path: root/runtime/zlib/contrib/inflate86/inffast.S

/*
 * inffast.S is a hand tuned assembler version of:
 *
 * inffast.c -- fast decoding
 * Copyright (C) 1995-2003 Mark Adler
 * For conditions of distribution and use, see copyright notice in zlib.h
 *
 * Copyright (C) 2003 Chris Anderson <christop@charm.net>
 * Please use the copyright conditions above.
 *
 * This version (Jan-23-2003) of inflate_fast was coded and tested under
 * GNU/Linux on a pentium 3, using the gcc-3.2 compiler distribution.  On that
 * machine, I found that gzip style archives decompressed about 20% faster than
 * the gcc-3.2 -O3 -fomit-frame-pointer compiled version.  Your results will
 * depend on how large of a buffer is used for z_stream.next_in & next_out
 * (8K-32K worked best for my 256K cpu cache) and how much overhead there is in
 * stream processing I/O and crc32/addler32.  In my case, this routine used
 * 70% of the cpu time and crc32 used 20%.
 *
 * I am confident that this version will work in the general case, but I have
 * not tested a wide variety of datasets or a wide variety of platforms.
 *
 * Jan-24-2003 -- Added -DUSE_MMX define for slightly faster inflating.
 * It should be a runtime flag instead of compile time flag...
 *
 * Jan-26-2003 -- Added runtime check for MMX support with cpuid instruction.
 * With -DUSE_MMX, only MMX code is compiled.  With -DNO_MMX, only non-MMX code
 * is compiled.  Without either option, runtime detection is enabled.  Runtime
 * detection should work on all modern cpus and the recomended algorithm (flip
 * ID bit on eflags and then use the cpuid instruction) is used in many
 * multimedia applications.  Tested under win2k with gcc-2.95 and gas-2.12
 * distributed with cygwin3.  Compiling with gcc-2.95 -c inffast.S -o
 * inffast.obj generates a COFF object which can then be linked with MSVC++
 * compiled code.  Tested under FreeBSD 4.7 with gcc-2.95.
 *
 * Jan-28-2003 -- Tested Athlon XP... MMX mode is slower than no MMX (and
 * slower than compiler generated code).  Adjusted cpuid check to use the MMX
 * code only for Pentiums < P4 until I have more data on the P4.  Speed
 * improvment is only about 15% on the Athlon when compared with code generated
 * with MSVC++.  Not sure yet, but I think the P4 will also be slower using the
 * MMX mode because many of it's x86 ALU instructions execute in .5 cycles and
 * have less latency than MMX ops.  Added code to buffer the last 11 bytes of
 * the input stream since the MMX code grabs bits in chunks of 32, which
 * differs from the inffast.c algorithm.  I don't think there would have been
 * read overruns where a page boundary was crossed (a segfault), but there
 * could have been overruns when next_in ends on unaligned memory (unintialized
 * memory read).
 *
 * Mar-13-2003 -- P4 MMX is slightly slower than P4 NO_MMX.  I created a C
 * version of the non-MMX code so that it doesn't depend on zstrm and zstate
 * structure offsets which are hard coded in this file.  This was last tested
 * with zlib-1.2.0 which is currently in beta testing, newer versions of this
 * and inffas86.c can be found at http://www.eetbeetee.com/zlib/ and
 * http://www.charm.net/~christop/zlib/
 */


/*
 * if you have underscore linking problems (_inflate_fast undefined), try
 * using -DGAS_COFF
 */
#if ! defined( GAS_COFF ) && ! defined( GAS_ELF )

#if defined( WIN32 ) || defined( __CYGWIN__ )
#define GAS_COFF /* windows object format */
#else
#define GAS_ELF
#endif

#endif /* ! GAS_COFF && ! GAS_ELF */


#if defined( GAS_COFF )

/* coff externals have underscores */
#define inflate_fast _inflate_fast
#define inflate_fast_use_mmx _inflate_fast_use_mmx

#endif /* GAS_COFF */


.file "inffast.S"

.globl inflate_fast

.text
.align 4,0
.L_invalid_literal_length_code_msg:
.string "invalid literal/length code"

.align 4,0
.L_invalid_distance_code_msg:
.string "invalid distance code"

.align 4,0
.L_invalid_distance_too_far_msg:
.string "invalid distance too far back"

#if ! defined( NO_MMX )
.align 4,0
.L_mask: /* mask[N] = ( 1 << N ) - 1 */
.long 0
.long 1
.long 3
.long 7
.long 15
.long 31
.long 63
.long 127
.long 255