/***********************************************************************
// compress a block of data.
************************************************************************/
static NOINLINE unsigned
do_compress(const uint8_t* in, unsigned in_len,
uint8_t* out, unsigned* out_len,
void* wrkmem)
{
register const uint8_t* ip;
uint8_t* op;
const uint8_t* const in_end = in + in_len;
const uint8_t* const ip_end = in + in_len - M2_MAX_LEN - 5;
const uint8_t* ii;
const void* *const dict = (const void**) wrkmem;
op = out;
ip = in;
ii = ip;
ip += 4;
for (;;) {
register const uint8_t* m_pos;
unsigned m_off;
unsigned m_len;
unsigned dindex;
D_INDEX1(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))
goto literal;
#if 1
if (m_off <= M2_MAX_OFFSET || m_pos[3] == ip[3])
goto try_match;
D_INDEX2(dindex,ip);
#endif
GINDEX(m_pos,m_off,dict,dindex,in);
if (LZO_CHECK_MPOS_NON_DET(m_pos,m_off,in,ip,M4_MAX_OFFSET))
goto literal;
if (m_off <= M2_MAX_OFFSET || m_pos[3] == ip[3])
goto try_match;
goto literal;
try_match:
#if 1 && defined(LZO_UNALIGNED_OK_2)
if (* (const lzo_ushortp) m_pos != * (const lzo_ushortp) ip)
#else
if (m_pos[0] != ip[0] || m_pos[1] != ip[1])
#endif
{
} else {
if (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] |= (uint8_t)(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)
#else
if (m_off <= M1_MAX_OFFSET && lit == 3)
#endif
{
register unsigned t;
t = lit;
assert(op - 2 > out);
op[-2] |= (uint8_t)(t);
do *op++ = *ii++;
while (--t > 0);
assert(ii == ip);
m_off -= 1;
*op++ = (uint8_t)(M1_MARKER | ((m_off & 3) << 2));
*op++ = (uint8_t)(m_off >> 2);
ip += 2;
goto match_done;
}
#endif
}
}
/* a literal */
literal:
UPDATE_I(dict, 0, dindex, ip, in);
++ip;
if (ip >= ip_end)
break;
continue;
/* a match */
match:
UPDATE_I(dict, 0, dindex, ip, in);
/* store current literal run */
if (pd(ip, ii) > 0) {
register unsigned t = pd(ip, ii);
if (t <= 3) {
assert(op - 2 > out);
op[-2] |= (uint8_t)(t);
}
else if (t <= 18)
*op++ = (uint8_t)(t - 3);
else {
register unsigned tt = t - 18;
*op++ = 0;
while (tt > 255) {
tt -= 255;
*op++ = 0;
}
assert(tt > 0);
*op++ = (uint8_t)(tt);
}
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++
#endif
) {
--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;
#if defined(LZO1X)
*op++ = (uint8_t)(((m_len - 1) << 5) | ((m_off & 7) << 2));
*op++ = (uint8_t)(m_off >> 3);
#elif defined(LZO1Y)
*op++ = (uint8_t)(((m_len + 1) << 4) | ((m_off & 3) << 2));
*op++ = (uint8_t)(m_off >> 2);
#endif
}
else if (m_off <= M3_MAX_OFFSET) {
static __lzo_noinline
int do_compress ( const lzo_bytep in , lzo_uint in_len,
lzo_bytep out, lzo_uintp out_len,
lzo_voidp wrkmem )
{
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 - 9;
const lzo_bytep ii;
lzo_dict_p const dict = (lzo_dict_p) wrkmem;
op = out;
ip = in;
ii = ip;
ip++;
for (;;)
{
const lzo_bytep m_pos;
LZO_DEFINE_UNINITIALIZED_VAR(lzo_uint, m_off, 0);
lzo_uint m_len;
lzo_uint dindex;
lzo_uint lit;
DINDEX1(dindex,ip);
GINDEX(m_pos,m_off,dict,dindex,in);
if (LZO_CHECK_MPOS_NON_DET(m_pos,m_off,in,ip,M3_MAX_OFFSET))
goto literal;
#if 1
if (m_off <= M2_MAX_OFFSET || m_pos[3] == ip[3])
goto try_match;
DINDEX2(dindex,ip);
#endif
GINDEX(m_pos,m_off,dict,dindex,in);
if (LZO_CHECK_MPOS_NON_DET(m_pos,m_off,in,ip,M3_MAX_OFFSET))
goto literal;
if (m_off <= M2_MAX_OFFSET || m_pos[3] == ip[3])
goto try_match;
goto literal;
try_match:
