D’s Newfangled Name Mangling

Posted on

Rainer Schuetze is the creator and maintainer of Visual D, the D plugin for Visual Studio. Recently, he implemented a new name mangling algorithm for the D frontend, which was released in DMD 2.077.0. In this post, he explains why it was needed and what it does.

What is symbol name mangling?

D embraces the separate compilation model that compiles D source code to object files and uses a linker to bind the object files to an executable binary file. This allows the reuse of precompiled object files and libraries, speeding up the build process. As the linker is usually one that’s also used for other languages with the same compilation model, e.g. C/C++ or Fortran, mixing object files from different languages is straightforward.

In an object file, a symbol name is assigned to each function or global variable, both when it is defined and when it is used via a call or access. The linker uses these names to connect references to definitions of the same name with only very bare knowledge about the symbol. For example, the symbol for this C function declaration,

extern(C) const(char)* find(int ch, const(char)* str);

does not tell the linker anything about function arguments or return type, as the C language uses the plain function name find as the symbol name (some platforms prepend a _ to the symbol). If you later change the order of the arguments to

extern(C) const(char)* find(const(char)* str, int ch);

but fail to update and recompile all source files that use the new declarartion, the linker will happily bind the resulting object files. In that case, the program is likely to crash, since a character passed to the function will be interpreted as a string pointer and vice versa.

D and C++ avoid this problem by adding more information to the symbol name, i.e. they encode into a symbol name the scope in which the symbol is defined, the function argument types, and the return type. Even if the linker does not interpret this information, linking fails with an undefined symbol error if the definitions used to build the object files don’t match. For example, the D function declaration

module test;
extern(D) const(char)* find(int ch, const(char)* str);

has a symbol name _D4test4findFiPxaZPxa, where _D is a prefix to identify the symbol as being generated from a D source symbol, 4test4find encodes the “fully qualified name” find in module test, and FiPxaZPxa describes the function type with an integer argument (designated by i) and the C-style string pointer type Pxa by just concatenating the encodings for argument types. Z terminates the function argument list and is followed by the encoding for the return type, again Pxa for a C-style string pointer. In contrast,

extern(D) const(char)* find(const(char)* str, int ch);

is encoded as _D4test4findFPxaiZPxa, making it a different symbol with the argument types reversed. The encoding ensures a normalized representation of types and scopes while also providing shorter symbols than minimal source code. This encoding is called “name mangling”.

Ed: Note that extern(C) and extern(D) are linkage attributes. When a function is declared in D without an explicit linkage attribute, extern(D) is the default.

In D, some function attributes are also mangled into the symbol name, e.g. @safe, nothrow, pure and @nogc. In theory, mangling could also cover parameter names, user defined attributes, or even contracts, but that is currently considered excessive.

Please note that even though name mangling can detect some mismatches in the binary interface of functions (i.e. how arguments are passed in registers or on the stack), it won’t catch every error; for example, structs, classes and other user defined types are mangled by name only, so that a change to their definition will still pass unnoticed by the linker.

The mangled name of a symbol is also available during compilation using the .mangleof property. This used to be exploited to provide type reflection of the symbol at compile time. This should no longer be necessary due to the introduction of new __traits that make this information accessible faster and more convenient, for example,



__traits(getFunctionAttributes, symbol);

Thus, usage of .mangleof is not recommended except for debugging purposes.

When reversing the mangling process in the “demangler”, all the encoded information is kept to make it available to the user, but that does not always yield correct D syntax. The first definition above demangles as

const(char)* test.find(int, const(char)*)

i.e. the module name test is added to the function name.

Template symbols

The two definitions of find shown above can coexist in D and C++, so name mangling is not only a way to detect errors at link time but also a necessity to represent overloads. It should at least contain enough information to distinguish different overloads of the same scoped identifier.

This becomes even more obvious when considering templates that usually instantiate different functions or variable definitions for each argument type. In D, the template instantiation information is added to the qualified name of a symbol.

Consider expression templates, a common example of meta programming used for delayed evaluation of expressions:

module expr;
struct Mul(X,Y)
    X x;
    Y y;
struct Add(X,Y)
    X x;
    Y y;

auto mul(X,Y)(X x, Y y) { return Mul!(X,Y)(x, y); }
auto add(X,Y)(X x, Y y) { return Add!(X,Y)(x, y); }

A function template is lowered by the compiler to an eponymous template:

template mul(X, Y)
    auto mul(X x, Y y) { return Mul!(X,Y)(x, y); }

The template name is part of the qualified function name, expr.mul!(X,Y).mul, and the auto return type is inferred to be Mul!(X,Y). This causes the symbol to reference the types X and Y three times. The demangled mangled name of an instantiation with types double and float of this template is

expr.Mul!(double,float) expr.mul!(double,float).mul(double,float)

