[dandotway]

Question for JIT experts: JS and Python are extremely hard to optimize because they both allow redefining anything at any time, yet V8 crushes Python by an order of magnitude in many benchmarks[1]:

         All times in seconds (lower is better)
  
  benchmark          Node.js     Python 3    Py takes x times longer
  ==================================================================
  regex-redux          5.06         1.34     ** Py3 is faster (PCRE C)
  pidigits             1.14         1.16     0.02
  reverse-complement   2.59         6.62     2.56
  k-nucleotide        15.84        46.31     2.92
  binary-trees         7.13        44.70     6.27
  fasta                1.91        36.90     19.3
  fannkuch-redux      11.31       341.45     30.19 (wut?)
  mandelbrot           4.04       177.35     43.9  (srsly?)
  n-body               8.42       541.34     64.3  (no numpy fortran cheat?)
  spectral-norm        1.67       112.97     67.65 (Python for Science[TM])

(If Python is allowed to call fast C code (PCRE) for regex-redux, I don't see why Python shouldn't be allowed to call fast Fortran BLAS/etc for n-body, but rules are rules, I guess. V8 doesn't cheat at spectral-norm, it's 100% legit JS.)

Both ecosystems have billions invested by corporations worth trillions; bottomless money exists to make Python faster. So why isn't Python faster?

V8's tactics include dynamically watching which loops/calls get run more than (say) 10,000 times and then speculatively generating[2] native machine instructions on the assumption the types don't change ("yep, foo(a,b) is only called with a and b both float64, generate greased x86-64 fastpath"), but gracefully falling back if the types later do change ("our greased x86-64 'foo(float64,float64)' routine will be passed a string! Fall back to slowpath! Fall back!"). Why doesn't Python do this? Is it because Google recruited the only genius unobtainium experts who could write such a thing? Google is a massive Python user, too.

[1] https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

[2] https://ponyfoo.com/articles/an-introduction-to-speculative-...

EDIT: HN commenter Jasper_ perhaps has the answer in another post[3]: "The current CPython maintainers believe that a JIT would be far too much complexity to maintain, and would drive away new contributors. I don't agree with their analysis; I would argue the bigger thing driving new contributors away is them actively doing so. People show up all the time with basic patches for things other language runtimes have been doing for 20+ years and get roasted alive for daring to suggest something as arcane as method dispatch caches. The horror!"

[3] https://news.ycombinator.com/item?id=30047289#30050248

[igouy]

> If Python is allowed to call fast C code (PCRE) for regex-redux, I don't see why Python shouldn't be allowed to call fast Fortran BLAS/etc for n-body…

The regex and arbitrary precision number libraries were allowed for regex-redux and pidigits because some language implementations wrapped those libraries rather than implement their own thing.