yading@11: ;*****************************************************************************
yading@11: ;* x86inc.asm: x264asm abstraction layer
yading@11: ;*****************************************************************************
yading@11: ;* Copyright (C) 2005-2012 x264 project
yading@11: ;*
yading@11: ;* Authors: Loren Merritt <lorenm@u.washington.edu>
yading@11: ;*          Anton Mitrofanov <BugMaster@narod.ru>
yading@11: ;*          Jason Garrett-Glaser <darkshikari@gmail.com>
yading@11: ;*          Henrik Gramner <hengar-6@student.ltu.se>
yading@11: ;*
yading@11: ;* Permission to use, copy, modify, and/or distribute this software for any
yading@11: ;* purpose with or without fee is hereby granted, provided that the above
yading@11: ;* copyright notice and this permission notice appear in all copies.
yading@11: ;*
yading@11: ;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
yading@11: ;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
yading@11: ;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
yading@11: ;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
yading@11: ;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
yading@11: ;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
yading@11: ;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
yading@11: ;*****************************************************************************
yading@11: 
yading@11: ; This is a header file for the x264ASM assembly language, which uses
yading@11: ; NASM/YASM syntax combined with a large number of macros to provide easy
yading@11: ; abstraction between different calling conventions (x86_32, win64, linux64).
yading@11: ; It also has various other useful features to simplify writing the kind of
yading@11: ; DSP functions that are most often used in x264.
yading@11: 
yading@11: ; Unlike the rest of x264, this file is available under an ISC license, as it
yading@11: ; has significant usefulness outside of x264 and we want it to be available
yading@11: ; to the largest audience possible.  Of course, if you modify it for your own
yading@11: ; purposes to add a new feature, we strongly encourage contributing a patch
yading@11: ; as this feature might be useful for others as well.  Send patches or ideas
yading@11: ; to x264-devel@videolan.org .
yading@11: 
yading@11: %ifndef private_prefix
yading@11:     %define private_prefix x264
yading@11: %endif
yading@11: 
yading@11: %ifndef public_prefix
yading@11:     %define public_prefix private_prefix
yading@11: %endif
yading@11: 
yading@11: %define WIN64  0
yading@11: %define UNIX64 0
yading@11: %if ARCH_X86_64
yading@11:     %ifidn __OUTPUT_FORMAT__,win32
yading@11:         %define WIN64  1
yading@11:     %elifidn __OUTPUT_FORMAT__,win64
yading@11:         %define WIN64  1
yading@11:     %else
yading@11:         %define UNIX64 1
yading@11:     %endif
yading@11: %endif
yading@11: 
yading@11: %ifdef PREFIX
yading@11:     %define mangle(x) _ %+ x
yading@11: %else
yading@11:     %define mangle(x) x
yading@11: %endif
yading@11: 
yading@11: ; Name of the .rodata section.
yading@11: %macro SECTION_RODATA 0-1 16
yading@11:     ; Kludge: Something on OS X fails to align .rodata even given an align
yading@11:     ; attribute, so use a different read-only section. This has been fixed in
yading@11:     ; yasm 0.8.0 and nasm 2.6.
yading@11:     %ifdef __YASM_VERSION_ID__
yading@11:         %if __YASM_VERSION_ID__ < 00080000h
yading@11:             %define NEED_MACHO_RODATA_KLUDGE
yading@11:         %endif
yading@11:     %elifdef __NASM_VERSION_ID__
yading@11:         %if __NASM_VERSION_ID__ < 02060000h
yading@11:             %define NEED_MACHO_RODATA_KLUDGE
yading@11:         %endif
yading@11:     %endif
yading@11: 
yading@11:     %ifidn __OUTPUT_FORMAT__,aout
yading@11:         section .text
yading@11:     %else
yading@11:         %ifndef NEED_MACHO_RODATA_KLUDGE
yading@11:             SECTION .rodata align=%1
yading@11:         %else
yading@11:             %ifidn __OUTPUT_FORMAT__,macho64
yading@11:                 SECTION .text align=%1
yading@11:             %elifidn __OUTPUT_FORMAT__,macho
yading@11:                 SECTION .text align=%1
yading@11:                 fakegot:
yading@11:             %else
yading@11:                 SECTION .rodata align=%1
yading@11:             %endif
yading@11:         %endif
yading@11:     %endif
yading@11: 
yading@11:     %undef NEED_MACHO_RODATA_KLUDGE
yading@11: %endmacro
yading@11: 
yading@11: ; aout does not support align=
yading@11: %macro SECTION_TEXT 0-1 16
yading@11:     %ifidn __OUTPUT_FORMAT__,aout
yading@11:         SECTION .text
yading@11:     %else
yading@11:         SECTION .text align=%1
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %if WIN64
yading@11:     %define PIC
yading@11: %elif ARCH_X86_64 == 0
yading@11: ; x86_32 doesn't require PIC.
yading@11: ; Some distros prefer shared objects to be PIC, but nothing breaks if
yading@11: ; the code contains a few textrels, so we'll skip that complexity.
yading@11:     %undef PIC
yading@11: %endif
yading@11: %ifdef PIC
yading@11:     default rel
yading@11: %endif
yading@11: 
yading@11: %macro CPUNOP 1
yading@11:     %if HAVE_CPUNOP
yading@11:         CPU %1
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ; Always use long nops (reduces 0x90 spam in disassembly on x86_32)
yading@11: CPUNOP amdnop
yading@11: 
yading@11: ; Macros to eliminate most code duplication between x86_32 and x86_64:
yading@11: ; Currently this works only for leaf functions which load all their arguments
yading@11: ; into registers at the start, and make no other use of the stack. Luckily that
yading@11: ; covers most of x264's asm.
yading@11: 
yading@11: ; PROLOGUE:
yading@11: ; %1 = number of arguments. loads them from stack if needed.
yading@11: ; %2 = number of registers used. pushes callee-saved regs if needed.
yading@11: ; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
yading@11: ; %4 = (optional) stack size to be allocated. If not aligned (x86-32 ICC 10.x,
yading@11: ;      MSVC or YMM), the stack will be manually aligned (to 16 or 32 bytes),
yading@11: ;      and an extra register will be allocated to hold the original stack
yading@11: ;      pointer (to not invalidate r0m etc.). To prevent the use of an extra
yading@11: ;      register as stack pointer, request a negative stack size.
yading@11: ; %4+/%5+ = list of names to define to registers
yading@11: ; PROLOGUE can also be invoked by adding the same options to cglobal
yading@11: 
yading@11: ; e.g.
yading@11: ; cglobal foo, 2,3,0, dst, src, tmp
yading@11: ; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
yading@11: 
yading@11: ; TODO Some functions can use some args directly from the stack. If they're the
yading@11: ; last args then you can just not declare them, but if they're in the middle
yading@11: ; we need more flexible macro.
yading@11: 
yading@11: ; RET:
yading@11: ; Pops anything that was pushed by PROLOGUE, and returns.
yading@11: 
yading@11: ; REP_RET:
yading@11: ; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
yading@11: ; which are slow when a normal ret follows a branch.