The mangling process of DMD before version 2.077 walks the abstract syntax tree of the declaration and emits the mangled representation of the types whenever it is hit. Now consider stacking operations, e.g.

auto square(X)(X x) { return mul(x, x); }

auto len = square("var");
pragma(msg, len.square.mangleof);
// S4expr66__T3MulTS4expr16__T3MulTAyaTAyaZ3MulTS4expr16__T3MulTAyaTAyaZ3MulZ3Mul

pragma(msg, typeof(len).mangleof.length);
pragma(msg, len.square.mangleof.length);
pragma(msg, len.square.square.mangleof.length);
pragma(msg, len.square.square.square.mangleof.length);
pragma(msg, len.square.square.square.square.mangleof.length);
pragma(msg, len.square.square.square.square.square.mangleof.length);
pragma(msg, len.square.square.square.square.square.square.mangleof.length);

With DMD 2.076 or earlier, this displays 28u, 78u, 179u, 381u, 785u, 1594u, 3212u, showing exponential growth of the mangled symbol name length even though the expression in the source code just grows linearly. This happens because types like Mul!(Mul!(string, string), Mul!(string, string)) are combined and the mangling repeats their full representation every time they are referenced.

Create a chain of 12 calls to square above and the symbol length increases to 207,114. Even worse, the resulting object file for COFF/64-bit is larger than 15 MB and the time to compile increases from 0.1 seconds to about 1 minute. Most of that time is spent generating code for functions only used at compile time.

Voldemort types returned from template functions can be similar, as they carry the function signature including the template arguments as part of the type name. These can also show a dramatic increase in build times without generating as much code as in the example.

Symbol compression to the rescue

In early 2016, a couple of attempts were made to shorten these long symbols:

  • cut off symbol names if they exceed a given threshold, but append a checksum of the full symbol instead. This was already done with an MD5 hash when emitting symbols for the DigitalMars C compiler tool chain as the OMF object file format does not allow symbols longer than 255 characters. The downside to this is that these symbols can no longer be demangled, so that symbols in linker messages cannot be translated back into human digestible names.
  • apply binary compression to the symbol name. Standard techniques use part of the full symbol name as a dictionary to encode repetitions within the name. This is usually done by encoding a position-length pair using characters outside the normal identifier set. Again, this is already in use when DMD tries to fit symbols into the OMF limit of 255 characters (before applying the MD5 hash trick), but it also has shown some disadvantages: when using characters above the ASCII range, this interferes with UTF8 encoded characters that are also allowed as symbol characters in the D language. It can also break linker output as the console might misinterpret it as a locale-specific character encoding. Avoiding this by applying a binary to ASCII conversion like base64 to the symbols would obfuscate the actual symbol name even more.
  • extend the mangling grammar by allowing references to entities already encoded. This is similar to binary compression, but does not need to encode match length as the entities have terminators embedded into the grammar. The most prominent entities are types. This is the road C++ has taken, as it is also affected by the issues described here. In C++, name mangling is not standardised by the language, but by the compiler or the platform. GNU g++ uses Itanium C++ ABI mangling, which does a pretty good job with C++ code similar to the example above or in the Voldemort issue. Even though Microsoft’s Visual C++ can encode recurring types as well, it still generates very long names because it limits the encoding to the first 10 types of an argument list.

The first attempts at applying the latter scheme to the mangling of D symbols showed disappointing results. As it turned out, these implementations missed a subtle detail of the mangler in the DMD front end at that time; it reused cached representations of mangled type names to combine them to more complex types. This fails to find repetitions of the types from which a cached type name was built.

This is where I stepped in to create a proof-of-concept version of the mangling without these omissions. Early results were promising, so I looked for more opportunities to reduce symbol length:

  • with fully qualified names always containing the package and module names of a symbol, identifiers tend to appear often in a mangled name.
  • qualified names are likely to come from the same module or package, so it would be nice to encode them as a single entity.

The unit tests of the Phobos runtime library are benchmark candidates, as they contain a lot of symbols for template-heavy code. At the given time there were 127,172 symbols found in the map file of the Windows build. These were the results of the different manglings:

back references max length average length
none 416133 369
types 2095 157
types+identifiers 1263 128
types+identifiers+qualified names 1114 117

(This has been measured with the implementation at the time, which is not exactly the same as the final mangling; it used special characters for the different back reference types, but this turned out not to be a good idea. The D mangling is supposed to be the same on all platforms, but these characters will have a special meaning to the linker on one of them.)

