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
43 ; FIXME: All of the 64bit asm functions that take a stride as an argument
44 ; via register, assume that the high dword of that register is filled with 0.
45 ; This is true in practice (since we never do any 64bit arithmetic on strides,
46 ; and x264's strides are all positive), but is not guaranteed by the ABI.
48 ; Name of the .rodata section.
49 ; Kludge: Something on OS X fails to align .rodata even given an align attribute,
50 ; so use a different read-only section.
51 %macro SECTION_RODATA 0-1 16
52 %ifidn __OUTPUT_FORMAT__,macho64
53 SECTION .text align=%1
54 %elifidn __OUTPUT_FORMAT__,macho
55 SECTION .text align=%1
58 SECTION .rodata align=%1
63 ; x86_64 can't fit 64bit address literals in most instruction types,
64 ; so shared objects (under the assumption that they might be anywhere
65 ; in memory) must use an address mode that does fit.
66 ; So all accesses to global variables must use this macro, e.g.
67 ; mov eax, [foo GLOBAL]
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.
81 %define GLOBAL wrt rip
86 ; Macros to eliminate most code duplication between x86_32 and x86_64:
87 ; Currently this works only for leaf functions which load all their arguments
88 ; into registers at the start, and make no other use of the stack. Luckily that
89 ; covers most of x264's asm.
92 ; %1 = number of arguments. loads them from stack if needed.
93 ; %2 = number of registers used. pushes callee-saved regs if needed.
94 ; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
95 ; %4 = list of names to define to registers
96 ; PROLOGUE can also be invoked by adding the same options to cglobal
99 ; cglobal foo, 2,3,0, dst, src, tmp
100 ; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
102 ; TODO Some functions can use some args directly from the stack. If they're the
103 ; last args then you can just not declare them, but if they're in the middle
104 ; we need more flexible macro.
107 ; Pops anything that was pushed by PROLOGUE
110 ; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
111 ; which are slow when a normal ret follows a branch.
114 ; rN and rNq are the native-size register holding function argument N
115 ; rNd, rNw, rNb are dword, word, and byte size
116 ; rNm is the original location of arg N (a register or on the stack), dword
117 ; rNmp is native size
125 %ifid %6 ; i.e. it's a register
127 %elifdef ARCH_X86_64 ; memory
128 %define r%1mp qword %6
130 %define r%1mp dword %6
135 %macro DECLARE_REG_SIZE 2
149 DECLARE_REG_SIZE ax, al
150 DECLARE_REG_SIZE bx, bl
151 DECLARE_REG_SIZE cx, cl
152 DECLARE_REG_SIZE dx, dl
153 DECLARE_REG_SIZE si, sil
154 DECLARE_REG_SIZE di, dil
155 DECLARE_REG_SIZE bp, bpl
157 ; t# defines for when per-arch register allocation is more complex than just function arguments
159 %macro DECLARE_REG_TMP 1-*
162 CAT_XDEFINE t, %%i, r%1
168 %macro DECLARE_REG_TMP_SIZE 0-*
170 %define t%1q t%1 %+ q
171 %define t%1d t%1 %+ d
172 %define t%1w t%1 %+ w
173 %define t%1b t%1 %+ b
178 DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7
188 %assign stack_offset stack_offset+gprsize
193 %assign stack_offset stack_offset-gprsize
199 %assign stack_offset stack_offset+(%2)
206 %assign stack_offset stack_offset-(%2)
216 %macro movsxdifnidn 2
228 %macro DEFINE_ARGS 0-*
232 CAT_UNDEF arg_name %+ %%i, q
233 CAT_UNDEF arg_name %+ %%i, d
234 CAT_UNDEF arg_name %+ %%i, w
235 CAT_UNDEF arg_name %+ %%i, b
236 CAT_UNDEF arg_name %+ %%i, m
237 CAT_UNDEF arg_name, %%i
244 %xdefine %1q r %+ %%i %+ q
245 %xdefine %1d r %+ %%i %+ d
246 %xdefine %1w r %+ %%i %+ w
247 %xdefine %1b r %+ %%i %+ b
248 %xdefine %1m r %+ %%i %+ m
249 CAT_XDEFINE arg_name, %%i, %1
253 %assign n_arg_names %%i
256 %ifdef WIN64 ; Windows x64 ;=================================================
258 DECLARE_REG 0, rcx, ecx, cx, cl, ecx
259 DECLARE_REG 1, rdx, edx, dx, dl, edx
260 DECLARE_REG 2, r8, r8d, r8w, r8b, r8d
261 DECLARE_REG 3, r9, r9d, r9w, r9b, r9d
262 DECLARE_REG 4, rdi, edi, di, dil, [rsp + stack_offset + 40]
263 DECLARE_REG 5, rsi, esi, si, sil, [rsp + stack_offset + 48]
264 DECLARE_REG 6, rax, eax, ax, al, [rsp + stack_offset + 56]
265 %define r7m [rsp + stack_offset + 64]
266 %define r8m [rsp + stack_offset + 72]
268 %macro LOAD_IF_USED 2 ; reg_id, number_of_args
270 mov r%1, [rsp + stack_offset + 8 + %1*8]
274 %macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
277 ASSERT regs_used <= 7
279 %assign xmm_regs_used %3
281 %assign xmm_regs_used 0
283 ASSERT xmm_regs_used <= 16
287 %assign stack_offset stack_offset+16
289 %if xmm_regs_used > 6
290 sub rsp, (xmm_regs_used-6)*16+16
291 %assign stack_offset stack_offset+(xmm_regs_used-6)*16+16
292 %assign %%i xmm_regs_used
293 %rep (xmm_regs_used-6)
295 movdqa [rsp + (%%i-6)*16+8], xmm %+ %%i