yading@11: 
yading@11: ; registers:
yading@11: ; rN and rNq are the native-size register holding function argument N
yading@11: ; rNd, rNw, rNb are dword, word, and byte size
yading@11: ; rNh is the high 8 bits of the word size
yading@11: ; rNm is the original location of arg N (a register or on the stack), dword
yading@11: ; rNmp is native size
yading@11: 
yading@11: %macro DECLARE_REG 2-3
yading@11:     %define r%1q %2
yading@11:     %define r%1d %2d
yading@11:     %define r%1w %2w
yading@11:     %define r%1b %2b
yading@11:     %define r%1h %2h
yading@11:     %define %2q %2
yading@11:     %if %0 == 2
yading@11:         %define r%1m  %2d
yading@11:         %define r%1mp %2
yading@11:     %elif ARCH_X86_64 ; memory
yading@11:         %define r%1m [rstk + stack_offset + %3]
yading@11:         %define r%1mp qword r %+ %1 %+ m
yading@11:     %else
yading@11:         %define r%1m [rstk + stack_offset + %3]
yading@11:         %define r%1mp dword r %+ %1 %+ m
yading@11:     %endif
yading@11:     %define r%1  %2
yading@11: %endmacro
yading@11: 
yading@11: %macro DECLARE_REG_SIZE 3
yading@11:     %define r%1q r%1
yading@11:     %define e%1q r%1
yading@11:     %define r%1d e%1
yading@11:     %define e%1d e%1
yading@11:     %define r%1w %1
yading@11:     %define e%1w %1
yading@11:     %define r%1h %3
yading@11:     %define e%1h %3
yading@11:     %define r%1b %2
yading@11:     %define e%1b %2
yading@11: %if ARCH_X86_64 == 0
yading@11:     %define r%1  e%1
yading@11: %endif
yading@11: %endmacro
yading@11: 
yading@11: DECLARE_REG_SIZE ax, al, ah
yading@11: DECLARE_REG_SIZE bx, bl, bh
yading@11: DECLARE_REG_SIZE cx, cl, ch
yading@11: DECLARE_REG_SIZE dx, dl, dh
yading@11: DECLARE_REG_SIZE si, sil, null
yading@11: DECLARE_REG_SIZE di, dil, null
yading@11: DECLARE_REG_SIZE bp, bpl, null
yading@11: 
yading@11: ; t# defines for when per-arch register allocation is more complex than just function arguments
yading@11: 
yading@11: %macro DECLARE_REG_TMP 1-*
yading@11:     %assign %%i 0
yading@11:     %rep %0
yading@11:         CAT_XDEFINE t, %%i, r%1
yading@11:         %assign %%i %%i+1
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro DECLARE_REG_TMP_SIZE 0-*
yading@11:     %rep %0
yading@11:         %define t%1q t%1 %+ q
yading@11:         %define t%1d t%1 %+ d
yading@11:         %define t%1w t%1 %+ w
yading@11:         %define t%1h t%1 %+ h
yading@11:         %define t%1b t%1 %+ b
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
yading@11: 
yading@11: %if ARCH_X86_64
yading@11:     %define gprsize 8
yading@11: %else
yading@11:     %define gprsize 4
yading@11: %endif
yading@11: 
yading@11: %macro PUSH 1
yading@11:     push %1
yading@11:     %ifidn rstk, rsp
yading@11:         %assign stack_offset stack_offset+gprsize
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro POP 1
yading@11:     pop %1
yading@11:     %ifidn rstk, rsp
yading@11:         %assign stack_offset stack_offset-gprsize
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro PUSH_IF_USED 1-*
yading@11:     %rep %0
yading@11:         %if %1 < regs_used
yading@11:             PUSH r%1
yading@11:         %endif
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro POP_IF_USED 1-*
yading@11:     %rep %0
yading@11:         %if %1 < regs_used
yading@11:             pop r%1
yading@11:         %endif
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro LOAD_IF_USED 1-*
yading@11:     %rep %0
yading@11:         %if %1 < num_args
yading@11:             mov r%1, r %+ %1 %+ mp
yading@11:         %endif
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro SUB 2
yading@11:     sub %1, %2
yading@11:     %ifidn %1, rstk
yading@11:         %assign stack_offset stack_offset+(%2)
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro ADD 2
yading@11:     add %1, %2
yading@11:     %ifidn %1, rstk
yading@11:         %assign stack_offset stack_offset-(%2)
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro movifnidn 2
yading@11:     %ifnidn %1, %2
yading@11:         mov %1, %2
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro movsxdifnidn 2
yading@11:     %ifnidn %1, %2
yading@11:         movsxd %1, %2
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro ASSERT 1
yading@11:     %if (%1) == 0
yading@11:         %error assert failed
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro DEFINE_ARGS 0-*
yading@11:     %ifdef n_arg_names
yading@11:         %assign %%i 0
yading@11:         %rep n_arg_names
yading@11:             CAT_UNDEF arg_name %+ %%i, q
yading@11:             CAT_UNDEF arg_name %+ %%i, d
yading@11:             CAT_UNDEF arg_name %+ %%i, w
yading@11:             CAT_UNDEF arg_name %+ %%i, h
yading@11:             CAT_UNDEF arg_name %+ %%i, b
yading@11:             CAT_UNDEF arg_name %+ %%i, m
yading@11:             CAT_UNDEF arg_name %+ %%i, mp
yading@11:             CAT_UNDEF arg_name, %%i
yading@11:             %assign %%i %%i+1
yading@11:         %endrep
yading@11:     %endif
yading@11: 
yading@11:     %xdefine %%stack_offset stack_offset
yading@11:     %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
yading@11:     %assign %%i 0
yading@11:     %rep %0
yading@11:         %xdefine %1q r %+ %%i %+ q
yading@11:         %xdefine %1d r %+ %%i %+ d
yading@11:         %xdefine %1w r %+ %%i %+ w
yading@11:         %xdefine %1h r %+ %%i %+ h
yading@11:         %xdefine %1b r %+ %%i %+ b
yading@11:         %xdefine %1m r %+ %%i %+ m
yading@11:         %xdefine %1mp r %+ %%i %+ mp
yading@11:         CAT_XDEFINE arg_name, %%i, %1
yading@11:         %assign %%i %%i+1
yading@11:         %rotate 1
yading@11:     %endrep
yading@11:     %xdefine stack_offset %%stack_offset
yading@11:     %assign n_arg_names %0
yading@11: %endmacro
yading@11: 
yading@11: %macro ALLOC_STACK 1-2 0 ; stack_size, n_xmm_regs (for win64 only)
yading@11:     %ifnum %1
yading@11:         %if %1 != 0
yading@11:             %assign %%stack_alignment ((mmsize + 15) & ~15)
yading@11:             %assign stack_size %1
yading@11:             %if stack_size < 0
yading@11:                 %assign stack_size -stack_size
yading@11:             %endif
yading@11:             %if mmsize != 8
yading@11:                 %assign xmm_regs_used %2
yading@11:             %endif
yading@11:             %if mmsize <= 16 && HAVE_ALIGNED_STACK
yading@11:                 %assign stack_size_padded stack_size + %%stack_alignment - gprsize - (stack_offset & (%%stack_alignment - 1))
yading@11:                 %if xmm_regs_used > 6