It’s rather simple in DMD to determine the identity of identifiers and types, as the latter are merged according to their mangling anyway. Qualified names and their associated symbols turned out to introduce a number of complications, though. Namely, mangleFunc in core.demangle allows building a mangled name of a function from a fully qualified name given as a string function argument and a type specified as a template argument. Implementing this for run-time usage requires copying the full mangling machinery and introspection capabilities of the compiler, which is unrealistic. Considering the limited benefit shown by the above Phobos statistics, the idea of encoding qualified names was dropped.

Here are some details about the new mangling:

  • Back references are now encoded by the character Q, followed by the relative position of the original appearance of the same identifier or type. These positions are encoded with respect to base 26, with the last digit encoded by a lowercase letter and the other digits encoded by an uppercase letter. That way, most back references are 2 or 3 characters long, 4 in extreme cases. Using a different encoding for the last digit allows determining the end of a number without looking at the next character. This helps to avoid ambiguities. (The Itanium C++ ABI mangling uses base 36 encoding by combining numbers and letters, but need a termination character _.)
  • Counting encodable entities as in the C++ mangling would result in slightly shorter mangled names, but needs the mangler to keep a dynamic list of respective positions. The current demangler is designed not to allocate as long as the supplied output buffer is large enough.
  • Relative positions are chosen instead of absolute positions to allow prepending the _D prefix without having to re-encode the symbol. Some platforms also prepend an additional underscore, for which the relative positions are agnostic.
  • The mangling grammar sometimes allows types and identifiers at the same position, so a demangler needs to distinguish between the two even if given by a back reference. That’s why a lookup to the referenced position is necessary to continue demangling; an identifier always starts with a number, while a type always starts with a letter.
  • Using Q for back references grabs the last free letter used to encode types, but there is at least one type defined in the mangling grammar that is not supposed to appear in a mangling anyway (namely TypeIdent), so it can be resurrected if the necessity appears.

For example, the expression template type shown above now mangles as

pragma(msg, len.square.mangleof);
// S4expr__T3MulTSQo__TQlTAyaTQeZQvTQtZQBb
//                ^^   ^^     ^^ ^^ ^^ ^^^ decode to:
//                |    |      |  |  |  |
//                |    |      |  |  |  +- 3Mul
//                |    |      |  |  +---- S4expr__T3MulTAyaTAyaZ3Mul
//                |    |      |  +------- 3Mul
//                |    |      +---------- Aya
//                |    +----------------- 3Mul
//                +---------------------- 4expr

with a length of 39 instead of 78 without back references. The resulting sizes are 23, 39, 57, 76, 95, 114, 133 showing linear growth. The chain of 12 calls to square shrinks from 207,114 characters to 247, i.e. by a factor of more than 800.

Implementing mangleFunc mentioned above for the mangling with back referencing identifiers still is not obvious; while the fully qualified name is not supposed to contain any types (e.g. as a struct template argument) identifiers in the mangled name can appear again in the function type. This was solved by extending the demangler to use “Design by Introspection” (DbI) (as coined by Andrei Alexandrescu):

  • make the Demangle struct a template that parameterizes on a struct that supplies a couple of hooks
    struct NoHooks {}  // supports: static bool parseLName(ref Demangle); ...
    private struct Demangle(Hooks = NoHooks)
    Hooks hooks;
        // ...
        void parseLName()
            static if(__traits(hasMember, Hooks, "parseLName"))
                if (hooks.parseLName(this))
                // normal decode...
  • create a hook that replaces a reoccurring identifier with the appropriate back reference
    struct RemangleHooks
        char[] result;
        size_t[const(char)[]] idpos;
        // ...
        bool parseLName(ref Demangler!RemangleHooks d)
            // flush input so far to result[]
            if (d.front == 'Q')
                // re-encode back reference...
            else if (auto ppos = currentIdentifier in idpos)
                // encode back reference to identifier at *ppos
                idpos[currentIdentifier] = currentPos;
            return true;
  • combine the qualified name and the type as before (core.demangle is still capable of decoding it) and run it through the hooked demangler
    char[] mangleFunc(FuncType)(const(char)[] qualifiedName)
        const(char)mangledQualifiedName = encodeLNames(qualifiedName);
        const(char)mangled = mangledQualifiedName ~ FuncType.mangleof;
        auto d = Demangle!RemangleHooks(mangled, null);
        d.mute = true; // no demangled output
        return d.hooks.result;

Is the new mangling sound?

The back references encoded into the mangling extend the existing mangling. Unfortunately, the latter had ambiguities reported to the D issue tracking system, with more of these likely yet to be uncovered. The demangler in core.demangle rejected about 3% of the unmodified symbols from the Phobos unit tests, while 15% were demangled only partially.

It’s tough to verify the soundness of an addition to an already complex and fragile definition, as a change to the mangling would need an update to the tooling (demangler, debuggers). Anyway, it was a good opportunity to get rid of these, too.

