFORM(reg, W##i)

  885. 

  886. // Macros for converting vldm/vstm ranges.

  887. #define QQ(qlo, qhi) D0(qlo)-D1(qhi)

  888. 

  889. // Stack management and unwinding.

  890. .macro	setfp	fp=r11, offset=0

  891.   .if \offset == 0

  892. 	mov	\fp, sp

  893. 	  .setfp \fp, sp

  894.   .else

  895. 	add	\fp, sp, #\offset

  896. 	  .setfp \fp, sp, #\offset

  897.   .endif

  898. 	.macro dropfp; _dropfp	\fp, \offset; .endm

  899. 	.L$_frameptr_p = -1

  900. .endm

  901. 

  902. .macro	_dropfp	fp, offset=0

  903.   .if \offset == 0

  904. 	mov	sp, \fp

  905.   .else

  906. 	sub	sp, \fp, #\offset

  907.   .endif

  908. 	.purgem	dropfp

  909. 	.L$_frameptr_p = 0

  910. .endm

  911. 

  912. .macro	stalloc	n

  913. 	sub	sp, sp, #\n

  914. 	  .pad #\n

  915. .endm

  916. 

  917. .macro	stfree	n

  918. 	add	sp, sp, #\n

  919. 	  .pad #-\n

  920. .endm

  921. 

  922. .macro	pushreg	rr:vararg

  923. 	push	{\rr}

  924. 	  .save {\rr}

  925. .endm

  926. 

  927. .macro	popreg	rr:vararg

  928. 	pop	{\rr}

  929. .endm

  930. 

  931. .macro	pushvfp rr:vararg

  932. 	vstmdb	sp!, {\rr}

  933. 	  .vsave {\rr}

  934. .endm

  935. 

  936. .macro	popvfp rr:vararg

  937. 	vldmia	sp!, {\rr}

  938. .endm

  939. 

  940. .macro	endprologue

  941. .endm

  942. 

  943. // No need for prologue markers on ARM.

  944. #define FUNC_POSTHOOK(_) .L$_prologue_p = -1

  945. 

  946. #endif

  947. 

  948. ///--------------------------------------------------------------------------

  949. /// AArch64-specific hacking.

  950. 

  951. #if CPUFAM_ARM64

  952. 

  953. // Set the function hooks.

  954. #define FUNC_PREHOOK(_) .balign 4

  955. #define FUNC_POSTHOOK(_) .cfi_startproc; .L$_prologue_p = -1

  956. #define ENDFUNC_HOOK(_) .cfi_endproc

  957. 

  958. // Call external subroutine at ADDR, possibly via PLT.

  959. .macro	callext addr

  960. 	bl	\addr

  961. .endm

  962. 

  963. // Load address of external symbol ADDR into REG.

  964. .macro	leaext	reg, addr

  965. #if WANT_PIC

  966. 	adrp	\reg, :got:\addr

  967. 	ldr	\reg, [\reg, #:got_lo12:\addr]

  968. #else

  969. 	adrp	\reg, \addr

  970. 	add	\reg, \reg, #:lo12:\addr

  971. #endif

  972. .endm

  973. 

  974. .macro	vzero	vz=v31

  975. 	// Set VZ (default v31) to zero.

  976. 	dup	\vz\().4s, wzr

  977. .endm

  978. 

  979. .macro	vshl128	vd, vn, nbit, vz=v31

  980. 	// Set VD to VN shifted left by NBIT.  Assume VZ (default v31) is

  981. 	// all-bits-zero.  NBIT must be a multiple of 8.

  982.   .if \nbit&3 != 0

  983. 	.error	"shift quantity must be whole number of bytes"

  984.   .endif

  985. 	ext	\vd\().16b, \vz\().16b, \vn\().16b, #16 - (\nbit >> 3)

  986. .endm

  987. 

  988. .macro	vshr128	vd, vn, nbit, vz=v31

  989. 	// Set VD to VN shifted right by NBIT.  Assume VZ (default v31) is

  990. 	// all-bits-zero.  NBIT must be a multiple of 8.

