aboutsummaryrefslogtreecommitdiff
path: root/zlib/contrib/asm686
diff options
context:
space:
mode:
authormarha <marha@users.sourceforge.net>2009-06-28 22:07:26 +0000
committermarha <marha@users.sourceforge.net>2009-06-28 22:07:26 +0000
commit3562e78743202e43aec8727005182a2558117eca (patch)
tree8f9113a77d12470c5c851a2a8e4cb02e89df7d43 /zlib/contrib/asm686
downloadvcxsrv-3562e78743202e43aec8727005182a2558117eca.tar.gz
vcxsrv-3562e78743202e43aec8727005182a2558117eca.tar.bz2
vcxsrv-3562e78743202e43aec8727005182a2558117eca.zip
Checked in the following released items:
xkeyboard-config-1.4.tar.gz ttf-bitstream-vera-1.10.tar.gz font-alias-1.0.1.tar.gz font-sun-misc-1.0.0.tar.gz font-sun-misc-1.0.0.tar.gz font-sony-misc-1.0.0.tar.gz font-schumacher-misc-1.0.0.tar.gz font-mutt-misc-1.0.0.tar.gz font-misc-misc-1.0.0.tar.gz font-misc-meltho-1.0.0.tar.gz font-micro-misc-1.0.0.tar.gz font-jis-misc-1.0.0.tar.gz font-isas-misc-1.0.0.tar.gz font-dec-misc-1.0.0.tar.gz font-daewoo-misc-1.0.0.tar.gz font-cursor-misc-1.0.0.tar.gz font-arabic-misc-1.0.0.tar.gz font-winitzki-cyrillic-1.0.0.tar.gz font-misc-cyrillic-1.0.0.tar.gz font-cronyx-cyrillic-1.0.0.tar.gz font-screen-cyrillic-1.0.1.tar.gz font-xfree86-type1-1.0.1.tar.gz font-adobe-utopia-type1-1.0.1.tar.gz font-ibm-type1-1.0.0.tar.gz font-bitstream-type1-1.0.0.tar.gz font-bitstream-speedo-1.0.0.tar.gz font-bh-ttf-1.0.0.tar.gz font-bh-type1-1.0.0.tar.gz font-bitstream-100dpi-1.0.0.tar.gz font-bh-lucidatypewriter-100dpi-1.0.0.tar.gz font-bh-100dpi-1.0.0.tar.gz font-adobe-utopia-100dpi-1.0.1.tar.gz font-adobe-100dpi-1.0.0.tar.gz font-util-1.0.1.tar.gz font-bitstream-75dpi-1.0.0.tar.gz font-bh-lucidatypewriter-75dpi-1.0.0.tar.gz font-adobe-utopia-75dpi-1.0.1.tar.gz font-bh-75dpi-1.0.0.tar.gz bdftopcf-1.0.1.tar.gz font-adobe-75dpi-1.0.0.tar.gz mkfontscale-1.0.6.tar.gz openssl-0.9.8k.tar.gz bigreqsproto-1.0.2.tar.gz xtrans-1.2.2.tar.gz resourceproto-1.0.2.tar.gz inputproto-1.4.4.tar.gz compositeproto-0.4.tar.gz damageproto-1.1.0.tar.gz zlib-1.2.3.tar.gz xkbcomp-1.0.5.tar.gz freetype-2.3.9.tar.gz pthreads-w32-2-8-0-release.tar.gz pixman-0.12.0.tar.gz kbproto-1.0.3.tar.gz evieext-1.0.2.tar.gz fixesproto-4.0.tar.gz recordproto-1.13.2.tar.gz randrproto-1.2.2.tar.gz scrnsaverproto-1.1.0.tar.gz renderproto-0.9.3.tar.gz xcmiscproto-1.1.2.tar.gz fontsproto-2.0.2.tar.gz xextproto-7.0.3.tar.gz xproto-7.0.14.tar.gz libXdmcp-1.0.2.tar.gz libxkbfile-1.0.5.tar.gz libfontenc-1.0.4.tar.gz libXfont-1.3.4.tar.gz libX11-1.1.5.tar.gz libXau-1.0.4.tar.gz libxcb-1.1.tar.gz xorg-server-1.5.3.tar.gz
Diffstat (limited to 'zlib/contrib/asm686')
-rw-r--r--zlib/contrib/asm686/README.68634
-rw-r--r--zlib/contrib/asm686/match.S329
2 files changed, 363 insertions, 0 deletions
diff --git a/zlib/contrib/asm686/README.686 b/zlib/contrib/asm686/README.686
new file mode 100644
index 000000000..a593f23af
--- /dev/null
+++ b/zlib/contrib/asm686/README.686
@@ -0,0 +1,34 @@
+This is a patched version of zlib, modified to use
+Pentium-Pro-optimized assembly code in the deflation algorithm. The
+files changed/added by this patch are:
+
+README.686
+match.S
+
+The speedup that this patch provides varies, depending on whether the
+compiler used to build the original version of zlib falls afoul of the
+PPro's speed traps. My own tests show a speedup of around 10-20% at
+the default compression level, and 20-30% using -9, against a version
+compiled using gcc 2.7.2.3. Your mileage may vary.
