Mercurial > sdl-ios-xcode
view src/video/mmx.h @ 5067:61d53410eb41
Fixed bug #859
CREATE_SUBDIRS helps a lot if browsing HTML documentation in a file browser.
ALWAYS_DETAILED_SEC makes sure everything has at least the automatic
documentation like function prototype and source references.
STRIP_FROM_PATH allows you to include only the relevant portions of the files'
paths, cleaning up both the file list and directory tree, though you need to
change the path listed here to match wherever you put SDL.
ALIASES avoids some warnings generated by
C:\source\svn.libsdl.org\trunk\SDL\src\joystick\darwin\10.3.9-FIX\IOHIDLib.h.
It seems Apple uses a few commands which are not normally supported by Doxygen.
BUILTIN_STL_SUPPORT adds support for parsing code which makes use of the
standard template library. There isn't a lot of C++ in SDL (some in bwindow at
least), but this still seems like a good idea.
TYPEDEF_HIDES_STRUCT means that for code like this:
typedef struct A {int B;} C;
C is documented as a structure containing B instead of a typedef mapped to A.
EXTRACT_ALL, EXTRACT_PRIVATE, EXTRACT_STATIC, EXTRACT_LOCAL_METHODS,
EXTRACT_ANON_NSPACES and INTERNAL_DOCS make sure that _everything_ is
documented.
CASE_SENSE_NAMES = NO avoids potential conflicts when building documentation on
case insensitive file systems like NTFS and FAT32.
WARN_NO_PARAMDOC lets you know when you have documented some, but not all, of
the parameters of a function. This is useful when you're working on adding
such documentation since it makes partially documented functions easier to
spot.
WARN_LOGFILE writes warnings to a seperate file instead of mixing them in with
stdout. When not running in quiet mode, these warnings can be hard to spot
without this flag.
I added *.h.in and *.h.default to FILE_PATTERNS to generate documentation for
config.h.in and config.h.default.
RECURSIVE tells doxygen to look not only in the input directory, but also in
subfolders.
EXCLUDE avoids documenting things like test programs, examples and templates
which need to be documented separately.
I've used EXCLUDE_PATTERNS to exclude non-source subdirectories that often find
their way into source folders (such as obj or .svn).
EXAMPLE_PATH lists directories doxygen will search to find included example
code. So far, SDL doesn't really use this feature, but I've listed some likely
locations.
SOURCE_BROWSER adds syntax highlighted source code to the HTML output.
USE_HTAGS is nice, but not available on Windows.
INLINE_SOURCES adds the body of a function to it's documentation so you can
quickly see exactly what it does.
ALPHABETICAL_INDEX generates an alphabetical list of all structures, functions,
etc., which makes it much easier to find what you're looking for.
IGNORE_PREFIX skips the SDL_ prefix when deciding which index page to place an
item on so you don't have everything show up under "S".
HTML_DYNAMIC_SECTIONS hides the includes/included by diagrams by default and
adds JavaScript to allow the user to show and hide them by clicking a link.
ENUM_VALUES_PER_LINE = 1 makes enums easier to read by placing each value on
it's own line.
GENERATE_TREEVIEW produces a two frame index page with a navigation tree on the
left.
I have LaTeX and man pages turned off to speed up doxygen, you may want to turn
them back on yourself.
I added _WIN32=1 to PREDEFINED to cause SDL to output documentation related to
Win32 builds of SDL. Normally, doxygen gets confused since there are multiple
definitions for various structures and formats that vary by platform. Without
this doxygen can produce broken documentation or, if you're lucky, output
documentation only for the dummy drivers, which isn't very useful. You need to
pick a platform.
GENERATE_TAGFILE produces a file which can be used to link other doxygen
documentation to the SDL documentation.
CLASS_DIAGRAMS turns on class diagrams even when dot is not available.
HAVE_DOT tells doxygen to try to use dot to generate diagrams.
TEMPLATE_RELATIONS and INCLUDE_GRAPH add additional diagrams to the
documentation.
DOT_MULTI_TARGETS speeds up dot.
