[PATCH 1 of 3 v4] store: implement fncache path mangling code in C
Bryan O'Sullivan
bos at serpentine.com
Thu Sep 6 18:31:16 CDT 2012
# HG changeset patch
# User Bryan O'Sullivan <bryano at fb.com>
# Date 1346971115 25200
# Node ID 831e86e4a9e9621c32c9c3fd314790d7a0276f22
# Parent c11d7eac6ee384f8b3e7d0f005cd2556d2d0b29c
store: implement fncache path mangling code in C
(This is not yet enabled; it will be turned on in a followup patch.)
The Python path mangling code used by fncache is complex and (perhaps
surprisingly) slow enough to negatively affect the overall performance
of Mercurial.
The name mangling scheme used by fncache consists of two paths, both of
which are translated into C in this patch.
For a short path (< 120 bytes), the Python code can be reduced to a
fairly tractable state machine that mangles the path in a single pass.
For longer paths, the mangling scheme is much more complicated, has
poor data dependencies, and requires a name to be hashed using SHA-1.
It's almost impossible to see how to implement this without multiple
copying passes.
Raw performance: I measured in a repo containing 150,000 files in its tip
manifest, with a median path name length of 57 bytes, and 95th percentile
of 96 bytes.
In this repo, the Python code takes 3.95 seconds to encode all path names,
while the C code (called from Python) takes 0.15 seconds.
Across several other large repositories, I've measured the speedup from
the C code at between 26x and 40x. For path names above 120 bytes where
the hashed encoding method must be used, the speedup is about 16x.
Practical performance:
"hg clone -U --uncompressed" of mozilla-central over http
Python: 20.5 seconds
mixed: 17.5 (server C, client Python)
all C: 14.8
"hg update" from null to tip in mozilla-central
Python: 30.1 seconds
C: 28.7
* * *
store: enable native fncache encoding by default
diff --git a/mercurial/parsers.c b/mercurial/parsers.c
--- a/mercurial/parsers.c
+++ b/mercurial/parsers.c
@@ -1506,11 +1506,14 @@
static char parsers_doc[] = "Efficient content parsing.";
+PyObject *pathencode(PyObject *self, PyObject *args);
+
static PyMethodDef methods[] = {
{"pack_dirstate", pack_dirstate, METH_VARARGS, "pack a dirstate\n"},
{"parse_manifest", parse_manifest, METH_VARARGS, "parse a manifest\n"},
{"parse_dirstate", parse_dirstate, METH_VARARGS, "parse a dirstate\n"},
{"parse_index2", parse_index2, METH_VARARGS, "parse a revlog index\n"},
+ {"pathencode", pathencode, METH_VARARGS, "fncache-encode a path\n"},
{NULL, NULL}
};
diff --git a/mercurial/pathencode.c b/mercurial/pathencode.c
new file mode 100644
--- /dev/null
+++ b/mercurial/pathencode.c
@@ -0,0 +1,686 @@
+/*
+ pathencode.c - efficient path name encoding
+
+ Copyright 2012 Facebook
+
+ This software may be used and distributed according to the terms of
+ the GNU General Public License, incorporated herein by reference.
+*/
+
+/*
+ * An implementation of the name encoding scheme used by the fncache
+ * store. The common case is of a path < 120 bytes long, which is
+ * handled either in a single pass with no allocations or two passes
+ * with a single allocation. For longer paths, multiple passes are
+ * required.