304 %macro RESTORE_XMM_INTERNAL 1
305 %if xmm_regs_used > 6
306 %assign %%i xmm_regs_used
307 %rep (xmm_regs_used-6)
309 movdqa xmm %+ %%i, [%1 + (%%i-6)*16+8]
311 add %1, (xmm_regs_used-6)*16+16
316 RESTORE_XMM_INTERNAL %1
317 %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
318 %assign xmm_regs_used 0
322 RESTORE_XMM_INTERNAL rsp
331 %if regs_used > 4 || xmm_regs_used > 6
338 %elifdef ARCH_X86_64 ; *nix x64 ;=============================================
340 DECLARE_REG 0, rdi, edi, di, dil, edi
341 DECLARE_REG 1, rsi, esi, si, sil, esi
342 DECLARE_REG 2, rdx, edx, dx, dl, edx
343 DECLARE_REG 3, rcx, ecx, cx, cl, ecx
344 DECLARE_REG 4, r8, r8d, r8w, r8b, r8d
345 DECLARE_REG 5, r9, r9d, r9w, r9b, r9d
346 DECLARE_REG 6, rax, eax, ax, al, [rsp + stack_offset + 8]
347 %define r7m [rsp + stack_offset + 16]
348 %define r8m [rsp + stack_offset + 24]
350 %macro LOAD_IF_USED 2 ; reg_id, number_of_args
352 mov r%1, [rsp - 40 + %1*8]
356 %macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
371 %else ; X86_32 ;==============================================================
373 DECLARE_REG 0, eax, eax, ax, al, [esp + stack_offset + 4]
374 DECLARE_REG 1, ecx, ecx, cx, cl, [esp + stack_offset + 8]
375 DECLARE_REG 2, edx, edx, dx, dl, [esp + stack_offset + 12]
376 DECLARE_REG 3, ebx, ebx, bx, bl, [esp + stack_offset + 16]
377 DECLARE_REG 4, esi, esi, si, null, [esp + stack_offset + 20]
378 DECLARE_REG 5, edi, edi, di, null, [esp + stack_offset + 24]
379 DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28]
380 %define r7m [esp + stack_offset + 32]
381 %define r8m [esp + stack_offset + 36]
384 %macro PUSH_IF_USED 1 ; reg_id
387 %assign stack_offset stack_offset+4
391 %macro POP_IF_USED 1 ; reg_id
397 %macro LOAD_IF_USED 2 ; reg_id, number_of_args
399 mov r%1, [esp + stack_offset + 4 + %1*4]
403 %macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
406 ASSERT regs_used <= 7
437 %endif ;======================================================================
441 ;=============================================================================
442 ; arch-independent part
443 ;=============================================================================
445 %assign function_align 16
447 ; Symbol prefix for C linkage
452 %xdefine %1.skip_prologue %1 %+ .skip_prologue
453 %ifidn __OUTPUT_FORMAT__,elf
454 global %1:function hidden
460 RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
461 %assign stack_offset 0
474 ; This is needed for ELF, otherwise the GNU linker assumes the stack is
475 ; executable by default.
476 %ifidn __OUTPUT_FORMAT__,elf
477 SECTION .note.GNU-stack noalloc noexec nowrite progbits
491 %define RESET_MM_PERMUTATION INIT_MMX
500 CAT_XDEFINE m, %%i, mm %+ %%i
501 CAT_XDEFINE nmm, %%i, %%i
512 %define RESET_MM_PERMUTATION INIT_XMM
516 %define num_mmregs 16
521 %define movnt movntdq
524 CAT_XDEFINE m, %%i, xmm %+ %%i
525 CAT_XDEFINE nxmm, %%i, %%i
532 ; I often want to use macros that permute their arguments. e.g. there's no
533 ; efficient way to implement butterfly or transpose or dct without swapping some
536 ; I would like to not have to manually keep track of the permutations:
537 ; If I insert a permutation in the middle of a function, it should automatically
538 ; change everything that follows. For more complex macros I may also have multiple
539 ; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
541 ; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
542 ; permutes its arguments. It's equivalent to exchanging the contents of the
543 ; registers, except that this way you exchange the register names instead, so it
544 ; doesn't cost any cycles.
546 %macro PERMUTE 2-* ; takes a list of pairs to swap
561 %macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
567 CAT_XDEFINE n, m%1, %1
568 CAT_XDEFINE n, m%2, %2
570 ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
571 ; Be careful using this mode in nested macros though, as in some cases there may be
572 ; other copies of m# that have already been dereferenced and don't get updated correctly.
573 %xdefine %%n1 n %+ %1
574 %xdefine %%n2 n %+ %2
575 %xdefine tmp m %+ %%n1
576 CAT_XDEFINE m, %%n1, m %+ %%n2
577 CAT_XDEFINE m, %%n2, tmp
578 CAT_XDEFINE n, m %+ %%n1, %%n1
579 CAT_XDEFINE n, m %+ %%n2, %%n2
586 ; If SAVE_MM_PERMUTATION is placed at the end of a function and given the
587 ; function name, then any later calls to that function will automatically
588 ; load the permutation, so values can be returned in mmregs.
589 %macro SAVE_MM_PERMUTATION 1 ; name to save as
592 CAT_XDEFINE %1_m, %%i, m %+ %%i
597 %macro LOAD_MM_PERMUTATION 1 ; name to load from
600 CAT_XDEFINE m, %%i, %1_m %+ %%i
601 CAT_XDEFINE n, m %+ %%i, %%i
609 LOAD_MM_PERMUTATION %1
613 ; Substitutions that reduce instruction size but are functionally equivalent