yading@11:                     %assign stack_size_padded stack_size_padded + (xmm_regs_used - 6) * 16
yading@11:                 %endif
yading@11:                 SUB rsp, stack_size_padded
yading@11:             %else
yading@11:                 %assign %%reg_num (regs_used - 1)
yading@11:                 %xdefine rstk r %+ %%reg_num
yading@11:                 ; align stack, and save original stack location directly above
yading@11:                 ; it, i.e. in [rsp+stack_size_padded], so we can restore the
yading@11:                 ; stack in a single instruction (i.e. mov rsp, rstk or mov
yading@11:                 ; rsp, [rsp+stack_size_padded])
yading@11:                 mov  rstk, rsp
yading@11:                 %assign stack_size_padded stack_size
yading@11:                 %if xmm_regs_used > 6
yading@11:                     %assign stack_size_padded stack_size_padded + (xmm_regs_used - 6) * 16
yading@11:                     %if mmsize == 32 && xmm_regs_used & 1
yading@11:                         ; re-align to 32 bytes
yading@11:                         %assign stack_size_padded (stack_size_padded + 16)
yading@11:                     %endif
yading@11:                 %endif
yading@11:                 %if %1 < 0 ; need to store rsp on stack
yading@11:                     sub  rsp, gprsize+stack_size_padded
yading@11:                     and  rsp, ~(%%stack_alignment-1)
yading@11:                     %xdefine rstkm [rsp+stack_size_padded]
yading@11:                     mov rstkm, rstk
yading@11:                 %else ; can keep rsp in rstk during whole function
yading@11:                     sub  rsp, stack_size_padded
yading@11:                     and  rsp, ~(%%stack_alignment-1)
yading@11:                     %xdefine rstkm rstk
yading@11:                 %endif
yading@11:             %endif
yading@11:             %if xmm_regs_used > 6
yading@11:                 WIN64_PUSH_XMM
yading@11:             %endif
yading@11:         %endif
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro SETUP_STACK_POINTER 1
yading@11:     %ifnum %1
yading@11:         %if %1 != 0 && (HAVE_ALIGNED_STACK == 0 || mmsize == 32)
yading@11:             %if %1 > 0
yading@11:                 %assign regs_used (regs_used + 1)
yading@11:             %elif ARCH_X86_64 && regs_used == num_args && num_args <= 4 + UNIX64 * 2
yading@11:                 %warning "Stack pointer will overwrite register argument"
yading@11:             %endif
yading@11:         %endif
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro DEFINE_ARGS_INTERNAL 3+
yading@11:     %ifnum %2
yading@11:         DEFINE_ARGS %3
yading@11:     %elif %1 == 4
yading@11:         DEFINE_ARGS %2
yading@11:     %elif %1 > 4
yading@11:         DEFINE_ARGS %2, %3
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %if WIN64 ; Windows x64 ;=================================================
yading@11: 
yading@11: DECLARE_REG 0,  rcx
yading@11: DECLARE_REG 1,  rdx
yading@11: DECLARE_REG 2,  R8
yading@11: DECLARE_REG 3,  R9
yading@11: DECLARE_REG 4,  R10, 40
yading@11: DECLARE_REG 5,  R11, 48
yading@11: DECLARE_REG 6,  rax, 56
yading@11: DECLARE_REG 7,  rdi, 64
yading@11: DECLARE_REG 8,  rsi, 72
yading@11: DECLARE_REG 9,  rbx, 80
yading@11: DECLARE_REG 10, rbp, 88
yading@11: DECLARE_REG 11, R12, 96
yading@11: DECLARE_REG 12, R13, 104
yading@11: DECLARE_REG 13, R14, 112
yading@11: DECLARE_REG 14, R15, 120
yading@11: 
yading@11: %macro PROLOGUE 2-5+ 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
yading@11:     %assign num_args %1
yading@11:     %assign regs_used %2
yading@11:     ASSERT regs_used >= num_args
yading@11:     SETUP_STACK_POINTER %4
yading@11:     ASSERT regs_used <= 15
yading@11:     PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
yading@11:     ALLOC_STACK %4, %3
yading@11:     %if mmsize != 8 && stack_size == 0
yading@11:         WIN64_SPILL_XMM %3
yading@11:     %endif
yading@11:     LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
yading@11:     DEFINE_ARGS_INTERNAL %0, %4, %5
yading@11: %endmacro
yading@11: 
yading@11: %macro WIN64_PUSH_XMM 0
yading@11:     %assign %%i xmm_regs_used
yading@11:     %rep (xmm_regs_used-6)
yading@11:         %assign %%i %%i-1
yading@11:         movdqa [rsp + (%%i-6)*16 + stack_size + (~stack_offset&8)], xmm %+ %%i
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro WIN64_SPILL_XMM 1
yading@11:     %assign xmm_regs_used %1
yading@11:     ASSERT xmm_regs_used <= 16
yading@11:     %if xmm_regs_used > 6
yading@11:         SUB rsp, (xmm_regs_used-6)*16+16
yading@11:         WIN64_PUSH_XMM
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro WIN64_RESTORE_XMM_INTERNAL 1
yading@11:     %if xmm_regs_used > 6
yading@11:         %assign %%i xmm_regs_used
yading@11:         %rep (xmm_regs_used-6)
yading@11:             %assign %%i %%i-1
yading@11:             movdqa xmm %+ %%i, [%1 + (%%i-6)*16+stack_size+(~stack_offset&8)]
yading@11:         %endrep
yading@11:         %if stack_size_padded == 0
yading@11:             add %1, (xmm_regs_used-6)*16+16
yading@11:         %endif
yading@11:     %endif
yading@11:     %if stack_size_padded > 0
yading@11:         %if stack_size > 0 && (mmsize == 32 || HAVE_ALIGNED_STACK == 0)
yading@11:             mov rsp, rstkm
yading@11:         %else
yading@11:             add %1, stack_size_padded
yading@11:         %endif
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro WIN64_RESTORE_XMM 1
yading@11:     WIN64_RESTORE_XMM_INTERNAL %1
yading@11:     %assign stack_offset (stack_offset-stack_size_padded)
yading@11:     %assign xmm_regs_used 0
yading@11: %endmacro
yading@11: 
yading@11: %define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32 || stack_size > 0
yading@11: 
yading@11: %macro RET 0
yading@11:     WIN64_RESTORE_XMM_INTERNAL rsp
yading@11:     POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
yading@11: %if mmsize == 32
yading@11:     vzeroupper
yading@11: %endif
yading@11:     ret
yading@11: %endmacro
yading@11: 
yading@11: %elif ARCH_X86_64 ; *nix x64 ;=============================================
yading@11: 
yading@11: DECLARE_REG 0,  rdi
yading@11: DECLARE_REG 1,  rsi
yading@11: DECLARE_REG 2,  rdx
yading@11: DECLARE_REG 3,  rcx
yading@11: DECLARE_REG 4,  R8
yading@11: DECLARE_REG 5,  R9
yading@11: DECLARE_REG 6,  rax, 8
yading@11: DECLARE_REG 7,  R10, 16
yading@11: DECLARE_REG 8,  R11, 24
yading@11: DECLARE_REG 9,  rbx, 32
yading@11: DECLARE_REG 10, rbp, 40
yading@11: DECLARE_REG 11, R12, 48
yading@11: DECLARE_REG 12, R13, 56