So scrutiny of the existing definition was required. To do this mechanically, the mangling specification from the web site was converted into a grammar digestible by the bison parser generator. Bison can create LALR(1) parser tables, which basically means that, while scanning a mangled symbol, looking at a character and its successor is enough to determine whether the character adds to a previous entity or starts a new one. When conflicts are reported when processing a grammar, they might be resolvable with a larger context, but they can also hint at actual problems or undesirable complexity. Adding pseudo-tokens representing handcrafted parser control flow can avoid these conflicts.

This gist shows a grammar for the D mangling scheme without the back references. It still has a couple of conflicts when run through Bison, one of which was determined to be an actual ambiguity in the definition. Adding back references to
the grammar doesn’t add any conflicts.

In addition, core.demangle was fixed to work for all symbols but those exposing the known ambiguities.


Some of the implementations in std.traits used the mangling of a symbol to introspect compile-time properties, for example, to determine the linkage. This was done using a simplified demangler. With the introduction of back references, these
didn’t work any more except for simple symbol names. Using a solution as for core.mangleFunc is feasible, but can slow down compilation considerably as the demangling needs to be executed via CTFE. Fortunately, new __traits have been added which cover all information that can be found in the mangling.

While most users will not notice any changes to their programs other than smaller object and executable file sizes, the new mangling can be a breaking change to external tools like the linker or a debugger. These will continue to work, but until they are updated, be prepared to eventually see the new mangled names for a little while instead of the demangled ones.

Thanks to Mike Parker, Walter Bright and Steven Schveighoffer for review.

DCompute: Running D on the GPU

Posted on

Nicholas Wilson is a student at Murdoch University, studying for his BEng (Hons)/BSc in Industrial Computer Systems (Hons) and Instrumentation & Control/ Molecular Biology & Genetics and Biomedical Science. He just finished his thesis on low-cost defect detection of solar cells by electroluminescence imaging, which gives him time to work on DCompute and write about it for the D Blog.He plays the piano, ice skates, and has spent 7 years putting D to use on number bashing, automation, and anything else that he could make a computer do for him.

DCompute is a framework and compiler extension to support writing native kernels for OpenCL and CUDA in D to utilize GPUs and other accelerators for computationally intensive code. Its compute API drivers automate the interactions between user code and the tedious and error prone APIs with the goal of enabling the rapid development of high performance D libraries and applications.


This is the second article on DCompute. In the previous article, we looked at the development of DCompute and some trivial examples. While we were able to successfully build kernels, there was no way to run them short of using them with an existing framework or doing everything yourself. This is no longer the case. As of v0.1.0, DCompute now comes with native wrappers for both OpenCL and CUDA, enabling kernel dispatch as easily as CUDA.

In order to run a kernel we need to pass it off to the appropriate compute API, either CUDA or OpenCL. While these APIs both try to achieve similar things they are different enough that to squeeze that last bit of performance out of them you need to treat each API separately. But there is sufficient overlap that we can make the interface reasonably consistent between the two. The C bindings to these APIs, however, are very low level and trying to use them is very tedious and extremely prone to error (yay void*).
In addition to the tedium and error proneness, you have to redundantly specify a lot of information, which further compounds the problem. Fortunately this is D and we can remove a lot of the redundancy through introspection and code generation.

The drivers wrap the C API, providing a clean and consistent interface that’s easy to use. While the documentation is a little sparse at the moment, the source code is for the most part straightforward (if you’re familiar with the C APIs, looking where a function is used is a good place to start). There is the occasional piece of magic to achieve a sane API.

Taming the beasts

OpenCL’s clGet*Info functions are the way to access properties of the class hidden behind the void*. A typical call looks like

cl_foo* foo = ...; 
cl_int refCount;
clGetFooInfo(foo, CL_FOO_REFERENCE_COUNT, refCount.sizeof, &refCount,null);

And that’s not even one for which you have to call, to figure out how much memory you need to allocate, then call again with the allocated buffer (and $DEITY help you if you want to get a cl_program’s binaries).

Using D, I have been able to turn that into this:

struct Foo
    void* raw;
    static struct Info
        @(0x1234) int referenceCount;
    mixin(generateGetInfo!(Info, clGetFooInfo));

Foo foo  = ...;
int refCount = foo.referenceCount;

All the magic is in generateGetInfo to generate a property for each member in Foo.Info, enabling much better scalability and bonus documentation.

CUDA also has properties exposed in a similar manner, however they are not essential (unlike OpenCL) for getting things done so their development has been deferred.

