michael@0: ;
michael@0: ; jsimdext.inc - common declarations
michael@0: ;
michael@0: ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0: ; Copyright 2010 D. R. Commander
michael@0: ;
michael@0: ; Based on
michael@0: ; x86 SIMD extension for IJG JPEG library - version 1.02
michael@0: ;
michael@0: ; Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0: ;
michael@0: ; This software is provided 'as-is', without any express or implied
michael@0: ; warranty.  In no event will the authors be held liable for any damages
michael@0: ; arising from the use of this software.
michael@0: ;
michael@0: ; Permission is granted to anyone to use this software for any purpose,
michael@0: ; including commercial applications, and to alter it and redistribute it
michael@0: ; freely, subject to the following restrictions:
michael@0: ;
michael@0: ; 1. The origin of this software must not be misrepresented; you must not
michael@0: ;    claim that you wrote the original software. If you use this software
michael@0: ;    in a product, an acknowledgment in the product documentation would be
michael@0: ;    appreciated but is not required.
michael@0: ; 2. Altered source versions must be plainly marked as such, and must not be
michael@0: ;    misrepresented as being the original software.
michael@0: ; 3. This notice may not be removed or altered from any source distribution.
michael@0: ;
michael@0: ; [TAB8]
michael@0: 
michael@0: ; ==========================================================================
michael@0: ;  System-dependent configurations
michael@0: 
michael@0: %ifdef WIN32	; ----(nasm -fwin32 -DWIN32 ...)--------
michael@0: ; * Microsoft Visual C++
michael@0: ; * MinGW (Minimalist GNU for Windows)
michael@0: ; * CygWin
michael@0: ; * LCC-Win32
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %ifdef __YASM_VER__
michael@0: %define SEG_TEXT    .text  align=16
michael@0: %define SEG_CONST   .rdata align=16
michael@0: %else
michael@0: %define SEG_TEXT    .text  align=16 public use32 class=CODE
michael@0: %define SEG_CONST   .rdata align=16 public use32 class=CONST
michael@0: %endif
michael@0: 
michael@0: %elifdef WIN64	; ----(nasm -fwin64 -DWIN64 ...)--------
michael@0: ; * Microsoft Visual C++
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %ifdef __YASM_VER__
michael@0: %define SEG_TEXT    .text  align=16
michael@0: %define SEG_CONST   .rdata align=16
michael@0: %else
michael@0: %define SEG_TEXT    .text  align=16 public use64 class=CODE
michael@0: %define SEG_CONST   .rdata align=16 public use64 class=CONST
michael@0: %endif
michael@0: %define EXTN(name)  name			; foo() -> foo
michael@0: 
michael@0: %elifdef OBJ32	; ----(nasm -fobj -DOBJ32 ...)----------
michael@0: ; * Borland C++ (Win32)
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %define SEG_TEXT    .text  align=16 public use32 class=CODE
michael@0: %define SEG_CONST   .data  align=16 public use32 class=DATA
michael@0: 
michael@0: %elifdef ELF	; ----(nasm -felf[64] -DELF ...)------------
michael@0: ; * Linux
michael@0: ; * *BSD family Unix using elf format
michael@0: ; * Unix System V, including Solaris x86, UnixWare and SCO Unix
michael@0: 
michael@0: ; mark stack as non-executable
michael@0: section .note.GNU-stack noalloc noexec nowrite progbits
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %ifdef __x86_64__
michael@0: %define SEG_TEXT    .text   progbits align=16
michael@0: %define SEG_CONST   .rodata progbits align=16
michael@0: %else
michael@0: %define SEG_TEXT    .text   progbits alloc exec   nowrite align=16
michael@0: %define SEG_CONST   .rodata progbits alloc noexec nowrite align=16
michael@0: %endif
michael@0: 
michael@0: ; To make the code position-independent, append -DPIC to the commandline
michael@0: ;
michael@0: %define GOT_SYMBOL  _GLOBAL_OFFSET_TABLE_	; ELF supports PIC
michael@0: %define EXTN(name)  name			; foo() -> foo
michael@0: 
michael@0: %elifdef AOUT	; ----(nasm -faoutb/aout -DAOUT ...)----
michael@0: ; * Older Linux using a.out format  (nasm -f aout -DAOUT ...)
michael@0: ; * *BSD family Unix using a.out format  (nasm -f aoutb -DAOUT ...)
