remove more unnecessary C-API usages

[scoder]

• list comprehensions are actually faster than repeated calls to PyList_Append()
• Py_XDECREF() works nicely on PyObject* (and the C compiler will usually drop the NULL check)
• isinstance(x, builtin) translates directly to a PyBuiltin_Check() call

[eriknw]

Thanks again!

I want to look more closely at a few of these changes so that I may learn from them :)

[mrocklin]

Whats the status on this? Can it be merged?

[eriknw]

Yeah, I'm sorry for keeping this PR lingering for so long. I really do appreciate your help, @scoder.

I think the majority of the changes in the PR are great and should be merged without question. Some of the changes, however, undo some of the performance tweaks I had made that were demonstrably faster. For example, I found that list comprehensions were slower when using data that I consider representative of typical usage.

Still, it is unreasonable for you, me, or anybody to manually test all these changes to check for performance regressions. benchtoolz (see #22) is usable, and I plan to upload it to github today. This will let us use a common suite of benchmarks that we can easily test against multiple versions of the same function. It also allows for rapid turnaround when a Cython file gets modified, because there is one file per function, and it gets compiled automatically.

Regarding this PR, I think we can merge it, and the onus will be on me to show that a particular tweak is worth it by adding it to benchmarkz (thus allowing others to test and modify the variations too).

I don't think cytoolz is a normal Cython project. For starters, it's generally not recommended to completely reproduce a package in Cython (although I think projects like toolz can be notable exceptions to this rule). If we can reproducibly get a 3% performance gain by using a slightly uglier statement, in general I would choose the 3% gain. toolz is where we keep the pretty code :)

Regarding isinstance(x, builtin) versus PyBuiltin_Check(), I know they produce the exact same C code, but I was undecided which to use. Using PyBuiltin_Check() is explicit in that it can't be optimized further. Most of my other uses of the C API were because I observed a performance increase. Again, the onus will be on me to show that a particular call to the C API is actually faster.

Unless there are comments to the contrary, I plan to merge this tomorrow. Thanks again, scoder, and sorry for the long delay. My hope is that benchtoolz will be worth the delay!