+
+Note that this code has been tailored for the PPro/PII in particular,
+and will not perform particuarly well on a Pentium.
+
+If you are using an assembler other than GNU as, you will have to
+translate match.S to use your assembler's syntax. (Have fun.)
+
+Brian Raiter
+breadbox@muppetlabs.com
+April, 1998
+
+
+Added for zlib 1.1.3:
+
+The patches come from
+http://www.muppetlabs.com/~breadbox/software/assembly.html
+
+To compile zlib with this asm file, copy match.S to the zlib directory
+then do:
+
+CFLAGS="-O3 -DASMV" ./configure
+make OBJA=match.o
diff --git a/zlib/contrib/asm686/match.S b/zlib/contrib/asm686/match.S
new file mode 100644
index 000000000..5c3e9ee36
--- /dev/null
+++ b/zlib/contrib/asm686/match.S
@@ -0,0 +1,329 @@
+/* match.s -- Pentium-Pro-optimized version of longest_match()
+ * Written for zlib 1.1.2
+ * Copyright (C) 1998 Brian Raiter <breadbox@muppetlabs.com>
+ *
+ * This is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License.
+ */
+
+#ifndef NO_UNDERLINE
+#define match_init _match_init
+#define longest_match _longest_match
+#endif
+
+#define MAX_MATCH (258)
+#define MIN_MATCH (3)
+#define MIN_LOOKAHEAD (MAX_MATCH + MIN_MATCH + 1)
+#define MAX_MATCH_8 ((MAX_MATCH + 7) & ~7)
+
+/* stack frame offsets */
+
+#define chainlenwmask 0 /* high word: current chain len */
+ /* low word: s->wmask */
+#define window 4 /* local copy of s->window */
+#define windowbestlen 8 /* s->window + bestlen */
+#define scanstart 16 /* first two bytes of string */
+#define scanend 12 /* last two bytes of string */
+#define scanalign 20 /* dword-misalignment of string */
+#define nicematch 24 /* a good enough match size */
+#define bestlen 28 /* size of best match so far */
+#define scan 32 /* ptr to string wanting match */
+
+#define LocalVarsSize (36)
+/* saved ebx 36 */
+/* saved edi 40 */
+/* saved esi 44 */
+/* saved ebp 48 */
+/* return address 52 */
+#define deflatestate 56 /* the function arguments */
+#define curmatch 60
+
+/* All the +zlib1222add offsets are due to the addition of fields
+ * in zlib in the deflate_state structure since the asm code was first written
+ * (if you compile with zlib 1.0.4 or older, use "zlib1222add equ (-4)").
+ * (if you compile with zlib between 1.0.5 and 1.2.2.1, use "zlib1222add equ 0").
+ * if you compile with zlib 1.2.2.2 or later , use "zlib1222add equ 8").
