summaryrefslogtreecommitdiff
path: root/src/lzo1x_c.ch
diff options
context:
space:
mode:
Diffstat (limited to 'src/lzo1x_c.ch')
-rw-r--r--src/lzo1x_c.ch430
1 files changed, 242 insertions, 188 deletions
diff --git a/src/lzo1x_c.ch b/src/lzo1x_c.ch
index 08d615e..805f71d 100644
--- a/src/lzo1x_c.ch
+++ b/src/lzo1x_c.ch
@@ -2,19 +2,7 @@
This file is part of the LZO real-time data compression library.
- Copyright (C) 2008 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2007 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2006 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2005 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2004 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2003 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2002 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2001 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 2000 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 1999 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 1998 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 1997 Markus Franz Xaver Johannes Oberhumer
- Copyright (C) 1996 Markus Franz Xaver Johannes Oberhumer
+ Copyright (C) 1996-2014 Markus Franz Xaver Johannes Oberhumer
All Rights Reserved.
The LZO library is free software; you can redistribute it and/or
@@ -39,6 +27,12 @@
+#if 1 && defined(DO_COMPRESS) && !defined(do_compress)
+ /* choose a unique name to better help PGO optimizations */
+# define do_compress LZO_PP_ECONCAT2(DO_COMPRESS,_core)
+#endif
+
+
/***********************************************************************
// compress a block of data.
************************************************************************/
@@ -46,12 +40,12 @@
static __lzo_noinline lzo_uint
do_compress ( const lzo_bytep in , lzo_uint in_len,
lzo_bytep out, lzo_uintp out_len,
- lzo_voidp wrkmem )
+ lzo_uint ti, lzo_voidp wrkmem)
{
- register const lzo_bytep ip;
+ const lzo_bytep ip;
lzo_bytep op;
const lzo_bytep const in_end = in + in_len;
- const lzo_bytep const ip_end = in + in_len - M2_MAX_LEN - 5;
+ const lzo_bytep const ip_end = in + in_len - 20;
const lzo_bytep ii;
lzo_dict_p const dict = (lzo_dict_p) wrkmem;
@@ -59,14 +53,17 @@ do_compress ( const lzo_bytep in , lzo_uint in_len,
ip = in;
ii = ip;
- ip += 4;
+ ip += ti < 4 ? 4 - ti : 0;
for (;;)
{
- register const lzo_bytep m_pos;
- lzo_uint m_off;
+ const lzo_bytep m_pos;
+#if !(LZO_DETERMINISTIC)
+ LZO_DEFINE_UNINITIALIZED_VAR(lzo_uint, m_off, 0);
lzo_uint m_len;
lzo_uint dindex;
-