Launching a kernel is a large point of pain when dealing with the C API of both OpenCL and (only marginally less horrible) CUDA, due to the complete lack of type safety and having to use the & operator into a void* far too much. In DCompute this incantation simply becomes

Event e = q.enqueue!(saxpy)([N])(b_res, alpha, b_x, b_y, N);

for OpenCL (1D with N work items), and

q.enqueue!(saxpy)([N, 1, 1], [1 ,1 ,1])(b_res, alpha, b_x, b_y, N);

for CUDA (equivalent to saxpy<<<N,1,0,q>>>(b_res,alpha,b_x,b_y, N);)

Where q is a queue, N is the length of buffers (b_res, b_x & b_y) and saxpy (single-precision a x plus y) is the kernel in this example. A full example may be found here, along with the magic that drives the OpenCL and CUDA enqueue functions.

The future of DCompute

While DCompute is functional, there is still much to do. The drivers still need some polish and user testing, and I need to set up continuous integration. A driver that unifies the different compute APIs is also in the works so that we can be even more cross-platform than the industry cross-platform standard.

Being able to convert SPIR-V into SPIR would enable targeting cl_khr_spir-capable 1.x and 2.0 CL implementations, dramatically increasing the number of devices that can run D kernel code (there’s nothing stopping you using the OpenCL driver for other kernels though).

On the compiler side of things, supporting OpenCL image and CUDA texture & surface operations in LDC would increase the applicability of the kernels that could be written.
I currently maintain a forward-ported fork of Khronos’s SPIR-V LLVM to generate SPIR-V from LLVM IR. I plan to use IWOCL to coordinate efforts to merge it into the LLVM trunk, and in doing so, remove the need for some of the hacks in place to deal with the oddities of the SPIR-V backend.

Using DCompute in your projects

If you want to use DCompute, you’ll need a recent LDC built against LLVM with the NVPTX (for CUDA) and/or SPIRV (for OpenCL 2.1+) targets enabled and should add "dcompute": "~>0.1.0" to your dub.json. LDC 1.4+ releases have NVPTX enabled. If you want to target OpenCL, you’ll need to build LDC yourself against my fork of LLVM.

Unit Testing In Action

Posted on

Mario Kröplin is a developer at Funkwerk AG, a German company whose passenger information system is developed in D and was recently highlighted on this blog. That post describes Funkwerk’s use of third-party unit testing frameworks and says, “the team recently discovered a way to combine xUnit testing with D’s built-in unittest, which may lead to another transition in their unit testing.” That’s Mario’s subject in this post.

There and Back Again

Ten years ago, programming in D was like starting over in our company. And, of course, unit testing was part of it right from the beginning. D’s built-in simple support made it easy to quickly write lots of unit tests. Until some of them failed. And soon, the failure became the rule. There’s always someone else to blame: D’s simple unit-test support is too simple. A look at Python reveals that the modules doctest and unittest live side by side in the standard library. We concluded that D’s unit test support corresponds to Python’s doctest, which means that there must be something else for the real unit testing.

Even back then, we immediately found such a unit testing framework in DUnit [An old D1 unit testing framework that you can read about at the old dsource.org – Ed.]. Thanks to good advice for xUnit testing, we were happy and content with this approach. At the end of life of D1, a replacement library for D2 was soon found. After a bumpy start, I found myself in the role of the maintainer of dunit [A D2 unit-testing framework that is separate from DUnit – Ed].

During DConf 2013, I copied a first example use of user-defined attributes to dunit. This allowed imitating JUnit 4, where, for example, test methods are annotated with @Test. By now, dunit imitates JUnit 5. So if you want to write unit tests in Java style, dunit is a good choice. But which D programmers would want to do that?

Recently, we reconsidered the weaknesses of D’s unit test support. Various solutions have been found to bypass the blockers (described in the following). On the other hand, good guidelines are added, for example, to use attributes even for unittest functions. So we decided to return to making use of D’s built-in unit test support. From our detour we retain some ideas to keep the test implementation maintainable.


Whenever a unit test fails at run time, the question is, why? The error message refers to the line number, where you find something like assert(answer == 42). But what is the value of answer if it isn’t 42? The irony is that this need is well understood. If you use a static assert instead, the error message reads like: static assert 54 == 42 is false. The fear of code bloat is the reason why you don’t get this automatically at run time.

If you look at the Language Reference, you will notice that the chapter Unit Tests covers primarily the special unittest function. It is assumed that assert is used for test verification, which is introduced in the chapter Contract Programming. In theory, it’s completely OK to reuse assert for test verification. Any failure reveals a programming error that must be fixed. In practice, however, test expectations are quite different from preconditions, postconditions, and invariants. While the expectations are usually specific (actual == expected) the contracts rather exclude specific values ​​(value != 0 or value !is null).