  991.   .if \nbit&3 != 0

  992. 	.error	"shift quantity must be whole number of bytes"

  993.   .endif

  994. 	ext	\vd\().16b, \vn\().16b, \vz\().16b, #\nbit >> 3

  995. .endm

  996. 

  997. // Stack management and unwinding.

  998. .macro	setfp	fp=x29, offset=0

  999.   // If you're just going through the motions with a fixed-size stack frame,

  1000.   // then you want to say `add x29, sp, #OFFSET' directly, which will avoid

  1001.   // pointlessly restoring sp later.

  1002.   .if \offset == 0

  1003. 	mov	\fp, sp

  1004. 	  .cfi_def_cfa_register \fp

  1005.   .else

  1006. 	add	\fp, sp, #\offset

  1007. 	  .cfi_def_cfa_register \fp

  1008. 	  .cfi_adjust_cfa_offset -\offset

  1009.   .endif

  1010. 	.macro dropfp; _dropfp	\fp, \offset; .endm

  1011. 	.L$_frameptr_p = -1

  1012. .endm

  1013. 

  1014. .macro	_dropfp	fp, offset=0

  1015.   .if \offset == 0

  1016. 	mov	sp, \fp

  1017. 	  .cfi_def_cfa_register sp

  1018.   .else

  1019. 	sub	sp, \fp, #\offset

  1020. 	  .cfi_def_cfa_register sp

  1021. 	  .cfi_adjust_cfa_offset +\offset

  1022.   .endif

  1023. 	.purgem	dropfp

  1024. 	.L$_frameptr_p = 0

  1025. .endm

  1026. 

  1027. .macro	stalloc	n

  1028. 	sub	sp, sp, #\n

  1029. 	  .cfi_adjust_cfa_offset +\n

  1030. .endm

  1031. 

  1032. .macro	stfree	n

  1033. 	add	sp, sp, #\n

  1034. 	  .cfi_adjust_cfa_offset -\n

  1035. .endm

  1036. 

  1037. .macro	pushreg	x, y=nil

  1038.   .ifeqs "\y", "nil"

  1039. 	str	\x, [sp, #-16]!

  1040. 	  .cfi_adjust_cfa_offset +16

  1041. 	  .cfi_rel_offset \x, 0

  1042.   .else

  1043. 	stp	\x, \y, [sp, #-16]!

  1044. 	  .cfi_adjust_cfa_offset +16

  1045. 	  .cfi_rel_offset \x, 0

  1046. 	  .cfi_rel_offset \y, 8

  1047.   .endif

  1048. .endm

  1049. 

  1050. .macro	popreg	x, y=nil

  1051.   .ifeqs "\y", "nil"

  1052. 	ldr	\x, [sp], #16

  1053. 	  .cfi_restore \x

  1054. 	  .cfi_adjust_cfa_offset -16

  1055.   .else

  1056. 	ldp	\x, \y, [sp], #16

  1057. 	  .cfi_restore \x

  1058. 	  .cfi_restore \y

  1059. 	  .cfi_adjust_cfa_offset -16

  1060.   .endif

  1061. .endm

  1062. 

  1063. .macro	savereg	x, y, z=nil

  1064.   .ifeqs "\z", "nil"

  1065. 	str	\x, [sp, \y]

  1066. 	  .cfi_rel_offset \x, \y

  1067.   .else

  1068. 	stp	\x, \y, [sp, #\z]

  1069. 	  .cfi_rel_offset \x, \z

  1070. 	  .cfi_rel_offset \y, \z + 8

  1071.   .endif

  1072. .endm

  1073. 

  1074. .macro	rstrreg	x, y, z=nil

  1075.   .ifeqs "\z", "nil"

  1076. 	ldr	\x, [sp, \y]

  1077. 	  .cfi_restore \x

  1078.   .else

  1079. 	ldp	\x, \y, [sp, #\z]

  1080. 	  .cfi_restore \x

  1081. 	  .cfi_restore \y

  1082.   .endif

  1083. .endm

  1084. 