+next:
+ if __lzo_unlikely(ip >= ip_end)
+ break;
DINDEX1(dindex,ip);
GINDEX(m_pos,m_off,dict,dindex,in);
if (LZO_CHECK_MPOS_NON_DET(m_pos,m_off,in,ip,M4_MAX_OFFSET))
@@ -83,210 +80,252 @@ do_compress ( const lzo_bytep in , lzo_uint in_len,
goto try_match;
goto literal;
-
try_match:
-#if 1 && defined(LZO_UNALIGNED_OK_2)
- if (* (const lzo_ushortp) m_pos != * (const lzo_ushortp) ip)
+#if (LZO_OPT_UNALIGNED32)
+ if (UA_GET_NE32(m_pos) != UA_GET_NE32(ip))
#else
- if (m_pos[0] != ip[0] || m_pos[1] != ip[1])
+ if (m_pos[0] != ip[0] || m_pos[1] != ip[1] || m_pos[2] != ip[2] || m_pos[3] != ip[3])
#endif
{
+ /* a literal */
+literal:
+ UPDATE_I(dict,0,dindex,ip,in);
+ ip += 1 + ((ip - ii) >> 5);
+ continue;
}
- else
- {
- if __lzo_likely(m_pos[2] == ip[2])
- {
-#if 0
- if (m_off <= M2_MAX_OFFSET)
- goto match;
- if (lit <= 3)
- goto match;
- if (lit == 3) /* better compression, but slower */
- {
- assert(op - 2 > out); op[-2] |= LZO_BYTE(3);
- *op++ = *ii++; *op++ = *ii++; *op++ = *ii++;
- goto code_match;
- }
- if (m_pos[3] == ip[3])
-#endif
- goto match;
- }
- else
- {
- /* still need a better way for finding M1 matches */
-#if 0
- /* a M1 match */
-#if 0
- if (m_off <= M1_MAX_OFFSET && lit > 0 && lit <= 3)
+/*match:*/
+ UPDATE_I(dict,0,dindex,ip,in);
#else
- if (m_off <= M1_MAX_OFFSET && lit == 3)
-#endif
- {
- register lzo_uint t;
-
- t = lit;
- assert(op - 2 > out); op[-2] |= LZO_BYTE(t);
- do *op++ = *ii++; while (--t > 0);
- assert(ii == ip);
- m_off -= 1;
- *op++ = LZO_BYTE(M1_MARKER | ((m_off & 3) << 2));
- *op++ = LZO_BYTE(m_off >> 2);
- ip += 2;
- goto match_done;
- }
-#endif
- }
- }
-
-
- /* a literal */
+ lzo_uint m_off;
+ lzo_uint m_len;
+ {
+ lzo_uint32_t dv;
+ lzo_uint dindex;
literal:
- UPDATE_I(dict,0,dindex,ip,in);
- ++ip;
+ ip += 1 + ((ip - ii) >> 5);
+next:
if __lzo_unlikely(ip >= ip_end)
break;
- continue;
-
+ dv = UA_GET_LE32(ip);
+ dindex = DINDEX(dv,ip);
+ GINDEX(m_off,m_pos,in+dict,dindex,in);
+ UPDATE_I(dict,0,dindex,ip,in);
+ if __lzo_unlikely(dv != UA_GET_LE32(m_pos))
+ goto literal;
+ }
+#endif
/* a match */
-match:
- UPDATE_I(dict,0,dindex,ip,in);
- /* store current literal run */
- if (pd(ip,ii) > 0)
- {
- register lzo_uint t = pd(ip,ii);
+ ii -= ti; ti = 0;
+ {
+ lzo_uint t = pd(ip,ii);
+ if (t != 0)
+ {
if (t <= 3)
{
- assert(op - 2 > out);
- op[-2] |= LZO_BYTE(t);
+ op[-2] = LZO_BYTE(op[-2] | t);
+#if (LZO_OPT_UNALIGNED32)
+ UA_COPY4(op, ii);
+ op += t;
+#else
+ { do *op++ = *ii++; while (--t > 0); }
+#endif
}
- else if (t <= 18)