yading@11: DECLARE_REG 13, R14, 64
yading@11: DECLARE_REG 14, R15, 72
yading@11: 
yading@11: %macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
yading@11:     %assign num_args %1
yading@11:     %assign regs_used %2
yading@11:     ASSERT regs_used >= num_args
yading@11:     SETUP_STACK_POINTER %4
yading@11:     ASSERT regs_used <= 15
yading@11:     PUSH_IF_USED 9, 10, 11, 12, 13, 14
yading@11:     ALLOC_STACK %4
yading@11:     LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
yading@11:     DEFINE_ARGS_INTERNAL %0, %4, %5
yading@11: %endmacro
yading@11: 
yading@11: %define has_epilogue regs_used > 9 || mmsize == 32 || stack_size > 0
yading@11: 
yading@11: %macro RET 0
yading@11: %if stack_size_padded > 0
yading@11: %if mmsize == 32 || HAVE_ALIGNED_STACK == 0
yading@11:     mov rsp, rstkm
yading@11: %else
yading@11:     add rsp, stack_size_padded
yading@11: %endif
yading@11: %endif
yading@11:     POP_IF_USED 14, 13, 12, 11, 10, 9
yading@11: %if mmsize == 32
yading@11:     vzeroupper
yading@11: %endif
yading@11:     ret
yading@11: %endmacro
yading@11: 
yading@11: %else ; X86_32 ;==============================================================
yading@11: 
yading@11: DECLARE_REG 0, eax, 4
yading@11: DECLARE_REG 1, ecx, 8
yading@11: DECLARE_REG 2, edx, 12
yading@11: DECLARE_REG 3, ebx, 16
yading@11: DECLARE_REG 4, esi, 20
yading@11: DECLARE_REG 5, edi, 24
yading@11: DECLARE_REG 6, ebp, 28
yading@11: %define rsp esp
yading@11: 
yading@11: %macro DECLARE_ARG 1-*
yading@11:     %rep %0
yading@11:         %define r%1m [rstk + stack_offset + 4*%1 + 4]
yading@11:         %define r%1mp dword r%1m
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
yading@11: 
yading@11: %macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
yading@11:     %assign num_args %1
yading@11:     %assign regs_used %2
yading@11:     ASSERT regs_used >= num_args
yading@11:     %if num_args > 7
yading@11:         %assign num_args 7
yading@11:     %endif
yading@11:     %if regs_used > 7
yading@11:         %assign regs_used 7
yading@11:     %endif
yading@11:     SETUP_STACK_POINTER %4
yading@11:     ASSERT regs_used <= 7
yading@11:     PUSH_IF_USED 3, 4, 5, 6
yading@11:     ALLOC_STACK %4
yading@11:     LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
yading@11:     DEFINE_ARGS_INTERNAL %0, %4, %5
yading@11: %endmacro
yading@11: 
yading@11: %define has_epilogue regs_used > 3 || mmsize == 32 || stack_size > 0
yading@11: 
yading@11: %macro RET 0
yading@11: %if stack_size_padded > 0
yading@11: %if mmsize == 32 || HAVE_ALIGNED_STACK == 0
yading@11:     mov rsp, rstkm
yading@11: %else
yading@11:     add rsp, stack_size_padded
yading@11: %endif
yading@11: %endif
yading@11:     POP_IF_USED 6, 5, 4, 3
yading@11: %if mmsize == 32
yading@11:     vzeroupper
yading@11: %endif
yading@11:     ret
yading@11: %endmacro
yading@11: 
yading@11: %endif ;======================================================================
yading@11: 
yading@11: %if WIN64 == 0
yading@11: %macro WIN64_SPILL_XMM 1
yading@11: %endmacro
yading@11: %macro WIN64_RESTORE_XMM 1
yading@11: %endmacro
yading@11: %macro WIN64_PUSH_XMM 0
yading@11: %endmacro
yading@11: %endif
yading@11: 
yading@11: %macro REP_RET 0
yading@11:     %if has_epilogue
yading@11:         RET
yading@11:     %else
yading@11:         rep ret
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro TAIL_CALL 2 ; callee, is_nonadjacent
yading@11:     %if has_epilogue
yading@11:         call %1
yading@11:         RET
yading@11:     %elif %2
yading@11:         jmp %1
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ;=============================================================================
yading@11: ; arch-independent part
yading@11: ;=============================================================================
yading@11: 
yading@11: %assign function_align 16
yading@11: 
yading@11: ; Begin a function.
yading@11: ; Applies any symbol mangling needed for C linkage, and sets up a define such that
yading@11: ; subsequent uses of the function name automatically refer to the mangled version.
yading@11: ; Appends cpuflags to the function name if cpuflags has been specified.
yading@11: ; The "" empty default parameter is a workaround for nasm, which fails if SUFFIX
yading@11: ; is empty and we call cglobal_internal with just %1 %+ SUFFIX (without %2).
yading@11: %macro cglobal 1-2+ "" ; name, [PROLOGUE args]
yading@11:     cglobal_internal 1, %1 %+ SUFFIX, %2
yading@11: %endmacro
yading@11: %macro cvisible 1-2+ "" ; name, [PROLOGUE args]
yading@11:     cglobal_internal 0, %1 %+ SUFFIX, %2
yading@11: %endmacro
yading@11: %macro cglobal_internal 2-3+
yading@11:     %if %1
yading@11:         %xdefine %%FUNCTION_PREFIX private_prefix
yading@11:         %xdefine %%VISIBILITY hidden
yading@11:     %else
yading@11:         %xdefine %%FUNCTION_PREFIX public_prefix
yading@11:         %xdefine %%VISIBILITY
yading@11:     %endif
yading@11:     %ifndef cglobaled_%2
yading@11:         %xdefine %2 mangle(%%FUNCTION_PREFIX %+ _ %+ %2)
yading@11:         %xdefine %2.skip_prologue %2 %+ .skip_prologue
yading@11:         CAT_XDEFINE cglobaled_, %2, 1
yading@11:     %endif
yading@11:     %xdefine current_function %2
yading@11:     %ifidn __OUTPUT_FORMAT__,elf
yading@11:         global %2:function %%VISIBILITY
yading@11:     %else
yading@11:         global %2
yading@11:     %endif
yading@11:     align function_align
yading@11:     %2:
yading@11:     RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
yading@11:     %xdefine rstk rsp
yading@11:     %assign stack_offset 0
yading@11:     %assign stack_size 0
yading@11:     %assign stack_size_padded 0
yading@11:     %assign xmm_regs_used 0
yading@11:     %ifnidn %3, ""
yading@11:         PROLOGUE %3
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro cextern 1
yading@11:     %xdefine %1 mangle(private_prefix %+ _ %+ %1)
yading@11:     CAT_XDEFINE cglobaled_, %1, 1
yading@11:     extern %1
yading@11: %endmacro
yading@11: 
yading@11: ; like cextern, but without the prefix
yading@11: %macro cextern_naked 1
yading@11:     %xdefine %1 mangle(%1)
yading@11:     CAT_XDEFINE cglobaled_, %1, 1
yading@11:     extern %1
yading@11: %endmacro
yading@11: 
yading@11: %macro const 2+
yading@11:     %xdefine %1 mangle(private_prefix %+ _ %+ %1)
yading@11:     global %1
yading@11:     %1: %2
yading@11: %endmacro
yading@11: 
yading@11: ; This is needed for ELF, otherwise the GNU linker assumes the stack is
yading@11: ; executable by default.