OUTPUT_DIRECTORY, INPUT and other paths reflect the fact that this Doxyfile is
intended to process src as well as include and is being run from a separate
subdirectory. Doxygen produces several temporary files while it's running and
if interrupted, can leave those files behind. It's easier to clean up if there
aren't a hundred or so files in the same folder. I typically run doxygen in
SDL/doxy and set the output directory to '.'. Since doxygen puts it's output
in subfolders by type, this keeps things pretty well organised. You could use
'../doc' instead and get the same results.
| author | Sam Lantinga <slouken@libsdl.org> |
|---|---|
| date | Fri, 21 Jan 2011 12:57:01 -0800 |
| parents | c121d94672cb |
| children | f2c2f0ecba5f |
line wrap: on
line source
/* mmx.h MultiMedia eXtensions GCC interface library for IA32. To use this library, simply include this header file and compile with GCC. You MUST have inlining enabled in order for mmx_ok() to work; this can be done by simply using -O on the GCC command line. Compiling with -DMMX_TRACE will cause detailed trace output to be sent to stderr for each mmx operation. This adds lots of code, and obviously slows execution to a crawl, but can be very useful for debugging. THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR ANY PARTICULAR PURPOSE. 1997-99 by H. Dietz and R. Fisher Notes: It appears that the latest gas has the pand problem fixed, therefore I'll undefine BROKEN_PAND by default. */ #ifndef _MMX_H #define _MMX_H /* Warning: at this writing, the version of GAS packaged with most Linux distributions does not handle the parallel AND operation mnemonic correctly. If the symbol BROKEN_PAND is defined, a slower alternative coding will be used. If execution of mmxtest results in an illegal instruction fault, define this symbol. */ #undef BROKEN_PAND /* The type of an value that fits in an MMX register (note that long long constant values MUST be suffixed by LL and unsigned long long values by ULL, lest they be truncated by the compiler) */ typedef union { long long q; /* Quadword (64-bit) value */ unsigned long long uq; /* Unsigned Quadword */ int d[2]; /* 2 Doubleword (32-bit) values */ unsigned int ud[2]; /* 2 Unsigned Doubleword */ short w[4]; /* 4 Word (16-bit) values */ unsigned short uw[4]; /* 4 Unsigned Word */ char b[8]; /* 8 Byte (8-bit) values */ unsigned char ub[8]; /* 8 Unsigned Byte */ float s[2]; /* Single-precision (32-bit) value */ } __attribute__ ((aligned(8))) mmx_t; /* On an 8-byte (64-bit) boundary */ #if 0 /* Function to test if multimedia instructions are supported... */ inline extern int mm_support(void) { /* Returns 1 if MMX instructions are supported, 3 if Cyrix MMX and Extended MMX instructions are supported 5 if AMD MMX and 3DNow! instructions are supported 0 if hardware does not support any of these */ register int rval = 0; __asm__ __volatile__( /* See if CPUID instruction is supported ... */ /* ... Get copies of EFLAGS into eax and ecx */ "pushf\n\t" "popl %%eax\n\t" "movl %%eax, %%ecx\n\t" /* ... Toggle