+#if (LZO_OPT_UNALIGNED32) || (LZO_OPT_UNALIGNED64)
+ else if (t <= 16)
+ {
*op++ = LZO_BYTE(t - 3);
+ UA_COPY8(op, ii);
+ UA_COPY8(op+8, ii+8);
+ op += t;
+ }
+#endif
else
{
- register lzo_uint tt = t - 18;
-
- *op++ = 0;
- while (tt > 255)
+ if (t <= 18)
+ *op++ = LZO_BYTE(t - 3);
+ else
{
- tt -= 255;
+ lzo_uint tt = t - 18;
*op++ = 0;
+ while __lzo_unlikely(tt > 255)
+ {
+ tt -= 255;
+ UA_SET1(op, 0);
+ op++;
+ }
+ assert(tt > 0);
+ *op++ = LZO_BYTE(tt);
}
- assert(tt > 0);
- *op++ = LZO_BYTE(tt);
+#if (LZO_OPT_UNALIGNED32) || (LZO_OPT_UNALIGNED64)
+ do {
+ UA_COPY8(op, ii);
+ UA_COPY8(op+8, ii+8);
+ op += 16; ii += 16; t -= 16;
+ } while (t >= 16); if (t > 0)
+#endif
+ { do *op++ = *ii++; while (--t > 0); }
}
- do *op++ = *ii++; while (--t > 0);
}
-
- /* code the match */
- assert(ii == ip);
- ip += 3;
- if (m_pos[3] != *ip++ || m_pos[4] != *ip++ || m_pos[5] != *ip++ ||
- m_pos[6] != *ip++ || m_pos[7] != *ip++ || m_pos[8] != *ip++
-#ifdef LZO1Y
- || m_pos[ 9] != *ip++ || m_pos[10] != *ip++ || m_pos[11] != *ip++
- || m_pos[12] != *ip++ || m_pos[13] != *ip++ || m_pos[14] != *ip++
+ }
+ m_len = 4;
+ {
+#if (LZO_OPT_UNALIGNED64)
+ lzo_uint64_t v;
+ v = UA_GET_NE64(ip + m_len) ^ UA_GET_NE64(m_pos + m_len);
+ if __lzo_unlikely(v == 0) {
+ do {
+ m_len += 8;
+ v = UA_GET_NE64(ip + m_len) ^ UA_GET_NE64(m_pos + m_len);
+ if __lzo_unlikely(ip + m_len >= ip_end)
+ goto m_len_done;
+ } while (v == 0);
+ }
+#if (LZO_ABI_BIG_ENDIAN) && defined(lzo_bitops_ctlz64)
+ m_len += lzo_bitops_ctlz64(v) / CHAR_BIT;
+#elif (LZO_ABI_BIG_ENDIAN)
+ if ((v >> (64 - CHAR_BIT)) == 0) do {
+ v <<= CHAR_BIT;
+ m_len += 1;
+ } while ((v >> (64 - CHAR_BIT)) == 0);
+#elif (LZO_ABI_LITTLE_ENDIAN) && defined(lzo_bitops_cttz64)
+ m_len += lzo_bitops_cttz64(v) / CHAR_BIT;
+#elif (LZO_ABI_LITTLE_ENDIAN)
+ if ((v & UCHAR_MAX) == 0) do {
+ v >>= CHAR_BIT;
+ m_len += 1;
+ } while ((v & UCHAR_MAX) == 0);
+#else
+ if (ip[m_len] == m_pos[m_len]) do {
+ m_len += 1;
+ } while (ip[m_len] == m_pos[m_len]);
+#endif
+#elif (LZO_OPT_UNALIGNED32)
+ lzo_uint32_t v;
+ v = UA_GET_NE32(ip + m_len) ^ UA_GET_NE32(m_pos + m_len);
+ if __lzo_unlikely(v == 0) {
+ do {
+ m_len += 4;
+ v = UA_GET_NE32(ip + m_len) ^ UA_GET_NE32(m_pos + m_len);
+ if (v != 0)
+ break;
+ m_len += 4;
+ v = UA_GET_NE32(ip + m_len) ^ UA_GET_NE32(m_pos + m_len);
+ if __lzo_unlikely(ip + m_len >= ip_end)
+ goto m_len_done;
+ } while (v == 0);
+ }
+#if (LZO_ABI_BIG_ENDIAN) && defined(lzo_bitops_ctlz32)
+ m_len += lzo_bitops_ctlz32(v) / CHAR_BIT;
+#elif (LZO_ABI_BIG_ENDIAN)
