mercurial/mpatch.c

rust-discovery: using from Python code

As previously done in other topics, the Rust version is used if it's been
built.

The version fully in Rust of the partialdiscovery class has the performance
advantage over the Python version (actually using the Rust MissingAncestor) if
the undecided set is big enough. Otherwise no sampling occurs, and the
discovery is reasonably fast anyway.

Note: it's hard to predict the size of the initial undecided set, it can
depend on the kind of topological changes between the local and remote graphs.
The point of the Rust version is to make the bad cases acceptable.

More specifically, the performance advantages are:

- faster sampling, especially takefullsample()
- much faster addmissings() in almost all cases (see commit message in
  grandparent of the present changeset)
- no conversion cost of the undecided set at the interface between Rust and
  Python

== Measurements with big undecided sets

For an extreme example, discovery between mozilla-try and mozilla-unified
(over one million undecided revisions, same case as in dbd0fcca6dfc), we
get roughly a x2.5/x3 better performance:

Growing sample size (5% starting with 200): time goes down from
210 to 72 seconds.
Constant sample size of 200: time down from 1853 to 659 seconds.

With a sample size computed from number of roots and heads of the
undecided set (`respectsize` is `False`), here are perfdiscovery results:

Before ! wall 9.358729 comb 9.360000 user 9.310000 sys 0.050000 (median of 50)
After  ! wall 3.793819 comb 3.790000 user 3.750000 sys 0.040000 (median of 50)

In that later case, the sample sizes are routinely in the hundreds of
thousands of revisions. While still faster, the Rust iteration in
addmissings has less of an advantage than with smaller sample sizes, but
one sees addcommons becoming faster, probably a consequence of not having
to copy big sets back and forth.

This example is not a goal in itself, but it showcases several different
areas in which the process can become slow, due to different factors, and
how this full Rust version can help.

== Measurements with small undecided sets

In cases the undecided set is small enough than no sampling occurs,
the Rust version has a disadvantage at init if `targetheads` is really big
(some time is lost in the translation to Rust data structures),
and that is compensated by the faster `addmissings()`.

On a private repository with over one million commits, we still get a minor
improvement, of 6.8%:

  Before ! wall 0.593585 comb 0.590000 user 0.550000 sys 0.040000 (median of 50)
  After  ! wall 0.553035 comb 0.550000 user 0.520000 sys 0.030000 (median of 50)

What's interesting in that case is the first addinfo() at 180ms for Rust and
233ms for Python+C, mostly due to add_missings and the children cache
computation being done in less than 0.2ms on the Rust side vs over 40ms on the
Python side.

The worst case we have on hand is with mozilla-try, prepared with
discovery-helper.sh for 10 heads and depth 10, time goes up 2.2% on the median.
In this case `targetheads` is really huge with 165842 server heads.

  Before ! wall 0.823884 comb 0.810000 user 0.790000 sys 0.020000 (median of 50)
  After  ! wall 0.842607 comb 0.840000 user 0.800000 sys 0.040000 (median of 50)

If that would be considered a problem, more adjustments can be made, which are
prematurate at this stage: cooking special variants of methods of the inner
MissingAncestors object, retrieving local heads directly from Rust to avoid
the cost of conversion. Effort would probably be better spent at this point
improving the surroundings if needed.

Here's another data point with a smaller repository, pypy, where performance
is almost identical

 Before ! wall 0.015121 comb 0.030000 user 0.020000 sys 0.010000 (median of 186)
 After  ! wall 0.015009 comb 0.010000 user 0.010000 sys 0.000000 (median of 184)

Differential Revision: https://phab.mercurial-scm.org/D6430

/*
 mpatch.c - efficient binary patching for Mercurial

 This implements a patch algorithm that's O(m + nlog n) where m is the
 size of the output and n is the number of patches.

 Given a list of binary patches, it unpacks each into a hunk list,
 then combines the hunk lists with a treewise recursion to form a
 single hunk list. This hunk list is then applied to the original
 text.

 The text (or binary) fragments are copied directly from their source
 Python objects into a preallocated output string to avoid the
 allocation of intermediate Python objects. Working memory is about 2x
 the total number of hunks.

