Oct 04, 2019

Beating std::visit Without Really Trying

One of my current projects is a sum type implementation for the D programming language. It bills itself as having "zero overhead compared to hand-written C," and while there are sound theoretical reasons to think that's true, I'd never actually sat down and verified it empirically. Was I really getting the optimal performance I thought I was?

I was also curious to see how my efforts stacked up next to C++17's <variant> and D's Algebraic, especially after reading a few blog posts and reddit comments about the challenges faced by the C++ implementation. Could D, and my library, SumType, in particular, do better?


To answer these questions, I designed a simple program and implemented it in C, in C++, and in D with both Algebraic and SumType, using the most natural and idomatic code for each language and library. My goal was to find out how well a production-grade, optimizing compiler would handle each sum type implementation in typical use. The compilers I chose were clang 8, clang++ 8, and ldc 1.17.0, all of which use LLVM 8 for their backends. All programs were compiled with optimization level -O2.

Each test program does the following things:

  • Defines 10 empty struct types, named T0 through T9.
  • Defines a sum type with 10 members, one of each of those types.
  • Defines a function that takes an instance of that sum type as an argument, and returns a unique integer for each possible type the sum type can hold.

For illustration's sake, here's the C++ version, slightly abridged:

#include <variant>

// Source: https://en.cppreference.com/w/cpp/utility/variant/visit
template<class... Ts> struct overloaded : Ts... { using Ts::operator()...; };
template<class... Ts> overloaded(Ts...) -> overloaded<Ts...>;

struct T0 {};
// ...
struct T9 {};

using example_variant = std::variant<T0, T1, T2, T3, T4, T5, T6, T7, T8, T9>;

int do_visit(example_variant v)
    return std::visit(overloaded {
        [](T0 val) { return 3; },
        [](T1 val) { return 5; },
        [](T2 val) { return 8; },
        // ...
        [](T9 val) { return 233; },
    }, v);

After compiling each program, I used objdump to disassemble them, and compared the generated assembly for each version of do_visit.

All of the code used, along with the commands used for compilation and disassembly, are available on Github in the repository pbackus/variant-comparison.



The C version uses a union with an enum tag as its sum type implementation, and a switch statement for its "visit" function. It's the straightforward, obvious implementation you'd use if you were writing the code by hand, without worrying about making it generic, so it makes a good baseline to compare the other versions to.

00000000000000a0 <do_visit>:
  a0:    cmp    $0xa,%edi         # check if type index is in-bounds
  a3:    jae    b0
  a5:    movslq %edi,%rax
  a8:    mov    0x0(,%rax,4),%eax # get result from global array
  af:    retq   

  # Error path
  b0:    push   %rax
  b1:    mov    $0x0,%edi
  b6:    mov    $0x0,%esi
  bb:    mov    $0x0,%ecx
  c0:    mov    $0x45,%edx
  c5:    callq  ca                # __assert_fail

As you might expect, the compiler is able to optimize the entire thing down into a single array lookup. It's so compact, in fact, that the code for calling libc's assertion-failure function takes up more space than the actual logic.


The C++ version uses std::variant and std::visit, with the overloaded helper template from cppreference.com to allow passing a set of labmda functions as a visitor. This is standard-library code, so we should expect it to be as well-optimized as the best and brightest C++ developers around can make it.

0000000000000000 <do_visit(std::__1::example_variant)>:
   0:    sub    $0x18,%rsp
   4:    mov    %rdi,%rax
   7:    mov    %rdi,0x8(%rsp)
   c:    shr    $0x20,%rax
  10:    mov    $0xffffffff,%ecx
  15:    cmp    %rcx,%rax        # check if type index is in-bounds
  18:    je     38
  1a:    mov    %rsp,%rcx
  1d:    mov    %rcx,0x10(%rsp)
  22:    lea    0x10(%rsp),%rdi
  27:    lea    0x8(%rsp),%rsi
  2c:    callq  *0x0(,%rax,8)    # call function pointer in global array
  33:    add    $0x18,%rsp
  37:    retq   

  # Error path
  38:    callq  3d               # __throw_bad_variant_access

The main thing to notice here is that, unlike in the C version, the compiler is not able to inline the individual visitor functions. Instead, it generates an indirect call to a function pointer stored in a global array. (The individual lambda functions, of course, all consist of a single mov followed by a return.)

I'm not enough of a C++ expert to understand the implementation of std::visit, so I can only guess why this is the case. Maybe the optimizer just gives up after too many layers of template-metaprogramming gunk? Regardless, it's bad news for fans of zero-cost abstractions, since there's a clear overhead here compared to the C assembly.

[UPDATE: Reddit user matthieum suggests that the function pointers themselves are likely to blame.]

D, with Algebraic

This version uses the sum type implementation from D's standard library, Phobos. Rather than create a dedicated sum type, the Phobos developers opted to make Algebraic a thin wrapper around Variant, D's equivalent of C++17's std::any. That choice turns out to have far-reaching repercussions, as we're about to see.

0000000000000000 <int dalgebraic.do_visit(ExampleAlgebraic)>:
   0:    jmpq   5 # int std.variant.visitImpl!(true, ExampleAlgebraic, ...)

