• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;*****************************************************************************
2;* x86inc.asm: x264asm abstraction layer
3;*****************************************************************************
4;* Copyright (C) 2005-2018 x264 project
5;*
6;* Authors: Loren Merritt <lorenm@u.washington.edu>
7;*          Henrik Gramner <henrik@gramner.com>
8;*          Anton Mitrofanov <BugMaster@narod.ru>
9;*          Fiona Glaser <fiona@x264.com>
10;*
11;* Permission to use, copy, modify, and/or distribute this software for any
12;* purpose with or without fee is hereby granted, provided that the above
13;* copyright notice and this permission notice appear in all copies.
14;*
15;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
16;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
17;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
18;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
19;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
20;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
21;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
22;*****************************************************************************
23
24; This is a header file for the x264ASM assembly language, which uses
25; NASM/YASM syntax combined with a large number of macros to provide easy
26; abstraction between different calling conventions (x86_32, win64, linux64).
27; It also has various other useful features to simplify writing the kind of
28; DSP functions that are most often used in x264.
29
30; Unlike the rest of x264, this file is available under an ISC license, as it
31; has significant usefulness outside of x264 and we want it to be available
32; to the largest audience possible.  Of course, if you modify it for your own
33; purposes to add a new feature, we strongly encourage contributing a patch
34; as this feature might be useful for others as well.  Send patches or ideas
35; to x264-devel@videolan.org .
36
37%ifndef private_prefix
38    %define private_prefix gst
39%endif
40
41%ifndef public_prefix
42    %define public_prefix private_prefix
43%endif
44
45%ifndef STACK_ALIGNMENT
46    %if ARCH_X86_64
47        %define STACK_ALIGNMENT 16
48    %else
49        %define STACK_ALIGNMENT 4
50    %endif
51%endif
52
53%define WIN64  0
54%define UNIX64 0
55%if ARCH_X86_64
56    %ifidn __OUTPUT_FORMAT__,win32
57        %define WIN64  1
58    %elifidn __OUTPUT_FORMAT__,win64
59        %define WIN64  1
60    %elifidn __OUTPUT_FORMAT__,x64
61        %define WIN64  1
62    %else
63        %define UNIX64 1
64    %endif
65%endif
66
67; Only 1 for yasm. Workaround here.
68%define HAVE_CPUNOP 0
69
70%define FORMAT_ELF 0
71%ifidn __OUTPUT_FORMAT__,elf
72    %define FORMAT_ELF 1
73%elifidn __OUTPUT_FORMAT__,elf32
74    %define FORMAT_ELF 1
75%elifidn __OUTPUT_FORMAT__,elf64
76    %define FORMAT_ELF 1
77%endif
78
79%ifdef PREFIX
80    %define mangle(x) _ %+ x
81%else
82    %define mangle(x) x
83%endif
84
85; aout does not support align=
86; NOTE: This section is out of sync with x264, in order to
87; keep supporting OS/2.
88%macro SECTION_RODATA 0-1 16
89    %ifidn __OUTPUT_FORMAT__,aout
90        SECTION .text
91    %elifidn __OUTPUT_FORMAT__,coff
92        SECTION .text
93    %elifidn __OUTPUT_FORMAT__,win32
94        SECTION .rdata align=%1
95    %elif WIN64
96        SECTION .rdata align=%1
97    %else
98        SECTION .rodata align=%1
99    %endif
100%endmacro
101
102%if WIN64
103    %define PIC
104%elif ARCH_X86_64 == 0
105; x86_32 doesn't require PIC.
106; Some distros prefer shared objects to be PIC, but nothing breaks if
107; the code contains a few textrels, so we'll skip that complexity.
108    %undef PIC
109%endif
110%ifdef PIC
111    default rel
112%endif
113
114%macro CPUNOP 1
115    %if HAVE_CPUNOP
116        CPU %1
117    %endif
118%endmacro
119
120; Macros to eliminate most code duplication between x86_32 and x86_64:
121; Currently this works only for leaf functions which load all their arguments
122; into registers at the start, and make no other use of the stack. Luckily that
123; covers most of x264's asm.
124
125; PROLOGUE:
126; %1 = number of arguments. loads them from stack if needed.
127; %2 = number of registers used. pushes callee-saved regs if needed.
128; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
129; %4 = (optional) stack size to be allocated. The stack will be aligned before
130;      allocating the specified stack size. If the required stack alignment is
131;      larger than the known stack alignment the stack will be manually aligned
132;      and an extra register will be allocated to hold the original stack
133;      pointer (to not invalidate r0m etc.). To prevent the use of an extra
134;      register as stack pointer, request a negative stack size.
135; %4+/%5+ = list of names to define to registers
136; PROLOGUE can also be invoked by adding the same options to cglobal
137
138; e.g.
139; cglobal foo, 2,3,7,0x40, dst, src, tmp
140; declares a function (foo) that automatically loads two arguments (dst and
141; src) into registers, uses one additional register (tmp) plus 7 vector
142; registers (m0-m6) and allocates 0x40 bytes of stack space.
143
144; TODO Some functions can use some args directly from the stack. If they're the
145; last args then you can just not declare them, but if they're in the middle
146; we need more flexible macro.
147
148; RET:
149; Pops anything that was pushed by PROLOGUE, and returns.
150
151; REP_RET:
152; Use this instead of RET if it's a branch target.
153
154; registers:
155; rN and rNq are the native-size register holding function argument N
156; rNd, rNw, rNb are dword, word, and byte size
157; rNh is the high 8 bits of the word size
158; rNm is the original location of arg N (a register or on the stack), dword
159; rNmp is native size
160
161%macro DECLARE_REG 2-3
162    %define r%1q %2
163    %define r%1d %2d
164    %define r%1w %2w
165    %define r%1b %2b
166    %define r%1h %2h
167    %define %2q %2
168    %if %0 == 2
169        %define r%1m  %2d
170        %define r%1mp %2
171    %elif ARCH_X86_64 ; memory
172        %define r%1m [rstk + stack_offset + %3]
173        %define r%1mp qword r %+ %1 %+ m
174    %else
175        %define r%1m [rstk + stack_offset + %3]
176        %define r%1mp dword r %+ %1 %+ m
177    %endif
178    %define r%1  %2
179%endmacro
180
181%macro DECLARE_REG_SIZE 3
182    %define r%1q r%1
183    %define e%1q r%1
184    %define r%1d e%1
185    %define e%1d e%1
186    %define r%1w %1
187    %define e%1w %1
188    %define r%1h %3
189    %define e%1h %3
190    %define r%1b %2
191    %define e%1b %2
192    %if ARCH_X86_64 == 0
193        %define r%1 e%1
194    %endif
195%endmacro
196
197DECLARE_REG_SIZE ax, al, ah
198DECLARE_REG_SIZE bx, bl, bh
199DECLARE_REG_SIZE cx, cl, ch
200DECLARE_REG_SIZE dx, dl, dh
201DECLARE_REG_SIZE si, sil, null
202DECLARE_REG_SIZE di, dil, null
203DECLARE_REG_SIZE bp, bpl, null
204
205; t# defines for when per-arch register allocation is more complex than just function arguments
206
207%macro DECLARE_REG_TMP 1-*
208    %assign %%i 0
209    %rep %0
210        CAT_XDEFINE t, %%i, r%1
211        %assign %%i %%i+1
212        %rotate 1
213    %endrep
214%endmacro
215
216%macro DECLARE_REG_TMP_SIZE 0-*
217    %rep %0
218        %define t%1q t%1 %+ q
219        %define t%1d t%1 %+ d
220        %define t%1w t%1 %+ w
221        %define t%1h t%1 %+ h
222        %define t%1b t%1 %+ b
223        %rotate 1
224    %endrep
225%endmacro
226
227DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
228
229%if ARCH_X86_64
230    %define gprsize 8
231%else
232    %define gprsize 4
233%endif
234
235%macro PUSH 1
236    push %1
237    %ifidn rstk, rsp
238        %assign stack_offset stack_offset+gprsize
239    %endif
240%endmacro
241
242%macro POP 1
243    pop %1
244    %ifidn rstk, rsp
