Wrap up casts into macro queso_cast

[pbauman]

After seeing #212 reported, we should remove all direct calls to static_cast and dynamic_cast and wrap them up in a macro so that dynamic_cast is done in debug mode (to catch bad casts) and static_cast otherwise (for performance). This is what libMesh and GRINS does and has worked very well.

[RhysU]

How far down in the compute loops are such casts used? Needing to macro this up for performance has a design smell to it.

[pbauman]

No idea, but wanted to get the thought down before I forgot about it. "Design smell" means "too much design" or "needs better design than macro" or other?

[RhysU]

Feels odd to need to perform a dynamic cast in a critical compute loop. I'd expect the expensive cast to always be done safely outside the loop.

Admittedly they can be relatively expensive (just tested locally on GCC 4.4.6 using source from http://tinodidriksen.com/2010/04/14/cpp-dynamic-cast-performance/):

OPERATION                                 |    MINTICK    AVGTICK  STDTICK |  RELATIVE
----------------------------------------- --------------------------------------------
reinterpret_cast known-type:              |    7017200    7022472     7468 |     1.00
virtual function + reinterpret_cast:      |   11023060   11077007   129030 |     1.57
member variable  + reinterpret_cast:      |    8021136    8045386    47212 |     1.14
dynamic_cast same-type-base       success |    7008016    7046708    49169 |     1.00
dynamic_cast same-type-level1     success |   36055128   36457611   732101 |     5.19
dynamic_cast same-type-level2     success |   38296548   38307029    10689 |     5.45
dynamic_cast same-type-level3     success |   36052380   36100992   113992 |     5.14
dynamic_cast level1-to-base       success |    6015188    6018479     3274 |     0.85
dynamic_cast level2-to-base       success |    6008468    6014180     4758 |     0.85
dynamic_cast level3-to-base       success |    6014668    6039942    56148 |     0.86
dynamic_cast level2-to-level1     success |   76261584   76997832  1035831 |    10.96
dynamic_cast level3-to-level1     success |  113149636  113461864   488122 |    16.15
dynamic_cast level3-to-level2     success |   76140680   77068236  1543087 |    10.97
dynamic_cast onebase-to-twobase   fail    |   28047192   28108865   142782 |     4.00
dynamic_cast onelevel1-to-twobase fail    |   73099688   73127042    47443 |    10.41
dynamic_cast onelevel2-to-twobase fail    |  179229428  179483756   596535 |    25.55
dynamic_cast onelevel3-to-twobase fail    |  282686416  283363397   870670 |    40.35

[RhysU]

Aight, grain of salt on that benchmark from having read some more critique of it. Comments about why the dynamic cast in the critical section still hold.

[roystgnr]

http://tinodidriksen.com/2010/04/14/cpp-dynamic-cast-performance/):

That source stuffs everything into main(); thanks to the lack of exported functions, the lack of function calls with possible side effects in those loops, and the lack of multiple inheritance, I wouldn't be too surprised if it was legal for the C++ compiler to optimize away much of the work in the heavyweight casts there. 10x and 20x slowdown may be a lower limit.

[RhysU]

Agreed @roystgnr. Sorry for the noise. Still, rather than macroing around things, I'll soapbox one last time that it feels weird if that sort of macro produces an appreciable change. If you need zero-cost polymorphism down in the guts, then CRTP. But stomaching what that means to a design is unpleasant and I'd still just recommend casting higher up in the loops.

[roystgnr]

Agreed on the design red flag. For the record, the motivation behind libmesh_cast was not "we have a bunch of dynamic_casts in inner loops", it was "we want to turn every dynamic_cast failure into a libmesh_error() exception so it's easier to debug". And then once we had to wrap our casts in inline functions anyway, and we had a static_cast fallback path for !defined(LIBMESH_HAVE_RTTI) anyway, it seemed sensible enough to use that same path in opt mode.

[RhysU]

Ah. That no-RTTI mode makes sense. Soapboxing withdrawn.