1 ;*****************************************************************************
3 ;*****************************************************************************
4 ;* Copyright (C) 2005-2008 x264 project
6 ;* Authors: Loren Merritt <lorenm@u.washington.edu>
7 ;* Anton Mitrofanov <BugMaster@narod.ru>
9 ;* Permission to use, copy, modify, and/or distribute this software for any
10 ;* purpose with or without fee is hereby granted, provided that the above
11 ;* copyright notice and this permission notice appear in all copies.
13 ;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
14 ;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
15 ;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
16 ;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
17 ;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
18 ;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
19 ;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
20 ;*****************************************************************************
22 ; This is a header file for the x264ASM assembly language, which uses
23 ; NASM/YASM syntax combined with a large number of macros to provide easy
24 ; abstraction between different calling conventions (x86_32, win64, linux64).
25 ; It also has various other useful features to simplify writing the kind of
26 ; DSP functions that are most often used in x264.
28 ; Unlike the rest of x264, this file is available under an ISC license, as it
29 ; has significant usefulness outside of x264 and we want it to be available
30 ; to the largest audience possible. Of course, if you modify it for your own
31 ; purposes to add a new feature, we strongly encourage contributing a patch
32 ; as this feature might be useful for others as well. Send patches or ideas
33 ; to x264-devel@videolan.org .
36 %ifidn __OUTPUT_FORMAT__,win32
44 %define mangle(x) _ %+ x
49 ; FIXME: All of the 64bit asm functions that take a stride as an argument
50 ; via register, assume that the high dword of that register is filled with 0.
51 ; This is true in practice (since we never do any 64bit arithmetic on strides,
52 ; and x264's strides are all positive), but is not guaranteed by the ABI.
54 ; Name of the .rodata section.
55 ; Kludge: Something on OS X fails to align .rodata even given an align attribute,
56 ; so use a different read-only section.
57 %macro SECTION_RODATA 0-1 16
58 %ifidn __OUTPUT_FORMAT__,macho64
59 SECTION .text align=%1
60 %elifidn __OUTPUT_FORMAT__,macho
61 SECTION .text align=%1
64 SECTION .rodata align=%1
71 ; x86_32 doesn't require PIC.
72 ; Some distros prefer shared objects to be PIC, but nothing breaks if
73 ; the code contains a few textrels, so we'll skip that complexity.
80 ; Macros to eliminate most code duplication between x86_32 and x86_64:
81 ; Currently this works only for leaf functions which load all their arguments
82 ; into registers at the start, and make no other use of the stack. Luckily that
83 ; covers most of x264's asm.
86 ; %1 = number of arguments. loads them from stack if needed.
87 ; %2 = number of registers used. pushes callee-saved regs if needed.
88 ; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
89 ; %4 = list of names to define to registers
90 ; PROLOGUE can also be invoked by adding the same options to cglobal
93 ; cglobal foo, 2,3,0, dst, src, tmp
94 ; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
96 ; TODO Some functions can use some args directly from the stack. If they're the
97 ; last args then you can just not declare them, but if they're in the middle
98 ; we need more flexible macro.
101 ; Pops anything that was pushed by PROLOGUE
104 ; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
105 ; which are slow when a normal ret follows a branch.