+ */
+
+#define zlib1222add (8)
+
+#define dsWSize (36+zlib1222add)
+#define dsWMask (44+zlib1222add)
+#define dsWindow (48+zlib1222add)
+#define dsPrev (56+zlib1222add)
+#define dsMatchLen (88+zlib1222add)
+#define dsPrevMatch (92+zlib1222add)
+#define dsStrStart (100+zlib1222add)
+#define dsMatchStart (104+zlib1222add)
+#define dsLookahead (108+zlib1222add)
+#define dsPrevLen (112+zlib1222add)
+#define dsMaxChainLen (116+zlib1222add)
+#define dsGoodMatch (132+zlib1222add)
+#define dsNiceMatch (136+zlib1222add)
+
+
+.file "match.S"
+
+.globl match_init, longest_match
+
+.text
+
+/* uInt longest_match(deflate_state *deflatestate, IPos curmatch) */
+
+longest_match:
+
+/* Save registers that the compiler may be using, and adjust %esp to */
+/* make room for our stack frame. */
+
+ pushl %ebp
+ pushl %edi
+ pushl %esi
+ pushl %ebx
+ subl $LocalVarsSize, %esp
+
+/* Retrieve the function arguments. %ecx will hold cur_match */
+/* throughout the entire function. %edx will hold the pointer to the */
+/* deflate_state structure during the function's setup (before */
+/* entering the main loop). */
+
+ movl deflatestate(%esp), %edx
+ movl curmatch(%esp), %ecx
+
+/* uInt wmask = s->w_mask; */
+/* unsigned chain_length = s->max_chain_length; */
+/* if (s->prev_length >= s->good_match) { */
+/* chain_length >>= 2; */
+/* } */
+
+ movl dsPrevLen(%edx), %eax
+ movl dsGoodMatch(%edx), %ebx
+ cmpl %ebx, %eax
+ movl dsWMask(%edx), %eax
+ movl dsMaxChainLen(%edx), %ebx
+ jl LastMatchGood
+ shrl $2, %ebx
+LastMatchGood:
+
+/* chainlen is decremented once beforehand so that the function can */
+/* use the sign flag instead of the zero flag for the exit test. */
+/* It is then shifted into the high word, to make room for the wmask */
+/* value, which it will always accompany. */
+
+ decl %ebx
+ shll $16, %ebx
+ orl %eax, %ebx
+ movl %ebx, chainlenwmask(%esp)
+
+/* if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead; */
+
+ movl dsNiceMatch(%edx), %eax
+ movl dsLookahead(%edx), %ebx
+ cmpl %eax, %ebx
+ jl LookaheadLess
+ movl %eax, %ebx
+LookaheadLess: movl %ebx, nicematch(%esp)
+
+/* register Bytef *scan = s->window + s->strstart; */
+
+ movl dsWindow(%edx), %esi
+ movl %esi, window(%esp)
+ movl dsStrStart(%edx), %ebp
+ lea (%esi,%ebp), %edi
+ movl %edi, scan(%esp)
+
+/* Determine how many bytes the scan ptr is off from being */
+/* dword-aligned. */
+
+ movl %edi, %eax
+ negl %eax
+ andl $3, %eax
+ movl %eax, scanalign(%esp)
+
+/* IPos limit = s->strstart > (IPos)MAX_DIST(s) ? */
+/* s->strstart - (IPos)MAX_DIST(s) : NIL; */
+
+ movl dsWSize(%edx), %eax
+ subl $MIN_LOOKAHEAD, %eax
+ subl %eax, %ebp
+ jg LimitPositive
+ xorl %ebp, %ebp
+LimitPositive:
+
+/* int best_len = s->prev_length; */
+
+ movl dsPrevLen(%edx), %eax
+ movl %eax, bestlen(%esp)
+
+/* Store the sum of s->window + best_len in %esi locally, and in %esi. */
+
+ addl %eax, %esi
+ movl %esi, windowbestlen(%esp)
+
+/* register ush scan_start = *(ushf*)scan; */
+/* register ush scan_end = *(ushf*)(scan+best_len-1); */
+/* Posf *prev = s->prev; */
+
+ movzwl (%edi), %ebx
+ movl %ebx, scanstart(%esp)
+ movzwl -1(%edi,%eax), %ebx
+ movl %ebx, scanend(%esp)
+ movl dsPrev(%edx), %edi
+
+/* Jump into the main loop. */
+
+ movl chainlenwmask(%esp), %edx
+ jmp LoopEntry
+
+.balign 16
+
+/* do {
+ * match = s->window + cur_match;
+ * if (*(ushf*)(match+best_len-1) != scan_end ||
+ * *(ushf*)match != scan_start) continue;
+ * [...]
+ * } while ((cur_match = prev[cur_match & wmask]) > limit
+ * && --chain_length != 0);
+ *
+ * Here is the inner loop of the function. The function will spend the
+ * majority of its time in this loop, and majority of that time will
+ * be spent in the first ten instructions.