yading@11: %ifidn __OUTPUT_FORMAT__,elf
yading@11: SECTION .note.GNU-stack noalloc noexec nowrite progbits
yading@11: %endif
yading@11: 
yading@11: ; cpuflags
yading@11: 
yading@11: %assign cpuflags_mmx      (1<<0)
yading@11: %assign cpuflags_mmx2     (1<<1) | cpuflags_mmx
yading@11: %assign cpuflags_3dnow    (1<<2) | cpuflags_mmx
yading@11: %assign cpuflags_3dnowext (1<<3) | cpuflags_3dnow
yading@11: %assign cpuflags_sse      (1<<4) | cpuflags_mmx2
yading@11: %assign cpuflags_sse2     (1<<5) | cpuflags_sse
yading@11: %assign cpuflags_sse2slow (1<<6) | cpuflags_sse2
yading@11: %assign cpuflags_sse3     (1<<7) | cpuflags_sse2
yading@11: %assign cpuflags_ssse3    (1<<8) | cpuflags_sse3
yading@11: %assign cpuflags_sse4     (1<<9) | cpuflags_ssse3
yading@11: %assign cpuflags_sse42    (1<<10)| cpuflags_sse4
yading@11: %assign cpuflags_avx      (1<<11)| cpuflags_sse42
yading@11: %assign cpuflags_xop      (1<<12)| cpuflags_avx
yading@11: %assign cpuflags_fma4     (1<<13)| cpuflags_avx
yading@11: %assign cpuflags_avx2     (1<<14)| cpuflags_avx
yading@11: %assign cpuflags_fma3     (1<<15)| cpuflags_avx
yading@11: 
yading@11: %assign cpuflags_cache32  (1<<16)
yading@11: %assign cpuflags_cache64  (1<<17)
yading@11: %assign cpuflags_slowctz  (1<<18)
yading@11: %assign cpuflags_lzcnt    (1<<19)
yading@11: %assign cpuflags_misalign (1<<20)
yading@11: %assign cpuflags_aligned  (1<<21) ; not a cpu feature, but a function variant
yading@11: %assign cpuflags_atom     (1<<22)
yading@11: %assign cpuflags_bmi1     (1<<23)
yading@11: %assign cpuflags_bmi2     (1<<24)|cpuflags_bmi1
yading@11: %assign cpuflags_tbm      (1<<25)|cpuflags_bmi1
yading@11: 
yading@11: %define    cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
yading@11: %define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
yading@11: 
yading@11: ; Takes up to 2 cpuflags from the above list.
yading@11: ; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
yading@11: ; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
yading@11: %macro INIT_CPUFLAGS 0-2
yading@11:     CPUNOP amdnop
yading@11:     %if %0 >= 1
yading@11:         %xdefine cpuname %1
yading@11:         %assign cpuflags cpuflags_%1
yading@11:         %if %0 >= 2
yading@11:             %xdefine cpuname %1_%2
yading@11:             %assign cpuflags cpuflags | cpuflags_%2
yading@11:         %endif
yading@11:         %xdefine SUFFIX _ %+ cpuname
yading@11:         %if cpuflag(avx)
yading@11:             %assign avx_enabled 1
yading@11:         %endif
yading@11:         %if mmsize == 16 && notcpuflag(sse2)
yading@11:             %define mova movaps
yading@11:             %define movu movups
yading@11:             %define movnta movntps
yading@11:         %endif
yading@11:         %if cpuflag(aligned)
yading@11:             %define movu mova
yading@11:         %elifidn %1, sse3
yading@11:             %define movu lddqu
yading@11:         %endif
yading@11:         %if notcpuflag(sse2)
yading@11:             CPUNOP basicnop
yading@11:         %endif
yading@11:     %else
yading@11:         %xdefine SUFFIX
yading@11:         %undef cpuname
yading@11:         %undef cpuflags
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ; merge mmx and sse*
yading@11: 
yading@11: %macro CAT_XDEFINE 3
yading@11:     %xdefine %1%2 %3
yading@11: %endmacro
yading@11: 
yading@11: %macro CAT_UNDEF 2
yading@11:     %undef %1%2
yading@11: %endmacro
yading@11: 
yading@11: %macro INIT_MMX 0-1+
yading@11:     %assign avx_enabled 0
yading@11:     %define RESET_MM_PERMUTATION INIT_MMX %1
yading@11:     %define mmsize 8
yading@11:     %define num_mmregs 8
yading@11:     %define mova movq
yading@11:     %define movu movq
yading@11:     %define movh movd
yading@11:     %define movnta movntq
yading@11:     %assign %%i 0
yading@11:     %rep 8
yading@11:     CAT_XDEFINE m, %%i, mm %+ %%i
yading@11:     CAT_XDEFINE nmm, %%i, %%i
yading@11:     %assign %%i %%i+1
yading@11:     %endrep
yading@11:     %rep 8
yading@11:     CAT_UNDEF m, %%i
yading@11:     CAT_UNDEF nmm, %%i
yading@11:     %assign %%i %%i+1
yading@11:     %endrep
yading@11:     INIT_CPUFLAGS %1
yading@11: %endmacro
yading@11: 
yading@11: %macro INIT_XMM 0-1+
yading@11:     %assign avx_enabled 0
yading@11:     %define RESET_MM_PERMUTATION INIT_XMM %1
yading@11:     %define mmsize 16
yading@11:     %define num_mmregs 8
yading@11:     %if ARCH_X86_64
yading@11:     %define num_mmregs 16
yading@11:     %endif
yading@11:     %define mova movdqa
yading@11:     %define movu movdqu
yading@11:     %define movh movq
yading@11:     %define movnta movntdq
yading@11:     %assign %%i 0
yading@11:     %rep num_mmregs
yading@11:     CAT_XDEFINE m, %%i, xmm %+ %%i
yading@11:     CAT_XDEFINE nxmm, %%i, %%i
yading@11:     %assign %%i %%i+1
yading@11:     %endrep
yading@11:     INIT_CPUFLAGS %1
yading@11: %endmacro
yading@11: 
yading@11: ; FIXME: INIT_AVX can be replaced by INIT_XMM avx
yading@11: %macro INIT_AVX 0
yading@11:     INIT_XMM
yading@11:     %assign avx_enabled 1
yading@11:     %define PALIGNR PALIGNR_SSSE3
yading@11:     %define RESET_MM_PERMUTATION INIT_AVX
yading@11: %endmacro
yading@11: 
yading@11: %macro INIT_YMM 0-1+
yading@11:     %assign avx_enabled 1
yading@11:     %define RESET_MM_PERMUTATION INIT_YMM %1
yading@11:     %define mmsize 32
yading@11:     %define num_mmregs 8
yading@11:     %if ARCH_X86_64
yading@11:     %define num_mmregs 16
yading@11:     %endif
yading@11:     %define mova vmovaps
yading@11:     %define movu vmovups
yading@11:     %undef movh
yading@11:     %define movnta vmovntps
yading@11:     %assign %%i 0
yading@11:     %rep num_mmregs
yading@11:     CAT_XDEFINE m, %%i, ymm %+ %%i
yading@11:     CAT_XDEFINE nymm, %%i, %%i
yading@11:     %assign %%i %%i+1
yading@11:     %endrep
yading@11:     INIT_CPUFLAGS %1
yading@11: %endmacro
yading@11: 
yading@11: INIT_XMM
yading@11: 
yading@11: ; I often want to use macros that permute their arguments. e.g. there's no
yading@11: ; efficient way to implement butterfly or transpose or dct without swapping some
yading@11: ; arguments.