+ */
+
+#include <Python.h>
+#include <assert.h>
+#include <ctype.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+
+/* state machine for the fast path */
+enum path_state {
+ START, /* first byte of a path component */
+ A, /* "AUX" */
+ AU,
+ THIRD, /* third of a 3-byte sequence, e.g. "AUX", "NUL" */
+ C, /* "CON" or "COMn" */
+ CO,
+ COMLPT, /* "COM" or "LPT" */
+ COMLPTn,
+ L,
+ LP,
+ N,
+ NU,
+ P, /* "PRN" */
+ PR,
+ LDOT, /* leading '.' */
+ DOT, /* '.' in a non-leading position */
+ H, /* ".h" */
+ HGDI, /* ".hg", ".d", or ".i" */
+ SPACE,
+ DEFAULT, /* byte of a path component after the first */
+};
+
+/* state machine for dir-encoding */
+enum dir_state {
+ DDOT,
+ DH,
+ DHGDI,
+ DDEFAULT,
+};
+
+static inline int isset(const uint32_t bitset[], char c)
+{
+ return bitset[((uint8_t) c) >> 5] & (1 << (((uint8_t) c) & 31));
+}
+
+static inline void charcopy(char *dest, Py_ssize_t *destlen, size_t destsize,
+ char c)
+{
+ if (dest) {
+ assert(*destlen < destsize);
+ dest[*destlen] = c;
+ }
+ (*destlen)++;
+}
+
+static inline void memcopy(char *dest, Py_ssize_t *destlen, size_t destsize,
+ const void *src, Py_ssize_t len)
+{
+ if (dest) {
+ assert(*destlen + len < destsize);
+ memcpy((void *)&dest[*destlen], src, len);
+ }
+ *destlen += len;
+}
+
+static inline void hexencode(char *dest, Py_ssize_t *destlen, size_t destsize,
+ uint8_t c)
+{
+ static const char hexdigit[] = "0123456789abcdef";
+
+ charcopy(dest, destlen, destsize, hexdigit[c >> 4]);
+ charcopy(dest, destlen, destsize, hexdigit[c & 15]);
+}
+
+/* 3-byte escape: tilde followed by two hex digits */
+static inline void escape3(char *dest, Py_ssize_t *destlen, size_t destsize,
+ char c)
+{
+ charcopy(dest, destlen, destsize, '~');
+ hexencode(dest, destlen, destsize, c);
+}
+
+static Py_ssize_t _encode(const uint32_t twobytes[8], const uint32_t onebyte[8],
+ char *dest, Py_ssize_t destlen, size_t destsize,
+ const char *src, Py_ssize_t len,
+ int encodedir)
+{
+ enum path_state state = START;
+ Py_ssize_t i = 0;
+
+ /*
+ * Python strings end with a zero byte, which we use as a
+ * terminal token as they are not valid inside path names.
+ */
+
+ while (i < len) {
+ switch (state) {
+ case START:
+ switch (src[i]) {
+ case '/':
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case '.':
+ state = LDOT;
+ escape3(dest, &destlen, destsize, src[i++]);
+ break;
+ case ' ':
+ state = DEFAULT;
+ escape3(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'a':
+ state = A;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'c':
+ state = C;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'l':
+ state = L;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'n':
+ state = N;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'p':
+ state = P;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ default:
+ state = DEFAULT;
+ break;
+ }
+ break;
+ case A:
+ if (src[i] == 'u') {
+ state = AU;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case AU:
+ if (src[i] == 'x') {
+ state = THIRD;
+ i++;
+ }
+ else state = DEFAULT;
+ break;
+ case THIRD:
+ state = DEFAULT;
+ switch (src[i]) {
+ case '.':
+ case '/':
+ case '\0':
+ escape3(dest, &destlen, destsize, src[i - 1]);
+ break;
+ default:
+ i--;
+ break;
+ }
+ break;
+ case C:
+ if (src[i] == 'o') {
+ state = CO;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case CO:
+ if (src[i] == 'm') {
+ state = COMLPT;
+ i++;
+ }
+ else if (src[i] == 'n') {
+ state = THIRD;
+ i++;
+ }
+ else state = DEFAULT;
+ break;
+ case COMLPT:
+ switch (src[i]) {
+ case '1': case '2': case '3': case '4': case '5':
+ case '6': case '7': case '8': case '9':
+ state = COMLPTn;
+ i++;
+ break;
+ default:
+ state = DEFAULT;
+ charcopy(dest, &destlen, destsize, src[i - 1]);
+ break;
+ }
+ break;
+ case COMLPTn:
+ state = DEFAULT;
+ switch (src[i]) {
+ case '.':
+ case '/':
+ case '\0':
+ escape3(dest, &destlen, destsize, src[i - 2]);
+ charcopy(dest, &destlen, destsize, src[i - 1]);
+ break;
+ default:
+ memcopy(dest, &destlen, destsize,
+ &src[i - 2], 2);
+ break;