the ID bit in one copy and store */ /* to the EFLAGS reg */ "xorl $0x200000, %%eax\n\t" "push %%eax\n\t" "popf\n\t" /* ... Get the (hopefully modified) EFLAGS */ "pushf\n\t" "popl %%eax\n\t" /* ... Compare and test result */ "xorl %%eax, %%ecx\n\t" "testl $0x200000, %%ecx\n\t" "jz NotSupported1\n\t" /* CPUID not supported */ /* Get standard CPUID information, and go to a specific vendor section */ "movl $0, %%eax\n\t" "cpuid\n\t" /* Check for Intel */ "cmpl $0x756e6547, %%ebx\n\t" "jne TryAMD\n\t" "cmpl $0x49656e69, %%edx\n\t" "jne TryAMD\n\t" "cmpl $0x6c65746e, %%ecx\n" "jne TryAMD\n\t" "jmp Intel\n\t" /* Check for AMD */ "\nTryAMD:\n\t" "cmpl $0x68747541, %%ebx\n\t" "jne TryCyrix\n\t" "cmpl $0x69746e65, %%edx\n\t" "jne TryCyrix\n\t" "cmpl $0x444d4163, %%ecx\n" "jne TryCyrix\n\t" "jmp AMD\n\t" /* Check for Cyrix */ "\nTryCyrix:\n\t" "cmpl $0x69727943, %%ebx\n\t" "jne NotSupported2\n\t" "cmpl $0x736e4978, %%edx\n\t" "jne NotSupported3\n\t" "cmpl $0x64616574, %%ecx\n\t" "jne NotSupported4\n\t" /* Drop through to Cyrix... */ /* Cyrix Section */ /* See if extended CPUID level 80000001 is supported */ /* The value of CPUID/80000001 for the 6x86MX is undefined according to the Cyrix CPU Detection Guide (Preliminary Rev. 1.01 table 1), so we'll check the value of eax for CPUID/0 to see if standard CPUID level 2 is supported. According to the table, the only CPU which supports level 2 is also the only one which supports extended CPUID levels. */ "cmpl $0x2, %%eax\n\t" "jne MMXtest\n\t" /* Use standard CPUID instead */ /* Extended CPUID supported (in theory), so get extended features */ "movl $0x80000001, %%eax\n\t" "cpuid\n\t" "testl $0x00800000, %%eax\n\t" /* Test for MMX */ "jz NotSupported5\n\t" /* MMX not supported */ "testl $0x01000000, %%eax\n\t" /* Test for Ext'd MMX */ "jnz EMMXSupported\n\t" "movl $1, %0:\n\n\t" /* MMX Supported */ "jmp Return\n\n" "EMMXSupported:\n\t" "movl $3, %0:\n\n\t" /* EMMX and MMX Supported */ "jmp Return\n\t" /* AMD Section */ "AMD:\n\t" /* See if extended CPUID is supported */ "movl $0x80000000, %%eax\n\t" "cpuid\n\t" "cmpl $0x80000000, %%eax\n\t" "jl MMXtest\n\t" /* Use standard CPUID instead */ /* Extended CPUID supported, so get extended features */ "movl $0x80000001, %%eax\n\t" "cpuid\n\t" "testl $0x00800000, %%edx\n\t" /* Test for MMX */ "jz NotSupported6\n\t" /* MMX not supported */ "testl $0x80000000, %%edx\n\t" /* Test for 3DNow! */ "jnz ThreeDNowSupported\n\t" "movl $1, %0:\n\n\t" /* MMX Supported */ "jmp Return\n\n" "ThreeDNowSupported:\n\t" "movl $5, %0:\n\n\t" /* 3DNow! and MMX Supported */ "jmp Return\n\t" /* Intel Section */ "Intel:\n\t" /* Check for MMX */ "MMXtest:\n\t" "movl $1, %%eax\n\t" "cpuid\n\t" "testl $0x00800000, %%edx\n\t" /* Test for MMX */ "jz NotSupported7\n\t" /* MMX Not supported */ "movl $1, %0:\n\n\t" /* MMX Supported */ "jmp Return\n\t" /* Nothing supported */ "\nNotSupported1:\n\t" "#movl $101, %0:\n\n\t" "\nNotSupported2:\n\t" "#movl $102, %0:\n\n\t" "\nNotSupported3:\n\t" "#movl $103, %0:\n\n\t" "\nNotSupported4:\n\t" "#movl $104, %0:\n\n\t" "\nNotSupported5:\n\t" "#movl $105, %0:\n\n\t" "\nNotSupported6:\n\t" "#movl $106, %0:\n\n\t" "\nNotSupported7:\n\t" "#movl $107, %0:\n\n\t" "movl $0, %0:\n\n\t" "Return:\n\t":"=a"(rval): /* no input */ :"eax", "ebx", "ecx", "edx"); /* Return */ return (rval); } /* Function to test if mmx instructions are supported... */ inline extern int mmx_ok(void) { /* Returns 1 if MMX instructions are supported, 0 otherwise */ return (mm_support() & 0x1); } #endif /* Helper functions for the instruction macros that follow... (note that memory-to-register, m2r, instructions are nearly as efficient as register-to-register, r2r, instructions; however, memory-to-memory instructions are really simulated as a convenience, and are only 1/3 as efficient) */ #ifdef MMX_TRACE /* Include the stuff for printing a trace to stderr... */ #define mmx_i2r(op, imm, reg) \ { \ mmx_t mmx_trace; \ mmx_trace.uq = (imm); \ printf(#op "_i2r(" #imm "=0x%08x%08x, ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ ("movq %%" #reg ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#reg "=0x%08x%08x) => ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ (#op " %0, %%" #reg \ : /* nothing */ \ : "X" (imm)); \ __asm__ __volatile__ ("movq %%" #reg ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#reg "=0x%08x%08x\n", \ mmx_trace.d[1], mmx_trace.d[0]); \ } #define mmx_m2r(op, mem, reg) \ { \ mmx_t mmx_trace; \ mmx_trace = (mem); \ printf(#op "_m2r(" #mem "=0x%08x%08x, ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ ("movq %%" #reg ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#reg "=0x%08x%08x) => ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ (#op " %0, %%" #reg \ : /* nothing */ \ : "X" (mem)); \ __asm__ __volatile__ ("movq %%" #reg ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#reg "=0x%08x%08x\n", \ mmx_trace.d[1], mmx_trace.d[0]); \ } #define mmx_r2m(op, reg, mem) \ { \ mmx_t mmx_trace; \ __asm__ __volatile__ ("movq %%" #reg ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#op "_r2m(" #reg "=0x%08x%08x, ", \ mmx_trace.d[1], mmx_trace.d[0]); \ mmx_trace = (mem); \ printf(#mem "=0x%08x%08x) => ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ (#op " %%" #reg ", %0" \ : "=X" (mem) \ : /* nothing */ ); \ mmx_trace = (mem); \ printf(#mem "=0x%08x%08x\n", \ mmx_trace.d[1], mmx_trace.d[0]); \ } #define mmx_r2r(op, regs, regd) \ { \ mmx_t mmx_trace; \ __asm__ __volatile__ ("movq %%" #regs ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#op "_r2r(" #regs "=0x%08x%08x, ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ ("movq %%" #regd ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#regd "=0x%08x%08x) => ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ (#op " %" #regs ", %" #regd); \ __asm__ __volatile__ ("movq %%" #regd ", %0" \ : "=X" (mmx_trace) \ : /* nothing */ ); \ printf(#regd "=0x%08x%08x\n", \ mmx_trace.d[1], mmx_trace.d[0]); \ } #define mmx_m2m(op, mems, memd) \ { \ mmx_t mmx_trace; \ mmx_trace = (mems); \ printf(#op "_m2m(" #mems "=0x%08x%08x, ", \ mmx_trace.d[1], mmx_trace.d[0]); \ mmx_trace = (memd); \ printf(#memd "=0x%08x%08x) => ", \ mmx_trace.d[1], mmx_trace.d[0]); \ __asm__ __volatile__ ("movq %0, %%mm0\n\t" \ #op " %1, %%mm0\n\t" \ "movq %%mm0, %0" \ : "=X" (memd) \ : "X" (mems)); \ mmx_trace = (memd); \ printf(#memd "=0x%08x%08x\n", \ mmx_trace.d[1], mmx_trace.d[0]); \ } #else /* These macros are a lot simpler without the tracing... */ #define mmx_i2r(op, imm, reg) \ __asm__ __volatile__ (#op " %0, %%" #reg \ : /* nothing */ \ : "X" (imm) ) #define mmx_m2r(op, mem, reg) \ __asm__ __volatile__ (#op " %0, %%" #reg \ : /* nothing */ \ : "m" (mem)) #define mmx_r2m(op, reg, mem) \ __asm__ __volatile__ (#op " %%" #reg ", %0" \ : "=X" (mem) \ : /* nothing */ ) #define mmx_r2r(op, regs, regd) \ __asm__ __volatile__ (#op " %" #regs ", %" #regd) #define mmx_m2m(op, mems, memd) \ __asm__ __volatile__ ("movq %0, %%mm0\n\t" \ #op " %1, %%mm0\n\t" \ "movq %%mm0, %0" \ : "=X" (memd) \ : "X" (mems)) #endif /* 1x64 MOVe Quadword (this is both a load and a store... in fact, it is the only way to store) */ #define movq_m2r(var, reg) mmx_m2r(movq, var, reg) #define movq_r2m(reg, var) mmx_r2m(movq, reg, var) #define movq_r2r(regs, regd) mmx_r2r(movq, regs, regd) #define movq(vars, vard) \ __asm__ __volatile__ ("movq %1, %%mm0\n\t" \ "movq %%mm0, %0" \ : "=X" (vard) \ : "X" (vars)) /* 1x32 MOVe Doubleword (like movq, this is both load and store... but is most useful for moving things between mmx registers and ordinary registers) */ #define movd_m2r(var, reg) mmx_m2r(movd, var, reg) #define movd_r2m(reg, var) mmx_r2m(movd, reg, var) #define movd_r2r(regs, regd) mmx_r2r(movd, regs, regd) #define movd(vars, vard) \ __asm__ __volatile__ ("movd %1, %%mm0\n\t" \ "movd %%mm0, %0" \ : "=X" (vard) \ : "X" (vars)) /* 2x32, 4x16, and 8x8 Parallel ADDs */ #define paddd_m2r(var, reg) mmx_m2r(paddd, var, reg) #define paddd_r2r(regs, regd) mmx_r2r(paddd, regs, regd) #define paddd(vars, vard) mmx_m2m(paddd, vars, vard) #define paddw_m2r(var, reg) mmx_m2r(paddw, var, reg) #define paddw_r2r(regs, regd) mmx_r2r(paddw, regs, regd) #define paddw(vars, vard) mmx_m2m(paddw, vars, vard) #define paddb_m2r(var, reg) mmx_m2r(paddb, var, reg) #define paddb_r2r(regs, regd) mmx_r2r(paddb, regs, regd) #define paddb(vars, vard) mmx_m2m(paddb, vars, vard) /* 4x16 and 8x8 Parallel ADDs using Saturation arithmetic */ #define paddsw_m2r(var, reg) mmx_m2r(paddsw, var, reg) #define paddsw_r2r(regs, regd) mmx_r2r(paddsw, regs, regd) #define paddsw(vars, vard) mmx_m2m(paddsw, vars, vard) #define paddsb_m2r(var, reg) mmx_m2r(paddsb, var, reg) #define paddsb_r2r(regs, regd) mmx_r2r(paddsb, regs, regd) #define paddsb(vars, vard) mmx_m2m(paddsb, vars, vard) /* 4x16 and 8x8 Parallel ADDs using Unsigned Saturation arithmetic */ #define paddusw_m2r(var, reg) mmx_m2r(paddusw, var, reg) #define paddusw_r2r(regs, regd) mmx_r2r(paddusw, regs, regd) #define paddusw(vars, vard) mmx_m2m(paddusw, vars, vard) #define paddusb_m2r(var, reg) mmx_m2r(paddusb, var, reg) #define