245        %assign stack_offset stack_offset-gprsize
246    %endif
247%endmacro
248
249%macro PUSH_IF_USED 1-*
250    %rep %0
251        %if %1 < regs_used
252            PUSH r%1
253        %endif
254        %rotate 1
255    %endrep
256%endmacro
257
258%macro POP_IF_USED 1-*
259    %rep %0
260        %if %1 < regs_used
261            pop r%1
262        %endif
263        %rotate 1
264    %endrep
265%endmacro
266
267%macro LOAD_IF_USED 1-*
268    %rep %0
269        %if %1 < num_args
270            mov r%1, r %+ %1 %+ mp
271        %endif
272        %rotate 1
273    %endrep
274%endmacro
275
276%macro SUB 2
277    sub %1, %2
278    %ifidn %1, rstk
279        %assign stack_offset stack_offset+(%2)
280    %endif
281%endmacro
282
283%macro ADD 2
284    add %1, %2
285    %ifidn %1, rstk
286        %assign stack_offset stack_offset-(%2)
287    %endif
288%endmacro
289
290%macro movifnidn 2
291    %ifnidn %1, %2
292        mov %1, %2
293    %endif
294%endmacro
295
296%macro movsxdifnidn 2
297    %ifnidn %1, %2
298        movsxd %1, %2
299    %endif
300%endmacro
301
302%macro ASSERT 1
303    %if (%1) == 0
304        %error assertion ``%1'' failed
305    %endif
306%endmacro
307
308%macro DEFINE_ARGS 0-*
309    %ifdef n_arg_names
310        %assign %%i 0
311        %rep n_arg_names
312            CAT_UNDEF arg_name %+ %%i, q
313            CAT_UNDEF arg_name %+ %%i, d
314            CAT_UNDEF arg_name %+ %%i, w
315            CAT_UNDEF arg_name %+ %%i, h
316            CAT_UNDEF arg_name %+ %%i, b
317            CAT_UNDEF arg_name %+ %%i, m
318            CAT_UNDEF arg_name %+ %%i, mp
319            CAT_UNDEF arg_name, %%i
320            %assign %%i %%i+1
321        %endrep
322    %endif
323
324    %xdefine %%stack_offset stack_offset
325    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
326    %assign %%i 0
327    %rep %0
328        %xdefine %1q r %+ %%i %+ q
329        %xdefine %1d r %+ %%i %+ d
330        %xdefine %1w r %+ %%i %+ w
331        %xdefine %1h r %+ %%i %+ h
332        %xdefine %1b r %+ %%i %+ b
333        %xdefine %1m r %+ %%i %+ m
334        %xdefine %1mp r %+ %%i %+ mp
335        CAT_XDEFINE arg_name, %%i, %1
336        %assign %%i %%i+1
337        %rotate 1
338    %endrep
339    %xdefine stack_offset %%stack_offset
340    %assign n_arg_names %0
341%endmacro
342
343%define required_stack_alignment ((mmsize + 15) & ~15)
344%define vzeroupper_required (mmsize > 16 && (ARCH_X86_64 == 0 || xmm_regs_used > 16 || notcpuflag(avx512)))
345%define high_mm_regs (16*cpuflag(avx512))
346
347%macro ALLOC_STACK 1-2 0 ; stack_size, n_xmm_regs (for win64 only)
348    %ifnum %1
349        %if %1 != 0
350            %assign %%pad 0
351            %assign stack_size %1
352            %if stack_size < 0
353                %assign stack_size -stack_size
354            %endif
355            %if WIN64
356                %assign %%pad %%pad + 32 ; shadow space
357                %if mmsize != 8
358                    %assign xmm_regs_used %2
359                    %if xmm_regs_used > 8
360                        %assign %%pad %%pad + (xmm_regs_used-8)*16 ; callee-saved xmm registers
361                    %endif
362                %endif
363            %endif
364            %if required_stack_alignment <= STACK_ALIGNMENT
365                ; maintain the current stack alignment
366                %assign stack_size_padded stack_size + %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
367                SUB rsp, stack_size_padded
368            %else
369                %assign %%reg_num (regs_used - 1)
370                %xdefine rstk r %+ %%reg_num
371                ; align stack, and save original stack location directly above
372                ; it, i.e. in [rsp+stack_size_padded], so we can restore the
373                ; stack in a single instruction (i.e. mov rsp, rstk or mov
374                ; rsp, [rsp+stack_size_padded])
375                %if %1 < 0 ; need to store rsp on stack
376                    %xdefine rstkm [rsp + stack_size + %%pad]
377                    %assign %%pad %%pad + gprsize
378                %else ; can keep rsp in rstk during whole function
379                    %xdefine rstkm rstk
380                %endif
381                %assign stack_size_padded stack_size + ((%%pad + required_stack_alignment-1) & ~(required_stack_alignment-1))
382                mov rstk, rsp
383                and rsp, ~(required_stack_alignment-1)
384                sub rsp, stack_size_padded
385                movifnidn rstkm, rstk
386            %endif
387            WIN64_PUSH_XMM
388        %endif
389    %endif
390%endmacro
391
392%macro SETUP_STACK_POINTER 1
393    %ifnum %1
394        %if %1 != 0 && required_stack_alignment > STACK_ALIGNMENT
395            %if %1 > 0
396                ; Reserve an additional register for storing the original stack pointer, but avoid using
397                ; eax/rax for this purpose since it can potentially get overwritten as a return value.
398                %assign regs_used (regs_used + 1)
399                %if ARCH_X86_64 && regs_used == 7
400                    %assign regs_used 8
401                %elif ARCH_X86_64 == 0 && regs_used == 1
402                    %assign regs_used 2
403                %endif
404            %endif
405            %if ARCH_X86_64 && regs_used < 5 + UNIX64 * 3
406                ; Ensure that we don't clobber any registers containing arguments. For UNIX64 we also preserve r6 (rax)
407                ; since it's used as a hidden argument in vararg functions to specify the number of vector registers used.
408                %assign regs_used 5 + UNIX64 * 3
409            %endif
410        %endif
411    %endif
412%endmacro
413
414%macro DEFINE_ARGS_INTERNAL 3+
415    %ifnum %2
416        DEFINE_ARGS %3
417    %elif %1 == 4
418        DEFINE_ARGS %2
419    %elif %1 > 4
420        DEFINE_ARGS %2, %3
421    %endif
422%endmacro
423
424%if WIN64 ; Windows x64 ;=================================================
425
426DECLARE_REG 0,  rcx
427DECLARE_REG 1,  rdx
428DECLARE_REG 2,  R8
429DECLARE_REG 3,  R9
430DECLARE_REG 4,  R10, 40
431DECLARE_REG 5,  R11, 48
432DECLARE_REG 6,  rax, 56
433DECLARE_REG 7,  rdi, 64
434DECLARE_REG 8,  rsi, 72
435DECLARE_REG 9,  rbx, 80
436DECLARE_REG 10, rbp, 88
437DECLARE_REG 11, R14, 96
438DECLARE_REG 12, R15, 104
439DECLARE_REG 13, R12, 112
440DECLARE_REG 14, R13, 120
441
442%macro PROLOGUE 2-5+ 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
443    %assign num_args %1
444    %assign regs_used %2
445    ASSERT regs_used >= num_args
446    SETUP_STACK_POINTER %4
447    ASSERT regs_used <= 15
448    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
449    ALLOC_STACK %4, %3
450    %if mmsize != 8 && stack_size == 0
451        WIN64_SPILL_XMM %3
452    %endif
453    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
454    DEFINE_ARGS_INTERNAL %0, %4, %5
455%endmacro
456
457%macro WIN64_PUSH_XMM 0
458    ; Use the shadow space to store XMM6 and XMM7, the rest needs stack space allocated.
459    %if xmm_regs_used > 6 + high_mm_regs
460        movaps [rstk + stack_offset +  8], xmm6
461    %endif
462    %if xmm_regs_used > 7 + high_mm_regs
463        movaps [rstk + stack_offset + 24], xmm7
464    %endif
465    %assign %%xmm_regs_on_stack xmm_regs_used - high_mm_regs - 8
466    %if %%xmm_regs_on_stack > 0
467        %assign %%i 8
468        %rep %%xmm_regs_on_stack
469            movaps [rsp + (%%i-8)*16 + stack_size + 32], xmm %+ %%i
470            %assign %%i %%i+1
471        %endrep
472    %endif
473%endmacro
474
475%macro WIN64_SPILL_XMM 1
476    %assign xmm_regs_used %1
477    ASSERT xmm_regs_used <= 16 + high_mm_regs
478    %assign %%xmm_regs_on_stack xmm_regs_used - high_mm_regs - 8
479    %if %%xmm_regs_on_stack > 0
480        ; Allocate stack space for callee-saved xmm registers plus shadow space and align the stack.