 Copyright 2005, 2006 Matt Mackall <[e-mail]>

 This software may be used and distributed according to the terms
 of the GNU General Public License, incorporated herein by reference.
*/

#include <limits.h>
#include <stdlib.h>
#include <string.h>

#include "bitmanipulation.h"
#include "compat.h"
#include "mpatch.h"

/* VC9 doesn't include bool and lacks stdbool.h based on cext/util.h */
#if defined(_MSC_VER) || __STDC_VERSION__ < 199901L
#define true 1
#define false 0
typedef unsigned char bool;
#else
#include <stdbool.h>
#endif

static struct mpatch_flist *lalloc(ssize_t size)
{
	struct mpatch_flist *a = NULL;

	if (size < 1) {
		size = 1;
	}

	a = (struct mpatch_flist *)malloc(sizeof(struct mpatch_flist));
	if (a) {
		a->base = (struct mpatch_frag *)malloc(
		    sizeof(struct mpatch_frag) * size);
		if (a->base) {
			a->head = a->tail = a->base;
			return a;
		}
		free(a);
	}
	return NULL;
}

void mpatch_lfree(struct mpatch_flist *a)
{
	if (a) {
		free(a->base);
		free(a);
	}
}

static ssize_t lsize(struct mpatch_flist *a)
{
	return a->tail - a->head;
}

/* add helper to add src and *dest iff it won't overflow */
static inline bool safeadd(int src, int *dest)
{
	if ((src > 0) == (*dest > 0)) {
		if (*dest > 0) {
			if (src > (INT_MAX - *dest)) {
				return false;
			}
		} else {
			if (src < (INT_MIN - *dest)) {
				return false;
			}
		}
	}
	*dest += src;
	return true;
}

/* subtract src from dest and store result in dest */
static inline bool safesub(int src, int *dest)
{
	if (((src > 0) && (*dest < INT_MIN + src)) ||
	    ((src < 0) && (*dest > INT_MAX + src))) {
		return false;
	}
	*dest -= src;
	return true;
}

/* move hunks in source that are less cut to dest, compensating
   for changes in offset. the last hunk may be split if necessary.
*/
static int gather(struct mpatch_flist *dest, struct mpatch_flist *src, int cut,
                  int offset)
{
	struct mpatch_frag *d = dest->tail, *s = src->head;
	int postend, c, l;

	while (s != src->tail) {
		int soffset = s->start;
		if (!safeadd(offset, &soffset)) {
			break; /* add would overflow, oh well */
		}
		if (soffset >= cut) {
			break; /* we've gone far enough */
		}

		postend = offset;
		if (!safeadd(s->start, &postend) ||
		    !safeadd(s->len, &postend)) {
			break;
		}
		if (postend <= cut) {
			/* save this hunk */
			int tmp = s->start;
			if (!safesub(s->end, &tmp)) {
				break;
			}
			if (!safeadd(s->len, &tmp)) {
				break;
			}
			if (!safeadd(tmp, &offset)) {
				break; /* add would overflow, oh well */
			}
			*d++ = *s++;
		} else {
			/* break up this hunk */
			c = cut;
			if (!safesub(offset, &c)) {
				break;
			}
			if (s->end < c) {
				c = s->end;
			}
			l = cut - offset - s->start;
			if (s->len < l) {
				l = s->len;
			}

			offset += s->start + l - c;

			d->start = s->start;
			d->end = c;
			d->len = l;
			d->data = s->data;
			d++;
			s->start = c;
			s->len = s->len - l;
			s->data = s->data + l;

			break;
		}
	}

	dest->tail = d;
	src->head = s;
	return offset;
}

/* like gather, but with no output list */
static int discard(struct mpatch_flist *src, int cut, int offset)
{
	struct mpatch_frag *s = src->head;
	int postend, c, l;

	while (s != src->tail) {
		int cmpcut = s->start;
		if (!safeadd(offset, &cmpcut)) {
			break;
		}
		if (cmpcut >= cut) {
			break;
		}

		postend = offset;
		if (!safeadd(s->start, &postend)) {
			break;
		}
		if (!safeadd(s->len, &postend)) {
			break;
		}
		if (postend <= cut) {
			/* do the subtraction first to avoid UB integer overflow
			 */
			int tmp = s->start;
			if (!safesub(s->end, &tmp)) {
				break;
			}
			if (!safeadd(s->len, &tmp)) {
				break;
			}
			if (!safeadd(tmp, &offset)) {
				break;
			}
			s++;
		} else {
			c = cut;
			if (!safesub(offset, &c)) {
				break;
			}
			if (s->end < c) {
				c = s->end;
			}
			l = cut - offset - s->start;
			if (s->len < l) {
				l = s->len;
			}

			offset += s->start + l - c;
			s->start = c;
			s->len = s->len - l;
			s->data = s->data + l;

			break;
		}
	}

	src->head = s;
	return offset;
}

/* combine hunk lists a and b, while adjusting b for offset changes in a/
   this deletes a and b and returns the resultant list. */
static struct mpatch_flist *combine(struct mpatch_flist *a,
                                    struct mpatch_flist *b)
{
	struct mpatch_flist *c = NULL;
	struct mpatch_frag *bh, *ct;
	int offset = 0, post;

	if (a && b) {
		c = lalloc((lsize(a) + lsize(b)) * 2);
	}

	if (c) {

		for (bh = b->head; bh != b->tail; bh++) {
			/* save old hunks */
			offset = gather(c, a, bh->start, offset);