+ }
+ break;
+ case L:
+ if (src[i] == 'p') {
+ state = LP;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case LP:
+ if (src[i] == 't') {
+ state = COMLPT;
+ i++;
+ }
+ else state = DEFAULT;
+ break;
+ case N:
+ if (src[i] == 'u') {
+ state = NU;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case NU:
+ if (src[i] == 'l') {
+ state = THIRD;
+ i++;
+ }
+ else state = DEFAULT;
+ break;
+ case P:
+ if (src[i] == 'r') {
+ state = PR;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case PR:
+ if (src[i] == 'n') {
+ state = THIRD;
+ i++;
+ }
+ else state = DEFAULT;
+ break;
+ case LDOT:
+ switch (src[i]) {
+ case 'd':
+ case 'i':
+ state = HGDI;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'h':
+ state = H;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ default:
+ state = DEFAULT;
+ break;
+ }
+ break;
+ case DOT:
+ switch (src[i]) {
+ case '/':
+ case '\0':
+ state = START;
+ memcopy(dest, &destlen, destsize, "~2e", 3);
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'd':
+ case 'i':
+ state = HGDI;
+ charcopy(dest, &destlen, destsize, '.');
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'h':
+ state = H;
+ memcopy(dest, &destlen, destsize, ".h", 2);
+ i++;
+ break;
+ default:
+ state = DEFAULT;
+ charcopy(dest, &destlen, destsize, '.');
+ break;
+ }
+ break;
+ case H:
+ if (src[i] == 'g') {
+ state = HGDI;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case HGDI:
+ if (src[i] == '/') {
+ state = START;
+ if (encodedir)
+ memcopy(dest, &destlen, destsize, ".hg",
+ 3);
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DEFAULT;
+ break;
+ case SPACE:
+ switch (src[i]) {
+ case '/':
+ case '\0':
+ state = START;
+ memcopy(dest, &destlen, destsize, "~20", 3);
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ default:
+ state = DEFAULT;
+ charcopy(dest, &destlen, destsize, ' ');
+ break;
+ }
+ break;
+ case DEFAULT:
+ while (isset(onebyte, src[i])) {
+ charcopy(dest, &destlen, destsize, src[i++]);
+ if (i == len)
+ goto done;
+ }
+ switch (src[i]) {
+ case '.':
+ state = DOT;
+ i++;
+ break;
+ case ' ':
+ state = SPACE;
+ i++;
+ break;
+ case '/':
+ state = START;
+ charcopy(dest, &destlen, destsize, '/');
+ i++;
+ break;
+ default:
+ if (isset(onebyte, src[i])) {
+ do {
+ charcopy(dest, &destlen,
+ destsize, src[i++]);
+ } while (i < len &&
+ isset(onebyte, src[i]));
+ }
+ else if (isset(twobytes, src[i])) {
+ char c = src[i++];
+ charcopy(dest, &destlen, destsize, '_');
+ charcopy(dest, &destlen, destsize,
+ c == '_' ? '_' : c + 32);
+ }
+ else
+ escape3(dest, &destlen, destsize,
+ src[i++]);
+ break;
+ }
+ break;
+ }
+ }
+done:
+ return destlen;
+}
+
+static Py_ssize_t basicencode(char *dest, size_t destsize,
+ const char *src, Py_ssize_t len)
+{
+ static const uint32_t twobytes[8] = { 0, 0, 0x87fffffe };
+
+ static const uint32_t onebyte[8] = {
+ 1, 0x2bff3bfa, 0x68000001, 0x2fffffff,
+ };
+
+ Py_ssize_t destlen = 0;
+
+ if (len < 5 || memcmp(src, "data/", 5) != 0) {
+ memcopy(dest, &destlen, destsize, src, len);
+ return destlen;
+ }
+
+ memcopy(dest, &destlen, destsize, "data/", 5);
+
+ return _encode(twobytes, onebyte, dest, destlen, destsize,
+ src + 5, len - 5, 1);
+}
+
+static Py_ssize_t direncode(char *dest, size_t destsize,
+ const char *src, Py_ssize_t len)
+{
+ enum dir_state state = DDEFAULT;
+ Py_ssize_t i = 5, destlen = 0;
+
+ memcopy(dest, &destlen, destsize, "data/", 5);
+
+ while (i < len) {
+ switch (state) {
+ case DDOT:
+ switch (src[i]) {
+ case 'd':
+ case 'i':
+ state = DHGDI;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ case 'h':
+ state = DH;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ default:
+ state = DDEFAULT;
+ break;
+ }
+ break;
+ case DH:
+ if (src[i] == 'g') {
+ state = DHGDI;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ else state = DDEFAULT;
+ break;
+ case DHGDI:
+ if (src[i] == '/') {
+ memcopy(dest, &destlen, destsize, ".hg", 3);
+ charcopy(dest, &destlen, destsize, src[i++]);
+ }
+ state = DDEFAULT;
+ break;
+ case DDEFAULT:
+ if (src[i] == '.')