So there are lots of implementations of templates like assertEquals or test!"==". The problem shows up if you want to have the most helpful error messages: expected 42 but got 54. For this, assertEquals is too symmetrical. In fact, JUnit’s assertEquals(expected, actual) was turned into TestNG’s assertEquals(actual, expected). Even with UFCS (Uniform Function Call Syntax), it is not clear how a.assertEquals(b) should be used. From time to time, programmers don’t write the arguments in the intended order. Then the error messages are the opposite of helpful. They are misleading: expected 54 but got 42.

Fluent assertions avoid this symmetry problem: actual.should.eq(expected) or expect(actual).to.eq(expected) are harder to use incorrectly. Thanks to UFCS and lazy parameters, the implementation in D is no problem. The common criticism is “the natural language formulation is too verbose”, or just “too many dots”. Currently, however, this seems to be the only way to get the most helpful error messages.

The next problem is that string comparisons are seldom as simple as: expected foo but got bar. Non-printable characters or lengthy texts, such as XML or JSON representations, sabotage error messages that were meant to be helpful. This can be avoided by escaping special characters and by showing differences. Finally, this is what the fluent-asserts library does.

Test Execution

At large, we want to get as much information as possible from a failed test run. How many test cases fail? Which test cases fail? Does the happy path fail or rather edge cases? Is it worth addressing the failures, or is it better to undo the change? The approach of stopping on the first error is contrary to these needs. The original idea was to run the unit tests before the start of the actual program. By now, however, separate test runners are often used, which continue in case of a failure. To emphasize this, test expectations usually throw their own exceptions, instead of the unrecoverable AssertError. This change already shows how many test cases fail.

Finding out what’s tested in the failing test cases is more difficult. At best, there are corresponding comments for documented unit tests. But an empty DDoc comment, ///, is all that’s needed to include the body of the unittest function as an example in the documentation. In the worst case, the unit test goes on and on verifying this and that.

The idea of the Sentence Style For Naming Unit Tests is that the name of the test function describes the test case. In D, however, the unittest functions are anonymous. On the other hand, D has user-defined attributes so that you can even use strings for the test description instead of CamelCase names. unit-threaded, for example, shows these string attributes so that you get a good impression of the extent of the problem in case of a failure. In addition, unit-threaded satisfies the requirement to execute test cases selectively. For example, only the one problematic test case or all tests except those tagged as “slow”. It’s promising to use unit-threaded as needed. You let D run the unittest functions as long as they pass. Only for troubleshooting should you switch to unit-threaded. You have to be careful, however, to only use compatible features.

By the way: the parallel test execution (from it’s name, the main goal of unit-threaded) was quite problematic with the first test suite we converted. On the other hand, the speedup was just 10%.


The D compiler has built-in code-coverage analysis. The ratio of the lines executed in the test is often used as an indicator for the quality of the tests. (See: Testing in the D Standard Library) A coverage of 100% cannot be achieved, for example, if you have an assert(0). Lower thresholds for the coverage can always be achieved by cheating. The fact that the unittest functions are also incorporated in the coverage is questionable. Imagine that a single line that has not yet been executed requires a lengthy unit test. As a consequence, this new unit test could significantly raise the coverage.

In order to avoid such measurement errors, we decided from the beginning to extract non-trivial unit tests to separate modules. We place these in parallel to the src tree in a unittest directory. Test utilities are also placed in the unittest directory, so that reading the actual code is not encumbered by large version (unittest) sections. (We also have test directories for customer tests.) For the coverage, we only count the modules under src. Code-coverage analysis creates a report file for each module. For a summary, which we output at the end of each successful test run, we have written a simple script. By now, covered is a ready-made solution.

In order to fully exploit the code-coverage analysis, an unusual formatting is required, for example, for the short-circuit evaluation of expressions with &&, ||, and ?:. We hope that dfmt can be changed to reformat the code temporarily.


What can you do to prevent the test implementation from getting out of control? After all, test code is also code that needs to be maintained. Sometimes the test implementation is more obscure than the code being tested.

As a solution the xUnit patterns suggest a structuring of the test implementation as a Four-Phase Test: fixture setup, exercise system under test, result verification, fixture teardown. The term fixture refers to the test context. For JUnit, this is the test class with attributes that are available to all test methods. A method with the annotation @BeforeEach initializes the attributes. This is the fixture setup. Another method with the annotation @AfterEach implements the fixture teardown. All methods annotated with @Test focus on exercise and verification.

At first glance, this approach seems to be incompatible with D’s unittest functions. The unittest functions do not get automatic access to the attributes of a class, even if they are defined in the context of a class. On the other hand, one can mimic the approach, for example, by implementing the fixtures next to the unittest functions as a struct:

    Fixture fixture;
    scope (exit) fixture.teardown;
    (fixture.x * fixture.y).should.eq(42);

The test implementation can be improved by executing the fixture setup in the constructor (or in opCall(), since default constructors are disallowed in structs) and the fixture teardown in the destructor:

    with (Fixture())
        (x * y).should.eq(42);

The with (Fixture()) pulls the context, in which test methods are executed implicitly in JUnit, explicitly into the unittest function. With this simple pattern you can structure unit tests in a tried and trusted way without having to use a framework for test classes ever again.