			/* discard replaced hunks */
			post = discard(a, bh->end, offset);

			/* insert new hunk */
			ct = c->tail;
			ct->start = bh->start;
			ct->end = bh->end;
			if (!safesub(offset, &(ct->start)) ||
			    !safesub(post, &(ct->end))) {
				/* It was already possible to exit
				 * this function with a return value
				 * of NULL before the safesub()s were
				 * added, so this should be fine. */
				mpatch_lfree(c);
				c = NULL;
				goto done;
			}
			ct->len = bh->len;
			ct->data = bh->data;
			c->tail++;
			offset = post;
		}

		/* hold on to tail from a */
		memcpy(c->tail, a->head, sizeof(struct mpatch_frag) * lsize(a));
		c->tail += lsize(a);
	}
done:
	mpatch_lfree(a);
	mpatch_lfree(b);
	return c;
}

/* decode a binary patch into a hunk list */
int mpatch_decode(const char *bin, ssize_t len, struct mpatch_flist **res)
{
	struct mpatch_flist *l;
	struct mpatch_frag *lt;
	int pos = 0;

	/* assume worst case size, we won't have many of these lists */
	l = lalloc(len / 12 + 1);
	if (!l) {
		return MPATCH_ERR_NO_MEM;
	}

	lt = l->tail;

	/* We check against len-11 to ensure we have at least 12 bytes
	   left in the patch so we can read our three be32s out of it. */
	while (pos >= 0 && pos < (len - 11)) {
		lt->start = getbe32(bin + pos);
		lt->end = getbe32(bin + pos + 4);
		lt->len = getbe32(bin + pos + 8);
		if (lt->start < 0 || lt->start > lt->end || lt->len < 0) {
			break; /* sanity check */
		}
		if (!safeadd(12, &pos)) {
			break;
		}
		lt->data = bin + pos;
		if (!safeadd(lt->len, &pos)) {
			break;
		}
		lt++;
	}

	if (pos != len) {
		mpatch_lfree(l);
		return MPATCH_ERR_CANNOT_BE_DECODED;
	}

	l->tail = lt;
	*res = l;
	return 0;
}

/* calculate the size of resultant text */
ssize_t mpatch_calcsize(ssize_t len, struct mpatch_flist *l)
{
	ssize_t outlen = 0, last = 0;
	struct mpatch_frag *f = l->head;

	while (f != l->tail) {
		if (f->start < last || f->end > len) {
			return MPATCH_ERR_INVALID_PATCH;
		}
		outlen += f->start - last;
		last = f->end;
		outlen += f->len;
		f++;
	}

	outlen += len - last;
	return outlen;
}

int mpatch_apply(char *buf, const char *orig, ssize_t len,
                 struct mpatch_flist *l)
{
	struct mpatch_frag *f = l->head;
	int last = 0;
	char *p = buf;

	while (f != l->tail) {
		if (f->start < last || f->start > len || f->end > len ||
		    last < 0) {
			return MPATCH_ERR_INVALID_PATCH;
		}
		memcpy(p, orig + last, f->start - last);
		p += f->start - last;
		memcpy(p, f->data, f->len);
		last = f->end;
		p += f->len;
		f++;
	}
	if (last < 0) {
		return MPATCH_ERR_INVALID_PATCH;
	}
	memcpy(p, orig + last, len - last);
	return 0;
}

/* recursively generate a patch of all bins between start and end */
struct mpatch_flist *
mpatch_fold(void *bins, struct mpatch_flist *(*get_next_item)(void *, ssize_t),
            ssize_t start, ssize_t end)
{
	ssize_t len;

	if (start + 1 == end) {
		/* trivial case, output a decoded list */
		return get_next_item(bins, start);
	}

	/* divide and conquer, memory management is elsewhere */
	len = (end - start) / 2;
	return combine(mpatch_fold(bins, get_next_item, start, start + len),
	               mpatch_fold(bins, get_next_item, start + len, end));
}