481        %assign %%pad %%xmm_regs_on_stack*16 + 32
482        %assign stack_size_padded %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
483        SUB rsp, stack_size_padded
484    %endif
485    WIN64_PUSH_XMM
486%endmacro
487
488%macro WIN64_RESTORE_XMM_INTERNAL 0
489    %assign %%pad_size 0
490    %assign %%xmm_regs_on_stack xmm_regs_used - high_mm_regs - 8
491    %if %%xmm_regs_on_stack > 0
492        %assign %%i xmm_regs_used - high_mm_regs
493        %rep %%xmm_regs_on_stack
494            %assign %%i %%i-1
495            movaps xmm %+ %%i, [rsp + (%%i-8)*16 + stack_size + 32]
496        %endrep
497    %endif
498    %if stack_size_padded > 0
499        %if stack_size > 0 && required_stack_alignment > STACK_ALIGNMENT
500            mov rsp, rstkm
501        %else
502            add rsp, stack_size_padded
503            %assign %%pad_size stack_size_padded
504        %endif
505    %endif
506    %if xmm_regs_used > 7 + high_mm_regs
507        movaps xmm7, [rsp + stack_offset - %%pad_size + 24]
508    %endif
509    %if xmm_regs_used > 6 + high_mm_regs
510        movaps xmm6, [rsp + stack_offset - %%pad_size +  8]
511    %endif
512%endmacro
513
514%macro WIN64_RESTORE_XMM 0
515    WIN64_RESTORE_XMM_INTERNAL
516    %assign stack_offset (stack_offset-stack_size_padded)
517    %assign stack_size_padded 0
518    %assign xmm_regs_used 0
519%endmacro
520
521%define has_epilogue regs_used > 7 || stack_size > 0 || vzeroupper_required || xmm_regs_used > 6+high_mm_regs
522
523%macro RET 0
524    WIN64_RESTORE_XMM_INTERNAL
525    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
526    %if vzeroupper_required
527        vzeroupper
528    %endif
529    AUTO_REP_RET
530%endmacro
531
532%elif ARCH_X86_64 ; *nix x64 ;=============================================
533
534DECLARE_REG 0,  rdi
535DECLARE_REG 1,  rsi
536DECLARE_REG 2,  rdx
537DECLARE_REG 3,  rcx
538DECLARE_REG 4,  R8
539DECLARE_REG 5,  R9
540DECLARE_REG 6,  rax, 8
541DECLARE_REG 7,  R10, 16
542DECLARE_REG 8,  R11, 24
543DECLARE_REG 9,  rbx, 32
544DECLARE_REG 10, rbp, 40
545DECLARE_REG 11, R14, 48
546DECLARE_REG 12, R15, 56
547DECLARE_REG 13, R12, 64
548DECLARE_REG 14, R13, 72
549
550%macro PROLOGUE 2-5+ 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
551    %assign num_args %1
552    %assign regs_used %2
553    %assign xmm_regs_used %3
554    ASSERT regs_used >= num_args
555    SETUP_STACK_POINTER %4
556    ASSERT regs_used <= 15
557    PUSH_IF_USED 9, 10, 11, 12, 13, 14
558    ALLOC_STACK %4
559    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
560    DEFINE_ARGS_INTERNAL %0, %4, %5
561%endmacro
562
563%define has_epilogue regs_used > 9 || stack_size > 0 || vzeroupper_required
564
565%macro RET 0
566    %if stack_size_padded > 0
567        %if required_stack_alignment > STACK_ALIGNMENT
568            mov rsp, rstkm
569        %else
570            add rsp, stack_size_padded
571        %endif
572    %endif
573    POP_IF_USED 14, 13, 12, 11, 10, 9
574    %if vzeroupper_required
575        vzeroupper
576    %endif
577    AUTO_REP_RET
578%endmacro
579
580%else ; X86_32 ;==============================================================
581
582DECLARE_REG 0, eax, 4
583DECLARE_REG 1, ecx, 8
584DECLARE_REG 2, edx, 12
585DECLARE_REG 3, ebx, 16
586DECLARE_REG 4, esi, 20
587DECLARE_REG 5, edi, 24
588DECLARE_REG 6, ebp, 28
589%define rsp esp
590
591%macro DECLARE_ARG 1-*
592    %rep %0
593        %define r%1m [rstk + stack_offset + 4*%1 + 4]
594        %define r%1mp dword r%1m
595        %rotate 1
596    %endrep
597%endmacro
598
599DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
600
601%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
602    %assign num_args %1
603    %assign regs_used %2
604    ASSERT regs_used >= num_args
605    %if num_args > 7
606        %assign num_args 7
607    %endif
608    %if regs_used > 7
609        %assign regs_used 7
610    %endif
611    SETUP_STACK_POINTER %4
612    ASSERT regs_used <= 7
613    PUSH_IF_USED 3, 4, 5, 6
614    ALLOC_STACK %4
615    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
616    DEFINE_ARGS_INTERNAL %0, %4, %5
617%endmacro
618
619%define has_epilogue regs_used > 3 || stack_size > 0 || vzeroupper_required
620
621%macro RET 0
622    %if stack_size_padded > 0
623        %if required_stack_alignment > STACK_ALIGNMENT
624            mov rsp, rstkm
625        %else
626            add rsp, stack_size_padded
627        %endif
628    %endif
629    POP_IF_USED 6, 5, 4, 3
630    %if vzeroupper_required
631        vzeroupper
632    %endif
633    AUTO_REP_RET
634%endmacro
635
636%endif ;======================================================================
637
638%if WIN64 == 0
639    %macro WIN64_SPILL_XMM 1
640    %endmacro
641    %macro WIN64_RESTORE_XMM 0
642    %endmacro
643    %macro WIN64_PUSH_XMM 0
644    %endmacro
645%endif
646
647; On AMD cpus <=K10, an ordinary ret is slow if it immediately follows either
648; a branch or a branch target. So switch to a 2-byte form of ret in that case.
649; We can automatically detect "follows a branch", but not a branch target.
650; (SSSE3 is a sufficient condition to know that your cpu doesn't have this problem.)
651%macro REP_RET 0
652    %if has_epilogue || cpuflag(ssse3)
653        RET
654    %else
655        rep ret
656    %endif
657    annotate_function_size
658%endmacro
659
660%define last_branch_adr $$
661%macro AUTO_REP_RET 0
662    %if notcpuflag(ssse3)
663        times ((last_branch_adr-$)>>31)+1 rep ; times 1 iff $ == last_branch_adr.
664    %endif
665    ret
666    annotate_function_size
667%endmacro
668
669%macro BRANCH_INSTR 0-*
670    %rep %0
671        %macro %1 1-2 %1
672            %2 %1
673            %if notcpuflag(ssse3)
674                %%branch_instr equ $
675                %xdefine last_branch_adr %%branch_instr
676            %endif
677        %endmacro
678        %rotate 1
679    %endrep
680%endmacro
681
682BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae, jna, jnae, jb, jbe, jnb, jnbe, jc, jnc, js, jns, jo, jno, jp, jnp
683
684%macro TAIL_CALL 2 ; callee, is_nonadjacent
685    %if has_epilogue
686        call %1
687        RET
688    %elif %2
689        jmp %1
690    %endif
691    annotate_function_size
692%endmacro
693
694;=============================================================================
695; arch-independent part
696;=============================================================================
697
698%assign function_align 16
699
700; Begin a function.
701; Applies any symbol mangling needed for C linkage, and sets up a define such that
702; subsequent uses of the function name automatically refer to the mangled version.
703; Appends cpuflags to the function name if cpuflags has been specified.
704; The "" empty default parameter is a workaround for nasm, which fails if SUFFIX
705; is empty and we call cglobal_internal with just %1 %+ SUFFIX (without %2).
706%macro cglobal 1-2+ "" ; name, [PROLOGUE args]
707    cglobal_internal 1, %1 %+ SUFFIX, %2
708%endmacro
709%macro cvisible 1-2+ "" ; name, [PROLOGUE args]
710    cglobal_internal 0, %1 %+ SUFFIX, %2
711%endmacro
712%macro cglobal_internal 2-3+
713    annotate_function_size
714    %if %1
715        %xdefine %%FUNCTION_PREFIX private_prefix
716        %xdefine %%VISIBILITY hidden
717    %else
718        %xdefine %%FUNCTION_PREFIX public_prefix
719        %xdefine %%VISIBILITY
720    %endif
721    %ifndef cglobaled_%2
722        %xdefine %2 mangle(%%FUNCTION_PREFIX %+ _ %+ %2)
723        %xdefine %2.skip_prologue %2 %+ .skip_prologue
724        CAT_XDEFINE cglobaled_, %2, 1
725    %endif
726    %xdefine current_function %2
727    %xdefine current_function_section __SECT__
728    %if FORMAT_ELF
729        global %2:function %%VISIBILITY
730    %else
731        global %2
732    %endif
733    align function_align
734    %2:
735    RESET_MM_PERMUTATION        ; needed for x86-64, also makes disassembly somewhat nicer
736    %xdefine rstk rsp           ; copy of the original stack pointer, used when greater alignment than the known stack alignment is required
737    %assign stack_offset 0      ; stack pointer offset relative to the return address
738    %assign stack_size 0        ; amount of stack space that can be freely used inside a function
739    %assign stack_size_padded 0 ; total amount of allocated stack space, including space for callee-saved xmm registers on WIN64 and alignment padding
740    %assign xmm_regs_used 0     ; number of XMM registers requested, used for dealing with callee-saved registers on WIN64 and vzeroupper
741    %ifnidn %3, ""
742        PROLOGUE %3
743    %endif
744%endmacro
745
746; Create a global symbol from a local label with the correct name mangling and type
747%macro cglobal_label 1
748    %if FORMAT_ELF
749        global current_function %+ %1:function hidden
750    %else
751        global current_function %+ %1
752    %endif
753    %1:
754%endmacro
755
756%macro cextern 1
757    %xdefine %1 mangle(private_prefix %+ _ %+ %1)
758    CAT_XDEFINE cglobaled_, %1, 1
759    extern %1
760%endmacro
761
762; like cextern, but without the prefix
763%macro cextern_naked 1
764    %ifdef PREFIX
765        %xdefine %1 mangle(%1)
766    %endif
767    CAT_XDEFINE cglobaled_, %1, 1
768    extern %1
769%endmacro
770
771%macro const 1-2+
772    %xdefine %1 mangle(private_prefix %+ _ %+ %1)
773    %if FORMAT_ELF
774        global %1:data hidden
775    %else
776        global %1
777    %endif
778    %1: %2
779%endmacro
780
781; This is needed for ELF, otherwise the GNU linker assumes the stack is executable by default.