+ if ((v >> (32 - CHAR_BIT)) == 0) do {
+ v <<= CHAR_BIT;
+ m_len += 1;
+ } while ((v >> (32 - CHAR_BIT)) == 0);
+#elif (LZO_ABI_LITTLE_ENDIAN) && defined(lzo_bitops_cttz32)
+ m_len += lzo_bitops_cttz32(v) / CHAR_BIT;
+#elif (LZO_ABI_LITTLE_ENDIAN)
+ if ((v & UCHAR_MAX) == 0) do {
+ v >>= CHAR_BIT;
+ m_len += 1;
+ } while ((v & UCHAR_MAX) == 0);
+#else
+ if (ip[m_len] == m_pos[m_len]) do {
+ m_len += 1;
+ } while (ip[m_len] == m_pos[m_len]);
+#endif
+#else
+ if __lzo_unlikely(ip[m_len] == m_pos[m_len]) {
+ do {
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if (ip[m_len] != m_pos[m_len])
+ break;
+ m_len += 1;
+ if __lzo_unlikely(ip + m_len >= ip_end)
+ goto m_len_done;
+ } while (ip[m_len] == m_pos[m_len]);
+ }
#endif
- )
+ }
+m_len_done:
+ m_off = pd(ip,m_pos);
+ ip += m_len;
+ ii = ip;
+ if (m_len <= M2_MAX_LEN && m_off <= M2_MAX_OFFSET)
{
- --ip;
- m_len = pd(ip, ii);
- assert(m_len >= 3); assert(m_len <= M2_MAX_LEN);
-
- if (m_off <= M2_MAX_OFFSET)
- {
- m_off -= 1;
+ m_off -= 1;
#if defined(LZO1X)
- *op++ = LZO_BYTE(((m_len - 1) << 5) | ((m_off & 7) << 2));
- *op++ = LZO_BYTE(m_off >> 3);
+ *op++ = LZO_BYTE(((m_len - 1) << 5) | ((m_off & 7) << 2));
+ *op++ = LZO_BYTE(m_off >> 3);
#elif defined(LZO1Y)
- *op++ = LZO_BYTE(((m_len + 1) << 4) | ((m_off & 3) << 2));
- *op++ = LZO_BYTE(m_off >> 2);
+ *op++ = LZO_BYTE(((m_len + 1) << 4) | ((m_off & 3) << 2));
+ *op++ = LZO_BYTE(m_off >> 2);
#endif
- }
- else if (m_off <= M3_MAX_OFFSET)
- {
- m_off -= 1;
+ }
+ else if (m_off <= M3_MAX_OFFSET)
+ {
+ m_off -= 1;
+ if (m_len <= M3_MAX_LEN)
*op++ = LZO_BYTE(M3_MARKER | (m_len - 2));
- goto m3_m4_offset;
- }
else
-#if defined(LZO1X)
{
- m_off -= 0x4000;
- assert(m_off > 0); assert(m_off <= 0x7fff);
- *op++ = LZO_BYTE(M4_MARKER |
- ((m_off & 0x4000) >> 11) | (m_len - 2));
- goto m3_m4_offset;
+ m_len -= M3_MAX_LEN;
+ *op++ = M3_MARKER | 0;
+ while __lzo_unlikely(m_len > 255)
+ {
+ m_len -= 255;
+ UA_SET1(op, 0);
+ op++;
+ }
+ *op++ = LZO_BYTE(m_len);
}
-#elif defined(LZO1Y)
- goto m4_match;
-#endif
+ *op++ = LZO_BYTE(m_off << 2);
+ *op++ = LZO_BYTE(m_off >> 6);
}
else
{
- {
- const lzo_bytep end = in_end;
- const lzo_bytep m = m_pos + M2_MAX_LEN + 1;
- while (ip < end && *m == *ip)
- m++, ip++;
- m_len = pd(ip, ii);
- }
- assert(m_len > M2_MAX_LEN);
-
- if (m_off <= M3_MAX_OFFSET)
- {
- m_off -= 1;
- if (m_len <= 33)
- *op++ = LZO_BYTE(M3_MARKER | (m_len - 2));
- else
- {
- m_len -= 33;
- *op++ = M3_MARKER | 0;
- goto m3_m4_len;
- }
- }
+ m_off -= 0x4000;
