Attached patch removes all uses of the C "register" keyword in the CPython sources.

The benchmark suite doesn't show any significant speedup or slowdown generally:

Report on Linux fsol 3.8.0-27-generic #40-Ubuntu SMP Tue Jul 9 00:17:05 UTC 2013 x86_64 x86_64
Total CPU cores: 4

### django_v2 ###
Min: 0.672909 -> 0.659464: 1.02x faster
Avg: 0.673821 -> 0.660258: 1.02x faster
Significant (t=32.72)
Stddev: 0.00071 -> 0.00060: 1.1869x smaller

### fastunpickle ###
Min: 0.578535 -> 0.593218: 1.03x slower
Avg: 0.578700 -> 0.593440: 1.03x slower
Significant (t=-185.83)
Stddev: 0.00011 -> 0.00014: 1.1950x larger

### float ###
Min: 0.309679 -> 0.302615: 1.02x faster
Avg: 0.312457 -> 0.304713: 1.03x faster
Significant (t=4.33)
Stddev: 0.00326 -> 0.00232: 1.4039x smaller

### go ###
Min: 0.602035 -> 0.578129: 1.04x faster
Avg: 0.604634 -> 0.581584: 1.04x faster
Significant (t=12.50)
Stddev: 0.00253 -> 0.00325: 1.2855x larger

### nqueens ###
Min: 0.309602 -> 0.316767: 1.02x slower
Avg: 0.310222 -> 0.317734: 1.02x slower
Significant (t=-21.22)
Stddev: 0.00057 -> 0.00055: 1.0295x smaller

### richards ###
Min: 0.184306 -> 0.178887: 1.03x faster
Avg: 0.184969 -> 0.180215: 1.03x faster
Significant (t=8.40)
Stddev: 0.00076 -> 0.00101: 1.3380x larger

### silent_logging ###
Min: 0.080233 -> 0.073160: 1.10x faster
Avg: 0.080594 -> 0.073170: 1.10x faster
Significant (t=69.28)
Stddev: 0.00024 -> 0.00001: 21.7820x smaller

### threaded_count ###
Min: 0.234882 -> 0.226954: 1.03x faster
Avg: 0.241049 -> 0.231458: 1.04x faster
Significant (t=3.04)
Stddev: 0.00567 -> 0.00420: 1.3519x smaller

### unpack_sequence ###
Min: 0.000046 -> 0.000048: 1.05x slower
Avg: 0.000046 -> 0.000049: 1.06x slower
Significant (t=-299.85)
Stddev: 0.00000 -> 0.00000: 1.0256x larger

The following not significant results are hidden, use -v to show them:
call_method, call_method_slots, call_method_unknown, call_simple, chaos, fannkuch, fastpickle, formatted_logging, iterative_count, json_dump_v2, json_load, meteor_contest, nbody, normal_startup, pathlib, pidigits, raytrace, regex_compile, regex_effbot, regex_v8, simple_logging, spectral_norm, startup_nosite, telco.

(amusingly, the two Python executables generated by gcc - with or without register - have exactly the same size, although their md5 hashes differ)

+1 for removal of the register keyword.

I hope that nowadays, no compiler needs register hint to be able to optimize correctly a function. GCC knows better than me which variables must be "optimized".

+1

> have exactly the same size

Binaries contain paddings.

> although their md5 hashes differ

Timestamps.

I'm interesting if someone builds CPython with more simple compilers (i.e. PCC [1]).

[1] http://en.wikipedia.org/wiki/Portable_C_Compiler

> Serhiy Storchaka added the comment:
> 
> > have exactly the same size
> 
> Binaries contain paddings.

I used the "size" command, which gives you the exact size of the various sections
(rather than the physical file size).

+1  The era of "register" has passed.

Is PCC used by any Linux or BSD distro? BSD prefers Clang.

New changeset e7f6cef7a4cc by Antoine Pitrou in branch 'default':
Issue #18722: Remove uses of the "register" keyword in C code.
http://hg.python.org/cpython/rev/e7f6cef7a4cc

Committed. I don't think clang or MSVC would bother more than gcc does.

I did a review of some of the changes in <https://bugs.python.org/issue26130#msg258346>. In particular, there are a couple of comments /* Help register allocation */ which were mangled to /* Help allocation */.

But it is not clear if the leftover variable copies are still helping (in which case the comments should be reverted), or if the copies and comments can be removed.

Agree that comments like /* Help allocation */ are confused and misleading.