782%if FORMAT_ELF
783    [SECTION .note.GNU-stack noalloc noexec nowrite progbits]
784%endif
785
786; Tell debuggers how large the function was.
787; This may be invoked multiple times per function; we rely on later instances overriding earlier ones.
788; This is invoked by RET and similar macros, and also cglobal does it for the previous function,
789; but if the last function in a source file doesn't use any of the standard macros for its epilogue,
790; then its size might be unspecified.
791%macro annotate_function_size 0
792    %ifdef __YASM_VER__
793        %ifdef current_function
794            %if FORMAT_ELF
795                current_function_section
796                %%ecf equ $
797                size current_function %%ecf - current_function
798                __SECT__
799            %endif
800        %endif
801    %endif
802%endmacro
803
804; cpuflags
805
806%assign cpuflags_mmx      (1<<0)
807%assign cpuflags_mmx2     (1<<1) | cpuflags_mmx
808%assign cpuflags_3dnow    (1<<2) | cpuflags_mmx
809%assign cpuflags_3dnowext (1<<3) | cpuflags_3dnow
810%assign cpuflags_sse      (1<<4) | cpuflags_mmx2
811%assign cpuflags_sse2     (1<<5) | cpuflags_sse
812%assign cpuflags_sse2slow (1<<6) | cpuflags_sse2
813%assign cpuflags_lzcnt    (1<<7) | cpuflags_sse2
814%assign cpuflags_sse3     (1<<8) | cpuflags_sse2
815%assign cpuflags_ssse3    (1<<9) | cpuflags_sse3
816%assign cpuflags_sse4     (1<<10)| cpuflags_ssse3
817%assign cpuflags_sse42    (1<<11)| cpuflags_sse4
818%assign cpuflags_aesni    (1<<12)| cpuflags_sse42
819%assign cpuflags_avx      (1<<13)| cpuflags_sse42
820%assign cpuflags_xop      (1<<14)| cpuflags_avx
821%assign cpuflags_fma4     (1<<15)| cpuflags_avx
822%assign cpuflags_fma3     (1<<16)| cpuflags_avx
823%assign cpuflags_bmi1     (1<<17)| cpuflags_avx|cpuflags_lzcnt
824%assign cpuflags_bmi2     (1<<18)| cpuflags_bmi1
825%assign cpuflags_avx2     (1<<19)| cpuflags_fma3|cpuflags_bmi2
826%assign cpuflags_avx512   (1<<20)| cpuflags_avx2 ; F, CD, BW, DQ, VL
827
828%assign cpuflags_cache32  (1<<21)
829%assign cpuflags_cache64  (1<<22)
830%assign cpuflags_aligned  (1<<23) ; not a cpu feature, but a function variant
831%assign cpuflags_atom     (1<<24)
832
833; Returns a boolean value expressing whether or not the specified cpuflag is enabled.
834%define    cpuflag(x) (((((cpuflags & (cpuflags_ %+ x)) ^ (cpuflags_ %+ x)) - 1) >> 31) & 1)
835%define notcpuflag(x) (cpuflag(x) ^ 1)
836
837; Takes an arbitrary number of cpuflags from the above list.
838; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
839; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
840%macro INIT_CPUFLAGS 0-*
841    %xdefine SUFFIX
842    %undef cpuname
843    %assign cpuflags 0
844
845    %if %0 >= 1
846        %rep %0
847            %ifdef cpuname
848                %xdefine cpuname cpuname %+ _%1
849            %else
850                %xdefine cpuname %1
851            %endif
852            %assign cpuflags cpuflags | cpuflags_%1
853            %rotate 1
854        %endrep
855        %xdefine SUFFIX _ %+ cpuname
856
857        %if cpuflag(avx)
858            %assign avx_enabled 1
859        %endif
860        %if (mmsize == 16 && notcpuflag(sse2)) || (mmsize == 32 && notcpuflag(avx2))
861            %define mova movaps
862            %define movu movups
863            %define movnta movntps
864        %endif
865        %if cpuflag(aligned)
866            %define movu mova
867        %elif cpuflag(sse3) && notcpuflag(ssse3)
868            %define movu lddqu
869        %endif
870    %endif
871
872    %if ARCH_X86_64 || cpuflag(sse2)
873        CPUNOP amdnop
874    %else
875        CPUNOP basicnop
876    %endif
877%endmacro
878
879; Merge mmx, sse*, and avx*
880; m# is a simd register of the currently selected size
881; xm# is the corresponding xmm register if mmsize >= 16, otherwise the same as m#
882; ym# is the corresponding ymm register if mmsize >= 32, otherwise the same as m#
883; zm# is the corresponding zmm register if mmsize >= 64, otherwise the same as m#
884; (All 4 remain in sync through SWAP.)
885
886%macro CAT_XDEFINE 3
887    %xdefine %1%2 %3
888%endmacro
889
890%macro CAT_UNDEF 2
891    %undef %1%2
892%endmacro
893
894%macro DEFINE_MMREGS 1 ; mmtype
895    %assign %%prev_mmregs 0
896    %ifdef num_mmregs
897        %assign %%prev_mmregs num_mmregs
898    %endif
899
900    %assign num_mmregs 8
901    %if ARCH_X86_64 && mmsize >= 16
902        %assign num_mmregs 16
903        %if cpuflag(avx512) || mmsize == 64
904            %assign num_mmregs 32
905        %endif
906    %endif
907
908    %assign %%i 0
909    %rep num_mmregs
910        CAT_XDEFINE m, %%i, %1 %+ %%i
911        CAT_XDEFINE nn%1, %%i, %%i
912        %assign %%i %%i+1
913    %endrep
914    %if %%prev_mmregs > num_mmregs
915        %rep %%prev_mmregs - num_mmregs
916            CAT_UNDEF m, %%i
917            CAT_UNDEF nn %+ mmtype, %%i
918            %assign %%i %%i+1
919        %endrep
920    %endif
921    %xdefine mmtype %1
922%endmacro
923
924; Prefer registers 16-31 over 0-15 to avoid having to use vzeroupper
925%macro AVX512_MM_PERMUTATION 0-1 0 ; start_reg
926    %if ARCH_X86_64 && cpuflag(avx512)
927        %assign %%i %1
928        %rep 16-%1
929            %assign %%i_high %%i+16
930            SWAP %%i, %%i_high
931            %assign %%i %%i+1
932        %endrep
933    %endif
934%endmacro
935
936%macro INIT_MMX 0-1+
937    %assign avx_enabled 0
938    %define RESET_MM_PERMUTATION INIT_MMX %1
939    %define mmsize 8
940    %define mova movq
941    %define movu movq
942    %define movh movd
943    %define movnta movntq
944    INIT_CPUFLAGS %1
945    DEFINE_MMREGS mm
946%endmacro
947
948%macro INIT_XMM 0-1+
949    %assign avx_enabled 0
950    %define RESET_MM_PERMUTATION INIT_XMM %1
951    %define mmsize 16
952    %define mova movdqa
953    %define movu movdqu
954    %define movh movq
955    %define movnta movntdq
956    INIT_CPUFLAGS %1
957    DEFINE_MMREGS xmm
958    %if WIN64
959        AVX512_MM_PERMUTATION 6 ; Swap callee-saved registers with volatile registers
960    %endif
961%endmacro
962
963%macro INIT_YMM 0-1+
964    %assign avx_enabled 1
965    %define RESET_MM_PERMUTATION INIT_YMM %1
966    %define mmsize 32
967    %define mova movdqa
968    %define movu movdqu
969    %undef movh
970    %define movnta movntdq
971    INIT_CPUFLAGS %1
972    DEFINE_MMREGS ymm
973    AVX512_MM_PERMUTATION
974%endmacro
975
976%macro INIT_ZMM 0-1+
977    %assign avx_enabled 1
978    %define RESET_MM_PERMUTATION INIT_ZMM %1
979    %define mmsize 64
980    %define mova movdqa
981    %define movu movdqu
982    %undef movh
983    %define movnta movntdq
984    INIT_CPUFLAGS %1
985    DEFINE_MMREGS zmm
986    AVX512_MM_PERMUTATION
987%endmacro
988
989INIT_XMM
990
991%macro DECLARE_MMCAST 1
992    %define  mmmm%1   mm%1
993    %define  mmxmm%1  mm%1
994    %define  mmymm%1  mm%1
995    %define  mmzmm%1  mm%1
996    %define xmmmm%1   mm%1
997    %define xmmxmm%1 xmm%1
998    %define xmmymm%1 xmm%1
999    %define xmmzmm%1 xmm%1
1000    %define ymmmm%1   mm%1
1001    %define ymmxmm%1 xmm%1
1002    %define ymmymm%1 ymm%1
1003    %define ymmzmm%1 ymm%1
1004    %define zmmmm%1   mm%1
1005    %define zmmxmm%1 xmm%1
1006    %define zmmymm%1 ymm%1
1007    %define zmmzmm%1 zmm%1
1008    %define xm%1 xmm %+ m%1
1009    %define ym%1 ymm %+ m%1
1010    %define zm%1 zmm %+ m%1
1011%endmacro
1012
1013%assign i 0
1014%rep 32
1015    DECLARE_MMCAST i
1016    %assign i i+1
1017%endrep
1018
1019; I often want to use macros that permute their arguments. e.g. there's no
1020; efficient way to implement butterfly or transpose or dct without swapping some
1021; arguments.
