Close

# Overcoming python's recursion limit.

In most programming languages, you can segfault your program by going too deep in a recursive function.

In CPython (the reference implementation that you are probably using), recursion is limited to a fixed number of consecutive recursive calls. The default maximum recursion can be checked by calling sys.getrecursionlimit(). On my machine, this returns 1000 for python 2 and 2000 for python 3.

If you exceed this depth, you will get a runtime error. e.g.:

$ipython Python 3.6.2 (default, Jul 17 2017, 16:44:45) Type 'copyright', 'credits' or 'license' for more information IPython 6.1.0 -- An enhanced Interactive Python. Type '?' for help. In [1]: def factorial(n): ...: if n <= 1: ...: # punting on the negative factorial question here ... ...: return 1 ...: return n * factorial(n - 1) ...: In [2]: factorial(2000) --------------------------------------------------------------------------- RecursionError Traceback (most recent call last) <ipython-input-2-c744509a6378> in <module>() ----> 1 factorial(2000) <ipython-input-1-55dcfb97cc6c> in factorial(n) 2 if n <= 1: 3 return n ----> 4 return n * factorial(n - 1) ... last 1 frames repeated, from the frame below ... <ipython-input-1-55dcfb97cc6c> in factorial(n) 2 if n <= 1: 3 return n ----> 4 return n * factorial(n - 1) RecursionError: maximum recursion depth exceeded in comparison  You can change the recursion depth by calling sys.setrecursionlimit with whatever you need the limit to be. However, if you are willing to use Cython you can completely bypass the recursion limit by using C function calling semantics. Let me explain by way of example: $ ipython
Python 3.6.2 (default, Jul 17 2017, 16:44:45)
IPython 6.1.0 -- An enhanced Interactive Python. Type '?' for help.

In [2]: %%cython
...: cpdef factorial(n):
...:     if n <= 1:
...:         return 1
...:     return n * factorial(n - 1)
...:


This version of factorial is not limited by the python interpreter’s rules regarding recursion. In Cython, if a function is declared a cdef or cpdef, Cython will generate a C function and that function will be called with the C calling convention for your machine within Cython code.

Note: cdef functions are ‘C only’ – they cannot be called from python. On the other hand, Cython will generate two version of cpdef functions. One pure C and a Python wrapper for it so you can call it from python code.

This Cython version of factorial will happily compute large factorials:

In [3]: factorial(2000)