0000000000000000 <int std.variant.visitImpl!(true, ExampleAlgebraic, ...)
   0:    push   %rbx
   1:    sub    $0x10,%rsp
   5:    cmp    0x0(%rip),%rdi    # check for uninitialized Variant
   c:    je     217               # if so, go to error path
  12:    mov    %rdi,%rbx

  # This part is repeated 10 times, once for each type
  15:    movq   $0x0,0x8(%rsp)    # prepare arguments for VariantN.handler
  1e:    lea    0x8(%rsp),%rdi
  23:    xor    %esi,%esi
  25:    xor    %edx,%edx
  27:    callq  *%rbx             # VariantN.handler: get TypeInfo for current type
  29:    mov    0x8(%rsp),%rsi
  2e:    mov    0x0(%rip),%rdi    # get TypeInfo for T0
  35:    callq  3a                # Object.opEquals: compare TypeInfos
  3a:    mov    %eax,%ecx
  3c:    mov    $0x3,%eax         # load return value for T0 into %eax
  41:    test   %cl,%cl           # check if Object.opEquals returned true
  43:    jne    211               # if so, go to return
 ...:    ...    ...

 # After checking for T9
 20f:    je     283               # if none of the types matched, assert(false)
 211:    add    $0x10,%rsp
 215:    pop    %rbx
 216:    retq   

 # Error path
 217:    mov    0x0(%rip),%rdi    # get ClassInfo for VariantException
 21e:    callq  223               # _d_allocclass: allocate VariantException
 223:    mov    %rax,%rbx
 226:    mov    0x0(%rip),%rax    # initialize VariantException vtable
 22d:    mov    %rax,(%rbx)
 230:    movq   $0x0,0x8(%rbx)
 238:    mov    0x0(%rip),%rax    # initialize VariantException
 23f:    movups 0x50(%rax),%xmm0
 243:    movups %xmm0,0x50(%rbx)
 247:    movups 0x10(%rax),%xmm0
 24b:    movups 0x20(%rax),%xmm1
 24f:    movups 0x30(%rax),%xmm2
 253:    movups 0x40(%rax),%xmm3
 257:    movups %xmm3,0x40(%rbx)
 25b:    movups %xmm2,0x30(%rbx)
 25f:    movups %xmm1,0x20(%rbx)
 263:    movups %xmm0,0x10(%rbx)
 267:    lea    0x0(%rip),%rdx    # get message for VariantException
 26e:    mov    $0x2f,%esi
 273:    mov    %rbx,%rdi
 276:    callq  27b               # VariantException.__ctor
 27b:    mov    %rbx,%rdi
 27e:    callq  283               # _d_throw_exception
 283:    lea    0x0(%rip),%rsi    # get error message for assert
 28a:    mov    $0x4b,%edi
 28f:    mov    $0xa1c,%edx
 294:    callq  299               # call D runtime assert function

The good part is that ldc is able to inline the lambdas into the body of visitImpl. The bad part is, well, everything else.

Because Algebraic is implemented using Variant, it has to rely on runtime type information (TypeInfo) to check the type of the current value. And because it uses TypeInfo, rather than an integer index like the C and C++ versions, these checks can't be condensed down into a single array lookup, or even a jump table. Finally, as if that wasn't enough, each TypeInfo comparison requires a virtual method call to Object.opEquals.

Beyond these obvious pessimizations, the other thing that stands out is the sheer volume of the generated code. It's an order of magnitude larger than both the C and C++ versions, with significant bloat in both the normal and error paths. Not only is this bad in isolation, since it puts unnecessary pressure on the instrution cache, it also means that potential optimization opportunities exposed by inlining the C and C++ versions of do_visit won't be available to D code using Algebraic.

D, with SumType

This version uses my own implementation of a sum type in D. SumType does not rely at all on runtime type information, and instead uses D's powerful compile-time reflection and metaprogramming capabilities to try to generate code as close to the C version as possible. Let's see how well it does.

0000000000000000 <int dsumtype.do_visit(ExampleSumType)>:
   0:    push   %rax
   1:    movsbq 0x10(%rsp),%rax    # get type index from stack
   7:    cmp    $0xa,%rax          # check if type index is in-bounds
   b:    jae    19
   d:    lea    0x0(%rip),%rcx     # load address of global array
  14:    mov    (%rcx,%rax,4),%eax # get result from global array
  17:    pop    %rcx
  18:    retq   

  # Error path
  19:    lea    0x0(%rip),%rdx     # get error message
  20:    mov    $0x4ca,%edi
  25:    mov    $0x9,%esi
  2a:    callq  2f                 # __switch_error

As it turns out, it's almost exactly the same as the C version: a single array lookup, plus some code for error handling. In fact, as far as I can tell, the only difference has to do with the way the function argument is passed in registers—the C version is able to avoid spilling anything to the stack, whereas this version has a push and a pop at the beginning and end, respectively.

Still, I think it's fair to say that this is a true zero-cost abstraction, and that the claim of "zero overhead compared to hand-written C" is justified.

[UPDATE: Reddit user ais52 explains that the purpose of the push/pop pair is to adjust the alignment of the stack pointer.]


If creating a zero-overhead generic sum type is so easy that even I can do it, why doesn't C++ have one? Are the libc++ developers a bunch of dummies?

No, of course not—but they are working with a significant handicap: C++.

The source code for variant is terrifyingly complex. In order to get the results shown above, the authors have had to use every template metaprogramming trick in the book, and then some. It's clearly the result of immense effort by some very skilled programmers.

By contrast, the source code of SumType is almost embarrassingly simple. The reason match, SumType's equivalent of std::visit, is able to optimize down to the same code as a C switch statement is that it literally is a switch statement, with a static foreach loop inside to generate the cases.

To be honest, I wasn't even thinking about optimization when I coded SumType. I wanted to make sure the interface was nice, and figured I could worry about performance later. But D makes it so easy to write simple, straightforward code that the compiler understands, I ended up falling into the pit of success and beating std::visit entirely by accident.