yading@11: ;
yading@11: ; I would like to not have to manually keep track of the permutations:
yading@11: ; If I insert a permutation in the middle of a function, it should automatically
yading@11: ; change everything that follows. For more complex macros I may also have multiple
yading@11: ; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
yading@11: ;
yading@11: ; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
yading@11: ; permutes its arguments. It's equivalent to exchanging the contents of the
yading@11: ; registers, except that this way you exchange the register names instead, so it
yading@11: ; doesn't cost any cycles.
yading@11: 
yading@11: %macro PERMUTE 2-* ; takes a list of pairs to swap
yading@11: %rep %0/2
yading@11:     %xdefine tmp%2 m%2
yading@11:     %xdefine ntmp%2 nm%2
yading@11:     %rotate 2
yading@11: %endrep
yading@11: %rep %0/2
yading@11:     %xdefine m%1 tmp%2
yading@11:     %xdefine nm%1 ntmp%2
yading@11:     %undef tmp%2
yading@11:     %undef ntmp%2
yading@11:     %rotate 2
yading@11: %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
yading@11: %rep %0-1
yading@11: %ifdef m%1
yading@11:     %xdefine tmp m%1
yading@11:     %xdefine m%1 m%2
yading@11:     %xdefine m%2 tmp
yading@11:     CAT_XDEFINE n, m%1, %1
yading@11:     CAT_XDEFINE n, m%2, %2
yading@11: %else
yading@11:     ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
yading@11:     ; Be careful using this mode in nested macros though, as in some cases there may be
yading@11:     ; other copies of m# that have already been dereferenced and don't get updated correctly.
yading@11:     %xdefine %%n1 n %+ %1
yading@11:     %xdefine %%n2 n %+ %2
yading@11:     %xdefine tmp m %+ %%n1
yading@11:     CAT_XDEFINE m, %%n1, m %+ %%n2
yading@11:     CAT_XDEFINE m, %%n2, tmp
yading@11:     CAT_XDEFINE n, m %+ %%n1, %%n1
yading@11:     CAT_XDEFINE n, m %+ %%n2, %%n2
yading@11: %endif
yading@11:     %undef tmp
yading@11:     %rotate 1
yading@11: %endrep
yading@11: %endmacro
yading@11: 
yading@11: ; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
yading@11: ; calls to that function will automatically load the permutation, so values can
yading@11: ; be returned in mmregs.
yading@11: %macro SAVE_MM_PERMUTATION 0-1
yading@11:     %if %0
yading@11:         %xdefine %%f %1_m
yading@11:     %else
yading@11:         %xdefine %%f current_function %+ _m
yading@11:     %endif
yading@11:     %assign %%i 0
yading@11:     %rep num_mmregs
yading@11:         CAT_XDEFINE %%f, %%i, m %+ %%i
yading@11:     %assign %%i %%i+1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: %macro LOAD_MM_PERMUTATION 1 ; name to load from
yading@11:     %ifdef %1_m0
yading@11:         %assign %%i 0
yading@11:         %rep num_mmregs
yading@11:             CAT_XDEFINE m, %%i, %1_m %+ %%i
yading@11:             CAT_XDEFINE n, m %+ %%i, %%i
yading@11:         %assign %%i %%i+1
yading@11:         %endrep
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
yading@11: %macro call 1
yading@11:     call_internal %1 %+ SUFFIX, %1
yading@11: %endmacro
yading@11: %macro call_internal 2
yading@11:     %xdefine %%i %2
yading@11:     %ifndef cglobaled_%2
yading@11:         %ifdef cglobaled_%1
yading@11:             %xdefine %%i %1
yading@11:         %endif
yading@11:     %endif
yading@11:     call %%i
yading@11:     LOAD_MM_PERMUTATION %%i
yading@11: %endmacro
yading@11: 
yading@11: ; Substitutions that reduce instruction size but are functionally equivalent
yading@11: %macro add 2
yading@11:     %ifnum %2
yading@11:         %if %2==128
yading@11:             sub %1, -128
yading@11:         %else
yading@11:             add %1, %2
yading@11:         %endif
yading@11:     %else
yading@11:         add %1, %2
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: %macro sub 2
yading@11:     %ifnum %2
yading@11:         %if %2==128
yading@11:             add %1, -128
yading@11:         %else
yading@11:             sub %1, %2
yading@11:         %endif
yading@11:     %else
yading@11:         sub %1, %2
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ;=============================================================================
yading@11: ; AVX abstraction layer
yading@11: ;=============================================================================
yading@11: 
yading@11: %assign i 0
yading@11: %rep 16
yading@11:     %if i < 8
yading@11:         CAT_XDEFINE sizeofmm, i, 8
yading@11:     %endif
yading@11:     CAT_XDEFINE sizeofxmm, i, 16
yading@11:     CAT_XDEFINE sizeofymm, i, 32
yading@11: %assign i i+1
yading@11: %endrep
yading@11: %undef i
yading@11: 
yading@11: %macro CHECK_AVX_INSTR_EMU 3-*
yading@11:     %xdefine %%opcode %1
yading@11:     %xdefine %%dst %2
yading@11:     %rep %0-2
yading@11:         %ifidn %%dst, %3
yading@11:             %error non-avx emulation of ``%%opcode'' is not supported
yading@11:         %endif
yading@11:         %rotate 1
yading@11:     %endrep
yading@11: %endmacro
yading@11: 
yading@11: ;%1 == instruction
yading@11: ;%2 == 1 if float, 0 if int
yading@11: ;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
yading@11: ;%4 == number of operands given
yading@11: ;%5+: operands
yading@11: %macro RUN_AVX_INSTR 6-7+
yading@11:     %ifid %6
yading@11:         %define %%sizeofreg sizeof%6
yading@11:     %elifid %5
yading@11:         %define %%sizeofreg sizeof%5
yading@11:     %else
yading@11:         %define %%sizeofreg mmsize
yading@11:     %endif
yading@11:     %if %%sizeofreg==32
yading@11:         %if %4>=3
yading@11:             v%1 %5, %6, %7
yading@11:         %else
yading@11:             v%1 %5, %6
yading@11:         %endif
yading@11:     %else
yading@11:         %if %%sizeofreg==8
yading@11:             %define %%regmov movq
yading@11:         %elif %2
yading@11:             %define %%regmov movaps
yading@11:         %else
yading@11:             %define %%regmov movdqa
yading@11:         %endif
yading@11: 
yading@11:         %if %4>=3+%3
yading@11:             %ifnidn %5, %6
yading@11:                 %if avx_enabled && %%sizeofreg==16
yading@11:                     v%1 %5, %6, %7
yading@11:                 %else
yading@11:                     CHECK_AVX_INSTR_EMU {%1 %5, %6, %7}, %5, %7
yading@11:                     %%regmov %5, %6
yading@11:                     %1 %5, %7
yading@11:                 %endif
yading@11:             %else
yading@11:                 %1 %5, %7
yading@11:             %endif
yading@11:         %elif %4>=3
yading@11:             %1 %5, %6, %7
yading@11:         %else
yading@11:             %1 %5, %6
yading@11:         %endif
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ; 3arg AVX ops with a memory arg can only have it in src2,
yading@11: ; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov).