  1085. .macro	endprologue

  1086. .endm

  1087. 

  1088. // cmov RD, RN, CC: set RD to RN if CC is satisfied, otherwise do nothing

  1089. .macro	cmov	rd, rn, cc

  1090. 	csel	\rd, \rn, \rd, \cc

  1091. .endm

  1092. 

  1093. // Notational improvement: write `csel.CC' etc., rather than `csel ..., CC'.

  1094. #define _COND(_)							\

  1095. 	_(eq) _(ne) _(cs) _(cc) _(vs) _(vc) _(mi) _(pl)			\

  1096. 	_(ge) _(lt) _(gt) _(le) _(hi) _(ls) _(al) _(nv)			\

  1097. 	_(hs) _(lo)

  1098. #define _INST(_)							\

  1099. 	_(ccmp) _(ccmn)							\

  1100. 	_(csel) _(cmov)							\

  1101. 	_(csinc) _(cinc) _(cset)					\

  1102. 	_(csneg) _(cneg)						\

  1103. 	_(csinv) _(cinv) _(csetm)

  1104. #define _CONDVAR(cc) _definstvar cc;

  1105. #define _INSTVARS(inst)							\

  1106. 	.macro _definstvar cc;						\

  1107. 	  .macro inst.\cc args:vararg; inst \args, \cc; .endm;		\

  1108. 	.endm;								\

  1109. 	_COND(_CONDVAR);						\

  1110. 	.purgem _definstvar;

  1111. 	_INST(_INSTVARS)

  1112. #undef _COND

  1113. #undef _INST

  1114. #undef _CONDVAR

  1115. #undef _INSTVARS

  1116. 

  1117. // Flag bits for `ccmp' and friends.

  1118. #define CCMP_N 8

  1119. #define CCMP_Z 4

  1120. #define CCMP_C 2

  1121. #define CCMP_V 1

  1122. 

  1123. // Flag settings for satisfying conditions.

  1124. #define CCMP_MI CCMP_N

  1125. #define CCMP_PL 0

  1126. #define CCMP_EQ CCMP_Z

  1127. #define CCMP_NE 0

  1128. #define CCMP_CS CCMP_C

  1129. #define CCMP_HS CCMP_C

  1130. #define CCMP_CC 0

  1131. #define CCMP_LO 0

  1132. #define CCMP_VS CCMP_V

  1133. #define CCMP_VC 0

  1134. #define CCMP_HI CCMP_C

  1135. #define CCMP_LS 0

  1136. #define CCMP_LT CCMP_N

  1137. #define CCMP_GE 0

  1138. #define CCMP_LE CCMP_N

  1139. #define CCMP_GT 0

  1140. 

  1141. #endif

  1142. 

  1143. ///--------------------------------------------------------------------------

  1144. /// Final stuff.

  1145. 

  1146. // Default values for the various hooks.

  1147. #ifndef FUNC_PREHOOK

  1148. #  define FUNC_PREHOOK(_)

  1149. #endif

  1150. #ifndef FUNC_POSTHOOK

  1151. #  define FUNC_POSTHOOK(_)

  1152. #endif

  1153. #ifndef ENDFUNC_HOOK

  1154. #  define ENDFUNC_HOOK(_)

  1155. #endif

  1156. 

  1157. #ifndef F

  1158. #  ifdef SYM_USCORE

  1159. #    define F(name) _##name

  1160. #  else

  1161. #    define F(name) name

  1162. #  endif

  1163. #endif

  1164. 

  1165. #ifndef TYPE_FUNC

  1166. #  define TYPE_FUNC(name)

  1167. #endif

  1168. 

  1169. #ifndef SIZE_OBJ

  1170. #  define SIZE_OBJ(name)

  1171. #endif

  1172. 

  1173. #if __ELF__ && !defined(WANT_EXECUTABLE_STACK)

  1174. 	.pushsection .note.GNU-stack, "", _SECTTY(progbits)

  1175. 	.popsection

  1176. #endif

  1177. 

  1178. ///----- That's all, folks --------------------------------------------------

  1179. 

  1180. #endif