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %define SEG_TEXT    .text
michael@0: %define SEG_CONST   .data
michael@0: 
michael@0: ; To make the code position-independent, append -DPIC to the commandline
michael@0: ;
michael@0: %define GOT_SYMBOL  __GLOBAL_OFFSET_TABLE_	; BSD-style a.out supports PIC
michael@0: 
michael@0: %elifdef MACHO	; ----(nasm -fmacho -DMACHO ...)--------
michael@0: ; * NeXTstep/OpenStep/Rhapsody/Darwin/MacOS X (Mach-O format)
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %define SEG_TEXT    .text  ;align=16	; nasm doesn't accept align=16. why?
michael@0: %define SEG_CONST   .rodata align=16
michael@0: 
michael@0: ; The generation of position-independent code (PIC) is the default on Darwin.
michael@0: ;
michael@0: %define PIC
michael@0: %define GOT_SYMBOL  _MACHO_PIC_		; Mach-O style code-relative addressing
michael@0: 
michael@0: %else		; ----(Other case)----------------------
michael@0: 
michael@0: ; -- segment definition --
michael@0: ;
michael@0: %define SEG_TEXT    .text
michael@0: %define SEG_CONST   .data
michael@0: 
michael@0: %endif	; ----------------------------------------------
michael@0: 
michael@0: ; ==========================================================================
michael@0: 
michael@0: ; --------------------------------------------------------------------------
michael@0: ;  Common types
michael@0: ;
michael@0: %ifdef __x86_64__
michael@0: %define POINTER                 qword           ; general pointer type
michael@0: %define SIZEOF_POINTER          SIZEOF_QWORD    ; sizeof(POINTER)
michael@0: %define POINTER_BIT             QWORD_BIT       ; sizeof(POINTER)*BYTE_BIT
michael@0: %else
michael@0: %define POINTER                 dword           ; general pointer type
michael@0: %define SIZEOF_POINTER          SIZEOF_DWORD    ; sizeof(POINTER)
michael@0: %define POINTER_BIT             DWORD_BIT       ; sizeof(POINTER)*BYTE_BIT
michael@0: %endif
michael@0: 
michael@0: %define INT                     dword           ; signed integer type
michael@0: %define SIZEOF_INT              SIZEOF_DWORD    ; sizeof(INT)
michael@0: %define INT_BIT                 DWORD_BIT       ; sizeof(INT)*BYTE_BIT
michael@0: 
michael@0: %define FP32                    dword           ; IEEE754 single
michael@0: %define SIZEOF_FP32             SIZEOF_DWORD    ; sizeof(FP32)
michael@0: %define FP32_BIT                DWORD_BIT       ; sizeof(FP32)*BYTE_BIT
michael@0: 
michael@0: %define MMWORD                  qword           ; int64  (MMX register)
michael@0: %define SIZEOF_MMWORD           SIZEOF_QWORD    ; sizeof(MMWORD)
michael@0: %define MMWORD_BIT              QWORD_BIT       ; sizeof(MMWORD)*BYTE_BIT
michael@0: 
michael@0: ; NASM is buggy and doesn't properly handle operand sizes for SSE
michael@0: ; instructions, so for now we have to define XMMWORD as blank.
michael@0: %define XMMWORD                                 ; int128 (SSE register)
michael@0: %define SIZEOF_XMMWORD          SIZEOF_OWORD    ; sizeof(XMMWORD)
michael@0: %define XMMWORD_BIT             OWORD_BIT       ; sizeof(XMMWORD)*BYTE_BIT
michael@0: 
michael@0: ; Similar hacks for when we load a dword or MMWORD into an xmm# register
michael@0: %define XMM_DWORD
michael@0: %define XMM_MMWORD
michael@0: 
michael@0: %define SIZEOF_BYTE             1               ; sizeof(BYTE)
michael@0: %define SIZEOF_WORD             2               ; sizeof(WORD)
michael@0: %define SIZEOF_DWORD            4               ; sizeof(DWORD)
michael@0: %define SIZEOF_QWORD            8               ; sizeof(QWORD)
michael@0: %define SIZEOF_OWORD            16              ; sizeof(OWORD)
michael@0: 
michael@0: %define BYTE_BIT                8               ; CHAR_BIT in C
michael@0: %define WORD_BIT                16              ; sizeof(WORD)*BYTE_BIT
michael@0: %define DWORD_BIT               32              ; sizeof(DWORD)*BYTE_BIT
michael@0: %define QWORD_BIT               64              ; sizeof(QWORD)*BYTE_BIT
michael@0: %define OWORD_BIT               128             ; sizeof(OWORD)*BYTE_BIT
michael@0: 
michael@0: ; --------------------------------------------------------------------------
michael@0: ;  External Symbol Name
michael@0: ;
michael@0: %ifndef EXTN
michael@0: %define EXTN(name)   _ %+ name		; foo() -> _foo
michael@0: %endif
michael@0: 
michael@0: ; --------------------------------------------------------------------------
michael@0: ;  Macros for position-independent code (PIC) support
michael@0: ;
michael@0: %ifndef GOT_SYMBOL
michael@0: %undef PIC
michael@0: %endif
michael@0: 
michael@0: %ifdef PIC ; -------------------------------------------
michael@0: 
michael@0: %ifidn GOT_SYMBOL,_MACHO_PIC_ ; --------------------
michael@0: 
michael@0: ; At present, nasm doesn't seem to support PIC generation for Mach-O.