+ *
+ * Within this loop:
+ * %ebx = scanend
+ * %ecx = curmatch
+ * %edx = chainlenwmask - i.e., ((chainlen << 16) | wmask)
+ * %esi = windowbestlen - i.e., (window + bestlen)
+ * %edi = prev
+ * %ebp = limit
+ */
+LookupLoop:
+ andl %edx, %ecx
+ movzwl (%edi,%ecx,2), %ecx
+ cmpl %ebp, %ecx
+ jbe LeaveNow
+ subl $0x00010000, %edx
+ js LeaveNow
+LoopEntry: movzwl -1(%esi,%ecx), %eax
+ cmpl %ebx, %eax
+ jnz LookupLoop
+ movl window(%esp), %eax
+ movzwl (%eax,%ecx), %eax
+ cmpl scanstart(%esp), %eax
+ jnz LookupLoop
+
+/* Store the current value of chainlen. */
+
+ movl %edx, chainlenwmask(%esp)
+
+/* Point %edi to the string under scrutiny, and %esi to the string we */
+/* are hoping to match it up with. In actuality, %esi and %edi are */
+/* both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and %edx is */
+/* initialized to -(MAX_MATCH_8 - scanalign). */
+
+ movl window(%esp), %esi
+ movl scan(%esp), %edi
+ addl %ecx, %esi
+ movl scanalign(%esp), %eax
+ movl $(-MAX_MATCH_8), %edx
+ lea MAX_MATCH_8(%edi,%eax), %edi
+ lea MAX_MATCH_8(%esi,%eax), %esi
+
+/* Test the strings for equality, 8 bytes at a time. At the end,
+ * adjust %edx so that it is offset to the exact byte that mismatched.
+ *
+ * We already know at this point that the first three bytes of the
+ * strings match each other, and they can be safely passed over before
+ * starting the compare loop. So what this code does is skip over 0-3
+ * bytes, as much as necessary in order to dword-align the %edi
+ * pointer. (%esi will still be misaligned three times out of four.)
+ *
+ * It should be confessed that this loop usually does not represent
+ * much of the total running time. Replacing it with a more
+ * straightforward "rep cmpsb" would not drastically degrade
+ * performance.
+ */
+LoopCmps:
+ movl (%esi,%edx), %eax
+ xorl (%edi,%edx), %eax
+ jnz LeaveLoopCmps
+ movl 4(%esi,%edx), %eax
+ xorl 4(%edi,%edx), %eax
+ jnz LeaveLoopCmps4
+ addl $8, %edx
+ jnz LoopCmps
+ jmp LenMaximum
+LeaveLoopCmps4: addl $4, %edx
+LeaveLoopCmps: testl $0x0000FFFF, %eax
+ jnz LenLower
+ addl $2, %edx
+ shrl $16, %eax
+LenLower: subb $1, %al
+ adcl $0, %edx
+
+/* Calculate the length of the match. If it is longer than MAX_MATCH, */
+/* then automatically accept it as the best possible match and leave. */
+
+ lea (%edi,%edx), %eax
+ movl scan(%esp), %edi
+ subl %edi, %eax
+ cmpl $MAX_MATCH, %eax
+ jge LenMaximum
+
+/* If the length of the match is not longer than the best match we */
+/* have so far, then forget it and return to the lookup loop. */
+
+ movl deflatestate(%esp), %edx
+ movl bestlen(%esp), %ebx
+ cmpl %ebx, %eax
+ jg LongerMatch
+ movl windowbestlen(%esp), %esi
+ movl dsPrev(%edx), %edi
+ movl scanend(%esp), %ebx
+ movl chainlenwmask(%esp), %edx
+ jmp LookupLoop
+
+/* s->match_start = cur_match; */
+/* best_len = len; */
+/* if (len >= nice_match) break; */
+/* scan_end = *(ushf*)(scan+best_len-1); */
+
+LongerMatch: movl nicematch(%esp), %ebx
+ movl %eax, bestlen(%esp)
+ movl %ecx, dsMatchStart(%edx)
+ cmpl %ebx, %eax
+ jge LeaveNow
+ movl window(%esp), %esi
+ addl %eax, %esi
+ movl %esi, windowbestlen(%esp)
+ movzwl -1(%edi,%eax), %ebx
+ movl dsPrev(%edx), %edi
+ movl %ebx, scanend(%esp)
+ movl chainlenwmask(%esp), %edx
+ jmp LookupLoop
+
+/* Accept the current string, with the maximum possible length. */
+
+LenMaximum: movl deflatestate(%esp), %edx
+ movl $MAX_MATCH, bestlen(%esp)
+ movl %ecx, dsMatchStart(%edx)
+
+/* if ((uInt)best_len <= s->lookahead) return (uInt)best_len; */
+/* return s->lookahead; */
+
+LeaveNow:
+ movl deflatestate(%esp), %edx
+ movl bestlen(%esp), %ebx
+ movl dsLookahead(%edx), %eax
+ cmpl %eax, %ebx
+ jg LookaheadRet
+ movl %ebx, %eax
+LookaheadRet:
+
+/* Restore the stack and return from whence we came. */
+
+ addl $LocalVarsSize, %esp
+ popl %ebx
+ popl %esi
+ popl %edi
+ popl %ebp
+match_init: ret