+ state = DDOT;
+ charcopy(dest, &destlen, destsize, src[i++]);
+ break;
+ }
+ }
+
+ return destlen;
+}
+
+static Py_ssize_t auxencode(char *dest, size_t destsize,
+ const char *src, Py_ssize_t len)
+{
+ static const uint32_t twobytes[8];
+
+ static const uint32_t onebyte[8] = {
+ ~0, 0xffffbffe, ~0, ~0, ~0, ~0, ~0, ~0,
+ };
+
+ return _encode(twobytes, onebyte, dest, 0, destsize, src, len, 0);
+}
+
+static const Py_ssize_t maxstorepathlen = 120;
+
+static Py_ssize_t lowerencode(char *dest, size_t destsize,
+ const char *src, Py_ssize_t len)
+{
+ static const uint32_t onebyte[8] = {
+ 1, 0x2bfffbfb, 0xe8000001, 0x2fffffff
+ };
+
+ static const uint32_t lower[8] = { 0, 0, 0x7fffffe };
+
+ Py_ssize_t i = 0, destlen = 0;
+
+ while (i < len) {
+ if (isset(onebyte, src[i]))
+ charcopy(dest, &destlen, destsize, src[i++]);
+ else if (isset(lower, src[i]))
+ charcopy(dest, &destlen, destsize, src[i++] + 32);
+ else
+ escape3(dest, &destlen, destsize, src[i++]);
+ }
+
+ return destlen;
+}
+
+static PyObject *hashmangle(const char *src, Py_ssize_t len, const char sha[20])
+{
+ static const Py_ssize_t dirprefixlen = 8;
+ static const Py_ssize_t maxshortdirslen = 68;
+ char *dest;
+ PyObject *ret;
+
+ Py_ssize_t i, d, p, lastslash = len - 1, lastdot = -1;
+ Py_ssize_t destsize, destlen = 0, slop, used;
+
+ while (lastslash >= 0 && src[lastslash] != '/') {
+ if (src[lastslash] == '.' && lastdot == -1)
+ lastdot = lastslash;
+ lastslash--;
+ }
+
+ destsize = 170;
+ if (lastdot >= 0)
+ destsize += len - lastdot - 1;
+
+ ret = PyString_FromStringAndSize(NULL, destsize);
+ if (ret == NULL)
+ return NULL;
+
+ dest = PyString_AS_STRING(ret);
+ memcopy(dest, &destlen, destsize, "dh/", 3);
+
+ for (i = d = p = 0; i < lastslash; i++, p++) {
+ if (src[i] == '/') {
+ char d = destlen > 0 ? dest[destlen - 1] : 0;
+ /* After truncation, a directory name may end
+ in a space or dot. */
+ if (d == '.' || d == ' ') {
+ destlen--;
+ charcopy(dest, &destlen, destsize, '_');
+ }
+ if (destlen > maxshortdirslen)
+ break;
+ charcopy(dest, &destlen, destsize, src[i]);
+ p = -1;
+ }
+ else if (p < dirprefixlen)
+ charcopy(dest, &destlen, destsize, src[i]);
+ }
+
+ if (destlen > maxshortdirslen + 3)
+ do {
+ destlen--;
+ } while (destlen > 0 && dest[destlen] != '/');
+
+ if (destlen > 3)
+ charcopy(dest, &destlen, destsize, '/');
+
+ used = destlen + 40;
+ if (lastdot >= 0)
+ used += len - lastdot - 1;
+ slop = maxstorepathlen - used;
+ if (slop > 0) {
+ Py_ssize_t basenamelen =
+ lastslash >= 0 ? len - lastslash - 2 : len - 1;
+
+ if (basenamelen > slop)
+ basenamelen = slop;
+ if (basenamelen > 0)
+ memcopy(dest, &destlen, destsize, &src[lastslash + 1],
+ basenamelen);
+ }
+
+ for (i = 0; i < 20; i++)
+ hexencode(dest, &destlen, destsize, sha[i]);
+
+ if (lastdot >= 0)
+ memcopy(dest, &destlen, destsize, &src[lastdot],
+ len - lastdot - 1);
+
+ PyString_GET_SIZE(ret) = destlen;
+
+ return ret;
+}
+
+/*
+ * Avoiding a trip through Python would improve performance by 50%,
+ * but we don't encounter enough long names to be worth the code.