108 ; rN and rNq are the native-size register holding function argument N
109 ; rNd, rNw, rNb are dword, word, and byte size
110 ; rNm is the original location of arg N (a register or on the stack), dword
111 ; rNmp is native size
119 %ifid %6 ; i.e. it's a register
121 %elifdef ARCH_X86_64 ; memory
122 %define r%1mp qword %6
124 %define r%1mp dword %6
129 %macro DECLARE_REG_SIZE 2
143 DECLARE_REG_SIZE ax, al
144 DECLARE_REG_SIZE bx, bl
145 DECLARE_REG_SIZE cx, cl
146 DECLARE_REG_SIZE dx, dl
147 DECLARE_REG_SIZE si, sil
148 DECLARE_REG_SIZE di, dil
149 DECLARE_REG_SIZE bp, bpl
151 ; t# defines for when per-arch register allocation is more complex than just function arguments
153 %macro DECLARE_REG_TMP 1-*
156 CAT_XDEFINE t, %%i, r%1
162 %macro DECLARE_REG_TMP_SIZE 0-*
164 %define t%1q t%1 %+ q
165 %define t%1d t%1 %+ d
166 %define t%1w t%1 %+ w
167 %define t%1b t%1 %+ b
172 DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7
182 %assign stack_offset stack_offset+gprsize
187 %assign stack_offset stack_offset-gprsize
193 %assign stack_offset stack_offset+(%2)
200 %assign stack_offset stack_offset-(%2)
210 %macro movsxdifnidn 2
222 %macro DEFINE_ARGS 0-*
226 CAT_UNDEF arg_name %+ %%i, q
227 CAT_UNDEF arg_name %+ %%i, d
228 CAT_UNDEF arg_name %+ %%i, w
229 CAT_UNDEF arg_name %+ %%i, b
230 CAT_UNDEF arg_name %+ %%i, m
231 CAT_UNDEF arg_name, %%i
238 %xdefine %1q r %+ %%i %+ q
239 %xdefine %1d r %+ %%i %+ d
240 %xdefine %1w r %+ %%i %+ w
241 %xdefine %1b r %+ %%i %+ b
242 %xdefine %1m r %+ %%i %+ m
243 CAT_XDEFINE arg_name, %%i, %1
247 %assign n_arg_names %%i
250 %ifdef WIN64 ; Windows x64 ;=================================================
252 DECLARE_REG 0, rcx, ecx, cx, cl, ecx
253 DECLARE_REG 1, rdx, edx, dx, dl, edx
254 DECLARE_REG 2, r8, r8d, r8w, r8b, r8d
255 DECLARE_REG 3, r9, r9d, r9w, r9b, r9d
256 DECLARE_REG 4, rdi, edi, di, dil, [rsp + stack_offset + 40]
257 DECLARE_REG 5, rsi, esi, si, sil, [rsp + stack_offset + 48]
258 DECLARE_REG 6, rax, eax, ax, al, [rsp + stack_offset + 56]
259 %define r7m [rsp + stack_offset + 64]
260 %define r8m [rsp + stack_offset + 72]
262 %macro LOAD_IF_USED 2 ; reg_id, number_of_args
264 mov r%1, [rsp + stack_offset + 8 + %1*8]
268 %macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
271 ASSERT regs_used <= 7
272 %assign xmm_regs_used %3
273 ASSERT xmm_regs_used <= 16
277 %assign stack_offset stack_offset+16
279 %if xmm_regs_used > 6
280 sub rsp, (xmm_regs_used-6)*16+16
281 %assign stack_offset stack_offset+(xmm_regs_used-6)*16+16
282 %assign %%i xmm_regs_used
283 %rep (xmm_regs_used-6)
285 movdqa [rsp + (%%i-6)*16+8], xmm %+ %%i
294 %macro RESTORE_XMM_INTERNAL 1
295 %if xmm_regs_used > 6
296 %assign %%i xmm_regs_used
297 %rep (xmm_regs_used-6)
299 movdqa xmm %+ %%i, [%1 + (%%i-6)*16+8]
301 add %1, (xmm_regs_used-6)*16+16
306 RESTORE_XMM_INTERNAL %1
307 %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
308 %assign xmm_regs_used 0
312 RESTORE_XMM_INTERNAL rsp
321 %if regs_used > 4 || xmm_regs_used > 6
328 %elifdef ARCH_X86_64 ; *nix x64 ;=============================================
330 DECLARE_REG 0, rdi, edi, di, dil, edi
331 DECLARE_REG 1, rsi, esi, si, sil, esi
332 DECLARE_REG 2, rdx, edx, dx, dl, edx
333 DECLARE_REG 3, rcx, ecx, cx, cl, ecx
334 DECLARE_REG 4, r8, r8d, r8w, r8b, r8d
335 DECLARE_REG 5, r9, r9d, r9w, r9b, r9d
336 DECLARE_REG 6, rax, eax, ax, al, [rsp + stack_offset + 8]
337 %define r7m [rsp + stack_offset + 16]
338 %define r8m [rsp + stack_offset + 24]
340 %macro LOAD_IF_USED 2 ; reg_id, number_of_args
342 mov r%1, [rsp - 40 + %1*8]
346 %macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
361 %else ; X86_32 ;==============================================================
