[cfe-dev] Constexpr evaluation speed

[David Rector via cfe-dev]

One more vein of thought re constexpr evaluation speed, the new constexpr interpreter, and in particular how the C++ standard could be tweaked to allow us to use existing LLVM optimizations to really turbocharge code with heavy constexpr usage.

Caveat: my knowledge is mostly about the AST; I know very little about the ABI/calling conventions/LLVM etc, which are implicated here.

Motivation: it seems to me that constexpr evaluations will get very complex in the future, and we should plan for it.  In particular, I have been fighting a bit of a battle on the SG7 list to ensure the reflection + injection facilities will be sufficiently general to allow most design patterns to be rendered obsolete via metafunctions, and I think we came to a rough agreement that this is indeed a worthy and viable goal.  

But this will involve some very complex metaprogramming.  Indeed, constexpr programming may well become more complex than non-constexpr programming once reflection + injection get involved, and in fact *that would be the ideal*.  Let the user automate the tasks which make C++ programming complex, but via customizable libraries from which they can pick and choose and modify, rather than endless fixed appendages to the language.

But as users constexpr all the things, so must compilers consider how best to constexpr all their optimizations.  Why rewrite optimizations specific to constexpr functions which have already been written into LLVM?

The RFC for the new constexpr intepreter has some good discussion, and seems to present an opportunity.  In particular I’m looking at the comments here: 
https://lists.llvm.org/pipermail/cfe-dev/2019-July/062807.html <https://lists.llvm.org/pipermail/cfe-dev/2019-July/062807.html>, and in particular this response to Richard from Nandor:

> >  * The current APValue representation is extremely bloated. Each instance is 72 bytes, and every subobject of an object is stored as a distinct APValue, so for instance a single char[128] variable will often occupy 9288 bytes of storage and incurs 128 distinct memory allocations.
>  Arrays and structures will be stored in a compact, contiguous form in memory, so we can save a lot of space here.

Suppose we go a short step further and store that data in compliance with the ABI just like it were run-time data (with the proper alignment, offsets etc of subobjects).  

Then, suppose this were permissible code:

```
// foo.h
struct A { constexpr A() {} … };
constexpr int foo(A a);

// foo.cpp
constexpr int foo(A a) { 
  // Complicated functions, lots of loops etc. which
  // can be well-optimized by LLVM…
}

// main.cpp (must be compiled AFTER foo.cpp, or error)
#include "foo.h"
template<int N> class Dummy {};
int main() {
  constexpr A a{};
  Dummy<foo(a)> d;
}
```

Upon being called to evaluate `foo(a)`, the interpreter would determine that this function has already been fully compiled into binary, *and* was marked constexpr (so we know there is no funny business in the definition), and therefore it can simply call that function directly, in whatever manner LLVM would call such a function, i.e. passing the raw data in accordance with the ABI.

This would seemingly allow us to benefit from optimizations written only for lowered versions of that code, for particularly-complicated constexpr code that warrant it — the user need only a) put the definitions in separate cpps/libraries, and b) ensure they are built before any translation units which depend on them.  For trivial constexpr evaluation, nothing need change.

Thoughts?  Is this remotely viable/sensible?