1022;
1023; I would like to not have to manually keep track of the permutations:
1024; If I insert a permutation in the middle of a function, it should automatically
1025; change everything that follows. For more complex macros I may also have multiple
1026; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
1027;
1028; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
1029; permutes its arguments. It's equivalent to exchanging the contents of the
1030; registers, except that this way you exchange the register names instead, so it
1031; doesn't cost any cycles.
1032
1033%macro PERMUTE 2-* ; takes a list of pairs to swap
1034    %rep %0/2
1035        %xdefine %%tmp%2 m%2
1036        %rotate 2
1037    %endrep
1038    %rep %0/2
1039        %xdefine m%1 %%tmp%2
1040        CAT_XDEFINE nn, m%1, %1
1041        %rotate 2
1042    %endrep
1043%endmacro
1044
1045%macro SWAP 2+ ; swaps a single chain (sometimes more concise than pairs)
1046    %ifnum %1 ; SWAP 0, 1, ...
1047        SWAP_INTERNAL_NUM %1, %2
1048    %else ; SWAP m0, m1, ...
1049        SWAP_INTERNAL_NAME %1, %2
1050    %endif
1051%endmacro
1052
1053%macro SWAP_INTERNAL_NUM 2-*
1054    %rep %0-1
1055        %xdefine %%tmp m%1
1056        %xdefine m%1 m%2
1057        %xdefine m%2 %%tmp
1058        CAT_XDEFINE nn, m%1, %1
1059        CAT_XDEFINE nn, m%2, %2
1060        %rotate 1
1061    %endrep
1062%endmacro
1063
1064%macro SWAP_INTERNAL_NAME 2-*
1065    %xdefine %%args nn %+ %1
1066    %rep %0-1
1067        %xdefine %%args %%args, nn %+ %2
1068        %rotate 1
1069    %endrep
1070    SWAP_INTERNAL_NUM %%args
1071%endmacro
1072
1073; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
1074; calls to that function will automatically load the permutation, so values can
1075; be returned in mmregs.
1076%macro SAVE_MM_PERMUTATION 0-1
1077    %if %0
1078        %xdefine %%f %1_m
1079    %else
1080        %xdefine %%f current_function %+ _m
1081    %endif
1082    %assign %%i 0
1083    %rep num_mmregs
1084        CAT_XDEFINE %%f, %%i, m %+ %%i
1085        %assign %%i %%i+1
1086    %endrep
1087%endmacro
1088
1089%macro LOAD_MM_PERMUTATION 1 ; name to load from
1090    %ifdef %1_m0
1091        %assign %%i 0
1092        %rep num_mmregs
1093            CAT_XDEFINE m, %%i, %1_m %+ %%i
1094            CAT_XDEFINE nn, m %+ %%i, %%i
1095            %assign %%i %%i+1
1096        %endrep
1097    %endif
1098%endmacro
1099
1100; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
1101%macro call 1
1102    %ifid %1
1103        call_internal %1 %+ SUFFIX, %1
1104    %else
1105        call %1
1106    %endif
1107%endmacro
1108%macro call_internal 2
1109    %xdefine %%i %2
1110    %ifndef cglobaled_%2
1111        %ifdef cglobaled_%1
1112            %xdefine %%i %1
1113        %endif
1114    %endif
1115    call %%i
1116    LOAD_MM_PERMUTATION %%i
1117%endmacro
1118
1119; Substitutions that reduce instruction size but are functionally equivalent
1120%macro add 2
1121    %ifnum %2
1122        %if %2==128
1123            sub %1, -128
1124        %else
1125            add %1, %2
1126        %endif
1127    %else
1128        add %1, %2
1129    %endif
1130%endmacro
1131
1132%macro sub 2
1133    %ifnum %2
1134        %if %2==128
1135            add %1, -128
1136        %else
1137            sub %1, %2
1138        %endif
1139    %else
1140        sub %1, %2
1141    %endif
1142%endmacro
1143
1144;=============================================================================
1145; AVX abstraction layer
1146;=============================================================================
1147
1148%assign i 0
1149%rep 32
1150    %if i < 8
1151        CAT_XDEFINE sizeofmm, i, 8
1152        CAT_XDEFINE regnumofmm, i, i
1153    %endif
1154    CAT_XDEFINE sizeofxmm, i, 16
1155    CAT_XDEFINE sizeofymm, i, 32
1156    CAT_XDEFINE sizeofzmm, i, 64
1157    CAT_XDEFINE regnumofxmm, i, i
1158    CAT_XDEFINE regnumofymm, i, i
1159    CAT_XDEFINE regnumofzmm, i, i
1160    %assign i i+1
1161%endrep
1162%undef i
1163
1164%macro CHECK_AVX_INSTR_EMU 3-*
1165    %xdefine %%opcode %1
1166    %xdefine %%dst %2
1167    %rep %0-2
1168        %ifidn %%dst, %3
1169            %error non-avx emulation of ``%%opcode'' is not supported
1170        %endif
1171        %rotate 1
1172    %endrep
1173%endmacro
1174
1175;%1 == instruction
1176;%2 == minimal instruction set
1177;%3 == 1 if float, 0 if int
1178;%4 == 1 if 4-operand emulation, 0 if 3-operand emulation, 255 otherwise (no emulation)
1179;%5 == 1 if commutative (i.e. doesn't matter which src arg is which), 0 if not
1180;%6+: operands
1181%macro RUN_AVX_INSTR 6-9+
1182    %ifnum sizeof%7
1183        %assign __sizeofreg sizeof%7
1184    %elifnum sizeof%6
1185        %assign __sizeofreg sizeof%6
1186    %else
1187        %assign __sizeofreg mmsize
1188    %endif
1189    %assign __emulate_avx 0
1190    %if avx_enabled && __sizeofreg >= 16
1191        %xdefine __instr v%1
1192    %else
1193        %xdefine __instr %1
1194        %if %0 >= 8+%4
1195            %assign __emulate_avx 1
1196        %endif
1197    %endif
1198    %ifnidn %2, fnord
1199        %ifdef cpuname
1200            %if notcpuflag(%2)
1201                %error use of ``%1'' %2 instruction in cpuname function: current_function
1202            %elif cpuflags_%2 < cpuflags_sse && notcpuflag(sse2) && __sizeofreg > 8
1203                %error use of ``%1'' sse2 instruction in cpuname function: current_function
1204            %endif
1205        %endif
1206    %endif
1207
1208    %if __emulate_avx
1209        %xdefine __src1 %7
1210        %xdefine __src2 %8
1211        %if %5 && %4 == 0
1212            %ifnidn %6, %7
1213                %ifidn %6, %8
1214                    %xdefine __src1 %8
1215                    %xdefine __src2 %7
1216                %elifnnum sizeof%8
1217                    ; 3-operand AVX instructions with a memory arg can only have it in src2,
1218                    ; whereas SSE emulation prefers to have it in src1 (i.e. the mov).
1219                    ; So, if the instruction is commutative with a memory arg, swap them.