363 DECLARE_REG 0, eax, eax, ax, al, [esp + stack_offset + 4]
364 DECLARE_REG 1, ecx, ecx, cx, cl, [esp + stack_offset + 8]
365 DECLARE_REG 2, edx, edx, dx, dl, [esp + stack_offset + 12]
366 DECLARE_REG 3, ebx, ebx, bx, bl, [esp + stack_offset + 16]
367 DECLARE_REG 4, esi, esi, si, null, [esp + stack_offset + 20]
368 DECLARE_REG 5, edi, edi, di, null, [esp + stack_offset + 24]
369 DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28]
370 %define r7m [esp + stack_offset + 32]
371 %define r8m [esp + stack_offset + 36]
374 %macro PUSH_IF_USED 1 ; reg_id
377 %assign stack_offset stack_offset+4
381 %macro POP_IF_USED 1 ; reg_id
387 %macro LOAD_IF_USED 2 ; reg_id, number_of_args
389 mov r%1, [esp + stack_offset + 4 + %1*4]
393 %macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
396 ASSERT regs_used <= 7
427 %endif ;======================================================================
431 ;=============================================================================
432 ; arch-independent part
433 ;=============================================================================
435 %assign function_align 16
437 ; Symbol prefix for C linkage
439 %xdefine %1 mangle(%1)
440 %xdefine %1.skip_prologue %1 %+ .skip_prologue
441 %ifidn __OUTPUT_FORMAT__,elf
442 global %1:function hidden
448 RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
449 %assign stack_offset 0
456 %xdefine %1 mangle(%1)
460 ; This is needed for ELF, otherwise the GNU linker assumes the stack is
461 ; executable by default.
462 %ifidn __OUTPUT_FORMAT__,elf
463 SECTION .note.GNU-stack noalloc noexec nowrite progbits
477 %define RESET_MM_PERMUTATION INIT_MMX
486 CAT_XDEFINE m, %%i, mm %+ %%i
487 CAT_XDEFINE nmm, %%i, %%i
498 %define RESET_MM_PERMUTATION INIT_XMM
502 %define num_mmregs 16
507 %define movnt movntdq
510 CAT_XDEFINE m, %%i, xmm %+ %%i
511 CAT_XDEFINE nxmm, %%i, %%i
518 ; I often want to use macros that permute their arguments. e.g. there's no
519 ; efficient way to implement butterfly or transpose or dct without swapping some
522 ; I would like to not have to manually keep track of the permutations:
523 ; If I insert a permutation in the middle of a function, it should automatically
524 ; change everything that follows. For more complex macros I may also have multiple
525 ; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
527 ; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
528 ; permutes its arguments. It's equivalent to exchanging the contents of the
529 ; registers, except that this way you exchange the register names instead, so it
530 ; doesn't cost any cycles.
532 %macro PERMUTE 2-* ; takes a list of pairs to swap
547 %macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
553 CAT_XDEFINE n, m%1, %1
554 CAT_XDEFINE n, m%2, %2
556 ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
557 ; Be careful using this mode in nested macros though, as in some cases there may be
558 ; other copies of m# that have already been dereferenced and don't get updated correctly.
559 %xdefine %%n1 n %+ %1
560 %xdefine %%n2 n %+ %2
561 %xdefine tmp m %+ %%n1
562 CAT_XDEFINE m, %%n1, m %+ %%n2
563 CAT_XDEFINE m, %%n2, tmp
564 CAT_XDEFINE n, m %+ %%n1, %%n1
565 CAT_XDEFINE n, m %+ %%n2, %%n2
572 ; If SAVE_MM_PERMUTATION is placed at the end of a function and given the
573 ; function name, then any later calls to that function will automatically
574 ; load the permutation, so values can be returned in mmregs.
575 %macro SAVE_MM_PERMUTATION 1 ; name to save as
578 CAT_XDEFINE %1_m, %%i, m %+ %%i
583 %macro LOAD_MM_PERMUTATION 1 ; name to load from
586 CAT_XDEFINE m, %%i, %1_m %+ %%i
587 CAT_XDEFINE n, m %+ %%i, %%i
595 LOAD_MM_PERMUTATION %1
599 ; Substitutions that reduce instruction size but are functionally equivalent