github
/
duktape
mirror of https://github.com/svaarala/duktape.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Prune and clean up string charlen algorithms

pull/422/head
Sami Vaarala 9 years ago
parent
1bddde8daf
commit
bc7a732002
1 changed files with 39 additions and 444 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 483
      src/duk_unicode_support.c

483
src/duk_unicode_support.c
View File

@ -257,101 +257,21 @@ DUK_INTERNAL duk_ucodepoint_t duk_unicode_decode_xutf8_checked(duk_hthread *thr,
/* Compute (extended) utf-8 length without codepoint encoding validation, /* Compute (extended) utf-8 length without codepoint encoding validation,
* used for string interning. * used for string interning.
* *
* NOTE: This algorithm is performance critical (more so than string hashing * NOTE: This algorithm is performance critical, more so than string hashing
* in some cases): it is needed when interning a string and it needs to scan * in some cases. It is needed when interning a string and needs to scan
* every byte of the string with no skipping. Having an ASCII fast path * every byte of the string with no skipping. Having an ASCII fast path
* would be useful (if possible in the algorithm). Several variants are * is useful if possible in the algorithm. The current algorithms were
* left below, commented out; the active algorithm was chosen on x64 based * chosen from several variants, based on x64 gcc -O2 testing. See:
* on gcc -O2 testing. * https://github.com/svaarala/duktape/pull/422
*/ */
const duk_uint8_t duk__ncont_incr[256] = { #if defined(DUK_USE_PREFER_SIZE)
/* 10xxxxxx = continuation chars (0x80...0xbf), above /* Small variant; roughly 150 bytes smaller than the fast variant. */
* and below that initial bytes. DUK_INTERNAL duk_size_t duk_unicode_unvalidated_utf8_length(const duk_uint8_t *data, duk_size_t blen) {
*/
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
};
const duk_uint8_t duk__nchar_incr[256] = {
/* 10xxxxxx = continuation chars (0x80...0xbf), above
* and below that initial bytes.
*/
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1
};
DUK_LOCAL duk_size_t duk__unicode_utf8clen_simple1(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p;
const duk_uint8_t *p_end;
duk_size_t clen;
p = data;
p_end = data + blen;
clen = 0;
while (p != p_end) {
duk_uint8_t x = *p++;
if (DUK_LIKELY(x < 0x80 || x >= 0xc0)) {
/* 10xxxxxx = continuation chars (0x80...0xbf), above
* and below that initial bytes.
*/
clen++;
}
}
return clen;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_simple2(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p;
const duk_uint8_t *p_end;
duk_size_t ncont;
p = data;
p_end = data + blen;
ncont = 0;
while (p != p_end) {
duk_uint8_t x = *p++;
if (DUK_UNLIKELY(x >= 0x80 && x <= 0xbf)) {
ncont++;
}
}
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_simple3(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p; const duk_uint8_t *p;
const duk_uint8_t *p_end; const duk_uint8_t *p_end;
duk_size_t ncont; duk_size_t ncont;
duk_size_t clen;
p = data; p = data;
p_end = data + blen; p_end = data + blen;
@ -359,156 +279,53 @@ DUK_LOCAL duk_size_t duk__unicode_utf8clen_simple3(const duk_uint8_t *data, duk_
while (p != p_end) { while (p != p_end) {
duk_uint8_t x; duk_uint8_t x;
x = *p++; x = *p++;
ncont += ((x & 0xc0) == 0x80) ? 1 : 0; if (DUK_UNLIKELY(x >= 0x80 && x <= 0xbf)) {
} ncont++;
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_simple4(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p;
const duk_uint8_t *p_end;
duk_size_t ncont;
p = data;
p_end = data + blen;
ncont = 0;
while (p != p_end) {
/* Bit trick:
* 10xxxxxx ^ 01000000 = 11xxxxxx (and other bit patterns are 10xxxxxx or less)
* + 01000000 = 1 00xxxxxx (and other bit patterns won't overflow to 9 bits)
* >>> 8 = 1
*/
duk_small_uint_t x;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
} }
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_lookup1(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p;
const duk_uint8_t *p_end;
duk_size_t clen;
p = data;
p_end = data + blen;
clen = 0;
while (p != p_end) {
duk_uint8_t x;
x = *p++;
clen += duk__nchar_incr[x];
} }
DUK_ASSERT(ncont <= blen);
clen = blen - ncont;
DUK_ASSERT(clen <= blen);
return clen; return clen;
} }
#else /* DUK_USE_PREFER_SIZE */
DUK_LOCAL duk_size_t duk__unicode_utf8clen_lookup2(const duk_uint8_t *data, duk_size_t blen) { /* This seems like a good overall approach. Fast path for ASCII in 4 byte
* blocks.
