revlog: avoid some unnecessary seek/tell syscalls

manifest: speed up creation of the manifestdict - fold iteration and rawset into parse - avoid creating extra new strings with [:] where possible - speed up node.bin

lazyparser.findnode: fix typo and s/rfind/find/ There's no reason to use reverse string search and it's slightly slower

lazyindex: speed up __getitem__ This function is fairly performance sensitive, so we make a couple ugly tweaks: - keep all entries packed so we needn't test entry types - fold index lookup/load into unpack call to eliminate local variable setting

lazyparser: up the blocksize from 512 bytes to 64k

revlog: implement a fast path for heads

revlogio: speed up parsing - precalcuate ending offset - pull some variables into local scope - separate inline and out of line code paths

revlog: eliminate diff and patches functions call mdiff variants directly

revlog: speed up chunkcache - use a reasonable cache size - avoid an extra copy when we pull in big revs

revlog: some basic code reordering

revlog: abstract out index entry packing This lets us remove most of the remaining v0/v1 logic from revlog

revlog: simplify revlog.__init__ - move stat into io helper - get rid of self.defversion and self.indexstat - fold _load into __init__

revlog: mark cache private

revlog: pull chunkcache back into revlog

revlog: change _inline from a function to a variable

revlog: simplify addrevision - remove unused defaults for p1, p2, and text - reduce some if/else - use better variable names - remove some extra variables - remove some obsolete corner tests - simply first entry handling for revlogng - simply inline vs outofline writeout

revlog: some codingstyle cleanups

revlog: add a magic null revision to our index We expand our index by one entry so that index[nullrev] points to a unique entry, the null revision. This naturally eliminates numerous extra tests in the performance-sensitive index access functions, most of which are now trivial again. Adding new entries is now done with insert(-1, e) rather than append(e).

revlog: change accesses to index entry elements to use positive offsets

revlog: parse revlogv0 indexes into v1 internally This lets us eliminate lots of special case code in our performance-sensitive index accessors.

revlog: only allow lazy parsing with revlogng files This will allow us to store indices in memory in a single entry format

revlog: simplify the v1 immediate parser - read all the data at once (large files are handled by the lazy parser) - cache the entire file for inline revlogs - simplify looping

revlog: set the threshold for lazy parsing higher A typical machine can parse a 1MB index in well under a second

revlog: simplify the v0 parser

revlog: add revlogio interface This lets us separate the logic for different on-disk revlog formats from the main revlog logic.

revlog: regroup parsing code

transactions: avoid late tear-down (issue641) We use weak references (ugh) to avoid having to manually delete transaction references out of each call frame when an exception occurs.

hgk: enable mouse wheel under Windows. For some reason, MouseWheel events are not routed under windows even in latest ActiveTcl 8.4.15 while they are under linux and macosx. These events are activated using code supplied with Tcl Tip 171: <http://www.tcl.tk/cgi-bin/tct/tip/171.html>. Strangely, the Tip code almost work but generates some unexpected infinite loop which is fixed using a simple boolean to check reentrancy.

hgk: enable mouse wheel on MouseWheel events. At this point, mouse wheel still does not work under Windows.

merge with crew-stable

addremove: print meaningful error message if --similar not numeric

Merge with crew