C++ modules are slated to be the biggest change in C++ since its inception. The design of modules has several essential goals in mind:

  1. Top-down isolation - The “importer” of a module cannot affect the content of the module being imported. The state of the compiler (preprocessor) in the importing source has no bearing on the processing of the imported code.
  2. Bottom-up isolation - The content of a module does not affect the state of the preprocessor in the importing code.
  3. Lateral isolation - If two modules are imported by the same file, there is no “cross-talk” between them. The ordering of the import statements is insignificant.
  4. Physical encapsulation - Only entities which are explicitly declared as exported by a module will be visible to consumers. Non-exported entities within a module will not affect name lookup in other modules (barring some possible strangeness with ADL. Long story…)
  5. Modular interfaces - The current module design enforces that for any given module, the entire public interface of that module is declared in a single TU called the “module-interface unit” (MIU). The implementation of subsets of the module interface may be defined in different TUs called “partitions.”

If you’ve been hoping for modules for as long as many have, you’ll note that “compilation speed” is missing from this list. Nevertheless, this is one of the biggest promises of modules. The possible speedup from modules is merely a consequence of the above design aspects.

This author can identify several aspects of C++ compilation that can greatly benefit from the design of modules. In order of most-to-least-obvious:

  1. Tokenization caching - Because of TU isolation, tokenizing a TU can be cached when the module is later imported into another TU.
  2. Parse-tree caching - Same as above. Tokenization and parsing are some of the most expensive operations in compiling C++. In my own tests, parsing can consume up to 30% of compilation time for files with a large preprocessed output.
  3. Lazy re-generation - If foo imports bar, and we later make changes to the implementation of bar, we may be able to omit recompilation of foo. Only changes to the interface of bar necessitate recompilation of foo.
  4. Template specialization memoization - This one is a bit more subtle and may take more work to implement, but the potential speedups are enormous. In short: A specialization of a class or function template appearing in the module interface unit can be cached and loaded from disk later.
  5. Inline function codegen caching - Codegen results for inline functions (including function templates and member functions of class templates) can be cached and later re-loaded by the compiler backend.
  6. Inline function codegen elision - extern template allows the compiler to omit performing codegen for function and class templates. This is hugely beneficial for the linker when it performs de-dup. Modules may allow the compiler to perform more extern template-style optimizations implicitly.

All in all, it’s looking pretty good, yeah?

But there’s something missing. A horrible, terrible, no-good, very bad flaw.

Remember the… Fortran?

Fortran implemented a module system that bears a slight resemblance to the design proposed for C++. A few short months ago, a paper p1300 was written by the SG15 group for review in San Diego. As far as I can tell, the paper was not discussed nor reviewed by any relevant eyes.

The gist of it is this:

  1. We have module foo and module bar, defined by foo.cpp and bar.cpp respectively.
  2. bar.cpp has an import foo; statement.
  3. How do we make sure that import foo will resolve when compiling bar.cpp? The current design and implementations require there to exist what is known as the “binary module interface” (abbreviated as BMI) defined for foo. The BMI is a file on the filesystem that describes the exported interface of the module foo. I’ll call that BMI foo.bmi for now. The extension isn’t important.
  4. Creation of foo.bmi is a byproduct of the compilation of foo.cpp. When compiling foo.cpp, the compiler will emit a foo.o and foo.bmi. As a consequence of this design, foo.cpp must be compiled before bar.cpp!

If alarm bells aren’t ringing already, let me discuss the way we currently work using header files:

  1. We have a “module” foo defined by foo.cpp and foo.hpp, and a “module” bar defined by bar.cpp and bar.hpp. Easy to understand.
  2. bar.cpp contains an #include <foo.hpp> preprocessor statement.
  3. How do we make sure that #include <foo.hpp> resolves when compiling bar.cpp? It’s simple: Make sure foo.hpp is present in a directory on the header search path list. We do not need to do any additional pre-processing.
  4. There is no ordering requirement between the compilation of the “modules” foo and bar. They can be processed in parallel.

Parallelization is probably the single most important aspect of increasing build performance. At this point, it isn’t even something you think about when you are optimizing your build because it is already there.

Modules change that. The importing of a module creates a “happens-before” dependency where #include did not. (I discuss this ordering in my Building Like (a) Ninja post).

The consequences of this design were explored very recently by Rene Rivera in Are modules fast?.

Spoiler alert: No. Or, more accurately: It’s subtle, but mostly no. The current module implementation used in that paper is extremely primitive but is still a valuable benchmark to understand what modules may look like performance-wise. Expectedly, as hardware parallelism increases, headers’ lead over modules becomes more and more pronounced. There is also a relationship between the DAG-depth (i.e. The length of the chain of modules that import each other). As this depth increases, modules grow slower and slower, while headers remain fairly constant for even “extreme” depths approaching 300).

A Sisyphean Scanning Task

Suppose I have this source file:

import greetings;
import std.iostream; int main() { std::cout << greeting::english() << '\n';
}

This is pretty simple. Since we import some modules, we will need to compile greetings and std.iostream before we can compile this file.

So, let’s do that…

Uh…

How?

We’ve been given a source file with two imports. That’s it. We don’t have anything else. Where is greetings defined? We need to find a file that contains a module greetings; statement.

This file located on the other side of the galaxy, talk.cpp, looks promising:

module; #ifdef FROMBULATE
#include <hello.h>
#endif

#ifndef ABSYNTH
export module something.pie;
#endif

import std.string; export namespace greeting { std::string english(); }

It defines that greeting::english function that we want. But how do we know that this is the right file? It doesn’t contain a module greetings; line!