Parameterized Tests

A parameterized test is a means to reuse a test implementation with different values ​​or with different types. Within a unittest function this would be no problem. Our goal, however, is to get as much information as possible from a failing test run. For which values ​​or which types does the test fail? unit-threaded provides support for parameterized tests with @Values ​​and @Types. If unit-threaded is not used to run the unittest functions, these test cases do not work at all.

With the new static foreach feature however, it is easy to implement parameterized tests without the support of a framework:

static foreach (i; 0 .. 2)
    static foreach (j; 0 .. 2)
        @(format!"%s + %s == 1"(i, j))
            (i + j).should.eq(1);

And if you run the failing test with unit-threaded, the descriptions of the failing test cases reveal the problem without the need to take a look at the test implementation:

0 + 0 == 1: expected 1 but got 0
1 + 1 == 1: expected 1 but got 2


D’s built-in unit test support works best when there are no failures. As shown, however, you do not need to change too much to be able to work properly in situations where you rely on helpful error messages. The imitation of a solution from another programming language is often easy in D. Nevertheless, one should reconsider such solutions from time to time.

If we had a wish, we would want separate libraries for expectations and for test execution. Currently, you get frameworks where not all features are great, or they are overloaded with alternative solutions. Such a separation should probably be supported by the Phobos runtime library. Currently, each framework defines expectations with its own unit test exceptions. In order to combine them, ugly interdependencies are required to match the exceptions thrown in one library to the exceptions caught in another library. A unit test exception in Phobos could avoid this problem.

The Making of ‘D Web Development’

Posted on

A long-time contributor to the D community, Kai Nacke is the author of ‘D Web Development‘ and the maintainer of LDC, the LLVM D Compiler. In this post, he tells the story of how his book came together. Currently, the eBook version is on sale for USD $10.00 as part of the publisher’s Back to School sale, as are ‘D Cookbook‘ by Adam Ruppe and ‘Learning D‘ by Michael Parker.

At the beginning of 2014, I was asked by Packt Publishing if I wanted to review the D Cookbook by Adam Ruppe. Of course I wanted to!

The review was stressful, but it was a lot of fun. At the end of the year came a surprising question for me: would I be willing to switch sides and write a book myself? Here, I hesitated. Sure, writing your own book is a dream, but is this at all possible on top of a regular job? The proposed topic, D Web Development, was interesting. Web technologies I knew, of course, but the vibe.d framework was for me only a large unit test for each LDC release.

My interest was awakened and I created a chapter overview, based solely on my experience as a developer and the online documentation of vibe.d. The result came out well and I was offered a contract. It came with an immediate challenge: I should set up a small project plan. How do you plan to write a book?!?

Without any experience in this area, I stuck to the following rules. For each chapter, I planned a little time frame. Each should include at least one weekend, for the larger chapters perhaps even two. I reserved some time for the Easter holiday, too. The first version of the book would therefore be ready at the beginning of July, when I started writing in mid-February.

Even the first chapter showed that this plan was much too optimistic. The writing went off quickly – as soon as I had something I could write about. But experimenting and testing took a lot of time. For one thing, I didn’t have much experience with vibe.d. There were sample programs that I wanted to develop Saturday to write about on Sunday. However, I was still searching for errors on Monday, without having written a single line!

On the other hand, there were still a few rough edges in vibe.d at the time, but I did not want to write that these would be changed or implemented in later versions of the library. So I developed a few patches for vibe.d, e.g. digest authentication. By the way, there were also new LDC releases to create. Fortunately, the LDC team had expanded, so I just took care of the release itself (thanks so much, folks!). The result was, of course, that I missed many of my milestones.

In May, the first chapters came back from the review process. Other content also had to be written, such as the text for the back of the book. In mid-December, the last chapter was finished and almost all review notes on the other chapters were incorporated. After a little Christmas break, the remaining notes were quickly incorporated and the pre-final version of each chapter was created in January. And then, on February 1, 2016, the news came that my book was now published. I’d done it! Almost exactly one year after I had started with the first chapter.

Was the work worth it? In any case, it was a very special experience. Would I do it again? Yes! Right now, I am playing with the idea of updating the book and expanding a chapter. Let’s see what happens…

The Evolution of the accessors Library

Posted on

Ronny Spiegel is a developer at Funkwerk AG, a German company whose passenger information system is developed in D and was recently highlighted on this blog. In this post, Ronny tells the story of the company’s open source accessors library, which provides a mechanism for users to automatically generate property getters and setters using D’s robust compile-time features.

