~bzr-pqm/bzr/bzr.dev : changes

~bzr-pqm/bzr/bzr.dev » Changes from revision 4371.3.38

Rev	Summary	Authors	Date
4371.3.38	Add a failing test for handling nodes that are in the same l... Add a failing test for handling nodes that are in the same linear chain. It fails because the ancestry skipping causes us to miss the fact that the two nodes are actually directly related. We could check at the beginning, as the code used to do, but I think that will be incomplete for the more-than-two heads cases.	John Arbash Meinel	15 years ago
4371.3.37	cleanup passes. cleanup passes. Write some helper functions that avoid having to retype the ugly inline code everywhere. Get rid of the special-case exit when popping candidate nodes. At best it just avoids adding a couple parents to the heap before we hit the while loop and notice we are done anyway. And it allows us to remove some extraneous if/break blocks and clean up the code a bit.	John Arbash Meinel	15 years ago
4371.3.36	Cache heads() results from dominator lookups. Cache heads() results from dominator lookups.	John Arbash Meinel	15 years ago
4371.3.35	Changing the heapq algorithm actually ends up with better ov... Changing the heapq algorithm actually ends up with better overall performance. First, use the same 'replace instead of pop+push' trick. Second, fix a small issue when gdfo ordering lets us number further than I thought we could. Clarify some code paths, which even makes it a little bit faster. (currently at 68.3ms for bzr.dev, 771ms for OOo)	John Arbash Meinel	15 years ago
4371.3.34	Changing the code to allow a special case for not always pop... Changing the code to allow a special case for not always popping the last item, and instead re-using that slot for the next item. It resolves the issue with OOo (w/ OOo being so linear, my guess is 90% of it is numbered with a single entry in the list, which probably caused cycles of deallocating the buffer, and reallocating it on every node.) Not sure if it worth the ugliness, though. Certainly the code is more prone to mistakes now. I'll probably go back to the heapq version. Approximately as fast, and much more elegant.	John Arbash Meinel	15 years ago
4371.3.33	Use a hybrid pass to walk all nodes. Use a hybrid pass to walk all nodes. This avoids a heapq, which we didn't really need anyway. It is ~ the same number of passes, only we trade off a simplistic 'find_tails' call, with doing more work on the 1-pass over the dict. It seems to be slightly faster for bzr, and slightly slower for OOo. Not particularly noticeable in the noise of heads(), but measurable for __init__ performance.	John Arbash Meinel	15 years ago
4371.3.32	A comment about rewriting _find_gdfo A comment about rewriting _find_gdfo	John Arbash Meinel	15 years ago
4371.3.31	Change _KnownGraph.parents to be a tuple rather than a list. Change _KnownGraph.parents to be a tuple rather than a list. We know the size as soon as we walk to a map entry. This way, we save a malloc, and have slightly smaller objects. Not a huge difference, but a small one. This changes KnownGraph(OOo_parent_map) time from 922ms down to 848ms, or about a 10% savings for __init__ time.	John Arbash Meinel	15 years ago
4371.3.30	Clean up the annotate code while using the new functionality... Clean up the annotate code while using the new functionality.	John Arbash Meinel	15 years ago
4371.3.29	A couple quick cleanups A couple quick cleanups	John Arbash Meinel	15 years ago
4371.3.28	Finish converting away from 'cdef list' syntax. Finish converting away from 'cdef list' syntax. Also, I found that casting a PyObject * directly to _KnownGraphNode avoids the TypeTest. Since PyDict_GetItem and PyList_GET_ITEM return borrowed references anyway, we need a cast (to auto-incref the pointer), so we may as well use them for speed, and avoid the type check at the same time. Also using PyDict_Next everywhere instead of 'for x in dict.iterX'. The nice thing about PyDict_Next is that it always gives you items, but you can chose what you actually care about. Also, by doing all these manipulations directly, we shave some more time out of building a KnownGraph and out of the heads() calls.	John Arbash Meinel	15 years ago
4371.3.27	Remove some += and unused code anyway. Remove some += and unused code anyway.	John Arbash Meinel	15 years ago
4371.3.26	A few more cleanups. Start moving away from pyrex 0.9.8isms A few more cleanups. Start moving away from pyrex 0.9.8isms	John Arbash Meinel	15 years ago
4371.3.25	Remove the time_graph.py tool, it has served its purpose. Remove the time_graph.py tool, it has served its purpose.	John Arbash Meinel	15 years ago
4371.3.24	A couple of cleanups. A couple of cleanups. left_parent/right_parent didn't show a real-world win, so they have been removed. The code is quite a bit clearer without them.	John Arbash Meinel	15 years ago
4371.3.23	Using PySet_Add in the inner loop is a little better, not Using PySet_Add in the inner loop is a little better, not tremendously, though	John Arbash Meinel	15 years ago
4371.3.22	The slowdown was that _to_cleanup wasn't getting reset. The slowdown was that _to_cleanup wasn't getting reset. The other slow part of the pyrex code was that we weren't properly skipping to linear_dominator. Now .pyx is 2x faster than .py.	John Arbash Meinel	15 years ago
4371.3.21	Attempt to make things better by splitting out left_parent a... Attempt to make things better by splitting out left_parent and right_parent into individual attributes (to avoid extra type checks, etc.) However, this seems to make things much worse rather than better.	John Arbash Meinel	15 years ago
4371.3.20	A few tweaks to the pyrex version. A few tweaks to the pyrex version. It is now 'as fast' as the python one for heads() checks on bzr ancestry. It is about 2s faster for heads() checks for annotating NEWS.	John Arbash Meinel	15 years ago
4371.3.19	Get an initial pyrex implementation. Get an initial pyrex implementation. Initial results show it to actually be slightly slower than pure python.	John Arbash Meinel	15 years ago

Older »