Dave

> On Mar 2, 2021, at 11:28 PM, David Rector <davrecthreads at gmail.com> wrote:
> 
> 
> 
>> On Mar 2, 2021, at 10:03 PM, Richard Smith <richard at metafoo.co.uk <mailto:richard at metafoo.co.uk>> wrote:
>> 
>> On Tue, 2 Mar 2021 at 16:46, David Rector via cfe-dev <cfe-dev at lists.llvm.org <mailto:cfe-dev at lists.llvm.org>> wrote:
>> I was reading http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1240r1.pdf <http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1240r1.pdf>, which describes proposed reflection features currently under advanced consideration.  
>> 
>> On page 5, the authors give their rationale for defining `reflexpr(x)` to return an object of an opaque builtin type called `meta::info`, which is useful only when passed to other builtin functions able to access its various properties for different reflected entities.  This design is favored over the alternative of defining it to return an AST-class-like object specific to each reflected entity (ReflectedDeclaration, ReflectedStatement etc.).  
>> 
>> In other words given this design the user must write e.g. `meta::parameters_of(reflexpr(somefunc))` instead of e.g. `reflexpr(somefunc)->parameters()`.
>> 
>> One rationale the authors give for this choice is that they found that accessing subobjects of a constexpr class object is significantly slower than accessing values which are not subobjects, all else being equal.  The authors present an example on pp 5-6.  
>> 
>> I tried to reproduce this example and their results in CompilerExplorer, and was mostly just shocked at the apparent orders-of-magnitude-differences between GCC (by far the fastest), Clang, and MSVC (by far the slowest) at both constant-evaluation tasks.
>> 
>> Example A (NB `f()` deals only with complete objects):
>> https://godbolt.org/z/TM5Wb6 <https://godbolt.org/z/TM5Wb6>
>> 
>> Example B (same as A, except `f()` now defined to dig through subobjects to get the data it needs):
>> https://godbolt.org/z/5dPq3W <https://godbolt.org/z/5dPq3W>
>> 
>> Results of 5 trials:
>>                     __Compilation_times__
>> A
>> GCC  (1793B):   337,   408,   325,   435,   242 ms
>> Clang (218B):  9361,  5173,  4066,  5698,  4263 ms
>> MSVC  (306B): 21850, 24616, 24957, 24925, 32323 ms
>> 
>> B
>> GCC  (2295B):   471,   319,   403,   309,   323 ms
>> Clang (218B): 17073, 15540, 17281, 13385, 18540 ms
>> MSVC   (n/a): always >60k ms
>> 
>> Takeaways:
>> 1. Clang performs constant evaluation 10-50 times slower than GCC.
>> 2. While Clang performs B ~3 times slower than it performs A, it is not clear that GCC is likewise affected by having to dig through subobjects (if it is, the effect is slight).
>> 
>> Questions:
>> 1. What am I missing?  Are there flags which might improve clang’s performance?  Is GCC somehow gaining an unfair advantage?  (Potential clue: note that the executable is the same small size, 218 bytes, for each of clang’s results, but it is larger and differs for GCC’s results…meaningful?)  
>> 
>> Yes, GCC is "cheating" (you're not testing what you think you are). GCC memoizes constant evaluations (at least when it's correct to do so). Here's a slightly modified version of your A that doesn't permit memoization: https://godbolt.org/z/3q3TYr <https://godbolt.org/z/3q3TYr>
>> 
>> 5 trials with that and GCC (1793B):  12340, 11106, 10204, 9983, 10771ms
>> 
>> (Times for Clang and MSVC seem similar to your measurements.) I don't know if compiler explorer uses the same machines for all compiles, or if all compilers are built in fully-optimized mode, but if so, that suggests that Clang is about 2x faster than GCC for this particular testcase.
>> 
>> 2. Given the "constexpr all the things" zeitgeist, and the constant evaluation speeds GCC has apparently realized, should the design of ExprConstant.cpp/APValue/etc. be reconsidered?
>> 
>> Ignoring the part about GCC, yes. We have a -fexperimental-new-constant-interpreter flag that enables a new interpreter, which was built to be substantially faster. Unfortunately it's not complete yet (the version in trunk doesn't support any looping constructs yet) but the early indications are very promising.
>> 
>> 3. If ExprConstant.cpp etc were overhauled for optimal speed, will it still ultimately be true that programs which dig through subobjects of compile time objects are necessarily slower than equivalent programs which deal only with complete compile-time objects?
>> 
>> I don't think that would necessarily be the case. I've also mentioned that in committee, but "one compiler can do X" doesn't necessarily translate into "everyone will do X", and folks representing other compilers have indicated they expect the more "bare-bones" approach will remain faster for their implementations.
>> 
> 
> Thanks Richard, very thorough answer; that adjustment does indeed turn the tables on GCC.  Clang takes the gold, and that’s before even the fancy new interpreter is deployed.  Looking forward to it.
> 
> Assuming like machines are running the various compilers on CompilerExplorer, I think we can all hazard a pretty good guess at this point which particular compiler really benefits from keeping things as bare bones as possible ;)
> 
> Dave

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20210304/5c82cd4d/attachment-0001.html>