A little bit of history.

We’ve always used UML tools to visualize the internal structure and document details of software. That’s true for me not only at Funkwerk, but also in the companies I worked before I joined the team here in Karlsfeld. One of the major issues of documentation is that at some point in time it will diverge from the actual implementation and become outdated. Additionally, if you have to support old versions of your components you will have to take care of old versions of your documentation as well.

The first approach to connecting code and model is to generate code from the model, which requires the model to reflect the current implementation. When I joined Funkwerk we were using ArgoUML to manage class diagrams which were used as input to generate code. Not only class or struct skeletons were generated (existing code was kept), but also methods to access members which were not even part of the model. In order to control whether a member should be accessible, annotations, similar to UDAs (User-Defined Attributes), were used as part of the member documentation. So it was very common for us to annotate a member with @Read or @Write even though it was only in the documentation. The tool which we used to generate code was powerful enough to create the implementation of these field accessor methods supported by annotations to synchronize access, or to automatically use invariants for pre- and post-conditions as well.

Anyway, the approach of using the model as a base for code generation always suffers from the same problem: it is very hard to merge models!

So we reversed the whole thing and decided to create documentation from code. We could still use code which had been generated before, but all the new classes had to be supplied with accessor functions. You can imagine that this was very annoying.

public class Journey
    private Leg[] legs_;

    public Leg[] legs()
	return this.legs_.dup;


(Yes, we’ve been writing Java and compiling as D.)

Code which was generated before still had these @Read and @Write annotations next to the fields. So I thought, “These look like UDAs. Why not just use those to generate the methods automatically?” I’d always wanted to use mixins and compile-time introspection in order to move forward with a more D-like development approach.

A first draft…

The very first version of the accessors library was able to generate basic read- and write-accessor methods using the allMembers trait, filtering by UDAs, and generating some basic code like:

public final Leg[] legs() { return this.legs_.dup; }

It works… Yes, it does.

We did not replace all existing accessor methods at once, but working on a large project at that time we introduced many of them. The automated generation of accessor methods was really a simplification for us.

…always has some issues.

The first implementation looked so simple – there must have been issues. And yes, there were. I cannot list all of them because I do not remember anymore, but some of these issues were:

Explicitly defined properties suppressed generated ones

We ran into a situation where we explicitly defined a setter method (e.g. because it had to notify an observer) but wanted to use the generated getter method. The result was that the defined setter method could be used but accessing the generated getter method (with the same name) was impossible.

According to the specification, the compiler places mixins in a nested scope and then imports them into the surrounding scope. If a function with the same name already exists in the surrounding scope, then this function overwrites the function from the mixin. So if there is a field with a @Read annotation and another explicitly defined mutating field accessor, then the @Read accessor is overwritten by the defined one.

The solution to this issue was rather simple. We had to use a string mixin to import the generated code into the class where it shall be used.


We have a guideline to avoid magic bools wherever possible and use much more verbose flags instead. So a simple attribute like:

private bool isExtraJourney_;


private Flag!”isExtraJourney” isExtraJourney_;

This approach has two advantages. Providing a value with Yes.isExtraJourney is much more verbose than just a true, and it creates a type. When there are two or more flags as part of a method signature, you cannot change the order of the flags (by accident) as you could with bools.

To generate the type of the return value (or in case of mutable access of the parameter) we used T.stringof, where T is the type of the field. Unfortunately, this does not work as expected for Flags.

Flag!”foo” fooFlag;

static assert(`Flag!”foo”`, typeof(fooFlag).stringof); // Fails!
static assert(`Flag`, typeof(fooFlag).stringof); // Succeeds!

Unit Tests

When using the mixin in private types defined in unit tests, a similar issue arose. Classes defined in unittest blocks have a prefix like __unittestL526_8. It was necessary to strip this prefix from the used type string.

Private Classes

While iterating over members of private classes, we stumbled across the issue that the allMembers (or derivedMembers) trait returns, in addition to __ctor, an unaccessible member called this. This issue remains unsolved.

The current implementation…

The currently released version covers the aforementioned issues, although there is still room for new features.

An example might be to provide a predicate which is then used for synchronizing access to the field. That was possible using the old version of the code generator by annotating it with @GuardedBy(“this”). Fortunately, we’ve advanced in our D coding skills and have moved away from Java code compiled with DMD to a more D-like style by using structs wherever we need value semantics (and we don’t have to deal with thousands of copies of that value). So at the moment, this doesn’t really hurt that much.

Another interesting (and still open issue) is to create accessors for aliased imported types. The generated code still refers to the real name of the type, which is then unknown to the compile unit where the code is mixed in.

…has room for improvement!

If you’re interested in dealing with this kind of problem and want to dive into CTFE and compile-time introspection, we welcome contributions!