Out[3]: 331627509245063324117539338057632403828111720810578039457193543706038077905600822400273230859732592255402352941225834109258084817415293796131386633526343688905634058556163940605117252571870647856393544045405243957467037674108722970434684158343752431580877533645127487995436859247408032408946561507233250652797655757179671536718689359056112815871601717232657156110004214012420433842573712700175883547796899921283528996665853405579854903657366350133386550401172012152635488038268152152246920995206031564418565480675946497051552288205234899995726450814065536678969532101467622671332026831552205194494461618239275204026529722631502574752048296064750927394165856283531779574482876314596450373991327334177263608852490093506621610144459709412707821313732563831572302019949914958316470942774473870327985549674298608839376326824152478834387469595829257740574539837501585815468136294217949972399813599481016556563876034227312912250384709872909626622461971076605931550201895135583165357871492290916779049702247094611937607785165110684432255905648736266530377384650390788049524600712549402614566072254136302754913671583406097831074945282217490781347709693241556111339828051358600690594619965257310741177081519922564516778571458056602185654760952377463016679422488444485798349801548032620829890965857381751888619376692828279888453584639896594213952984465291092009103710046149449915828588050761867924946385180879874512891408019340074625920057098729578599643650655895612410231018690556060308783629110505601245908998383410799367902052076858669183477906558544700148692656924631933337612428097420067172846361939249698628468719993450393889367270487127172734561700354867477509102955523953547941107421913301356819541091941462766417542161587625262858089801222443890248677182054959415751991701271767571787495861619665931878855141835782092601482071777331735396034304969082070589958701381980813035590160762908388574561288217698136182483576739218303118414719133986892842344000779246691209766731651433494437473235636572048844478331854941693030124531676232745367879322847473824485092283139952509732505979127031047683601481191102229253372697693823670057565612400290576043852852902937606479533458179666123839605262549107186663869354766108455046198102084050635827676526589492393249519685954171672419329530683673495544004586359838161043059449826627530605423580755894108278880427825951089880635410567917950974017780688782869810219010900148352061688883720250310665922068601483649830532782088263536558043605686781284169217133047141176312175895777122637584753123517230990549829210134687304205898014418063875382664169897704237759406280877253702265426530580862379301422675821187143502918637636340300173251818262076039747369595202642632364145446851113427202150458383851010136941313034856221916631623892632765815355011276307825059969158824533457435437863683173730673296589355199694458236873508830278657700879749889992343555566240682834763784685183844973648873952475103224222110561201295829657191368108693825475764118886879346725191246192151144738836269591643672490071653428228152661247800463922544945170363723627940757784542091048305461656190622174286981602973324046520201992813854882681951007282869701070737500927666487502174775372742351508748246720274170031581122805896178122160747437947510950620938556674581252518376682157712807861499255876132352950422346387878954850885764466136290394127665978044202092281337987115900896264878942413210454925003566670632909441579372986743421470507213588932019580723064781498429522595589012754823971773325722910325760929790733299545056388362640474650245080809469116072632087494143973000704111418595530278827357654819182002449697761111346318195282761590964189790958117338627206088910432945244978535147014112442143055486089639578378347325323595763291438925288393986256273242862775563140463830389168421633113445636309571965978466338551492316196335675355138403425804162919837822266909521770153175338730284610841886554138329171951332117895728541662084823682817932512931237521541926970269703299477643823386483008871530373405666383868294088487730721762268849023084934661194260180272613802108005078215741006054848201347859578102770707780655512772540501674332396066253216415004808772403047611929032210154385353138685538486425570790795341176519571188683739880683895792743749683498142923292196309777090143936843655333359307820181312993455024206044563340578606962471961505603394899523321800434359967256623927196435402872055475012079854331970674797313126813523653744085662263206768837585132782896252333284341812977624697079543436003492343159239674763638912115285406657783646213911247447051255226342701239527018127045491648045932248108858674600952306793175967755581011679940005249806303763141344412269037034987355799916009259248075052485541568266281760815446308305406677412630124441864204108373119093130001154470560277773724378067188899770851056727276781247198832857695844217588895160467868204810010047816462358220838532488134270834079868486632162720208823308727819085378845469131556021728873121907393965209260229101477527080930865364979858554010577450279289814603688431821508637246216967872282169347370599286277112447690920902988320166830170273420259765671709863311216349502171264426827119650264054228231759630874475301847194095524263411498469508073390080000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000


In fact, on my laptop it will compute 20,000 and even 200,000. I won’t include the output of factorial(200000) here, it is a 973,351 digit number!

## Tail Call Optimization

Since recursive functions grow the stack, the usefulness of recursive functions is limited to computations which are short enough to not require more memory than is available on the stack. This can be a serious limitation. As a result, many compilers support what is known as ‘Tail Call Optimization’. Basically, if the last statement in your function body is a function call, you can re-use the stack frame. This is because you are not using the result of the function call in the body of your function. This optimization can be made for any tail call, but the main use case is for ‘tail recursive’ functions. Recursive functions are tail recursive if the last statement is a recursive call to the function.

Our factorial function is not tail recursive because the last line:

return n * factorial(n - 1)


uses the return value of factorial. We can rewrite factorial to be tail recursive:

In [1]: %%cython
...: def factorial(n):
...:     return _factorial(n, 1)
...:
...: cdef _factorial(n, a):
...:     if n <= 1:
...:         return a
...:     return _factorial(n - 1, n * a)


Since we must make the last statement the recursive call, we need to pass the state around as a second parameter (known as an accumulator). This might seem a little clunky, but the benefits are worth it. You can recurse as deep as you want in constant stack space. Due to the awkwardness of the accumulator parameter it is common to make a wrapper for the recursive function to hide the accumulator as we have done.

When I started playing around with this, I was pretty hopeful that you could enjoy the benefits of tail call optimization by using Cython since it is generating real C functions and gcc (and most C compilers) implement tail call optimization. Unfortunately, even though our code ends with a tail call, the generated C code does not and so we don’t get the benefits of tail call optimization even if we make our cdef functions tail recursive. Cython is updated frequently, so it is possible that this will be fixed in the future.