michael@0: ; The PIC support code below is a little tricky.
michael@0: 
michael@0: 	SECTION	SEG_CONST
michael@0: const_base:
michael@0: 
michael@0: %define GOTOFF(got,sym) (got) + (sym) - const_base
michael@0: 
michael@0: %imacro get_GOT	1
michael@0: 	; NOTE: this macro destroys ecx resister.
michael@0: 	call	%%geteip
michael@0: 	add	ecx, byte (%%ref - $)
michael@0: 	jmp	short %%adjust
michael@0: %%geteip:
michael@0: 	mov	ecx, POINTER [esp]
michael@0: 	ret
michael@0: %%adjust:
michael@0: 	push	ebp
michael@0: 	xor	ebp,ebp		; ebp = 0
michael@0: %ifidni %1,ebx	; (%1 == ebx)
michael@0: 	; db 0x8D,0x9C + jmp near const_base =
michael@0: 	;   lea ebx, [ecx+ebp*8+(const_base-%%ref)] ; 8D,9C,E9,(offset32)
michael@0: 	db	0x8D,0x9C		; 8D,9C
michael@0: 	jmp	near const_base		; E9,(const_base-%%ref)
michael@0: %%ref:
michael@0: %else  ; (%1 != ebx)
michael@0: 	; db 0x8D,0x8C + jmp near const_base =
michael@0: 	;   lea ecx, [ecx+ebp*8+(const_base-%%ref)] ; 8D,8C,E9,(offset32)
michael@0: 	db	0x8D,0x8C		; 8D,8C
michael@0: 	jmp	near const_base		; E9,(const_base-%%ref)
michael@0: %%ref:	mov	%1, ecx
michael@0: %endif ; (%1 == ebx)
michael@0: 	pop	ebp
michael@0: %endmacro
michael@0: 
michael@0: %else	; GOT_SYMBOL != _MACHO_PIC_ ----------------
michael@0: 
michael@0: %define GOTOFF(got,sym) (got) + (sym) wrt ..gotoff
michael@0: 
michael@0: %imacro get_GOT	1
michael@0: 	extern	GOT_SYMBOL
michael@0: 	call	%%geteip
michael@0: 	add	%1, GOT_SYMBOL + $$ - $ wrt ..gotpc
michael@0: 	jmp	short %%done
michael@0: %%geteip:
michael@0: 	mov	%1, POINTER [esp]
michael@0: 	ret
michael@0: %%done:
michael@0: %endmacro
michael@0: 
michael@0: %endif	; GOT_SYMBOL == _MACHO_PIC_ ----------------
michael@0: 
michael@0: %imacro pushpic	1.nolist
michael@0: 	push	%1
michael@0: %endmacro
michael@0: %imacro poppic	1.nolist
michael@0: 	pop	%1
michael@0: %endmacro
michael@0: %imacro movpic	2.nolist
michael@0: 	mov	%1,%2
michael@0: %endmacro
michael@0: 
michael@0: %else	; !PIC -----------------------------------------
michael@0: 
michael@0: %define GOTOFF(got,sym) (sym)
michael@0: 
michael@0: %imacro get_GOT	1.nolist
michael@0: %endmacro
michael@0: %imacro pushpic	1.nolist
michael@0: %endmacro
michael@0: %imacro poppic	1.nolist
michael@0: %endmacro
michael@0: %imacro movpic	2.nolist
michael@0: %endmacro
michael@0: 
michael@0: %endif	;  PIC -----------------------------------------
michael@0: 
michael@0: ; --------------------------------------------------------------------------
michael@0: ;  Align the next instruction on {2,4,8,16,..}-byte boundary.