1220                    %xdefine __src1 %8
1221                    %xdefine __src2 %7
1222                %endif
1223            %endif
1224        %endif
1225        %ifnidn %6, __src1
1226            %if %0 >= 9
1227                CHECK_AVX_INSTR_EMU {%1 %6, %7, %8, %9}, %6, __src2, %9
1228            %else
1229                CHECK_AVX_INSTR_EMU {%1 %6, %7, %8}, %6, __src2
1230            %endif
1231            %if __sizeofreg == 8
1232                MOVQ %6, __src1
1233            %elif %3
1234                MOVAPS %6, __src1
1235            %else
1236                MOVDQA %6, __src1
1237            %endif
1238        %endif
1239        %if %0 >= 9
1240            %1 %6, __src2, %9
1241        %else
1242            %1 %6, __src2
1243        %endif
1244    %elif %0 >= 9
1245        __instr %6, %7, %8, %9
1246    %elif %0 == 8
1247        __instr %6, %7, %8
1248    %elif %0 == 7
1249        __instr %6, %7
1250    %else
1251        __instr %6
1252    %endif
1253%endmacro
1254
1255;%1 == instruction
1256;%2 == minimal instruction set
1257;%3 == 1 if float, 0 if int
1258;%4 == 1 if 4-operand emulation, 0 if 3-operand emulation, 255 otherwise (no emulation)
1259;%5 == 1 if commutative (i.e. doesn't matter which src arg is which), 0 if not
1260%macro AVX_INSTR 1-5 fnord, 0, 255, 0
1261    %macro %1 1-10 fnord, fnord, fnord, fnord, %1, %2, %3, %4, %5
1262        %ifidn %2, fnord
1263            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1
1264        %elifidn %3, fnord
1265            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2
1266        %elifidn %4, fnord
1267            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3
1268        %elifidn %5, fnord
1269            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3, %4
1270        %else
1271            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3, %4, %5
1272        %endif
1273    %endmacro
1274%endmacro
1275
1276; Instructions with both VEX/EVEX and legacy encodings
1277; Non-destructive instructions are written without parameters
1278AVX_INSTR addpd, sse2, 1, 0, 1
1279AVX_INSTR addps, sse, 1, 0, 1
1280AVX_INSTR addsd, sse2, 1, 0, 0
1281AVX_INSTR addss, sse, 1, 0, 0
1282AVX_INSTR addsubpd, sse3, 1, 0, 0
1283AVX_INSTR addsubps, sse3, 1, 0, 0
1284AVX_INSTR aesdec, aesni, 0, 0, 0
1285AVX_INSTR aesdeclast, aesni, 0, 0, 0
1286AVX_INSTR aesenc, aesni, 0, 0, 0
1287AVX_INSTR aesenclast, aesni, 0, 0, 0
1288AVX_INSTR aesimc, aesni
1289AVX_INSTR aeskeygenassist, aesni
1290AVX_INSTR andnpd, sse2, 1, 0, 0
1291AVX_INSTR andnps, sse, 1, 0, 0
1292AVX_INSTR andpd, sse2, 1, 0, 1
1293AVX_INSTR andps, sse, 1, 0, 1
1294AVX_INSTR blendpd, sse4, 1, 1, 0
1295AVX_INSTR blendps, sse4, 1, 1, 0
1296AVX_INSTR blendvpd, sse4 ; can't be emulated
1297AVX_INSTR blendvps, sse4 ; can't be emulated
1298AVX_INSTR cmpeqpd, sse2, 1, 0, 1
1299AVX_INSTR cmpeqps, sse, 1, 0, 1
1300AVX_INSTR cmpeqsd, sse2, 1, 0, 0
1301AVX_INSTR cmpeqss, sse, 1, 0, 0
1302AVX_INSTR cmplepd, sse2, 1, 0, 0
1303AVX_INSTR cmpleps, sse, 1, 0, 0
1304AVX_INSTR cmplesd, sse2, 1, 0, 0
1305AVX_INSTR cmpless, sse, 1, 0, 0
1306AVX_INSTR cmpltpd, sse2, 1, 0, 0
1307AVX_INSTR cmpltps, sse, 1, 0, 0
1308AVX_INSTR cmpltsd, sse2, 1, 0, 0
1309AVX_INSTR cmpltss, sse, 1, 0, 0
1310AVX_INSTR cmpneqpd, sse2, 1, 0, 1
1311AVX_INSTR cmpneqps, sse, 1, 0, 1
1312AVX_INSTR cmpneqsd, sse2, 1, 0, 0
1313AVX_INSTR cmpneqss, sse, 1, 0, 0
1314AVX_INSTR cmpnlepd, sse2, 1, 0, 0
1315AVX_INSTR cmpnleps, sse, 1, 0, 0
1316AVX_INSTR cmpnlesd, sse2, 1, 0, 0
1317AVX_INSTR cmpnless, sse, 1, 0, 0
1318AVX_INSTR cmpnltpd, sse2, 1, 0, 0
1319AVX_INSTR cmpnltps, sse, 1, 0, 0
1320AVX_INSTR cmpnltsd, sse2, 1, 0, 0
1321AVX_INSTR cmpnltss, sse, 1, 0, 0
1322AVX_INSTR cmpordpd, sse2 1, 0, 1
1323AVX_INSTR cmpordps, sse 1, 0, 1
1324AVX_INSTR cmpordsd, sse2 1, 0, 0
1325AVX_INSTR cmpordss, sse 1, 0, 0
1326AVX_INSTR cmppd, sse2, 1, 1, 0
1327AVX_INSTR cmpps, sse, 1, 1, 0
1328AVX_INSTR cmpsd, sse2, 1, 1, 0
1329AVX_INSTR cmpss, sse, 1, 1, 0
1330AVX_INSTR cmpunordpd, sse2, 1, 0, 1
1331AVX_INSTR cmpunordps, sse, 1, 0, 1
1332AVX_INSTR cmpunordsd, sse2, 1, 0, 0
1333AVX_INSTR cmpunordss, sse, 1, 0, 0
1334AVX_INSTR comisd, sse2
1335AVX_INSTR comiss, sse
1336AVX_INSTR cvtdq2pd, sse2
1337AVX_INSTR cvtdq2ps, sse2
1338AVX_INSTR cvtpd2dq, sse2
1339AVX_INSTR cvtpd2ps, sse2
1340AVX_INSTR cvtps2dq, sse2
1341AVX_INSTR cvtps2pd, sse2
1342AVX_INSTR cvtsd2si, sse2
1343AVX_INSTR cvtsd2ss, sse2, 1, 0, 0
1344AVX_INSTR cvtsi2sd, sse2, 1, 0, 0
1345AVX_INSTR cvtsi2ss, sse, 1, 0, 0
1346AVX_INSTR cvtss2sd, sse2, 1, 0, 0
1347AVX_INSTR cvtss2si, sse
1348AVX_INSTR cvttpd2dq, sse2
1349AVX_INSTR cvttps2dq, sse2
1350AVX_INSTR cvttsd2si, sse2
1351AVX_INSTR cvttss2si, sse
1352AVX_INSTR divpd, sse2, 1, 0, 0
1353AVX_INSTR divps, sse, 1, 0, 0
1354AVX_INSTR divsd, sse2, 1, 0, 0
1355AVX_INSTR divss, sse, 1, 0, 0
1356AVX_INSTR dppd, sse4, 1, 1, 0
1357AVX_INSTR dpps, sse4, 1, 1, 0
1358AVX_INSTR extractps, sse4
1359AVX_INSTR haddpd, sse3, 1, 0, 0
1360AVX_INSTR haddps, sse3, 1, 0, 0
1361AVX_INSTR hsubpd, sse3, 1, 0, 0
1362AVX_INSTR hsubps, sse3, 1, 0, 0
1363AVX_INSTR insertps, sse4, 1, 1, 0
1364AVX_INSTR lddqu, sse3
1365AVX_INSTR ldmxcsr, sse
1366AVX_INSTR maskmovdqu, sse2
1367AVX_INSTR maxpd, sse2, 1, 0, 1
1368AVX_INSTR maxps, sse, 1, 0, 1
1369AVX_INSTR maxsd, sse2, 1, 0, 0
1370AVX_INSTR maxss, sse, 1, 0, 0
1371AVX_INSTR minpd, sse2, 1, 0, 1
1372AVX_INSTR minps, sse, 1, 0, 1
1373AVX_INSTR minsd, sse2, 1, 0, 0
1374AVX_INSTR minss, sse, 1, 0, 0
1375AVX_INSTR movapd, sse2
1376AVX_INSTR movaps, sse
1377AVX_INSTR movd, mmx