But it does. Sometimes. It turns out when we compile with -DFROMBULATE, then the file hello.h is pasted into the source file. What’s in there?

#ifdef __SOME_BUILTIN_MACRO__
# define MODULE_NAME greetings
#else // Legacy module name
# define MODULE_NAME salutations
#endif

export module MODULE_NAME;

Oh.

Oh no.

This is fine. This is fine… Don’t worry. All we need to do is… run the preprocessor to check if the file comes out as module salutations or module greetings.

This is okay, but… There are 4,201 files that could define modules that can be imported, and any one of them could be module greetings;

Also, we can’t use our own implementation of the preprocessor: We need to run exactly the preprocessor that will be compiling this code. See that __SOME_BUILTIN_MACRO__? We have no clue what that is. If we don’t get it exactly right, the compilation will fail, or, even worse, we may miscompile the file.

So what can we do? We could cache the names of all the modules after preprocessing all the files, right? Well, where do we store that mapping? And what happens when we want to compile with a different compiler that results in a different mapping? What if we add new files that need to be scanned? Do we need to search every directory that contains these thousands of source files every time we build, just to check if any modules were added, removed, or renamed? On systems where process startup and/or filesystem access is not cheap, these costs will add up.

Possible Solutions

These two problems are distinct but related in that I (and many others) believe that one change to the modules design will fix them both: Module interface unit locations must be deterministic.

There are two alternative ideas to enforce this:

  1. Force MIU filenames to derive from the module’s name. This mimics the design of header filenames being directly related to how they are found from an #include directive.
  2. Provide a “manifest” or “mapping” file that describes the filepath to an MIU based on the module name. This file will need to be user-provided, or we are back in the scanning problem.

With MIU lookup deterministic and easily defined, we can then go to the next essential step: The BMI of a module must be lazily generated.

The compilation ordering between TUs will kill module adoption dead in its tracks. Even relatively shallow DAG depths are much slower than the equivalent with header files. The only answer is that TU compilation must be parallelizable, even in the face of importing other TUs.

In this respect, C++ would be best to mimic Python’s import implementation: When a new import statement is encountered, Python will first find a source file corresponding to that module, and then look for a pre-compiled version in a deterministic fashion. If the pre-compiled version already exists and is up-to-date, it will be used. If no pre-compiled version exists, the source file will be compiled and the resulting bytecode will be written to disk. The bytecode is then loaded. If two interpreter instances encounter the same un-compiled source file at the same time, they will race to write the bytecode. The race doesn’t matter, though: They will both come to the same conclusion and write the same file to disk.

In order to facilitate parallel compilation of TUs in the DAG, C++ modules must be implemented in the same way. Ahead-of-time BMI compilation is a non-starter. Instead, a compiler should lazily generate the BMI when it first encounters an import statement for the module in question. The build system should not concern itself with BMIs at all.

All of this can only work if the location of an MIU is deterministic for the compiler.

I Have Little Hope

There was a recent upset on the Twitter-verse. The Pre-Kona mailing was posted on January 25th. Amongst the many papers posted you will find p1427, Concerns about module toolability. Amongst its authors and contributors are names of build system and tooling engineers from around the industry. Am I appealing to authority here? Yes I am, but I feel that these are some of the most qualified people to provide feedback on module toolability.

This paper was born from the concerns of many tool authors and collaborators (more than what is named on the paper itself, including myself) who have felt that their concerns about modules have been ignored for months and years.

People outside of SG15 have been keen to shoot down discussion on the issues with module toolability, claiming that SG15 does not have the necessary implementation experience to make useful statements regarding modules.

SG15 has only had face-to-face meetings. The last meeting, in San Diego, was useless as the chair was absent and people were too busy getting caught up since the prior meetings to have any useful discussions. With no SG15 meetings outside of those at the official WG21 convenings, the members thereof have difficulty staying up-to-date and collaborating on work. In addition, many times that SG15 has attempted to raise issues they have been shot down as their work is considered “out-of-scope” for the C++ language.

A Tweet about the pre-Kona mailings spawned discussion of C++ modules and p1427. Questions were raised about who to trust regarding module toolability.

This discussion culminated in an eventual call for SG15 to “STFU” unless they can provide code samples that prove the problems they outline. This is a request for code that cannot be implemented in any current compiler and cannot be implemented in any current build system. Even if these were to exist, the request is for proving a negative: A task which cannot be done empirically. As such, this request for code is a goal that cannot be met.

The issues were not discussed. The issues were not disproven. No one even mentioned the problems outlined in p1427. We are told to simply trust some big names to know better than we do (an appeal to authority).

People backing the current modules design have not proven that modules work at scale, yet also demand proof from SG15 that they do not work at scale. Existing “modules” deployments do not use the current design, and do not use the automated module scanning that would be required for use with real build systems in the wild.

If modules are merged and it turns out that they cannot be implemented in a well-performing and flexible fashion, people will not use modules. If a broken modules proposal is merged into C++, it may be irrecoverable and C++ will never see the promises of modules realized.

Is it possible for the current modules proposal to be implemented successfully? I can’t answer with a definitive “no”, but me and many others feel that there are significant issues that need to be addressed.

But, judging by the behavior of others, it may seem that it doesn’t matter what SG15 thinks: They are being shot down at every turn by people with very little experience in C++ tooling, and the SG15 chair is completely absent through this entire discussion. Anything SG15 does is declared “unsubstantiated” and “out-of-scope.”

I was afraid to call out this behavior: I’m not keen on interpersonal conflict. Nevertheless, I’m more afraid that C++ will end up with a permanently useless modules design.