michael@0: ;  ".balign n,,m" in GNU as
michael@0: ;
michael@0: %define MSKLE(x,y)  (~(((y) & 0xFFFF) - ((x) & 0xFFFF)) >> 16)
michael@0: %define FILLB(b,n)  (($$-(b)) & ((n)-1))
michael@0: 
michael@0: %imacro alignx 1-2.nolist 0xFFFF
michael@0: %%bs:	times MSKLE(FILLB(%%bs,%1),%2) & MSKLE(16,FILLB($,%1)) & FILLB($,%1) \
michael@0: 	       db 0x90                               ; nop
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/9 \
michael@0: 	       db 0x8D,0x9C,0x23,0x00,0x00,0x00,0x00 ; lea ebx,[ebx+0x00000000]
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/7 \
michael@0: 	       db 0x8D,0xAC,0x25,0x00,0x00,0x00,0x00 ; lea ebp,[ebp+0x00000000]
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/6 \
michael@0: 	       db 0x8D,0xAD,0x00,0x00,0x00,0x00      ; lea ebp,[ebp+0x00000000]
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/4 \
michael@0: 	       db 0x8D,0x6C,0x25,0x00                ; lea ebp,[ebp+0x00]
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/3 \
michael@0: 	       db 0x8D,0x6D,0x00                     ; lea ebp,[ebp+0x00]
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/2 \
michael@0: 	       db 0x8B,0xED                          ; mov ebp,ebp
michael@0: 	times MSKLE(FILLB(%%bs,%1),%2) & FILLB($,%1)/1 \
michael@0: 	       db 0x90                               ; nop
michael@0: %endmacro
michael@0: 
michael@0: ; Align the next data on {2,4,8,16,..}-byte boundary.
michael@0: ;
michael@0: %imacro alignz 1.nolist
michael@0: 	align %1, db 0		; filling zeros
michael@0: %endmacro
michael@0: 
michael@0: %ifdef __x86_64__
michael@0: 
michael@0: %ifdef WIN64
michael@0: 
michael@0: %imacro collect_args 0
michael@0: 	push r12
michael@0: 	push r13
michael@0: 	push r14
michael@0: 	push r15
michael@0: 	mov r10, rcx
michael@0: 	mov r11, rdx
michael@0: 	mov r12, r8
michael@0: 	mov r13, r9
michael@0: 	mov r14, [rax+48]
michael@0: 	mov r15, [rax+56]
michael@0: 	push rsi
michael@0: 	push rdi
michael@0: 	sub     rsp, SIZEOF_XMMWORD
michael@0: 	movaps  XMMWORD [rsp], xmm6
michael@0: 	sub     rsp, SIZEOF_XMMWORD
michael@0: 	movaps  XMMWORD [rsp], xmm7
michael@0: %endmacro
michael@0: 
michael@0: %imacro uncollect_args 0
michael@0: 	movaps  xmm7, XMMWORD [rsp]
michael@0: 	add     rsp, SIZEOF_XMMWORD
michael@0: 	movaps  xmm6, XMMWORD [rsp]
michael@0: 	add     rsp, SIZEOF_XMMWORD
michael@0: 	pop rdi
michael@0: 	pop rsi
michael@0: 	pop r15
michael@0: 	pop r14
michael@0: 	pop r13
michael@0: 	pop r12
michael@0: %endmacro
michael@0: 
michael@0: %else
michael@0: 
michael@0: %imacro collect_args 0
michael@0: 	push r10
michael@0: 	push r11
michael@0: 	push r12
michael@0: 	push r13
michael@0: 	push r14
michael@0: 	push r15
michael@0: 	mov r10, rdi
michael@0: 	mov r11, rsi
michael@0: 	mov r12, rdx
michael@0: 	mov r13, rcx
michael@0: 	mov r14, r8
michael@0: 	mov r15, r9
michael@0: %endmacro
michael@0: 
michael@0: %imacro uncollect_args 0
michael@0: 	pop r15
michael@0: 	pop r14
michael@0: 	pop r13
michael@0: 	pop r12
michael@0: 	pop r11
michael@0: 	pop r10
michael@0: %endmacro
michael@0: 
michael@0: %endif
michael@0: 
michael@0: %endif
michael@0: 
michael@0: ; --------------------------------------------------------------------------
michael@0: ;  Defines picked up from the C headers
michael@0: ;
michael@0: %include "jsimdcfg.inc"
michael@0: 
michael@0: ; --------------------------------------------------------------------------