1378AVX_INSTR movddup, sse3
1379AVX_INSTR movdqa, sse2
1380AVX_INSTR movdqu, sse2
1381AVX_INSTR movhlps, sse, 1, 0, 0
1382AVX_INSTR movhpd, sse2, 1, 0, 0
1383AVX_INSTR movhps, sse, 1, 0, 0
1384AVX_INSTR movlhps, sse, 1, 0, 0
1385AVX_INSTR movlpd, sse2, 1, 0, 0
1386AVX_INSTR movlps, sse, 1, 0, 0
1387AVX_INSTR movmskpd, sse2
1388AVX_INSTR movmskps, sse
1389AVX_INSTR movntdq, sse2
1390AVX_INSTR movntdqa, sse4
1391AVX_INSTR movntpd, sse2
1392AVX_INSTR movntps, sse
1393AVX_INSTR movq, mmx
1394AVX_INSTR movsd, sse2, 1, 0, 0
1395AVX_INSTR movshdup, sse3
1396AVX_INSTR movsldup, sse3
1397AVX_INSTR movss, sse, 1, 0, 0
1398AVX_INSTR movupd, sse2
1399AVX_INSTR movups, sse
1400AVX_INSTR mpsadbw, sse4, 0, 1, 0
1401AVX_INSTR mulpd, sse2, 1, 0, 1
1402AVX_INSTR mulps, sse, 1, 0, 1
1403AVX_INSTR mulsd, sse2, 1, 0, 0
1404AVX_INSTR mulss, sse, 1, 0, 0
1405AVX_INSTR orpd, sse2, 1, 0, 1
1406AVX_INSTR orps, sse, 1, 0, 1
1407AVX_INSTR pabsb, ssse3
1408AVX_INSTR pabsd, ssse3
1409AVX_INSTR pabsw, ssse3
1410AVX_INSTR packsswb, mmx, 0, 0, 0
1411AVX_INSTR packssdw, mmx, 0, 0, 0
1412AVX_INSTR packuswb, mmx, 0, 0, 0
1413AVX_INSTR packusdw, sse4, 0, 0, 0
1414AVX_INSTR paddb, mmx, 0, 0, 1
1415AVX_INSTR paddw, mmx, 0, 0, 1
1416AVX_INSTR paddd, mmx, 0, 0, 1
1417AVX_INSTR paddq, sse2, 0, 0, 1
1418AVX_INSTR paddsb, mmx, 0, 0, 1
1419AVX_INSTR paddsw, mmx, 0, 0, 1
1420AVX_INSTR paddusb, mmx, 0, 0, 1
1421AVX_INSTR paddusw, mmx, 0, 0, 1
1422AVX_INSTR palignr, ssse3, 0, 1, 0
1423AVX_INSTR pand, mmx, 0, 0, 1
1424AVX_INSTR pandn, mmx, 0, 0, 0
1425AVX_INSTR pavgb, mmx2, 0, 0, 1
1426AVX_INSTR pavgw, mmx2, 0, 0, 1
1427AVX_INSTR pblendvb, sse4 ; can't be emulated
1428AVX_INSTR pblendw, sse4, 0, 1, 0
1429AVX_INSTR pclmulqdq, fnord, 0, 1, 0
1430AVX_INSTR pclmulhqhqdq, fnord, 0, 0, 0
1431AVX_INSTR pclmulhqlqdq, fnord, 0, 0, 0
1432AVX_INSTR pclmullqhqdq, fnord, 0, 0, 0
1433AVX_INSTR pclmullqlqdq, fnord, 0, 0, 0
1434AVX_INSTR pcmpestri, sse42
1435AVX_INSTR pcmpestrm, sse42
1436AVX_INSTR pcmpistri, sse42
1437AVX_INSTR pcmpistrm, sse42
1438AVX_INSTR pcmpeqb, mmx, 0, 0, 1
1439AVX_INSTR pcmpeqw, mmx, 0, 0, 1
1440AVX_INSTR pcmpeqd, mmx, 0, 0, 1
1441AVX_INSTR pcmpeqq, sse4, 0, 0, 1
1442AVX_INSTR pcmpgtb, mmx, 0, 0, 0
1443AVX_INSTR pcmpgtw, mmx, 0, 0, 0
1444AVX_INSTR pcmpgtd, mmx, 0, 0, 0
1445AVX_INSTR pcmpgtq, sse42, 0, 0, 0
1446AVX_INSTR pextrb, sse4
1447AVX_INSTR pextrd, sse4
1448AVX_INSTR pextrq, sse4
1449AVX_INSTR pextrw, mmx2
1450AVX_INSTR phaddw, ssse3, 0, 0, 0
1451AVX_INSTR phaddd, ssse3, 0, 0, 0
1452AVX_INSTR phaddsw, ssse3, 0, 0, 0
1453AVX_INSTR phminposuw, sse4
1454AVX_INSTR phsubw, ssse3, 0, 0, 0
1455AVX_INSTR phsubd, ssse3, 0, 0, 0
1456AVX_INSTR phsubsw, ssse3, 0, 0, 0
1457AVX_INSTR pinsrb, sse4, 0, 1, 0
1458AVX_INSTR pinsrd, sse4, 0, 1, 0
1459AVX_INSTR pinsrq, sse4, 0, 1, 0
1460AVX_INSTR pinsrw, mmx2, 0, 1, 0
1461AVX_INSTR pmaddwd, mmx, 0, 0, 1
1462AVX_INSTR pmaddubsw, ssse3, 0, 0, 0
1463AVX_INSTR pmaxsb, sse4, 0, 0, 1
1464AVX_INSTR pmaxsw, mmx2, 0, 0, 1
1465AVX_INSTR pmaxsd, sse4, 0, 0, 1
1466AVX_INSTR pmaxub, mmx2, 0, 0, 1
1467AVX_INSTR pmaxuw, sse4, 0, 0, 1
1468AVX_INSTR pmaxud, sse4, 0, 0, 1
1469AVX_INSTR pminsb, sse4, 0, 0, 1
1470AVX_INSTR pminsw, mmx2, 0, 0, 1
1471AVX_INSTR pminsd, sse4, 0, 0, 1
1472AVX_INSTR pminub, mmx2, 0, 0, 1
1473AVX_INSTR pminuw, sse4, 0, 0, 1
1474AVX_INSTR pminud, sse4, 0, 0, 1
1475AVX_INSTR pmovmskb, mmx2
1476AVX_INSTR pmovsxbw, sse4
1477AVX_INSTR pmovsxbd, sse4
1478AVX_INSTR pmovsxbq, sse4
1479AVX_INSTR pmovsxwd, sse4
1480AVX_INSTR pmovsxwq, sse4
1481AVX_INSTR pmovsxdq, sse4
1482AVX_INSTR pmovzxbw, sse4
1483AVX_INSTR pmovzxbd, sse4
1484AVX_INSTR pmovzxbq, sse4
1485AVX_INSTR pmovzxwd, sse4
1486AVX_INSTR pmovzxwq, sse4
1487AVX_INSTR pmovzxdq, sse4
1488AVX_INSTR pmuldq, sse4, 0, 0, 1
1489AVX_INSTR pmulhrsw, ssse3, 0, 0, 1
1490AVX_INSTR pmulhuw, mmx2, 0, 0, 1
1491AVX_INSTR pmulhw, mmx, 0, 0, 1
1492AVX_INSTR pmullw, mmx, 0, 0, 1
1493AVX_INSTR pmulld, sse4, 0, 0, 1
1494AVX_INSTR pmuludq, sse2, 0, 0, 1
1495AVX_INSTR por, mmx, 0, 0, 1
1496AVX_INSTR psadbw, mmx2, 0, 0, 1
1497AVX_INSTR pshufb, ssse3, 0, 0, 0
1498AVX_INSTR pshufd, sse2
1499AVX_INSTR pshufhw, sse2
1500AVX_INSTR pshuflw, sse2
1501AVX_INSTR psignb, ssse3, 0, 0, 0
1502AVX_INSTR psignw, ssse3, 0, 0, 0
1503AVX_INSTR psignd, ssse3, 0, 0, 0
1504AVX_INSTR psllw, mmx, 0, 0, 0
1505AVX_INSTR pslld, mmx, 0, 0, 0
1506AVX_INSTR psllq, mmx, 0, 0, 0
1507AVX_INSTR pslldq, sse2, 0, 0, 0
1508AVX_INSTR psraw, mmx, 0, 0, 0
1509AVX_INSTR psrad, mmx, 0, 0, 0
1510AVX_INSTR psrlw, mmx, 0, 0, 0
1511AVX_INSTR psrld, mmx, 0, 0, 0
1512AVX_INSTR psrlq, mmx, 0, 0, 0
1513AVX_INSTR psrldq, sse2, 0, 0, 0
1514AVX_INSTR psubb, mmx, 0, 0, 0
1515AVX_INSTR psubw, mmx, 0, 0, 0
1516AVX_INSTR psubd, mmx, 0, 0, 0
1517AVX_INSTR psubq, sse2, 0, 0, 0
1518AVX_INSTR psubsb, mmx, 0, 0, 0
1519AVX_INSTR psubsw, mmx, 0, 0, 0
1520AVX_INSTR psubusb, mmx, 0, 0, 0
1521AVX_INSTR psubusw, mmx, 0, 0, 0
1522AVX_INSTR ptest, sse4
1523AVX_INSTR punpckhbw, mmx, 0, 0, 0
1524AVX_INSTR punpckhwd, mmx, 0, 0, 0
1525AVX_INSTR punpckhdq, mmx, 0, 0, 0
1526AVX_INSTR punpckhqdq, sse2, 0, 0, 0
1527AVX_INSTR punpcklbw, mmx, 0, 0, 0
1528AVX_INSTR punpcklwd, mmx, 0, 0, 0
1529AVX_INSTR punpckldq, mmx, 0, 0, 0
1530AVX_INSTR punpcklqdq, sse2, 0, 0, 0
1531AVX_INSTR pxor, mmx, 0, 0, 1
1532AVX_INSTR rcpps, sse
1533AVX_INSTR rcpss, sse, 1, 0, 0
1534AVX_INSTR roundpd, sse4
1535AVX_INSTR roundps, sse4
1536AVX_INSTR roundsd, sse4, 1, 1, 0
1537AVX_INSTR roundss, sse4, 1, 1, 0
1538AVX_INSTR rsqrtps, sse
1539AVX_INSTR rsqrtss, sse, 1, 0, 0
1540AVX_INSTR shufpd, sse2, 1, 1, 0
1541AVX_INSTR shufps, sse, 1, 1, 0