paddusb_r2r(regs, regd) mmx_r2r(paddusb, regs, regd) #define paddusb(vars, vard) mmx_m2m(paddusb, vars, vard) /* 2x32, 4x16, and 8x8 Parallel SUBs */ #define psubd_m2r(var, reg) mmx_m2r(psubd, var, reg) #define psubd_r2r(regs, regd) mmx_r2r(psubd, regs, regd) #define psubd(vars, vard) mmx_m2m(psubd, vars, vard) #define psubw_m2r(var, reg) mmx_m2r(psubw, var, reg) #define psubw_r2r(regs, regd) mmx_r2r(psubw, regs, regd) #define psubw(vars, vard) mmx_m2m(psubw, vars, vard) #define psubb_m2r(var, reg) mmx_m2r(psubb, var, reg) #define psubb_r2r(regs, regd) mmx_r2r(psubb, regs, regd) #define psubb(vars, vard) mmx_m2m(psubb, vars, vard) /* 4x16 and 8x8 Parallel SUBs using Saturation arithmetic */ #define psubsw_m2r(var, reg) mmx_m2r(psubsw, var, reg) #define psubsw_r2r(regs, regd) mmx_r2r(psubsw, regs, regd) #define psubsw(vars, vard) mmx_m2m(psubsw, vars, vard) #define psubsb_m2r(var, reg) mmx_m2r(psubsb, var, reg) #define psubsb_r2r(regs, regd) mmx_r2r(psubsb, regs, regd) #define psubsb(vars, vard) mmx_m2m(psubsb, vars, vard) /* 4x16 and 8x8 Parallel SUBs using Unsigned Saturation arithmetic */ #define psubusw_m2r(var, reg) mmx_m2r(psubusw, var, reg) #define psubusw_r2r(regs, regd) mmx_r2r(psubusw, regs, regd) #define psubusw(vars, vard) mmx_m2m(psubusw, vars, vard) #define psubusb_m2r(var, reg) mmx_m2r(psubusb, var, reg) #define psubusb_r2r(regs, regd) mmx_r2r(psubusb, regs, regd) #define psubusb(vars, vard) mmx_m2m(psubusb, vars, vard) /* 4x16 Parallel MULs giving Low 4x16 portions of results */ #define pmullw_m2r(var, reg) mmx_m2r(pmullw, var, reg) #define pmullw_r2r(regs, regd) mmx_r2r(pmullw, regs, regd) #define pmullw(vars, vard) mmx_m2m(pmullw, vars, vard) /* 4x16 Parallel MULs giving High 4x16 portions of results */ #define pmulhw_m2r(var, reg) mmx_m2r(pmulhw, var, reg) #define pmulhw_r2r(regs, regd) mmx_r2r(pmulhw, regs, regd) #define pmulhw(vars, vard) mmx_m2m(pmulhw, vars, vard) /* 4x16->2x32 Parallel Mul-ADD (muls like pmullw, then adds adjacent 16-bit fields in the multiply result to make the final 2x32 result) */ #define pmaddwd_m2r(var, reg) mmx_m2r(pmaddwd, var, reg) #define pmaddwd_r2r(regs, regd) mmx_r2r(pmaddwd, regs, regd) #define pmaddwd(vars, vard) mmx_m2m(pmaddwd, vars, vard) /* 1x64 bitwise AND */ #ifdef BROKEN_PAND #define pand_m2r(var, reg) \ { \ mmx_m2r(pandn, (mmx_t) -1LL, reg); \ mmx_m2r(pandn, var, reg); \ } #define pand_r2r(regs, regd) \ { \ mmx_m2r(pandn, (mmx_t) -1LL, regd); \ mmx_r2r(pandn, regs, regd) \ } #define pand(vars, vard) \ { \ movq_m2r(vard, mm0); \ mmx_m2r(pandn, (mmx_t) -1LL, mm0); \ mmx_m2r(pandn, vars, mm0); \ movq_r2m(mm0, vard); \ } #else #define pand_m2r(var, reg) mmx_m2r(pand, var, reg) #define pand_r2r(regs, regd) mmx_r2r(pand, regs, regd) #define pand(vars, vard) mmx_m2m(pand, vars, vard) #endif /* 1x64 bitwise AND with Not the destination */ #define pandn_m2r(var, reg) mmx_m2r(pandn, var, reg) #define pandn_r2r(regs, regd) mmx_r2r(pandn, regs, regd) #define pandn(vars, vard) mmx_m2m(pandn, vars, vard) /* 1x64 bitwise OR */ #define por_m2r(var, reg) mmx_m2r(por, var, reg) #define por_r2r(regs, regd) mmx_r2r(por, regs, regd) #define por(vars, vard) mmx_m2m(por, vars, vard) /* 1x64 bitwise eXclusive OR */ #define pxor_m2r(var, reg) mmx_m2r(pxor, var, reg) #define pxor_r2r(regs, regd) mmx_r2r(pxor, regs, regd) #define pxor(vars, vard) mmx_m2m(pxor, vars, vard) /* 2x32, 4x16, and 8x8 Parallel CoMPare for EQuality (resulting fields are either 0 or -1) */ #define pcmpeqd_m2r(var, reg) mmx_m2r(pcmpeqd, var, reg) #define pcmpeqd_r2r(regs, regd) mmx_r2r(pcmpeqd, regs, regd) #define pcmpeqd(vars, vard) mmx_m2m(pcmpeqd, vars, vard) #define pcmpeqw_m2r(var, reg) mmx_m2r(pcmpeqw, var, reg) #define pcmpeqw_r2r(regs, regd) mmx_r2r(pcmpeqw, regs, regd) #define pcmpeqw(vars, vard) mmx_m2m(pcmpeqw, vars, vard) #define pcmpeqb_m2r(var, reg) mmx_m2r(pcmpeqb, var, reg) #define pcmpeqb_r2r(regs, regd) mmx_r2r(pcmpeqb, regs, regd) #define pcmpeqb(vars, vard) mmx_m2m(pcmpeqb, vars, vard) /* 2x32, 4x16, and 8x8 Parallel CoMPare for Greater Than (resulting fields are either 0 or -1) */ #define pcmpgtd_m2r(var, reg) mmx_m2r(pcmpgtd, var, reg) #define pcmpgtd_r2r(regs, regd) mmx_r2r(pcmpgtd, regs, regd) #define pcmpgtd(vars, vard) mmx_m2m(pcmpgtd, vars, vard) #define pcmpgtw_m2r(var, reg) mmx_m2r(pcmpgtw, var, reg) #define pcmpgtw_r2r(regs, regd) mmx_r2r(pcmpgtw, regs, regd) #define pcmpgtw(vars, vard) mmx_m2m(pcmpgtw, vars, vard) #define pcmpgtb_m2r(var, reg) mmx_m2r(pcmpgtb, var, reg) #define pcmpgtb_r2r(regs, regd) mmx_r2r(pcmpgtb, regs, regd) #define pcmpgtb(vars, vard) mmx_m2m(pcmpgtb, vars, vard) /* 1x64, 2x32, and 4x16 Parallel Shift Left Logical */ #define psllq_i2r(imm, reg) mmx_i2r(psllq, imm, reg) #define psllq_m2r(var, reg) mmx_m2r(psllq, var, reg) #define psllq_r2r(regs, regd) mmx_r2r(psllq, regs, regd) #define psllq(vars, vard) mmx_m2m(psllq, vars, vard) #define pslld_i2r(imm, reg) mmx_i2r(pslld, imm, reg) #define pslld_m2r(var, reg) mmx_m2r(pslld, var, reg) #define pslld_r2r(regs, regd) mmx_r2r(pslld, regs, regd) #define pslld(vars, vard) mmx_m2m(pslld, vars, vard) #define psllw_i2r(imm, reg) mmx_i2r(psllw, imm, reg) #define psllw_m2r(var, reg) mmx_m2r(psllw, var, reg) #define psllw_r2r(regs, regd) mmx_r2r(psllw, regs, regd) #define psllw(vars, vard) mmx_m2m(psllw, vars, vard) /* 1x64, 2x32, and 4x16 Parallel Shift Right Logical */ #define psrlq_i2r(imm, reg) mmx_i2r(psrlq, imm, reg) #define psrlq_m2r(var, reg) mmx_m2r(psrlq, var, reg) #define psrlq_r2r(regs, regd) mmx_r2r(psrlq, regs, regd) #define psrlq(vars, vard) mmx_m2m(psrlq, vars, vard) #define psrld_i2r(imm, reg) mmx_i2r(psrld, imm, reg) #define psrld_m2r(var, reg) mmx_m2r(psrld, var, reg) #define psrld_r2r(regs, regd) mmx_r2r(psrld, regs, regd) #define psrld(vars, vard) mmx_m2m(psrld, vars, vard) #define psrlw_i2r(imm, reg) mmx_i2r(psrlw, imm, reg) #define psrlw_m2r(var, reg) mmx_m2r(psrlw, var, reg) #define psrlw_r2r(regs, regd) mmx_r2r(psrlw, regs, regd) #define psrlw(vars, vard) mmx_m2m(psrlw, vars, vard) /* 2x32 and 4x16 Parallel Shift Right