*/
DUK_INTERNAL duk_size_t duk_unicode_unvalidated_utf8_length(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p; const duk_uint8_t *p;
const duk_uint8_t *p_end; const duk_uint8_t *p_end;
const duk_uint32_t *p32_end;
const duk_uint32_t *p32;
duk_size_t ncont; duk_size_t ncont;
duk_size_t clen;
ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
p = data; p = data;
p_end = data + blen; p_end = data + blen;
ncont = 0;
while (p != p_end) {
duk_uint8_t x;
x = *p++;
ncont += duk__ncont_incr[x];
}
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll1(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
const duk_uint32_t *p32;
if (blen < 16) { if (blen < 16) {
goto skip_fastpath; goto skip_fastpath;
} }
/* Align 'p' to 4. */
while (((duk_small_uint_t) (duk_uintptr_t) (void *) p) & 0x03) {
duk_uint8_t x;
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
}
/* Full, aligned 4-byte reads. */
p_end = data + blen;
p_end = p + ((duk_size_t) (p_end - p) & (duk_size_t) (~0x03));
p32 = (const duk_uint32_t *) p;
while (p32 != (const duk_uint32_t *) p_end) {
duk_uint32_t x;
x = *p32++;
if ((x & 0x80808080UL) == 0) {
; /* ASCII fast path */
} else {
if ((x & 0xc0000000UL) == 0x80000000UL) {
ncont++;
}
if ((x & 0x00c00000UL) == 0x00800000UL) {
ncont++;
}
if ((x & 0x0000c000UL) == 0x00008000UL) {
ncont++;
}
if ((x & 0x000000c0UL) == 0x00000080UL) {
ncont++;
}
}
}
p = (const duk_uint8_t *) p32;
/* Fall through to handle the rest. */
skip_fastpath:
p_end = data + blen;
while (p != p_end) {
duk_uint8_t x;
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
}
DUK_ASSERT(ncont <= blen);
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll2(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
const duk_uint32_t *p32;
if (blen < 16) { /* Align 'p' to 4; the input data may have arbitrary alignment.
goto skip_fastpath; * End of string check not needed because blen >= 16.
} */
/* Align 'p' to 4. */ while (((duk_small_uint_t) (duk_uintptr_t) (const void *) p) & 0x03) {
while (((duk_small_uint_t) (duk_uintptr_t) (void *) p) & 0x03) {
duk_uint8_t x; duk_uint8_t x;
x = *p++; x = *p++;
if (x < 0x80 || x >= 0xc0) { if (DUK_UNLIKELY(x >= 0x80 && x <= 0xbf)) {
;
} else {
ncont++; ncont++;
} }
} }
/* Full, aligned 4-byte reads. */ /* Full, aligned 4-byte reads. */
p_end = data + blen; p32_end = (const duk_uint32_t *) (p + ((duk_size_t) (p_end - p) & (duk_size_t) (~0x03)));
p_end = p + ((duk_size_t) (p_end - p) & (duk_size_t) (~0x03));
p32 = (const duk_uint32_t *) p; p32 = (const duk_uint32_t *) p;
while (p32 != (const duk_uint32_t *) p_end) { while (p32 != (const duk_uint32_t *) p32_end) {
duk_uint32_t x; duk_uint32_t x;
x = *p32++; x = *p32++;
if ((x & 0x80808080UL) == 0) { if (DUK_LIKELY((x & 0x80808080UL) == 0)) {
; /* ASCII fast path */ ; /* ASCII fast path */
} else { } else {
/* Flip highest bit of each byte which changes /* Flip highest bit of each byte which changes
@ -516,260 +333,38 @@ DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll2(const duk_uint8_t *data, duk_
* allows an easy bit mask test. * allows an easy bit mask test.
*/ */
x ^= 0x80808080UL; x ^= 0x80808080UL;
if (!(x & 0xc0000000UL)) { if (DUK_UNLIKELY(!(x & 0xc0000000UL))) {
ncont++; ncont++;
} }
if (!(x & 0x00c00000UL)) { if (DUK_UNLIKELY(!(x & 0x00c00000UL))) {
ncont++; ncont++;
} }
if (!(x & 0x0000c000UL)) { if (DUK_UNLIKELY(!(x & 0x0000c000UL))) {
ncont++; ncont++;
} }
if (!(x & 0x000000c0UL)) { if (DUK_UNLIKELY(!(x & 0x000000c0UL))) {
ncont++; ncont++;
} }
} }
} }
p = (const duk_uint8_t *) p32; p = (const duk_uint8_t *) p32;
/* Fall through to handle the rest. */ /* Fall through to handle the rest. */
skip_fastpath:
p_end = data + blen;
while (p != p_end) {
duk_uint8_t x;
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
}
DUK_ASSERT(ncont <= blen);
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll3(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
const duk_uint32_t *p32;
if (blen < 16) {
goto skip_fastpath;
}
/* Align 'p' to 4. */
while (((duk_small_uint_t) (duk_uintptr_t) (void *) p) & 0x03) {
duk_uint8_t x;
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
}
/* Full, aligned 4-byte reads. */
p_end = data + blen;
p_end = p + ((duk_size_t) (p_end - p) & (duk_size_t) (~0x03));
p32 = (const duk_uint32_t *) p;
while (p32 != (const duk_uint32_t *) p_end) {
/* Bit tricks to work 4 bytes at a time, similar to the bit trick below.