1542AVX_INSTR sqrtpd, sse2
1543AVX_INSTR sqrtps, sse
1544AVX_INSTR sqrtsd, sse2, 1, 0, 0
1545AVX_INSTR sqrtss, sse, 1, 0, 0
1546AVX_INSTR stmxcsr, sse
1547AVX_INSTR subpd, sse2, 1, 0, 0
1548AVX_INSTR subps, sse, 1, 0, 0
1549AVX_INSTR subsd, sse2, 1, 0, 0
1550AVX_INSTR subss, sse, 1, 0, 0
1551AVX_INSTR ucomisd, sse2
1552AVX_INSTR ucomiss, sse
1553AVX_INSTR unpckhpd, sse2, 1, 0, 0
1554AVX_INSTR unpckhps, sse, 1, 0, 0
1555AVX_INSTR unpcklpd, sse2, 1, 0, 0
1556AVX_INSTR unpcklps, sse, 1, 0, 0
1557AVX_INSTR xorpd, sse2, 1, 0, 1
1558AVX_INSTR xorps, sse, 1, 0, 1
1559
1560; 3DNow instructions, for sharing code between AVX, SSE and 3DN
1561AVX_INSTR pfadd, 3dnow, 1, 0, 1
1562AVX_INSTR pfsub, 3dnow, 1, 0, 0
1563AVX_INSTR pfmul, 3dnow, 1, 0, 1
1564
1565; base-4 constants for shuffles
1566%assign i 0
1567%rep 256
1568    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
1569    %if j < 10
1570        CAT_XDEFINE q000, j, i
1571    %elif j < 100
1572        CAT_XDEFINE q00, j, i
1573    %elif j < 1000
1574        CAT_XDEFINE q0, j, i
1575    %else
1576        CAT_XDEFINE q, j, i
1577    %endif
1578    %assign i i+1
1579%endrep
1580%undef i
1581%undef j
1582
1583%macro FMA_INSTR 3
1584    %macro %1 4-7 %1, %2, %3
1585        %if cpuflag(xop)
1586            v%5 %1, %2, %3, %4
1587        %elifnidn %1, %4
1588            %6 %1, %2, %3
1589            %7 %1, %4
1590        %else
1591            %error non-xop emulation of ``%5 %1, %2, %3, %4'' is not supported
1592        %endif
1593    %endmacro
1594%endmacro
1595
1596FMA_INSTR  pmacsww,  pmullw, paddw
1597FMA_INSTR  pmacsdd,  pmulld, paddd ; sse4 emulation
1598FMA_INSTR pmacsdql,  pmuldq, paddq ; sse4 emulation
1599FMA_INSTR pmadcswd, pmaddwd, paddd
1600
1601; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
1602; This lets us use tzcnt without bumping the yasm version requirement yet.
1603%define tzcnt rep bsf
1604
1605; Macros for consolidating FMA3 and FMA4 using 4-operand (dst, src1, src2, src3) syntax.
1606; FMA3 is only possible if dst is the same as one of the src registers.
1607; Either src2 or src3 can be a memory operand.
1608%macro FMA4_INSTR 2-*
1609    %push fma4_instr
1610    %xdefine %$prefix %1
1611    %rep %0 - 1
1612        %macro %$prefix%2 4-6 %$prefix, %2
1613            %if notcpuflag(fma3) && notcpuflag(fma4)
1614                %error use of ``%5%6'' fma instruction in cpuname function: current_function
1615            %elif cpuflag(fma4)
1616                v%5%6 %1, %2, %3, %4
1617            %elifidn %1, %2
1618                ; If %3 or %4 is a memory operand it needs to be encoded as the last operand.
1619                %ifnum sizeof%3
1620                    v%{5}213%6 %2, %3, %4
1621                %else
1622                    v%{5}132%6 %2, %4, %3
1623                %endif
1624            %elifidn %1, %3
1625                v%{5}213%6 %3, %2, %4
1626            %elifidn %1, %4
1627                v%{5}231%6 %4, %2, %3
1628            %else
1629                %error fma3 emulation of ``%5%6 %1, %2, %3, %4'' is not supported
1630            %endif
1631        %endmacro
1632        %rotate 1
1633    %endrep
1634    %pop
1635%endmacro
1636
1637FMA4_INSTR fmadd,    pd, ps, sd, ss
1638FMA4_INSTR fmaddsub, pd, ps
1639FMA4_INSTR fmsub,    pd, ps, sd, ss
1640FMA4_INSTR fmsubadd, pd, ps
1641FMA4_INSTR fnmadd,   pd, ps, sd, ss
1642FMA4_INSTR fnmsub,   pd, ps, sd, ss
1643
1644; Macros for converting VEX instructions to equivalent EVEX ones.
1645%macro EVEX_INSTR 2-3 0 ; vex, evex, prefer_evex
1646    %macro %1 2-7 fnord, fnord, %1, %2, %3
1647        %ifidn %3, fnord
1648            %define %%args %1, %2
1649        %elifidn %4, fnord
1650            %define %%args %1, %2, %3
1651        %else
1652            %define %%args %1, %2, %3, %4
1653        %endif
1654        %assign %%evex_required cpuflag(avx512) & %7
1655        %ifnum regnumof%1
1656            %if regnumof%1 >= 16 || sizeof%1 > 32
1657                %assign %%evex_required 1
1658            %endif
1659        %endif
1660        %ifnum regnumof%2
1661            %if regnumof%2 >= 16 || sizeof%2 > 32
1662                %assign %%evex_required 1
1663            %endif
1664        %endif
1665        %if %%evex_required
1666            %6 %%args
1667        %else
1668            %5 %%args ; Prefer VEX over EVEX due to shorter instruction length
1669        %endif
1670    %endmacro
1671%endmacro
1672
1673EVEX_INSTR vbroadcastf128, vbroadcastf32x4
1674EVEX_INSTR vbroadcasti128, vbroadcasti32x4
1675EVEX_INSTR vextractf128,   vextractf32x4
1676EVEX_INSTR vextracti128,   vextracti32x4
1677EVEX_INSTR vinsertf128,    vinsertf32x4
1678EVEX_INSTR vinserti128,    vinserti32x4
1679EVEX_INSTR vmovdqa,        vmovdqa32
1680EVEX_INSTR vmovdqu,        vmovdqu32
1681EVEX_INSTR vpand,          vpandd
1682EVEX_INSTR vpandn,         vpandnd
1683EVEX_INSTR vpor,           vpord
1684EVEX_INSTR vpxor,          vpxord
1685EVEX_INSTR vrcpps,         vrcp14ps,   1 ; EVEX versions have higher precision
1686EVEX_INSTR vrcpss,         vrcp14ss,   1
1687EVEX_INSTR vrsqrtps,       vrsqrt14ps, 1
1688EVEX_INSTR vrsqrtss,       vrsqrt14ss, 1
1689
1690; workaround: vpbroadcastq is broken in x86_32 due to a yasm bug (fixed in 1.3.0)
1691%ifdef __YASM_VER__
1692    %if __YASM_VERSION_ID__ < 0x01030000 && ARCH_X86_64 == 0
1693        %macro vpbroadcastq 2
1694            %if sizeof%1 == 16
1695                movddup %1, %2
1696            %else
1697                vbroadcastsd %1, %2
1698            %endif
1699        %endmacro
1700    %endif
1701%endif
1702