+ if (m_len <= M4_MAX_LEN)
+ *op++ = LZO_BYTE(M4_MARKER | ((m_off >> 11) & 8) | (m_len - 2));
else
{
-#if defined(LZO1Y)
-m4_match:
-#endif
- m_off -= 0x4000;
- assert(m_off > 0); assert(m_off <= 0x7fff);
- if (m_len <= M4_MAX_LEN)
- *op++ = LZO_BYTE(M4_MARKER |
- ((m_off & 0x4000) >> 11) | (m_len - 2));
- else
+ m_len -= M4_MAX_LEN;
+ *op++ = LZO_BYTE(M4_MARKER | ((m_off >> 11) & 8));
+ while __lzo_unlikely(m_len > 255)
{
- m_len -= M4_MAX_LEN;
- *op++ = LZO_BYTE(M4_MARKER | ((m_off & 0x4000) >> 11));
-m3_m4_len:
- while (m_len > 255)
- {
- m_len -= 255;
- *op++ = 0;
- }
- assert(m_len > 0);
- *op++ = LZO_BYTE(m_len);
+ m_len -= 255;
+ UA_SET1(op, 0);
+ op++;
}
+ *op++ = LZO_BYTE(m_len);
}
-
-m3_m4_offset:
- *op++ = LZO_BYTE((m_off & 63) << 2);
+ *op++ = LZO_BYTE(m_off << 2);
*op++ = LZO_BYTE(m_off >> 6);
}
-
-#if 0
-match_done:
-#endif
- ii = ip;
- if __lzo_unlikely(ip >= ip_end)
- break;
+ goto next;
}
*out_len = pd(op, out);
- return pd(in_end,ii);
+ return pd(in_end,ii-ti);
}
@@ -299,16 +338,30 @@ DO_COMPRESS ( const lzo_bytep in , lzo_uint in_len,
lzo_bytep out, lzo_uintp out_len,
lzo_voidp wrkmem )
{
+ const lzo_bytep ip = in;
lzo_bytep op = out;
- lzo_uint t;
+ lzo_uint l = in_len;
+ lzo_uint t = 0;
- if __lzo_unlikely(in_len <= M2_MAX_LEN + 5)
- t = in_len;
- else
+ while (l > 20)
{
- t = do_compress(in,in_len,op,out_len,wrkmem);
+ lzo_uint ll = l;
+ lzo_uintptr_t ll_end;
+#if 0 || (LZO_DETERMINISTIC)
+ ll = LZO_MIN(ll, 49152);
+#endif
+ ll_end = (lzo_uintptr_t)ip + ll;
+ if ((ll_end + ((t + ll) >> 5)) <= ll_end || (const lzo_bytep)(ll_end + ((t + ll) >> 5)) <= ip + ll)
+ break;
+#if (LZO_DETERMINISTIC)
+ lzo_memset(wrkmem, 0, ((lzo_uint)1 << D_BITS) * sizeof(lzo_dict_t));
+#endif
+ t = do_compress(ip,ll,op,out_len,t,wrkmem);
+ ip += ll;
op += *out_len;
+ l -= ll;
}
+ t += l;
if (t > 0)
{
@@ -317,7 +370,7 @@ DO_COMPRESS ( const lzo_bytep in , lzo_uint in_len,
if (op == out && t <= 238)
*op++ = LZO_BYTE(17 + t);
else if (t <= 3)
- op[-2] |= LZO_BYTE(t);
+ op[-2] = LZO_BYTE(op[-2] | t);
else if (t <= 18)
*op++ = LZO_BYTE(t - 3);
else
@@ -328,12 +381,14 @@ DO_COMPRESS ( const lzo_bytep in , lzo_uint in_len,
while (tt > 255)
{
tt -= 255;
- *op++ = 0;
+ UA_SET1(op, 0);
+ op++;
}
assert(tt > 0);
*op++ = LZO_BYTE(tt);
}
- do *op++ = *ii++; while (--t > 0);
+ UA_COPYN(op, ii, t);
+ op += t;
}
*op++ = M4_MARKER | 1;
@@ -348,4 +403,3 @@ DO_COMPRESS ( const lzo_bytep in , lzo_uint in_len,
/*
vi:ts=4:et
*/
-
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-29 18:42:34 +0000