*
* 10xxxxxx 10xxxxxx 10xxxxxxx 10xxxxxx
* ^ 0x40404040 11xxxxxx 11xxxxxx 11xxxxxxx 11xxxxxx
* >> 6 00000011 00000011 000000011 00000011
* + 0x01010101 00000100 00000100 000000100 00000100
* ^ ^ ^ ^
* `--------+---------+--------+---- carry if cont byte [+]
*/
duk_uint32_t x;
x = *p32++;
x = ((x ^ 0x40404040UL) >> 6) + 0x01010101UL;
x &= 0x04040404UL;
x = (x & 0xffffUL) + (x >> 16); /* two step sum of carries */
x = (x & 0xffUL) + (x >> 8);
ncont += x >> 2;
}
/* Fall through to handle the rest. */
skip_fastpath: skip_fastpath:
p_end = data + blen;
while (p != p_end) { while (p != p_end) {
duk_uint8_t x; duk_uint8_t x;
x = *p++; x = *p++;
if (x < 0x80 || x >= 0xc0) { if (DUK_UNLIKELY(x >= 0x80 && x <= 0xbf)) {
;
} else {
ncont++;
}
}
DUK_ASSERT(ncont <= blen);
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll4(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
p_end = data + (blen & ((duk_size_t) (~0x03)));
while (p < p_end) {
duk_uint8_t x;
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++;
}
}
p_end = data + blen;
while (p < p_end) {
duk_uint8_t x;
x = *p++;
if (x < 0x80 || x >= 0xc0) {
;
} else {
ncont++; ncont++;
} }
} }
DUK_ASSERT(ncont <= blen); DUK_ASSERT(ncont <= blen);
return blen - ncont; clen = blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll5(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
p_end = data + (blen & ((duk_size_t) (~0x03)));
while (p != p_end) {
duk_uint8_t x;
x = *p++;
ncont += duk__ncont_incr[x];
x = *p++;
ncont += duk__ncont_incr[x];
x = *p++;
ncont += duk__ncont_incr[x];
x = *p++;
ncont += duk__ncont_incr[x];
}
p_end = data + blen;
while (p != p_end) {
duk_uint8_t x;
x = *p++;
ncont += duk__ncont_incr[x];
}
DUK_ASSERT(ncont <= blen);
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll6(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
p_end = data + (blen & ((duk_size_t) (~0x03)));
while (p != p_end) {
duk_small_uint_t x;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
}
p_end = data + blen;
while (p != p_end) {
duk_small_uint_t x;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
}
DUK_ASSERT(ncont <= blen);
return blen - ncont;
}
DUK_LOCAL duk_size_t duk__unicode_utf8clen_unroll7(const duk_uint8_t *data, duk_size_t blen) {
const duk_uint8_t *p = data;
const duk_uint8_t *p_end;
duk_size_t ncont = 0; /* number of continuation (non-initial) bytes in [0x80,0xbf] */
p_end = data + (blen & ((duk_size_t) (~0x03)));
while (p != p_end) {
/* Similar bit trick as above, but postpone the shift.
* This means we need to avoid overflows from the lower
* bits and need the "x & 0xc0".
*/
duk_small_uint_t x;
duk_small_uint_t tmp = 0;
x = *p++;
tmp += ((x & 0xc0) ^ 0x40) + 0x40;
x = *p++;
tmp += ((x & 0xc0) ^ 0x40) + 0x40;
x = *p++;
tmp += ((x & 0xc0) ^ 0x40) + 0x40;
x = *p++;
tmp += ((x & 0xc0) ^ 0x40) + 0x40;
ncont += tmp;
}
p_end = data + blen;
while (p != p_end) {
duk_small_uint_t x;
x = *p++;
ncont += ((x ^ 0x40) + 0x40) >> 8;
}
DUK_ASSERT(ncont <= blen);
return blen - ncont;
}
DUK_INTERNAL duk_size_t duk_unicode_unvalidated_utf8_length(const duk_uint8_t *data, duk_size_t blen) {
duk_size_t clen;
#if 0
clen = duk__unicode_utf8clen_simple1(data, blen);
clen = duk__unicode_utf8clen_simple2(data, blen);
clen = duk__unicode_utf8clen_simple3(data, blen);
clen = duk__unicode_utf8clen_simple4(data, blen);
clen = duk__unicode_utf8clen_lookup1(data, blen);
clen = duk__unicode_utf8clen_lookup2(data, blen);
clen = duk__unicode_utf8clen_unroll1(data, blen);
clen = duk__unicode_utf8clen_unroll2(data, blen);
clen = duk__unicode_utf8clen_unroll3(data, blen);
clen = duk__unicode_utf8clen_unroll4(data, blen);
clen = duk__unicode_utf8clen_unroll5(data, blen);
clen = duk__unicode_utf8clen_unroll6(data, blen);
clen = duk__unicode_utf8clen_unroll7(data, blen);
#endif
clen = duk__unicode_utf8clen_unroll1(data, blen);
DUK_ASSERT(clen <= blen); DUK_ASSERT(clen <= blen);
return clen; return clen;
} }
#endif /* DUK_USE_PREFER_SIZE */
/* /*
* Unicode range matcher * Unicode range matcher

Write Preview
Loading…
Cancel
Save