Arithmetic */ #define psrad_i2r(imm, reg) mmx_i2r(psrad, imm, reg) #define psrad_m2r(var, reg) mmx_m2r(psrad, var, reg) #define psrad_r2r(regs, regd) mmx_r2r(psrad, regs, regd) #define psrad(vars, vard) mmx_m2m(psrad, vars, vard) #define psraw_i2r(imm, reg) mmx_i2r(psraw, imm, reg) #define psraw_m2r(var, reg) mmx_m2r(psraw, var, reg) #define psraw_r2r(regs, regd) mmx_r2r(psraw, regs, regd) #define psraw(vars, vard) mmx_m2m(psraw, vars, vard) /* 2x32->4x16 and 4x16->8x8 PACK and Signed Saturate (packs source and dest fields into dest in that order) */ #define packssdw_m2r(var, reg) mmx_m2r(packssdw, var, reg) #define packssdw_r2r(regs, regd) mmx_r2r(packssdw, regs, regd) #define packssdw(vars, vard) mmx_m2m(packssdw, vars, vard) #define packsswb_m2r(var, reg) mmx_m2r(packsswb, var, reg) #define packsswb_r2r(regs, regd) mmx_r2r(packsswb, regs, regd) #define packsswb(vars, vard) mmx_m2m(packsswb, vars, vard) /* 4x16->8x8 PACK and Unsigned Saturate (packs source and dest fields into dest in that order) */ #define packuswb_m2r(var, reg) mmx_m2r(packuswb, var, reg) #define packuswb_r2r(regs, regd) mmx_r2r(packuswb, regs, regd) #define packuswb(vars, vard) mmx_m2m(packuswb, vars, vard) /* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK Low (interleaves low half of dest with low half of source as padding in each result field) */ #define punpckldq_m2r(var, reg) mmx_m2r(punpckldq, var, reg) #define punpckldq_r2r(regs, regd) mmx_r2r(punpckldq, regs, regd) #define punpckldq(vars, vard) mmx_m2m(punpckldq, vars, vard) #define punpcklwd_m2r(var, reg) mmx_m2r(punpcklwd, var, reg) #define punpcklwd_r2r(regs, regd) mmx_r2r(punpcklwd, regs, regd) #define punpcklwd(vars, vard) mmx_m2m(punpcklwd, vars, vard) #define punpcklbw_m2r(var, reg) mmx_m2r(punpcklbw, var, reg) #define punpcklbw_r2r(regs, regd) mmx_r2r(punpcklbw, regs, regd) #define punpcklbw(vars, vard) mmx_m2m(punpcklbw, vars, vard) /* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK High (interleaves high half of dest with high half of source as padding in each result field) */ #define punpckhdq_m2r(var, reg) mmx_m2r(punpckhdq, var, reg) #define punpckhdq_r2r(regs, regd) mmx_r2r(punpckhdq, regs, regd) #define punpckhdq(vars, vard) mmx_m2m(punpckhdq, vars, vard) #define punpckhwd_m2r(var, reg) mmx_m2r(punpckhwd, var, reg) #define punpckhwd_r2r(regs, regd) mmx_r2r(punpckhwd, regs, regd) #define punpckhwd(vars, vard) mmx_m2m(punpckhwd, vars, vard) #define punpckhbw_m2r(var, reg) mmx_m2r(punpckhbw, var, reg) #define punpckhbw_r2r(regs, regd) mmx_r2r(punpckhbw, regs, regd) #define punpckhbw(vars, vard) mmx_m2m(punpckhbw, vars, vard) /* Empty MMx State (used to clean-up when going from mmx to float use of the registers that are shared by both; note that there is no float-to-mmx operation needed, because only the float tag word info is corruptible) */ #ifdef MMX_TRACE #define emms() \ { \ printf("emms()\n"); \ __asm__ __volatile__ ("emms"); \ } #else #define emms() __asm__ __volatile__ ("emms") #endif #endif /* vi: set ts=4 sw=4 expandtab: */