+ */
+static int sha1hash(char hash[20], const char *str, Py_ssize_t len)
+{
+ static PyObject *shafunc;
+ PyObject *shaobj, *hashobj;
+
+ if (shafunc == NULL) {
+ PyObject *util, *name = PyString_FromString("mercurial.util");
+
+ if (name == NULL)
+ return -1;
+
+ util = PyImport_Import(name);
+ Py_DECREF(name);
+
+ if (util == NULL) {
+ PyErr_SetString(PyExc_ImportError, "mercurial.util");
+ return -1;
+ }
+ shafunc = PyObject_GetAttrString(util, "sha1");
+ Py_DECREF(util);
+
+ if (shafunc == NULL) {
+ PyErr_SetString(PyExc_AttributeError,
+ "module 'mercurial.util' has no "
+ "attribute 'sha1'");
+ return -1;
+ }
+ }
+
+ shaobj = PyObject_CallFunction(shafunc, "s#", str, len);
+
+ if (shaobj == NULL)
+ return -1;
+
+ hashobj = PyObject_CallMethod(shaobj, "digest", "");
+ Py_DECREF(shaobj);
+
+ if (!PyString_Check(hashobj) || PyString_GET_SIZE(hashobj) != 20) {
+ PyErr_SetString(PyExc_TypeError,
+ "result of digest is not a 20-byte hash");
+ Py_DECREF(hashobj);
+ return -1;
+ }
+
+ memcpy(hash, PyString_AS_STRING(hashobj), 20);
+ Py_DECREF(hashobj);
+ return 0;
+}
+
+static PyObject *hashencode(const char *src, Py_ssize_t len)
+{
+ const Py_ssize_t baselen = (len - 5) * 3;
+#ifndef _MSC_VER
+ /* alloca is surprisingly slow, so avoid with sane compilers */
+ char dired[baselen];
+ char lowered[baselen];
+ char auxed[baselen];
+#else
+ char *dired = alloca(baselen);
+ char *lowered = alloca(baselen);
+ char *auxed = alloca(baselen);
+#endif
+ Py_ssize_t dirlen, lowerlen, auxlen;
+ char sha[20];
+
+ dirlen = direncode(dired, baselen, src, len);
+ if (sha1hash(sha, dired, dirlen - 1) == -1)
+ return NULL;
+ lowerlen = lowerencode(lowered, baselen, dired + 5, dirlen - 5);
+ auxlen = auxencode(auxed, baselen, lowered, lowerlen);
+ return hashmangle(auxed, auxlen, sha);
+}
+
+PyObject *pathencode(PyObject *self, PyObject *args)
+{
+ Py_ssize_t len, newlen;
+ PyObject *pathobj, *newobj;
+ char *path;
+
+ if (!PyArg_ParseTuple(args, "O:pathencode", &pathobj))
+ return NULL;
+
+ if (PyString_AsStringAndSize(pathobj, &path, &len) == -1) {
+ PyErr_SetString(PyExc_TypeError, "expected a string");
+ return NULL;
+ }
+
+ newlen = len ? basicencode(NULL, 0, path, len + 1) : 1;
+
+ if (newlen <= maxstorepathlen + 1) {
+ if (newlen == len + 1) {
+ Py_INCREF(pathobj);
+ return pathobj;
+ }
+
+ newobj = PyString_FromStringAndSize(NULL, newlen);
+
+ if (newobj) {
+ PyString_GET_SIZE(newobj)--;
+ basicencode(PyString_AS_STRING(newobj), newlen, path,
+ len + 1);
+ }
+ }
+ else
+ newobj = hashencode(path, len + 1);
+
+ return newobj;
+}
diff --git a/setup.py b/setup.py
--- a/setup.py
+++ b/setup.py
@@ -421,7 +421,8 @@
Extension('mercurial.bdiff', ['mercurial/bdiff.c']),
Extension('mercurial.diffhelpers', ['mercurial/diffhelpers.c']),
Extension('mercurial.mpatch', ['mercurial/mpatch.c']),
- Extension('mercurial.parsers', ['mercurial/parsers.c']),
+ Extension('mercurial.parsers', ['mercurial/parsers.c',
+ 'mercurial/pathencode.c']),
]
osutil_ldflags = []
More information about the Mercurial-devel
mailing list