yading@11: ; So, if the op is symmetric and the wrong one is memory, swap them.
yading@11: %macro RUN_AVX_INSTR1 8
yading@11:     %assign %%swap 0
yading@11:     %if avx_enabled
yading@11:         %ifnid %6
yading@11:             %assign %%swap 1
yading@11:         %endif
yading@11:     %elifnidn %5, %6
yading@11:         %ifnid %7
yading@11:             %assign %%swap 1
yading@11:         %endif
yading@11:     %endif
yading@11:     %if %%swap && %3 == 0 && %8 == 1
yading@11:         RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6
yading@11:     %else
yading@11:         RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7
yading@11:     %endif
yading@11: %endmacro
yading@11: 
yading@11: ;%1 == instruction
yading@11: ;%2 == 1 if float, 0 if int
yading@11: ;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
yading@11: ;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not
yading@11: %macro AVX_INSTR 4
yading@11:     %macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4
yading@11:         %ifidn %3, fnord
yading@11:             RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
yading@11:         %elifidn %4, fnord
yading@11:             RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9
yading@11:         %elifidn %5, fnord
yading@11:             RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
yading@11:         %else
yading@11:             RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5
yading@11:         %endif
yading@11:     %endmacro
yading@11: %endmacro
yading@11: 
yading@11: AVX_INSTR addpd, 1, 0, 1
yading@11: AVX_INSTR addps, 1, 0, 1
yading@11: AVX_INSTR addsd, 1, 0, 1
yading@11: AVX_INSTR addss, 1, 0, 1
yading@11: AVX_INSTR addsubpd, 1, 0, 0
yading@11: AVX_INSTR addsubps, 1, 0, 0
yading@11: AVX_INSTR andpd, 1, 0, 1
yading@11: AVX_INSTR andps, 1, 0, 1
yading@11: AVX_INSTR andnpd, 1, 0, 0
yading@11: AVX_INSTR andnps, 1, 0, 0
yading@11: AVX_INSTR blendpd, 1, 0, 0
yading@11: AVX_INSTR blendps, 1, 0, 0
yading@11: AVX_INSTR blendvpd, 1, 0, 0
yading@11: AVX_INSTR blendvps, 1, 0, 0
yading@11: AVX_INSTR cmppd, 1, 1, 0
yading@11: AVX_INSTR cmpps, 1, 1, 0
yading@11: AVX_INSTR cmpsd, 1, 1, 0
yading@11: AVX_INSTR cmpss, 1, 1, 0
yading@11: AVX_INSTR cvtdq2ps, 1, 0, 0
yading@11: AVX_INSTR cvtpd2dq, 1, 0, 0
yading@11: AVX_INSTR cvtps2dq, 1, 0, 0
yading@11: AVX_INSTR divpd, 1, 0, 0
yading@11: AVX_INSTR divps, 1, 0, 0
yading@11: AVX_INSTR divsd, 1, 0, 0
yading@11: AVX_INSTR divss, 1, 0, 0
yading@11: AVX_INSTR dppd, 1, 1, 0
yading@11: AVX_INSTR dpps, 1, 1, 0
yading@11: AVX_INSTR haddpd, 1, 0, 0
yading@11: AVX_INSTR haddps, 1, 0, 0
yading@11: AVX_INSTR hsubpd, 1, 0, 0
yading@11: AVX_INSTR hsubps, 1, 0, 0
yading@11: AVX_INSTR maxpd, 1, 0, 1
yading@11: AVX_INSTR maxps, 1, 0, 1
yading@11: AVX_INSTR maxsd, 1, 0, 1
yading@11: AVX_INSTR maxss, 1, 0, 1
yading@11: AVX_INSTR minpd, 1, 0, 1
yading@11: AVX_INSTR minps, 1, 0, 1
yading@11: AVX_INSTR minsd, 1, 0, 1
yading@11: AVX_INSTR minss, 1, 0, 1
yading@11: AVX_INSTR movhlps, 1, 0, 0
yading@11: AVX_INSTR movlhps, 1, 0, 0
yading@11: AVX_INSTR movsd, 1, 0, 0
yading@11: AVX_INSTR movss, 1, 0, 0
yading@11: AVX_INSTR mpsadbw, 0, 1, 0
yading@11: AVX_INSTR mulpd, 1, 0, 1
yading@11: AVX_INSTR mulps, 1, 0, 1
yading@11: AVX_INSTR mulsd, 1, 0, 1
yading@11: AVX_INSTR mulss, 1, 0, 1
yading@11: AVX_INSTR orpd, 1, 0, 1
yading@11: AVX_INSTR orps, 1, 0, 1
yading@11: AVX_INSTR pabsb, 0, 0, 0
yading@11: AVX_INSTR pabsw, 0, 0, 0
yading@11: AVX_INSTR pabsd, 0, 0, 0
yading@11: AVX_INSTR packsswb, 0, 0, 0
yading@11: AVX_INSTR packssdw, 0, 0, 0
yading@11: AVX_INSTR packuswb, 0, 0, 0
yading@11: AVX_INSTR packusdw, 0, 0, 0
yading@11: AVX_INSTR paddb, 0, 0, 1
yading@11: AVX_INSTR paddw, 0, 0, 1
yading@11: AVX_INSTR paddd, 0, 0, 1
yading@11: AVX_INSTR paddq, 0, 0, 1
yading@11: AVX_INSTR paddsb, 0, 0, 1
yading@11: AVX_INSTR paddsw, 0, 0, 1
yading@11: AVX_INSTR paddusb, 0, 0, 1
yading@11: AVX_INSTR paddusw, 0, 0, 1
yading@11: AVX_INSTR palignr, 0, 1, 0
yading@11: AVX_INSTR pand, 0, 0, 1
yading@11: AVX_INSTR pandn, 0, 0, 0
yading@11: AVX_INSTR pavgb, 0, 0, 1
yading@11: AVX_INSTR pavgw, 0, 0, 1
yading@11: AVX_INSTR pblendvb, 0, 0, 0
yading@11: AVX_INSTR pblendw, 0, 1, 0
yading@11: AVX_INSTR pcmpestri, 0, 0, 0
yading@11: AVX_INSTR pcmpestrm, 0, 0, 0
yading@11: AVX_INSTR pcmpistri, 0, 0, 0
yading@11: AVX_INSTR pcmpistrm, 0, 0, 0
yading@11: AVX_INSTR pcmpeqb, 0, 0, 1
yading@11: AVX_INSTR pcmpeqw, 0, 0, 1
yading@11: AVX_INSTR pcmpeqd, 0, 0, 1
yading@11: AVX_INSTR pcmpeqq, 0, 0, 1
yading@11: AVX_INSTR pcmpgtb, 0, 0, 0
yading@11: AVX_INSTR pcmpgtw, 0, 0, 0