#if 0 && (LZO_OPT_UNALIGNED16)
if (UA_GET_NE16(m_pos) != UA_GET_NE16(ip))
#else
if (m_pos[0] != ip[0] || m_pos[1] != ip[1])
#endif
{
}
else
{
if (m_pos[2] == ip[2])
{
m_pos += 3;
#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 == ip[3])
#endif
goto match;
}
}
/* a literal */
literal:
UPDATE_I(dict,0,dindex,ip,in);
if (++ip >= ip_end)
break;
continue;
/* a match */
match:
UPDATE_I(dict,0,dindex,ip,in);
/* store current literal run */
lit = pd(ip,ii);
if (lit > 0)
{
lzo_uint t = lit;
if (t < 4 && op > out)
op[-2] = LZO_BYTE(op[-2] | t);
else if (t <= 31)
*op++ = LZO_BYTE(t);
else
{
lzo_uint tt = t - 31;
*op++ = 0;
while (tt > 255)
{
tt -= 255;
UA_SET1(op, 0);
op++;
}
assert(tt > 0);
*op++ = LZO_BYTE(tt);
}
do *op++ = *ii++; while (--t > 0);
}
assert(ii == ip);
/* code the match */
ip += 3;
if (*m_pos++ != *ip++ || *m_pos++ != *ip++ || *m_pos++ != *ip++ ||
*m_pos++ != *ip++ || *m_pos++ != *ip++ || *m_pos++ != *ip++)
{
--ip;
m_len = pd(ip, ii);
assert(m_len >= 3); assert(m_len <= 8);
if (m_off <= M2_MAX_OFFSET)
{
m_off -= 1;
*op++ = LZO_BYTE(((m_len - 2) << 5) | ((m_off & 7) << 2));
*op++ = LZO_BYTE(m_off >> 3);
}
else if (m_len == 3 && m_off <= 2*M2_MAX_OFFSET && lit > 0)
static
int do_compress ( const lzo_byte *in , lzo_uint in_len,
lzo_byte *out, lzo_uint *out_len,
lzo_voidp wrkmem )
{
#if 1 && defined(__GNUC__) && defined(__i386__)
register const lzo_byte *ip __asm__("%esi");
#else
register const lzo_byte *ip;
#endif
lzo_uint32 dv;
lzo_byte *op;
const lzo_byte * const in_end = in + in_len;
const lzo_byte * const ip_end = in + in_len - 9;
const lzo_byte *ii;
const lzo_bytepp const dict = (const lzo_bytepp) wrkmem;
op = out;
ip = in;
ii = ip;
DVAL_FIRST(dv,ip);
UPDATE_D(dict,cycle,dv,ip);
ip++;
DVAL_NEXT(dv,ip);
while (1)
{
#if 1 && defined(__GNUC__) && defined(__i386__)
register const lzo_byte *m_pos __asm__("%edi");
#else
register const lzo_byte *m_pos;
#endif
lzo_uint m_len;
lzo_ptrdiff_t m_off;
lzo_uint lit;
{
lzo_uint dindex = DINDEX(dv,ip);
m_pos = dict[dindex];
UPDATE_I(dict,cycle,dindex,ip);
}
if (LZO_CHECK_MPOS_NON_DET(m_pos,m_off,in,ip,M3_MAX_OFFSET))
{
}
#if defined(LZO_UNALIGNED_OK_2)
else if (* (lzo_ushortp) m_pos != * (lzo_ushortp) ip)
#else
else if (m_pos[0] != ip[0] || m_pos[1] != ip[1])
#endif
{
}
else
{
if (m_pos[2] == ip[2])
{
m_pos += 3;
if (m_off <= M2_MAX_OFFSET)
goto match;
#if 1
if (ip - ii <= 3)
goto match;
#else
if (ip - ii == 3) /* better compression, but slower */
goto match;
#endif
if (*m_pos == ip[3])
goto match;
}
}
/* a literal */
++ip;
if (ip >= ip_end)
break;
DVAL_NEXT(dv,ip);
continue;
/* a match */
match:
/* store current literal run */
lit = ip - ii;
if (lit > 0)
{
register lzo_uint t = lit;
if (t < 4 && op > out)
op[-2] |= LZO_BYTE(t);
else if (t <= 31)
*op++ = LZO_BYTE(t);
else
{
register lzo_uint tt = t - 31;
*op++ = 0;
while (tt > 255)
{
tt -= 255;
*op++ = 0;
}
assert(tt > 0);
*op++ = LZO_BYTE(tt);
}
do *op++ = *ii++; while (--t > 0);
}
assert(ii == ip);
/* code the match */
ip += 3;
if (*m_pos++ != *ip++ || *m_pos++ != *ip++ || *m_pos++ != *ip++ ||
*m_pos++ != *ip++ || *m_pos++ != *ip++ || *m_pos++ != *ip++)
{
--ip;
m_len = ip - ii;
assert(m_len >= 3); assert(m_len <= 8);
if (m_off <= M2_MAX_OFFSET)
{
m_off -= 1;
*op++ = LZO_BYTE(((m_len - 2) << 5) | ((m_off & 7) << 2));
*op++ = LZO_BYTE(m_off >> 3);
}
else if (m_len == 3 && m_off <= 2*M2_MAX_OFFSET && lit > 0)