yading@11: AVX_INSTR pcmpgtd, 0, 0, 0
yading@11: AVX_INSTR pcmpgtq, 0, 0, 0
yading@11: AVX_INSTR phaddw, 0, 0, 0
yading@11: AVX_INSTR phaddd, 0, 0, 0
yading@11: AVX_INSTR phaddsw, 0, 0, 0
yading@11: AVX_INSTR phsubw, 0, 0, 0
yading@11: AVX_INSTR phsubd, 0, 0, 0
yading@11: AVX_INSTR phsubsw, 0, 0, 0
yading@11: AVX_INSTR pmaddwd, 0, 0, 1
yading@11: AVX_INSTR pmaddubsw, 0, 0, 0
yading@11: AVX_INSTR pmaxsb, 0, 0, 1
yading@11: AVX_INSTR pmaxsw, 0, 0, 1
yading@11: AVX_INSTR pmaxsd, 0, 0, 1
yading@11: AVX_INSTR pmaxub, 0, 0, 1
yading@11: AVX_INSTR pmaxuw, 0, 0, 1
yading@11: AVX_INSTR pmaxud, 0, 0, 1
yading@11: AVX_INSTR pminsb, 0, 0, 1
yading@11: AVX_INSTR pminsw, 0, 0, 1
yading@11: AVX_INSTR pminsd, 0, 0, 1
yading@11: AVX_INSTR pminub, 0, 0, 1
yading@11: AVX_INSTR pminuw, 0, 0, 1
yading@11: AVX_INSTR pminud, 0, 0, 1
yading@11: AVX_INSTR pmovmskb, 0, 0, 0
yading@11: AVX_INSTR pmulhuw, 0, 0, 1
yading@11: AVX_INSTR pmulhrsw, 0, 0, 1
yading@11: AVX_INSTR pmulhw, 0, 0, 1
yading@11: AVX_INSTR pmullw, 0, 0, 1
yading@11: AVX_INSTR pmulld, 0, 0, 1
yading@11: AVX_INSTR pmuludq, 0, 0, 1
yading@11: AVX_INSTR pmuldq, 0, 0, 1
yading@11: AVX_INSTR por, 0, 0, 1
yading@11: AVX_INSTR psadbw, 0, 0, 1
yading@11: AVX_INSTR pshufb, 0, 0, 0
yading@11: AVX_INSTR pshufd, 0, 1, 0
yading@11: AVX_INSTR pshufhw, 0, 1, 0
yading@11: AVX_INSTR pshuflw, 0, 1, 0
yading@11: AVX_INSTR psignb, 0, 0, 0
yading@11: AVX_INSTR psignw, 0, 0, 0
yading@11: AVX_INSTR psignd, 0, 0, 0
yading@11: AVX_INSTR psllw, 0, 0, 0
yading@11: AVX_INSTR pslld, 0, 0, 0
yading@11: AVX_INSTR psllq, 0, 0, 0
yading@11: AVX_INSTR pslldq, 0, 0, 0
yading@11: AVX_INSTR psraw, 0, 0, 0
yading@11: AVX_INSTR psrad, 0, 0, 0
yading@11: AVX_INSTR psrlw, 0, 0, 0
yading@11: AVX_INSTR psrld, 0, 0, 0
yading@11: AVX_INSTR psrlq, 0, 0, 0
yading@11: AVX_INSTR psrldq, 0, 0, 0
yading@11: AVX_INSTR psubb, 0, 0, 0
yading@11: AVX_INSTR psubw, 0, 0, 0
yading@11: AVX_INSTR psubd, 0, 0, 0
yading@11: AVX_INSTR psubq, 0, 0, 0
yading@11: AVX_INSTR psubsb, 0, 0, 0
yading@11: AVX_INSTR psubsw, 0, 0, 0
yading@11: AVX_INSTR psubusb, 0, 0, 0
yading@11: AVX_INSTR psubusw, 0, 0, 0
yading@11: AVX_INSTR ptest, 0, 0, 0
yading@11: AVX_INSTR punpckhbw, 0, 0, 0
yading@11: AVX_INSTR punpckhwd, 0, 0, 0
yading@11: AVX_INSTR punpckhdq, 0, 0, 0
yading@11: AVX_INSTR punpckhqdq, 0, 0, 0
yading@11: AVX_INSTR punpcklbw, 0, 0, 0
yading@11: AVX_INSTR punpcklwd, 0, 0, 0
yading@11: AVX_INSTR punpckldq, 0, 0, 0
yading@11: AVX_INSTR punpcklqdq, 0, 0, 0
yading@11: AVX_INSTR pxor, 0, 0, 1
yading@11: AVX_INSTR shufps, 1, 1, 0
yading@11: AVX_INSTR subpd, 1, 0, 0
yading@11: AVX_INSTR subps, 1, 0, 0
yading@11: AVX_INSTR subsd, 1, 0, 0
yading@11: AVX_INSTR subss, 1, 0, 0
yading@11: AVX_INSTR unpckhpd, 1, 0, 0
yading@11: AVX_INSTR unpckhps, 1, 0, 0
yading@11: AVX_INSTR unpcklpd, 1, 0, 0
yading@11: AVX_INSTR unpcklps, 1, 0, 0
yading@11: AVX_INSTR xorpd, 1, 0, 1
yading@11: AVX_INSTR xorps, 1, 0, 1
yading@11: 
yading@11: ; 3DNow instructions, for sharing code between AVX, SSE and 3DN
yading@11: AVX_INSTR pfadd, 1, 0, 1
yading@11: AVX_INSTR pfsub, 1, 0, 0
yading@11: AVX_INSTR pfmul, 1, 0, 1
yading@11: 
yading@11: ; base-4 constants for shuffles
yading@11: %assign i 0
yading@11: %rep 256
yading@11:     %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
yading@11:     %if j < 10
yading@11:         CAT_XDEFINE q000, j, i
yading@11:     %elif j < 100
yading@11:         CAT_XDEFINE q00, j, i
yading@11:     %elif j < 1000
yading@11:         CAT_XDEFINE q0, j, i
yading@11:     %else
yading@11:         CAT_XDEFINE q, j, i
yading@11:     %endif
yading@11: %assign i i+1
yading@11: %endrep
yading@11: %undef i
yading@11: %undef j
yading@11: 
yading@11: %macro FMA_INSTR 3
yading@11:     %macro %1 5-8 %1, %2, %3
yading@11:         %if cpuflag(xop) || cpuflag(fma4)
yading@11:             v%6 %1, %2, %3, %4
yading@11:         %else
yading@11:             %ifidn %1, %4
yading@11:                 %7 %5, %2, %3
yading@11:                 %8 %1, %4, %5
yading@11:             %else
yading@11:                 %7 %1, %2, %3
yading@11:                 %8 %1, %4
yading@11:             %endif
yading@11:         %endif
yading@11:     %endmacro
yading@11: %endmacro
yading@11: 
yading@11: FMA_INSTR  fmaddps,   mulps, addps
yading@11: FMA_INSTR  pmacsdd,  pmulld, paddd
yading@11: FMA_INSTR  pmacsww,  pmullw, paddw
yading@11: FMA_INSTR pmadcswd, pmaddwd, paddd
yading@11: 
yading@11: ; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
yading@11: ; This lets us use tzcnt without bumping the yasm version requirement yet.
yading@11: %define tzcnt rep bsf